feat: /add-audit skill — opt-in ncl-surface audit log on the guard seam

Ships the local audit log as an install skill, scoped to the ncl command
surface: one withAudit(dispatch) composition at the definition site covers
both transports and the grant-carrying approved replay; holds are recorded
as pending events correlated by approval id (recovered from the hold row),
so --correlation returns the whole gated chain. src/audit/ installs as a
domain-free leaf — schema v1, recursive key-pattern redactor (JSON string
args parsed so inner secret keys mask), NDJSON day-files under data/audit/,
fail-open + loud emit, post-write exporter hooks, boot writability refusal,
self-contained retention prune (boot + daily, unref'd timer). `ncl audit
list` is host + global-scope only (off the group-scope allowlist, fails
closed) and errors clearly on a disabled box. Core reach-ins: the dispatch
composition and the resource-barrel import, both guarded by a shipped
AST/behavior wiring test; guard conformance stays clean at boot and in CI.
Approval-lifecycle events, OneCLI credential holds, and channel/sender
events are later increments that attach as pure adds.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

.claude/skills/add-audit/REMOVE.md

@@ -0,0 +1,52 @@
+# Remove /add-audit
+
+Reverses every change the skill made. Safe to re-run even if some pieces are
+already gone. Run from the NanoClaw project root.
+
+## 1. Delete the copied files
+
+```bash
+rm -rf src/audit
+rm -f src/cli/dispatch.audit.ts src/cli/dispatch.audit.test.ts
+rm -f src/cli/resources/audit.ts
+rm -f src/audit-wiring.test.ts
+```
+
+## 2. Revert the dispatch composition
+
+In `src/cli/dispatch.ts`, delete (not comment out) the three edits the skill
+made:
+
+1. The import line: `import { withAudit } from './dispatch.audit.js';`
+2. The composition block (comment + `export const dispatch = withAudit(dispatchInner);`)
+3. Rename the dispatcher back — change `async function dispatchInner(` to
+   `export async function dispatch(`
+
+## 3. Unregister the resource
+
+Delete the `import './audit.js';` line from `src/cli/resources/index.ts`.
+
+## 4. Remove the settings
+
+Delete the `AUDIT_ENABLED` and `AUDIT_RETENTION_DAYS` lines from `.env`.
+
+## 5. Rebuild and restart
+
+```bash
+pnpm run build
+source setup/lib/install-slug.sh
+launchctl kickstart -k gui/$(id -u)/$(launchd_label)   # macOS
+# systemctl --user restart $(systemd_unit)             # Linux
+```
+
+## Day-files
+
+`data/audit/*.ndjson` are the operator's records and are deliberately left in
+place. To purge them too:
+
+```bash
+rm -rf data/audit
+```
+
+No dependency was added, nothing under `container/` was touched, and no DB
+schema changed — there is nothing else to undo.

.claude/skills/add-audit/SKILL.md

@@ -0,0 +1,211 @@
+---
+name: add-audit
+description: Add an opt-in local audit log for the ncl command surface — every dispatch outcome on both transports (host socket and container), including scope denials, approval holds, and approved replays, written as SIEM-shaped append-only NDJSON day-files under data/audit/. Read back with `ncl audit list`; plug exporters in via registerAuditHook. Off until AUDIT_ENABLED=true.
+---
+
+# /add-audit — Local Audit Log (ncl surface)
+
+Records one canonical audit event for every command that reaches the `ncl`
+dispatcher — human over the host socket or agent over the container transport
+— including denials. Both transports converge on the exported `dispatch`, so
+one composition covers the whole surface: there is no second door.
+
+```
+ncl (host socket) ──┐
+                    ├─→ dispatch = withAudit(dispatchInner) ─→ one NDJSON line
+container cli_request ┘        │                               data/audit/<UTC-day>.ndjson
+approved replay (grant) ───────┘
+```
+
+What one event carries: `actor` (`host:<os-user>` or the agent group),
+`origin` (transport, session, channel), dotted `action` from the guard
+catalog (e.g. `groups.config.add-mcp-server`), touched/attempted `resources`,
+`outcome` (`success · failure · denied · pending`), `correlation_id` (the
+approval id on gated chains — the pending event and the approved replay share
+it, so `--correlation <id>` returns the whole chain), and redacted `details`.
+
+Scope of this increment: the ncl dispatch surface only. Approval-lifecycle
+events (request/decision as their own events), OneCLI credential holds, and
+channel/sender events are later increments — they attach to `src/audit/` as
+pure adds (new `*.audit.ts` adapters and observer subscriptions).
+
+## Steps
+
+### 0. Pre-flight (idempotency)
+
+The apply is safe to re-run; every step below is guarded. Skip to
+**Enable** if all of these already hold:
+
+- `src/audit/` exists
+- `src/cli/resources/index.ts` contains `import './audit.js';`
+- `src/cli/dispatch.ts` contains `export const dispatch = withAudit(dispatchInner);`
+
+Before editing, verify the reach-in targets still exist: `src/cli/dispatch.ts`
+must contain `export async function dispatch(` (or the already-applied
+composition), and `src/cli/resources/index.ts` must be the resource barrel (a
+list of `import './<resource>.js';` lines). If either has moved, stop and
+adapt rather than guessing.
+
+### 1. Copy the payload
+
+From the NanoClaw project root:
+
+```bash
+cp -R "${CLAUDE_SKILL_DIR}/add/src/." src/
+```
+
+What lands (mirrors of the destination paths):
+
+- `src/audit/` — the domain-free leaf: event schema (`types.ts`), env config
+  (`config.ts`), redactor, NDJSON day-file store, emit seam, post-write hooks,
+  reader, boot/maintenance wiring, vocabulary — plus its tests.
+- `src/cli/dispatch.audit.ts` — the CLI adapter: `withAudit` middleware and
+  the actor/origin/resource mapping (+ `dispatch.audit.test.ts`).
+- `src/cli/resources/audit.ts` — the read-only `ncl audit` resource.
+- `src/audit-wiring.test.ts` — goes red if either core edit below is deleted
+  or drifts.
+
+### 2. Register the resource
+
+Append to `src/cli/resources/index.ts` (skip if the line is already present):
+
+```typescript
+import './audit.js';
+```
+
+### 3. Compose the dispatch middleware
+
+This is the skill's one functional reach-in, in `src/cli/dispatch.ts`. Three
+small edits:
+
+1. Add the import (next to the other `./` imports):
+
+   ```typescript
+   import { withAudit } from './dispatch.audit.js';
+   ```
+
+2. Rename the dispatcher declaration — change
+
+   ```typescript
+   export async function dispatch(
+   ```
+
+   to
+
+   ```typescript
+   async function dispatchInner(
+   ```
+
+3. Directly after that function's closing brace (before the
+   `registerApprovalHandler('cli_command', …)` block), add:
+
+   ```typescript
+   // Audit middleware (installed by /add-audit): the exported dispatch is the
+   // wrapped function, so both transports and the approved replay below all
+   // pass the one composition.
+   export const dispatch = withAudit(dispatchInner);
+   ```
+
+The composition must live at the definition site (not at the import sites):
+the approved-replay handler in the same file calls `dispatch(...)` too, and
+only the wrapped export covers it. `src/audit-wiring.test.ts` asserts exactly
+this shape via the TypeScript AST.
+
+Loading `dispatch.audit.ts` also boots the audit log: on an enabled box it
+asserts `data/audit/` is writable (refusing to start beats a silent audit
+gap), runs the boot retention prune, and arms an unref'd maintenance timer.
+Nothing else in core changes — no `index.ts` or `host-sweep.ts` edits.
+
+### 4. Enable
+
+Add the two settings to `.env` (idempotent — overwrite or append):
+
+```bash
+grep -q '^AUDIT_ENABLED=' .env && sed -i.bak 's/^AUDIT_ENABLED=.*/AUDIT_ENABLED=true/' .env && rm -f .env.bak || echo 'AUDIT_ENABLED=true' >> .env
+grep -q '^AUDIT_RETENTION_DAYS=' .env || echo 'AUDIT_RETENTION_DAYS=90' >> .env
+```
+
+`AUDIT_ENABLED` is the master switch — off (or absent) means `emitAuditEvent`
+is a no-op and `data/audit/` is never created. `AUDIT_RETENTION_DAYS`: day
+files strictly older than the horizon are hard-deleted (unlinked) at boot and
+once per UTC day; `0` = keep forever; unset = 90.
+
+### 5. Build and test
+
+Run `build` before the tests — it's what catches a missed copy or a drifted
+import path across the whole composed tree:
+
+```bash
+pnpm run build
+pnpm exec vitest run src/audit src/cli/dispatch.audit.test.ts src/audit-wiring.test.ts
+pnpm test
+```
+
+### 6. Restart the service
+
+```bash
+source setup/lib/install-slug.sh
+launchctl kickstart -k gui/$(id -u)/$(launchd_label)   # macOS
+# systemctl --user restart $(systemd_unit)             # Linux
+```
+
+### 7. Verify (runtime smoke)
+
+```bash
+ncl groups list                     # any command — this one is now an audit event
+cat data/audit/$(date -u +%F).ndjson | tail -1
+ncl audit list --limit 5
+```
+
+The last line of the day-file is the `groups.list` event you just caused.
+
+## Reading it back
+
+```
+ncl audit list [--actor <id>] [--action <name-or-prefix>] [--resource <id-or-type>]
+               [--outcome success|failure|denied|pending|approved|rejected]
+               [--since 7d|24h|30m|ISO] [--until …] [--correlation <approval-id>]
+               [--limit N] [--format ndjson]
+```
+
+Newest first, default limit 100. `--action groups.config` matches the whole
+dotted subtree. `--format ndjson` streams the stored lines verbatim — pipe to
+a file for SIEM import. On a disabled box the command errors with
+`audit log is disabled — set AUDIT_ENABLED=true` rather than returning an
+empty list that would read as "no actions happened".
+
+The resource is deliberately **not** on the group-scope allowlist: audit
+spans agent groups, so group-scoped agents are refused before the handler
+(fails closed). Host callers and `cli_scope: global` agents only.
+
+## Redaction, failure posture, exporters
+
+- Every event's `details` passes a recursive key-pattern redactor
+  (`/(token|secret|key|password|credential|auth|bearer)/i` → `[REDACTED]`,
+  ~2 KB per-value cap) at the single emit seam.
+- Fail-open + loud: a failed append is `log.error`'d and the audited action
+  proceeds. At boot, an enabled box refuses to start if `data/audit/` isn't
+  writable.
+- In-process exporters register via `registerAuditHook` (from
+  `src/audit/index.js`) — post-write hooks that fire only after the local
+  append succeeds, so an external system can never be ahead of the source of
+  truth. Ship one as its own `/add-*` skill; declare `/add-audit` as its
+  dependency.
+
+## Troubleshooting
+
+- **Host won't start, banner names the audit directory** — `AUDIT_ENABLED=true`
+  but `data/audit/` isn't writable. Fix permissions or disable audit.
+- **`audit log is disabled — set AUDIT_ENABLED=true`** — the read-back guard;
+  enable in `.env` and restart.
+- **An agent gets "CLI access is scoped to this agent group" for `ncl audit`**
+  — by design; grant the group `cli_scope: global` only if it should read
+  cross-group history.
+- **`pnpm test` on an enabled box adds a few events to today's day-file** —
+  expected: the base dispatch tests drive real dispatches. Harmless noise;
+  the audit test suites themselves write only to temp dirs or capture buffers.
+
+## Removal
+
+See [REMOVE.md](REMOVE.md) — reverses every change; existing day-files are
+left in place (they're the operator's records).

.claude/skills/add-audit/add/src/audit-wiring.test.ts

@@ -0,0 +1,136 @@
+/**
+ * Wiring tests for the /add-audit skill's two core edits — they go red if
+ * either edit is deleted or drifts:
+ *
+ *  1. src/cli/dispatch.ts — the exported dispatch must be the composed
+ *     `withAudit(dispatchInner)` (AST check: only the definition-site
+ *     composition covers both transports AND the in-module approved replay).
+ *  2. src/cli/resources/index.ts — the audit resource must register through
+ *     the real barrel (behavior check), stay OFF the group-scope allowlist,
+ *     and leave the guard conformance walk clean.
+ *
+ * Plus the emit invariant the module's discipline rests on: emitAuditEvent
+ * appears only in src/audit/ and *.audit.ts adapter files.
+ */
+import fs from 'fs';
+import path from 'path';
+import ts from 'typescript';
+import { describe, expect, it } from 'vitest';
+
+// Same production barrels the boot conformance walk sees (mirrors
+// src/guard/conformance.test.ts).
+import './cli/commands/index.js';
+import './modules/index.js';
+import './cli/delivery-action.js';
+import './cli/dispatch.js';
+
+import { commandGuard, GROUP_SCOPE_RESOURCES, lookup } from './cli/registry.js';
+import { grantContinuationGaps } from './guard-conformance.js';
+
+describe('audit resource registration (barrel wiring)', () => {
+  it('registers audit-list through the real resource barrel', () => {
+    const cmd = lookup('audit-list');
+    expect(cmd).toBeDefined();
+    expect(cmd?.action).toBe('audit.list');
+    expect(cmd?.access).toBe('open');
+  });
+
+  it('derives a guard-catalog entry for the audit command', () => {
+    const guard = commandGuard('audit-list');
+    expect(guard.action).toBe('audit.list');
+  });
+
+  it('is NOT on the group-scope allowlist — group-scoped agents fail closed', () => {
+    expect(GROUP_SCOPE_RESOURCES.has('audit')).toBe(false);
+  });
+
+  it('leaves the boot conformance walk clean with the audit resource registered', () => {
+    expect(grantContinuationGaps()).toEqual([]);
+  });
+});
+
+describe('dispatch composition (AST wiring)', () => {
+  const source = fs.readFileSync(path.resolve('src/cli/dispatch.ts'), 'utf8');
+  const sf = ts.createSourceFile('dispatch.ts', source, ts.ScriptTarget.Latest, true);
+
+  const hasExportModifier = (node: ts.HasModifiers): boolean =>
+    (ts.getModifiers(node) ?? []).some((m) => m.kind === ts.SyntaxKind.ExportKeyword);
+
+  let importsWithAudit = false;
+  let innerDeclaredUnexported = false;
+  let exportsWrappedDispatch = false;
+  let exportsUnwrappedDispatchFn = false;
+
+  sf.forEachChild((node) => {
+    if (
+      ts.isImportDeclaration(node) &&
+      ts.isStringLiteral(node.moduleSpecifier) &&
+      node.moduleSpecifier.text === './dispatch.audit.js'
+    ) {
+      const named = node.importClause?.namedBindings;
+      if (named && ts.isNamedImports(named) && named.elements.some((e) => e.name.text === 'withAudit')) {
+        importsWithAudit = true;
+      }
+    }
+    if (ts.isFunctionDeclaration(node) && node.name?.text === 'dispatchInner') {
+      innerDeclaredUnexported = !hasExportModifier(node);
+    }
+    if (ts.isFunctionDeclaration(node) && node.name?.text === 'dispatch' && hasExportModifier(node)) {
+      exportsUnwrappedDispatchFn = true;
+    }
+    if (ts.isVariableStatement(node) && hasExportModifier(node)) {
+      for (const decl of node.declarationList.declarations) {
+        if (
+          ts.isIdentifier(decl.name) &&
+          decl.name.text === 'dispatch' &&
+          decl.initializer &&
+          ts.isCallExpression(decl.initializer) &&
+          ts.isIdentifier(decl.initializer.expression) &&
+          decl.initializer.expression.text === 'withAudit' &&
+          decl.initializer.arguments.length === 1 &&
+          ts.isIdentifier(decl.initializer.arguments[0]) &&
+          decl.initializer.arguments[0].text === 'dispatchInner'
+        ) {
+          exportsWrappedDispatch = true;
+        }
+      }
+    }
+  });
+
+  it('imports withAudit from the audit adapter', () => {
+    expect(importsWithAudit).toBe(true);
+  });
+
+  it('keeps the inner dispatcher unexported (the guarded path is the only path)', () => {
+    expect(innerDeclaredUnexported).toBe(true);
+  });
+
+  it('exports dispatch as withAudit(dispatchInner)', () => {
+    expect(exportsWrappedDispatch).toBe(true);
+    expect(exportsUnwrappedDispatchFn).toBe(false);
+  });
+});
+
+describe('emit invariant', () => {
+  it('emitAuditEvent appears only in src/audit/ and *.audit.ts files', () => {
+    const srcRoot = path.resolve('src');
+    const offenders: string[] = [];
+    const walk = (dir: string): void => {
+      for (const entry of fs.readdirSync(dir, { withFileTypes: true })) {
+        const full = path.join(dir, entry.name);
+        if (entry.isDirectory()) {
+          walk(full);
+          continue;
+        }
+        if (!entry.name.endsWith('.ts')) continue;
+        const rel = path.relative(srcRoot, full).split(path.sep).join('/');
+        if (rel === 'audit-wiring.test.ts') continue; // this file names the symbol
+        if (rel.startsWith('audit/')) continue;
+        if (/\.audit(\.test)?\.ts$/.test(rel)) continue;
+        if (fs.readFileSync(full, 'utf8').includes('emitAuditEvent')) offenders.push(rel);
+      }
+    };
+    walk(srcRoot);
+    expect(offenders).toEqual([]);
+  });
+});

.claude/skills/add-audit/add/src/audit/config.ts

@@ -0,0 +1,30 @@
+/**
+ * Audit env config (installed by /add-audit) — the two audit vars, read the
+ * same way src/config.ts reads every host setting (readEnvFile from .env,
+ * with process.env taking precedence). Kept audit-owned so installing the
+ * skill never edits core config.ts: the feature's whole footprint in core is
+ * the dispatch composition in dispatch.ts and the resource-barrel import.
+ */
+import { readEnvFile } from '../env.js';
+
+const envConfig = readEnvFile(['AUDIT_ENABLED', 'AUDIT_RETENTION_DAYS']);
+
+/**
+ * Master switch. Off by default — nothing is persisted (and data/audit/ is
+ * never created) until an operator sets AUDIT_ENABLED=true.
+ */
+export const AUDIT_ENABLED = (process.env.AUDIT_ENABLED || envConfig.AUDIT_ENABLED) === 'true';
+
+/**
+ * Day-file retention horizon in days; consulted only when audit is enabled.
+ * Unset or unparseable → 90. An explicit 0 (or negative) = keep forever.
+ */
+function parseRetentionDays(raw: string | undefined): number {
+  if (raw === undefined || raw.trim() === '') return 90;
+  const n = Number.parseInt(raw, 10);
+  return Number.isNaN(n) ? 90 : n;
+}
+
+export const AUDIT_RETENTION_DAYS = parseRetentionDays(
+  process.env.AUDIT_RETENTION_DAYS || envConfig.AUDIT_RETENTION_DAYS,
+);

.claude/skills/add-audit/add/src/audit/emit.ts

@@ -0,0 +1,43 @@
+/**
+ * The single emit seam. Adapters import emitAuditEvent from here directly
+ * (it is deliberately not re-exported by the barrel): only src/audit/ and
+ * `*.audit.ts` adapter files may call it — grep holds the invariant.
+ *
+ * The opt-in check lives here, so the whole feature switches at one point.
+ * Fail-open + loud: a failed append (or a throwing input thunk) is
+ * log.error'd and the audited action proceeds.
+ */
+import { randomUUID } from 'crypto';
+
+import { log } from '../log.js';
+import { AUDIT_ENABLED } from './config.js';
+import { notifyAuditHooks } from './hooks.js';
+import { redactDetails } from './redact.js';
+import { appendAuditLine } from './store.js';
+import type { AuditEvent, AuditEventInput } from './types.js';
+
+export function emitAuditEvent(input: AuditEventInput | (() => AuditEventInput)): void {
+  if (!AUDIT_ENABLED) return; // The one opt-in check — the whole feature switches here.
+  try {
+    if (typeof input === 'function') input = input();
+    const event: AuditEvent = {
+      event_id: randomUUID(),
+      time: new Date().toISOString(),
+      schema_version: 1,
+      actor: { ...input.actor, email: null, user_id: null, group_ids: null },
+      origin: input.origin,
+      action: input.action,
+      resources: input.resources,
+      outcome: input.outcome,
+      correlation_id: input.correlationId ?? null,
+      details: redactDetails(input.details ?? {}),
+    };
+    const line = JSON.stringify(event);
+    appendAuditLine(line);
+    notifyAuditHooks(event, line);
+    // eslint-disable-next-line no-catch-all/no-catch-all -- fail-open is the posture: an audit failure must never take down the audited action
+  } catch (err) {
+    const action = typeof input === 'function' ? undefined : input.action;
+    log.error('Audit append failed — action proceeding (fail-open)', { action, err });
+  }
+}

.claude/skills/add-audit/add/src/audit/hooks.test.ts

@@ -0,0 +1,203 @@
+/**
+ * Post-write hook contract: hooks observe the LOG (fire only after a
+ * successful append, exported ⊆ written), failures are isolated everywhere,
+ * and the lifecycle (init/maintain/shutdown) behaves.
+ */
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const state = vi.hoisted(() => ({ enabled: true, appendThrows: false, appended: [] as string[] }));
+
+vi.mock('../config.js', async (importOriginal) => {
+  const actual = await importOriginal<typeof import('../config.js')>();
+  return {
+    ...actual,
+    DATA_DIR: '/tmp/nanoclaw-test-hooks-unused',
+  };
+});
+
+vi.mock('./config.js', () => ({
+  get AUDIT_ENABLED() {
+    return state.enabled;
+  },
+  AUDIT_RETENTION_DAYS: 90,
+}));
+
+vi.mock('./store.js', async (importOriginal) => {
+  const actual = await importOriginal<typeof import('./store.js')>();
+  return {
+    ...actual,
+    appendAuditLine: (line: string) => {
+      if (state.appendThrows) throw new Error('disk full');
+      state.appended.push(line);
+    },
+  };
+});
+
+vi.mock('../log.js', () => ({
+  log: { debug: vi.fn(), info: vi.fn(), warn: vi.fn(), error: vi.fn(), fatal: vi.fn() },
+}));
+
+let hooks: typeof import('./hooks.js');
+let emit: typeof import('./emit.js');
+let log: (typeof import('../log.js'))['log'];
+
+beforeEach(async () => {
+  state.enabled = true;
+  state.appendThrows = false;
+  state.appended.length = 0;
+  vi.resetModules(); // fresh hook registry per test
+  hooks = await import('./hooks.js');
+  emit = await import('./emit.js');
+  log = (await import('../log.js')).log;
+  vi.clearAllMocks();
+});
+
+afterEach(() => {
+  vi.clearAllMocks();
+});
+
+const EVENT_INPUT = {
+  actor: { type: 'human' as const, id: 'host:test' },
+  origin: { transport: 'socket' as const },
+  action: 'groups.list',
+  resources: [{ type: 'agent_group' }],
+  outcome: 'success' as const,
+  details: { limit: 5 },
+};
+
+describe('post-write notification', () => {
+  it('calls a registered hook with the parsed event and the exact stored line', () => {
+    const seen: Array<{ event: import('./types.js').AuditEvent; line: string }> = [];
+    hooks.registerAuditHook({ name: 'demo', onEvent: (event, line) => seen.push({ event, line }) });
+
+    emit.emitAuditEvent(EVENT_INPUT);
+
+    expect(state.appended).toHaveLength(1);
+    expect(seen).toHaveLength(1);
+    expect(seen[0].line).toBe(state.appended[0]);
+    expect(JSON.parse(seen[0].line)).toEqual(seen[0].event);
+    expect(seen[0].event.action).toBe('groups.list');
+  });
+
+  it('does NOT call hooks when the local append fails — exported ⊆ written', () => {
+    const onEvent = vi.fn();
+    hooks.registerAuditHook({ name: 'demo', onEvent });
+    state.appendThrows = true;
+
+    expect(() => emit.emitAuditEvent(EVENT_INPUT)).not.toThrow(); // action still proceeds
+    expect(onEvent).not.toHaveBeenCalled();
+    expect(log.error).toHaveBeenCalledWith(expect.stringContaining('Audit append failed'), expect.anything());
+  });
+
+  it('does NOT call hooks when audit is disabled', () => {
+    const onEvent = vi.fn();
+    hooks.registerAuditHook({ name: 'demo', onEvent });
+    state.enabled = false;
+
+    emit.emitAuditEvent(EVENT_INPUT);
+
+    expect(state.appended).toHaveLength(0);
+    expect(onEvent).not.toHaveBeenCalled();
+  });
+
+  it('isolates a throwing hook: the write survives, later hooks still run, the action proceeds', () => {
+    const second = vi.fn();
+    hooks.registerAuditHook({
+      name: 'broken',
+      onEvent: () => {
+        throw new Error('exporter exploded');
+      },
+    });
+    hooks.registerAuditHook({ name: 'healthy', onEvent: second });
+
+    expect(() => emit.emitAuditEvent(EVENT_INPUT)).not.toThrow();
+
+    expect(state.appended).toHaveLength(1); // the log has the event regardless
+    expect(second).toHaveBeenCalledTimes(1);
+    expect(log.error).toHaveBeenCalledWith(
+      expect.stringContaining('Audit hook threw'),
+      expect.objectContaining({ hook: 'broken', action: 'groups.list' }),
+    );
+  });
+});
+
+describe('lifecycle', () => {
+  it('initAuditHooks surfaces a failing init as a fatal error naming the hook', () => {
+    hooks.registerAuditHook({ name: 'ok', onEvent: () => {}, init: vi.fn() });
+    hooks.registerAuditHook({
+      name: 'bad-boot',
+      onEvent: () => {},
+      init: () => {
+        throw new Error('no route to collector');
+      },
+    });
+
+    expect(() => hooks.initAuditHooks()).toThrow(/audit hook "bad-boot" failed to initialize.*no route/);
+  });
+
+  it('maintainAuditHooks calls every maintain and isolates throws', () => {
+    const m1 = vi.fn(() => {
+      throw new Error('flush failed');
+    });
+    const m2 = vi.fn();
+    hooks.registerAuditHook({ name: 'a', onEvent: () => {}, maintain: m1 });
+    hooks.registerAuditHook({ name: 'b', onEvent: () => {}, maintain: m2 });
+    hooks.registerAuditHook({ name: 'c', onEvent: () => {} }); // no maintain — fine
+
+    expect(() => hooks.maintainAuditHooks()).not.toThrow();
+    expect(m1).toHaveBeenCalledTimes(1);
+    expect(m2).toHaveBeenCalledTimes(1);
+    expect(log.error).toHaveBeenCalledWith(
+      expect.stringContaining('maintenance failed'),
+      expect.objectContaining({ hook: 'a' }),
+    );
+  });
+
+  it('shutdownAuditHooks awaits async shutdowns and isolates throws', async () => {
+    const order: string[] = [];
+    hooks.registerAuditHook({
+      name: 'a',
+      onEvent: () => {},
+      shutdown: async () => {
+        await Promise.resolve();
+        order.push('a');
+      },
+    });
+    hooks.registerAuditHook({
+      name: 'b',
+      onEvent: () => {},
+      shutdown: () => {
+        throw new Error('handle already closed');
+      },
+    });
+    hooks.registerAuditHook({
+      name: 'c',
+      onEvent: () => {},
+      shutdown: () => {
+        order.push('c');
+      },
+    });
+
+    await hooks.shutdownAuditHooks();
+
+    expect(order).toEqual(['a', 'c']);
+    expect(log.error).toHaveBeenCalledWith(
+      expect.stringContaining('shutdown failed'),
+      expect.objectContaining({ hook: 'b' }),
+    );
+  });
+
+  it('maintainAudit skips hook maintenance when audit is disabled', async () => {
+    const init = await import('./init.js');
+    const maintain = vi.fn();
+    hooks.registerAuditHook({ name: 'a', onEvent: () => {}, maintain });
+
+    state.enabled = false;
+    init.maintainAudit();
+    expect(maintain).not.toHaveBeenCalled();
+
+    state.enabled = true;
+    init.maintainAudit();
+    expect(maintain).toHaveBeenCalledTimes(1);
+  });
+});

.claude/skills/add-audit/add/src/audit/hooks.ts

@@ -0,0 +1,92 @@
+/**
+ * Post-write audit hooks — the in-process extension seam.
+ *
+ * A hook observes the audit LOG, not the event stream: `onEvent` fires only
+ * after an event has been durably appended to the local day-file, so anything
+ * a hook exports is guaranteed to exist in the source of truth
+ * (exported ⊆ written). If the local append fails, hooks are not called —
+ * and a hook that misses events (crash, restart) catches up by reading the
+ * day-files, which is the at-least-once story.
+ *
+ * Registration follows the tree's observer idiom (registerApprovalResolvedHandler,
+ * registerResponseHandler, …): an in-tree or skill-installed module calls
+ * `registerAuditHook(...)` at import time — no core edits, and credentials or
+ * transport for an external system live in that module, never here.
+ */
+import { log } from '../log.js';
+import type { AuditEvent } from './types.js';
+
+export interface AuditHook {
+  /** Short identifier used in logs and lifecycle errors. */
+  name: string;
+  /**
+   * Called after a successful local append. `line` is the exact stored bytes
+   * (one NDJSON line, no trailing newline); `event` is the parsed record.
+   * MUST be fast and non-blocking — this runs on the audited action's call
+   * path. A real exporter buffers here and does its IO from `maintain`/its own
+   * timers. Throwing is tolerated: isolated and logged, never propagated.
+   */
+  onEvent(event: AuditEvent, line: string): void;
+  /** Boot hook, called only when audit is enabled. Throw = host refuses to start. */
+  init?(): void;
+  /** Periodic maintenance — called from the audit maintenance timer (enabled boxes only). */
+  maintain?(): void;
+  /** Graceful-shutdown hook (flush buffers, close handles). */
+  shutdown?(): void | Promise<void>;
+}
+
+const hooks: AuditHook[] = [];
+
+export function registerAuditHook(hook: AuditHook): void {
+  hooks.push(hook);
+}
+
+/** Fan out one written event to every hook, isolating failures per hook. */
+export function notifyAuditHooks(event: AuditEvent, line: string): void {
+  for (const hook of hooks) {
+    try {
+      hook.onEvent(event, line);
+      // eslint-disable-next-line no-catch-all/no-catch-all -- isolation is the contract: one bad hook must not affect the log, other hooks, or the audited action
+    } catch (err) {
+      log.error('Audit hook threw — event is safely in the log', { hook: hook.name, action: event.action, err });
+    }
+  }
+}
+
+/** Boot lifecycle. A hook that can't start is a silent-export-gap risk — fatal. */
+export function initAuditHooks(): void {
+  for (const hook of hooks) {
+    try {
+      hook.init?.();
+    } catch (err) {
+      throw new Error(
+        `audit hook "${hook.name}" failed to initialize: ${err instanceof Error ? err.message : String(err)}`,
+        { cause: err },
+      );
+    }
+  }
+}
+
+/** Periodic lifecycle — maintenance, isolated per hook. */
+export function maintainAuditHooks(): void {
+  for (const hook of hooks) {
+    try {
+      hook.maintain?.();
+      // eslint-disable-next-line no-catch-all/no-catch-all -- one hook's maintenance failure must not stop the others (or the timer)
+    } catch (err) {
+      log.error('Audit hook maintenance failed', { hook: hook.name, err });
+    }
+  }
+}
+
+/** Shutdown lifecycle — awaited by the host's graceful shutdown. */
+export async function shutdownAuditHooks(): Promise<void> {
+  for (const hook of hooks) {
+    try {
+      await hook.shutdown?.();
+      // eslint-disable-next-line no-catch-all/no-catch-all -- shutdown must drain every hook even when one throws
+    } catch (err) {
+      log.error('Audit hook shutdown failed', { hook: hook.name, err });
+    }
+  }
+}

.claude/skills/add-audit/add/src/audit/index.ts

@@ -0,0 +1,18 @@
+/**
+ * Opt-in local audit log — installed by /add-audit.
+ *
+ * This directory is a domain-free leaf: the event schema, the emit seam, the
+ * store, the reader, post-write hooks, and shared vocabulary. What gets
+ * audited — and how each domain describes itself — lives in the domain-owned
+ * `*.audit.ts` adapter files next to the code they observe. Business logic
+ * contains zero audit calls.
+ *
+ * emitAuditEvent is deliberately NOT re-exported here: adapters import it
+ * from './emit.js' directly, and only src/audit/ + `*.audit.ts` may call it.
+ */
+export * from './types.js';
+export { redactDetails } from './redact.js';
+export { AUDIT_DIR } from './store.js';
+export { AUDIT_ENABLED, AUDIT_RETENTION_DAYS } from './config.js';
+export { initAuditLog, maintainAudit } from './init.js';
+export { type AuditHook, registerAuditHook } from './hooks.js';

.claude/skills/add-audit/add/src/audit/init.ts

@@ -0,0 +1,54 @@
+/**
+ * Boot-time audit wiring, self-contained (installed by /add-audit).
+ *
+ * initAuditLog() runs at module load of the CLI audit adapter
+ * (cli/dispatch.audit.ts) — i.e. during the host's barrel phase, before the
+ * CLI server or any delivery poll accepts work — because dispatch.ts, which
+ * both transports import, composes withAudit at module scope. When enabled:
+ * assert data/audit/ is writable (refusing to start beats running with a
+ * silent audit gap), run the boot prune, start the registered post-write
+ * hooks, and arm the maintenance timer.
+ *
+ * The timer owns the daily cadence in-module (checked hourly; the prune
+ * itself fires at most once per UTC day) so installing the skill touches
+ * dispatch.ts and the resource barrel only — no host-sweep edit. It is
+ * unref'd: it never keeps the process alive.
+ */
+import { log } from '../log.js';
+import { onShutdown } from '../response-registry.js';
+import { AUDIT_ENABLED, AUDIT_RETENTION_DAYS } from './config.js';
+import { initAuditHooks, maintainAuditHooks, shutdownAuditHooks } from './hooks.js';
+import { assertAuditWritable, AUDIT_DIR, markPrunedToday, pruneAuditLog, pruneAuditLogIfDue } from './store.js';
+
+const MAINTENANCE_INTERVAL_MS = 60 * 60 * 1000;
+
+let initialized = false;
+
+export function initAuditLog(): void {
+  if (!AUDIT_ENABLED || initialized) return;
+  initialized = true;
+  try {
+    assertAuditWritable();
+  } catch (err) {
+    throw new Error(
+      `AUDIT_ENABLED=true but the audit directory is not writable: ${AUDIT_DIR} (${err instanceof Error ? err.message : String(err)})`,
+      { cause: err },
+    );
+  }
+  pruneAuditLog();
+  markPrunedToday();
+  initAuditHooks(); // throw → boot fails, same posture as the writability assert
+  onShutdown(() => shutdownAuditHooks());
+  setInterval(maintainAudit, MAINTENANCE_INTERVAL_MS).unref();
+  log.info('Audit log enabled', { dir: AUDIT_DIR, retentionDays: AUDIT_RETENTION_DAYS });
+}
+
+/**
+ * One maintenance tick: retention prune (throttled to once per UTC day
+ * internally) plus every hook's periodic maintenance. No-op when audit is
+ * disabled. Exported so a future scheduler can drive it too.
+ */
+export function maintainAudit(): void {
+  pruneAuditLogIfDue();
+  if (AUDIT_ENABLED) maintainAuditHooks();
+}

.claude/skills/add-audit/add/src/audit/reader.test.ts

@@ -0,0 +1,153 @@
+import fs from 'fs';
+import os from 'os';
+import path from 'path';
+
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const state = vi.hoisted(() => ({ dataDir: '', enabled: true }));
+
+vi.mock('../config.js', () => ({
+  get DATA_DIR() {
+    return state.dataDir;
+  },
+}));
+
+vi.mock('./config.js', () => ({
+  get AUDIT_ENABLED() {
+    return state.enabled;
+  },
+  AUDIT_RETENTION_DAYS: 90,
+}));
+
+vi.mock('../log.js', () => ({
+  log: { debug: vi.fn(), info: vi.fn(), warn: vi.fn(), error: vi.fn(), fatal: vi.fn() },
+}));
+
+let reader: typeof import('./reader.js');
+let store: typeof import('./store.js');
+
+beforeEach(async () => {
+  state.dataDir = fs.mkdtempSync(path.join(os.tmpdir(), 'audit-reader-'));
+  state.enabled = true;
+  vi.resetModules();
+  store = await import('./store.js');
+  reader = await import('./reader.js');
+});
+
+afterEach(() => {
+  fs.rmSync(state.dataDir, { recursive: true, force: true });
+});
+
+function dayString(daysAgo: number): string {
+  return store.utcDay(new Date(Date.now() - daysAgo * 24 * 60 * 60 * 1000));
+}
+
+function isoAt(daysAgo: number, tag: number): string {
+  const base = new Date(Date.now() - daysAgo * 24 * 60 * 60 * 1000);
+  return `${base.toISOString().slice(0, 10)}T10:00:0${tag}.000Z`;
+}
+
+let seq = 0;
+function seedEvent(daysAgo: number, overrides: Record<string, unknown> = {}): Record<string, unknown> {
+  seq += 1;
+  const event = {
+    event_id: `e-${seq}`,
+    time: isoAt(daysAgo, seq % 10),
+    schema_version: 1,
+    actor: { type: 'human', id: 'host:moshe', email: null, user_id: null, group_ids: null },
+    origin: { transport: 'socket' },
+    action: 'groups.list',
+    resources: [{ type: 'agent_group', id: 'ag-1' }],
+    outcome: 'success',
+    correlation_id: null,
+    details: {},
+    ...overrides,
+  };
+  const dir = path.join(state.dataDir, 'audit');
+  fs.mkdirSync(dir, { recursive: true });
+  fs.appendFileSync(path.join(dir, `${dayString(daysAgo)}.ndjson`), JSON.stringify(event) + '\n');
+  return event;
+}
+
+describe('listAuditEvents', () => {
+  it('throws the disabled error rather than returning an empty list', () => {
+    state.enabled = false;
+    expect(() => reader.listAuditEvents({})).toThrow('audit log is disabled — set AUDIT_ENABLED=true');
+  });
+
+  it('returns flat rows newest-first across day-files, honoring --limit', () => {
+    seedEvent(2, { event_id: 'old' });
+    seedEvent(1, { event_id: 'mid-a' });
+    seedEvent(1, { event_id: 'mid-b' });
+    seedEvent(0, { event_id: 'new' });
+
+    const rows = reader.listAuditEvents({}) as Array<Record<string, unknown>>;
+    expect(rows.map((r) => r.event_id)).toEqual(['new', 'mid-b', 'mid-a', 'old']);
+    expect(rows[0]).toMatchObject({ actor: 'host:moshe', action: 'groups.list', outcome: 'success' });
+    expect(rows[0].resources).toBe('agent_group:ag-1');
+
+    const limited = reader.listAuditEvents({ limit: 2 }) as Array<Record<string, unknown>>;
+    expect(limited.map((r) => r.event_id)).toEqual(['new', 'mid-b']);
+  });
+
+  it('filters by actor, outcome, correlation, and resource (id or type)', () => {
+    seedEvent(0, { event_id: 'a', actor: { type: 'agent', id: 'ag-1' }, outcome: 'denied' });
+    seedEvent(0, { event_id: 'b', correlation_id: 'appr-9', resources: [{ type: 'approval', id: 'appr-9' }] });
+    seedEvent(0, { event_id: 'c', resources: [{ type: 'user', id: 'slack:U1' }] });
+
+    expect((reader.listAuditEvents({ actor: 'ag-1' }) as unknown[]).length).toBe(1);
+    expect((reader.listAuditEvents({ outcome: 'denied' }) as unknown[]).length).toBe(1);
+    expect((reader.listAuditEvents({ correlation: 'appr-9' }) as unknown[]).length).toBe(1);
+    expect((reader.listAuditEvents({ resource: 'slack:U1' }) as unknown[]).length).toBe(1);
+    expect((reader.listAuditEvents({ resource: 'approval' }) as unknown[]).length).toBe(1);
+  });
+
+  it('matches actions exactly or by dotted prefix', () => {
+    seedEvent(0, { event_id: 'cfg', action: 'groups.config.add-mcp-server' });
+    seedEvent(0, { event_id: 'list', action: 'groups.list' });
+    seedEvent(0, { event_id: 'other', action: 'sessions.list' });
+
+    expect((reader.listAuditEvents({ action: 'groups' }) as unknown[]).length).toBe(2);
+    expect((reader.listAuditEvents({ action: 'groups.config' }) as unknown[]).length).toBe(1);
+    expect((reader.listAuditEvents({ action: 'groups.list' }) as unknown[]).length).toBe(1);
+    // Prefix means dotted segments, not substrings.
+    expect((reader.listAuditEvents({ action: 'group' }) as unknown[]).length).toBe(0);
+  });
+
+  it('applies --since/--until with relative and ISO forms', () => {
+    seedEvent(5, { event_id: 'old' });
+    seedEvent(0, { event_id: 'recent' });
+
+    const relative = reader.listAuditEvents({ since: '2d' }) as Array<Record<string, unknown>>;
+    expect(relative.map((r) => r.event_id)).toEqual(['recent']);
+
+    const iso = reader.listAuditEvents({ until: dayString(2) }) as Array<Record<string, unknown>>;
+    expect(iso.map((r) => r.event_id)).toEqual(['old']);
+
+    expect(() => reader.listAuditEvents({ since: 'yesterdayish' })).toThrow('invalid --since');
+  });
+
+  it('--format ndjson returns the stored lines verbatim', () => {
+    const seeded = seedEvent(0, { event_id: 'x1' });
+    const out = reader.listAuditEvents({ format: 'ndjson' });
+    expect(typeof out).toBe('string');
+    expect(JSON.parse(out as string)).toEqual(seeded);
+    expect(() => reader.listAuditEvents({ format: 'csv' })).toThrow('invalid --format');
+  });
+
+  it('skips malformed stored lines and still returns the rest', () => {
+    seedEvent(0, { event_id: 'good' });
+    fs.appendFileSync(path.join(state.dataDir, 'audit', `${dayString(0)}.ndjson`), 'not-json\n');
+
+    const rows = reader.listAuditEvents({}) as Array<Record<string, unknown>>;
+    expect(rows.map((r) => r.event_id)).toEqual(['good']);
+  });
+
+  it('rejects an unknown --outcome value', () => {
+    expect(() => reader.listAuditEvents({ outcome: 'meh' })).toThrow('invalid --outcome');
+  });
+
+  it('returns empty when nothing has been recorded yet', () => {
+    expect(reader.listAuditEvents({})).toEqual([]);
+  });
+});

.claude/skills/add-audit/add/src/audit/reader.ts

@@ -0,0 +1,160 @@
+/**
+ * Read-back for `ncl audit list` — a newest-first stream-scan over the
+ * day-files. No index: fine at v1 volume, and adding one later doesn't change
+ * the store. NDJSON export returns the stored lines verbatim.
+ */
+import fs from 'fs';
+import path from 'path';
+
+import { log } from '../log.js';
+import { AUDIT_ENABLED } from './config.js';
+import { AUDIT_DIR, utcDay } from './store.js';
+import type { AuditEvent, AuditOutcome } from './types.js';
+
+const OUTCOMES: ReadonlySet<string> = new Set(['success', 'failure', 'denied', 'pending', 'approved', 'rejected']);
+const DAY_FILE_RE = /^(\d{4}-\d{2}-\d{2})\.ndjson$/;
+const DEFAULT_LIMIT = 100;
+
+export interface AuditQuery {
+  actor?: string;
+  /** Exact action or dotted prefix (`groups` matches `groups.config.update`). */
+  action?: string;
+  /** Matches any resources[] entry by id or by type. */
+  resource?: string;
+  outcome?: AuditOutcome;
+  sinceMs?: number;
+  untilMs?: number;
+  correlation?: string;
+  limit: number;
+}
+
+/** `7d` / `24h` / `30m` relative to now, or an ISO date/datetime (UTC). */
+export function parseTimeFlag(value: string, flag: string): number {
+  const rel = /^(\d+)([dhm])$/.exec(value);
+  if (rel) {
+    const n = Number(rel[1]);
+    const unitMs = rel[2] === 'd' ? 86_400_000 : rel[2] === 'h' ? 3_600_000 : 60_000;
+    return Date.now() - n * unitMs;
+  }
+  const abs = Date.parse(value);
+  if (!Number.isNaN(abs)) return abs;
+  throw new Error(`invalid ${flag} value "${value}" — use e.g. 7d, 24h, 30m, or an ISO date`);
+}
+
+/** Newest first across files and within each file, up to q.limit. */
+export function queryAuditEvents(q: AuditQuery): { events: AuditEvent[]; lines: string[] } {
+  const events: AuditEvent[] = [];
+  const lines: string[] = [];
+  let malformed = 0;
+
+  for (const { day, file } of dayFilesNewestFirst()) {
+    if (events.length >= q.limit) break;
+    // Whole-day skip: a file can't match a window its day lies outside.
+    if (q.sinceMs !== undefined && day < utcDay(new Date(q.sinceMs))) continue;
+    if (q.untilMs !== undefined && day > utcDay(new Date(q.untilMs))) continue;
+
+    let content: string;
+    try {
+      content = fs.readFileSync(file, 'utf8');
+      // eslint-disable-next-line no-catch-all/no-catch-all -- a torn/pruned day-file must not fail the whole query
+    } catch (err) {
+      log.warn('Audit reader failed to read day-file', { file, err });
+      continue;
+    }
+    const fileLines = content.split('\n').filter((l) => l.trim() !== '');
+    // Lines within a file are chronological — walk backwards for newest-first.
+    for (let i = fileLines.length - 1; i >= 0 && events.length < q.limit; i--) {
+      let event: AuditEvent;
+      try {
+        event = JSON.parse(fileLines[i]) as AuditEvent;
+        // eslint-disable-next-line no-catch-all/no-catch-all -- malformed stored lines are skipped (counted + warned below)
+      } catch {
+        malformed++;
+        continue;
+      }
+      if (!matches(event, q)) continue;
+      events.push(event);
+      lines.push(fileLines[i]);
+    }
+  }
+
+  if (malformed > 0) {
+    log.warn('Audit reader skipped malformed lines', { malformed });
+  }
+  return { events, lines };
+}
+
+function dayFilesNewestFirst(): Array<{ day: string; file: string }> {
+  let entries: string[];
+  try {
+    entries = fs.readdirSync(AUDIT_DIR);
+    // eslint-disable-next-line no-catch-all/no-catch-all -- no audit dir yet means no events, not an error
+  } catch {
+    return [];
+  }
+  return entries
+    .map((e) => DAY_FILE_RE.exec(e))
+    .filter((m): m is RegExpExecArray => m !== null)
+    .map((m) => ({ day: m[1], file: path.join(AUDIT_DIR, m[0]) }))
+    .sort((a, b) => (a.day < b.day ? 1 : -1));
+}
+
+function matches(event: AuditEvent, q: AuditQuery): boolean {
+  if (q.actor !== undefined && event.actor?.id !== q.actor) return false;
+  if (q.action !== undefined && event.action !== q.action && !event.action?.startsWith(q.action + '.')) return false;
+  if (q.outcome !== undefined && event.outcome !== q.outcome) return false;
+  if (q.correlation !== undefined && event.correlation_id !== q.correlation) return false;
+  if (q.resource !== undefined) {
+    const hit = (event.resources ?? []).some((r) => r.id === q.resource || r.type === q.resource);
+    if (!hit) return false;
+  }
+  const t = Date.parse(event.time ?? '');
+  if (q.sinceMs !== undefined && !(t >= q.sinceMs)) return false;
+  if (q.untilMs !== undefined && !(t <= q.untilMs)) return false;
+  return true;
+}
+
+/**
+ * `ncl audit list` handler. Disabled → an explicit error: an empty list would
+ * read as "no actions happened", which is a different truth than "not
+ * recording". `--format ndjson` returns the stored lines verbatim (the human
+ * formatter passes strings through); default returns flat rows for the table.
+ */
+export function listAuditEvents(args: Record<string, unknown>): string | Array<Record<string, unknown>> {
+  if (!AUDIT_ENABLED) {
+    throw new Error('audit log is disabled — set AUDIT_ENABLED=true');
+  }
+
+  const format = args.format !== undefined ? String(args.format) : '';
+  if (format && format !== 'ndjson') {
+    throw new Error(`invalid --format "${format}" — only "ndjson" is supported`);
+  }
+  const outcome = args.outcome !== undefined ? String(args.outcome) : undefined;
+  if (outcome !== undefined && !OUTCOMES.has(outcome)) {
+    throw new Error(`invalid --outcome "${outcome}" — one of: ${[...OUTCOMES].join(', ')}`);
+  }
+
+  const q: AuditQuery = {
+    actor: args.actor !== undefined ? String(args.actor) : undefined,
+    action: args.action !== undefined ? String(args.action) : undefined,
+    resource: args.resource !== undefined ? String(args.resource) : undefined,
+    outcome: outcome as AuditOutcome | undefined,
+    correlation: args.correlation !== undefined ? String(args.correlation) : undefined,
+    sinceMs: args.since !== undefined ? parseTimeFlag(String(args.since), '--since') : undefined,
+    untilMs: args.until !== undefined ? parseTimeFlag(String(args.until), '--until') : undefined,
+    limit: args.limit !== undefined ? Math.max(1, Number(args.limit) || DEFAULT_LIMIT) : DEFAULT_LIMIT,
+  };
+
+  const { events, lines } = queryAuditEvents(q);
+  if (format === 'ndjson') return lines.join('\n');
+
+  return events.map((e) => ({
+    time: e.time,
+    actor: e.actor?.id ?? '',
+    action: e.action,
+    resources: (e.resources ?? []).map((r) => (r.id ? `${r.type}:${r.id}` : r.type)).join(' '),
+    outcome: e.outcome,
+    correlation: e.correlation_id ?? '',
+    event_id: e.event_id,
+  }));
+}

.claude/skills/add-audit/add/src/audit/redact.test.ts

@@ -0,0 +1,69 @@
+import { describe, expect, it } from 'vitest';
+
+import { redactDetails } from './redact.js';
+
+describe('redactDetails', () => {
+  it('masks sensitive keys at any depth, case-insensitively', () => {
+    const out = redactDetails({
+      name: 'notion',
+      env: { NOTION_TOKEN: 'secret-value', SAFE_VALUE: 'ok' },
+      nested: { Authorization: 'Bearer abc', list: [{ 'api-key': 'k' }, { plain: 'p' }] },
+      password: 'hunter2',
+    });
+    expect(out).toEqual({
+      name: 'notion',
+      env: { NOTION_TOKEN: '[REDACTED]', SAFE_VALUE: 'ok' },
+      nested: { Authorization: '[REDACTED]', list: [{ 'api-key': '[REDACTED]' }, { plain: 'p' }] },
+      password: '[REDACTED]',
+    });
+  });
+
+  it('matches the documented key pattern (token|secret|key|password|credential|auth|bearer)', () => {
+    const out = redactDetails({
+      access_token: 'x',
+      clientSecret: 'x',
+      ssh_key: 'x',
+      credentials: 'x',
+      oauth_flow: 'x',
+      bearerValue: 'x',
+      username: 'moshe',
+    });
+    expect(Object.entries(out).filter(([, v]) => v === '[REDACTED]')).toHaveLength(6);
+    expect(out.username).toBe('moshe');
+  });
+
+  it('never recurses into a masked key — the whole value is replaced', () => {
+    const out = redactDetails({ auth: { inner: 'visible?' } });
+    expect(out.auth).toBe('[REDACTED]');
+  });
+
+  it('truncates strings over 2 KB post-redaction', () => {
+    const long = 'a'.repeat(5000);
+    const out = redactDetails({ blob: long, short: 'b' });
+    expect(out.blob).toBe('a'.repeat(2048) + '…[truncated]');
+    expect(out.short).toBe('b');
+  });
+
+  it('passes non-string scalars through untouched', () => {
+    const out = redactDetails({ n: 42, b: true, z: null, u: undefined });
+    expect(out).toEqual({ n: 42, b: true, z: null, u: undefined });
+  });
+
+  it('caps depth (cycle guard) without throwing', () => {
+    const cyclic: Record<string, unknown> = { a: 1 };
+    cyclic.self = cyclic;
+    const out = redactDetails(cyclic);
+    let cursor: unknown = out;
+    for (let i = 0; i < 20 && typeof cursor === 'object' && cursor !== null; i++) {
+      cursor = (cursor as Record<string, unknown>).self;
+    }
+    expect(cursor).toBe('[MAX_DEPTH]');
+  });
+
+  it('does not mutate the input', () => {
+    const input = { token: 'x', nested: { list: ['a'.repeat(3000)] } };
+    const snapshot = JSON.parse(JSON.stringify(input));
+    redactDetails(input);
+    expect(input).toEqual(snapshot);
+  });
+});

.claude/skills/add-audit/add/src/audit/redact.ts

@@ -0,0 +1,32 @@
+/**
+ * The one redactor, applied to every event's `details` at the emit seam —
+ * guarding current and future surfaces by default. Key-pattern mask plus a
+ * per-value size cap; message bodies are never passed in (emit sites record
+ * shape only — the audit log is a governance record, not a chat archive).
+ */
+
+const SENSITIVE_KEY = /(token|secret|key|password|credential|auth|bearer)/i;
+const MAX_VALUE_CHARS = 2048;
+const MAX_DEPTH = 8;
+
+export function redactDetails(details: Record<string, unknown>): Record<string, unknown> {
+  return redactObject(details, 0);
+}
+
+function redactObject(obj: Record<string, unknown>, depth: number): Record<string, unknown> {
+  const out: Record<string, unknown> = {};
+  for (const [k, v] of Object.entries(obj)) {
+    out[k] = SENSITIVE_KEY.test(k) ? '[REDACTED]' : redactValue(v, depth + 1);
+  }
+  return out;
+}
+
+function redactValue(value: unknown, depth: number): unknown {
+  if (typeof value === 'string') {
+    return value.length > MAX_VALUE_CHARS ? `${value.slice(0, MAX_VALUE_CHARS)}…[truncated]` : value;
+  }
+  if (value === null || typeof value !== 'object') return value;
+  if (depth > MAX_DEPTH) return '[MAX_DEPTH]';
+  if (Array.isArray(value)) return value.map((v) => redactValue(v, depth + 1));
+  return redactObject(value as Record<string, unknown>, depth);
+}

.claude/skills/add-audit/add/src/audit/store.test.ts

@@ -0,0 +1,187 @@
+import fs from 'fs';
+import os from 'os';
+import path from 'path';
+
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+// Core config (DATA_DIR) and the audit config (toggles) are mocked so
+// store/emit resolve a per-test temp DATA_DIR and audit switches. Getters
+// keep the values live across vi.resetModules().
+const state = vi.hoisted(() => ({ dataDir: '', enabled: true, retention: 90 }));
+
+vi.mock('../config.js', () => ({
+  get DATA_DIR() {
+    return state.dataDir;
+  },
+}));
+
+vi.mock('./config.js', () => ({
+  get AUDIT_ENABLED() {
+    return state.enabled;
+  },
+  get AUDIT_RETENTION_DAYS() {
+    return state.retention;
+  },
+}));
+
+vi.mock('../log.js', () => ({
+  log: { debug: vi.fn(), info: vi.fn(), warn: vi.fn(), error: vi.fn(), fatal: vi.fn() },
+}));
+
+let store: typeof import('./store.js');
+let emit: typeof import('./emit.js');
+let log: (typeof import('../log.js'))['log'];
+
+beforeEach(async () => {
+  state.dataDir = fs.mkdtempSync(path.join(os.tmpdir(), 'audit-store-'));
+  state.enabled = true;
+  state.retention = 90;
+  vi.resetModules();
+  store = await import('./store.js');
+  emit = await import('./emit.js');
+  log = (await import('../log.js')).log;
+  vi.clearAllMocks();
+});
+
+afterEach(() => {
+  fs.chmodSync(path.join(state.dataDir), 0o700);
+  const auditDir = path.join(state.dataDir, 'audit');
+  if (fs.existsSync(auditDir)) fs.chmodSync(auditDir, 0o700);
+  fs.rmSync(state.dataDir, { recursive: true, force: true });
+});
+
+function auditDir(): string {
+  return path.join(state.dataDir, 'audit');
+}
+
+function writeDayFile(day: string, lines = 1): void {
+  fs.mkdirSync(auditDir(), { recursive: true });
+  fs.writeFileSync(path.join(auditDir(), `${day}.ndjson`), '{"x":1}\n'.repeat(lines));
+}
+
+function dayString(daysAgo: number): string {
+  return store.utcDay(new Date(Date.now() - daysAgo * 24 * 60 * 60 * 1000));
+}
+
+const EVENT_INPUT = {
+  actor: { type: 'human' as const, id: 'host:test' },
+  origin: { transport: 'socket' as const },
+  action: 'groups.list',
+  resources: [{ type: 'agent_group' }],
+  outcome: 'success' as const,
+};
+
+describe('appendAuditLine', () => {
+  it("appends one line to today's UTC day-file, creating the directory lazily", () => {
+    expect(fs.existsSync(auditDir())).toBe(false);
+    store.appendAuditLine('{"a":1}');
+    store.appendAuditLine('{"b":2}');
+    const file = path.join(auditDir(), `${store.utcDay()}.ndjson`);
+    expect(fs.readFileSync(file, 'utf8')).toBe('{"a":1}\n{"b":2}\n');
+  });
+});
+
+describe('emitAuditEvent', () => {
+  it('writes a schema_version-1 record with envelope fields and null enrichment', () => {
+    emit.emitAuditEvent({ ...EVENT_INPUT, details: { limit: 100 } });
+    const file = path.join(auditDir(), `${store.utcDay()}.ndjson`);
+    const record = JSON.parse(fs.readFileSync(file, 'utf8').trim());
+    expect(record).toMatchObject({
+      schema_version: 1,
+      actor: { type: 'human', id: 'host:test', email: null, user_id: null, group_ids: null },
+      origin: { transport: 'socket' },
+      action: 'groups.list',
+      outcome: 'success',
+      correlation_id: null,
+      details: { limit: 100 },
+    });
+    expect(record.event_id).toMatch(/^[0-9a-f-]{36}$/);
+    expect(new Date(record.time).toISOString()).toBe(record.time);
+  });
+
+  it('redacts details at the emit seam', () => {
+    emit.emitAuditEvent({ ...EVENT_INPUT, details: { env: { API_TOKEN: 'x' } } });
+    const file = path.join(auditDir(), `${store.utcDay()}.ndjson`);
+    expect(fs.readFileSync(file, 'utf8')).toContain('"API_TOKEN":"[REDACTED]"');
+  });
+
+  it('is a no-op when audit is disabled — the directory is never created', () => {
+    state.enabled = false;
+    emit.emitAuditEvent(EVENT_INPUT);
+    expect(fs.existsSync(auditDir())).toBe(false);
+  });
+
+  it('fails open and loud when the append fails', () => {
+    fs.mkdirSync(auditDir(), { recursive: true });
+    fs.chmodSync(auditDir(), 0o500);
+    expect(() => emit.emitAuditEvent(EVENT_INPUT)).not.toThrow();
+    expect(log.error).toHaveBeenCalledWith(
+      expect.stringContaining('Audit append failed'),
+      expect.objectContaining({ action: 'groups.list' }),
+    );
+  });
+});
+
+describe('assertAuditWritable', () => {
+  it('creates the directory and probes it with a zero-byte append', () => {
+    store.assertAuditWritable();
+    expect(fs.existsSync(path.join(auditDir(), `${store.utcDay()}.ndjson`))).toBe(true);
+  });
+
+  it('throws when the directory is not writable', () => {
+    fs.mkdirSync(auditDir(), { recursive: true });
+    fs.chmodSync(auditDir(), 0o500);
+    expect(() => store.assertAuditWritable()).toThrow();
+  });
+});
+
+describe('pruneAuditLog', () => {
+  it('unlinks only day-files strictly older than the horizon', () => {
+    writeDayFile(dayString(100));
+    writeDayFile(dayString(91));
+    writeDayFile(dayString(90));
+    writeDayFile(dayString(1));
+    fs.writeFileSync(path.join(auditDir(), 'not-a-day-file.txt'), 'keep');
+    fs.writeFileSync(path.join(auditDir(), '2026-01-01.ndjson.bak'), 'keep');
+    store.pruneAuditLog(90);
+    const left = fs.readdirSync(auditDir()).sort();
+    expect(left).toEqual(
+      [`${dayString(90)}.ndjson`, `${dayString(1)}.ndjson`, '2026-01-01.ndjson.bak', 'not-a-day-file.txt'].sort(),
+    );
+  });
+
+  it('keeps forever when retention is 0', () => {
+    writeDayFile(dayString(400));
+    store.pruneAuditLog(0);
+    expect(fs.existsSync(path.join(auditDir(), `${dayString(400)}.ndjson`))).toBe(true);
+  });
+
+  it('is a no-op when the audit directory does not exist', () => {
+    expect(() => store.pruneAuditLog(90)).not.toThrow();
+  });
+});
+
+describe('pruneAuditLogIfDue', () => {
+  it('prunes at most once per UTC day', () => {
+    writeDayFile(dayString(100));
+    store.pruneAuditLogIfDue();
+    expect(fs.existsSync(path.join(auditDir(), `${dayString(100)}.ndjson`))).toBe(false);
+    writeDayFile(dayString(100));
+    store.pruneAuditLogIfDue();
+    expect(fs.existsSync(path.join(auditDir(), `${dayString(100)}.ndjson`))).toBe(true);
+  });
+
+  it('does nothing when audit is disabled', () => {
+    state.enabled = false;
+    writeDayFile(dayString(100));
+    store.pruneAuditLogIfDue();
+    expect(fs.existsSync(path.join(auditDir(), `${dayString(100)}.ndjson`))).toBe(true);
+  });
+
+  it('does nothing after markPrunedToday until the day rolls over', () => {
+    store.markPrunedToday();
+    writeDayFile(dayString(100));
+    store.pruneAuditLogIfDue();
+    expect(fs.existsSync(path.join(auditDir(), `${dayString(100)}.ndjson`))).toBe(true);
+  });
+});

.claude/skills/add-audit/add/src/audit/store.ts

@@ -0,0 +1,81 @@
+/**
+ * Day-file store — daily NDJSON files under data/audit/, host process is the
+ * single writer. Append-only is structural: nothing here can update a line,
+ * and retention is unlinking whole files (a literal hard delete). Both
+ * transports converge host-side, so single-writer holds by construction.
+ *
+ * This module throws on fs failure; the fail-open posture (log.error, action
+ * proceeds) lives in emit.ts, and the boot-time strictness (refuse to start)
+ * lives in init.ts.
+ */
+import fs from 'fs';
+import path from 'path';
+
+import { DATA_DIR } from '../config.js';
+import { log } from '../log.js';
+import { AUDIT_ENABLED, AUDIT_RETENTION_DAYS } from './config.js';
+
+export const AUDIT_DIR = path.join(DATA_DIR, 'audit');
+const DAY_FILE_RE = /^(\d{4}-\d{2}-\d{2})\.ndjson$/;
+const DAY_MS = 24 * 60 * 60 * 1000;
+
+export function utcDay(d: Date = new Date()): string {
+  return d.toISOString().slice(0, 10);
+}
+
+export function dayFilePath(day: string): string {
+  return path.join(AUDIT_DIR, `${day}.ndjson`);
+}
+
+/** The single append point. Throws on fs failure — emitAuditEvent catches. */
+export function appendAuditLine(line: string): void {
+  fs.mkdirSync(AUDIT_DIR, { recursive: true });
+  fs.appendFileSync(dayFilePath(utcDay()), line + '\n');
+}
+
+/** Boot-time writability assert — a zero-byte append is a true write probe. */
+export function assertAuditWritable(): void {
+  fs.mkdirSync(AUDIT_DIR, { recursive: true });
+  fs.appendFileSync(dayFilePath(utcDay()), '');
+}
+
+/** Unlink day-files strictly older than (today UTC − retentionDays). 0 or negative = keep forever. */
+export function pruneAuditLog(retentionDays: number = AUDIT_RETENTION_DAYS): void {
+  if (retentionDays <= 0) return;
+  let entries: string[];
+  try {
+    entries = fs.readdirSync(AUDIT_DIR);
+    // eslint-disable-next-line no-catch-all/no-catch-all -- no audit dir yet means nothing to prune, not an error
+  } catch {
+    return;
+  }
+  const horizon = utcDay(new Date(Date.now() - retentionDays * DAY_MS));
+  let pruned = 0;
+  for (const entry of entries) {
+    const m = DAY_FILE_RE.exec(entry);
+    if (!m || m[1] >= horizon) continue;
+    try {
+      fs.unlinkSync(path.join(AUDIT_DIR, entry));
+      pruned++;
+      // eslint-disable-next-line no-catch-all/no-catch-all -- one stuck file must not stop the prune of the others
+    } catch (err) {
+      log.error('Audit prune failed to unlink day-file', { file: entry, err });
+    }
+  }
+  if (pruned > 0) log.info('Audit retention pruned day-files', { pruned, retentionDays });
+}
+
+let lastPruneDay: string | null = null;
+
+/** Retention prune, throttled to once per UTC day — safe to call every tick. */
+export function pruneAuditLogIfDue(): void {
+  if (!AUDIT_ENABLED || AUDIT_RETENTION_DAYS <= 0) return;
+  const today = utcDay();
+  if (lastPruneDay === today) return;
+  lastPruneDay = today;
+  pruneAuditLog();
+}
+
+export function markPrunedToday(): void {
+  lastPruneDay = utcDay();
+}

.claude/skills/add-audit/add/src/audit/types.ts

@@ -0,0 +1,60 @@
+/**
+ * Canonical audit event — one flat, SIEM-shaped record per action.
+ *
+ * Fields are chosen so they project losslessly onto OCSF and Elastic ECS
+ * (the two schemas SIEMs converge on); the store keeps the neutral shape and
+ * a future forwarder pays the translation once, at the edge. `schema_version`
+ * covers evolution — additive changes only within a version.
+ */
+
+export type AuditActorType = 'human' | 'agent' | 'system';
+
+export interface AuditActor {
+  type: AuditActorType;
+  /** `host:<os-user>` | `<channel>:<handle>` | agent group id | `host` (system). */
+  id: string;
+}
+
+export interface AuditOrigin {
+  transport: 'socket' | 'container' | 'channel';
+  session_id?: string;
+  messaging_group_id?: string;
+  channel?: string;
+}
+
+export interface AuditResource {
+  type: string;
+  /** Omitted when only the attempted type is known (e.g. a denied list). */
+  id?: string;
+}
+
+export type AuditOutcome = 'success' | 'failure' | 'denied' | 'pending' | 'approved' | 'rejected';
+
+/** What emit sites provide. Envelope fields are stamped by emitAuditEvent. */
+export interface AuditEventInput {
+  actor: AuditActor;
+  origin: AuditOrigin;
+  /** Dotted namespaced verb, e.g. `groups.config.add-mcp-server`. */
+  action: string;
+  resources: AuditResource[];
+  outcome: AuditOutcome;
+  /** The approval id on gated chains; null/omitted otherwise. */
+  correlationId?: string | null;
+  details?: Record<string, unknown>;
+}
+
+export interface AuditEvent {
+  event_id: string;
+  time: string;
+  schema_version: 1;
+  actor: AuditActor & { email: null; user_id: null; group_ids: null };
+  origin: AuditOrigin;
+  action: string;
+  resources: AuditResource[];
+  outcome: AuditOutcome;
+  correlation_id: string | null;
+  details: Record<string, unknown>;
+}
+
+/** Actor for timer/sweep-driven outcomes (expiries, startup sweeps). */
+export const SYSTEM_ACTOR: AuditActor = { type: 'system', id: 'host' };

.claude/skills/add-audit/add/src/audit/vocab.ts

@@ -0,0 +1,39 @@
+/**
+ * Domain-free event vocabulary — actor and origin constructors shared by the
+ * domain-owned `*.audit.ts` adapters (the CLI adapter today; approval and
+ * channel adapters attach here in later increments). Pure derivation; nothing
+ * here writes the log.
+ *
+ * Leaf rule: this module (like the rest of src/audit/) may depend on node,
+ * config/log, shared types, and the db read layer — never on src/cli/* or
+ * src/modules/*. Domain-specific mapping (CLI resources, approval payloads)
+ * lives in the adapter file of the domain that owns it.
+ */
+import os from 'os';
+
+import { getMessagingGroup } from '../db/messaging-groups.js';
+import type { AuditOrigin } from './types.js';
+
+/**
+ * Host callers stamp `host:<install user>` daemon-side: the ncl socket is
+ * 0600 and owned by the install user, so the identity is accurate by
+ * construction without peer credentials.
+ */
+export function hostUser(): string {
+  try {
+    return os.userInfo().username;
+    // eslint-disable-next-line no-catch-all/no-catch-all -- os.userInfo throws on exotic hosts; a fallback actor id beats no audit event
+  } catch {
+    return process.env.USER || 'unknown';
+  }
+}
+
+export function containerOrigin(sessionId: string, messagingGroupId: string | null): AuditOrigin {
+  const origin: AuditOrigin = { transport: 'container', session_id: sessionId };
+  if (messagingGroupId) {
+    origin.messaging_group_id = messagingGroupId;
+    const channel = getMessagingGroup(messagingGroupId)?.channel_type;
+    if (channel) origin.channel = channel;
+  }
+  return origin;
+}

.claude/skills/add-audit/add/src/cli/dispatch.audit.test.ts

@@ -0,0 +1,297 @@
+/**
+ * Audit middleware behavior of the exported dispatch — what gets recorded,
+ * for whom, and how gated chains correlate. Drives the real wrapped dispatch
+ * (real registry, real guard); audit is force-enabled and the store's append
+ * is captured. DB reads and approval delivery are mocked.
+ */
+import os from 'os';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+import type { PendingApproval } from '../types.js';
+
+const appended = vi.hoisted(() => ({ lines: [] as string[] }));
+const pendingRows = vi.hoisted(() => ({ rows: [] as unknown[] }));
+
+vi.mock('../audit/config.js', () => ({
+  AUDIT_ENABLED: true,
+  AUDIT_RETENTION_DAYS: 90,
+}));
+
+// Neutralize the adapter's module-scope boot (writability assert, prune,
+// maintenance timer) — the middleware is the unit under test here.
+vi.mock('../audit/init.js', () => ({
+  initAuditLog: vi.fn(),
+  maintainAudit: vi.fn(),
+}));
+
+vi.mock('../audit/store.js', async (importOriginal) => {
+  const actual = await importOriginal<typeof import('../audit/store.js')>();
+  return {
+    ...actual,
+    appendAuditLine: (line: string) => {
+      appended.lines.push(line);
+    },
+  };
+});
+
+vi.mock('../log.js', () => ({
+  log: { debug: vi.fn(), info: vi.fn(), warn: vi.fn(), error: vi.fn(), fatal: vi.fn() },
+}));
+
+const mockGetContainerConfig = vi.fn();
+vi.mock('../db/container-configs.js', () => ({
+  getContainerConfig: (...args: unknown[]) => mockGetContainerConfig(...args),
+}));
+
+vi.mock('../db/agent-groups.js', () => ({
+  getAgentGroup: vi.fn(() => ({ id: 'g1', name: 'Group One' })),
+}));
+
+const mockGetPendingApproval = vi.fn();
+vi.mock('../db/sessions.js', () => ({
+  getSession: vi.fn(() => ({ id: 's1', agent_group_id: 'g1', messaging_group_id: 'mg1' })),
+  getPendingApproval: (...args: unknown[]) => mockGetPendingApproval(...args),
+  getPendingApprovalsByAction: () => pendingRows.rows,
+}));
+
+vi.mock('../db/messaging-groups.js', () => ({
+  getMessagingGroup: vi.fn(() => ({ channel_type: 'slack' })),
+}));
+
+const mockGetResource = vi.fn();
+vi.mock('./crud.js', () => ({
+  getResource: (...args: unknown[]) => mockGetResource(...args),
+}));
+
+vi.mock('../modules/approvals/index.js', () => ({
+  registerApprovalHandler: vi.fn(),
+  requestApproval: vi.fn(async () => undefined),
+}));
+
+import { register } from './registry.js';
+
+register({
+  name: 'groups-test',
+  description: 'echo command on the groups resource',
+  action: 'groups.test',
+  resource: 'groups',
+  access: 'open',
+  parseArgs: (raw) => raw,
+  handler: async (args) => ({ echo: args }),
+});
+
+register({
+  name: 'groups-get',
+  description: 'echo command for dash-joined id resolution',
+  action: 'groups.get',
+  resource: 'groups',
+  access: 'open',
+  parseArgs: (raw) => raw,
+  handler: async (args) => ({ echo: args }),
+});
+
+register({
+  name: 'wirings-list',
+  description: 'not on the group-scope allowlist',
+  action: 'wirings.list',
+  resource: 'wirings',
+  access: 'open',
+  parseArgs: (raw) => raw,
+  handler: async () => [],
+});
+
+register({
+  name: 'groups-fail',
+  description: 'handler that throws',
+  action: 'groups.fail',
+  resource: 'groups',
+  access: 'open',
+  parseArgs: (raw) => raw,
+  handler: async () => {
+    throw new Error('boom');
+  },
+});
+
+register({
+  name: 'groups-gated',
+  description: 'approval-gated command',
+  action: 'groups.gated',
+  resource: 'groups',
+  access: 'approval',
+  parseArgs: (raw) => raw,
+  handler: async () => 'ran',
+});
+
+import { dispatch } from './dispatch.js';
+import type { CallerContext } from './frame.js';
+
+const AGENT_CTX: CallerContext = { caller: 'agent', sessionId: 's1', agentGroupId: 'g1', messagingGroupId: 'mg1' };
+
+function grantRow(frameId: string, command: string): PendingApproval {
+  return {
+    approval_id: 'appr-123-abc',
+    session_id: 's1',
+    request_id: 'appr-123-abc',
+    action: 'cli_command',
+    payload: JSON.stringify({ frame: { id: frameId, command, args: {} }, callerContext: AGENT_CTX }),
+    created_at: new Date().toISOString(),
+    agent_group_id: 'g1',
+    channel_type: null,
+    platform_id: null,
+    platform_message_id: null,
+    expires_at: null,
+    status: 'pending',
+    title: 'CLI: groups-gated',
+    options_json: '[]',
+    approver_user_id: null,
+  };
+}
+
+function events(): Array<Record<string, any>> {
+  return appended.lines.map((l) => JSON.parse(l));
+}
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  appended.lines.length = 0;
+  pendingRows.rows = [];
+  mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+  mockGetResource.mockImplementation((plural: string) => (plural === 'groups' ? { scopeField: 'id' } : undefined));
+});
+
+describe('withAudit(dispatch)', () => {
+  it('records a success event for a host caller with socket origin and host actor', async () => {
+    const resp = await dispatch({ id: '1', command: 'groups-test', args: { foo: 'bar' } }, { caller: 'host' });
+
+    expect(resp.ok).toBe(true);
+    const [event] = events();
+    expect(event).toMatchObject({
+      schema_version: 1,
+      actor: { type: 'human', id: `host:${os.userInfo().username}`, email: null },
+      origin: { transport: 'socket' },
+      action: 'groups.test',
+      outcome: 'success',
+      correlation_id: null,
+      details: { foo: 'bar' },
+    });
+  });
+
+  it('records effective args after group auto-fill, with container origin and channel', async () => {
+    await dispatch({ id: '1', command: 'groups-test', args: {} }, AGENT_CTX);
+
+    const [event] = events();
+    expect(event.actor).toMatchObject({ type: 'agent', id: 'g1' });
+    expect(event.origin).toEqual({
+      transport: 'container',
+      session_id: 's1',
+      messaging_group_id: 'mg1',
+      channel: 'slack',
+    });
+    expect(event.details).toMatchObject({ id: 'g1', agent_group_id: 'g1', group: 'g1' });
+    expect(event.resources).toContainEqual({ type: 'agent_group', id: 'g1' });
+  });
+
+  it('records a denied event for a scope denial, naming the attempted resource type', async () => {
+    const resp = await dispatch({ id: '1', command: 'wirings-list', args: {} }, AGENT_CTX);
+
+    expect(resp.ok).toBe(false);
+    const [event] = events();
+    expect(event).toMatchObject({
+      action: 'wirings.list',
+      outcome: 'denied',
+      resources: [{ type: 'wirings' }],
+      details: { error: 'forbidden' },
+    });
+    expect(event.details.reason).toContain('scoped');
+  });
+
+  it('records a failure event when the handler throws', async () => {
+    await dispatch({ id: '1', command: 'groups-fail', args: {} }, { caller: 'host' });
+
+    const [event] = events();
+    expect(event).toMatchObject({ action: 'groups.fail', outcome: 'failure', details: { error: 'handler-error' } });
+    expect(event.details.reason).toContain('boom');
+  });
+
+  it('records a hold as a pending event correlated to the approval row it created', async () => {
+    pendingRows.rows = [grantRow('1', 'groups-gated')];
+
+    const resp = await dispatch({ id: '1', command: 'groups-gated', args: {} }, AGENT_CTX);
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) expect(resp.error.code).toBe('approval-pending');
+    const [event] = events();
+    expect(event).toMatchObject({
+      action: 'groups.gated',
+      outcome: 'pending',
+      correlation_id: 'appr-123-abc',
+    });
+    expect(event.resources).toContainEqual({ type: 'approval', id: 'appr-123-abc' });
+    expect(event.details.error).toBeUndefined();
+  });
+
+  it('records an uncorrelated pending event when no approval row was created (no approver)', async () => {
+    pendingRows.rows = [];
+
+    await dispatch({ id: '1', command: 'groups-gated', args: {} }, AGENT_CTX);
+
+    const [event] = events();
+    expect(event).toMatchObject({ outcome: 'pending', correlation_id: null });
+  });
+
+  it('records an approved replay as success with the grant approval id as correlation_id', async () => {
+    const grant = grantRow('9', 'groups-gated');
+    mockGetPendingApproval.mockReturnValue(grant);
+
+    const resp = await dispatch({ id: '9', command: 'groups-gated', args: {} }, AGENT_CTX, { grant });
+
+    expect(resp.ok).toBe(true);
+    const [event] = events();
+    expect(event).toMatchObject({
+      action: 'groups.gated',
+      outcome: 'success',
+      correlation_id: 'appr-123-abc',
+    });
+    expect(event.resources).toContainEqual({ type: 'approval', id: 'appr-123-abc' });
+  });
+
+  it('records unknown commands as cli.unknown-command with the raw name in details', async () => {
+    await dispatch({ id: '1', command: 'nope-nothing', args: {} }, { caller: 'host' });
+
+    const [event] = events();
+    expect(event).toMatchObject({
+      action: 'cli.unknown-command',
+      outcome: 'failure',
+      resources: [],
+      details: { command: 'nope-nothing', error: 'unknown-command' },
+    });
+  });
+
+  it('records the resolved command and id for dash-joined positional ids', async () => {
+    const uuid = '550e8400-e29b-41d4-a716-446655440000';
+    await dispatch({ id: '1', command: `groups-get-${uuid}`, args: {} }, { caller: 'host' });
+
+    const [event] = events();
+    expect(event).toMatchObject({ action: 'groups.get', outcome: 'success' });
+    expect(event.resources).toContainEqual({ type: 'agent_group', id: uuid });
+    expect(event.details.id).toBe(uuid);
+  });
+
+  it('normalizes hyphenated arg keys in details', async () => {
+    await dispatch({ id: '1', command: 'groups-test', args: { 'dry-run': 'true' } }, { caller: 'host' });
+
+    const [event] = events();
+    expect(event.details).toMatchObject({ dry_run: 'true' });
+  });
+
+  it('parses JSON string args so the redactor reaches their inner keys', async () => {
+    await dispatch(
+      { id: '1', command: 'groups-test', args: { env: '{"NOTION_TOKEN":"tok-123","SAFE":"ok"}', note: '{not json' } },
+      { caller: 'host' },
+    );
+
+    const [event] = events();
+    expect(event.details.env).toEqual({ NOTION_TOKEN: '[REDACTED]', SAFE: 'ok' });
+    expect(event.details.note).toBe('{not json');
+  });
+});

.claude/skills/add-audit/add/src/cli/dispatch.audit.ts

@@ -0,0 +1,238 @@
+/**
+ * CLI audit adapter (installed by /add-audit) — owns how the dispatcher
+ * describes itself to the audit log: the dispatch middleware plus the
+ * CLI-specific actor/origin/resource mapping. Composed in dispatch.ts as
+ * `export const dispatch = withAudit(dispatchInner)`; business logic there
+ * contains zero audit calls.
+ *
+ * Loading this module also boots the audit log (writability assert, boot
+ * prune, hook lifecycle, maintenance timer): dispatch.ts is imported by both
+ * transports during the host's barrel phase, so initAuditLog() runs before
+ * any command is accepted — and an enabled box with an unwritable
+ * data/audit/ refuses to start.
+ */
+import { emitAuditEvent } from '../audit/emit.js';
+import { initAuditLog } from '../audit/init.js';
+import { type AuditActor, type AuditOrigin, type AuditOutcome, type AuditResource } from '../audit/types.js';
+import { containerOrigin, hostUser } from '../audit/vocab.js';
+import { getContainerConfig } from '../db/container-configs.js';
+import { getPendingApprovalsByAction } from '../db/sessions.js';
+import type { PendingApproval } from '../types.js';
+import { getResource } from './crud.js';
+import type { CallerContext, RequestFrame, ResponseFrame } from './frame.js';
+import { type CommandDef, lookup } from './registry.js';
+
+initAuditLog();
+
+// ── CLI mapping ──
+
+/**
+ * Host callers stamp `host:<install user>` daemon-side (the ncl socket is
+ * 0600 and owned by the install user); container callers are their agent group.
+ */
+export function actorForCaller(ctx: CallerContext): AuditActor {
+  return ctx.caller === 'host' ? { type: 'human', id: `host:${hostUser()}` } : { type: 'agent', id: ctx.agentGroupId };
+}
+
+export function originForCaller(ctx: CallerContext): AuditOrigin {
+  if (ctx.caller === 'host') return { transport: 'socket' };
+  return containerOrigin(ctx.sessionId, ctx.messagingGroupId || null);
+}
+
+/**
+ * Frame-level args use `--hyphen-keys`; recorded details use the same
+ * underscore form the parsed handlers see. Mirrors crud's normalizeArgs
+ * (kept local so audit doesn't depend on a module tests commonly mock).
+ */
+export function normalizeArgKeys(raw: Record<string, unknown>): Record<string, unknown> {
+  const out: Record<string, unknown> = {};
+  for (const [k, v] of Object.entries(raw)) {
+    out[k.replace(/-/g, '_')] = v;
+  }
+  return out;
+}
+
+/** CLI resource plural → audit resource type, where the singular isn't it. */
+const RESOURCE_TYPE_OVERRIDES: Record<string, string> = {
+  groups: 'agent_group',
+  'messaging-groups': 'messaging_group',
+  'dropped-messages': 'dropped_message',
+  'user-dms': 'user_dm',
+};
+
+/**
+ * Derive touched/attempted resources from a command's effective args. Generic
+ * by design: `id` → the command's own resource, group/user args → their
+ * types, and a bare `{type}` entry when nothing else is known (a denied
+ * `users list` still names what was attempted).
+ */
+export function resourcesForCli(cmd: CommandDef, args: Record<string, unknown>): AuditResource[] {
+  if (!cmd.resource) return [];
+  const type = RESOURCE_TYPE_OVERRIDES[cmd.resource] ?? getResource(cmd.resource)?.name ?? cmd.resource;
+
+  const out: AuditResource[] = [];
+  const push = (t: string, id: unknown): void => {
+    if (typeof id !== 'string' || !id) return;
+    if (!out.some((r) => r.type === t && r.id === id)) out.push({ type: t, id });
+  };
+  push(type, args.id);
+  push('agent_group', args.agent_group_id ?? args.group);
+  push('user', args.user);
+  if (out.length === 0) out.push({ type });
+  return out;
+}
+
+// ── Dispatch mechanics, mirrored for the record ──
+// Dispatch resolves the command and auto-fills args on a local copy of the
+// frame that never leaves it, so the middleware mirrors the two documented
+// mechanics below. The copies are mechanical, and drift only ever degrades a
+// record's detail (a fallback action name, a missing auto-filled arg) — never
+// dispatch behavior, and never an outcome.
+
+/**
+ * Mirror of dispatch's command resolution: exact lookup, then the longest
+ * registered dash-prefix with the remainder recorded as --id.
+ */
+function resolveForRecord(req: RequestFrame): { cmd?: CommandDef; args: Record<string, unknown> } {
+  const direct = lookup(req.command);
+  if (direct) return { cmd: direct, args: req.args };
+  let shortened = req.command;
+  let idx: number;
+  while ((idx = shortened.lastIndexOf('-')) > 0) {
+    shortened = shortened.slice(0, idx);
+    const fallback = lookup(shortened);
+    if (fallback) {
+      const tail = req.command.slice(shortened.length + 1);
+      return { cmd: fallback, args: { ...req.args, id: req.args.id ?? tail } };
+    }
+  }
+  return { args: req.args };
+}
+
+/**
+ * Mirror of dispatch's group-scope arg auto-fill, so the record shows the
+ * effective args the handler saw (e.g. which agent group a bare
+ * `sessions list` actually listed).
+ */
+function effectiveArgs(
+  cmd: CommandDef | undefined,
+  args: Record<string, unknown>,
+  ctx: CallerContext,
+): Record<string, unknown> {
+  if (ctx.caller !== 'agent') return args;
+  if ((getContainerConfig(ctx.agentGroupId)?.cli_scope ?? 'group') !== 'group') return args;
+  const fill: Record<string, unknown> = {
+    agent_group_id: args.agent_group_id ?? ctx.agentGroupId,
+    group: args.group ?? ctx.agentGroupId,
+  };
+  if (cmd?.resource === 'groups' || cmd?.resource === 'destinations') {
+    fill.id = args.id ?? ctx.agentGroupId;
+  }
+  return { ...args, ...fill };
+}
+
+/**
+ * CLI args arrive as strings, so a JSON-object/array value (e.g.
+ * `--env '{"NOTION_TOKEN":"…"}'`) would reach the redactor as one opaque
+ * string under an innocent key and its inner secret keys would survive.
+ * Recording the parsed form lets the redactor walk inside — the audit log's
+ * "secrets never land" property depends on it for exactly this flow.
+ */
+function parseJsonishValues(args: Record<string, unknown>): Record<string, unknown> {
+  const out: Record<string, unknown> = {};
+  for (const [k, v] of Object.entries(args)) {
+    if (typeof v === 'string' && (v.startsWith('{') || v.startsWith('['))) {
+      try {
+        out[k] = JSON.parse(v);
+        continue;
+        // eslint-disable-next-line no-catch-all/no-catch-all -- not JSON after all: record the string as-is
+      } catch {
+        /* fall through */
+      }
+    }
+    out[k] = v;
+  }
+  return out;
+}
+
+/**
+ * The approval row a hold just created for this frame — it gives the pending
+ * event the same correlation_id the approved replay will carry as its guard
+ * grant. requestApproval keeps the minted id internal, so the row is
+ * recovered by the frame id it stored in its payload; no row (e.g. no
+ * configured approver) → the hold is still recorded, uncorrelated.
+ */
+function holdApprovalIdFor(frameId: string): string | null {
+  const rows = getPendingApprovalsByAction('cli_command');
+  for (let i = rows.length - 1; i >= 0; i--) {
+    try {
+      const payload = JSON.parse(rows[i].payload) as { frame?: { id?: string } };
+      if (payload.frame?.id === frameId) return rows[i].approval_id;
+      // eslint-disable-next-line no-catch-all/no-catch-all -- a row with an unparseable payload is simply not this frame's hold
+    } catch {
+      continue;
+    }
+  }
+  return null;
+}
+
+// ── The dispatch middleware ──
+
+type DispatchInner = (
+  req: RequestFrame,
+  ctx: CallerContext,
+  opts?: { grant?: PendingApproval },
+) => Promise<ResponseFrame>;
+
+/**
+ * Dispatch middleware — the exported `dispatch` is the wrapped function, so
+ * the socket server, the container delivery-action, and the in-module
+ * approved replay are all covered by the one composition.
+ *
+ * Outcome derives from the response frame: ok → success, forbidden → denied
+ * (captures pre-handler scope denials), approval-pending → pending (the
+ * record of a hold), anything else → failure. Correlation is the approval id:
+ * an approved replay carries the approval row as its guard grant
+ * (opts.grant), and a fresh hold is correlated by recovering the row it just
+ * created — so `--correlation <approval-id>` returns the whole gated chain.
+ */
+export function withAudit(inner: DispatchInner): DispatchInner {
+  return async (req, ctx, opts = {}) => {
+    const res = await inner(req, ctx, opts);
+    emitAuditEvent(() => {
+      const resolved = resolveForRecord(req);
+      const cmd = resolved.cmd;
+      const pending = !res.ok && res.error.code === 'approval-pending';
+      const outcome: AuditOutcome = res.ok
+        ? 'success'
+        : res.error.code === 'forbidden'
+          ? 'denied'
+          : pending
+            ? 'pending'
+            : 'failure';
+      // A denial records the attempt as asked (raw args): nothing ran, so the
+      // auto-fill never conceptually happened — and a cross-group attempt
+      // must show the foreign id the caller passed, not a filled-in own-group.
+      const args = normalizeArgKeys(outcome === 'denied' ? resolved.args : effectiveArgs(cmd, resolved.args, ctx));
+      const correlationId = opts.grant?.approval_id ?? (pending ? holdApprovalIdFor(req.id) : null);
+      const details: Record<string, unknown> = parseJsonishValues(args);
+      if (!res.ok && !pending) {
+        details.error = res.error.code;
+        details.reason = res.error.message;
+      }
+      if (!cmd) details.command = req.command;
+      const resources = cmd ? resourcesForCli(cmd, args) : [];
+      if (correlationId) resources.push({ type: 'approval', id: correlationId });
+      return {
+        actor: actorForCaller(ctx),
+        origin: originForCaller(ctx),
+        action: cmd ? (cmd.action ?? `cli.${cmd.name}`) : 'cli.unknown-command',
+        resources,
+        outcome,
+        correlationId,
+        details,
+      };
+    });
+    return res;
+  };
+}

.claude/skills/add-audit/add/src/cli/resources/audit.ts

@@ -0,0 +1,56 @@
+import { listAuditEvents } from '../../audit/reader.js';
+import { registerResource } from '../crud.js';
+
+/**
+ * Read-only audit resource (installed by /add-audit). The "table" is the
+ * NDJSON day-file store — no generic CRUD verbs are declared, so it is never
+ * queried as SQL; `list` is a custom operation backed by the audit reader.
+ *
+ * Deliberately NOT on the group-scope allowlist (GROUP_SCOPE_RESOURCES):
+ * audit spans agent groups, so group-scoped agents are refused before the
+ * handler — the resource is host + `cli_scope: global` only, by omission
+ * (fails closed).
+ */
+registerResource({
+  name: 'audit_event',
+  plural: 'audit',
+  table: '(data/audit/*.ndjson)',
+  description: 'Local audit log — one event per ncl command. Newest first. Requires AUDIT_ENABLED=true.',
+  idColumn: 'event_id',
+  columns: [
+    { name: 'actor', type: 'string', description: 'Filter: exact actor id (host:<user>, <channel>:<handle>, agent group id)' },
+    { name: 'action', type: 'string', description: 'Filter: exact action or dotted prefix (e.g. groups.config)' },
+    { name: 'resource', type: 'string', description: 'Filter: matches any event resource by id or type' },
+    {
+      name: 'outcome',
+      type: 'string',
+      description: 'Filter: event outcome',
+      enum: ['success', 'failure', 'denied', 'pending', 'approved', 'rejected'],
+    },
+    { name: 'since', type: 'string', description: 'Window start: 7d / 24h / 30m relative, or ISO date' },
+    { name: 'until', type: 'string', description: 'Window end: same formats as --since' },
+    { name: 'correlation', type: 'string', description: 'Filter: approval id tying a gated chain together' },
+    { name: 'limit', type: 'number', description: 'Max events (default 100, newest first)' },
+    { name: 'format', type: 'string', description: 'Output format', enum: ['ndjson'] },
+  ],
+  operations: {},
+  customOperations: {
+    list: {
+      access: 'open',
+      description: 'Query audit events, newest first. --format ndjson streams the stored lines for SIEM export.',
+      examples: [
+        'ncl audit list --outcome denied --since 7d',
+        'ncl audit list --correlation appr-1751970000000-x1y2z3',
+        'ncl audit list --action groups.config --format ndjson',
+      ],
+      handler: async (args) => listAuditEvents(args),
+      formatHuman: (data) => {
+        // NDJSON export prints the stored lines verbatim; for the table view
+        // this throws so dispatch falls back to the row objects, which every
+        // client already renders.
+        if (typeof data === 'string') return data;
+        throw new Error('render rows client-side');
+      },
+    },
+  },
+});

.claude/skills/add-mnemon/SKILL.md

@@ -9,13 +9,13 @@ Installs [mnemon](https://github.com/mnemon-dev/mnemon) in the agent container i
 
 ## Provider Compatibility
 
-mnemon hooks fire only under `--target claude-code`. Use this skill on agent groups that run the default Claude provider (`AGENT_PROVIDER=claude`). Confirm the provider before applying:
+mnemon hooks fire only under `--target claude-code`. Use this skill on agent groups that run the default Claude provider. The provider is the materialized `provider` key in each group's `container.json` (absent or `claude` = default Claude provider). Confirm it before applying:
 
 ```bash
-grep AGENT_PROVIDER .env groups/*/container.json 2>/dev/null
+grep -H '"provider"' groups/*/container.json 2>/dev/null   # no match, or "provider": "claude" = Claude
 ```
 
-If a group uses a different provider (e.g. `AGENT_PROVIDER=opencode`), it spawns its own process and never invokes the `claude` CLI, so the hooks registered by `mnemon setup` do not run for that group.
+If a group sets a different provider (e.g. `"provider": "opencode"`), it spawns its own process and never invokes the `claude` CLI, so the hooks registered by `mnemon setup` do not run for that group.
 
 ## Phase 1: Pre-flight
 

.claude/skills/add-opencode/SKILL.md

@@ -1,11 +1,11 @@
 ---
 name: add-opencode
-description: Use OpenCode as an agent provider (AGENT_PROVIDER=opencode). OpenRouter, OpenAI, Google, DeepSeek, etc. via OpenCode config — not the Anthropic Agent SDK. Per-session and per-group via agent_provider; host passes OPENCODE_* and XDG mount when spawning containers.
+description: Use OpenCode as an agent provider. OpenRouter, OpenAI, Google, DeepSeek, etc. via OpenCode config — not the Anthropic Agent SDK. Per group via `ncl groups config update --provider opencode`; host passes OPENCODE_* and XDG mount when spawning containers.
 ---
 
 # OpenCode agent provider
 
-NanoClaw runs agents in a long-lived **poll loop** inside the container. The backend is selected with **`AGENT_PROVIDER`** (`claude` | `opencode` | `mock`).
+NanoClaw runs agents in a long-lived **poll loop** inside the container. The backend is selected per agent group by the **`provider`** key in that group's `container.json` (materialized from the `container_configs` table) — set it with `ncl groups config update --provider opencode`. Default is `claude`.
 
 Trunk ships with only the `claude` provider baked in. This skill copies the OpenCode provider files in from the `providers` branch, wires them into the host and container barrels, installs dependencies, and rebuilds the image.
 
@@ -148,7 +148,7 @@ done
 
 Set model/provider strings in the form OpenCode expects (often `provider/model-id`). **Put comments on their own lines** — a `#` inside a value is kept verbatim and breaks model IDs.
 
-These variables are read **on the host** and passed into the container only when the effective provider is `opencode`. They do not switch the provider by themselves; the DB still needs `agent_provider` set (below).
+These variables are read **on the host** and passed into the container only when the effective provider is `opencode`. They do not switch the provider by themselves; the group still needs `provider` set to `opencode` (see [Select the provider](#select-the-provider) below).
 
 - `OPENCODE_PROVIDER` — OpenCode provider id, e.g. `openrouter`, `anthropic`, `deepseek`.
 - `OPENCODE_MODEL` — full model id in `provider/model` form, e.g. `deepseek/deepseek-chat`.
@@ -215,7 +215,7 @@ OPENCODE_SMALL_MODEL=anthropic/claude-haiku-4-5-20251001
 
 Zen's HTTP API (e.g. `POST …/zen/v1/messages`) expects the key in the **`x-api-key`** header. If OneCLI injects **`Authorization: Bearer …`** only, Zen often returns **401 / "Missing API key"** even though the gateway is working.
 
-**Naming:** NanoClaw **`AGENT_PROVIDER=opencode`** (DB `agent_provider`) means "run the **OpenCode agent provider**." Separately, **`OPENCODE_PROVIDER=opencode`** in `.env` is OpenCode's **Zen provider id** inside the OpenCode config (see [Zen docs](https://opencode.ai/docs/zen/)).
+**Naming:** NanoClaw's **`provider: opencode`** (the `container.json` key, set via `ncl groups config update --provider opencode`) means "run the **OpenCode agent provider**." Separately, **`OPENCODE_PROVIDER=opencode`** in `.env` is OpenCode's **Zen provider id** inside the OpenCode config (see [Zen docs](https://opencode.ai/docs/zen/)).
 
 **Host `.env` (typical Zen shape):**
 
@@ -236,9 +236,16 @@ onecli secrets create --name "OpenCode Zen" --type generic \
   --header-name "x-api-key" --value-format "{value}"
 ```
 
-### Per group / per session
+### Select the provider
 
-Set `"provider": "opencode"` in the group's **`container.json`** (`groups/<folder>/container.json`) — the in-container runner reads `provider` from there, not from the DB. The DB columns **`agent_groups.agent_provider`** and **`sessions.agent_provider`** (session overrides group) only drive host-side provider contribution — per-session XDG mount, `OPENCODE_*` env passthrough — and do not propagate into `container.json` at spawn time. Set both, or just edit `container.json`; if they disagree, the runner uses `container.json` and the host-side resolver falls back through session → group → `container.json` → `'claude'`.
+Per group, from the host:
+
+```bash
+ncl groups config update --id <group-id> --provider opencode
+ncl groups restart --id <group-id>
+```
+
+`ncl groups config update --provider` writes the `provider` value into the `container_configs` table; the host materializes it into `groups/<folder>/container.json` at spawn time and the in-container runner reads `provider` from there (defaulting to `claude`). The restart picks up the change. Switching is an operator action — run it from the host. Memory does NOT carry over automatically between providers — run `/migrate-memory` to carry it across.
 
 Extra MCP servers still come from **`NANOCLAW_MCP_SERVERS`** / `container_config.mcpServers` on the host; the runner merges them into the same `mcpServers` object passed to **both** Claude and OpenCode providers.
 
@@ -250,6 +257,6 @@ Extra MCP servers still come from **`NANOCLAW_MCP_SERVERS`** / `container_config
 
 ## Next Steps
 
-The registration and Dockerfile guards in step 7 verify the wiring. To confirm an end-to-end round-trip, set `agent_provider = 'opencode'` (or `"provider": "opencode"` in the group's `container.json`) on a test group, register the matching provider key in OneCLI, and send a message. A clean exchange returns the model's reply with no `Unknown provider: opencode` error and no UUID/session warnings in the logs.
+The registration and Dockerfile guards in step 7 verify the wiring. To confirm an end-to-end round-trip, switch a test group with `ncl groups config update --id <group-id> --provider opencode && ncl groups restart --id <group-id>`, register the matching provider key in OneCLI, and send a message. A clean exchange returns the model's reply with no `Unknown provider: opencode` error and no UUID/session warnings in the logs.
 
 To remove this provider, see [REMOVE.md](REMOVE.md).

.claude/skills/manage-mounts/SKILL.md

@@ -13,18 +13,18 @@ Configure which host directories NanoClaw agent containers can access. The mount
 cat ~/.config/nanoclaw/mount-allowlist.json 2>/dev/null || echo "No mount allowlist configured"
 ```
 
-Show the current config to the user in a readable format: which directories are allowed, whether non-main agents are read-only.
+Show the current config to the user in a readable format: which directories are allowed, and whether each is read-only or read-write.
 
 ## Add Directories
 
 Ask which directories the user wants agents to access. For each path:
 - Validate the path exists
-- Ask if it should be read-only for non-main agents (default: yes)
+- Ask if it should be read-write (`allowReadWrite: true`) or read-only (`allowReadWrite: false`, the safer default)
 
 Build the JSON config and write it:
 
 ```bash
-pnpm exec tsx setup/index.ts --step mounts --force -- --json '{"allowedRoots":[{"path":"/path/to/dir","readOnly":false}],"blockedPatterns":[],"nonMainReadOnly":true}'
+pnpm exec tsx setup/index.ts --step mounts --force -- --json '{"allowedRoots":[{"path":"/path/to/dir","allowReadWrite":true}],"blockedPatterns":[]}'
 ```
 
 Use `--force` to overwrite the existing config.
@@ -34,7 +34,7 @@ Use `--force` to overwrite the existing config.
 Read the current config, show it, ask which entry to remove, then write the updated config through the same write path (build the trimmed JSON and pass it to `--step mounts --force -- --json`):
 
 ```bash
-pnpm exec tsx setup/index.ts --step mounts --force -- --json '{"allowedRoots":[],"blockedPatterns":[],"nonMainReadOnly":true}'
+pnpm exec tsx setup/index.ts --step mounts --force -- --json '{"allowedRoots":[],"blockedPatterns":[]}'
 ```
 
 ## Reset to Empty
@@ -45,12 +45,10 @@ pnpm exec tsx setup/index.ts --step mounts --force -- --empty
 
 ## After Changes
 
-Restart the service so containers pick up the new config (the unit/label names are per-install — see `setup/lib/install-slug.sh`).
+The allowlist is read fresh when a container is spawned, so new mounts apply to newly spawned containers automatically — no service restart needed.
 
-Run from your NanoClaw project root:
+To apply the new config to a group that already has a running container, restart just that group:
 
 ```bash
-source setup/lib/install-slug.sh
-launchctl kickstart -k gui/$(id -u)/$(launchd_label)  # macOS
-systemctl --user restart $(systemd_unit)              # Linux
+ncl groups restart --id <group-id>
 ```

.github/workflows/label-pr.yml

@@ -14,6 +14,7 @@ jobs:
     runs-on: ubuntu-latest
     permissions:
       pull-requests: write
+      issues: write # createLabel — auto-provisions the core-team label
     steps:
       - uses: actions/github-script@v7
         with:
@@ -30,6 +31,24 @@ jobs:
 
             if (body.includes('contributing-guide: v1')) labels.push('follows-guidelines');
 
+            // Lowercase GitHub logins; keep in sync with the core team roster.
+            const CORE_TEAM = ['gavrielc', 'koshkoshinsk', 'glifocat', 'gabi-simons', 'omri-maya', 'amit-shafnir', 'moshe-nanoco'];
+            const author = context.payload.pull_request.user.login.toLowerCase();
+            if (CORE_TEAM.includes(author)) {
+              labels.push('core-team');
+              try {
+                await github.rest.issues.createLabel({
+                  owner: context.repo.owner,
+                  repo: context.repo.repo,
+                  name: 'core-team',
+                  color: '1D76DB',
+                  description: 'PR opened by a core team member',
+                });
+              } catch (e) {
+                if (e.status !== 422) throw e; // 422: label already exists
+              }
+            }
+
             if (labels.length > 0) {
               await github.rest.issues.addLabels({
                 owner: context.repo.owner,

CHANGELOG.md

@@ -4,6 +4,9 @@ All notable changes to NanoClaw will be documented in this file.
 
 ## [Unreleased]
 
+- **New skill: `/add-audit` — opt-in local audit log for the `ncl` surface.** Installs an append-only, SIEM-shaped audit log: every command through the dispatcher (both transports — host socket and container — including scope denials, approval holds, and approved replays) becomes one canonical NDJSON event under `data/audit/<UTC-day>.ndjson`. Gated chains share a `correlation_id` (the approval id: the hold's `pending` event and the replay's terminal event carry the same one); `details` pass a recursive secret-key redactor (~2 KB value cap). `AUDIT_RETENTION_DAYS` (default 90; `0` = keep forever) hard-deletes day-files past the horizon at boot and once per UTC day. Read back with `ncl audit list [--actor --action --resource --outcome --since --until --correlation --limit] [--format ndjson]` — host + global-scope callers only (the resource stays off the group-scope allowlist, so group-scoped agents fail closed). In-process exporters plug in via `registerAuditHook` — post-write hooks that fire only after the local append succeeds. Off by default: nothing is persisted (and `data/audit/` is never created) until `AUDIT_ENABLED=true`, an enabled box refuses to boot if the directory isn't writable, and a disabled box answers `ncl audit list` with a clear error instead of an empty list. The skill's whole core footprint is the dispatch composition (`export const dispatch = withAudit(dispatchInner)`) plus the resource-barrel import, both guarded by a shipped wiring test. Approval-lifecycle events (request/decision as their own events), OneCLI credential holds, and channel/sender events are later increments on the same schema.
+- **The guard seam (guarded-actions phase 2).** Every privileged action crossing the container or channel boundary now passes ONE decision function — `guard()` in the new `src/guard/` leaf — before it executes: `allow`, `hold` (through the flow's existing approval mechanics), or `deny`. The four handler registries (ncl commands, delivery actions, response handlers, message interceptors) wrap their handlers at registration, so the guarded path is the only path by construction, and a registry-walking conformance test fails CI on any unmapped mutating entry. The structural baseline is today's checks verbatim (cli_scope enforcement, create_agent's scope branch, a2a destination ACL + `agent_message_policies` hold, self-mod unconditional hold, unknown_sender_policy, channel-registration click auth, host-as-trusted-caller in code). Policy-as-data (tighten-only rule sources composing with the baseline) is deferred to phase 3, where the generalized rules table arrives with its first operator-visible consumer. Approved replays now carry the verified approval row as a **grant** — the `approved: true` boolean is deleted — and the guard re-runs the structural baseline on every replay. Three outcome changes are inherent to that replay semantics and ship deliberately: **(a)** approving a held a2a message after its destination was revoked no longer delivers it (the live baseline re-check); **(b)** a forged, already-consumed, or wrong-action/wrong-payload grant refuses the replay instead of executing (previously any in-process caller passing `approved: true` executed); **(c)** the channel-registration free-text name reply re-checks approver eligibility at reply time — a privilege revoked between the click and the reply, or a registration that vanished meanwhile, now refuses instead of acting. Zero rules ship by default — the seeded posture is today's behavior. (The D1/D2/D4 click-auth fixes from the guarded-actions defect inventory are NOT in this change — they belong to the approval-contract PR.)
+- **Guard conformance is now a boot invariant, not just a CI test.** The host refuses to start (upgrade-tripwire posture: clear banner + exit 1) if any privileged command or delivery action is registered without a guard-catalog mapping. CI can't see skill-installed code — `/add-*` skills register handlers on machines where the test suite never runs — so the same registry walk the conformance test uses (`src/guard-conformance.ts`, one shared exemption list) now runs in `main()` before the host accepts a message, and a bad registration crashes at skill-install time instead of running unguarded. Companion hardening: re-registering a guard-wrapped delivery action *without* a guard spec now throws at registration — previously that silently replaced the guarded handler while the catalog still reported the action as guarded.
 - **Optional per-container resource caps.** `CONTAINER_CPU_LIMIT` and `CONTAINER_MEMORY_LIMIT` pass through to `docker run` as `--cpus` / `--memory` (`container-runner.ts`). Both empty by default — no flag added, spawn args byte-identical to today — so existing installs are unaffected. Set them to cap an agent container's CPU/memory so one agent can't monopolize the host (e.g. `CONTAINER_CPU_LIMIT=2`, `CONTAINER_MEMORY_LIMIT=8g`). Swap is intentionally not managed here: `--memory` is a hard cap on a swapless host.
 - [BREAKING] **Chat SDK pinned to `4.29.0` (was `4.26.0` via `^4.24.0`).** `chat` and the `@chat-adapter/*` channel adapters are version-locked — the adapter's `ChatInstance` must match the bridge's, so a mismatched pair fails to typecheck at `createChatSdkBridge(...)`. `chat` is therefore pinned exactly, and the channel-adapter install pins move with it — the `/add-<channel>` SKILL.md steps and `setup/*.sh` scripts on `main`, plus the adapter code on the `channels` branch. Core installs with no channel (only `cli`) are unaffected. **Migration:** if any channel is installed (Slack, Discord, Telegram, Teams, …), re-run its `/add-<channel>` skill to pull the matching `4.29.0` adapter.
 - **Budget/billing-exhausted LLM turns now reach the user instead of being silently dropped.** When a turn ends in a non-retryable provider error (e.g. an Anthropic `403 billing_error`) with no `<message>` wrapping, the agent-runner delivers the provider's notice to the originating channel and stops re-nudging the failing gateway. `providers/claude.ts` now surfaces the SDK's `is_error` flag (and the error subtype's `errors[]` text); `poll-loop.ts` delivers that text and skips the re-wrap retry. Fixes the case where a spend-limit notice produced silence plus a turn-after-turn retry loop.

CLAUDE.md

@@ -15,11 +15,11 @@ If you are a fresh install (you ran `git clone`, not `git pull`) and there are n
 
 # NanoClaw
 
-Personal Claude assistant. See [README.md](README.md) for philosophy and setup. Architecture lives in `docs/`.
+Personal AI assistant. See [README.md](README.md) for philosophy and setup. Architecture lives in `docs/`.
 
 ## Quick Context
 
-The host is a single Node process that orchestrates per-session agent containers. Platform messages land via channel adapters, route through an entity model (users → messaging groups → agent groups → sessions), get written into the session's inbound DB, and wake a container. The agent-runner inside the container polls the DB, calls Claude, and writes back to the outbound DB. The host polls the outbound DB and delivers through the same adapter.
+The host is a single Node process that orchestrates per-session agent containers. Platform messages land via channel adapters, route through an entity model (users → messaging groups → agent groups → sessions), get written into the session's inbound DB, and wake a container. The agent-runner inside the container polls the DB, calls the agent, and writes back to the outbound DB. The host polls the outbound DB and delivers through the same adapter.
 
 **Everything is a message.** There is no IPC, no file watcher, no stdin piping between host and container. The two session DBs are the sole IO surface.
 
@@ -32,7 +32,7 @@ agent_group_members (user_id, agent_group_id)    — unprivileged access gate
 user_dms (user_id, channel_type, messaging_group_id) — cold-DM cache
 
 agent_groups (workspace, memory, CLAUDE.md, personality, container config)
-    ↕ many-to-many via messaging_group_agents (session_mode, trigger_rules, priority)
+    ↕ many-to-many via messaging_group_agents (session_mode, engage_mode/engage_pattern, sender_scope, priority)
 messaging_groups (one chat/channel on one platform; instance = adapter-instance name, defaults to channel_type; unknown_sender_policy)
 
 sessions (agent_group_id + messaging_group_id + thread_id → per-session container)
@@ -44,8 +44,8 @@ Privilege is user-level (owner/admin), not agent-group-level. See [docs/isolatio
 
 Each session has **two** SQLite files under `data/v2-sessions/<session_id>/`:
 
-- `inbound.db` — host writes, container reads. `messages_in`, routing, destinations, pending_questions, processing_ack.
-- `outbound.db` — container writes, host reads. `messages_out`, session_state.
+- `inbound.db` — host writes, container reads. `messages_in`, delivered, destinations, session_routing.
+- `outbound.db` — container writes, host reads. `messages_out`, processing_ack, session_state, container_state.
 
 Exactly one writer per file — no cross-mount lock contention. Heartbeat is a file touch at `/workspace/.heartbeat`, not a DB update. Host uses even `seq` numbers, container uses odd.
 
@@ -65,13 +65,14 @@ For ad-hoc queries from skills or scripts, use the in-tree wrapper rather than t
 | `src/host-sweep.ts` | 60s sweep: `processing_ack` sync, stale detection, due-message wake, recurrence |
 | `src/session-manager.ts` | Resolves sessions; opens `inbound.db` / `outbound.db`; manages heartbeat path |
 | `src/container-runner.ts` | Spawns per-agent-group Docker containers with session DB + outbox mounts, OneCLI `ensureAgent` |
-| `src/container-runtime.ts` | Runtime selection (Docker vs Apple containers), orphan cleanup |
+| `src/container-runtime.ts` | Docker CLI wrapper (runtime binary, host-gateway args, mount args), orphan cleanup |
+| `src/guard/` | Privileged-action decision seam: `guard(action, input)` → allow \| hold \| deny. Domain-free leaf; domain baselines are `defineGuardedAction` values exported from module-edge `guard.ts` adapters (cli, agent-to-agent, self-mod, permissions) — consults hold the value, so a missing/typo'd wiring is a compile error, never a fail-open. All four handler registries require a guard spec or an explicit `unguarded(<reason>)` declaration at the registration site. Approved replays carry the approval row as a grant and re-check the structural baseline. Boot check = grant continuations only (`src/guard-conformance.ts`); Policy-as-data is deferred to guarded-actions phase 3. Conformance test: `src/guard/conformance.test.ts` |
 | `src/modules/permissions/access.ts` | `canAccessAgentGroup` — owner / global admin / scoped admin / member resolution against `user_roles` + `agent_group_members` |
 | `src/modules/approvals/primitive.ts` | `pickApprover`, `pickApprovalDelivery`, `requestApproval`, approval-handler registry |
 | `src/command-gate.ts` | Router-side admin command gate — queries `user_roles` directly (no env var, no container-side check) |
 | `src/modules/approvals/onecli-approvals.ts` | OneCLI credentialed-action approval bridge |
 | `src/modules/permissions/user-dm.ts` | Cold-DM resolution + `user_dms` cache |
-| `src/group-init.ts` | Per-agent-group filesystem scaffold (CLAUDE.md, skills, agent-runner-src overlay) |
+| `src/group-init.ts` | Per-agent-group filesystem scaffold (CLAUDE.md, skills) — agent-runner source is a shared read-only mount, not copied per group |
 | `src/db/container-configs.ts` | CRUD for `container_configs` table (per-group container runtime config) |
 | `src/backfill-container-configs.ts` | Migrates legacy `container.json` files into the DB on startup |
 | `src/container-restart.ts` | Kill + on-wake respawn for agent group containers |
@@ -79,8 +80,8 @@ For ad-hoc queries from skills or scripts, use the in-tree wrapper rather than t
 | `src/channels/` | Channel adapter infra (registry, Chat SDK bridge); specific channel adapters are skill-installed from the `channels` branch |
 | `src/providers/` | Host-side provider container-config (`claude` baked in; `opencode` etc. installed from the `providers` branch) |
 | `container/agent-runner/src/` | Agent-runner: poll loop, formatter, provider abstraction, MCP tools, destinations |
-| `container/skills/` | Container skills mounted into every agent session (`onecli-gateway`, `welcome`, `self-customize`, `agent-browser`, `slack-formatting`) |
-| `groups/<folder>/` | Per-agent-group filesystem (CLAUDE.md, skills, per-group `agent-runner-src/` overlay) |
+| `container/skills/` | Container skills mounted into every agent session (`agent-browser`, `frontend-engineer`, `onecli-gateway`, `self-customize`, `slack-formatting`, `vercel-cli`, `welcome`, `whatsapp-formatting`) |
+| `groups/<folder>/` | Per-agent-group filesystem (CLAUDE.md, skills) — agent-runner source is a shared read-only mount, not copied per group |
 | `scripts/init-first-agent.ts` | Bootstrap the first DM-wired agent (used by `/init-first-agent` skill) |
 | `migrate-v2.sh` + `setup/migrate-v2/` | v1→v2 migration. Standalone script: `bash migrate-v2.sh`. Seeds DB, copies groups/sessions, installs channels, builds container, offers service switchover, then hands off to `/migrate-from-v1` skill for owner setup and CLAUDE.md cleanup. See [docs/migration-dev.md](docs/migration-dev.md). |
 | `nanoclaw.sh --uninstall` + `setup/uninstall/` | Uninstall this copy only (slug-scoped): service, containers + image, `data/`, `logs/`, `groups/`, this copy's OneCLI agents. Confirms per group; `--dry-run` previews, `--yes` skips prompts. Other copies and the shared OneCLI app are untouched. Bypasses bootstrap entirely; `uninstall.sh` is a pointer that execs it. |
@@ -115,7 +116,7 @@ Key files: `src/cli/dispatch.ts` (dispatcher + approval handler), `src/cli/crud.
 
 Trunk does not ship any specific channel adapter or non-default agent provider. The codebase is the registry/infra; the actual adapters and providers live on long-lived sibling branches and get copied in by skills:
 
-- **`channels` branch** — Discord, Slack, Telegram, WhatsApp, Teams, Linear, GitHub, iMessage, Webex, Resend, Matrix, Google Chat, WhatsApp Cloud (+ helpers, tests, channel-specific setup steps). Installed via `/add-<channel>` skills.
+- **`channels` branch** — Discord, Slack, Telegram, WhatsApp, Teams, Linear, GitHub, iMessage, Webex, Resend, Matrix, Google Chat, WhatsApp Cloud, Signal, WeChat, DeltaChat, Emacs (+ helpers, tests, channel-specific setup steps). Installed via `/add-<channel>` skills.
 - **`providers` branch** — OpenCode (and any future non-default agent providers). Installed via `/add-opencode`.
 
 Each `/add-<name>` skill is idempotent: `git fetch origin <branch>` → copy module(s) into the standard paths → append a self-registration import to the relevant barrel → `pnpm install <pkg>@<pinned-version>` → build.
@@ -170,7 +171,7 @@ No container restart needed — the gateway looks up secrets per request.
 
 Approval-gating credentialed actions is a **two-sided** flow:
 
-- **Server-side** (OneCLI gateway): decides *when* to hold a request and emit a pending approval. As of `onecli@1.3.0`, the CLI does **not** expose this — `rules create --action` only accepts `block` or `rate_limit`, and `secrets create` has no approval flag. Approval policies must be configured via the OneCLI web UI at `http://127.0.0.1:10254`. If/when the CLI grows an `approve` action, this section needs updating.
+- **Server-side** (OneCLI gateway): decides *when* to hold a request and emit a pending approval. As of `onecli@2.2.5`, the CLI does **not** expose this — `rules create --action` only accepts `block` or `rate_limit`, and `secrets create` has no approval flag. Approval policies must be configured via the OneCLI web UI at `http://127.0.0.1:10254`. If/when the CLI grows an `approve` action, this section needs updating.
 - **Host-side** (nanoclaw): receives pending approvals and routes them to a human. `src/modules/approvals/onecli-approvals.ts` registers a callback via `onecli.configureManualApproval(cb)` (long-polls `GET /api/approvals/pending`). The callback uses `pickApprover` + `pickApprovalDelivery` from `src/modules/approvals/primitive.ts` to DM an approver. Approvers are resolved from the `user_roles` table — preference order: scoped admins for the agent group → global admins → owners. There is no env var like `NANOCLAW_ADMIN_USER_IDS`; roles are persisted in the central DB only.
 
 If approvals are configured server-side but the host callback isn't running (or throws), every credentialed call hangs until the gateway times out. Conversely, if the gateway has no rule asking for approval, the host callback never fires regardless of how it's wired.
@@ -182,7 +183,7 @@ Four types of skills. See [CONTRIBUTING.md](CONTRIBUTING.md) for the full taxono
 - **Channel/provider install skills** — copy the relevant module(s) in from the `channels` or `providers` branch, wire imports, install pinned deps (e.g. `/add-discord`, `/add-slack`, `/add-whatsapp`, `/add-opencode`).
 - **Utility skills** — ship code files alongside `SKILL.md` (e.g. a `scripts/` CLI or helper).
 - **Operational skills** — instruction-only workflows (`/setup`, `/debug`, `/customize`, `/init-first-agent`, `/manage-channels`, `/init-onecli`, `/update-nanoclaw`).
-- **Container skills** — loaded inside agent containers at runtime (`container/skills/`: `onecli-gateway`, `welcome`, `self-customize`, `agent-browser`, `slack-formatting`).
+- **Container skills** — loaded inside agent containers at runtime (`container/skills/`: `agent-browser`, `frontend-engineer`, `onecli-gateway`, `self-customize`, `slack-formatting`, `vercel-cli`, `welcome`, `whatsapp-formatting`).
 
 | Skill | When to Use |
 |-------|-------------|
@@ -216,7 +217,7 @@ Run commands directly — don't tell the user to run them.
 
 ```bash
 # Host (Node + pnpm)
-pnpm run dev          # Host with hot reload
+pnpm run dev          # Host via tsx (no watch)
 pnpm run build        # Compile host TypeScript (src/)
 ./container/build.sh  # Rebuild agent container image (nanoclaw-agent:latest)
 pnpm test             # Host tests (vitest)
@@ -297,7 +298,7 @@ The agent container runs on **Bun**; the host runs on **Node** (pnpm). They comm
 - **Writing a new named-param SQL insert/update in the container** → use `$name` in both SQL and JS keys: `.run({ $id: msg.id })`. `bun:sqlite` does not auto-strip the prefix the way `better-sqlite3` does on the host. Positional `?` params work normally.
 - **Adding a test in `container/agent-runner/src/`** → import from `bun:test`, not `vitest`. Vitest runs on Node and can't load `bun:sqlite`. `vitest.config.ts` excludes this tree.
 - **Adding a Node CLI the agent invokes at runtime** (like `agent-browser`, `claude-code`, `vercel`) → put it in the Dockerfile's pnpm global-install block, pinned to an exact version via a new `ARG`. Don't use `bun install -g` — that bypasses the pnpm supply-chain policy.
-- **Changing the Dockerfile entrypoint or the dynamic-spawn command** (`src/container-runner.ts` line ~301) → keep `exec bun ...` so signals forward cleanly. The image has no `/app/dist`; don't reintroduce a tsc build step.
+- **Changing the Dockerfile entrypoint or the dynamic-spawn command** (`src/container-runner.ts` line ~503) → keep `exec bun ...` so signals forward cleanly. The image has no `/app/dist`; don't reintroduce a tsc build step.
 - **Changing session-DB pragmas** (`container/agent-runner/src/db/connection.ts`) → `journal_mode=DELETE` is load-bearing for cross-mount visibility. Read the comment block at the top of the file first.
 
 ## CJK font support

CONTRIBUTING.md

@@ -75,7 +75,7 @@ Standalone tools that ship code files alongside the SKILL.md. The SKILL.md tells
 
 #### 3. Operational skills (instruction-only)
 
-Workflows and guides with no code changes. The SKILL.md is the entire skill — Claude follows the instructions to perform a task.
+Workflows and guides with no code changes. The SKILL.md is the entire skill — the coding agent follows the instructions to perform a task.
 
 **Location:** `.claude/skills/` on `main`
 
@@ -88,13 +88,13 @@ Workflows and guides with no code changes. The SKILL.md is the entire skill —
 
 #### 4. Container skills (agent runtime)
 
-Skills that run inside the agent container, not on the host. These teach the container agent how to use tools, format output, or perform tasks. They are synced into each group's `.claude/skills/` directory when a container starts.
+Skills that run inside the agent container, not on the host. These teach the NanoClaw agent how to use tools, format output, or perform tasks. They are synced into each group's `.claude/skills/` directory when a container starts.
 
 **Location:** `container/skills/<name>/`
 
-**Examples:** `agent-browser` (web browsing), `capabilities` (/capabilities command), `status` (/status command), `slack-formatting` (Slack mrkdwn syntax)
+**Examples:** `agent-browser` (web browsing), `frontend-engineer`, `onecli-gateway` (OneCLI proxy usage), `self-customize`, `slack-formatting` (Slack mrkdwn syntax), `vercel-cli`, `welcome`, `whatsapp-formatting`
 
-**Key difference:** These are NOT invoked by the user on the host. They're loaded by Claude Code inside the container and influence how the agent behaves.
+**Key difference:** You never invoke these from a coding-agent session on the host, the way you run `/setup` or `/update-nanoclaw` in Claude Code/Codex/OpenCode. They're mounted into the sandbox and loaded by the NanoClaw agent itself, shaping how it behaves when you chat with it.
 
 **Guidelines:**
 - Follow the same SKILL.md + frontmatter format

README.md

@@ -22,7 +22,7 @@
 
 [OpenClaw](https://github.com/openclaw/openclaw) is an impressive project, but I wouldn't have been able to sleep if I had given complex software I didn't understand full access to my life. OpenClaw has nearly half a million lines of code, 53 config files, and 70+ dependencies. Its security is at the application level (allowlists, pairing codes) rather than true OS-level isolation. Everything runs in one Node process with shared memory.
 
-NanoClaw provides that same core functionality, but in a codebase small enough to understand: one process and a handful of files. Claude agents run in their own Linux containers with filesystem isolation, not merely behind permission checks.
+NanoClaw provides that same core functionality, but in a codebase small enough to understand: one process and a handful of files. Agents run in their own Linux containers with filesystem isolation, not merely behind permission checks.
 
 ## Quick Start
 
@@ -49,7 +49,7 @@ bash migrate-v2.sh
 
 Run the script directly, not from inside a Claude session — the deterministic side needs interactive prompts and real shell I/O for Node/pnpm bootstrap, Docker, OneCLI, and the container build.
 
-**What it does:** merges `.env`, seeds the v2 DB from `registered_groups`, copies group folders + session data + scheduled tasks, installs the channel adapters you select, copies channel auth state (including Baileys keystore + LID mappings for WhatsApp), builds the agent container.
+**What it does:** merges `.env`, seeds the v2 DB from `registered_groups`, copies group folders + session data + scheduled tasks, installs the channel adapters you select, copies channel auth state (including the Baileys keystore for WhatsApp — LID mapping is now resolved per-message by the Baileys v7 adapter, not migrated), builds the agent container.
 
 **What it doesn't:** flip the system service. Pick *"switch to v2"* at the prompt, or do it manually after testing — your v1 install is left untouched.
 
@@ -78,11 +78,11 @@ See [docs/v1-to-v2-changes.md](docs/v1-to-v2-changes.md) for what's different an
 - **Multi-channel messaging** — WhatsApp, Telegram, Discord, Slack, Microsoft Teams, iMessage, Matrix, Google Chat, Webex, Linear, GitHub, WeChat, and email via Resend. Installed on demand with `/add-<channel>` skills. Run one or many at the same time.
 - **Flexible isolation** — connect each channel to its own agent for full privacy, share one agent across many channels for unified memory with separate conversations, or fold multiple channels into a single shared session so one conversation spans many surfaces. Pick per channel via `/manage-channels`. See [docs/isolation-model.md](docs/isolation-model.md).
 - **Per-agent workspace** — each agent group has its own `CLAUDE.md`, its own memory, its own container, and only the mounts you allow. Nothing crosses the boundary unless you wire it to.
-- **Scheduled tasks** — recurring jobs that run Claude and can message you back
+- **Scheduled tasks** — recurring jobs executed by the agent, which can message you the results
 - **Web access** — search and fetch content from the web
-- **Container isolation** — agents are sandboxed in Docker (macOS/Linux/WSL2), with optional [Docker Sandboxes](docs/docker-sandboxes.md) micro-VM isolation or Apple Container as a macOS-native opt-in
+- **Container isolation** — agents are sandboxed in Docker containers (macOS/Linux/WSL2)
 - **Credential security** — agents never hold raw API keys. Outbound requests route through [OneCLI's Agent Vault](https://github.com/onecli/onecli), which injects credentials at request time and enforces per-agent policies and rate limits.
-- **Agent templates**: stamp a ready-to-run agent (instructions + MCP tools + skills, no secrets) from a reusable bundle, via the setup wizard or `ncl groups create --template <ref>`. Load from the [public library](https://github.com/nanocoai/nanoclaw-templates), a local folder, or any git repo. See [docs/templates.md](docs/templates.md).
+- **Agent templates**: stamp a ready-to-run agent (instructions + MCP tools + skills, no secrets) from a reusable bundle via `ncl groups create --template <ref>`. Templates load from the local `templates/` folder; populate it by hand or by copying from the [public library](https://github.com/nanocoai/nanoclaw-templates). See [docs/templates.md](docs/templates.md).
 
 ## Usage
 
@@ -124,10 +124,7 @@ This keeps trunk as pure registry and infra, and every fork stays lean — users
 
 ### RFS (Request for Skills)
 
-Skills we'd like to see:
-
-**Communication Channels**
-- `/add-signal` — Add Signal as a channel
+No channel or provider skills are currently requested — propose one via an issue.
 
 ## Requirements
 
@@ -142,7 +139,7 @@ Skills we'd like to see:
 messaging apps → host process (router) → inbound.db → container (Bun, Claude Agent SDK) → outbound.db → host process (delivery) → messaging apps
 ```
 
-A single Node host orchestrates per-session agent containers. When a message arrives, the host routes it via the entity model (user → messaging group → agent group → session), writes it to the session's `inbound.db`, and wakes the container. The agent-runner inside the container polls `inbound.db`, runs Claude, and writes responses to `outbound.db`. The host polls `outbound.db` and delivers back through the channel adapter.
+A single Node host orchestrates per-session agent containers. When a message arrives, the host routes it via the entity model (user → messaging group → agent group → session), writes it to the session's `inbound.db`, and wakes the container. The agent-runner inside the container polls `inbound.db`, runs the agent, and writes responses to `outbound.db`. The host polls `outbound.db` and delivers back through the channel adapter.
 
 Two SQLite files per session, each with exactly one writer — no cross-mount contention, no IPC, no stdin piping. Channels and alternative providers self-register at startup; trunk ships the registry and the Chat SDK bridge, while the adapters themselves are skill-installed per fork.
 
@@ -165,7 +162,7 @@ Key files:
 
 **Why Docker?**
 
-Docker provides cross-platform support (macOS, Linux and Windows via WSL2) and a mature ecosystem. On macOS, Apple Container is also supported as a lighter-weight native runtime. For additional isolation, [Docker Sandboxes](docs/docker-sandboxes.md) run each container inside a micro VM.
+Docker provides cross-platform support (macOS, Linux and Windows via WSL2) and a mature ecosystem.
 
 **Can I run this on Linux or Windows?**
 

container/agent-runner/src/cli/ncl.ts

@@ -22,7 +22,9 @@ type RequestFrame = {
 };
 
 type ResponseFrame =
-  | { id: string; ok: true; data: unknown }
+  // `human` mirrors src/cli/frame.ts: an optional server-rendered string we
+  // print verbatim instead of running our own (drift-prone) formatter.
+  | { id: string; ok: true; data: unknown; human?: string }
   | { id: string; ok: false; error: { code: string; message: string } };
 
 // ---------------------------------------------------------------------------
@@ -244,6 +246,9 @@ if (!resp) {
 
 if (json) {
   process.stdout.write(JSON.stringify(resp, null, 2) + '\n');
+} else if (resp.ok && resp.human !== undefined) {
+  // Server-rendered view — print verbatim.
+  process.stdout.write(resp.human + '\n');
 } else {
   const output = formatHuman(resp);
   if (!resp.ok) {

container/agent-runner/src/current-batch.ts

@@ -1,29 +0,0 @@
-/**
- * Per-batch context the poll loop publishes for downstream consumers
- * (MCP tools, etc.) that don't sit on the poll-loop's call stack.
- *
- * Today the only field is `inReplyTo` — the id of the first inbound
- * message in the batch the agent is currently processing. MCP tools like
- * `send_message` and `send_file` read this and stamp it onto the outbound
- * row so the host's a2a return-path routing can correlate replies back to
- * the originating session.
- *
- * This is module-level state on purpose: the agent-runner is single-process
- * and processes one batch at a time. Poll-loop calls `setCurrentInReplyTo`
- * before invoking the provider and `clearCurrentInReplyTo` after the batch
- * completes (or errors out).
- */
-let currentInReplyTo: string | null = null;
-
-export function setCurrentInReplyTo(id: string | null): void {
-  currentInReplyTo = id;
-}
-
-export function clearCurrentInReplyTo(): void {
-  currentInReplyTo = null;
-}
-
-export function getCurrentInReplyTo(): string | null {
-  return currentInReplyTo;
-}
-

container/agent-runner/src/db/connection.ts

@@ -48,7 +48,11 @@ export function openInboundDb(): Database {
   // so the singleton survives for the rest of the test.
   if (_testMode && _inbound) {
     const db = _inbound;
-    return { prepare: (sql: string) => db.prepare(sql), exec: (sql: string) => db.exec(sql), close: () => {} } as unknown as Database;
+    return {
+      prepare: (sql: string) => db.prepare(sql),
+      exec: (sql: string) => db.exec(sql),
+      close: () => {},
+    } as unknown as Database;
   }
   const db = new Database(DEFAULT_INBOUND_PATH, { readonly: true });
   db.exec('PRAGMA busy_timeout = 5000');
@@ -260,11 +264,3 @@ export function closeSessionDb(): void {
   _outbound?.close();
   _outbound = null;
 }
-
-/**
- * @deprecated Use getInboundDb() / getOutboundDb() instead.
- * Kept for backward compatibility during migration.
- */
-export function getSessionDb(): Database {
-  return getInboundDb();
-}

container/agent-runner/src/db/index.ts

@@ -1,7 +1,6 @@
 export {
   getInboundDb,
   getOutboundDb,
-  getSessionDb,
   initTestSessionDb,
   closeSessionDb,
   touchHeartbeat,

container/agent-runner/src/db/session-state.ts

@@ -77,3 +77,47 @@ export function setContinuation(providerName: string, id: string): void {
 export function clearContinuation(providerName: string): void {
   deleteValue(continuationKey(providerName));
 }
+
+/**
+ * The a2a reply stamp: the id of the first inbound message in the batch the
+ * agent is currently processing. The poll loop publishes it at batch start;
+ * MCP tools (`send_message`, `send_file`) read it and stamp it onto outbound
+ * rows so the host's a2a return-path routing can correlate replies back to
+ * the originating session.
+ *
+ * This lives in outbound.db rather than module state because the MCP server
+ * runs as a separate stdio subprocess from the poll loop — module state set
+ * by the poll loop is invisible to it. Both processes open outbound.db
+ * (journal_mode=DELETE + busy_timeout make intra-container access safe).
+ */
+const IN_REPLY_TO_KEY = 'current_in_reply_to';
+
+/**
+ * Ignore a stamp older than this. The poll loop clears the stamp in a
+ * finally, but a container killed mid-batch (SIGKILL) can leave one behind;
+ * the guard stops a later out-of-batch read from picking up a dead stamp.
+ * Generous so a long-running batch's late sends still stamp correctly.
+ */
+const IN_REPLY_TO_MAX_AGE_MS = 30 * 60 * 1000;
+
+export function setCurrentInReplyTo(id: string | null): void {
+  if (id === null) {
+    clearCurrentInReplyTo();
+    return;
+  }
+  setValue(IN_REPLY_TO_KEY, id);
+}
+
+export function clearCurrentInReplyTo(): void {
+  deleteValue(IN_REPLY_TO_KEY);
+}
+
+export function getCurrentInReplyTo(): string | null {
+  const row = getOutboundDb()
+    .prepare('SELECT value, updated_at FROM session_state WHERE key = ?')
+    .get(IN_REPLY_TO_KEY) as { value: string; updated_at: string } | undefined;
+  if (!row) return null;
+  const age = Date.now() - new Date(row.updated_at).getTime();
+  if (!Number.isFinite(age) || age > IN_REPLY_TO_MAX_AGE_MS) return null;
+  return row.value;
+}

container/agent-runner/src/mcp-tools/core.test.ts

@@ -4,14 +4,31 @@
  * batch in poll-loop, and outbound writes from MCP tools (send_message,
  * send_file) must pick it up so a2a return-path routing on the host can
  * correlate replies back to the originating session.
+ *
+ * The stamp is published through session_state in outbound.db, not module
+ * state — the MCP server runs as a separate stdio subprocess from the poll
+ * loop, so it can only see the stamp through the shared DB. These tests seed
+ * it the same way the poll-loop process does (a direct DB write) rather than
+ * via any in-memory helper, so they exercise the real process boundary.
  */
 import { describe, it, expect, beforeEach, afterEach } from 'bun:test';
 
-import { initTestSessionDb, closeSessionDb, getInboundDb } from '../db/connection.js';
+import { initTestSessionDb, closeSessionDb, getInboundDb, getOutboundDb } from '../db/connection.js';
 import { getUndeliveredMessages } from '../db/messages-out.js';
-import { setCurrentInReplyTo, clearCurrentInReplyTo } from '../current-batch.js';
 import { sendMessage } from './core.js';
 
+/**
+ * Publish the a2a reply stamp the way the poll loop does: a direct write to
+ * session_state in outbound.db. `ageMs` back-dates updated_at to exercise the
+ * staleness guard MCP tools apply when reading it.
+ */
+function publishInReplyTo(id: string, ageMs = 0): void {
+  const updatedAt = new Date(Date.now() - ageMs).toISOString();
+  getOutboundDb()
+    .prepare('INSERT OR REPLACE INTO session_state (key, value, updated_at) VALUES (?, ?, ?)')
+    .run('current_in_reply_to', id, updatedAt);
+}
+
 beforeEach(() => {
   initTestSessionDb();
   // Seed a peer agent destination
@@ -24,13 +41,12 @@ beforeEach(() => {
 });
 
 afterEach(() => {
-  clearCurrentInReplyTo();
   closeSessionDb();
 });
 
 describe('send_message MCP tool — in_reply_to plumbing', () => {
-  it('stamps current batch in_reply_to on outbound rows', async () => {
-    setCurrentInReplyTo('inbound-msg-1');
+  it('stamps the batch in_reply_to (published via the DB) on outbound rows', async () => {
+    publishInReplyTo('inbound-msg-1');
 
     await sendMessage.handler({ to: 'peer', text: 'hello' });
 
@@ -40,7 +56,17 @@ describe('send_message MCP tool — in_reply_to plumbing', () => {
   });
 
   it('writes null when no batch is active', async () => {
-    // No setCurrentInReplyTo before this call — simulates ad-hoc / out-of-batch invocation.
+    // Nothing published to session_state — simulates ad-hoc / out-of-batch invocation.
+    await sendMessage.handler({ to: 'peer', text: 'hello' });
+
+    const out = getUndeliveredMessages();
+    expect(out).toHaveLength(1);
+    expect(out[0].in_reply_to).toBeNull();
+  });
+
+  it('ignores a stale stamp left behind by a killed container', async () => {
+    publishInReplyTo('inbound-msg-1', 60 * 60 * 1000); // an hour old
+
     await sendMessage.handler({ to: 'peer', text: 'hello' });
 
     const out = getUndeliveredMessages();

container/agent-runner/src/mcp-tools/core.ts

@@ -9,9 +9,9 @@
 import fs from 'fs';
 import path from 'path';
 
-import { getCurrentInReplyTo } from '../current-batch.js';
 import { findByName, getAllDestinations } from '../destinations.js';
 import { getMessageIdBySeq, getRoutingBySeq, writeMessageOut } from '../db/messages-out.js';
+import { getCurrentInReplyTo } from '../db/session-state.js';
 import { getSessionRouting } from '../db/session-routing.js';
 import { registerTools } from './server.js';
 import type { McpToolDefinition } from './types.js';

container/agent-runner/src/poll-loop.ts

@@ -2,8 +2,13 @@ import { findByName, getAllDestinations, type DestinationEntry } from './destina
 import { getPendingMessages, markProcessing, markCompleted, type MessageInRow } from './db/messages-in.js';
 import { writeMessageOut } from './db/messages-out.js';
 import { getInboundDb, touchHeartbeat, clearStaleProcessingAcks } from './db/connection.js';
-import { clearContinuation, migrateLegacyContinuation, setContinuation } from './db/session-state.js';
-import { clearCurrentInReplyTo, setCurrentInReplyTo } from './current-batch.js';
+import {
+  clearContinuation,
+  clearCurrentInReplyTo,
+  migrateLegacyContinuation,
+  setContinuation,
+  setCurrentInReplyTo,
+} from './db/session-state.js';
 import {
   formatMessages,
   extractRouting,

container/agent-runner/src/providers/claude.ts

@@ -449,7 +449,7 @@ export class ClaudeProvider implements AgentProvider {
           yield { type: 'result', text, isError: m.is_error === true };
         } else if (message.type === 'system' && (message as { subtype?: string }).subtype === 'api_retry') {
           yield { type: 'error', message: 'API retry', retryable: true };
-        } else if (message.type === 'system' && (message as { subtype?: string }).subtype === 'rate_limit_event') {
+        } else if (message.type === 'rate_limit_event') {
           yield { type: 'error', message: 'Rate limit', retryable: false, classification: 'quota' };
         } else if (message.type === 'system' && (message as { subtype?: string }).subtype === 'compact_boundary') {
           const meta = (message as { compact_metadata?: { pre_tokens?: number } }).compact_metadata;

docs/APPLE-CONTAINER-NETWORKING.md

@@ -1,90 +0,0 @@
-# Apple Container Networking Setup (macOS 26)
-
-Apple Container's vmnet networking requires manual configuration for containers to access the internet. Without this, containers can communicate with the host but cannot reach external services (DNS, HTTPS, APIs).
-
-## Quick Setup
-
-Run these two commands (requires `sudo`):
-
-```bash
-# 1. Enable IP forwarding so the host routes container traffic
-sudo sysctl -w net.inet.ip.forwarding=1
-
-# 2. Enable NAT so container traffic gets masqueraded through your internet interface
-echo "nat on en0 from 192.168.64.0/24 to any -> (en0)" | sudo pfctl -ef -
-```
-
-> **Note:** Replace `en0` with your active internet interface. Check with: `route get 8.8.8.8 | grep interface`
-
-## Making It Persistent
-
-These settings reset on reboot. To make them permanent:
-
-**IP Forwarding** — add to `/etc/sysctl.conf`:
-```
-net.inet.ip.forwarding=1
-```
-
-**NAT Rules** — add to `/etc/pf.conf` (before any existing rules):
-```
-nat on en0 from 192.168.64.0/24 to any -> (en0)
-```
-
-Then reload: `sudo pfctl -f /etc/pf.conf`
-
-## IPv6 DNS Issue
-
-By default, DNS resolvers return IPv6 (AAAA) records before IPv4 (A) records. Since our NAT only handles IPv4, Node.js applications inside containers will try IPv6 first and fail.
-
-The container image and runner are configured to prefer IPv4 via:
-```
-NODE_OPTIONS=--dns-result-order=ipv4first
-```
-
-This is set both in the `Dockerfile` and passed via `-e` flag in `container-runner.ts`.
-
-## Verification
-
-```bash
-# Check IP forwarding is enabled
-sysctl net.inet.ip.forwarding
-# Expected: net.inet.ip.forwarding: 1
-
-# Test container internet access
-container run --rm --entrypoint curl nanoclaw-agent:latest \
-  -s4 --connect-timeout 5 -o /dev/null -w "%{http_code}" https://api.anthropic.com
-# Expected: 404
-
-# Check bridge interface (only exists when a container is running)
-ifconfig bridge100
-```
-
-## Troubleshooting
-
-| Symptom | Cause | Fix |
-|---------|-------|-----|
-| `curl: (28) Connection timed out` | IP forwarding disabled | `sudo sysctl -w net.inet.ip.forwarding=1` |
-| HTTP works, HTTPS times out | IPv6 DNS resolution | Add `NODE_OPTIONS=--dns-result-order=ipv4first` |
-| `Could not resolve host` | DNS not forwarded | Check bridge100 exists, verify pfctl NAT rules |
-| Container hangs after output | Missing `process.exit(0)` in agent-runner | Rebuild container image |
-
-## How It Works
-
-```
-Container VM (192.168.64.x)
-    │
-    ├── eth0 → gateway 192.168.64.1
-    │
-bridge100 (192.168.64.1) ← host bridge, created by vmnet when container runs
-    │
-    ├── IP forwarding (sysctl) routes packets from bridge100 → en0
-    │
-    ├── NAT (pfctl) masquerades 192.168.64.0/24 → en0's IP
-    │
-en0 (your WiFi/Ethernet) → Internet
-```
-
-## References
-
-- [apple/container#469](https://github.com/apple/container/issues/469) — No network from container on macOS 26
-- [apple/container#656](https://github.com/apple/container/issues/656) — Cannot access internet URLs during building

docs/README.md

@@ -6,8 +6,5 @@ The files in this directory are original design documents and developer referenc
 
 | This directory | Documentation site |
 |---|---|
-| [SPEC.md](SPEC.md) | [Architecture](https://docs.nanoclaw.dev/concepts/architecture) |
 | [SECURITY.md](SECURITY.md) | [Security model](https://docs.nanoclaw.dev/concepts/security) |
 | [REQUIREMENTS.md](REQUIREMENTS.md) | [Introduction](https://docs.nanoclaw.dev/introduction) |
-| [docker-sandboxes.md](docker-sandboxes.md) | [Docker Sandboxes](https://docs.nanoclaw.dev/advanced/docker-sandboxes) |
-| [APPLE-CONTAINER-NETWORKING.md](APPLE-CONTAINER-NETWORKING.md) | [Container runtime](https://docs.nanoclaw.dev/advanced/container-runtime) |

docs/REQUIREMENTS.md

@@ -20,7 +20,7 @@ The entire codebase should be something you can read and understand. One Node.js
 
 ### Security Through True Isolation
 
-Instead of application-level permission systems trying to prevent agents from accessing things, agents run in actual Linux containers. The isolation is at the OS level. Agents can only see what's explicitly mounted. Bash access is safe because commands run inside the container, not on your Mac.
+Instead of application-level permission systems trying to prevent agents from accessing things, agents run in actual Linux containers. The isolation is at the OS level. Agents can only see what's explicitly mounted. Bash access is safe because commands run inside the container, not on your host.
 
 ### Built for the Individual User
 
@@ -47,23 +47,23 @@ When people contribute, they shouldn't add "Telegram support alongside WhatsApp.
 Skills we'd like to see contributed:
 
 ### Communication Channels
-- `/add-signal` - Add Signal as a channel
-- `/add-matrix` - Add Matrix integration
 
-> **Note:** Telegram, Slack, Discord, Gmail, and Apple Container skills already exist. See the [skills documentation](https://docs.nanoclaw.dev/integrations/skills-system) for the full list.
+None currently — Signal and Matrix have since shipped as skills.
+
+> **Note:** Telegram, Slack, Discord, Gmail, Signal, and Matrix skills already exist. See the [skills documentation](https://docs.nanoclaw.dev/integrations/skills-system) for the full list.
 
 ---
 
 ## Vision
 
-A personal Claude assistant accessible via messaging, with minimal custom code.
+A personal AI assistant accessible via messaging, with minimal custom code.
 
 **Core components:**
 - **Claude Agent SDK** as the core agent
-- **Containers** for isolated agent execution (Linux VMs)
+- **Containers** for isolated agent execution (Docker)
 - **Multi-channel messaging** (WhatsApp, Telegram, Discord, Slack, Gmail) — add exactly the channels you need
 - **Persistent memory** per conversation and globally
-- **Scheduled tasks** that run Claude and can message back
+- **Scheduled tasks** executed by the agent, which can message back
 - **Web access** for search and browsing
 - **Browser automation** via agent-browser
 
@@ -93,14 +93,14 @@ A personal Claude assistant accessible via messaging, with minimal custom code.
 - Sessions auto-compact when context gets too long, preserving critical information
 
 ### Container Isolation
-- All agents run inside containers (lightweight Linux VMs)
+- All agents run inside Docker containers
 - Each agent invocation spawns a container with mounted directories
 - Containers provide filesystem isolation - agents can only see mounted paths
 - Bash access is safe because commands run inside the container, not on the host
 - Browser automation via agent-browser with Chromium in the container
 
 ### Scheduled Tasks
-- Users can ask Claude to schedule recurring or one-time tasks from any group
+- Users can ask the agent to schedule recurring or one-time tasks from any group
 - Tasks run as full agents in the context of the group that created them
 - Tasks have access to all tools including Bash (safe in container)
 - Tasks can optionally send messages to their group via `send_message` tool, or complete silently

docs/SECURITY.md

@@ -1,70 +1,106 @@
 # NanoClaw Security Model
 
+> The canonical, continuously-verified version of this model lives at
+> [docs.nanoclaw.dev/concepts/security](https://docs.nanoclaw.dev/concepts/security).
+> This in-repo copy can drift; if the two disagree, verify against
+> `src/container-runner.ts` (`buildMounts`).
+
 ## Trust Model
 
+Privilege is **user-level**, persisted in the `user_roles` table (owner /
+admin, global or scoped to an agent group) plus `agent_group_members` (the
+unprivileged access gate).
+
 | Entity | Trust Level | Rationale |
 |--------|-------------|-----------|
-| Main group | Trusted | Private self-chat, admin control |
-| Non-main groups | Untrusted | Other users may be malicious |
-| Container agents | Sandboxed | Isolated execution environment |
-| Incoming messages | User input | Potential prompt injection |
+| Owners / admins (`user_roles`) | Trusted | Hold owner/admin roles; gate admin commands and approve credentialed actions |
+| Group members (`agent_group_members`) | Access-gated | Membership grants access to an agent group, but their messages are still untrusted input |
+| Unregistered senders | Untrusted | Subject to each messaging group's `unknown_sender_policy` |
+| Agent containers | Sandboxed | Long-lived per-session container; isolated by mounts, non-root, no host reach |
+| Incoming messages | User input | Potential prompt injection regardless of who sent them |
 
 ## Security Boundaries
 
 ### 1. Container Isolation (Primary Boundary)
 
-Agents execute in containers (lightweight Linux VMs), providing:
-- **Process isolation** - Container processes cannot affect the host
-- **Filesystem isolation** - Only explicitly mounted directories are visible
-- **Non-root execution** - Runs as unprivileged `node` user (uid 1000)
-- **Ephemeral containers** - Fresh environment per invocation (`--rm`)
+Agents execute in containers (Docker), providing:
+- **Process isolation** — container processes cannot affect the host
+- **Filesystem isolation** — only explicitly mounted directories are visible
+- **Non-root execution** — runs as an unprivileged user (`node`, uid 1000, or the host uid remapped in)
+- **Per-session containers** — one long-lived container per session polls that session's DBs and handles many messages, then is torn down (`--rm`) when the session goes idle.
 
-This is the primary security boundary. Rather than relying on application-level permission checks, the attack surface is limited by what's mounted.
+This is the primary security boundary. Rather than relying on application-level
+permission checks, the attack surface is limited by what's mounted.
 
 ### 2. Mount Security
 
-**External Allowlist** - Mount permissions stored at `~/.config/nanoclaw/mount-allowlist.json`, which is:
-- Outside project root
-- Never mounted into containers
-- Cannot be modified by agents
+`buildMounts` (`src/container-runner.ts`) composes a fixed set of mounts per
+spawn. For the default (Claude) provider these are:
 
-**Default Blocked Patterns:**
+| Container path | Host source | Mode | Purpose |
+|---|---|---|---|
+| `/workspace` | `data/v2-sessions/<group>/<session>/` | RW | Session folder — `inbound.db`, `outbound.db`, `outbox/`, `.claude/` |
+| `/workspace/agent` | `groups/<folder>/` | RW | Agent group working files + `CLAUDE.local.md` |
+| `/workspace/agent/container.json` | group `container.json` | RO | Container config — readable, not writable |
+| `/workspace/agent/CLAUDE.md` | composed `CLAUDE.md` | RO | Regenerated every spawn; agent edits would be clobbered |
+| `/workspace/agent/.claude-fragments` | group `.claude-fragments/` | RO | Composer skill/MCP fragments |
+| `/app/CLAUDE.md` | `container/CLAUDE.md` | RO | Shared base doc imported by the composed entry point |
+| `/home/node/.claude` | `data/v2-sessions/<group>/.claude-shared/` | RW | Claude state, settings, skill symlinks |
+| `/app/src` | `container/agent-runner/src/` | RO | Shared agent-runner source (same for all groups) |
+| `/app/skills` | `container/skills/` | RO | Shared container skills |
+| `/workspace/extra/<name>` | allowlisted host dir | RO (RW only if allowed) | Operator-configured additional mounts |
+
+The config mounts (`container.json`, `CLAUDE.md`, `.claude-fragments`) are
+**nested read-only mounts on top of the read-write group dir** — the agent can
+read its config but cannot modify it. The project root is **never mounted**: the
+container only ever sees the paths above plus any provider-contributed mounts
+(e.g. an OpenCode XDG dir). Host application source (`src/`, `dist/`,
+`package.json`) is not reachable.
+
+**Additional-mount allowlist** — extra mounts from a group's container config
+are validated against an allowlist at `~/.config/nanoclaw/mount-allowlist.json`,
+which is:
+- Outside the project root
+- Never mounted into containers
+- Not modifiable by agents
+
+Its schema:
+
+```json
+{
+  "allowedRoots": [
+    { "path": "~/projects", "allowReadWrite": true, "description": "Dev projects" },
+    { "path": "~/Documents/work", "allowReadWrite": false, "description": "Read-only" }
+  ],
+  "blockedPatterns": ["password", "secret", "token"]
+}
 ```
-.ssh, .gnupg, .aws, .azure, .gcloud, .kube, .docker,
-credentials, .env, .netrc, .npmrc, id_rsa, id_ed25519,
+
+**Default blocked patterns** (merged with any in the file):
+```
+.ssh, .gnupg, .gpg, .aws, .azure, .gcloud, .kube, .docker,
+credentials, .env, .netrc, .npmrc, .pypirc, id_rsa, id_ed25519,
 private_key, .secret
 ```
 
-**Protections:**
-- Symlink resolution before validation (prevents traversal attacks)
-- Container path validation (rejects `..` and absolute paths)
-- `nonMainReadOnly` option forces read-only for non-main groups
-
-**Read-Only Project Root:**
-
-The main group's project root is mounted read-only. Writable paths the agent needs (store, group folder, IPC, `.claude/`) are mounted separately. This prevents the agent from modifying host application code (`src/`, `dist/`, `package.json`, etc.) which would bypass the sandbox entirely on next restart. The `store/` directory is mounted read-write so the main agent can access the SQLite database directly.
+**Enforcement** (`src/modules/mount-security/index.ts`):
+- **No allowlist file ⇒ every additional mount is blocked** — the fixed mounts above are unaffected, but nothing extra is granted until the operator creates the file.
+- Symlinks are resolved to their real path (`realpathSync`) before any check, defeating traversal via symlink.
+- The real path is rejected if it matches a blocked pattern, and rejected unless it sits under one of `allowedRoots`.
+- The container path is validated: relative, non-empty, no `..`, no leading `/`, no `:` (blocks Docker `-v` option injection). It is mounted under `/workspace/extra/`.
+- **Read-write is granted only when the mount requests it (`readonly: false`) *and* the matched root has `allowReadWrite: true`.** Otherwise the mount is forced read-only.
 
 ### 3. Session Isolation
 
-Each group has isolated Claude sessions at `data/sessions/{group}/.claude/`:
-- Groups cannot see other groups' conversation history
-- Session data includes full message history and file contents read
-- Prevents cross-group information disclosure
+Per-session state lives under `data/v2-sessions/<agent-group>/<session>/`
+(`inbound.db`, `outbound.db`, `outbox/`, `.claude/`). Claude state
+(`.claude-shared`) and the working folder are scoped to the agent group, so:
+- Different agent groups cannot see each other's conversation history or files.
+- A group's sessions share that group's memory but keep separate message DBs.
 
-### 4. IPC Authorization
+This prevents cross-group information disclosure.
 
-Messages and task operations are verified against group identity:
-
-| Operation | Main Group | Non-Main Group |
-|-----------|------------|----------------|
-| Send message to own chat | ✓ | ✓ |
-| Send message to other chats | ✓ | ✗ |
-| Schedule task for self | ✓ | ✓ |
-| Schedule task for others | ✓ | ✗ |
-| View all tasks | ✓ | Own only |
-| Manage other groups | ✓ | ✗ |
-
-### 5. Credential Isolation (OneCLI Agent Vault)
+### 4. Credential Isolation (OneCLI Agent Vault)
 
 Real API credentials **never enter containers**. NanoClaw uses [OneCLI's Agent Vault](https://github.com/onecli/onecli) to proxy outbound requests and inject credentials at the gateway level.
 
@@ -77,13 +113,12 @@ Real API credentials **never enter containers**. NanoClaw uses [OneCLI's Agent V
 **Per-agent policies:**
 Each NanoClaw group gets its own OneCLI agent identity. This allows different credential policies per group (e.g. your sales agent vs. support agent). OneCLI supports rate limits, and time-bound access and approval flows are on the roadmap.
 
-**NOT Mounted:**
-- Channel auth sessions (`store/auth/`) — host only
-- Mount allowlist — external, never mounted
-- Any credentials matching blocked patterns
-- `.env` is shadowed with `/dev/null` in the project root mount
+**Never on the container filesystem:**
+- The project root and `.env` — never mounted; the container only receives the paths in the mount table above.
+- The mount allowlist — external (`~/.config/nanoclaw/…`), never mounted.
+- Real credentials — injected per request by the OneCLI gateway, never written into any mount.
 
-### 6. Egress Lockdown (Forced Proxy)
+### 5. Egress Lockdown (Forced Proxy)
 
 The `HTTPS_PROXY` env var only redirects *proxy-aware* clients — a tool that
 ignores it (or a raw socket) could reach the internet directly and bypass
@@ -111,31 +146,42 @@ no `host-gateway` route).
   exception: a heal failure there is logged but not fatal, since already-running
   agents stay on the internal net (no leak) until the gateway returns.
 
+**Default: egress is open.** Lockdown is **off** unless you opt in; by default
+the agent reaches the OneCLI gateway over the host-gateway path and outbound
+traffic is not confined to the internal network.
+
 **Configuration:**
 
 | Env | Default | Meaning |
 | --- | --- | --- |
-| `NANOCLAW_EGRESS_LOCKDOWN` | `false` | Set `true` to opt in (otherwise the host-gateway path is used). Enabled automatically by `/add-golden-registry`. |
+| `NANOCLAW_EGRESS_LOCKDOWN` | `false` | Set `true` to opt in (otherwise the host-gateway path is used). |
 | `NANOCLAW_EGRESS_NETWORK` | `nanoclaw-egress` | Network name. |
 | `ONECLI_GATEWAY_CONTAINER` | `onecli` | Gateway container to attach. |
 
+These variables are read from the **host process** environment (the service's
+environment / `.env`), not from inside the container. The agent container is
+started with only `TZ` and any provider-declared variables — host environment
+variables, including secrets, are never forwarded into the agent.
+
 **⚠ Behavior when enabled:** with lockdown on, agents have **no direct
 internet** — all traffic must go through OneCLI. Proxy-aware clients (npm, pnpm,
 pip, curl, node/bun with the proxy env) are unaffected. Any workflow that relies
 on a **non-proxy-aware** tool reaching the internet directly will fail by design.
 Lockdown is **off by default**; opt in with `NANOCLAW_EGRESS_LOCKDOWN=true`.
 
-## Privilege Comparison
+## Resource Limits
 
-| Capability | Main Group | Non-Main Group |
-|------------|------------|----------------|
-| Project root access | `/workspace/project` (ro) | None |
-| Store (SQLite DB) | `/workspace/project/store` (rw) | None |
-| Group folder | `/workspace/group` (rw) | `/workspace/group` (rw) |
-| Global memory | Implicit via project | `/workspace/global` (ro) |
-| Additional mounts | Configurable | Read-only unless allowed |
-| Network access | Unrestricted | Unrestricted |
-| MCP tools | All | All |
+Per-container CPU and memory caps are **opt-in and unset by default** — a runaway
+agent is not throttled unless the operator configures a limit:
+
+| Env | Default | Meaning |
+| --- | --- | --- |
+| `CONTAINER_CPU_LIMIT` | *(empty — unbounded)* | Passed to `--cpus` when set (e.g. `2`). |
+| `CONTAINER_MEMORY_LIMIT` | *(empty — unbounded)* | Passed to `--memory` when set (e.g. `8g`). |
+
+Only `--memory` is a container-level cap; whether it's a *hard* cap depends on
+the host having no swap (a deployment concern). On a swapless host a runaway is
+OOM-killed at the limit.
 
 ## Security Architecture Diagram
 
@@ -149,7 +195,7 @@ Lockdown is **off by default**; opt in with `NANOCLAW_EGRESS_LOCKDOWN=true`.
 ┌──────────────────────────────────────────────────────────────────┐
 │                     HOST PROCESS (TRUSTED)                        │
 │  • Message routing                                                │
-│  • IPC authorization                                              │
+│  • Role / access checks (user_roles, agent_group_members)        │
 │  • Mount validation (external allowlist)                          │
 │  • Container lifecycle                                            │
 │  • OneCLI Agent Vault (injects credentials, enforces policies)   │

docs/SPEC.md

@@ -1,5 +1,7 @@
 # NanoClaw Specification
 
+> **⚠️ Historical v1 spec.** This document describes the original NanoClaw v1 architecture — the single `store/messages.db`, the file-based IPC watcher, the `task-scheduler.ts` loop, the `MAX_CONCURRENT_CONTAINERS` cap, and the `groups/{channel}_{name}/` folder convention. **None of these exist in v2.** v2 replaced them with the two-DB session split (`inbound.db`/`outbound.db`), the entity model (users → messaging groups → agent groups → sessions), and the system-action delivery path. Kept for reference only. For the current architecture start at [architecture.md](architecture.md) and the root [CLAUDE.md](../CLAUDE.md); the v1→v2 diff is in [v1-to-v2-changes.md](v1-to-v2-changes.md).
+
 A personal Claude assistant with multi-channel support, persistent memory per conversation, scheduled tasks, and container-isolated agent execution.
 
 ---

docs/agent-runner-details.md

@@ -14,37 +14,62 @@ The boundary: the agent-runner decides **what** to send and **what to do** with
 
 ## AgentProvider Interface
 
+Provider-wide settings (MCP servers, env, additional directories, model, effort,
+assistant name) are passed to the provider **constructor** via `ProviderOptions`, not
+per query. `QueryInput` carries only what changes turn to turn: the prompt, the
+continuation token to resume, the working directory, and system context to inject.
+
 ```typescript
 interface AgentProvider {
+  /** True if the SDK handles slash commands natively and wants them passed
+   *  through raw. When false, the poll-loop formats them like any chat message. */
+  readonly supportsNativeSlashCommands: boolean;
+
+  /** Opt-in: scaffold a persistent memory/ tree at boot. Providers with native
+   *  memory (Claude's CLAUDE.local.md) omit it. Never gated on a provider name. */
+  readonly usesMemoryScaffold?: boolean;
+
+  /** Optional. Called after each completed exchange so providers whose harness
+   *  keeps no on-disk transcript can persist it themselves. Claude (the SDK
+   *  writes its own .jsonl) omits this. */
+  onExchangeComplete?(exchange: ProviderExchange): void;
+
   /** Start a new query. Returns a handle for streaming input and output. */
   query(input: QueryInput): AgentQuery;
+
+  /** True if the error means the stored continuation is invalid (missing
+   *  transcript, unknown session) and should be cleared. */
+  isSessionInvalid(err: unknown): boolean;
+
+  /** Optional pre-resume maintenance: given the stored continuation, return a
+   *  reason string to drop it and start fresh (e.g. transcript too large/old to
+   *  cold-resume before the host idle ceiling), or null to keep resuming. */
+  maybeRotateContinuation?(continuation: string, cwd: string): string | null;
+}
+
+interface ProviderOptions {
+  assistantName?: string;
+  mcpServers?: Record<string, McpServerConfig>;
+  env?: Record<string, string | undefined>;
+  additionalDirectories?: string[];
+  model?: string;   // alias (sonnet/opus/haiku) or full model ID
+  effort?: string;  // low | medium | high | xhigh | max
 }
 
 interface QueryInput {
-  /** Initial prompt (already formatted by agent-runner).
-   *  String for text-only. ContentBlock[] for multimodal (images, PDFs, audio). */
-  prompt: string | ContentBlock[];
+  /** Initial prompt, already formatted by the agent-runner into a string. */
+  prompt: string;
 
-  /** Session ID to resume, if any */
-  sessionId?: string;
+  /** Opaque continuation token from a previous query. The provider decides
+   *  what it means (session ID, thread ID, or nothing). */
+  continuation?: string;
 
-  /** Resume from a specific point in the session (provider-specific, may be ignored) */
-  resumeAt?: string;
-
-  /** Working directory inside the container */
+  /** Working directory inside the container. */
   cwd: string;
 
-  /** MCP server configurations (normalized format — provider translates) */
-  mcpServers: Record<string, McpServerConfig>;
-
-  /** System prompt / developer instructions */
-  systemPrompt?: string;
-
-  /** Environment variables for the SDK process */
-  env: Record<string, string | undefined>;
-
-  /** Additional directories the agent can access */
-  additionalDirectories?: string[];
+  /** System context to inject; the provider translates it into whatever its
+   *  SDK expects (preset append, full system prompt, per-turn injection). */
+  systemContext?: { instructions?: string };
 }
 
 interface McpServerConfig {
@@ -54,40 +79,42 @@ interface McpServerConfig {
 }
 
 interface AgentQuery {
-  /** Push a follow-up message into the active query */
+  /** Push a follow-up message into the active query. */
   push(message: string): void;
 
-  /** Signal that no more input will be sent */
+  /** Signal that no more input will be sent. */
   end(): void;
 
-  /** Output event stream */
+  /** Output event stream. */
   events: AsyncIterable<ProviderEvent>;
 
-  /** Force-stop the query (e.g., container shutting down) */
+  /** Force-stop the query (e.g., container shutting down). */
   abort(): void;
 }
 
 type ProviderEvent =
-  | { type: 'init'; sessionId: string }
-  | { type: 'result'; text: string | null }
+  | { type: 'init'; continuation: string }
+  | { type: 'result'; text: string | null; isError?: boolean }
   | { type: 'error'; message: string; retryable: boolean; classification?: string }
-  | { type: 'progress'; message: string };
+  | { type: 'progress'; message: string }
+  | { type: 'activity' };
 ```
 
 ### What the interface does NOT include
 
 - **Message formatting** — the agent-runner formats messages before passing to the provider. The provider receives a ready-to-send prompt string.
-- **Hooks** — Claude-specific. The Claude provider registers hooks internally (PreCompact, PreToolUse, etc.). Other providers don't need them.
-- **Tool allowlists** — Claude uses `allowedTools`. Codex uses `approvalPolicy`. OpenCode uses `permission`. Each provider configures this internally based on the same intent: "allow everything, no prompting."
-- **Session persistence** — Claude persists sessions to disk automatically. Codex and OpenCode manage their own session state. The agent-runner doesn't control this — it just passes `sessionId` and `resumeAt`.
+- **Hooks** — Claude-specific. The Claude provider registers hooks internally (PreToolUse, PostToolUse, PreCompact). Other providers don't need them.
+- **Tool allowlists** — Claude uses `allowedTools` + `disallowedTools`. Other SDKs use their own equivalents. Each provider configures this internally.
+- **Session persistence** — the agent-runner stores one opaque `continuation` token per provider (see [Session Resume](#session-resume)) and passes it back as `QueryInput.continuation`. What it means is provider-private; Claude persists its own `.jsonl` transcript on disk keyed by the continuation (session ID).
 - **Sandbox configuration** — provider-specific. Each provider configures its own sandbox internally.
 
 ### Provider event semantics
 
-- **`init`** — emitted once per query when the provider establishes or resumes a session. The agent-runner captures `sessionId` for future resume.
-- **`result`** — emitted when the agent produces a complete response. May be emitted multiple times per query (e.g., Claude's multi-turn with subagents). The agent-runner writes each result to messages_out.
-- **`error`** — emitted on failure. `retryable` indicates whether the agent-runner should retry. `classification` is optional detail (e.g., 'quota', 'auth', 'transport').
+- **`init`** — emitted once per query when the provider establishes or resumes a session. The agent-runner captures `continuation` and persists it for future resume.
+- **`result`** — emitted when the agent produces a complete response. May be emitted multiple times per query (e.g., Claude's multi-turn with subagents). `isError` is set when the SDK flagged the turn as an error (e.g. a non-retryable billing error) so the poll-loop still surfaces the text instead of dropping it. The agent-runner writes each result to messages_out.
+- **`error`** — emitted on failure. `retryable` indicates whether the agent-runner should retry. `classification` is optional detail (e.g., 'quota').
 - **`progress`** — optional, for logging. The agent-runner logs these but doesn't act on them.
+- **`activity`** — a liveness signal. Providers MUST yield it on every underlying SDK event (tool call, thinking, partial message) so the poll-loop's idle timer stays honest during long tool runs.
 
 ## Provider Implementations
 
@@ -97,58 +124,82 @@ Only the `claude` provider ships in trunk. The Codex and OpenCode sections below
 
 Wraps `@anthropic-ai/claude-agent-sdk`'s `query()`.
 
+The provider takes its settings (`mcpServers`, `env`, `additionalDirectories`,
+`model`, `effort`, `assistantName`) in its constructor via `ProviderOptions`; `query()`
+only reads the per-turn `QueryInput`.
+
 ```typescript
 class ClaudeProvider implements AgentProvider {
+  readonly supportsNativeSlashCommands = true;
+  // ...constructor stores options.mcpServers, .env, .additionalDirectories,
+  //    .model, .effort, .assistantName...
+
   query(input: QueryInput): AgentQuery {
     const stream = new MessageStream();  // AsyncIterable<SDKUserMessage>
     stream.push(input.prompt);
 
-    const sdkQuery = query({
+    const sdkResult = sdkQuery({
       prompt: stream,
       options: {
         cwd: input.cwd,
-        resume: input.sessionId,
-        resumeSessionAt: input.resumeAt,
-        systemPrompt: input.systemPrompt
-          ? { type: 'preset', preset: 'claude_code', append: input.systemPrompt }
+        additionalDirectories: this.additionalDirectories,
+        resume: input.continuation,
+        pathToClaudeCodeExecutable: '/pnpm/claude',
+        systemPrompt: input.systemContext?.instructions
+          ? { type: 'preset', preset: 'claude_code', append: input.systemContext.instructions }
           : undefined,
-        mcpServers: input.mcpServers,  // already the right shape
-        additionalDirectories: input.additionalDirectories,
-        env: input.env,
-        allowedTools: NANOCLAW_TOOL_ALLOWLIST,
+        // Base tools plus one `mcp__<server>__*` pattern per registered MCP
+        // server — without the explicit MCP patterns the SDK's allowedTools
+        // filter silently drops every MCP namespace.
+        allowedTools: [...TOOL_ALLOWLIST, ...Object.keys(this.mcpServers).map(mcpAllowPattern)],
+        disallowedTools: SDK_DISALLOWED_TOOLS,
+        env: this.env,
+        model: this.model,
+        effort: this.effort,
         permissionMode: 'bypassPermissions',
         allowDangerouslySkipPermissions: true,
+        settingSources: ['project', 'user', 'local'],
+        mcpServers: this.mcpServers,
         hooks: {
-          PreCompact: [{ hooks: [preCompactHook] }],
-          PreToolUse: [{ matcher: 'Bash', hooks: [sanitizeBashHook] }],
+          PreToolUse: [{ hooks: [preToolUseHook] }],
+          PostToolUse: [{ hooks: [postToolUseHook] }],
+          PostToolUseFailure: [{ hooks: [postToolUseHook] }],
+          PreCompact: [{ hooks: [createPreCompactHook(this.assistantName)] }],
         },
       },
     });
 
+    let aborted = false;
     return {
       push: (msg) => stream.push(msg),
       end: () => stream.end(),
-      abort: () => sdkQuery.close(),
-      events: translateClaudeEvents(sdkQuery),
+      // Abort doesn't call into the SDK — it flips a flag the event generator
+      // checks and ends the input stream so the query drains and stops.
+      abort: () => { aborted = true; stream.end(); },
+      events: translateEvents(sdkResult, () => aborted),
     };
   }
 }
 ```
 
-`translateClaudeEvents` is an async generator that maps SDK messages to `ProviderEvent`:
-- `message.type === 'system' && message.subtype === 'init'` → `{ type: 'init', sessionId }`
-- `message.type === 'result'` → `{ type: 'result', text }`
-- `message.type === 'system' && message.subtype === 'api_retry'` → `{ type: 'error', retryable: true }`
-- `message.type === 'system' && message.subtype === 'rate_limit_event'` → `{ type: 'error', retryable: false, classification: 'quota' }`
-- `message.type === 'system' && message.subtype === 'task_notification'` → `{ type: 'progress', message }`
-- Everything else → logged, not emitted
+`translateEvents` is an async generator that yields `{ type: 'activity' }` for **every**
+SDK message (so the idle timer stays honest) and maps recognized messages to `ProviderEvent`:
+- `system`/`init` → `{ type: 'init', continuation: session_id }`
+- `result` → `{ type: 'result', text, isError }` — `text` is `result.result`, or the joined `result.errors[]` on error subtypes (billing/quota), so the notice still reaches the user
+- `system`/`api_retry` → `{ type: 'error', retryable: true }`
+- `system`/`rate_limit_event` → `{ type: 'error', retryable: false, classification: 'quota' }`
+- `system`/`compact_boundary` → `{ type: 'result', text: 'Context compacted…' }`
+- `system`/`task_notification` → `{ type: 'progress', message }`
+- when the `aborted` flag is set → the generator returns immediately
 
-**Claude-specific features preserved inside the provider:**
-- `MessageStream` for async iterable input (push-based)
-- `resumeSessionAt` for resume at specific message UUID
-- PreCompact hook for transcript archiving
-- PreToolUse hook for sanitizing bash env vars
-- Full tool allowlist
+**Claude-specific behavior inside the provider:**
+- `MessageStream` for async iterable input (push-based follow-ups)
+- Resume via the SDK `resume` option keyed on the stored `continuation` (the SDK session ID) — no separate resume-at cursor
+- `TOOL_ALLOWLIST` (Bash, Read, Write, Edit, Glob, Grep, WebSearch, WebFetch, Task, Skill, …) extended at the call site with a `mcp__<server>__*` pattern per registered MCP server; `SDK_DISALLOWED_TOOLS` blocks SDK builtins that collide with NanoClaw's own scheduling/interaction model (CronCreate/Delete/List, ScheduleWakeup, AskUserQuestion, Enter/ExitPlanMode, Enter/ExitWorktree)
+- **PreToolUse hook** records the current tool + its declared timeout to `container_state` (so the host sweep widens its stuck tolerance while a long Bash runs) and, as defense-in-depth, blocks any `SDK_DISALLOWED_TOOLS` call that slips through. It does **not** sanitize bash env vars — there is no such hook.
+- **PostToolUse / PostToolUseFailure** hooks clear the in-flight tool
+- **PreCompact** hook archives the transcript to `conversations/` before compaction
+- `maybeRotateContinuation` drops an oversized/aged transcript (default caps 12 MB / 14 days, both operator-overridable) so a cold container isn't killed reloading days of `.jsonl` before the host idle ceiling; `isSessionInvalid` clears a continuation whose transcript is gone
 - `additionalDirectories` for multi-directory access
 
 ### Codex Provider
@@ -159,8 +210,8 @@ Wraps `@openai/codex-sdk`.
 class CodexProvider implements AgentProvider {
   query(input: QueryInput): AgentQuery {
     const codex = new Codex(this.buildOptions(input));
-    const thread = input.sessionId
-      ? codex.resumeThread(input.sessionId, this.threadOptions(input))
+    const thread = input.continuation
+      ? codex.resumeThread(input.continuation, this.threadOptions(input))
       : codex.startThread(this.threadOptions(input));
 
     const abortController = new AbortController();
@@ -188,13 +239,13 @@ class CodexProvider implements AgentProvider {
           signal: abortController.signal,
         });
 
-        let sessionId: string | undefined;
+        let continuation: string | undefined;
         let resultText = '';
 
         for await (const event of streamed.events) {
           if (event.type === 'thread.started') {
-            sessionId = event.thread_id;
-            yield { type: 'init', sessionId };
+            continuation = event.thread_id;
+            yield { type: 'init', continuation };
           }
           if (event.type === 'item.completed' && event.item.type === 'agent_message') {
             resultText = event.item.text || resultText;
@@ -264,7 +315,7 @@ class OpenCodeProvider implements AgentProvider {
 
   private async *run(client, server, stream, input, getPendingFollowUp): AsyncIterable<ProviderEvent> {
     const session = await client.session.create();
-    yield { type: 'init', sessionId: session.data.id };
+    yield { type: 'init', continuation: session.data.id };
 
     await client.session.promptAsync({
       path: { id: session.data.id },
@@ -356,62 +407,58 @@ The agent-runner transforms messages_in rows into a prompt string. The provider
 
 **Routing field stripping:** `platform_id`, `channel_type`, `thread_id` are never included in the prompt. They're stored as context for writing messages_out.
 
-**Single message formatting by kind:**
+Every kind renders to a single self-contained XML element. The `id` attribute is the
+message's `seq` (the agent-facing message ID it passes to `edit_message` / `add_reaction`).
+The `from` attribute is the origin destination name (resolved from the routing fields via
+the destination map), so the agent always knows where a message came from — routing fields
+themselves are never shown.
 
-- **`chat`** — format into message XML:
+- **`chat`** — one `<message>` per row:
   ```xml
-  <message sender="John" time="2024-01-01 10:00">
-    Check this PR
-  </message>
+  <message id="5" from="family" sender="John" time="Jan 1, 10:00 AM">Check this PR</message>
   ```
+  A reply carries a `reply_to` attribute and an inline `<quoted_message from="…">…</quoted_message>`.
 
-- **`chat-sdk`** — extract fields from serialized Chat SDK message:
+- **`chat-sdk`** — same `<message>` shape, fields extracted from the serialized Chat SDK
+  message. Attachments are appended inline: `[image: screenshot.png — saved to /workspace/…]`
+  or `[image: screenshot.png (https://signed-url…)]`. Images/PDFs that Claude handles
+  natively are also passed as content blocks (see Media Handling below).
+
+- **`task`** — a `<task>` element, script output first when present:
   ```xml
-  <message sender="John (john@slack)" time="2024-01-01 10:00">
-    Check this PR
-    [image: screenshot.png — https://signed-url...]
-  </message>
-  ```
-  Attachments are listed inline. Images/PDFs that Claude handles natively are passed as content blocks (see Media Handling below).
-
-- **`task`** — task prompt, optionally with script output:
-  ```
-  [SCHEDULED TASK]
-
-  Script output:
-  {"data": ...}
+  <task from="scheduler" time="Jan 1, 9:00 AM">Script output:
+  {"data": …}
 
   Instructions:
-  Review open PRs
+  Review open PRs</task>
   ```
 
-- **`webhook`** — webhook payload:
-  ```
-  [WEBHOOK: github/pull_request]
-
-  {"action": "opened", "pull_request": {...}}
+- **`webhook`** — a `<webhook>` element wrapping the JSON payload:
+  ```xml
+  <webhook from="github" source="github" event="pull_request">{"action": "opened", …}</webhook>
   ```
 
-- **`system`** — host action result (response to an earlier system request):
-  ```
-  [SYSTEM RESPONSE]
-
-  Action: register_agent_group
-  Status: success
-  Result: {"agent_group_id": "ag-456"}
+- **`system`** — host action result, rendered as `<system_response>`:
+  ```xml
+  <system_response from="host" action="create_agent" status="success">{"agent_group_id": "ag-456"}</system_response>
   ```
 
-**Batch formatting:** Multiple pending messages are combined into one prompt:
+**Batch formatting:** All pending messages are combined into one prompt. The prompt opens
+with a self-closing `<context timezone="<IANA>" />` header (so the agent interprets every
+timestamp — and every time it schedules — in the user's zone), then the chat messages
+concatenated as consecutive `<message>` blocks, then any task/webhook/system elements,
+joined by blank lines:
 
 ```xml
-<context timezone="America/Los_Angeles">
-<messages>
-<message sender="John" time="10:00">Check this PR</message>
-<message sender="Jane" time="10:01">Already on it</message>
-</messages>
+<context timezone="America/Los_Angeles" />
+<message id="2" from="family" sender="John" time="10:00">Check this PR</message>
+<message id="4" from="family" sender="Jane" time="10:01">Already on it</message>
 ```
 
-Mixed kinds (e.g., a chat message + a system response) are combined with clear delimiters. Each section is labeled by kind.
+There is **no** outer `<messages>` envelope — an earlier revision wrapped multi-message
+batches that way, but the Claude Agent SDK answered the wrapped shape with a synthetic
+"No response requested." stub instead of calling the API (#2555). Dropping the wrapper made
+the single-message path just the N=1 case of the same concatenation.
 
 **Command detection:** Messages starting with `/` are checked against a command list. Recognized commands bypass formatting and are passed raw to the provider (for Claude's slash command handling) or intercepted by the agent-runner (for NanoClaw-level commands like session reset).
 
@@ -430,54 +477,70 @@ interface RoutingContext {
 
 When writing messages_out (either from provider results or MCP tool calls), the agent-runner copies this routing context by default. The agent never sees routing fields — it just produces text. The routing is implicit: "respond to whoever sent the message."
 
-MCP tools that target a different destination (e.g., `send_to_agent`, `send_message` with explicit channel) override the routing context for that specific messages_out row.
+MCP tools that target a named destination (`send_message` / `send_file` with a `to`
+argument) resolve routing through the session's destination map instead of the default
+reply context — including agent-to-agent sends, which are just a `to` pointing at an
+`agent`-type destination.
 
 ### Status Management
 
-The agent-runner manages the `status` and `status_changed` fields on messages_in:
+`inbound.db` is a read-only mount inside the container, so the agent-runner never writes
+`messages_in`. It tracks processing status in the `processing_ack` table in the
+container-owned `outbound.db`; the host reads `processing_ack` and mirrors completion
+back onto `messages_in.status`.
 
 ```
-pending → processing → completed
-                    → failed (if provider returns error and max retries exhausted)
+processing_ack: (no row) → processing → completed
 ```
 
-- **Pick up:** `UPDATE messages_in SET status = 'processing', status_changed = now(), tries = tries + 1 WHERE id IN (...)`
-- **Complete:** `UPDATE messages_in SET status = 'completed', status_changed = now() WHERE id IN (...)`
-- **Error:** Agent-runner does NOT set `failed` — it leaves the message as `processing`. The host detects stale processing via `status_changed` and handles retry logic (reset to pending with backoff). This keeps retry policy on the host side.
+- **Pick up:** `INSERT OR REPLACE INTO processing_ack (message_id, status, status_changed) VALUES (?, 'processing', now())` for each claimed row (`markProcessing`). Pending queries skip any row already present in `processing_ack`.
+- **Complete:** same upsert with `status = 'completed'` (`markCompleted`). Every consumed batch ends here — error outcomes included. On a provider error the poll-loop writes an error **chat message** to `messages_out` (so the user sees it), then still acks the batch completed; errors surface as messages, not as an ack status. (A `markFailed` helper exists in `messages-in.ts` but currently has no callers.)
+- The host's `syncProcessingAcks` mirrors acked ids onto `messages_in.status = 'completed'`. Its stale/retry policy is driven off the `.heartbeat` file mtime and the `processing_ack` claim timestamps. On startup the agent-runner clears leftover `processing` acks (crash recovery) so orphaned claims re-process.
 
 ### MCP Tools
 
-The agent-runner runs an MCP server that exposes NanoClaw tools to the agent. All tools write to the session DB.
-
-**DB path:** The MCP server receives the session DB path via environment variable. It opens a second connection to the same SQLite file (WAL mode allows concurrent access).
+The agent-runner runs an MCP server (stdio) that exposes NanoClaw tools to the agent. The
+tool modules use the same two-DB connection layer as the rest of the runner
+(`container/agent-runner/src/db/connection.ts`): they read the host-written `inbound.db`
+at `/workspace/inbound.db` **read-only** (destinations, session routing, question
+responses, task lists) and write to the container-owned `outbound.db` at
+`/workspace/outbound.db`. There is no shared single-file connection and no WAL — both files
+are `journal_mode=DELETE` because WAL's memory-mapped `-shm` file does not stay coherent
+across the VirtioFS host↔container mount.
 
 #### send_message
 
-Send a chat message to the current conversation (or a specified destination).
+Send a chat message to a named destination. Agents address destinations by name, never by
+raw platform/channel/thread IDs — the destination map (`destinations` table in `inbound.db`,
+written by the host) resolves the name to routing fields.
 
 ```typescript
 {
   name: 'send_message',
   params: {
-    text: string,          // message content
-    channel?: string,      // optional: target channel type (default: reply to origin)
-    platformId?: string,   // optional: target platform ID
-    threadId?: string,     // optional: target thread ID
+    text: string,    // message content (required)
+    to?: string,     // destination name (e.g. "family", "worker-1").
+                     // Optional when the agent has exactly one destination.
   }
 }
 ```
 
-Implementation: write a `messages_out` row with `kind: 'chat'`. If channel/platformId/threadId are provided, use those as routing. Otherwise, copy from the current routing context.
+Implementation: `resolveRouting(to)` looks up the destination. With no `to`, it defaults to
+the session's own reply routing (`session_routing`); if the destination resolves to the same
+channel the session is bound to, the session's `thread_id` is preserved so the reply lands
+in-thread, otherwise `thread_id` is null. The tool then writes a `messages_out` row with
+`kind: 'chat'` and content `{ text }`, and returns the new `seq` as the message id.
 
 #### send_file
 
-Send a file to the current conversation.
+Send a file to a named destination (same destination model as `send_message`).
 
 ```typescript
 {
   name: 'send_file',
   params: {
-    path: string,          // file path (relative to /workspace/agent/ or absolute)
+    path: string,          // file path (relative to /workspace/agent/ or absolute) (required)
+    to?: string,           // destination name; optional if the agent has one destination
     text?: string,         // optional accompanying message
     filename?: string,     // display name (default: basename of path)
   }
@@ -485,10 +548,10 @@ Send a file to the current conversation.
 ```
 
 Implementation:
-1. Generate a message ID
-2. Create `outbox/{messageId}/` directory
-3. Copy the file into the outbox directory
-4. Write a `messages_out` row with `files: [filename]` in the content
+1. Resolve routing via `resolveRouting(to)` (as `send_message`)
+2. Generate a message ID and create `/workspace/outbox/{messageId}/`
+3. Copy the file into that outbox directory
+4. Write a `messages_out` row (`kind: 'chat'`) with content `{ text, files: [filename] }`
 
 #### send_card
 
@@ -523,11 +586,11 @@ Send an interactive question and wait for the user's response. This is a **block
 ```
 
 Implementation:
-1. Generate a `questionId`
-2. Write a `messages_out` row with `operation: 'ask_question'`, the question, options, and questionId
-3. Poll `messages_in` for a row with matching `questionId` in content
-4. When found, return the `selectedOption` as the tool result
-5. If timeout expires, return a timeout error as the tool result
+1. Generate a `questionId` and normalize each option to `{ label, selectedLabel, value }`
+2. Write a `messages_out` row with `kind: 'chat-sdk'` and content `{ type: 'ask_question', questionId, title, question, options }`
+3. Poll `inbound.db` (read-only) for a pending `messages_in` row whose content carries the matching `questionId` (`findQuestionResponse`), skipping any already in `processing_ack`
+4. When found, `markCompleted` the response row (a `processing_ack` write in `outbound.db`) and return its `selectedOption` as the tool result
+5. If the deadline passes, return a timeout error as the tool result
 
 The agent's execution is paused at this tool call. The provider's query keeps running (Claude holds the tool call open). The agent-runner polls for the response in a separate loop.
 
@@ -563,22 +626,13 @@ Add an emoji reaction to a message.
 
 Implementation: write a `messages_out` row with `operation: 'reaction'`.
 
-#### send_to_agent
+#### Agent-to-agent sends (no dedicated tool)
 
-Send a message to another agent group.
-
-```typescript
-{
-  name: 'send_to_agent',
-  params: {
-    agentGroupId: string,  // target agent group
-    text: string,          // message content
-    sessionId?: string,    // optional: target specific session
-  }
-}
-```
-
-Implementation: write a `messages_out` row with `channel_type: 'agent'`, `platform_id: agentGroupId`, `thread_id: sessionId`.
+There is no `send_to_agent` tool. Agents and channels share one destination namespace, so
+messaging another agent is just `send_message(to="<agent-name>")` where the named
+destination is of type `agent`. `resolveRouting` maps it to a `messages_out` row with
+`channel_type: 'agent'` and `platform_id` set to the target agent group id; the host
+validates the send and routes it into the target session's `inbound.db`.
 
 #### schedule_task
 
@@ -596,7 +650,7 @@ Schedule a one-shot or recurring task.
 }
 ```
 
-Implementation: write a `messages_in` row (to self) with `kind: 'task'`, `process_after`, and optionally `recurrence`. The host sweep picks it up when due.
+Implementation: the container can't write host-owned `inbound.db`, so this writes a `messages_out` row with `kind: 'system'` and `action: 'schedule_task'` (`container/agent-runner/src/mcp-tools/scheduling.ts`). During delivery the host's action handler (`src/modules/scheduling/actions.ts` → `insertTask()` in `src/modules/scheduling/db.ts`) inserts the `kind: 'task'` row into `inbound.db` with `process_after` and optionally `recurrence`. The host sweep picks it up when due.
 
 #### list_tasks
 
@@ -609,7 +663,7 @@ List active scheduled/recurring tasks.
 }
 ```
 
-Implementation: query `messages_in WHERE recurrence IS NOT NULL AND status != 'failed'`.
+Implementation: a read, not a write — the container may read the read-only `inbound.db` mount directly. Returns one row per series (the live pending/paused occurrence): `SELECT series_id AS id, ... FROM messages_in WHERE kind = 'task' AND status IN ('pending','paused') GROUP BY series_id`. See `container/agent-runner/src/mcp-tools/scheduling.ts`.
 
 #### cancel_task / pause_task / resume_task / update_task
 
@@ -625,27 +679,44 @@ Modify a scheduled task.
 // update_task: merge { prompt?, recurrence?, processAfter?, script? } into the live row
 ```
 
-Implementation: cancel/pause/resume update the live row(s) directly. update_task is sent as a system action — the host reads current content, merges supplied fields, and writes back. All four match by `(id = ? OR series_id = ?) AND kind='task' AND status IN ('pending','paused')`, so they reach the live next occurrence of a recurring task even when the agent passes the original (now-completed) id.
+Implementation: all four are sent as system actions (`messages_out`, `kind: 'system'`, `action: 'cancel_task' | 'pause_task' | 'resume_task' | 'update_task'`) — the container never writes `inbound.db`. The host's handlers in `src/modules/scheduling/actions.ts` apply the change against `inbound.db` via `src/modules/scheduling/db.ts`: cancel/pause/resume flip status on the live row(s); update_task reads current content, merges supplied fields, and writes back. All four match by `(id = ? OR series_id = ?) AND kind='task' AND status IN ('pending','paused')`, so they reach the live next occurrence of a recurring task even when the agent passes the original (now-completed) id.
 
-#### register_agent_group
+#### create_agent
 
-Register a new agent group (admin only).
+Create a long-lived companion sub-agent. The `name` becomes a destination the creating
+agent can address. (There is no `register_agent_group` tool — this replaced it.)
 
 ```typescript
 {
-  name: 'register_agent_group',
+  name: 'create_agent',
   params: {
-    name: string,
-    folder: string,
-    platformId: string,        // messaging group to wire to
-    channelType: string,
-    triggerRules?: object,
-    sessionMode?: 'shared' | 'per-thread',
+    name: string,           // human-readable name; also the destination name (required)
+    instructions?: string,  // CLAUDE.md content for the new agent (role, personality)
   }
 }
 ```
 
-Implementation: write a `messages_out` row with `kind: 'system'`, `action: 'register_agent_group'`. The host reads, validates admin permission, creates the entity rows in the central DB, and writes a `system` messages_in response.
+Implementation: fire-and-forget. Writes a `messages_out` row with `kind: 'system'`,
+`action: 'create_agent'`, `requestId`, `name`, and `instructions`. The container is
+untrusted and does not gate itself; the host authorizes by CLI scope — trusted owner groups
+(scope `global`) create directly, confined groups require admin approval
+(`src/modules/agent-to-agent/create-agent.ts`) — then creates the entity rows and notifies
+the agent via a chat message when the agent is ready.
+
+#### Self-modification: install_packages, add_mcp_server
+
+Two fire-and-forget system-action tools let an agent extend its own runtime (both require
+admin approval, applied host-side):
+
+- **`install_packages`** — `{ apt?: string[], npm?: string[], reason?: string }`. Package
+  names are validated at the tool boundary and re-validated on the host. On approval the
+  host rebuilds the per-agent image and restarts the container.
+- **`add_mcp_server`** — `{ name, command, args?, env? }`. Wires an existing third-party MCP
+  server into the agent's `container.json`; on approval the host updates the config and
+  restarts (no rebuild — Bun runs the TS directly).
+
+Both write a `messages_out` row with `kind: 'system'` and the matching `action`, then return
+immediately; the host notifies the agent when approval resolves.
 
 ### Media Handling
 
@@ -701,20 +772,28 @@ Archive location: `/workspace/agent/conversations/{date}-{summary}.md`
 
 ### Session Resume
 
-The agent-runner tracks `sessionId` and `resumeAt` across queries:
+The agent-runner tracks a single opaque `continuation` token per provider:
 
-- `sessionId` — captured from `ProviderEvent { type: 'init' }`. Passed back to `QueryInput.sessionId` on the next query.
-- `resumeAt` — Claude-specific (last assistant message UUID). Stored by the agent-runner, passed to `QueryInput.resumeAt`. Providers that don't support this ignore it.
+- Captured from `ProviderEvent { type: 'init', continuation }` and persisted to the
+  `session_state` table in `outbound.db` under the key `continuation:<provider>` (keyed per
+  provider because a continuation is provider-private — a Claude session id is meaningless to
+  another provider).
+- Passed back as `QueryInput.continuation` on the next query. For Claude that becomes the
+  SDK `resume` option; the SDK reloads its on-disk `.jsonl` transcript for that session id.
 
-These are ephemeral to the container's lifetime. When the container is killed and restarted, the host passes the stored `sessionId` from the central DB's sessions table. `resumeAt` is lost on container restart (the provider resumes from the end of the session).
+Because it lives in the session folder's `outbound.db`, the continuation survives container
+teardown and restart — a fresh container reads it back and resumes. `/clear` deletes the row
+to start a clean session. Before resuming, `maybeRotateContinuation` may archive and drop an
+oversized/aged transcript (so a cold container isn't killed reloading it), and
+`isSessionInvalid` clears a continuation whose backing transcript has gone missing.
 
 ### Container Startup
 
 The agent-runner receives configuration via:
 
-- **Environment variables:** `AGENT_PROVIDER` (claude/codex/opencode), `NANOCLAW_ADMIN_USER_ID`, provider-specific vars (API keys, model overrides), `TZ`
-- **Fixed mount paths:** Session DB at `/workspace/session.db`. Agent group folder at `/workspace/agent/`. System prompt from `/workspace/agent/CLAUDE.md` and `/workspace/global/CLAUDE.md`.
-- **Optional startup config:** Some config may be passed as a JSON file at a fixed path (e.g., `/workspace/config.json`) for things like the session ID to resume, assistant name, and admin user ID. This avoids overloading environment variables.
+- **`container.json`:** The provider name, model, assistant name, MCP servers, and other NanoClaw config are read from `/workspace/agent/container.json` (materialized by the host from the `container_configs` table), not from environment variables. See `container/agent-runner/src/config.ts`.
+- **Environment variables:** provider-specific vars only (API keys, model overrides), `TZ`.
+- **Fixed mount paths:** Host-written `inbound.db` (read-only) at `/workspace/inbound.db` and container-owned `outbound.db` at `/workspace/outbound.db`. Agent group folder at `/workspace/agent/`. System prompt from `/workspace/agent/CLAUDE.md` and `/workspace/global/CLAUDE.md`.
 
 The agent-runner reads config, creates the provider, and enters the poll loop. No stdin, no initial prompt — messages are already in the session DB.
 
@@ -731,7 +810,7 @@ function createProvider(name: ProviderName, config: ProviderConfig): AgentProvid
 }
 ```
 
-The provider name comes from the container's environment (`AGENT_PROVIDER` env var), set by the host based on `agent_groups.agent_provider` or `sessions.agent_provider`.
+The provider name comes from the `provider` key in `/workspace/agent/container.json` (defaulting to `'claude'`), which the host materializes from the `container_configs` table — set it with `ncl groups config update --provider`. It is not an environment variable.
 
 `ProviderConfig` contains provider-specific settings (API keys, model overrides, etc.) passed via environment variables — not via the interface. Each provider reads what it needs from `env`.
 

docs/api-details.md

@@ -13,6 +13,8 @@ interface ChannelSetup {
 
   // Host callbacks
   onInbound(platformId: string, threadId: string | null, message: InboundMessage): void;
+  // Admin-transport adapters (e.g. CLI) route to an arbitrary channel via this instead of onInbound
+  onInboundEvent(event: InboundEvent): void;
   onMetadata(platformId: string, name?: string, isGroup?: boolean): void;
 }
 
@@ -34,7 +36,7 @@ interface ChannelAdapter {
   isConnected(): boolean;
 
   // Outbound delivery
-  deliver(platformId: string, threadId: string | null, message: OutboundMessage): Promise<void>;
+  deliver(platformId: string, threadId: string | null, message: OutboundMessage): Promise<string | undefined>;
 
   // Optional
   setTyping?(platformId: string, threadId: string | null): Promise<void>;
@@ -307,7 +309,7 @@ function createWhatsAppChannel(): ChannelAdapter {
 **Ask user question:**
 ```json
 {
-  "operation": "ask_question",
+  "type": "ask_question",
   "questionId": "q-123",
   "title": "Failing Test",
   "question": "How should we handle the failing test?",

docs/architecture-diagram.html

@@ -311,7 +311,6 @@ erDiagram
     string name
     string folder
     string agent_provider
-    json container_config
   }
   messaging_groups {
     int id
@@ -344,14 +343,15 @@ erDiagram
     int messaging_group_id
     int agent_group_id
     string session_mode
-    json trigger_rules
+    string engage_mode "pattern | mention | mention-sticky"
+    string sender_scope "all | known"
     int priority
   }
   sessions {
     int id
     int agent_group_id
     int messaging_group_id
-    string sdk_session_id
+    string thread_id
     string status
   }
 </pre>
@@ -391,7 +391,7 @@ flowchart LR
   Container -->|writes · odd seq| Out
   Container -->|touch every poll| HB
   HostSweep[Host sweep] -->|stat mtime| HB
-  HostSweep -->|reads processing_ack| In
+  HostSweep -->|reads processing_ack| Out
 </pre>
         </div>
       </section>

docs/architecture-diagram.md

@@ -28,18 +28,18 @@ flowchart TB
     Approvals["configureManualApproval<br/>-> pending_approvals"]
   end
 
-  subgraph Session["Per-Session Container (Docker / Apple Container)"]
+  subgraph Session["Per-Session Container (Docker)"]
     direction TB
     PollLoop["Poll Loop<br/>(container/agent-runner)"]
-    Provider["Agent providers<br/>(claude, opencode; todo: codex)"]
+    Provider["Agent providers<br/>(claude — the only one registered in trunk;<br/>opencode ships via the /add-opencode skill)"]
     MCP["MCP Tools<br/>send_message, send_file, edit_message,<br/>add_reaction, send_card, ask_user_question,<br/>schedule_task, create_agent,<br/>install_packages, add_mcp_server"]
     Skills["Container Skills<br/>(container/skills/)"]
-    InDB[("inbound.db<br/>host writes<br/>even seq<br/>messages_in<br/>destinations<br/>processing_ack")]
-    OutDB[("outbound.db<br/>container writes<br/>odd seq<br/>messages_out<br/>heartbeat file")]
+    InDB[("inbound.db<br/>host writes<br/>even seq<br/>messages_in<br/>delivered<br/>destinations<br/>session_routing")]
+    OutDB[("outbound.db<br/>container writes<br/>odd seq<br/>messages_out<br/>processing_ack<br/>session_state<br/>container_state<br/>heartbeat file")]
   end
 
   subgraph Groups["Agent Group Filesystem (groups/*)"]
-    Folder["CLAUDE.md<br/>memory<br/>per-group skills<br/>container_config"]
+    Folder["CLAUDE.md<br/>memory<br/>per-group skills<br/>container.json (materialized from container_configs)"]
   end
 
   P1 & P2 & P3 & P4 & P5 --> Bridge
@@ -140,7 +140,6 @@ erDiagram
     string name
     string folder
     string agent_provider
-    json container_config
   }
   messaging_groups {
     int id
@@ -173,14 +172,15 @@ erDiagram
     int messaging_group_id
     int agent_group_id
     string session_mode "agent-shared | shared | per-thread"
-    json trigger_rules
+    string engage_mode "pattern | mention | mention-sticky"
+    string sender_scope "all | known"
     int priority
   }
   sessions {
     int id
     int agent_group_id
     int messaging_group_id
-    string sdk_session_id
+    string thread_id
     string status
   }
 ```
@@ -209,7 +209,7 @@ flowchart LR
   Container -->|"writes only<br/>(odd seq)"| Out
   Container -->|touch every poll| HB
   HostSweep[Host sweep] -->|stat mtime| HB
-  HostSweep -->|reads processing_ack| In
+  HostSweep -->|reads processing_ack| Out
 
   note1["Each file has exactly ONE writer.<br/>Eliminates SQLite cross-process write contention.<br/>Collision-free seq numbering."]
 ```

docs/architecture.md

@@ -1,8 +1,20 @@
 # NanoClaw Architecture (Draft)
 
+> **Draft — design intent, not a line-by-line spec.** Some passages predate the current implementation and can drift from it. The root [CLAUDE.md](../CLAUDE.md) and the cited source files (`src/`, `container/agent-runner/src/`) are the source of truth; when this doc and the code disagree, trust the code. Notably, scheduling MCP tools do **not** write `inbound.db` directly — they emit `messages_out` system actions that the host applies (see [agent-runner-details.md](agent-runner-details.md) and `src/modules/scheduling/`).
+
 ## Core Idea
 
-Each agent session has a mounted SQLite DB. The DB is the one and only IO mechanism between host and container. No IPC files, no stdin piping. Two tables: messages_in (host → agent-runner) and messages_out (agent-runner → host). Everything is a message.
+Each agent session has a **pair** of mounted SQLite DBs. They are the one and only IO
+mechanism between host and container. No IPC files, no stdin piping. `inbound.db` carries
+host → agent-runner messages (`messages_in`); `outbound.db` carries agent-runner → host
+messages (`messages_out`) plus the container's processing acks. Everything is a message.
+
+The split exists so each file has exactly one writer: the host writes `inbound.db` (the
+container opens it read-only) and the container writes `outbound.db` (the host opens it
+read-only). One writer per file means no cross-process lock contention over the
+host↔container mount. Both files run `journal_mode=DELETE`, **not** WAL: WAL's memory-mapped
+`-shm` coherency does not propagate across VirtioFS, so a WAL reader in the guest would
+freeze on an early snapshot and never see new host writes.
 
 ## Two-Level DB
 
@@ -11,15 +23,18 @@ Each agent session has a mounted SQLite DB. The DB is the one and only IO mechan
 - Maps platform IDs → agent groups → sessions
 - Channel adapters don't touch this directly — the host does the lookup
 
-**Per-session DB (mounted into container):**
-- messages_in (written by host, read by agent-runner)
-- messages_out (written by agent-runner, read by host)
-- Everything is a message: chat, tasks, webhooks, system actions, agent-to-agent — all use these two tables
-- One DB per session, not per agent group
+**Per-session DBs (mounted into container):**
+- `inbound.db` → `messages_in` (written by host, read-only in container) plus host-written
+  lookup tables the container reads live: `destinations`, `session_routing`, `delivered`
+- `outbound.db` → `messages_out` (written by agent-runner, read by host) plus
+  `processing_ack`, `session_state`, and `container_state` (all container-owned)
+- Everything is a message: chat, tasks, webhooks, system actions, agent-to-agent — all use
+  `messages_in` / `messages_out`
+- One pair per session, not per agent group
 
 ## Agent Groups vs Sessions
 
-An agent group has its own filesystem — folder, CLAUDE.md, skills, container config. Multiple sessions can share the same agent group (same filesystem, same skills) but each session gets its own DB mounted at a known path. Each session = a separate container with the same agent group's filesystem but a different session DB.
+An agent group has its own filesystem — folder, CLAUDE.md, skills, container config. Multiple sessions can share the same agent group (same filesystem, same skills) but each session gets its own `inbound.db`/`outbound.db` pair mounted at known paths. Each session = a separate container with the same agent group's filesystem but a different DB pair.
 
 ## Message Flow
 
@@ -28,13 +43,13 @@ Platform event
   → Channel adapter (trigger check, ID extraction)
   → Returns: { platformChannelId, platformThreadId, triggered }
   → Host maps platformChannelId + platformThreadId → agent group + session
-  → Host writes message to session's DB
+  → Host writes messages_in row to the session's inbound.db
   → Host calls wakeUpAgent(session)
   → Container spins up (or is already running)
-  → Agent-runner polls its session DB, finds new messages
-  → Agent-runner processes with Claude
-  → Agent-runner writes response to session DB
-  → Host polls active session DBs for responses
+  → Agent-runner polls inbound.db (read-only), finds new messages
+  → Agent-runner processes with the configured provider
+  → Agent-runner writes response to messages_out in outbound.db
+  → Host polls active sessions' outbound.db for responses
   → Host reads response, looks up conversation, delivers through channel adapter
 ```
 
@@ -128,9 +143,8 @@ Non-Chat-SDK channels (WhatsApp via Baileys, Gmail, custom integrations) impleme
 The host is an orchestrator:
 1. **Spawn** — when wakeUpAgent is called and no container exists for the session
 2. **Idle kill** — when a container has no unprocessed messages for some timeout period
-3. **Limits** — MAX_CONCURRENT_CONTAINERS caps active containers
 
-When a container spins up, the agent-runner immediately starts polling its session DB. Messages are already there waiting.
+When a container spins up, the agent-runner immediately starts polling `inbound.db`. Messages are already there waiting.
 
 ## Media Handling
 
@@ -184,50 +198,66 @@ Dedup is the channel adapter's responsibility. Chat SDK handles this internally.
 
 ## Session DB Schema
 
-Two tables. JSON blobs for content — schema-free, format varies by `kind`.
+Split across the two files. JSON blobs for content — schema-free, format varies by `kind`.
+`seq` is a global ordering counter with a **disjoint parity**: the host writes even seqs to
+`messages_in`, the container writes odd seqs to `messages_out`. Each side reads the other's
+MAX(seq) to pick its next value, so seq is a single monotonic message id across both tables —
+which is why the agent-facing message id it returns from `send_message` (and accepts in
+`edit_message` / `add_reaction`) is unambiguous.
 
 ```sql
--- Host writes, agent-runner reads
+-- inbound.db — host writes, container opens read-only
 CREATE TABLE messages_in (
   id             TEXT PRIMARY KEY,
+  seq            INTEGER UNIQUE,     -- even (host-assigned)
   kind           TEXT NOT NULL,      -- 'chat' | 'chat-sdk' | 'task' | 'webhook' | 'system'
   timestamp      TEXT NOT NULL,
-  status         TEXT DEFAULT 'pending',  -- 'pending' | 'processing' | 'completed' | 'failed'
-  status_changed TEXT,               -- ISO timestamp of last status change
+  status         TEXT DEFAULT 'pending',  -- host-owned; the host mirrors the container's
+                                          -- processing_ack terminal states onto this column
   process_after  TEXT,               -- ISO timestamp. NULL = process immediately.
   recurrence     TEXT,               -- cron expression. NULL = one-shot.
+  series_id      TEXT,               -- groups a recurring task's occurrences
   tries          INTEGER DEFAULT 0,  -- number of processing attempts
-
-  -- routing (agent-runner copies to messages_out; agent never sees these)
-  platform_id    TEXT,
+  trigger        INTEGER NOT NULL DEFAULT 1,  -- 0 = accumulated context (don't wake), 1 = wake
+  platform_id    TEXT,               -- routing (stripped before the agent sees content)
   channel_type   TEXT,
   thread_id      TEXT,
-
-  -- payload (structure depends on kind)
-  content        TEXT NOT NULL        -- JSON blob
+  content        TEXT NOT NULL,      -- JSON blob (structure depends on kind)
+  source_session_id TEXT,            -- a2a return path: source session that emitted the trigger
+  on_wake        INTEGER NOT NULL DEFAULT 0  -- 1 = only deliver on a container's first poll
 );
 
--- Agent-runner writes, host reads
+-- outbound.db — container writes, host opens read-only
 CREATE TABLE messages_out (
   id             TEXT PRIMARY KEY,
-  in_reply_to    TEXT,               -- references messages_in.id (optional)
+  seq            INTEGER UNIQUE,     -- odd (container-assigned)
+  in_reply_to    TEXT,              -- references messages_in.id (optional)
   timestamp      TEXT NOT NULL,
-  delivered      INTEGER DEFAULT 0,
-  deliver_after  TEXT,               -- ISO timestamp. NULL = deliver immediately.
-  recurrence     TEXT,               -- cron expression. NULL = one-shot.
-
-  -- routing (default: copied from messages_in by agent-runner)
-  kind           TEXT NOT NULL,      -- 'chat' | 'chat-sdk' | 'task' | 'webhook' | 'system'
-  platform_id    TEXT,
+  deliver_after  TEXT,              -- ISO timestamp. NULL = deliver immediately.
+  recurrence     TEXT,              -- cron expression. NULL = one-shot.
+  kind           TEXT NOT NULL,     -- copied from messages_in by default
+  platform_id    TEXT,              -- routing (default: copied from messages_in)
   channel_type   TEXT,
   thread_id      TEXT,
-
-  -- payload (format matches kind)
-  content        TEXT NOT NULL        -- JSON blob
+  content        TEXT NOT NULL      -- JSON blob (format matches kind)
 );
 
+-- outbound.db — container's processing status (it can't write inbound.db).
+-- Host reads this to drive the message lifecycle; stale 'processing' rows are
+-- cleared on container startup (crash recovery).
+CREATE TABLE processing_ack (
+  message_id     TEXT PRIMARY KEY,  -- references messages_in.id
+  status         TEXT NOT NULL,     -- 'processing' | 'completed' | 'failed'
+  status_changed TEXT NOT NULL
+);
 ```
 
+Delivery is tracked host-side in a `delivered` table in `inbound.db` (not a column on
+`messages_out`, which the host can't write). `inbound.db` also holds the host-written
+`destinations` and `session_routing` tables the container reads live; `outbound.db` also
+holds `session_state` (the resume continuation, per provider) and `container_state`
+(current tool in flight).
+
 ### Scheduling
 
 One-shot and recurring tasks use the same tables — no separate scheduler.
@@ -236,15 +266,15 @@ One-shot and recurring tasks use the same tables — no separate scheduler.
 
 **Recurring:** Same, plus a `recurrence` cron expression. After the host marks a row as handled/delivered, if `recurrence` is set, it inserts a new row with `process_after`/`deliver_after` advanced to the next cron occurrence. Next time is computed from the scheduled time (not wall clock) to prevent drift.
 
-**Host sweep** (every ~60s across all session DBs):
-- `messages_in WHERE status = 'pending' AND (process_after IS NULL OR process_after <= now())` → wake agent
-- `messages_in WHERE status = 'processing' AND status_changed < (now - stale_threshold)` → stale detection, increment tries, reset to pending with backoff
-- `messages_out WHERE delivered = 0 AND (deliver_after IS NULL OR deliver_after <= now())` → deliver
+**Host sweep** (every ~60s across all sessions):
+- `inbound.db` → `messages_in WHERE status = 'pending' AND (process_after IS NULL OR process_after <= now())` → wake agent
+- A `processing_ack` (in `outbound.db`) whose claim, or the `.heartbeat` mtime, is older than the stale threshold → stale detection, increment tries, reschedule `process_after` with backoff
+- `outbound.db` → due `messages_out` rows not yet in the host's `delivered` table (in `inbound.db`) → deliver
 - After completing/delivering a row with `recurrence`, insert next occurrence
 
 **Active container poll** (~1s) checks the same conditions but only for sessions with running containers.
 
-**Agent-runner creates schedules** by writing messages_in (to itself) or messages_out (reminders/notifications) with `process_after` and optionally `recurrence`.
+**Agent-runner creates schedules** by emitting a `messages_out` row with `kind: 'system'` and an `action` (`schedule_task`, `cancel_task`, …) — it cannot write host-owned `inbound.db` directly. The host applies the action during delivery (`src/modules/scheduling/actions.ts`), inserting/updating the `kind: 'task'` `messages_in` row with `process_after` and optionally `recurrence`.
 
 ### messages_in content by kind
 
@@ -331,7 +361,7 @@ Two patterns, both handled at the host level:
 
 In both cases, the approval and action execution happen on the host side, not the agent side.
 
-**Approval routing:** Privilege is a user-level concept. `user_roles` records `owner` (global only — first user to pair becomes owner) and `admin` (global or scoped to a specific `agent_group_id`). When an action requires approval, `pickApprover(agentGroupId)` returns candidates in order: scoped admins for that agent group → global admins → owners (deduplicated). `pickApprovalDelivery` then takes the first candidate reachable via `ensureUserDm` (with a same-channel-kind tie-break so a Discord approval request prefers a Discord-using approver). The approval card lands in the approver's DM messaging group, not the origin chat. Delivery is resolved through the Chat SDK's `openDM` for resolution-required channels (Discord/Slack/…) or the user's handle directly for direct-addressable channels (Telegram/WhatsApp/…), and the mapping is cached in `user_dms` for subsequent requests. See `src/access.ts`, `src/user-dm.ts`.
+**Approval routing:** Privilege is a user-level concept. `user_roles` records `owner` (global only — first user to pair becomes owner) and `admin` (global or scoped to a specific `agent_group_id`). When an action requires approval, `pickApprover(agentGroupId)` returns candidates in order: scoped admins for that agent group → global admins → owners (deduplicated). `pickApprovalDelivery` then takes the first candidate reachable via `ensureUserDm` (with a same-channel-kind tie-break so a Discord approval request prefers a Discord-using approver). The approval card lands in the approver's DM messaging group, not the origin chat. Delivery is resolved through the Chat SDK's `openDM` for resolution-required channels (Discord/Slack/…) or the user's handle directly for direct-addressable channels (Telegram/WhatsApp/…), and the mapping is cached in `user_dms` for subsequent requests. See `src/modules/permissions/access.ts` and `src/modules/permissions/user-dm.ts` (`ensureUserDm`); the approver-picking primitives live in `src/modules/approvals/primitive.ts`.
 
 **Editing a sent message:**
 
@@ -408,14 +438,16 @@ This is documented as a pattern, not a built-in feature.
 
 ## Design Decisions
 
-**Session DB location:** Not in the agent group folder. Separate directory (e.g., `sessions/{session_id}/`). Each session gets its own folder containing `session.db` and the Claude SDK's `.claude/` directory. The session identity IS the folder — no need to track Claude SDK session IDs.
+**Session DB location:** Not in the agent group folder. Separate directory (e.g., `sessions/{session_id}/`). Each session gets its own folder containing `inbound.db`, `outbound.db`, and — when using the claude provider — the SDK's `.claude/` directory. The session identity IS the folder. The SDK's resume token (session id) is persisted in `outbound.db`'s `session_state` table, so a fresh container picks the conversation back up without the host tracking it centrally.
 
 **Container mount structure:**
 
 ```
 /workspace/                 ← mount: session folder (read-write)
-  .claude/                  ← Claude SDK session data (auto-created)
-  session.db                ← session SQLite DB
+  .claude/                  ← Claude SDK session data / transcripts (auto-created)
+  inbound.db                ← host writes, container reads (opened read-only)
+  outbound.db               ← container writes, host reads (opened read-only)
+  .heartbeat                ← container touches this; host watches its mtime
   outbox/                   ← agent-runner writes outbound files here
   agent/                    ← mount: agent group folder (nested, read-write)
     CLAUDE.md               ← agent instructions
@@ -423,11 +455,17 @@ This is documented as a pattern, not a built-in feature.
     ... working files
 ```
 
-Two directory mounts: session folder at `/workspace`, agent group folder at `/workspace/agent/`. The agent-runner CDs into `/workspace/agent/` to run the agent. Claude SDK writes `.claude/` at `/workspace/.claude/` (root of the workspace). The session DB is at `/workspace/session.db`.
+Two directory mounts: session folder at `/workspace`, agent group folder at `/workspace/agent/`. The agent-runner CDs into `/workspace/agent/` to run the agent. Claude SDK writes `.claude/` at `/workspace/.claude/` (root of the workspace).
 
-This works on both Docker (nested bind mounts) and Apple Container (directory mounts only — no file-level mounts, but nested directory mounts are supported).
+The runtime is Docker (`src/container-runtime.ts` hardcodes the `docker` binary); nested bind mounts make this layout straightforward. The layout deliberately sticks to directory mounts (no file-level mounts) so it stays portable to runtimes that only support directory mounts.
 
-**Session DB concurrent access:** The host writes messages_in, the agent-runner writes messages_out. Both access the same SQLite file simultaneously. WAL mode handles this — SQLite allows concurrent readers, and the two sides write to different tables so writer contention is minimal. The host enables WAL mode when creating the session DB.
+**Cross-mount DB access:** The two files exist precisely so each has a single writer — the
+host writes `inbound.db`, the container writes `outbound.db` — which removes writer
+contention across the mount. Both files use `journal_mode=DELETE`, **not** WAL: WAL keeps its
+index in a memory-mapped `-shm` file, and VirtioFS does not propagate that mmap coherency
+from host to guest, so a WAL reader in the container would freeze on an early snapshot and
+silently never see new host writes. Readers that must see fresh host writes promptly (the
+`messages_in` poll) open `inbound.db` with `mmap_size = 0` to bypass SQLite's page cache.
 
 **Session management:** Host-managed. The host creates session folders and mounts them. The container only sees its own session folder.
 
@@ -438,7 +476,7 @@ This works on both Docker (nested bind mounts) and Apple Container (directory mo
 3. More messages arrive before container starts → host finds the existing session, writes to the same session DB
 4. Container starts, mounts the folder, agent-runner finds messages waiting
 
-The central DB session row creation is the serialization point. No Claude SDK session ID to coordinate — the SDK discovers its own session data in `.claude/` when the agent runs.
+The central DB session row creation is the serialization point. No provider session ID to coordinate — the SDK discovers its own session data in `.claude/` when the agent runs.
 
 **System actions:** The agent uses MCP tools (register group, reset session, schedule task, etc.). The agent-runner handles these tool calls and writes a structured, deterministic messages_out row with `kind: 'system'`. This is not natural language — it's a programmatic, structured payload that the host processes deterministically. Host validates permissions, executes, and writes the result back as a `system` messages_in row.
 
@@ -448,7 +486,9 @@ The central DB session row creation is the serialization point. No Claude SDK se
 
 ### Output Delivery
 
-NanoClaw does not stream tokens to users. The Claude Agent SDK's `query()` yields complete results. The agent-runner writes one complete message to messages_out per result. The host delivers complete messages to channels.
+NanoClaw does not stream tokens to users. The provider's query interface yields complete results, but a result's text is not delivered as-is: the agent-runner parses it for `<message to="name">...</message>` blocks (`dispatchResultText` in poll-loop.ts) and writes one messages_out row per block, addressed to that destination with its thread context resolved per destination. Everything outside a block — including `<internal>...</internal>` — is scratchpad: logged, never sent. A block naming an unknown destination is dropped into the scratchpad log.
+
+If a result produced text but no valid block, the agent-runner pushes a one-time `<system>` nudge into the live turn asking the agent to re-wrap its response. The exception is a non-retryable error result (e.g. a billing error) with no envelope, which is delivered as an error notice instead of being dropped as scratchpad. Mid-turn interim updates go out through the `send_message` MCP tool; the final-text envelope parsing is how a turn's reply reaches the user. The host delivers complete messages_out rows to channels.
 
 Message editing is supported as an explicit operation (agent calls an `edit_message` tool), not as a streaming mechanism.
 
@@ -456,21 +496,30 @@ Typing indicators: host sets typing when a container is active for a session, cl
 
 ### Message Batching
 
-When multiple messages arrive while the container is down, they accumulate as `handled = 0` rows in messages_in. When the container wakes up, the agent-runner queries all unhandled messages and processes them as a batch — multiple messages are formatted into a single `<messages>` XML block.
+When multiple messages arrive while the container is down, they accumulate as `status = 'pending'` rows in `messages_in`. When the container wakes up, the agent-runner reads all pending messages (those not yet in `processing_ack`) and processes them as a batch — formatted as a `<context timezone="…" />` header followed by the messages concatenated as consecutive `<message>` blocks. (There is no `<messages>` wrapper element; see [agent-runner-details.md](agent-runner-details.md#message-formatting).)
 
 ### Message Lifecycle
 
 ```
-pending → processing → completed
-                    → failed (after max retries)
+messages_in.status:   pending ──────────► completed (mirrored from ack)
+                              └─────────► failed (host-set, retries exhausted)
+processing_ack.status:         processing → completed
 ```
 
-- **pending**: Written by host. Ready to be picked up (if `process_after` is null or past).
-- **processing**: Agent-runner sets this when it picks up the message. `status_changed` is set to now. Prevents other polls from re-picking the same message.
-- **completed**: Agent-runner sets this after successful processing.
-- **failed**: Set after max retries exhausted.
+Because `inbound.db` is read-only in the container, the agent-runner never mutates
+`messages_in.status`. It records lifecycle in `processing_ack` (in `outbound.db`); the host
+reads that and mirrors completion back.
 
-**Stale detection**: If a message is `processing` but `status_changed` is too old (e.g., >10 minutes), the host assumes the container crashed. It resets the message to `pending`, increments `tries`, and sets `process_after` with exponential backoff.
+- **pending**: Host writes the `messages_in` row. Ready to be picked up (if `process_after` is null or past).
+- **processing**: Agent-runner upserts a `processing_ack` row (`status = 'processing'`) when it claims the message. Subsequent polls skip any id already in `processing_ack`, so it isn't re-picked.
+- **completed**: Agent-runner sets `processing_ack.status = 'completed'` for **every** consumed batch, error outcomes included — a provider error is surfaced to the user as an error chat message in `messages_out`, then the batch is still acked completed. The host's `syncProcessingAcks` copies it onto `messages_in.status`.
+- **failed**: Set by the **host** (sweep's `markMessageFailed`) when retries are exhausted — never by the container.
+
+**Liveness / stale detection**: The container touches a `/workspace/.heartbeat` file rather
+than writing the DB. The host sweep watches that mtime (widening its tolerance when
+`container_state` shows a long-declared Bash running) to decide a container has crashed, then
+increments `tries` and reschedules `process_after` with exponential backoff. On the next
+container startup, leftover `processing` acks are cleared so orphaned claims re-process.
 
 ### Error Handling and Retries
 
@@ -555,7 +604,7 @@ const DISCORD_TOKEN = process.env.DISCORD_BOT_TOKEN;
 const GMAIL_CREDS = process.env.GMAIL_CREDENTIALS_PATH;
 ```
 
-Shared config (DATA_DIR, TIMEZONE, MAX_CONCURRENT_CONTAINERS) stays in `config.ts`. Channel/skill-specific config stays in the module that uses it.
+Shared config (DATA_DIR, TIMEZONE) stays in `config.ts`. Channel/skill-specific config stays in the module that uses it.
 
 ### Code Style
 
@@ -594,13 +643,15 @@ src/db/
 - **No inline ALTER TABLE.** A migration runner with a `schema_version` table replaces `try { ALTER TABLE } catch { /* exists */ }` blocks. On startup, it checks the current version and applies pending migrations in order. Each migration is a function: `(db: Database) => void`.
 - **Skills add migrations.** A skill that needs a new column adds a new numbered migration file. No conflicts with other skills' migrations as long as numbers don't collide (use timestamps or high-enough numbers for skill branches).
 
-**Agent-runner session DB** uses the same pattern but lighter — no migrations needed since session DBs are created fresh by the host:
+**Agent-runner session DBs** use the same pattern but lighter — no migrations needed since the DB files are created fresh by the host:
 
 ```
 container/agent-runner/src/db/
-  connection.ts          ← open session.db at fixed path, WAL mode
-  messages-in.ts         ← read pending, update status
-  messages-out.ts        ← write results, outbox queries
+  connection.ts          ← open inbound.db (read-only) + outbound.db (DELETE mode) at fixed paths
+  messages-in.ts         ← read pending from inbound.db, ack via processing_ack in outbound.db
+  messages-out.ts        ← write results/outbox rows to outbound.db (odd seq)
+  session-state.ts       ← resume continuation, keyed per provider
+  session-routing.ts     ← read the host-written default reply routing
   index.ts               ← barrel
 ```
 
@@ -663,9 +714,13 @@ CREATE TABLE agent_groups (
   name             TEXT NOT NULL,
   folder           TEXT NOT NULL UNIQUE,
   agent_provider   TEXT,              -- default for sessions (null = system default)
-  container_config TEXT,              -- JSON: { additionalMounts, timeout }
   created_at       TEXT NOT NULL
 );
+-- Container config is NOT a column here — it lives in a separate container_configs
+-- table (migration 014), keyed by agent_group_id, with columns: provider, model,
+-- effort, image_tag, assistant_name, max_messages_per_prompt, cli_scope, and JSON
+-- columns skills / mcp_servers / packages_apt / packages_npm / additional_mounts.
+-- The host materializes it into /workspace/agent/container.json for the container.
 
 -- Platform groups/channels (WhatsApp group, Slack channel, Discord channel, email thread, etc.)
 -- One row per chat PER ADAPTER INSTANCE. instance defaults to channel_type
@@ -720,16 +775,21 @@ CREATE TABLE user_dms (
   PRIMARY KEY (user_id, channel_type)
 );
 
--- Which agent groups handle which messaging groups, with what rules
+-- Which agent groups handle which messaging groups, with what rules.
+-- The opaque trigger_rules JSON + response_scope enum were replaced (migration
+-- 010) by four orthogonal axes:
 CREATE TABLE messaging_group_agents (
-  id                 TEXT PRIMARY KEY,
-  messaging_group_id TEXT NOT NULL REFERENCES messaging_groups(id),
-  agent_group_id     TEXT NOT NULL REFERENCES agent_groups(id),
-  trigger_rules      TEXT,              -- JSON: { pattern, mentionOnly, excludeSenders, includeSenders }
-  response_scope     TEXT DEFAULT 'all',    -- 'all' | 'triggered' | 'allowlisted'
-  session_mode       TEXT DEFAULT 'shared', -- 'shared' | 'per-thread'
-  priority           INTEGER DEFAULT 0,     -- higher = checked first when multiple agents match
-  created_at         TEXT NOT NULL,
+  id                     TEXT PRIMARY KEY,
+  messaging_group_id     TEXT NOT NULL REFERENCES messaging_groups(id),
+  agent_group_id         TEXT NOT NULL REFERENCES agent_groups(id),
+  engage_mode            TEXT NOT NULL DEFAULT 'mention',  -- 'pattern' | 'mention' | 'mention-sticky'
+  engage_pattern         TEXT,                             -- regex; required for engage_mode='pattern'
+                                                           -- ('.' = match every message, the "always" flavor)
+  sender_scope           TEXT NOT NULL DEFAULT 'all',      -- 'all' | 'known'
+  ignored_message_policy TEXT NOT NULL DEFAULT 'drop',     -- 'drop' | 'accumulate'
+  session_mode           TEXT DEFAULT 'shared',            -- 'shared' | 'per-thread'
+  priority               INTEGER DEFAULT 0,                -- higher = checked first when multiple agents match
+  created_at             TEXT NOT NULL,
   UNIQUE(messaging_group_id, agent_group_id)
 );
 
@@ -794,7 +854,7 @@ stopped → running → idle → stopped
 
 ## Agent-Runner Architecture
 
-The agent-runner is the process inside the container. It mediates between the session DB and the Claude SDK — polling for work, formatting messages for the agent, translating tool calls into DB rows, and managing the agent lifecycle.
+The agent-runner is the process inside the container. It mediates between the session DB and the agent provider — polling for work, formatting messages for the agent, translating tool calls into DB rows, and managing the agent lifecycle.
 
 ### IO Model
 
@@ -807,50 +867,58 @@ All IO goes through the session DB. No stdin, no stdout markers, no IPC files.
 
 ### Poll Loop
 
-1. Query `messages_in WHERE status = 'pending' AND (process_after IS NULL OR process_after <= now())`
-2. If rows found: set `status = 'processing'`, `status_changed = now()` on each
+1. Query `inbound.db` (read-only) for `messages_in WHERE status = 'pending' AND (process_after IS NULL OR process_after <= now())`, skipping any id already in `processing_ack`
+2. If rows found: upsert `processing_ack` rows with `status = 'processing'` in `outbound.db` (the container can't write `messages_in`)
 3. Batch messages into a single prompt (strip routing fields, format by kind)
-4. Push into Claude SDK's MessageStream
+4. Push into the provider's input stream
 5. Process agent output → write `messages_out` rows
-6. Set processed messages to `status = 'completed'`
+6. Set the processed ids' `processing_ack.status = 'completed'` (the host mirrors that onto `messages_in.status`)
 7. Back to step 1. If no messages found, sleep briefly and re-poll (container stays warm for idle timeout)
 
 ### Message Formatting by Kind
 
 Agent-runner strips routing fields (`platform_id`, `channel_type`, `thread_id`) before formatting. The agent never sees routing info — it only sees content.
 
-- **`chat`** — format into `<messages>` XML block
-- **`chat-sdk`** — extract text, author, attachments from serialized message; format into `<messages>` XML
-- **`task`** — format as `[SCHEDULED TASK]` prefix + prompt. Run pre-script if present.
-- **`webhook`** — format as `[WEBHOOK: source/event]` + JSON payload
-- **`system`** — host action results (e.g., "register_group succeeded"). Format as system context, not chat.
+- **`chat`** — format into a `<message id="…" from="…" sender="…" time="…">` element
+- **`chat-sdk`** — extract text, author, attachments from serialized message; same `<message>` element
+- **`task`** — format as a `<task from="…" time="…">` element (script output first if present). Run pre-script if present.
+- **`webhook`** — format as a `<webhook source="…" event="…">` element wrapping the JSON payload
+- **`system`** — host action results, formatted as `<system_response action="…" status="…">`, not chat
 
 Mixed batches (e.g., a chat message + a system result both pending) are combined into one prompt with clear delimiters.
 
 ### MCP Tools
 
-MCP tools write directly to the session DB.
+MCP tools write to the container's own `outbound.db`. Anything that needs a change in host-owned `inbound.db` (schedule/cancel/pause/resume/update a task, create an agent, self-modify) is emitted as a `kind: 'system'` `messages_out` action that the host applies during delivery — the container never writes `inbound.db`.
 
-**Core tools:**
+**Messaging & interaction:**
 
 | Tool | What it does |
 |------|-------------|
-| `send_message` | Write `messages_out` row, `kind: 'chat'` |
-| `send_file` | Move file to `outbox/{msg_id}/`, write `messages_out` with filenames |
-| `schedule_task` | Write `messages_in` row (to self) with `process_after` + `recurrence`. Or `messages_out` with `deliver_after` for outbound reminders. |
-| `list_tasks` | Query `messages_in WHERE recurrence IS NOT NULL` |
-| `pause_task` / `resume_task` / `cancel_task` | Modify `messages_in` rows (update status, clear/set recurrence) |
-| `register_agent_group` | Write `messages_out`, `kind: 'system'`, `action: 'register_agent_group'` |
-
-**New tools:**
-
-| Tool | What it does |
-|------|-------------|
-| `ask_user_question` | Write `messages_out` with question card. Hold tool call open, poll `messages_in` for response matching `questionId`. Return selection as tool result. |
-| `edit_message` | Write `messages_out` with `operation: 'edit'` |
+| `send_message` | Resolve `to` (destination name) → routing, write `messages_out` row, `kind: 'chat'`. Omit `to` to reply in place. Also the agent-to-agent path: a `to` naming an `agent`-type destination. |
+| `send_file` | Copy file to `outbox/{msg_id}/`, write `messages_out` (`kind: 'chat'`) with filenames, same `to` resolution |
+| `send_card` | Write `messages_out`, `kind: 'chat-sdk'`, content `{ type: 'card', … }` |
+| `ask_user_question` | Write `messages_out` (`kind: 'chat-sdk'`, `type: 'ask_question'`). Hold tool call open, poll `inbound.db` for the response matching `questionId`. Return selection as tool result. |
+| `edit_message` | Write `messages_out` with `operation: 'edit'` (targets the original message's destination) |
 | `add_reaction` | Write `messages_out` with `operation: 'reaction'` |
-| `send_to_agent` | Write `messages_out` with `channel_type: 'agent'`, `platform_id: '{target}'` |
-| `send_card` | Write `messages_out` with card structure |
+
+(There is no `send_to_agent` tool — agent-to-agent is `send_message` to an `agent` destination.)
+
+**Scheduling** (all emit `kind: 'system'` actions except the read-only `list_tasks`):
+
+| Tool | What it does |
+|------|-------------|
+| `schedule_task` | `action: 'schedule_task'`; host inserts the `kind: 'task'` `messages_in` row with `process_after` + optional `recurrence` |
+| `list_tasks` | Read `inbound.db` (read-only) — one row per series: `kind = 'task' AND status IN ('pending','paused') GROUP BY series_id` |
+| `pause_task` / `resume_task` / `cancel_task` / `update_task` | matching `action`; host updates the live `messages_in` row(s) |
+
+**Central-DB / self-modification** (`kind: 'system'` actions; host authorizes, often via admin approval):
+
+| Tool | What it does |
+|------|-------------|
+| `create_agent` | `action: 'create_agent'` (name + instructions); host creates the agent group (replaces the old `register_agent_group`) |
+| `install_packages` | `action: 'install_packages'`; on approval host rebuilds the per-agent image and restarts |
+| `add_mcp_server` | `action: 'add_mcp_server'`; on approval host updates `container.json` and restarts |
 
 See [agent-runner-details.md](agent-runner-details.md) for full MCP tool parameter definitions.
 
@@ -886,11 +954,11 @@ The command lists are hardcoded in the agent-runner. Admin verification happens
 
 The agent-runner processes recurring task messages like any other messages_in row. After the agent-runner marks a recurring message as `completed`, the **host** handles inserting the next occurrence (new messages_in row with `process_after` advanced to next cron time). The agent-runner doesn't manage recurrence — it just processes what it finds.
 
-Pre-scripts: if a task message has a `script` field, run it first. If `wakeAgent = false`, mark completed without invoking Claude.
+Pre-scripts: if a task message has a `script` field, run it first. If `wakeAgent = false`, mark completed without invoking the provider.
 
 ### Agent-to-Agent Messaging
 
-**Outbound:** Agent calls `send_to_agent` tool → agent-runner writes messages_out with `channel_type: 'agent'`, `platform_id` = target agent group ID. Host validates permissions and writes to target session's messages_in.
+**Outbound:** Agent calls `send_message(to="<agent-name>")` where the named destination is of type `agent` → agent-runner writes messages_out with `channel_type: 'agent'`, `platform_id` = target agent group ID. Host validates permissions and writes to the target session's `inbound.db` (recording `source_session_id` so the reply routes back to this exact session).
 
 **Inbound:** Messages from other agents arrive as normal `chat` messages_in rows. The content includes `sender` and `senderId` (e.g., `"senderId": "agent:pr-admin"`). No special formatting — the agent sees it as a chat message.
 
@@ -906,7 +974,7 @@ Pre-scripts: if a task message has a `script` field, run it first. If `wakeAgent
 
 - **Approval routing** — how does the host find the admin's DM conversation? What if no DM channel exists? Is the approval list configurable per agent group or global?
 - **MCP server lifecycle** — does the MCP server process persist across multiple queries in the same container, or restart each time?
-- **Container startup config** — what config (if any) is passed to the container at launch beyond env vars? The session DB is at a fixed mount path. System prompt comes from CLAUDE.md. Provider name comes from env. What else?
+- **Container startup config** — what config (if any) is passed to the container at launch beyond env vars? The DB files are at fixed mount paths. System prompt comes from CLAUDE.md. Provider name comes from `container.json` (materialized from the `container_configs` table), not env. What else?
 - **Idle detection with pending questions** — when `ask_user_question` is waiting for a response, the container should not be considered idle. Also need to detect when the agent is still working (active tool calls, subagents) and avoid killing the container even if no messages_out have been written recently.
 
 ## Related Documents

docs/build-and-runtime.md

@@ -33,13 +33,13 @@ Both are committed. CI and the Dockerfile run `--frozen-lockfile` variants — a
 ## Supply chain
 
 - **Host + global CLIs** (pnpm): `minimumReleaseAge: 4320` (3-day hold on new versions), `onlyBuiltDependencies` allowlist for postinstall scripts. See `pnpm-workspace.yaml` and `docs/SECURITY.md`.
-- **Agent-runner** (Bun): no release-age policy — Bun doesn't have an equivalent today. The defenses are `bun.lock` pinning plus version-pinned CLIs/Bun itself via Dockerfile ARGs. When bumping `@anthropic-ai/claude-agent-sdk` or any runtime dep, review the release date on npm and bump deliberately, not via `bun update`.
+- **Agent-runner** (Bun): no release-age policy — Bun doesn't have an equivalent today. The defenses are `bun.lock` pinning plus a version-pinned Bun itself via a Dockerfile ARG (global CLIs are pinned separately in `container/cli-tools.json`). When bumping `@anthropic-ai/claude-agent-sdk` or any runtime dep, review the release date on npm and bump deliberately, not via `bun update`.
 
 ## Image build surface
 
 `container/Dockerfile` is a single-stage build on `node:22-slim`:
 
-- **Pinned ARGs** — `BUN_VERSION`, `CLAUDE_CODE_VERSION`, `AGENT_BROWSER_VERSION`, `VERCEL_VERSION`. Bump deliberately in PRs.
+- **Pinned ARGs** — `BUN_VERSION`, `PNPM_VERSION`, `INSTALL_CJK_FONTS`. Bump deliberately in PRs. Global CLI versions (`@anthropic-ai/claude-code`, `agent-browser`, `vercel`) are pinned separately in `container/cli-tools.json`, not as ARGs.
 - **CJK fonts** — `ARG INSTALL_CJK_FONTS=false`. `container/build.sh` reads `INSTALL_CJK_FONTS` from `.env` and passes it through. Default build saves ~200MB; opt in when the user works with Chinese/Japanese/Korean content.
 - **BuildKit cache mounts** — `/var/cache/apt`, `/var/lib/apt`, `/root/.bun/install/cache`, `/root/.cache/pnpm`. Rebuilds where `package.json`/`bun.lock` haven't changed are fast. Requires BuildKit (default on Docker 23+, Apple Container-compat).
 - **`tini` as init** — reaps Chromium zombies, forwards signals so in-flight `outbound.db` writes finalize on SIGTERM.
@@ -49,7 +49,7 @@ Both are committed. CI and the Dockerfile run `--frozen-lockfile` variants — a
 ## Session wake (two paths)
 
 1. **Base image ENTRYPOINT** — used for stdin-piped test invocations like the sample in `container/build.sh`: `tini --> entrypoint.sh` captures stdin to `/tmp/input.json`, then `exec bun run src/index.ts`.
-2. **Host-spawned session** — `src/container-runner.ts` at line ~301 uses `--entrypoint bash` with `-c 'exec bun run /app/src/index.ts'`. Bypasses tini (Docker's default PID 1 handling applies). Stdin is unused; all IO flows through the mounted session DBs.
+2. **Host-spawned session** — `src/container-runner.ts` at line ~503 uses `--entrypoint bash` with `-c 'exec bun run /app/src/index.ts'`. Bypasses tini (Docker's default PID 1 handling applies). Stdin is unused; all IO flows through the mounted session DBs.
 
 Both paths end with Bun running the same source file from `/app/src/index.ts`.
 

docs/db-central.md

@@ -2,7 +2,7 @@
 
 Complete reference for `data/v2.db`, the host-owned admin-plane database. Start with [db.md](db.md) for the three-DB overview, the map, and the cross-mount rules.
 
-Access layer: `src/db/`. Authoritative schema reference: `src/db/schema.ts` (comments only — actual creation runs via migrations in `src/db/migrations/`).
+Access layer: `src/db/`. `src/db/schema.ts`'s `SCHEMA` constant is a *reference copy* of the core tables for orientation — it is not exhaustive: several tables (`agent_destinations`, `pending_approvals`, `container_configs`, `agent_message_policies`, `pending_channel_approvals`, and others) exist only in their migration files under `src/db/migrations/`, which remain the actual source of truth for what's created at runtime.
 
 ---
 
@@ -55,20 +55,24 @@ Wiring: which agent group handles which messaging group. Many-to-many — the sa
 
 ```sql
 CREATE TABLE messaging_group_agents (
-  id                 TEXT PRIMARY KEY,
-  messaging_group_id TEXT NOT NULL REFERENCES messaging_groups(id),
-  agent_group_id     TEXT NOT NULL REFERENCES agent_groups(id),
-  trigger_rules      TEXT,
-  response_scope     TEXT DEFAULT 'all',
-  session_mode       TEXT DEFAULT 'shared',
-  priority           INTEGER DEFAULT 0,
-  created_at         TEXT NOT NULL,
+  id                     TEXT PRIMARY KEY,
+  messaging_group_id     TEXT NOT NULL REFERENCES messaging_groups(id),
+  agent_group_id         TEXT NOT NULL REFERENCES agent_groups(id),
+  engage_mode            TEXT NOT NULL DEFAULT 'mention',
+                         -- 'pattern' | 'mention' | 'mention-sticky'
+  engage_pattern         TEXT,   -- regex; required when engage_mode='pattern';
+                                 -- '.' means "match every message"
+  sender_scope           TEXT NOT NULL DEFAULT 'all',    -- 'all' | 'known'
+  ignored_message_policy TEXT NOT NULL DEFAULT 'drop',   -- 'drop' | 'accumulate'
+  session_mode           TEXT DEFAULT 'shared',
+  priority               INTEGER DEFAULT 0,
+  created_at             TEXT NOT NULL,
   UNIQUE(messaging_group_id, agent_group_id)
 );
 ```
 
 - `session_mode`: `shared` (one session per channel), `per-thread` (one per thread), `agent-shared` (one per agent group across all channels).
-- `trigger_rules`: JSON; e.g. regex for native channels.
+- `engage_mode` / `engage_pattern` / `sender_scope` / `ignored_message_policy`: four orthogonal axes (migration 010) that replaced v1's opaque `trigger_rules` JSON + `response_scope` enum. `engage_mode='pattern'` requires `engage_pattern` (`'.'` matches every message — the "always respond" flavor); `sender_scope='known'` restricts engagement to group members; `ignored_message_policy='accumulate'` keeps ignored messages as context instead of dropping them.
 - **Side effect:** creating a wiring must also populate `agent_destinations` — don't mutate one without the other (see §1.10).
 
 ### 1.4 `users`
@@ -323,6 +327,71 @@ CREATE TABLE container_configs (
 - **Readers:** `src/container-config.ts`, `src/container-runner.ts`, `src/cli/dispatch.ts` (scope enforcement), `src/claude-md-compose.ts`
 - **Writers:** `src/db/container-configs.ts`, `src/modules/self-mod/apply.ts`, `src/backfill-container-configs.ts`
 
+### 1.16 `pending_sender_approvals`
+
+In-flight state for the `unknown_sender_policy = 'request_approval'` flow. A row exists while an admin-approval card is outstanding for a first-time sender in a wired messaging group; `UNIQUE(messaging_group_id, sender_identity)` dedups concurrent attempts from the same sender instead of spamming the admin with repeat cards.
+
+```sql
+CREATE TABLE pending_sender_approvals (
+  id                 TEXT PRIMARY KEY,
+  messaging_group_id TEXT NOT NULL REFERENCES messaging_groups(id),
+  agent_group_id     TEXT NOT NULL REFERENCES agent_groups(id),
+  sender_identity    TEXT NOT NULL,    -- namespaced user id (channel_type:handle)
+  sender_name        TEXT,
+  original_message   TEXT NOT NULL,    -- JSON of the original InboundEvent
+  approver_user_id   TEXT NOT NULL,
+  created_at         TEXT NOT NULL,
+  title              TEXT NOT NULL DEFAULT '',       -- added by migration 013
+  options_json       TEXT NOT NULL DEFAULT '[]',     -- added by migration 013
+  UNIQUE(messaging_group_id, sender_identity)
+);
+```
+
+Deleted on admin approve (after adding the sender as a member) or deny.
+
+- Access layer: `src/modules/permissions/db/pending-sender-approvals.ts`
+- **Readers/writers:** `src/modules/permissions/sender-approval.ts`, `src/modules/permissions/index.ts`, `src/db/sessions.ts` (`getAskQuestionRender`), `src/cli/resources/groups.ts`
+
+### 1.17 `pending_channel_approvals`
+
+In-flight state for the unknown-channel registration flow. When a channel with no `messaging_group_agents` wiring receives a mention or DM, the router escalates to the owner; `PRIMARY KEY(messaging_group_id)` gives free in-flight dedup via `INSERT OR IGNORE` — a second mention while a card is pending drops silently.
+
+```sql
+CREATE TABLE pending_channel_approvals (
+  messaging_group_id TEXT PRIMARY KEY REFERENCES messaging_groups(id),
+  agent_group_id     TEXT NOT NULL REFERENCES agent_groups(id),
+                     -- agent the approved wiring will target (earliest
+                     -- agent_group by created_at, picked at request time)
+  original_message   TEXT NOT NULL,      -- JSON of the original InboundEvent
+  approver_user_id   TEXT NOT NULL,
+  created_at         TEXT NOT NULL,
+  title              TEXT NOT NULL DEFAULT '',       -- added by migration 013
+  options_json       TEXT NOT NULL DEFAULT '[]'      -- added by migration 013
+);
+```
+
+Approve creates the `messaging_group_agents` wiring and replays the triggering event; deny sets `messaging_groups.denied_at` so future messages on that channel drop without re-prompting. Either way, this row is deleted.
+
+- Access layer: `src/modules/permissions/db/pending-channel-approvals.ts`
+- **Readers/writers:** `src/modules/permissions/channel-approval.ts`, `src/modules/permissions/index.ts`, `src/router.ts`, `src/db/sessions.ts` (`getAskQuestionRender`), `src/cli/resources/groups.ts`
+
+### 1.18 `agent_message_policies`
+
+Per-message approval gate on an agent-to-agent connection between two agent groups. No row for a `(from, to)` pair means free flow (no approval required); a row names the `approver` who must sign off on each message.
+
+```sql
+CREATE TABLE agent_message_policies (
+  from_agent_group_id TEXT NOT NULL REFERENCES agent_groups(id),
+  to_agent_group_id   TEXT NOT NULL REFERENCES agent_groups(id),
+  approver            TEXT NOT NULL,
+  created_at          TEXT NOT NULL,
+  PRIMARY KEY (from_agent_group_id, to_agent_group_id)
+);
+```
+
+- Access layer: `src/modules/agent-to-agent/db/agent-message-policies.ts`
+- **Readers/writers:** `src/cli/resources/policies.ts`; approved messages create a row in `pending_approvals` (see §1.11) via the a2a send path.
+
 ---
 
 ## 2. Migration system
@@ -330,21 +399,33 @@ CREATE TABLE container_configs (
 Migrations live in `src/db/migrations/`, one file per migration. Runner: `runMigrations()` in `src/db/migrations/index.ts`. It:
 
 1. Creates `schema_version` if absent.
-2. Reads `MAX(version)` — call it `current`.
-3. For each migration with `version > current`, executes `up(db)` inside a transaction and appends a `schema_version` row.
+2. Reads every already-applied `name` from `schema_version` into a `Set` and filters the `migrations` barrel array down to the ones whose `name` isn't in that set — dedup is by **name**, not by the numeric `version` field.
+3. Runs each pending migration's `up(db)` inside a transaction, in the barrel array's literal order (which is *not* sorted by `version`), then inserts a `schema_version` row.
+4. The `version` column stored in `schema_version` is **not** the migration's own `version` field — it's `COALESCE(MAX(version), 0) + 1`, i.e. an auto-assigned applied-order number computed at insert time. The `version` field on the `Migration` object is just an ordering hint for humans reading the barrel file; it lets module migrations (installed later by skills) pick arbitrary numbers without coordinating with trunk.
 
-| # | File | Introduces |
-|---|------|------------|
-| 001 | `001-initial.ts` | Core tables: `agent_groups`, `messaging_groups`, `messaging_group_agents`, `users`, `user_roles`, `agent_group_members`, `user_dms`, `sessions`, `pending_questions` |
-| 002 | `002-chat-sdk-state.ts` | `chat_sdk_kv`, `chat_sdk_subscriptions`, `chat_sdk_locks`, `chat_sdk_lists` |
-| 003 | `003-pending-approvals.ts` | `pending_approvals` (session-bound + OneCLI fields) |
-| 004 | `004-agent-destinations.ts` | `agent_destinations` + backfill from existing `messaging_group_agents` wirings |
-| 007 | `007-pending-approvals-title-options.ts` | `ALTER TABLE pending_approvals` add `title`, `options_json` (retrofits DBs created between 003 and 007) |
-| 008 | `008-dropped-messages.ts` | `unregistered_senders` |
-| 009 | `009-drop-pending-credentials.ts` | Drop the defunct `pending_credentials` table |
-| 014 | `014-container-configs.ts` | `container_configs` — per-agent-group container runtime config |
-| 015 | `015-cli-scope.ts` | `ALTER TABLE container_configs ADD COLUMN cli_scope` |
+A few migrations also set `disableForeignKeys: true` (needed for table recreates — SQLite can't relax a table-level `UNIQUE` without DROP+RENAME, which fails FK integrity checks with live child rows). The runner toggles `PRAGMA foreign_keys` around the transaction and runs `PRAGMA foreign_key_check` inside it, snapshotting pre-existing violations so it only fails on violations the migration itself introduced.
 
-Numbers 005 and 006 are intentionally absent — migrations were renumbered during early development.
+Several early migrations were later renamed/retired and replaced by "module" files (their original `name` is retained on the new file so already-migrated DBs don't re-run them):
+
+| Ver. | Name (stored in `schema_version`) | File | Introduces |
+|---|---|------|------------|
+| 1 | `initial-v2-schema` | `001-initial.ts` | Core tables: `agent_groups`, `messaging_groups`, `messaging_group_agents` (with the original `trigger_rules`/`response_scope` columns — see v10), `users`, `user_roles`, `agent_group_members`, `user_dms`, `sessions`, `pending_questions` |
+| 2 | `chat-sdk-state` | `002-chat-sdk-state.ts` | `chat_sdk_kv`, `chat_sdk_subscriptions`, `chat_sdk_locks`, `chat_sdk_lists` |
+| 3 | `pending-approvals` | `module-approvals-pending-approvals.ts` | `pending_approvals` (session-bound + OneCLI fields) |
+| 4 | `agent-destinations` | `module-agent-to-agent-destinations.ts` | `agent_destinations` + backfill from existing `messaging_group_agents` wirings |
+| 7 | `pending-approvals-title-options` | `module-approvals-title-options.ts` | Retroactive `ALTER TABLE pending_approvals` add `title`, `options_json` for DBs that ran migration 3 before its `CREATE TABLE` was edited to include those columns |
+| 8 | `dropped-messages` | `008-dropped-messages.ts` | `unregistered_senders` |
+| 9 | `drop-pending-credentials` | `009-drop-pending-credentials.ts` | Drop the defunct `pending_credentials` table |
+| 10 | `engage-modes` | `010-engage-modes.ts` | `messaging_group_agents`: add `engage_mode`, `engage_pattern`, `sender_scope`, `ignored_message_policy`; backfill from `trigger_rules`/`response_scope`; drop those two legacy columns (see §1.3) |
+| 11 | `pending-sender-approvals` | `011-pending-sender-approvals.ts` | `pending_sender_approvals` (see §1.16) |
+| 12 | `channel-registration` | `012-channel-registration.ts` | `messaging_groups.denied_at` + `pending_channel_approvals` (see §1.17) |
+| 13 | `approval-render-metadata` | `013-approval-render-metadata.ts` | `title`, `options_json` columns on `pending_channel_approvals` and `pending_sender_approvals` |
+| 14 | `container-configs` | `014-container-configs.ts` | `container_configs` — per-agent-group container runtime config |
+| 15 | `cli-scope` | `015-cli-scope.ts` | `ALTER TABLE container_configs ADD COLUMN cli_scope` |
+| 16 | `messaging-group-instance` | `016-messaging-group-instance.ts` | `messaging_groups` gets an `instance` column (adapter-instance dimension); table recreate (`disableForeignKeys: true`) backfills `instance = channel_type` on every existing row and relaxes the `UNIQUE` to `(channel_type, platform_id, instance)` |
+| 17 | `agent-message-policies` | `017-agent-message-policies.ts` | `agent_message_policies` (see §1.18) |
+| 18 | `approvals-approver-user-id` | `018-approvals-approver-user-id.ts` | `pending_approvals.approver_user_id` — names a single required approver for a2a message-gate policies |
+
+Numbers 5 and 6 are intentionally absent — migrations were renumbered during early development.
 
 Session DB schemas (`INBOUND_SCHEMA`, `OUTBOUND_SCHEMA`) are **not** versioned here. They're `CREATE TABLE IF NOT EXISTS` so new columns land via the session-DB lazy migration helpers (`migrateDeliveredTable()` etc.) when a session file from an older build is reopened. See [db-session.md](db-session.md).

docs/db-session.md

@@ -10,14 +10,16 @@ Schemas live in `src/db/schema.ts` as the `INBOUND_SCHEMA` and `OUTBOUND_SCHEMA`
 
 ```
 data/v2-sessions/<agent_group_id>/<session_id>/
-  inbound.db              ← host writes, container reads (read-only mount)
+  inbound.db              ← host writes, container reads (read-only open)
   outbound.db             ← container writes, host reads (read-only open)
   .heartbeat              ← mtime touched by container (not a DB write)
   inbox/<message_id>/     ← user attachments, decoded from inbound message content
   outbox/<message_id>/    ← attachments the agent produced
 ```
 
-One session = one folder = one pair of DBs. The `agent_group_id` parent directory also holds per-group state (`.claude-shared/`, `agent-runner-src/`) that is shared across every session of that agent group.
+The session directory itself is mounted read-write into the container (`src/container-runner.ts`) — read-only is *not* a mount property. The container opens `inbound.db` with `{ readonly: true }` at the SQLite connection layer (`container/agent-runner/src/db/connection.ts`), so the container could technically write to the underlying file via another path, but every code path that touches `inbound.db` from inside the container goes through that read-only handle.
+
+One session = one folder = one pair of DBs. The `agent_group_id` parent directory also holds per-group state (`.claude-shared/`) that is shared across every session of that agent group. (The agent-runner source is not copied per group — it's a shared read-only mount from `container/agent-runner/src` into every container; see `src/container-runner.ts`.)
 
 Path helpers in `src/session-manager.ts`: `sessionDir()`, `inboundDbPath()`, `outboundDbPath()`, `heartbeatPath()`.
 
@@ -55,7 +57,7 @@ CREATE INDEX idx_messages_in_series ON messages_in(series_id);
 
 Content shapes: see [api-details.md §Session DB Schema Details](api-details.md#session-db-schema-details).
 
-**Writers (host):** `insertMessage()`, `insertTask()`, `insertRecurrence()` — all in `src/db/session-db.ts`. Each calls `nextEvenSeq()`.
+**Writers (host):** `insertMessage()` (and `nextEvenSeq()`) in `src/db/session-db.ts`; `insertTask()` and `insertRecurrence()` in `src/modules/scheduling/db.ts`. Each calls `nextEvenSeq()`.
 **Reader (container):** `container/agent-runner/src/db/messages-in.ts` — polls `status='pending' AND (process_after IS NULL OR process_after <= now)`.
 
 ### 2.2 `delivered`
@@ -177,6 +179,24 @@ CREATE TABLE session_state (
 
 Access: `container/agent-runner/src/db/session-state.ts`.
 
+### 4.4 `container_state`
+
+Single-row (`id=1`) tool-in-flight tracker. The container records the currently-running tool on `PreToolUse` and clears it on `PostToolUse`/`PostToolUseFailure`; the host reads it during the stale-container sweep to widen its stuck-tolerance window when `Bash` is running with a user-declared `timeout` over the normal threshold, so long-running scripts aren't killed as "stuck".
+
+```sql
+CREATE TABLE container_state (
+  id                       INTEGER PRIMARY KEY CHECK (id = 1),
+  current_tool             TEXT,
+  tool_declared_timeout_ms INTEGER,
+  tool_started_at          TEXT,
+  updated_at               TEXT NOT NULL
+);
+```
+
+- **Writer (container):** `setContainerToolInFlight()` / `clearContainerToolInFlight()` in `container/agent-runner/src/db/connection.ts`, called from the `preToolUseHook` / `postToolUseHook` in `container/agent-runner/src/providers/claude.ts`.
+- **Reader (host):** `getContainerState()` in `src/db/session-db.ts`; consumed by the sweep's `bashTimeoutMs()` helper in `src/host-sweep.ts`.
+- `CREATE TABLE IF NOT EXISTS` — forward-compatible with `outbound.db` files created before this table existed; `getContainerState()` returns `null` if the table or row is absent.
+
 ---
 
 ## 5. Schema evolution

docs/db.md

@@ -35,7 +35,6 @@ data/
   v2-sessions/
     <agent_group_id>/
       .claude-shared/                     ← shared Claude state for the agent group
-      agent-runner-src/                   ← per-group agent-runner overlay
       <session_id>/
         inbound.db                        ← host writes, container reads
         outbound.db                       ← container writes, host reads

docs/docker-sandboxes.md

@@ -1,359 +0,0 @@
-# Running NanoClaw in Docker Sandboxes (Manual Setup)
-
-This guide walks through setting up NanoClaw inside a [Docker Sandbox](https://docs.docker.com/ai/sandboxes/) from scratch — no install script, no pre-built fork. You'll clone the upstream repo, apply the necessary patches, and have agents running in full hypervisor-level isolation.
-
-## Architecture
-
-```
-Host (macOS / Windows WSL)
-└── Docker Sandbox (micro VM with isolated kernel)
-    ├── NanoClaw process (Node.js)
-    │   ├── Channel adapters (WhatsApp, Telegram, etc.)
-    │   └── Container spawner → nested Docker daemon
-    └── Docker-in-Docker
-        └── nanoclaw-agent containers
-            └── Claude Agent SDK
-```
-
-Each agent runs in its own container, inside a micro VM that is fully isolated from your host. Two layers of isolation: per-agent containers + the VM boundary.
-
-The sandbox provides a MITM proxy at `host.docker.internal:3128` that handles network access and injects your Anthropic API key automatically.
-
-> **Note:** This guide is based on a validated setup running on macOS (Apple Silicon) with WhatsApp. Other channels (Telegram, Slack, etc.) and environments (Windows WSL) may require additional proxy patches for their specific HTTP/WebSocket clients. The core patches (container runner, credential proxy, Dockerfile) apply universally — channel-specific proxy configuration varies.
-
-## Prerequisites
-
-- **Docker Desktop v4.40+** with Sandbox support
-- **Anthropic API key** (the sandbox proxy manages injection)
-- For **Telegram**: a bot token from [@BotFather](https://t.me/BotFather) and your chat ID
-- For **WhatsApp**: a phone with WhatsApp installed
-
-Verify sandbox support:
-```bash
-docker sandbox version
-```
-
-## Step 1: Create the Sandbox
-
-On your host machine:
-
-```bash
-# Create a workspace directory
-mkdir -p ~/nanoclaw-workspace
-
-# Create a shell sandbox with the workspace mounted
-docker sandbox create shell ~/nanoclaw-workspace
-```
-
-If you're using WhatsApp, configure proxy bypass so WhatsApp's Noise protocol isn't MITM-inspected:
-
-```bash
-docker sandbox network proxy shell-nanoclaw-workspace \
-  --bypass-host web.whatsapp.com \
-  --bypass-host "*.whatsapp.com" \
-  --bypass-host "*.whatsapp.net"
-```
-
-Telegram does not need proxy bypass.
-
-Enter the sandbox:
-```bash
-docker sandbox run shell-nanoclaw-workspace
-```
-
-## Step 2: Install Prerequisites
-
-Inside the sandbox:
-
-```bash
-sudo apt-get update && sudo apt-get install -y build-essential python3
-npm config set strict-ssl false
-```
-
-## Step 3: Clone and Install NanoClaw
-
-NanoClaw must live inside the workspace directory — Docker-in-Docker can only bind-mount from the shared workspace path.
-
-```bash
-# Clone to home first (virtiofs can corrupt git pack files during clone)
-cd ~
-git clone https://github.com/nanocoai/nanoclaw.git
-
-# Replace with YOUR workspace path (the host path you passed to `docker sandbox create`)
-WORKSPACE=/Users/you/nanoclaw-workspace
-
-# Move into workspace so DinD mounts work
-mv nanoclaw "$WORKSPACE/nanoclaw"
-cd "$WORKSPACE/nanoclaw"
-
-# Install dependencies
-pnpm install
-pnpm install https-proxy-agent
-```
-
-## Step 4: Apply Proxy and Sandbox Patches
-
-NanoClaw needs several patches to work inside a Docker Sandbox. These handle proxy routing, CA certificates, and Docker-in-Docker mount restrictions.
-
-### 4a. Dockerfile — proxy args for container image build
-
-`pnpm install` inside `docker build` fails with `SELF_SIGNED_CERT_IN_CHAIN` because the sandbox's MITM proxy presents its own certificate. Add proxy build args to `container/Dockerfile`:
-
-Add these lines after the `FROM` line:
-
-```dockerfile
-# Accept proxy build args
-ARG http_proxy
-ARG https_proxy
-ARG no_proxy
-ARG NODE_EXTRA_CA_CERTS
-ARG npm_config_strict_ssl=true
-RUN npm config set strict-ssl ${npm_config_strict_ssl}
-```
-
-And after the `RUN pnpm install` line:
-
-```dockerfile
-RUN npm config set strict-ssl true
-```
-
-### 4b. Build script — forward proxy args
-
-Patch `container/build.sh` to pass proxy env vars to `docker build`:
-
-Add these `--build-arg` flags to the `docker build` command:
-
-```bash
---build-arg http_proxy="${http_proxy:-$HTTP_PROXY}" \
---build-arg https_proxy="${https_proxy:-$HTTPS_PROXY}" \
---build-arg no_proxy="${no_proxy:-$NO_PROXY}" \
---build-arg npm_config_strict_ssl=false \
-```
-
-### 4c. Container runner — proxy forwarding, CA cert mount, /dev/null fix
-
-Three changes to `src/container-runner.ts`:
-
-**Replace `/dev/null` shadow mount.** The sandbox rejects `/dev/null` bind mounts. Find where `.env` is shadow-mounted to `/dev/null` and replace it with an empty file:
-
-```typescript
-// Create an empty file to shadow .env (Docker Sandbox rejects /dev/null mounts)
-const emptyEnvPath = path.join(DATA_DIR, 'empty-env');
-if (!fs.existsSync(emptyEnvPath)) fs.writeFileSync(emptyEnvPath, '');
-// Use emptyEnvPath instead of '/dev/null' in the mount
-```
-
-**Forward proxy env vars** to spawned agent containers. Add `-e` flags for `HTTP_PROXY`, `HTTPS_PROXY`, `NO_PROXY` and their lowercase variants.
-
-**Mount CA certificate.** If `NODE_EXTRA_CA_CERTS` or `SSL_CERT_FILE` is set, copy the cert into the project directory and mount it into agent containers:
-
-```typescript
-const caCertSrc = process.env.NODE_EXTRA_CA_CERTS || process.env.SSL_CERT_FILE;
-if (caCertSrc) {
-  const certDir = path.join(DATA_DIR, 'ca-cert');
-  fs.mkdirSync(certDir, { recursive: true });
-  fs.copyFileSync(caCertSrc, path.join(certDir, 'proxy-ca.crt'));
-  // Mount: certDir -> /workspace/ca-cert (read-only)
-  // Set NODE_EXTRA_CA_CERTS=/workspace/ca-cert/proxy-ca.crt in the container
-}
-```
-
-### 4d. Container runtime — prevent self-termination
-
-In `src/container-runtime.ts`, the `cleanupOrphans()` function matches containers by the `nanoclaw-` prefix. Inside a sandbox, the sandbox container itself may match (e.g., `nanoclaw-docker-sandbox`). Filter out the current hostname:
-
-```typescript
-// In cleanupOrphans(), filter out os.hostname() from the list of containers to stop
-```
-
-### 4e. Credential proxy — route through MITM proxy
-
-In `src/credential-proxy.ts`, upstream API requests need to go through the sandbox proxy. Add `HttpsProxyAgent` to outbound requests:
-
-```typescript
-import { HttpsProxyAgent } from 'https-proxy-agent';
-
-const proxyUrl = process.env.HTTPS_PROXY || process.env.https_proxy;
-const upstreamAgent = proxyUrl ? new HttpsProxyAgent(proxyUrl) : undefined;
-// Pass upstreamAgent to https.request() options
-```
-
-### 4f. Setup script — proxy build args
-
-Patch `setup/container.ts` to pass the same proxy `--build-arg` flags as `build.sh` (Step 4b).
-
-## Step 5: Build
-
-```bash
-pnpm run build
-bash container/build.sh
-```
-
-## Step 6: Add a Channel
-
-### Telegram
-
-```bash
-# Apply the Telegram skill
-pnpm exec tsx scripts/apply-skill.ts .claude/skills/add-telegram
-
-# Rebuild after applying the skill
-pnpm run build
-
-# Configure .env
-cat > .env << EOF
-TELEGRAM_BOT_TOKEN=<your-token-from-botfather>
-ASSISTANT_NAME=nanoclaw
-ANTHROPIC_API_KEY=proxy-managed
-EOF
-mkdir -p data/env && cp .env data/env/env
-
-# Register your chat
-pnpm exec tsx setup/index.ts --step register \
-  --jid "tg:<your-chat-id>" \
-  --name "My Chat" \
-  --trigger "@nanoclaw" \
-  --folder "telegram_main" \
-  --channel telegram \
-  --assistant-name "nanoclaw" \
-  --is-main \
-  --no-trigger-required
-```
-
-**To find your chat ID:** Send any message to your bot, then:
-```bash
-curl -s --proxy $HTTPS_PROXY "https://api.telegram.org/bot<TOKEN>/getUpdates" | python3 -m json.tool
-```
-
-**Telegram in groups:** Disable Group Privacy in @BotFather (`/mybots` > Bot Settings > Group Privacy > Turn off), then remove and re-add the bot.
-
-**Important:** If the Telegram skill creates `src/channels/telegram.ts`, you'll need to patch it for proxy support. Add an `HttpsProxyAgent` and pass it to grammy's `Bot` constructor via `baseFetchConfig.agent`. Then rebuild.
-
-### WhatsApp
-
-Make sure you configured proxy bypass in [Step 1](#step-1-create-the-sandbox) first.
-
-```bash
-# Apply the WhatsApp skill
-pnpm exec tsx scripts/apply-skill.ts .claude/skills/add-whatsapp
-
-# Rebuild
-pnpm run build
-
-# Configure .env
-cat > .env << EOF
-ASSISTANT_NAME=nanoclaw
-ANTHROPIC_API_KEY=proxy-managed
-EOF
-mkdir -p data/env && cp .env data/env/env
-
-# Authenticate (choose one):
-
-# QR code — scan with WhatsApp camera:
-pnpm exec tsx src/whatsapp-auth.ts
-
-# OR pairing code — enter code in WhatsApp > Linked Devices > Link with phone number:
-pnpm exec tsx src/whatsapp-auth.ts --pairing-code --phone <phone-number-no-plus>
-
-# Register your chat (JID = your phone number + @s.whatsapp.net)
-pnpm exec tsx setup/index.ts --step register \
-  --jid "<phone>@s.whatsapp.net" \
-  --name "My Chat" \
-  --trigger "@nanoclaw" \
-  --folder "whatsapp_main" \
-  --channel whatsapp \
-  --assistant-name "nanoclaw" \
-  --is-main \
-  --no-trigger-required
-```
-
-**Important:** The WhatsApp skill files (`src/channels/whatsapp.ts` and `src/whatsapp-auth.ts`) also need proxy patches — add `HttpsProxyAgent` for WebSocket connections and a proxy-aware version fetch. Then rebuild.
-
-### Both Channels
-
-Apply both skills, patch both for proxy support, combine the `.env` variables, and register each chat separately.
-
-## Step 7: Run
-
-```bash
-pnpm start
-```
-
-You don't need to set `ANTHROPIC_API_KEY` manually. The sandbox proxy intercepts requests and replaces `proxy-managed` with your real key automatically.
-
-## Networking Details
-
-### How the proxy works
-
-All traffic from the sandbox routes through the host proxy at `host.docker.internal:3128`:
-
-```
-Agent container → DinD bridge → Sandbox VM → host.docker.internal:3128 → Host proxy → api.anthropic.com
-```
-
-**"Bypass" does not mean traffic skips the proxy.** It means the proxy passes traffic through without MITM inspection. Node.js doesn't automatically use `HTTP_PROXY` env vars — you need explicit `HttpsProxyAgent` configuration in every HTTP/WebSocket client.
-
-### Shared paths for DinD mounts
-
-Only the workspace directory is available for Docker-in-Docker bind mounts. Paths outside the workspace fail with "path not shared":
-- `/dev/null` → replace with an empty file in the project dir
-- `/usr/local/share/ca-certificates/` → copy cert to project dir
-- `/home/agent/` → clone to workspace instead
-
-### Git clone and virtiofs
-
-The workspace is mounted via virtiofs. Git's pack file handling can corrupt over virtiofs during clone. Workaround: clone to `/home/agent` first, then `mv` into the workspace.
-
-## Troubleshooting
-
-### pnpm install fails with SELF_SIGNED_CERT_IN_CHAIN
-```bash
-npm config set strict-ssl false
-```
-
-### Container build fails with proxy errors
-```bash
-docker build \
-  --build-arg http_proxy=$http_proxy \
-  --build-arg https_proxy=$https_proxy \
-  -t nanoclaw-agent:latest container/
-```
-
-### Agent containers fail with "path not shared"
-All bind-mounted paths must be under the workspace directory. Check:
-- Is NanoClaw cloned into the workspace? (not `/home/agent/`)
-- Is the CA cert copied to the project root?
-- Has the empty `.env` shadow file been created?
-
-### Agent containers can't reach Anthropic API
-Verify proxy env vars are forwarded to agent containers. Check container logs for `HTTP_PROXY=http://host.docker.internal:3128`.
-
-### WhatsApp error 405
-The version fetch is returning a stale version. Make sure the proxy-aware `fetchWaVersionViaProxy` patch is applied — it fetches `sw.js` through `HttpsProxyAgent` and parses `client_revision`.
-
-### WhatsApp "Connection failed" immediately
-Proxy bypass not configured. From the **host**, run:
-```bash
-docker sandbox network proxy <sandbox-name> \
-  --bypass-host web.whatsapp.com \
-  --bypass-host "*.whatsapp.com" \
-  --bypass-host "*.whatsapp.net"
-```
-
-### Telegram bot doesn't receive messages
-1. Check the grammy proxy patch is applied (look for `HttpsProxyAgent` in `src/channels/telegram.ts`)
-2. Check Group Privacy is disabled in @BotFather if using in groups
-
-### Git clone fails with "inflate: data stream error"
-Clone to a non-workspace path first, then move:
-```bash
-cd ~ && git clone https://github.com/nanocoai/nanoclaw.git && mv nanoclaw /path/to/workspace/nanoclaw
-```
-
-### WhatsApp QR code doesn't display
-Run the auth command interactively inside the sandbox (not piped through `docker sandbox exec`):
-```bash
-docker sandbox run shell-nanoclaw-workspace
-# Then inside:
-pnpm exec tsx src/whatsapp-auth.ts
-```

docs/isolation-model.md

@@ -80,7 +80,7 @@ agent_groups (workspace, memory, CLAUDE.md, personality)
     ↕ many-to-many
 messaging_groups (a specific channel/chat/group on a platform)
     via
-messaging_group_agents (session_mode, trigger_rules, priority)
+messaging_group_agents (session_mode, engage_mode, engage_pattern, sender_scope, ignored_message_policy, priority)
 ```
 
 - **Shared session:** multiple messaging_groups → same agent_group, `session_mode = 'agent-shared'`

docs/onecli-upgrades.md

@@ -10,11 +10,11 @@ Find out what is running and what is required:
 
 ```bash
 cat versions.json                                   # the sanctioned pin
-curl -s http://127.0.0.1:10254/api/health           # reports the running gateway version
+curl -s http://127.0.0.1:10254/api/health           # liveness check; `version` field is typically "unknown", not the gateway version
 curl -s -o /dev/null -w '%{http_code}' http://127.0.0.1:10254/v1/health
 ```
 
-If the last command prints `404`, the server predates the `/v1` API that `@onecli-sh/sdk` 2.x requires — every SDK call will fail with 404s that look transient but are permanent. If your gateway is remote, substitute its host for `127.0.0.1` (it's in `.env` as `ONECLI_URL` / `NANOCLAW_ONECLI_API_HOST`).
+If the last command prints `404`, the server predates the `/v1` API that `@onecli-sh/sdk` 2.x requires — every SDK call will fail with 404s that look transient but are permanent. If your gateway is remote, substitute its host for `127.0.0.1` (it's `ONECLI_URL` in `.env`; `NANOCLAW_ONECLI_API_HOST` is a setup-time override only, not persisted to `.env`).
 
 Why gateways fall behind: the OneCLI installer's docker-compose tracks the `latest` image tag, but Docker never re-pulls a tag — the server freezes at whatever `latest` meant on install day.
 

docs/setup-flow.md

@@ -179,7 +179,7 @@ So during this step:
   marker.
 
 The level-2 log still gets an entry (`auth [interactive] → success`
-with the method — subscription / oauth-token / api-key). Level-3 captures
+with the method — subscription / oauth / api). Level-3 captures
 are optional here; mirroring `script -q` output is tricky and the risk of
 leaking the token to disk outweighs the debugging value.
 
@@ -190,7 +190,7 @@ leaking the token to disk outweighs the debugging value.
 | `nanoclaw.sh` | Top-level wrapper. Phase 1 (bootstrap) and phase 2 (setup:auto) orchestration. Writes bootstrap's raw log + progression entry. `--uninstall` bypasses bootstrap entirely — it execs setup:auto directly (the flow lives in `setup/uninstall/`), or prints manual-cleanup guidance and exits 1 when the TS toolchain is missing. |
 | `setup.sh` | Phase 1 bootstrap: Node, pnpm, native-module verify. Emits its own `BOOTSTRAP` status block (historically printed to stdout; now goes to the bootstrap raw log). |
 | `setup/auto.ts` | Phase 2 driver. Orchestrates the clack UI, step execution, user prompts, and writes to all three log levels for every step it spawns. |
-| `setup/logs.ts` | The logging primitives (`logStep`, `logUserInput`, `logComplete`, `stepRawLog`, `initSetupLog`). Single source of truth for level 2/3 formatting and file paths. |
+| `setup/logs.ts` | The logging primitives (`step`, `userInput`, `complete`, `stepRawLog`, `reset`). Single source of truth for level 2/3 formatting and file paths. |
 | `setup/<step>.ts` | Individual step implementations. Must emit one terminal status block; must not write directly to the terminal. |
 | `setup/register-claude-token.sh` | The Anthropic exception. Inherits stdio, prints its own UI, returns a status to the driver. |
 | `setup/add-telegram.sh` | Non-interactive adapter installer. Reads `TELEGRAM_BOT_TOKEN` from env; never prompts. User-facing bits live in `auto.ts`. |

docs/setup-wiring.md

@@ -14,7 +14,7 @@ Last updated: 2026-04-09
 - Container clears stale `processing_ack` entries on startup (crash recovery)
 - Files: `src/db/schema.ts` (INBOUND_SCHEMA + OUTBOUND_SCHEMA), `src/session-manager.ts`, `src/delivery.ts`, `src/host-sweep.ts`, `container/agent-runner/src/db/connection.ts`, `messages-in.ts`, `messages-out.ts`, `poll-loop.ts`, `mcp-tools/scheduling.ts`, `mcp-tools/interactive.ts`
 - Container image rebuilt with tsconfig (`container/agent-runner/tsconfig.json`)
-- E2E verified: host → Docker container → Claude responds → "E2E works!" ✓
+- E2E verified: host → Docker container → agent responds → "E2E works!" ✓
 
 ### OneCLI Integration
 - `ensureAgent()` call added before `applyContainerConfig()` in `src/container-runner.ts`
@@ -65,11 +65,11 @@ Added `session_mode: 'agent-shared'` for cross-channel shared sessions (e.g. Git
 
 ### Entity Model
 ```
-agent_groups (id, name, folder, agent_provider, container_config)
-    ↕ many-to-many
-messaging_groups (id, channel_type, platform_id, name, is_group, unknown_sender_policy)
+agent_groups (id, name, folder, agent_provider)
+    ↕ many-to-many                       (container runtime config lives in the separate container_configs table)
+messaging_groups (id, channel_type, platform_id, instance, name, is_group, unknown_sender_policy, denied_at)
     via
-messaging_group_agents (messaging_group_id, agent_group_id, trigger_rules, session_mode, priority)
+messaging_group_agents (messaging_group_id, agent_group_id, engage_mode, engage_pattern, sender_scope, ignored_message_policy, session_mode, priority)
 
 users (id, kind, display_name)          -- namespaced as "<channel>:<handle>"
 user_roles (user_id, role, agent_group_id)    -- owner / admin (global or scoped)

docs/templates.md

@@ -2,27 +2,21 @@
 
 A **template** is a reusable folder you stamp into a working agent group: it
 carries the agent's standing instructions, its MCP tool servers, and its skills,
-but **no secrets and no provider**. Point `ncl` (or the setup wizard) at one and
+but **no secrets and no provider**. Point `ncl` at one and
 you get a configured agent in seconds; you choose the runtime/provider
 separately.
 
-Templates are purely additive: no DB migration, no new dependency. **At runtime,
-templates are resolved only from a local directory**: `templates/` at the
+Templates are purely additive: no DB migration, no new dependency. **Templates
+are resolved only from a local directory**: `templates/` at the
 project root by default (committed but shipped empty), or whatever
-`NANOCLAW_TEMPLATES_DIR` points at (a local path only). The setup wizard can also
-discover templates from the public registry
+`NANOCLAW_TEMPLATES_DIR` points at (a local path only). The public registry
 ([`nanocoai/nanoclaw-templates`](https://github.com/nanocoai/nanoclaw-templates))
-and copy a chosen one into your local `templates/` before stamping.
+is a manual copy source — clone or download it yourself and copy the chosen
+template into your local `templates/` before stamping.
 
 ## Using a template
 
-**During install.** `bash nanoclaw.sh` opens the setup wizard. Choose **Template
-setup**, then either **NanoClaw template library** (clones the public registry,
-copies the template you pick into your local `templates/`) or **Local templates**
-(lists what's already in `templates/`). The normal auth step then picks the
-runtime, and the wizard stamps and wires your first agent.
-
-**Anytime, via the CLI:**
+**Via the CLI:**
 
 ```bash
 ncl groups create --template sales/sdr --name "SDR Agent"
@@ -40,8 +34,8 @@ e.g. `sales/sdr` → `templates/sales/sdr`.
 
 For safety the ref must stay inside the templates directory: absolute paths, a
 leading `~`, and `../` escapes are rejected. There is no `--source`, no git URL,
-and no remote fetch at `ncl` time. Populate `templates/` first (by hand, or via
-the setup wizard's library option), then stamp.
+and no remote fetch at `ncl` time. Populate `templates/` first (by hand, e.g.
+copying from the public registry), then stamp.
 
 `NANOCLAW_TEMPLATES_DIR` may point the library at another **local** directory; it
 is never a URL and never changes at runtime.

docs/v1-to-v2-changes.md

@@ -80,7 +80,7 @@ Tasks can exist before a session is awake — the host sweep creates/wakes the c
 
 **v2:** OneCLI Agent Vault. A separate local service at `http://127.0.0.1:10254` holds secrets. Agents are *scoped* to specific secrets and the vault injects them into approved API requests as they leave the container. The container never sees the raw secret value.
 
-Gotcha: auto-created agents default to `selective` secret mode — no secrets attached, even if matching secrets exist in the vault. See the "auto-created agents start in selective secret mode" section of the root CLAUDE.md for the fix (`onecli agents set-secret-mode --mode all`).
+Note: auto-created agents default to `all` secret mode — every vault secret whose host pattern matches is injected automatically. See the "Secret modes" section of the root CLAUDE.md if you want per-agent control (`onecli agents set-secret-mode --mode selective`).
 
 **What the automated migration does:** copies every v1 `.env` key verbatim into v2 `.env`, never overwriting existing v2 keys. The OneCLI vault migration is a separate step owned by the `/init-onecli` skill, which knows how to pull from `.env`.
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "nanoclaw",
-  "version": "2.1.33",
+  "version": "2.1.40",
   "description": "Personal Claude assistant. Lightweight, secure, customizable.",
   "type": "module",
   "packageManager": "pnpm@10.33.0",

repo-tokens/badge.svg

@@ -1,5 +1,5 @@
-<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="90" height="20" role="img" aria-label="207k tokens, 104% of context window">
-  <title>207k tokens, 104% of context window</title>
+<svg xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink" width="90" height="20" role="img" aria-label="213k tokens, 106% of context window">
+  <title>213k tokens, 106% of context window</title>
   <linearGradient id="s" x2="0" y2="100%">
     <stop offset="0" stop-color="#bbb" stop-opacity=".1"/>
     <stop offset="1" stop-opacity=".1"/>
@@ -15,8 +15,8 @@
       <g fill="#fff" text-anchor="middle" font-family="Verdana,Geneva,DejaVu Sans,sans-serif" font-size="11">
         <text aria-hidden="true" x="26" y="15" fill="#010101" fill-opacity=".3">tokens</text>
         <text x="26" y="14">tokens</text>
-        <text aria-hidden="true" x="71" y="15" fill="#010101" fill-opacity=".3">207k</text>
-        <text x="71" y="14">207k</text>
+        <text aria-hidden="true" x="71" y="15" fill="#010101" fill-opacity=".3">213k</text>
+        <text x="71" y="14">213k</text>
       </g>
     </g>
   </a>

scripts/test-v2-agent.ts

@@ -33,7 +33,9 @@ db.exec(`
 `);
 
 // Insert test message
-db.prepare(`INSERT INTO messages_in (id, kind, timestamp, status, content) VALUES (?, 'chat', datetime('now'), 'pending', ?)`).run(
+db.prepare(
+  `INSERT INTO messages_in (id, kind, timestamp, status, content) VALUES (?, 'chat', datetime('now'), 'pending', ?)`,
+).run(
   'test-1',
   JSON.stringify({ sender: 'Gavriel', text: 'Say "Hello from v2!" and nothing else. Do not use any tools.' }),
 );
@@ -44,7 +46,7 @@ db.close();
 process.env.SESSION_DB_PATH = DB_PATH;
 process.env.AGENT_PROVIDER = 'claude';
 
-const { getSessionDb, closeSessionDb } = await import('../container/agent-runner/src/db/connection.js');
+const { getInboundDb, closeSessionDb } = await import('../container/agent-runner/src/db/connection.js');
 const { getUndeliveredMessages } = await import('../container/agent-runner/src/db/messages-out.js');
 const { getPendingMessages } = await import('../container/agent-runner/src/db/messages-in.js');
 const { createProvider } = await import('../container/agent-runner/src/providers/factory.js');

scripts/test-v2-channel-e2e.ts

@@ -71,7 +71,7 @@ import { routeInbound } from '../src/router.js';
 import { setDeliveryAdapter, startActiveDeliveryPoll, stopDeliveryPolls } from '../src/delivery.js';
 import { getChannelAdapter, registerChannelAdapter, initChannelAdapters } from '../src/channels/channel-registry.js';
 import { findSession } from '../src/db/sessions.js';
-import { sessionDbPath } from '../src/session-manager.js';
+import { inboundDbPath } from '../src/session-manager.js';
 import type { ChannelAdapter, ChannelSetup, OutboundMessage } from '../src/channels/adapter.js';
 
 // Track delivered messages
@@ -99,7 +99,9 @@ const mockAdapter: ChannelAdapter = {
 
   async setTyping() {},
   async teardown() {},
-  isConnected() { return true; },
+  isConnected() {
+    return true;
+  },
 };
 
 // Register mock adapter
@@ -179,15 +181,19 @@ console.log(`✓ Container status: ${session.container_status}`);
 import { execSync } from 'child_process';
 const checkContainerLogs = () => {
   try {
-    const containers = execSync('docker ps -a --filter name=nanoclaw-v2-test-channel --format "{{.Names}}"').toString().trim();
+    const containers = execSync('docker ps -a --filter name=nanoclaw-v2-test-channel --format "{{.Names}}"')
+      .toString()
+      .trim();
     for (const name of containers.split('\n').filter(Boolean)) {
       console.log(`\nContainer logs (${name}):`);
       console.log(execSync(`docker logs ${name} 2>&1`).toString());
     }
-  } catch { /* ignore */ }
+  } catch {
+    /* ignore */
+  }
 };
 
-const sessDbPath = sessionDbPath('ag-chan', session.id);
+const sessDbPath = inboundDbPath('ag-chan', session.id);
 console.log(`✓ Session DB: ${sessDbPath}`);
 
 // --- Step 4: Wait for delivery through mock adapter ---
@@ -210,7 +216,9 @@ await new Promise<void>((resolve) => {
         console.log(`  messages_out rows: ${out.length}`);
         if (out.length > 0) console.log('  (messages exist but delivery failed)');
         db.close();
-      } catch { /* ignore */ }
+      } catch {
+        /* ignore */
+      }
       checkContainerLogs();
       cleanup();
       process.exit(1);

setup/add-discord.sh

@@ -1,7 +1,7 @@
 #!/usr/bin/env bash
 #
 # Install the Discord adapter, persist DISCORD_BOT_TOKEN / APPLICATION_ID /
-# PUBLIC_KEY to .env + data/env/env, and restart the service. Non-interactive —
+# PUBLIC_KEY to .env, and restart the service. Non-interactive —
 # the operator-facing "Create a bot" walkthrough, owner confirmation, and
 # server-invite step live in setup/channels/discord.ts. Credentials come in via
 # env vars: DISCORD_BOT_TOKEN, DISCORD_APPLICATION_ID, DISCORD_PUBLIC_KEY.
@@ -105,10 +105,6 @@ upsert_env DISCORD_BOT_TOKEN "$DISCORD_BOT_TOKEN"
 upsert_env DISCORD_APPLICATION_ID "$DISCORD_APPLICATION_ID"
 upsert_env DISCORD_PUBLIC_KEY "$DISCORD_PUBLIC_KEY"
 
-# Container reads from data/env/env (the host mounts it).
-mkdir -p data/env
-cp .env data/env/env
-
 log "Restarting service so the new adapter picks up the credentials…"
 # shellcheck source=setup/lib/install-slug.sh
 source "$PROJECT_ROOT/setup/lib/install-slug.sh"

setup/add-imessage.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 #
-# Install the iMessage adapter, persist mode/creds to .env + data/env/env,
+# Install the iMessage adapter, persist mode/creds to .env,
 # and restart the service. Non-interactive — the Full Disk Access walkthrough
 # (local mode) and Photon URL/key prompts (remote mode) live in
 # setup/channels/imessage.ts. Creds come in via env vars:
@@ -135,10 +135,6 @@ else
   remove_env IMESSAGE_ENABLED
 fi
 
-# Container reads from data/env/env (the host mounts it).
-mkdir -p data/env
-cp .env data/env/env
-
 log "Restarting service so the new adapter picks up the creds…"
 # shellcheck source=setup/lib/install-slug.sh
 source "$PROJECT_ROOT/setup/lib/install-slug.sh"

setup/add-slack.sh

@@ -2,7 +2,7 @@
 #
 # Install the Slack adapter, persist SLACK_BOT_TOKEN plus the mode-specific
 # secret (SLACK_APP_TOKEN for Socket Mode, SLACK_SIGNING_SECRET for webhook) to
-# .env + data/env/env, and restart the service. Non-interactive — the
+# .env, and restart the service. Non-interactive — the
 # operator-facing app creation walkthrough + credential paste live in
 # setup/channels/slack.ts. Credentials come in via env vars:
 # SLACK_BOT_TOKEN, and SLACK_APP_TOKEN and/or SLACK_SIGNING_SECRET.
@@ -108,10 +108,6 @@ if [ -n "${SLACK_SIGNING_SECRET:-}" ]; then
   upsert_env SLACK_SIGNING_SECRET "$SLACK_SIGNING_SECRET"
 fi
 
-# Container reads from data/env/env (the host mounts it).
-mkdir -p data/env
-cp .env data/env/env
-
 log "Restarting service so the new adapter picks up the credentials…"
 # shellcheck source=setup/lib/install-slug.sh
 source "$PROJECT_ROOT/setup/lib/install-slug.sh"

setup/add-teams.sh

@@ -1,7 +1,7 @@
 #!/usr/bin/env bash
 #
 # Install the Teams adapter, persist TEAMS_APP_ID / _PASSWORD / _TENANT_ID /
-# _TYPE to .env + data/env/env, and restart the service. Non-interactive —
+# _TYPE to .env, and restart the service. Non-interactive —
 # the operator-facing Azure portal walkthroughs live in
 # setup/channels/teams.ts. Credentials come in via env vars:
 #   TEAMS_APP_ID            (required)
@@ -114,10 +114,6 @@ if [ -n "${TEAMS_APP_TENANT_ID:-}" ]; then
   upsert_env TEAMS_APP_TENANT_ID "$TEAMS_APP_TENANT_ID"
 fi
 
-# Container reads from data/env/env (the host mounts it).
-mkdir -p data/env
-cp .env data/env/env
-
 log "Restarting service so the new adapter picks up the credentials…"
 # shellcheck source=setup/lib/install-slug.sh
 source "$PROJECT_ROOT/setup/lib/install-slug.sh"

setup/add-telegram.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 #
-# Install the Telegram adapter, persist the bot token to .env + data/env/env,
+# Install the Telegram adapter, persist the bot token to .env,
 # restart the service, and open the bot's chat page in the local Telegram
 # client. Non-interactive — the operator-facing "Create a bot" instructions
 # and token paste live in setup/auto.ts. The token comes in via the
@@ -134,10 +134,6 @@ if echo "$INFO" | grep -q '"ok":true'; then
   BOT_USERNAME=$(echo "$INFO" | sed -nE 's/.*"username":"([^"]+)".*/\1/p')
 fi
 
-# Container reads from data/env/env (the host mounts it).
-mkdir -p data/env
-cp .env data/env/env
-
 # Browser/app deep-link is done by the parent driver (setup/channels/telegram.ts)
 # BEFORE this script runs — gated on a clack confirm so focus-stealing doesn't
 # surprise the user. Keeping it out of here means this script stays pure

setup/channels/signal.ts

@@ -10,7 +10,7 @@
  *   2. Install the adapter + qrcode via setup/add-signal.sh (idempotent).
  *   3. Run the signal-auth step, rendering each SIGNAL_AUTH_QR block as
  *      a terminal QR the operator scans from Signal → Linked Devices.
- *   4. Persist SIGNAL_ACCOUNT to .env (+ data/env/env).
+ *   4. Persist SIGNAL_ACCOUNT to .env.
  *   5. Kick the service so the adapter picks up the new credentials.
  *   6. Ask operator role + agent name.
  *   7. Wire the agent via scripts/init-first-agent.ts; the existing welcome
@@ -333,7 +333,7 @@ async function renderQr(url: string): Promise<string[]> {
   }
 }
 
-/** Persist SIGNAL_ACCOUNT to .env and mirror to data/env/env for the container. */
+/** Persist SIGNAL_ACCOUNT to .env. */
 function writeSignalAccount(account: string): void {
   const envPath = path.join(process.cwd(), '.env');
   let contents = '';
@@ -353,10 +353,6 @@ function writeSignalAccount(account: string): void {
   }
   fs.writeFileSync(envPath, contents);
 
-  const containerEnvDir = path.join(process.cwd(), 'data', 'env');
-  fs.mkdirSync(containerEnvDir, { recursive: true });
-  fs.copyFileSync(envPath, path.join(containerEnvDir, 'env'));
-
   setupLog.userInput('signal_account', account);
 }
 

setup/channels/whatsapp.ts

@@ -433,7 +433,7 @@ async function askChatPhone(authedPhone: string): Promise<string> {
   return phone;
 }
 
-/** Persist ASSISTANT_HAS_OWN_NUMBER=true to .env and data/env/env. */
+/** Persist ASSISTANT_HAS_OWN_NUMBER=true to .env. */
 function writeAssistantHasOwnNumber(): void {
   const envPath = path.join(process.cwd(), '.env');
   let contents = '';
@@ -452,11 +452,6 @@ function writeAssistantHasOwnNumber(): void {
     contents += 'ASSISTANT_HAS_OWN_NUMBER=true\n';
   }
   fs.writeFileSync(envPath, contents);
-
-  // Container reads from data/env/env.
-  const containerEnvDir = path.join(process.cwd(), 'data', 'env');
-  fs.mkdirSync(containerEnvDir, { recursive: true });
-  fs.copyFileSync(envPath, path.join(containerEnvDir, 'env'));
 }
 
 async function resolveAgentName(): Promise<string> {

setup/migrate-v2/channel-auth.ts

@@ -116,15 +116,6 @@ function main(): void {
     channelsProcessed++;
   }
 
-  // Sync to data/env/env
-  if (fs.existsSync(v2EnvPath)) {
-    const containerEnvDir = path.join(process.cwd(), 'data', 'env');
-    try {
-      fs.mkdirSync(containerEnvDir, { recursive: true });
-      fs.copyFileSync(v2EnvPath, path.join(containerEnvDir, 'env'));
-    } catch { /* non-fatal */ }
-  }
-
   console.log(`OK:channels=${channelsProcessed},env_keys=${envKeysCopied},files=${filesCopied}`);
   if (missing.length > 0) {
     console.log(`MISSING:${missing.join(',')}`);

setup/migrate-v2/env.ts

@@ -65,15 +65,6 @@ function main(): void {
     fs.writeFileSync(v2EnvPath, result);
   }
 
-  // Sync to data/env/env (container reads from here)
-  const containerEnvDir = path.join(process.cwd(), 'data', 'env');
-  try {
-    fs.mkdirSync(containerEnvDir, { recursive: true });
-    fs.copyFileSync(v2EnvPath, path.join(containerEnvDir, 'env'));
-  } catch {
-    // Non-fatal
-  }
-
   console.log(`OK:copied=${copied.length},skipped=${skipped.length}`);
   if (copied.length > 0) console.log(`COPIED:${copied.join(',')}`);
 }

setup/set-env.ts

@@ -1,10 +1,9 @@
 /**
- * Step: set-env — Write or update a KEY=VALUE in .env, with optional sync to
- * data/env/env (the container-mounted copy).
+ * Step: set-env — Write or update a KEY=VALUE in .env.
  *
  * Usage:
  *   pnpm exec tsx setup/index.ts --step set-env -- \
- *     --key TELEGRAM_BOT_TOKEN --value "<token>" [--sync-container]
+ *     --key TELEGRAM_BOT_TOKEN --value "<token>"
  *
  * Exists so channel-install flows don't have to invent grep/sed/rm pipelines
  * (which can't be allowlisted tightly — sed can read any file, and each
@@ -21,7 +20,6 @@ import { emitStatus } from './status.js';
 export async function run(args: string[]): Promise<void> {
   const keyIdx = args.indexOf('--key');
   const valueIdx = args.indexOf('--value');
-  const syncContainer = args.includes('--sync-container');
 
   if (keyIdx === -1 || !args[keyIdx + 1]) {
     throw new Error('--key <KEY> is required');
@@ -59,19 +57,9 @@ export async function run(args: string[]): Promise<void> {
   fs.writeFileSync(envFile, content);
   log.info('Updated .env', { key, existed });
 
-  let synced = false;
-  if (syncContainer) {
-    const dataEnvDir = path.join(projectRoot, 'data', 'env');
-    fs.mkdirSync(dataEnvDir, { recursive: true });
-    fs.copyFileSync(envFile, path.join(dataEnvDir, 'env'));
-    synced = true;
-    log.info('Synced .env to container mount', { path: 'data/env/env' });
-  }
-
   emitStatus('SET_ENV', {
     KEY: key,
     EXISTED: existed,
-    SYNCED_TO_CONTAINER: synced,
     STATUS: 'success',
   });
 }

src/claude-md-compose.ts

@@ -11,8 +11,7 @@
  *
  * Runs on every spawn from `container-runner.buildMounts()`. Deterministic —
  * same inputs produce the same CLAUDE.md, and stale fragments are pruned.
- *
- * See `docs/claude-md-composition.md` for the full design.
+ * The composition order and fragment sources are documented inline above.
  */
 import fs from 'fs';
 import path from 'path';

src/cli/client.ts

@@ -43,8 +43,14 @@ async function main(): Promise<void> {
     process.exit(2);
   }
 
-  process.stdout.write(formatResponse(res, json ? 'json' : 'human'));
-  process.exit(res.ok ? 0 : 1);
+  const output =
+    !json && res.ok && res.human !== undefined
+      ? res.human + '\n' // server-rendered view — print verbatim
+      : formatResponse(res, json ? 'json' : 'human');
+  // Exit only after stdout drains: process.exit() discards buffered pipe
+  // writes, silently truncating any response past the 64KB pipe buffer
+  // (bit `ncl sessions list --json` at scale).
+  process.stdout.write(output, () => process.exit(res.ok ? 0 : 1));
 }
 
 function pickTransport(): Transport {

src/cli/commands/help.ts

@@ -5,11 +5,10 @@
  *   ncl groups help         — show group resource details (verbs, columns, enums)
  */
 import { getContainerConfig } from '../../db/container-configs.js';
-import { getResource, getResources } from '../crud.js';
+import { getResources } from '../crud.js';
+import { renderVerbHelp, summaryLine } from '../help-render.js';
 import type { CallerContext } from '../frame.js';
-import { listCommands, register } from '../registry.js';
-
-const GROUP_SCOPE_RESOURCES = new Set(['groups', 'sessions', 'destinations', 'members']);
+import { GROUP_SCOPE_RESOURCES, listCommands, register } from '../registry.js';
 
 function getCliScope(ctx: CallerContext): string | undefined {
   if (ctx.caller !== 'agent') return undefined;
@@ -76,13 +75,28 @@ export function registerResourceHelpCommands(): void {
     try {
       register({
         name: `${res.plural}-help`,
-        description: `Show ${res.name} resource details.`,
+        description: `Show ${res.name} resource details; \`${res.plural} help <verb>\` for one verb in depth.`,
         access: 'open',
         resource: res.plural,
-        parseArgs: () => ({}),
-        handler: async (_args, ctx) => {
+        parseArgs: (raw) => raw,
+        handler: async (args, ctx) => {
           const cliScope = getCliScope(ctx);
           const lines: string[] = [];
+
+          // `ncl <resource> help <verb>` arrives via the dispatcher's
+          // longest-prefix fallback (`groups-help-create` → `groups-help` +
+          // id=`create`) and renders that verb's deep help — full description,
+          // flags, examples. No new routing. Group-scope auto-fill also puts
+          // the caller's agent group ID into `id` on groups/destinations —
+          // that's not a verb request, so ignore it.
+          const autoFilled = ctx.caller === 'agent' && args.id === ctx.agentGroupId;
+          const verbArg = !autoFilled && typeof args.id === 'string' ? args.id : null;
+          if (verbArg) {
+            const deep = renderVerbHelp(res, verbArg);
+            if (!deep) throw new Error(`no verb "${verbArg}" on ${res.plural} — run \`ncl ${res.plural} help\``);
+            return deep;
+          }
+
           lines.push(`${res.plural}: ${res.description}`);
 
           if (cliScope === 'group' && GROUP_SCOPE_RESOURCES.has(res.plural)) {
@@ -92,23 +106,27 @@ export function registerResourceHelpCommands(): void {
 
           lines.push('');
 
-          // Verbs
+          // Verbs — one summary line each; deep help is a verb away. Only the
+          // exceptional access levels are tagged: `open` is the unmarked default.
           const idAutoFilled = cliScope === 'group' && (res.plural === 'groups' || res.plural === 'destinations');
           const idHint = idAutoFilled ? '' : ' <id>';
+          const tag = (access: string | undefined) => (!access || access === 'open' ? '' : ` [${access}]`);
           const verbs: string[] = [];
-          if (res.operations.list) verbs.push(`list [open]`);
-          if (res.operations.get) verbs.push(`get${idHint} [open]`);
-          if (res.operations.create) verbs.push(`create [approval]`);
-          if (res.operations.update) verbs.push(`update${idHint} [approval]`);
-          if (res.operations.delete) verbs.push(`delete${idHint} [approval]`);
+          if (res.operations.list) verbs.push(`list${tag(res.operations.list)}`);
+          if (res.operations.get) verbs.push(`get${idHint}${tag(res.operations.get)}`);
+          if (res.operations.create) verbs.push(`create${tag(res.operations.create)}`);
+          if (res.operations.update) verbs.push(`update${idHint}${tag(res.operations.update)}`);
+          if (res.operations.delete) verbs.push(`delete${idHint}${tag(res.operations.delete)}`);
           if (res.customOperations) {
             for (const [verb, op] of Object.entries(res.customOperations)) {
-              verbs.push(`${verb} [${op.access}] — ${op.description}`);
+              verbs.push(`${verb}${tag(op.access)} — ${summaryLine(op.description)}`);
             }
           }
           lines.push('Verbs:');
           for (const v of verbs) lines.push(`  ${v}`);
           lines.push('');
+          lines.push(`Run \`ncl ${res.plural} help <verb>\` (or add --help to any command) for flags and examples.`);
+          lines.push('');
 
           // Columns
           const autoFilledFields =

src/cli/crud-validate.test.ts

@@ -0,0 +1,173 @@
+import { describe, it, expect, vi } from 'vitest';
+
+vi.mock('../db/connection.js', () => ({ getDb: vi.fn() }));
+vi.mock('../db/container-configs.js', () => ({
+  getContainerConfig: vi.fn(() => ({ cli_scope: 'group' })),
+}));
+
+import { registerResource, validateArgs } from './crud.js';
+import { registerResourceHelpCommands } from './commands/help.js';
+import { lookup } from './registry.js';
+import type { CallerContext } from './frame.js';
+
+// --- validateArgs unit ---
+
+describe('validateArgs', () => {
+  const defs = [
+    { name: 'target', type: 'string' as const, description: 'Target.', required: true },
+    { name: 'count', type: 'number' as const, description: 'Count.', default: 1 },
+    { name: 'force', type: 'boolean' as const, description: 'Force.' },
+    { name: 'meta', type: 'json' as const, description: 'Meta.' },
+    { name: 'size', type: 'string' as const, description: 'Size.', enum: ['s', 'm', 'l'] },
+  ];
+
+  it('coerces types per declaration', () => {
+    const out = validateArgs(defs, { target: 'x', count: '5', force: 'true', meta: '{"a":1}' });
+    expect(out).toMatchObject({ target: 'x', count: 5, force: true, meta: { a: 1 } });
+  });
+
+  it('applies defaults for absent optional flags', () => {
+    expect(validateArgs(defs, { target: 'x' }).count).toBe(1);
+  });
+
+  it('rejects a missing required flag', () => {
+    expect(() => validateArgs(defs, {})).toThrow('--target is required');
+  });
+
+  it('rejects unknown flags', () => {
+    expect(() => validateArgs(defs, { target: 'x', bogus: '1' })).toThrow('unknown flag --bogus');
+  });
+
+  it('tolerates dispatch-injected keys without declaration', () => {
+    const out = validateArgs(defs, { target: 'x', id: 'ag-1', agent_group_id: 'ag-1', group: 'ag-1' });
+    expect(out.id).toBe('ag-1');
+  });
+
+  it('rejects enum violations', () => {
+    expect(() => validateArgs(defs, { target: 'x', size: 'xl' })).toThrow('--size must be one of: s, m, l');
+  });
+
+  it('rejects non-numeric values for number flags', () => {
+    expect(() => validateArgs(defs, { target: 'x', count: 'many' })).toThrow('--count must be a number');
+  });
+
+  it('rejects a value-less flag on a non-boolean (client sends true)', () => {
+    expect(() => validateArgs(defs, { target: true })).toThrow('--target requires a value');
+  });
+
+  it('accepts a value-less boolean flag', () => {
+    expect(validateArgs(defs, { target: 'x', force: true }).force).toBe(true);
+  });
+
+  it('rejects invalid JSON', () => {
+    expect(() => validateArgs(defs, { target: 'x', meta: '{nope' })).toThrow('--meta must be valid JSON');
+  });
+});
+
+// --- registerResource wiring: strict where declared, lenient otherwise ---
+
+registerResource({
+  name: 'widget',
+  plural: 'widgets',
+  table: 'widgets',
+  description: 'Test widgets.',
+  idColumn: 'id',
+  columns: [
+    { name: 'id', type: 'string', description: 'UUID.', generated: true },
+    { name: 'name', type: 'string', description: 'Display name.', required: true, updatable: true },
+  ],
+  operations: {},
+  customOperations: {
+    ping: {
+      access: 'open',
+      description: 'Ping a widget.',
+      args: [{ name: 'target', type: 'string', description: 'Where to ping.', required: true }],
+      examples: ['ncl widgets ping --target prod'],
+      handler: async (args) => ({ echo: args }),
+    },
+    legacy: {
+      access: 'open',
+      description: 'Legacy op without declared args.',
+      handler: async (args) => ({ echo: args }),
+    },
+  },
+});
+registerResourceHelpCommands();
+
+describe('strict validation wiring (declared args)', () => {
+  const parse = lookup('widgets-ping')!.parseArgs;
+
+  it('passes and coerces valid args', () => {
+    expect(parse({ target: 'prod' })).toMatchObject({ target: 'prod' });
+  });
+
+  it('failure carries the verb usage block (error + fix in one round-trip)', () => {
+    let message = '';
+    try {
+      parse({});
+    } catch (e) {
+      message = (e as Error).message;
+    }
+    expect(message).toContain('--target is required');
+    expect(message).toContain('ncl widgets ping'); // usage line
+    expect(message).toContain('Flags:');
+    expect(message).toContain('Examples:');
+  });
+
+  it('rejects unknown flags with the usage block', () => {
+    expect(() => parse({ target: 'prod', bogus: '1' })).toThrow(/unknown flag --bogus[\s\S]*Flags:/);
+  });
+
+  it('normalizes dashed flags before validating', () => {
+    // --target arrives as raw key "target"; a dashed alias like "tar-get" would
+    // normalize to underscores — prove normalize runs before validate.
+    expect(() => parse({ 'bogus-flag': '1', target: 'x' })).toThrow('unknown flag --bogus-flag');
+  });
+});
+
+describe('lenient ops (no declared args) keep legacy behavior', () => {
+  it('passes stray flags through untouched', () => {
+    const parse = lookup('widgets-legacy')!.parseArgs;
+    expect(parse({ anything: 'goes', 'dash-key': '1' })).toMatchObject({ anything: 'goes', dash_key: '1' });
+  });
+});
+
+// --- resource help: deep verb view + group-scope auto-fill guard ---
+
+describe('resource help command', () => {
+  const helpCmd = lookup('widgets-help')!;
+  const host: CallerContext = { caller: 'host' };
+  const agent: CallerContext = {
+    caller: 'agent',
+    sessionId: 'sess-1',
+    agentGroupId: 'ag-1',
+    messagingGroupId: 'mg-1',
+  };
+
+  it('renders the resource overview with verb summaries', async () => {
+    const out = (await helpCmd.handler(helpCmd.parseArgs({}), host)) as string;
+    expect(out).toContain('widgets: Test widgets.');
+    expect(out).toContain('ping — Ping a widget.');
+    expect(out).toContain('help <verb>');
+  });
+
+  it('renders deep help for `help <verb>` (id from prefix fallback)', async () => {
+    const out = (await helpCmd.handler(helpCmd.parseArgs({ id: 'ping' }), host)) as string;
+    expect(out).toContain('ncl widgets ping');
+    expect(out).toContain('--target');
+    expect(out).toContain('Examples:');
+  });
+
+  it('errors on an unknown verb', async () => {
+    await expect(helpCmd.handler(helpCmd.parseArgs({ id: 'bogus' }), host)).rejects.toThrow(
+      'no verb "bogus" on widgets',
+    );
+  });
+
+  it('treats an auto-filled agent group id as no verb (scoped agent, plain help)', async () => {
+    // dispatch auto-fills id=ctx.agentGroupId on groups/destinations; the
+    // handler must show the overview, not "no verb <uuid>".
+    const out = (await helpCmd.handler(helpCmd.parseArgs({ id: 'ag-1' }), agent)) as string;
+    expect(out).toContain('widgets: Test widgets.');
+  });
+});

src/cli/crud.ts

@@ -9,6 +9,7 @@
 import { randomUUID } from 'crypto';
 
 import { getDb } from '../db/connection.js';
+import { renderVerbHelp } from './help-render.js';
 import { register } from './registry.js';
 import type { Access } from './registry.js';
 import type { CallerContext } from './frame.js';
@@ -37,9 +38,21 @@ export interface ColumnDef {
 
 export interface CustomOperation {
   access: Access;
+  /** First line = one-line summary (resource help). Full text renders in the
+   *  per-verb deep help (`ncl <resource> help <verb>` / `--help`). */
   description: string;
+  /**
+   * Declaring args opts this verb into strict validation: required/enum/type
+   * checks plus unknown-flag rejection, with the verb's generated usage block
+   * appended to every failure. Omit to keep the legacy lenient behavior
+   * (handler validates by hand, stray flags ignored).
+   */
   args?: ColumnDef[];
+  /** Ready-to-paste invocations, rendered under EXAMPLES in deep help. */
+  examples?: string[];
   handler: (args: Record<string, unknown>, ctx: CallerContext) => Promise<unknown>;
+  /** Presentational renderer for human mode — see CommandDef.formatHuman. */
+  formatHuman?: (data: unknown) => string;
 }
 
 export interface ResourceDef {
@@ -110,8 +123,11 @@ function genericList(def: ResourceDef) {
     }
     const where = filters.length > 0 ? ` WHERE ${filters.join(' AND ')}` : '';
     params.push(limit);
+    // Newest first: without an ORDER BY the LIMIT silently hides the most
+    // recently inserted rows once a table outgrows it (bit `sessions list`
+    // past 200 sessions — a just-created session was invisible).
     return getDb()
-      .prepare(`SELECT ${cols} FROM ${def.table}${where} LIMIT ?`)
+      .prepare(`SELECT ${cols} FROM ${def.table}${where} ORDER BY rowid DESC LIMIT ?`)
       .all(...params);
   };
 }
@@ -222,6 +238,85 @@ function normalizeArgs(raw: Record<string, unknown>): Record<string, unknown> {
   return out;
 }
 
+// ---------------------------------------------------------------------------
+// Strict arg validation (opt-in via CustomOperation.args)
+// ---------------------------------------------------------------------------
+
+/**
+ * Keys the dispatcher may inject into `req.args` before parseArgs runs
+ * (group-scope auto-fill). Strict validation must tolerate them even when a
+ * verb doesn't declare them, or every scoped agent call breaks.
+ */
+const DISPATCH_INJECTED_KEYS = ['id', 'agent_group_id', 'group'] as const;
+
+/**
+ * Validate `args` (already underscore-normalized) against a ColumnDef list:
+ * unknown-flag rejection, required, enum, and type coercion per the declared
+ * type. Returns a coerced copy; throws with a focused message on the first
+ * problem. Works from any ColumnDef list so generic CRUD resources can opt
+ * into the same strictness later without a second validator.
+ */
+export function validateArgs(
+  defs: ColumnDef[],
+  args: Record<string, unknown>,
+  opts: { allowExtra?: readonly string[] } = {},
+): Record<string, unknown> {
+  const declared = new Map(defs.map((d) => [d.name, d]));
+  const allowed = new Set<string>([...declared.keys(), ...(opts.allowExtra ?? DISPATCH_INJECTED_KEYS)]);
+
+  for (const key of Object.keys(args)) {
+    if (!allowed.has(key)) {
+      throw new Error(`unknown flag --${key.replace(/_/g, '-')}`);
+    }
+  }
+
+  const out: Record<string, unknown> = { ...args };
+  for (const def of defs) {
+    const flag = `--${def.name.replace(/_/g, '-')}`;
+    const v = args[def.name];
+    if (v === undefined) {
+      if (def.required) throw new Error(`${flag} is required`);
+      if (def.default !== undefined) out[def.name] = def.default;
+      continue;
+    }
+    // The client parses a value-less `--flag` as boolean true.
+    if (v === true && def.type !== 'boolean') {
+      throw new Error(`${flag} requires a value`);
+    }
+    switch (def.type) {
+      case 'number': {
+        const n = Number(v);
+        if (Number.isNaN(n)) throw new Error(`${flag} must be a number, got "${v}"`);
+        out[def.name] = n;
+        break;
+      }
+      case 'boolean': {
+        if (v === true || v === 'true' || v === '1') out[def.name] = true;
+        else if (v === false || v === 'false' || v === '0') out[def.name] = false;
+        else throw new Error(`${flag} must be true or false, got "${v}"`);
+        break;
+      }
+      case 'json': {
+        if (typeof v === 'string') {
+          try {
+            out[def.name] = JSON.parse(v);
+          } catch {
+            throw new Error(`${flag} must be valid JSON`);
+          }
+        }
+        break;
+      }
+      case 'string':
+        out[def.name] = String(v);
+        break;
+    }
+    if (def.enum && !def.enum.includes(String(out[def.name]))) {
+      throw new Error(`${flag} must be one of: ${def.enum.join(', ')}`);
+    }
+  }
+  return out;
+}
+
 // ---------------------------------------------------------------------------
 // registerResource
 // ---------------------------------------------------------------------------
@@ -232,6 +327,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.list) {
     register({
       name: `${def.plural}-list`,
+      action: `${def.plural}.list`,
       description: `List all ${def.plural}.`,
       access: def.operations.list,
       resource: def.plural,
@@ -244,6 +340,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.get) {
     register({
       name: `${def.plural}-get`,
+      action: `${def.plural}.get`,
       description: `Get a ${def.name} by ID.`,
       access: def.operations.get,
       resource: def.plural,
@@ -256,6 +353,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.create) {
     register({
       name: `${def.plural}-create`,
+      action: `${def.plural}.create`,
       description: `Create a new ${def.name}.`,
       access: def.operations.create,
       resource: def.plural,
@@ -267,6 +365,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.update) {
     register({
       name: `${def.plural}-update`,
+      action: `${def.plural}.update`,
       description: `Update a ${def.name}.`,
       access: def.operations.update,
       resource: def.plural,
@@ -278,6 +377,7 @@ export function registerResource(def: ResourceDef): void {
   if (def.operations.delete) {
     register({
       name: `${def.plural}-delete`,
+      action: `${def.plural}.delete`,
       description: `Delete a ${def.name}.`,
       access: def.operations.delete,
       resource: def.plural,
@@ -286,16 +386,31 @@ export function registerResource(def: ResourceDef): void {
     });
   }
 
-  // Custom operations
+  // Custom operations. Declaring `args` opts the verb into strict validation;
+  // every failure carries the verb's usage block so a caller (human or agent)
+  // can fix the invocation without a second help round-trip.
   if (def.customOperations) {
     for (const [verb, op] of Object.entries(def.customOperations)) {
+      const declared = op.args;
       register({
         name: `${def.plural}-${verb.replace(/ /g, '-')}`,
+        action: `${def.plural}.${verb.replace(/ /g, '.')}`,
         description: op.description,
         access: op.access,
         resource: def.plural,
-        parseArgs: (raw) => normalizeArgs(raw),
+        parseArgs: declared
+          ? (raw) => {
+              try {
+                return validateArgs(declared, normalizeArgs(raw));
+              } catch (e) {
+                const usage = renderVerbHelp(def, verb);
+                const msg = e instanceof Error ? e.message : String(e);
+                throw new Error(usage ? `${msg}\n\n${usage}` : msg);
+              }
+            }
+          : (raw) => normalizeArgs(raw),
         handler: async (args, ctx) => op.handler(args as Record<string, unknown>, ctx),
+        formatHuman: op.formatHuman,
       });
     }
   }

src/cli/delivery-action.ts

@@ -8,52 +8,57 @@
 import type Database from 'better-sqlite3';
 
 import { registerDeliveryAction } from '../delivery.js';
+import { unguarded } from '../guard/index.js';
 import { insertMessage } from '../db/session-db.js';
 import { log } from '../log.js';
 import { dispatch } from './dispatch.js';
 import type { RequestFrame } from './frame.js';
 import type { Session } from '../types.js';
 
-registerDeliveryAction('cli_request', async (content, session, inDb) => {
-  const requestId = content.requestId as string;
-  const command = content.command as string;
-  const args = (content.args as Record<string, unknown>) ?? {};
+registerDeliveryAction(
+  'cli_request',
+  async (content, session, inDb) => {
+    const requestId = content.requestId as string;
+    const command = content.command as string;
+    const args = (content.args as Record<string, unknown>) ?? {};
 
-  if (!requestId || !command) {
-    log.warn('cli_request missing requestId or command', { sessionId: session.id });
-    return;
-  }
+    if (!requestId || !command) {
+      log.warn('cli_request missing requestId or command', { sessionId: session.id });
+      return;
+    }
 
-  const req: RequestFrame = { id: requestId, command, args };
-  const ctx = {
-    caller: 'agent' as const,
-    sessionId: session.id,
-    agentGroupId: session.agent_group_id,
-    messagingGroupId: session.messaging_group_id ?? '',
-  };
+    const req: RequestFrame = { id: requestId, command, args };
+    const ctx = {
+      caller: 'agent' as const,
+      sessionId: session.id,
+      agentGroupId: session.agent_group_id,
+      messagingGroupId: session.messaging_group_id ?? '',
+    };
 
-  log.info('CLI request from agent', { requestId, command, sessionId: session.id });
+    log.info('CLI request from agent', { requestId, command, sessionId: session.id });
 
-  const response = await dispatch(req, ctx);
+    const response = await dispatch(req, ctx);
 
-  // Write response to inbound.db so the container can read it.
-  // trigger=0: don't wake the agent — this is an inline response to a tool call.
-  insertMessage(inDb, {
-    id: `cli-resp-${requestId}`,
-    kind: 'system',
-    timestamp: new Date().toISOString(),
-    platformId: null,
-    channelType: null,
-    threadId: null,
-    content: JSON.stringify({
-      type: 'cli_response',
-      requestId,
-      frame: response,
-    }),
-    processAfter: null,
-    recurrence: null,
-    trigger: 0,
-  });
+    // Write response to inbound.db so the container can read it.
+    // trigger=0: don't wake the agent — this is an inline response to a tool call.
+    insertMessage(inDb, {
+      id: `cli-resp-${requestId}`,
+      kind: 'system',
+      timestamp: new Date().toISOString(),
+      platformId: null,
+      channelType: null,
+      threadId: null,
+      content: JSON.stringify({
+        type: 'cli_response',
+        requestId,
+        frame: response,
+      }),
+      processAfter: null,
+      recurrence: null,
+      trigger: 0,
+    });
 
-  log.info('CLI response written', { requestId, ok: response.ok, sessionId: session.id });
-});
+    log.info('CLI response written', { requestId, ok: response.ok, sessionId: session.id });
+  },
+  unguarded('transport envelope — every inner command is guarded at dispatch'),
+);

src/cli/dispatch.test.ts

@@ -9,6 +9,7 @@ const approvalState = vi.hoisted(() => ({
     | ((args: {
         session: unknown;
         payload: Record<string, unknown>;
+        approval: Record<string, unknown>;
         userId: string;
         notify: (text: string) => void;
       }) => Promise<void>),
@@ -18,6 +19,7 @@ const approvalState = vi.hoisted(() => ({
       handler: (args: {
         session: unknown;
         payload: Record<string, unknown>;
+        approval: Record<string, unknown>;
         userId: string;
         notify: (text: string) => void;
       }) => Promise<void>,
@@ -43,8 +45,11 @@ vi.mock('../db/agent-groups.js', () => ({
 }));
 
 const mockGetSession = vi.fn();
+// The guard's grant check re-fetches the approval row to prove it's live.
+const mockGetPendingApproval = vi.fn();
 vi.mock('../db/sessions.js', () => ({
   getSession: (...args: unknown[]) => mockGetSession(...args),
+  getPendingApproval: (...args: unknown[]) => mockGetPendingApproval(...args),
 }));
 
 // dispatch's post-handler looks up the resource's `scopeField` via getResource.
@@ -174,6 +179,17 @@ register({
   handler: async () => ({ agent_group_id: 'g1', model: 'opus' }),
 });
 
+// A dash-joined command whose custom-operation key contains spaces
+// ('config update') — used by the --help space/dash bridging test.
+register({
+  name: 'groups-config-update',
+  description: 'bare registry description (should not be the help answer)',
+  resource: 'groups',
+  access: 'open',
+  parseArgs: (raw) => raw,
+  handler: async (args) => ({ echo: args }),
+});
+
 // The real `sessions-get` name — triggers the pre-handler ownership check.
 register({
   name: 'sessions-get',
@@ -461,10 +477,20 @@ describe('CLI scope enforcement', () => {
       callerContext: ctx,
     });
 
+    // The approve path hands the handler the live approval row — the grant
+    // the replay carries back into dispatch.
+    const grantRow = {
+      approval_id: 'appr-t1',
+      action: 'cli_command',
+      payload: JSON.stringify(approval.payload),
+    };
+    mockGetPendingApproval.mockReturnValue(grantRow);
+
     expect(approvalState.approvalHandler).toBeTypeOf('function');
     await approvalState.approvalHandler!({
       session: { id: 's1', agent_group_id: 'g1', messaging_group_id: 'mg1' },
       payload: approval.payload,
+      approval: grantRow,
       userId: 'telegram:admin',
       notify: vi.fn(),
     });
@@ -473,6 +499,73 @@ describe('CLI scope enforcement', () => {
     expect(approvalState.requestApproval).toHaveBeenCalledTimes(1);
   });
 
+  // --- Grant-carrying replay (the `approved: true` boolean no longer exists) ---
+
+  it('replay with a dead grant (row deleted) refuses instead of re-holding', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    mockGetSession.mockReturnValue({ id: 's1', agent_group_id: 'g1', messaging_group_id: 'mg1' });
+    mockGetAgentGroup.mockReturnValue({ id: 'g1', name: 'Group One' });
+
+    const ctx = agentCtx();
+    await dispatch({ id: '1', command: 'approval-context-command', args: {} }, ctx);
+    const approval = approvalState.requestApproval.mock.calls[0][0] as { payload: Record<string, unknown> };
+
+    mockGetPendingApproval.mockReturnValue(undefined); // resolution already deleted the row
+    const notify = vi.fn();
+    await approvalState.approvalHandler!({
+      session: { id: 's1', agent_group_id: 'g1', messaging_group_id: 'mg1' },
+      payload: approval.payload,
+      approval: { approval_id: 'appr-dead', action: 'cli_command', payload: JSON.stringify(approval.payload) },
+      userId: 'telegram:admin',
+      notify,
+    });
+
+    expect(approvalState.observedContexts).toHaveLength(0); // handler never ran
+    expect(approvalState.requestApproval).toHaveBeenCalledTimes(1); // no second card
+    expect(notify.mock.calls[0][0]).toContain('failed');
+  });
+
+  it("a grant approved for one command doesn't transfer to another", async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+
+    // A live cli_command row, but held for a DIFFERENT command.
+    const grantRow = {
+      approval_id: 'appr-other',
+      action: 'cli_command',
+      payload: JSON.stringify({ frame: { command: 'members-add' } }),
+    };
+    mockGetPendingApproval.mockReturnValue(grantRow);
+
+    const resp = await dispatch({ id: '1', command: 'approval-context-command', args: {} }, agentCtx(), {
+      grant: grantRow as never,
+    });
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) {
+      expect(resp.error.code).toBe('forbidden');
+      expect(resp.error.message).toContain('grant');
+    }
+    expect(approvalState.observedContexts).toHaveLength(0);
+  });
+
+  it('a fabricated grant object without a live row is refused', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    mockGetPendingApproval.mockReturnValue(undefined);
+
+    const forged = {
+      approval_id: 'appr-forged',
+      action: 'cli_command',
+      payload: JSON.stringify({ frame: { command: 'approval-context-command' } }),
+    };
+    const resp = await dispatch({ id: '1', command: 'approval-context-command', args: {} }, agentCtx(), {
+      grant: forged as never,
+    });
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) expect(resp.error.code).toBe('forbidden');
+    expect(approvalState.requestApproval).not.toHaveBeenCalled(); // refusal, not a fresh hold
+  });
+
   // --- Post-handler filtering ---
 
   it('group: groups list filters out other groups', async () => {
@@ -595,18 +688,260 @@ describe('CLI scope enforcement', () => {
   });
 });
 
-// --- Dash-joined positional id resolution (generated ids contain dashes) ---
+// Multi-segment command, to prove the longest-prefix match (verb itself has dashes).
+register({
+  name: 'groups-cfg-get',
+  description: 'test multi-segment command',
+  resource: 'groups',
+  access: 'open',
+  parseArgs: (raw) => raw,
+  handler: async (args) => ({ echo: args }),
+});
 
-describe('dash-joined positional id resolution', () => {
-  it('resolves `groups-get-<uuid-with-dashes>` to (groups get, id=<uuid>)', async () => {
-    const uuid = '550e8400-e29b-41d4-a716-446655440000';
+describe('positional dashed-id resolution', () => {
+  const host = { caller: 'host' as const };
 
-    const resp = await dispatch({ id: '1', command: `groups-get-${uuid}`, args: {} }, { caller: 'host' });
+  it('resolves a long dashed id to command + intact id (no shredding)', async () => {
+    const id = 'task-374f0630-d3e0-4965-81da-fe4bf7a6a442';
+    const resp = await dispatch({ id: '1', command: `groups-test-${id}`, args: {} }, host);
+    expect(resp.ok).toBe(true);
+    if (resp.ok) expect(resp.data).toEqual({ echo: { id } });
+  });
+
+  it('matches the LONGEST command prefix when the verb itself has dashes', async () => {
+    const resp = await dispatch({ id: '2', command: 'groups-cfg-get-abc-123', args: {} }, host);
+    expect(resp.ok).toBe(true);
+    if (resp.ok) expect(resp.data).toEqual({ echo: { id: 'abc-123' } });
+  });
+
+  it('leaves a registered no-id command alone', async () => {
+    const resp = await dispatch({ id: '3', command: 'groups-test', args: {} }, host);
+    expect(resp.ok).toBe(true);
+    if (resp.ok) expect(resp.data).toEqual({ echo: {} });
+  });
+
+  it('does not override an explicit --id', async () => {
+    const resp = await dispatch({ id: '4', command: 'groups-test-tail', args: { id: 'explicit' } }, host);
+    expect(resp.ok).toBe(true);
+    if (resp.ok) expect(resp.data).toEqual({ echo: { id: 'explicit' } });
+  });
+});
+
+// --- `--help` interception: answer with generated help, execute nothing ---
+
+describe('--help interception', () => {
+  it('returns command help instead of executing (open command)', async () => {
+    const resp = await dispatch({ id: '1', command: 'test-cmd', args: { help: true } }, { caller: 'host' });
 
     expect(resp.ok).toBe(true);
     if (resp.ok) {
-      const data = resp.data as { echo: Record<string, unknown> };
-      expect(data.echo.id).toBe(uuid);
+      // No deep resource def for 'test' → falls back to the description.
+      expect(resp.data).toBe('test command (non-group resource)');
+    }
+  });
+
+  it('carries the help text in `human` so clients print it verbatim', async () => {
+    const resp = await dispatch({ id: '1', command: 'test-cmd', args: { help: true } }, { caller: 'host' });
+
+    expect(resp.ok).toBe(true);
+    if (resp.ok) {
+      expect(typeof resp.data).toBe('string');
+      expect(resp.human).toBe(resp.data);
+    }
+  });
+
+  it('never mints an approval card for --help on an approval-gated command', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    mockGetSession.mockReturnValue({ id: 's1', agent_group_id: 'g1', messaging_group_id: 'mg1' });
+
+    const resp = await dispatch({ id: '1', command: 'approval-context-command', args: { help: true } }, agentCtx());
+
+    expect(resp.ok).toBe(true);
+    expect(approvalState.requestApproval).not.toHaveBeenCalled();
+    expect(approvalState.observedContexts).toHaveLength(0); // handler never ran
+  });
+
+  it('renders deep verb help when the resource def is available', async () => {
+    mockGetResource.mockImplementation((plural: string) =>
+      plural === 'groups'
+        ? {
+            name: 'group',
+            plural: 'groups',
+            table: 'agent_groups',
+            description: 'Agent groups.',
+            idColumn: 'id',
+            scopeField: 'id',
+            columns: [],
+            operations: {},
+            customOperations: {
+              test: {
+                access: 'open',
+                description: 'Deep test op.',
+                args: [{ name: 'foo', type: 'string', description: 'A foo.', required: true }],
+                handler: async () => ({}),
+              },
+            },
+          }
+        : undefined,
+    );
+
+    const resp = await dispatch({ id: '1', command: 'groups-test', args: { help: true } }, { caller: 'host' });
+
+    expect(resp.ok).toBe(true);
+    if (resp.ok) {
+      expect(resp.data).toContain('ncl groups test');
+      expect(resp.data).toContain('--foo');
+      expect(resp.data).toContain('(required)');
+    }
+  });
+
+  it('renders deep verb help for a multi-word custom-operation key (spaces vs dashes)', async () => {
+    // registerResource stores the op under 'config update' but registers the
+    // command as 'groups-config-update'; help must bridge the two.
+    mockGetResource.mockImplementation((plural: string) =>
+      plural === 'groups'
+        ? {
+            name: 'group',
+            plural: 'groups',
+            table: 'agent_groups',
+            description: 'Agent groups.',
+            idColumn: 'id',
+            scopeField: 'id',
+            columns: [],
+            operations: {},
+            customOperations: {
+              'config update': {
+                access: 'open',
+                description: 'Update container config.',
+                args: [{ name: 'model', type: 'string', description: 'Model override.' }],
+                handler: async () => ({}),
+              },
+            },
+          }
+        : undefined,
+    );
+
+    const resp = await dispatch({ id: '1', command: 'groups-config-update', args: { help: true } }, { caller: 'host' });
+
+    expect(resp.ok).toBe(true);
+    if (resp.ok) {
+      expect(resp.data).toContain('ncl groups config update');
+      expect(resp.data).toContain('--model');
+      // Not the bare registry description fallback:
+      expect(resp.data).not.toBe('bare registry description (should not be the help answer)');
+    }
+  });
+
+  it('still enforces group scope before answering --help', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+
+    const resp = await dispatch({ id: '1', command: 'wirings-list', args: { help: true } }, agentCtx());
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) expect(resp.error.code).toBe('forbidden');
+  });
+});
+
+// --- Unknown-command errors carry their fix ---
+
+describe('unknown-command errors', () => {
+  it('lists the resource verbs when the command names a known resource', async () => {
+    mockGetResource.mockImplementation((plural: string) =>
+      plural === 'groups'
+        ? {
+            name: 'group',
+            plural: 'groups',
+            table: 'agent_groups',
+            description: 'Agent groups.',
+            idColumn: 'id',
+            columns: [],
+            operations: { list: 'open', get: 'open' },
+            customOperations: {
+              restart: { access: 'approval', description: 'Restart.', handler: async () => ({}) },
+            },
+          }
+        : undefined,
+    );
+
+    const resp = await dispatch({ id: '1', command: 'groups-restrat', args: {} }, { caller: 'host' });
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) {
+      expect(resp.error.code).toBe('unknown-command');
+      expect(resp.error.message).toContain('verbs for groups: list, get, restart');
+      expect(resp.error.message).toContain('ncl groups help');
+    }
+  });
+
+  it('suggests the closest command name for near-miss typos', async () => {
+    const resp = await dispatch({ id: '1', command: 'test-cm', args: {} }, { caller: 'host' });
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) {
+      expect(resp.error.code).toBe('unknown-command');
+      expect(resp.error.message).toContain('did you mean "test-cmd"?');
+    }
+  });
+
+  it('falls back to a plain pointer when nothing is close', async () => {
+    const resp = await dispatch({ id: '1', command: 'zzz-qqq-vvv', args: {} }, { caller: 'host' });
+
+    expect(resp.ok).toBe(false);
+    if (!resp.ok) {
+      expect(resp.error.message).toContain('no command "zzz-qqq-vvv"');
+      expect(resp.error.message).toContain('ncl help');
+      expect(resp.error.message).not.toContain('did you mean');
     }
   });
 });
+
+// --- formatHuman hook: server-rendered human view on the frame ---
+
+describe('formatHuman hook', () => {
+  register({
+    name: 'render-cmd',
+    description: 'command with a human renderer',
+    access: 'open',
+    parseArgs: (raw) => raw,
+    handler: async () => [{ id: 'x1', status: 'live' }],
+    formatHuman: (rows) => `TABLE(${(rows as { id: string }[]).map((r) => r.id).join(',')})`,
+  });
+
+  register({
+    name: 'render-throws',
+    description: 'command whose renderer throws',
+    access: 'open',
+    parseArgs: (raw) => raw,
+    handler: async () => ({ fine: true }),
+    formatHuman: () => {
+      throw new Error('renderer bug');
+    },
+  });
+
+  it('attaches human alongside data', async () => {
+    const resp = await dispatch({ id: '1', command: 'render-cmd', args: {} }, { caller: 'host' });
+
+    expect(resp.ok).toBe(true);
+    if (resp.ok) {
+      expect(resp.human).toBe('TABLE(x1)');
+      expect(resp.data).toEqual([{ id: 'x1', status: 'live' }]); // machine contract intact
+    }
+  });
+
+  it('a throwing renderer degrades to a plain frame, never an error', async () => {
+    const resp = await dispatch({ id: '1', command: 'render-throws', args: {} }, { caller: 'host' });
+
+    expect(resp.ok).toBe(true);
+    if (resp.ok) {
+      expect(resp.human).toBeUndefined();
+      expect(resp.data).toEqual({ fine: true });
+    }
+  });
+
+  it('commands without a renderer stay human-less', async () => {
+    const resp = await dispatch({ id: '1', command: 'test-cmd', args: {} }, { caller: 'host' });
+
+    expect(resp.ok).toBe(true);
+    if (resp.ok) expect(resp.human).toBeUndefined();
+  });
+});

src/cli/dispatch.ts

@@ -3,22 +3,37 @@
  * the per-session DB poller (container caller) call dispatch() with the
  * same frame and a transport-supplied CallerContext.
  *
- * Approval gating for risky calls from the container is the only branch
- * that differs by caller. Host callers and `open` commands run inline.
+ * Every command passes the guard before its handler runs — the decision
+ * (allow / hold / deny) comes from the command's catalog entry, derived at
+ * registration (see cli/guard.ts). Dispatch keeps the mechanics: arg
+ * auto-fill, the sessions-get existence oracle, `--help` interception,
+ * parseArgs, and post-handler row filtering. An approved replay re-enters
+ * here carrying the verified approval row as its grant — the guard re-checks
+ * the structural baseline live, and the `approved: true` boolean no longer
+ * exists.
  */
 import { getContainerConfig } from '../db/container-configs.js';
 import { getAgentGroup } from '../db/agent-groups.js';
 import { getSession } from '../db/sessions.js';
+import { guard, type GuardActor } from '../guard/index.js';
 import { registerApprovalHandler, requestApproval } from '../modules/approvals/index.js';
+import type { PendingApproval } from '../types.js';
 import type { CallerContext, ErrorCode, RequestFrame, ResponseFrame } from './frame.js';
 import { getResource } from './crud.js';
-import { lookup } from './registry.js';
+import { listVerbs, renderVerbHelp } from './help-render.js';
+import { commandGuard, listCommands, lookup } from './registry.js';
 
 type DispatchOptions = {
-  /** True when a command is being replayed after approval. */
-  approved?: boolean;
+  /** Verified approval row when a command is replayed after approval. */
+  grant?: PendingApproval;
 };
 
+function actorFor(ctx: CallerContext): GuardActor {
+  return ctx.caller === 'host'
+    ? { kind: 'host' }
+    : { kind: 'agent', agentGroupId: ctx.agentGroupId, sessionId: ctx.sessionId };
+}
+
 export async function dispatch(
   req: RequestFrame,
   ctx: CallerContext,
@@ -26,21 +41,22 @@ export async function dispatch(
 ): Promise<ResponseFrame> {
   let cmd = lookup(req.command);
 
-  // Fallback: if the full command isn't registered, split the dash-joined
-  // command and treat the longest registered prefix as the command, with the
-  // re-joined remainder as the target ID. Clients join all positional args
-  // with dashes (e.g. `ncl groups get abc123` → command "groups-get-abc123"),
-  // and generated ids (UUIDs, `sess-…`, `appr-…`) themselves contain dashes,
-  // so trimming a single trailing segment isn't enough — walk prefixes from
-  // longest to shortest so `groups-get-<uuid-with-dashes>` still resolves to
-  // "groups-get" + id "<uuid-with-dashes>".
+  // Fallback: if the full command isn't registered, find the LONGEST registered
+  // command that is a dash-prefix of req.command; the remainder is the target ID,
+  // kept intact (dashes and all). This lets clients join all positional args with
+  // dashes — e.g. `ncl groups get abc123` → "groups-get-abc123" → "groups-get" +
+  // id "abc123", and crucially `ncl tasks cancel task-374f-...-442` →
+  // "tasks-cancel" + id "task-374f-...-442" (a dashed id is no longer shredded).
+  // Trimming from the end (longest→shortest) means a multi-segment verb like
+  // "groups-config-add-mcp-server" still matches before any shorter prefix.
   if (!cmd) {
-    const parts = req.command.split('-');
-    for (let i = parts.length - 1; i > 0; i--) {
-      const shortened = parts.slice(0, i).join('-');
+    let shortened = req.command;
+    let idx: number;
+    while ((idx = shortened.lastIndexOf('-')) > 0) {
+      shortened = shortened.slice(0, idx);
       const fallback = lookup(shortened);
       if (fallback) {
-        const tail = parts.slice(i).join('-');
+        const tail = req.command.slice(shortened.length + 1); // full remainder = id, dashes intact
         cmd = fallback;
         req = { ...req, command: shortened, args: { ...req.args, id: req.args.id ?? tail } };
         break;
@@ -49,47 +65,16 @@ export async function dispatch(
   }
 
   if (!cmd) {
-    return err(req.id, 'unknown-command', `no command "${req.command}"`);
+    return err(req.id, 'unknown-command', unknownCommandMessage(req.command));
   }
 
-  // CLI scope enforcement for agent callers
+  // Group-scope mechanics for agent callers (visibility, not policy — the
+  // allow/hold/deny decisions live in the guard baseline, cli/guard.ts).
   if (ctx.caller === 'agent') {
     const configRow = getContainerConfig(ctx.agentGroupId);
     const cliScope = configRow?.cli_scope ?? 'group';
 
-    if (cliScope === 'disabled') {
-      return err(req.id, 'forbidden', 'CLI access is disabled for this agent group.');
-    }
-
     if (cliScope === 'group') {
-      const allowed = new Set(['groups', 'sessions', 'destinations', 'members']);
-      // Only allow whitelisted resources and general commands (no resource, like help)
-      if (cmd.resource && !allowed.has(cmd.resource)) {
-        return err(req.id, 'forbidden', `CLI access is scoped to this agent group. Cannot access "${cmd.resource}".`);
-      }
-
-      // Enforce group scope on all agent-group-related args.
-      // Different resources use different arg names for the agent group ID.
-      // Only check --id for resources where it IS the agent group ID.
-      const groupArgs = ['agent_group_id', 'group'] as const;
-      for (const key of groupArgs) {
-        if (req.args[key] && req.args[key] !== ctx.agentGroupId) {
-          return err(req.id, 'forbidden', 'CLI access is scoped to this agent group.');
-        }
-      }
-      if (
-        (cmd.resource === 'groups' || cmd.resource === 'destinations') &&
-        req.args.id &&
-        req.args.id !== ctx.agentGroupId
-      ) {
-        return err(req.id, 'forbidden', 'CLI access is scoped to this agent group.');
-      }
-
-      // Block cli_scope changes from group-scoped agents (privilege escalation)
-      if (req.args.cli_scope !== undefined || req.args['cli-scope'] !== undefined) {
-        return err(req.id, 'forbidden', 'Cannot change cli_scope from a group-scoped agent.');
-      }
-
       // Auto-fill agent-group-related args so the agent doesn't need
       // to pass its own group ID explicitly.
       const fill: Record<string, unknown> = {
@@ -115,7 +100,33 @@ export async function dispatch(
     }
   }
 
-  if (ctx.caller !== 'host' && cmd.access === 'approval' && !opts.approved) {
+  const decision = guard(commandGuard(cmd.name), {
+    actor: actorFor(ctx),
+    payload: req.args,
+    grant: opts.grant ?? null,
+  });
+
+  if (decision.effect === 'deny') {
+    return err(req.id, 'forbidden', decision.reason);
+  }
+
+  // `--help` interception: answer with the command's generated help instead of
+  // executing. Placed after the guard's deny (a group-scoped agent can't probe
+  // forbidden resources) and BEFORE hold execution — asking for help on an
+  // approval-gated verb must never mint an approval card.
+  if (req.args.help === true) {
+    // Carry the help text in `human` too, so both clients print it verbatim
+    // as clean multi-line text instead of a JSON-stringified blob.
+    const helpText = commandHelp(cmd.name, cmd.resource, cmd.description);
+    return { id: req.id, ok: true, data: helpText, human: helpText };
+  }
+
+  if (decision.effect === 'hold') {
+    if (ctx.caller !== 'agent') {
+      // Holds only arise for agent callers; anything else is a guard bug —
+      // fail closed rather than card a ghost.
+      return err(req.id, 'forbidden', decision.reason);
+    }
     const session = getSession(ctx.sessionId);
     if (!session) {
       return err(req.id, 'handler-error', 'Session not found.');
@@ -186,16 +197,27 @@ export async function dispatch(
       }
     }
 
+    // Server-render the human view once, so every transport — host CLI and
+    // the Bun container client (which can't import host formatters) — prints
+    // one canonical rendering. Runs after scope filtering; a throwing
+    // formatter degrades to plain `data`, never fails the response.
+    if (cmd.formatHuman) {
+      try {
+        return { id: req.id, ok: true, data, human: cmd.formatHuman(data) };
+      } catch {
+        // fall through to the plain frame
+      }
+    }
     return { id: req.id, ok: true, data };
   } catch (e) {
     return err(req.id, 'handler-error', errMsg(e));
   }
 }
 
-registerApprovalHandler('cli_command', async ({ payload, notify }) => {
+registerApprovalHandler('cli_command', async ({ payload, approval, notify }) => {
   const frame = payload.frame as RequestFrame;
   const callerContext = parseCallerContext(payload.callerContext) ?? { caller: 'host' };
-  const response = await dispatch(frame, callerContext, { approved: true });
+  const response = await dispatch(frame, callerContext, { grant: approval });
 
   if (response.ok) {
     const data = typeof response.data === 'string' ? response.data : JSON.stringify(response.data, null, 2);
@@ -225,6 +247,83 @@ function parseCallerContext(value: unknown): CallerContext | undefined {
   return undefined;
 }
 
+/** Help text for a resolved command: deep verb help when derivable, else description. */
+function commandHelp(name: string, resource: string | undefined, description: string): string {
+  if (resource && name.startsWith(`${resource}-`)) {
+    const res = getResource(resource);
+    const verb = name.slice(resource.length + 1);
+    const deep = res && renderVerbHelp(res, verb);
+    if (deep) return deep;
+    // Custom-operation KEYS may contain spaces ('config update') while command
+    // names are dash-joined ('groups-config-update'). Resolve by matching keys
+    // normalized the same way registerResource builds command names.
+    if (res?.customOperations) {
+      const spaced = Object.keys(res.customOperations).find((k) => k.replace(/ /g, '-') === verb);
+      const deepSpaced = spaced && renderVerbHelp(res, spaced);
+      if (deepSpaced) return deepSpaced;
+    }
+  }
+  return description;
+}
+
+/**
+ * Unknown-command error that carries its fix: if the command names a known
+ * resource, list that resource's verbs; otherwise suggest the closest
+ * registered command. Resource detection walks dash-prefixes longest-first,
+ * same as the ID fallback above, so multi-word plurals (messaging-groups,
+ * user-dms) resolve.
+ */
+function unknownCommandMessage(command: string): string {
+  const parts = command.split('-');
+  for (let i = parts.length; i > 0; i--) {
+    const prefix = parts.slice(0, i).join('-');
+    const res = getResource(prefix);
+    if (res) {
+      return (
+        `no command "${command}" — verbs for ${res.plural}: ${listVerbs(res).join(', ')}. ` +
+        `Run \`ncl ${res.plural} help <verb>\` for flags and examples.`
+      );
+    }
+  }
+  const names = listCommands()
+    .filter((c) => c.access !== 'hidden')
+    .map((c) => c.name);
+  const closest = closestName(command, names);
+  return `no command "${command}"${closest ? ` — did you mean "${closest}"?` : ''} Run \`ncl help\`.`;
+}
+
+/** Closest name by edit distance, only when convincingly close (≤2 edits). */
+function closestName(input: string, names: string[]): string | undefined {
+  let best: string | undefined;
+  let bestDist = 3;
+  for (const name of names) {
+    if (Math.abs(name.length - input.length) >= bestDist) continue;
+    const d = editDistance(input, name, bestDist);
+    if (d < bestDist) {
+      bestDist = d;
+      best = name;
+    }
+  }
+  return best;
+}
+
+function editDistance(a: string, b: string, cap: number): number {
+  const prev = new Array(b.length + 1).fill(0).map((_, i) => i);
+  for (let i = 1; i <= a.length; i++) {
+    let diag = prev[0];
+    prev[0] = i;
+    let rowMin = prev[0];
+    for (let j = 1; j <= b.length; j++) {
+      const tmp = prev[j];
+      prev[j] = Math.min(prev[j] + 1, prev[j - 1] + 1, diag + (a[i - 1] === b[j - 1] ? 0 : 1));
+      diag = tmp;
+      rowMin = Math.min(rowMin, prev[j]);
+    }
+    if (rowMin >= cap) return cap;
+  }
+  return prev[b.length];
+}
+
 function err(id: string, code: ErrorCode, message: string): ResponseFrame {
   return { id, ok: false, error: { code, message } };
 }

src/cli/frame.ts

@@ -18,7 +18,14 @@ export type RequestFrame = {
 };
 
 export type ResponseFrame =
-  | { id: string; ok: true; data: unknown }
+  // `human` is an optional server-rendered presentational string. It lets
+  // every transport — host CLI and the Bun container client — print one
+  // canonical rendering without importing host-only formatters (the two
+  // runtimes share no modules, so client-side formatters drift). `data`
+  // stays the machine contract; --json callers ignore `human`. Additive:
+  // old clients that don't know the field just fall back to their own
+  // rendering of `data`.
+  | { id: string; ok: true; data: unknown; human?: string }
   | { id: string; ok: false; error: { code: ErrorCode; message: string } };
 
 export type ErrorCode =

src/cli/guard.ts

@@ -0,0 +1,86 @@
+/**
+ * CLI guard adapter — the command registry's catalog derivation and
+ * structural baseline, moved verbatim out of dispatch.ts (guarded-actions
+ * phase 2). Declaration is registration: registry.register() derives one
+ * catalog entry per command from the CommandDef itself; no second file is
+ * edited when a command is added.
+ *
+ * The baseline carries today's decisions exactly:
+ *   host caller → allow (the 0600 socket is the auth story — in code,
+ *   unremovable by data);
+ *   cli_scope 'disabled' → deny; 'group' → resource allowlist, cross-group
+ *   arg denial, cli_scope-change denial;
+ *   access 'approval' for agent callers → hold for the group's admin chain.
+ *
+ * Arg auto-fill, the sessions-get existence oracle, and post-handler row
+ * filtering stay in dispatch.ts — mechanics, not policy.
+ */
+import { getContainerConfig } from '../db/container-configs.js';
+import { ALLOW, DENY, HOLD, type GuardedActionSpec, type GuardInput } from '../guard/index.js';
+import { GROUP_SCOPE_RESOURCES, type CommandDef } from './registry.js';
+
+/** Dotted catalog action name for a command. */
+export function commandGuardAction(cmd: Pick<CommandDef, 'name' | 'action'>): string {
+  return cmd.action ?? `cli.${cmd.name}`;
+}
+
+/** Catalog entry derived from a CommandDef at registration time. */
+export function commandGuardSpec(cmd: CommandDef): GuardedActionSpec {
+  return {
+    action: commandGuardAction(cmd),
+    approvalAction: cmd.access === 'approval' ? 'cli_command' : undefined,
+    // Bind a cli_command grant to the exact command it was approved for.
+    grantMatches: (grant) => {
+      try {
+        const payload = JSON.parse(grant.payload) as { frame?: { command?: string } };
+        return payload.frame?.command === cmd.name;
+      } catch {
+        return false;
+      }
+    },
+    baseline: (input) => commandBaseline(cmd, input),
+  };
+}
+
+function commandBaseline(cmd: CommandDef, input: GuardInput) {
+  const { actor } = input;
+  if (actor.kind === 'host') return ALLOW('host caller (trusted socket)');
+  if (actor.kind !== 'agent') return DENY('CLI commands accept host or agent callers only.');
+
+  const args = input.payload;
+  const cliScope = getContainerConfig(actor.agentGroupId)?.cli_scope ?? 'group';
+
+  if (cliScope === 'disabled') {
+    return DENY('CLI access is disabled for this agent group.');
+  }
+
+  if (cliScope === 'group') {
+    // Only allow whitelisted resources and general commands (no resource, like help)
+    if (cmd.resource && !GROUP_SCOPE_RESOURCES.has(cmd.resource)) {
+      return DENY(`CLI access is scoped to this agent group. Cannot access "${cmd.resource}".`);
+    }
+
+    // Enforce group scope on all agent-group-related args.
+    // Different resources use different arg names for the agent group ID.
+    // Only check --id for resources where it IS the agent group ID.
+    for (const key of ['agent_group_id', 'group'] as const) {
+      if (args[key] && args[key] !== actor.agentGroupId) {
+        return DENY('CLI access is scoped to this agent group.');
+      }
+    }
+    if ((cmd.resource === 'groups' || cmd.resource === 'destinations') && args.id && args.id !== actor.agentGroupId) {
+      return DENY('CLI access is scoped to this agent group.');
+    }
+
+    // Block cli_scope changes from group-scoped agents (privilege escalation)
+    if (args.cli_scope !== undefined || args['cli-scope'] !== undefined) {
+      return DENY('Cannot change cli_scope from a group-scoped agent.');
+    }
+  }
+
+  if (cmd.access === 'approval') {
+    return HOLD(`agent-initiated "${cmd.name}" requires admin approval`);
+  }
+
+  return ALLOW('open command');
+}

src/cli/help-render.test.ts

@@ -0,0 +1,95 @@
+import { describe, it, expect } from 'vitest';
+
+import type { ResourceDef } from './crud.js';
+import { listVerbs, renderVerbHelp, summaryLine } from './help-render.js';
+
+const res: ResourceDef = {
+  name: 'widget',
+  plural: 'widgets',
+  table: 'widgets',
+  description: 'Test widgets.',
+  idColumn: 'id',
+  columns: [
+    { name: 'id', type: 'string', description: 'UUID.', generated: true },
+    { name: 'name', type: 'string', description: 'Display name.', required: true, updatable: true },
+    { name: 'size', type: 'string', description: 'Widget size.', enum: ['s', 'm', 'l'], default: 'm' },
+  ],
+  operations: { list: 'open', get: 'open', create: 'approval', update: 'approval' },
+  customOperations: {
+    ping: {
+      access: 'open',
+      description: 'Ping a widget.\nLonger prose that only deep help shows.',
+      args: [
+        { name: 'target', type: 'string', description: 'Where to ping.', required: true },
+        { name: 'count', type: 'number', description: 'How many times.', default: 1 },
+      ],
+      examples: ['ncl widgets ping --target prod --count 3'],
+      handler: async () => ({}),
+    },
+  },
+};
+
+describe('listVerbs', () => {
+  it('lists enabled generics then custom verbs', () => {
+    expect(listVerbs(res)).toEqual(['list', 'get', 'create', 'update', 'ping']);
+  });
+});
+
+describe('renderVerbHelp — custom operation', () => {
+  it('renders usage, full description, flags with tags, and examples', () => {
+    const out = renderVerbHelp(res, 'ping')!;
+    expect(out).toContain('ncl widgets ping');
+    expect(out).toContain('Longer prose that only deep help shows.');
+    expect(out).toContain('--target');
+    expect(out).toContain('(required)');
+    expect(out).toContain('--count');
+    expect(out).toContain('default: 1');
+    expect(out).toContain('Examples:');
+    expect(out).toContain('ncl widgets ping --target prod --count 3');
+  });
+
+  it('tags non-open access on the usage line', () => {
+    const gated: ResourceDef = {
+      ...res,
+      customOperations: { ping: { ...res.customOperations!.ping, access: 'approval' } },
+    };
+    expect(renderVerbHelp(gated, 'ping')).toContain('ncl widgets ping [approval]');
+  });
+});
+
+describe('renderVerbHelp — generic verbs', () => {
+  it('create renders non-generated columns as flags', () => {
+    const out = renderVerbHelp(res, 'create')!;
+    expect(out).toContain('ncl widgets create [approval]');
+    expect(out).toContain('--name');
+    expect(out).toContain('(required)');
+    expect(out).toContain('--size');
+    expect(out).toContain('values: s | m | l');
+    expect(out).not.toContain('--id'); // generated
+  });
+
+  it('update renders only updatable columns and takes <id>', () => {
+    const out = renderVerbHelp(res, 'update')!;
+    expect(out).toContain('ncl widgets update <id> [approval]');
+    expect(out).toContain('--name');
+    expect(out).not.toContain('--size');
+  });
+
+  it('list renders filter flags plus --limit, never marked required', () => {
+    const out = renderVerbHelp(res, 'list')!;
+    expect(out).toContain('--limit');
+    expect(out).toContain('--name');
+    expect(out).not.toContain('(required)');
+  });
+
+  it('returns undefined for verbs the resource does not have', () => {
+    expect(renderVerbHelp(res, 'delete')).toBeUndefined(); // not in operations
+    expect(renderVerbHelp(res, 'bogus')).toBeUndefined();
+  });
+});
+
+describe('summaryLine', () => {
+  it('returns only the first line', () => {
+    expect(summaryLine('Ping a widget.\nLonger prose.')).toBe('Ping a widget.');
+  });
+});

src/cli/help-render.ts

@@ -0,0 +1,115 @@
+/**
+ * Pure renderers for command help. Single source for three surfaces that must
+ * never disagree:
+ *   - `ncl <resource> help [<verb>]` (commands/help.ts)
+ *   - `--help` on any command (dispatch interception)
+ *   - the usage block appended to invalid-args errors (crud.ts validation)
+ *
+ * Imports only types from crud.ts, so crud.ts can import these functions at
+ * runtime without a cycle.
+ */
+import type { ColumnDef, CustomOperation, ResourceDef } from './crud.js';
+
+const GENERIC_VERBS = ['list', 'get', 'create', 'update', 'delete'] as const;
+type GenericVerb = (typeof GENERIC_VERBS)[number];
+
+export function flagName(col: Pick<ColumnDef, 'name'>): string {
+  return `--${col.name.replace(/_/g, '-')}`;
+}
+
+/** First line of a possibly multi-paragraph description. */
+export function summaryLine(description: string): string {
+  return description.split('\n', 1)[0];
+}
+
+/** Indent every non-empty line of a block by `pad`. */
+export function indent(text: string, pad: string): string {
+  return text
+    .split('\n')
+    .map((l) => (l ? pad + l : l))
+    .join('\n');
+}
+
+function flagLine(col: ColumnDef, extraTags: string[] = []): string {
+  const tags: string[] = [...extraTags];
+  if (col.required) tags.push('required');
+  if (col.default !== undefined && col.default !== null) tags.push(`default: ${col.default}`);
+  if (col.enum) tags.push(`values: ${col.enum.join(' | ')}`);
+  const tagStr = tags.length > 0 ? ` (${tags.join(', ')})` : '';
+  return `  ${flagName(col).padEnd(28)} ${summaryLine(col.description)}${tagStr}`;
+}
+
+/** All verbs a resource exposes, generics first, in help order. */
+export function listVerbs(res: ResourceDef): string[] {
+  const verbs: string[] = GENERIC_VERBS.filter((v) => res.operations[v]);
+  if (res.customOperations) verbs.push(...Object.keys(res.customOperations));
+  return verbs;
+}
+
+/** Flags a generic verb accepts, derived from the resource's columns. */
+function genericFlags(res: ResourceDef, verb: GenericVerb): ColumnDef[] {
+  switch (verb) {
+    case 'create':
+      return res.columns.filter((c) => !c.generated);
+    case 'update':
+      return res.columns.filter((c) => c.updatable);
+    case 'list':
+      // Non-generated columns double as equality filters.
+      return [
+        ...res.columns.filter((c) => !c.generated).map((c) => ({ ...c, required: false })),
+        { name: 'limit', type: 'number', description: 'Max rows returned.', default: 200 } as ColumnDef,
+      ];
+    case 'get':
+    case 'delete':
+      return [];
+  }
+}
+
+function genericSummary(res: ResourceDef, verb: GenericVerb): string {
+  switch (verb) {
+    case 'list':
+      return `List ${res.plural}. Flags below act as equality filters.`;
+    case 'get':
+      return `Get a ${res.name} by ID.`;
+    case 'create':
+      return `Create a new ${res.name}.`;
+    case 'update':
+      return `Update a ${res.name} by ID. Provide at least one updatable flag.`;
+    case 'delete':
+      return `Delete a ${res.name} by ID.`;
+  }
+}
+
+/**
+ * Deep help for one verb: usage line, full description, flags, examples.
+ * `verb` is a custom-operation key or a generic CRUD verb. Returns undefined
+ * for a verb the resource doesn't have.
+ */
+export function renderVerbHelp(res: ResourceDef, verb: string): string | undefined {
+  const op: CustomOperation | undefined = res.customOperations?.[verb];
+  const generic = !op && (GENERIC_VERBS as readonly string[]).includes(verb) ? (verb as GenericVerb) : undefined;
+  if (!op && !generic) return undefined;
+  if (generic && !res.operations[generic]) return undefined;
+
+  const access = op ? op.access : res.operations[generic!];
+  const accessTag = access && access !== 'open' ? ` [${access}]` : '';
+  const needsId = generic === 'get' || generic === 'update' || generic === 'delete';
+
+  const lines: string[] = [];
+  lines.push(`ncl ${res.plural} ${verb}${needsId ? ' <id>' : ''}${accessTag}`);
+  lines.push('');
+  lines.push(op ? op.description : genericSummary(res, generic!));
+
+  const flags = op ? (op.args ?? []) : genericFlags(res, generic!);
+  if (flags.length > 0) {
+    lines.push('');
+    lines.push('Flags:');
+    for (const f of flags) lines.push(flagLine(f));
+  }
+  if (op?.examples?.length) {
+    lines.push('');
+    lines.push('Examples:');
+    for (const ex of op.examples) lines.push(indent(ex, '  '));
+  }
+  return lines.join('\n');
+}

src/cli/registry.ts

@@ -8,14 +8,30 @@
  * registers the help commands, so the registry is populated before the host's
  * CLI server accepts connections.
  */
+import { defineGuardedAction, type GuardedAction } from '../guard/index.js';
+import { commandGuardSpec } from './guard.js';
 import type { CallerContext } from './frame.js';
 
-export type Access = 'open' | 'approval';
+/**
+ * Resources an agent under `cli_scope=group` may touch. Single source —
+ * consumed by both dispatch enforcement and `ncl help` filtering, so the
+ * agent is never shown a resource the gate would reject (or vice versa).
+ */
+export const GROUP_SCOPE_RESOURCES = new Set(['groups', 'sessions', 'destinations', 'members']);
+
+export type Access = 'open' | 'approval' | 'hidden';
 
 export type CommandDef<TArgs = unknown, TData = unknown> = {
   name: string;
   description: string;
   access: Access;
+  /**
+   * Dotted guard-catalog action name (e.g. `roles.grant`,
+   * `groups.config.add-mcp-server`). Set by registerResource from the
+   * resource + verb; commands registered directly (help) fall back to
+   * `cli.<name>`.
+   */
+  action?: string;
   /**
    * The group-scope whitelist key. Under `cli_scope: 'group'` the dispatcher
    * only lets an agent run commands whose `resource` is on the whitelist
@@ -34,15 +50,37 @@ export type CommandDef<TArgs = unknown, TData = unknown> = {
   /** Validates `frame.args` and produces the typed handler input. Throws on invalid. */
   parseArgs: (raw: Record<string, unknown>) => TArgs;
   handler: (args: TArgs, ctx: CallerContext) => Promise<TData>;
+  /**
+   * Optional presentational renderer. When set, dispatch attaches its output
+   * as the response frame's `human` field (server-rendered once, printed
+   * verbatim by every client in human mode). Runs after post-handler scope
+   * filtering, so it only ever sees data the caller is allowed to see. A
+   * throwing formatter is ignored — clients fall back to rendering `data`.
+   */
+  formatHuman?: (data: TData) => string;
 };
 
 const registry = new Map<string, CommandDef>();
+const commandGuards = new Map<string, GuardedAction>();
 
 export function register<TArgs, TData>(def: CommandDef<TArgs, TData>): void {
   if (registry.has(def.name)) {
     throw new Error(`CLI command "${def.name}" already registered`);
   }
   registry.set(def.name, def as CommandDef);
+  // Declaration is registration: every command gets a guard-catalog entry
+  // derived from its own definition, in the same call that registers it — a
+  // command cannot exist without a guard, and dispatch consults it by value.
+  commandGuards.set(def.name, defineGuardedAction(commandGuardSpec(def as CommandDef)));
+}
+
+/** The guard defined for a registered command — total for anything register() accepted. */
+export function commandGuard(name: string): GuardedAction {
+  const g = commandGuards.get(name);
+  if (!g) {
+    throw new Error(`CLI command "${name}" has no guard — was it registered through register()?`);
+  }
+  return g;
 }
 
 export function lookup(name: string): CommandDef | undefined {

src/container-runner.ts

@@ -3,9 +3,10 @@
  * Spawns agent containers with session folder + agent group folder mounts.
  * The container runs the v2 agent-runner which polls the session DB.
  */
-import { ChildProcess, execSync, spawn } from 'child_process';
+import { ChildProcess, exec, spawn } from 'child_process';
 import fs from 'fs';
 import path from 'path';
+import { promisify } from 'util';
 
 import { OneCLI } from '@onecli-sh/sdk';
 
@@ -504,6 +505,8 @@ async function buildContainerArgs(
   return args;
 }
 
+const execAsync = promisify(exec);
+
 /** Build a per-agent-group Docker image with custom packages. */
 export async function buildAgentGroupImage(agentGroupId: string): Promise<void> {
   const agentGroup = getAgentGroup(agentGroupId);
@@ -539,9 +542,12 @@ export async function buildAgentGroupImage(agentGroupId: string): Promise<void>
   const tmpDockerfile = path.join(DATA_DIR, `Dockerfile.${agentGroupId}`);
   fs.writeFileSync(tmpDockerfile, dockerfile);
   try {
-    execSync(`${CONTAINER_RUNTIME_BIN} build -t ${imageTag} -f ${tmpDockerfile} .`, {
+    // Awaited async exec so the single-threaded host stays responsive during
+    // the build (can take minutes) instead of blocking on execSync. exec buffers
+    // stdout/stderr (matching the old stdio: 'pipe') and rejects on a non-zero
+    // exit, so error propagation is unchanged.
+    await execAsync(`${CONTAINER_RUNTIME_BIN} build -t ${imageTag} -f ${tmpDockerfile} .`, {
       cwd: DATA_DIR,
-      stdio: 'pipe',
       timeout: 900_000,
     });
   } finally {

src/db/session-db.ts

@@ -181,14 +181,6 @@ export function syncProcessingAcks(inDb: Database.Database, outDb: Database.Data
   })();
 }
 
-export function getStuckProcessingIds(outDb: Database.Database): string[] {
-  return (
-    outDb.prepare("SELECT message_id FROM processing_ack WHERE status = 'processing'").all() as Array<{
-      message_id: string;
-    }>
-  ).map((r) => r.message_id);
-}
-
 export interface ProcessingClaim {
   message_id: string;
   status_changed: string;

src/delivery-actions.test.ts

@@ -4,7 +4,9 @@
  * `registerDeliveryAction` is the hook modules use to handle system-kind
  * outbound messages; `getDeliveryAction` is the read side that makes those
  * registrations behavior-testable. Goes red if either half of the registry
- * is removed or the two stop sharing the same map.
+ * is removed or the two stop sharing the same map. Every registration now
+ * carries a guard spec or an explicit unguarded(<reason>) declaration —
+ * omission is a type error.
  */
 import { describe, it, expect, vi } from 'vitest';
 
@@ -16,11 +18,14 @@ vi.mock('./container-runner.js', () => ({
 }));
 
 import { registerDeliveryAction, getDeliveryAction, type DeliveryActionHandler } from './delivery.js';
+import { defineGuardedAction, HOLD, unguarded } from './guard/index.js';
+
+const testUnguarded = unguarded('test — registry mechanics only');
 
 describe('delivery action registry', () => {
   it('getDeliveryAction returns the handler registerDeliveryAction registered', () => {
     const handler: DeliveryActionHandler = async () => {};
-    registerDeliveryAction('test_registry_action', handler);
+    registerDeliveryAction('test_registry_action', handler, testUnguarded);
     expect(getDeliveryAction('test_registry_action')).toBe(handler);
   });
 
@@ -31,8 +36,34 @@ describe('delivery action registry', () => {
   it('re-registering an action overwrites the previous handler', () => {
     const first: DeliveryActionHandler = async () => {};
     const second: DeliveryActionHandler = async () => {};
-    registerDeliveryAction('test_overwrite_action', first);
-    registerDeliveryAction('test_overwrite_action', second);
+    registerDeliveryAction('test_overwrite_action', first, testUnguarded);
+    registerDeliveryAction('test_overwrite_action', second, testUnguarded);
     expect(getDeliveryAction('test_overwrite_action')).toBe(second);
   });
+
+  it('refuses to replace a guard-wrapped action with an unguarded handler', () => {
+    const guardAction = defineGuardedAction({
+      action: 'test.guarded-overwrite',
+      baseline: () => HOLD('t'),
+    });
+    registerDeliveryAction('test_guarded_overwrite', async () => {}, {
+      guardAction,
+      requestHold: async () => {},
+    });
+
+    // Disarming the guard by re-registering unguarded must throw — otherwise
+    // the action's catalog entry would still exist while the live path runs
+    // unguarded.
+    expect(() => registerDeliveryAction('test_guarded_overwrite', async () => {}, testUnguarded)).toThrow(
+      /disarm the guard/,
+    );
+
+    // Re-registering WITH a spec stays allowed (a legitimate replacement
+    // keeps the action guarded).
+    registerDeliveryAction('test_guarded_overwrite', async () => {}, {
+      guardAction,
+      requestHold: async () => {},
+    });
+    expect(getDeliveryAction('test_guarded_overwrite')).toBeDefined();
+  });
 });

src/delivery.ts

@@ -20,12 +20,13 @@ import {
   markDeliveryFailed,
   migrateDeliveredTable,
 } from './db/session-db.js';
+import { guard, type GuardedAction, type Unguarded } from './guard/index.js';
 import { log } from './log.js';
 import { normalizeOptions } from './channels/ask-question.js';
 import { clearOutbox, openInboundDb, openOutboundDb, readOutboxFiles } from './session-manager.js';
 import { pauseTypingRefreshAfterDelivery, setTypingAdapter } from './modules/typing/index.js';
 import type { OutboundFile } from './channels/adapter.js';
-import type { Session } from './types.js';
+import type { PendingApproval, Session } from './types.js';
 
 const ACTIVE_POLL_MS = 1000;
 const SWEEP_POLL_MS = 60_000;
@@ -393,14 +394,19 @@ async function deliverMessage(
  * Delivery action registry.
  *
  * Modules register handlers for system-kind outbound message actions via
- * `registerDeliveryAction`. Core checks the registry first in
- * `handleSystemAction` and falls through to the inline switch when no
- * handler is registered. The switch will shrink as modules are extracted
- * (scheduling, approvals, agent-to-agent) and eventually only its default
- * branch remains.
+ * `registerDeliveryAction`. Unknown actions log "Unknown system action".
  *
- * Default when no handler registered and the switch doesn't match: log
- * "Unknown system action" and return.
+ * Privileged delivery actions (create_agent, install_packages,
+ * add_mcp_server) register with a guard spec: every path to the handler body
+ * — dispatch, approved replay, test lookup — goes through the guard consult
+ * (allow / hold / deny), so there is no unguarded route to it. On approve,
+ * the continuation re-enters the same entry carrying the approval row as its
+ * grant (`reenterGuardedDeliveryAction`), so the structural baseline is
+ * re-checked live. Plain actions (scheduling self-actions, the cli_request
+ * bridge — its inner commands are guarded at dispatch) register with an
+ * explicit `unguarded(<reason>)` declaration instead of a spec — omission is
+ * not representable, so the decision to run unguarded is visible, and
+ * justified, at the registration site.
  */
 export type DeliveryActionHandler = (
   content: Record<string, unknown>,
@@ -408,18 +414,115 @@ export type DeliveryActionHandler = (
   inDb: Database.Database,
 ) => Promise<void>;
 
-const actionHandlers = new Map<string, DeliveryActionHandler>();
+/** Handler shape for guard-wrapped actions — must not touch inDb (replays run without one). */
+export type GuardedDeliveryHandler = (content: Record<string, unknown>, session: Session) => Promise<void>;
 
-export function registerDeliveryAction(action: string, handler: DeliveryActionHandler): void {
-  if (actionHandlers.has(action)) {
-    log.warn('Delivery action handler overwritten', { action });
-  }
-  actionHandlers.set(action, handler);
+export interface DeliveryGuardSpec {
+  /** Guard action consulted before the handler runs — the defined value, not a name. */
+  guardAction: GuardedAction;
+  /**
+   * Domain validation that runs before the guard — malformed requests are
+   * answered (notify) without ever creating a hold. Return false to stop.
+   */
+  precheck?: (content: Record<string, unknown>, session: Session) => boolean | Promise<boolean>;
+  /** Create the hold (the domain's requestApproval call — card text lives with the domain). */
+  requestHold: (content: Record<string, unknown>, session: Session) => Promise<void>;
+  /** Tell the requester about a deny. */
+  onDeny?: (content: Record<string, unknown>, session: Session, reason: string) => void;
 }
 
-/** Look up a registered delivery-action handler. Lets module registrations be behavior-tested. */
+type DeliveryEntry =
+  | { guard: Unguarded; handler: DeliveryActionHandler }
+  | { guard: DeliveryGuardSpec; handler: GuardedDeliveryHandler };
+
+const deliveryActions = new Map<string, DeliveryEntry>();
+
+function isUnguardedEntry(entry: DeliveryEntry): entry is Extract<DeliveryEntry, { guard: Unguarded }> {
+  return 'reason' in entry.guard;
+}
+
+export function registerDeliveryAction(action: string, handler: DeliveryActionHandler, unguardedDecl: Unguarded): void;
+export function registerDeliveryAction(action: string, handler: GuardedDeliveryHandler, spec: DeliveryGuardSpec): void;
+export function registerDeliveryAction(
+  action: string,
+  handler: DeliveryActionHandler | GuardedDeliveryHandler,
+  guardDecl: DeliveryGuardSpec | Unguarded,
+): void {
+  const existing = deliveryActions.get(action);
+  if (existing) {
+    // Replacing a guard-wrapped action with an unguarded handler would
+    // disarm the guard while its catalog entry still exists — refuse. A
+    // skill that wants to extend a guarded action must compose at the
+    // module's exported functions instead, or re-register with a guard spec
+    // of its own.
+    if ('reason' in guardDecl && !isUnguardedEntry(existing)) {
+      throw new Error(
+        `delivery action "${action}" is guard-wrapped; re-registering it without a guard spec would disarm the guard`,
+      );
+    }
+    log.warn('Delivery action handler overwritten', { action });
+  }
+  // The overloads pair each handler shape with its declaration; the merged
+  // implementation signature erases that pairing, hence the one cast.
+  deliveryActions.set(action, { guard: guardDecl, handler } as DeliveryEntry);
+}
+
+async function runGuarded(
+  action: string,
+  entry: Extract<DeliveryEntry, { guard: DeliveryGuardSpec }>,
+  content: Record<string, unknown>,
+  session: Session,
+  grant: PendingApproval | null,
+): Promise<void> {
+  const spec = entry.guard;
+  if (spec.precheck && !(await spec.precheck(content, session))) return;
+
+  const decision = guard(spec.guardAction, {
+    actor: { kind: 'agent', agentGroupId: session.agent_group_id, sessionId: session.id },
+    payload: content,
+    grant,
+  });
+
+  if (decision.effect === 'deny') {
+    log.warn('Delivery action denied by guard', { action, reason: decision.reason });
+    spec.onDeny?.(content, session, decision.reason);
+    return;
+  }
+  if (decision.effect === 'hold') {
+    await spec.requestHold(content, session);
+    return;
+  }
+  await entry.handler(content, session);
+}
+
+/**
+ * Approve continuation for a guard-wrapped delivery action: re-enter the
+ * entry with the approval row as the grant. The guard treats the grant as
+ * hold-satisfied but re-runs the structural baseline, so approve-then-revoke
+ * does not execute. Domains register this as their approval handler in the
+ * same line that registers the action.
+ */
+export function reenterGuardedDeliveryAction(action: string) {
+  return async (ctx: { session: Session; payload: Record<string, unknown>; approval: PendingApproval }) => {
+    const entry = deliveryActions.get(action);
+    if (!entry || isUnguardedEntry(entry)) {
+      log.warn('Approved replay for an action that is not guard-wrapped — dropping', { action });
+      return;
+    }
+    await runGuarded(action, entry, ctx.payload, ctx.session, ctx.approval);
+  };
+}
+
+/**
+ * The invocable for a registered action — the raw handler for unguarded
+ * entries, the guard-consulting path for guarded ones. Dispatch and tests
+ * both come through here; there is no route around the guard.
+ */
 export function getDeliveryAction(action: string): DeliveryActionHandler | undefined {
-  return actionHandlers.get(action);
+  const entry = deliveryActions.get(action);
+  if (!entry) return undefined;
+  if (isUnguardedEntry(entry)) return entry.handler;
+  return (content, session) => runGuarded(action, entry, content, session, null);
 }
 
 /**
@@ -435,7 +538,7 @@ async function handleSystemAction(
   const action = content.action as string;
   log.info('System action from agent', { sessionId: session.id, action });
 
-  const registered = actionHandlers.get(action);
+  const registered = getDeliveryAction(action);
   if (registered) {
     await registered(content, session, inDb);
     return;

src/guard-conformance.ts

@@ -0,0 +1,70 @@
+/**
+ * Boot-time guard sanity — the one cross-registry invariant left to check
+ * after all import-time registrations have run.
+ *
+ * The old registry walk is gone: everything it detected is now
+ * unconstructible at the API level.
+ *   - A consult site cannot name a missing catalog entry — guard() takes the
+ *     GuardedAction VALUE returned by defineGuardedAction, so a dropped
+ *     module-edge import or a typo'd action is a compile error (and a forged
+ *     value is denied at runtime), not a silent allow.
+ *   - A handler cannot register unguarded by omission — every registry
+ *     (delivery actions, response handlers, interceptors, CLI commands)
+ *     requires a guard spec or an explicit unguarded(<reason>) declaration
+ *     at the registration site.
+ *
+ * What remains is completeness ACROSS registries: a guarded action that
+ * holds via `approvalAction` needs a registered approval handler, or an
+ * approved card resolves into nothing — the hold has no continuation. That
+ * pairing only exists once every module has loaded (catalog entries and
+ * approval handlers register from different modules), so it stays a boot
+ * check with the fail-closed posture: the host refuses to start, surfacing
+ * the mis-composition at skill-install time instead of at the first
+ * approved card.
+ */
+import { listGuardedActions } from './guard/index.js';
+import { log } from './log.js';
+import { getApprovalHandler } from './modules/approvals/primitive.js';
+
+/** Holding actions with no approve continuation. Empty = conformant. */
+export function grantContinuationGaps(): string[] {
+  return listGuardedActions()
+    .filter((spec) => spec.approvalAction && !getApprovalHandler(spec.approvalAction))
+    .map(
+      (spec) =>
+        `guarded action "${spec.action}" holds via approval action "${spec.approvalAction}" ` +
+        'but no approval handler is registered — an approved hold would have no continuation',
+    );
+}
+
+/**
+ * Boot check: refuse to start when a holding action has no continuation.
+ * Call after all import-time registrations (any point in main()).
+ */
+export function enforceGuardConformance(): void {
+  const gaps = grantContinuationGaps();
+  if (gaps.length === 0) return;
+
+  console.error(
+    [
+      '',
+      '='.repeat(64),
+      'NanoClaw stopped: guard conformance failure',
+      '='.repeat(64),
+      'A guarded action can hold for approval, but no approval handler is',
+      'registered for its approval action — an admin could click Approve',
+      'and nothing would execute. This usually means a module (or skill)',
+      'defined a holding baseline without registering its continuation.',
+      '',
+      ...gaps.map((g) => `  - ${g}`),
+      '',
+      'Register the approval handler (registerApprovalHandler) in the same',
+      'module that defines the guarded action, or drop approvalAction from',
+      'the definition if the action can never hold.',
+      '='.repeat(64),
+      '',
+    ].join('\n'),
+  );
+  log.error('Guard conformance failure — refusing to start', { gaps });
+  process.exit(1);
+}

src/guard/conformance.test.ts

@@ -0,0 +1,84 @@
+/**
+ * Guard conformance — the boot invariant, checked with the real registries.
+ *
+ * The old registry walk is gone: an unmapped consult or an undeclared
+ * unguarded registration is now unconstructible — guard() takes the defined
+ * GuardedAction value (a dropped module-edge import or typo'd name is a
+ * compile error), and every registry requires a guard spec or an explicit
+ * unguarded(<reason>) declaration. What's left to verify structurally is the
+ * cross-registry pairing the compiler can't see: every holding action has a
+ * registered approve continuation. The check runs here in CI and at every
+ * boot (enforceGuardConformance refuses to start) — CI can't see
+ * skill-installed registrations, the boot check can.
+ */
+import { describe, expect, it } from 'vitest';
+
+// Production barrels — side-effect imports populate the real registries.
+import '../cli/commands/index.js';
+import '../modules/index.js';
+import '../cli/delivery-action.js';
+import '../cli/dispatch.js'; // registers the cli_command approval handler
+
+import { commandGuard, listCommands } from '../cli/registry.js';
+import { grantContinuationGaps } from '../guard-conformance.js';
+import { getApprovalHandler } from '../modules/approvals/primitive.js';
+import { defineGuardedAction, listGuardedActions } from './guard-actions.js';
+import { HOLD } from './types.js';
+
+describe('guard conformance', () => {
+  it('the grant-continuation check (shared with the boot check) reports zero gaps', () => {
+    expect(grantContinuationGaps()).toEqual([]);
+  });
+
+  it('every holding action pairs with a registered approval handler', () => {
+    const holding = listGuardedActions().filter((spec) => spec.approvalAction);
+    expect(holding.length).toBeGreaterThan(0);
+
+    const dangling = holding.filter((spec) => !getApprovalHandler(spec.approvalAction as string));
+    expect(dangling.map((s) => s.action)).toEqual([]);
+  });
+
+  it('every mutating ncl command derives a guard that holds via cli_command', () => {
+    const mutating = listCommands().filter((cmd) => cmd.access === 'approval');
+    expect(mutating.length).toBeGreaterThan(0);
+
+    const wrong = mutating.filter((cmd) => commandGuard(cmd.name).approvalAction !== 'cli_command');
+    expect(wrong.map((c) => c.name)).toEqual([]);
+  });
+
+  it('the domain catalog entries are defined once the module barrels load', () => {
+    const actions = new Set(listGuardedActions().map((s) => s.action));
+    for (const expected of [
+      'agents.create',
+      'a2a.send',
+      'self_mod.install_packages',
+      'self_mod.add_mcp_server',
+      'senders.admit',
+      'channels.register',
+    ]) {
+      expect(actions.has(expected), `catalog is missing "${expected}"`).toBe(true);
+    }
+  });
+
+  it('defining the same action twice throws — names are the catalog key', () => {
+    defineGuardedAction({ action: 'test.dup-define', baseline: () => HOLD('x') });
+    expect(() => defineGuardedAction({ action: 'test.dup-define', baseline: () => HOLD('x') })).toThrow(
+      /already defined/,
+    );
+  });
+
+  // KEEP LAST: defines a holding action with no continuation into the shared
+  // per-worker catalog, so every gap check after this point sees it.
+  it('the check names a holding action with no approve continuation (what boot refuses on)', () => {
+    defineGuardedAction({
+      action: 'test.dangling-hold',
+      approvalAction: 'test_dangling_hold_approved',
+      baseline: () => HOLD('always'),
+    });
+
+    const gaps = grantContinuationGaps();
+    expect(gaps).toHaveLength(1);
+    expect(gaps[0]).toContain('test.dangling-hold');
+    expect(gaps[0]).toContain('no approval handler');
+  });
+});

src/guard/guard-actions.ts

@@ -0,0 +1,73 @@
+/**
+ * The action catalog — the enforcement boundary.
+ *
+ * An action either is defined here (and every consult passes its decision)
+ * or cannot be consulted at all: guard() takes the GuardedAction VALUE
+ * returned by defineGuardedAction, so the wiring between a consult site and
+ * its baseline is a symbol reference the compiler checks. A dropped
+ * module-edge import or a typo'd action name is a build error, not a
+ * runtime fail-open — there is no lookup that can miss.
+ *
+ * Definitions are still recorded by name so boot can enumerate them: the
+ * grant-continuation check (src/guard-conformance.ts) pairs every holding
+ * action with its registered approval handler, and duplicate names are
+ * refused at definition time (grants match on the name).
+ */
+import type { GuardDecision, GuardInput } from './types.js';
+import type { PendingApproval } from '../types.js';
+
+export interface GuardedActionSpec {
+  /** Dotted action name, e.g. 'roles.grant', 'agents.create', 'a2a.send'. */
+  action: string;
+  /**
+   * Today's structural checks for this action, verbatim — the only source of
+   * allow. Runs on every consult, including approved replays (a grant
+   * satisfies a hold, never a deny).
+   */
+  baseline: (input: GuardInput) => GuardDecision;
+  /**
+   * The pending_approvals.action its holds resolve through — a grant is only
+   * accepted when its row carries this action. Omit for actions that can
+   * never be held (deny/allow-only baselines).
+   */
+  approvalAction?: string;
+  /**
+   * Extra domain binding between a grant and the replayed input (e.g. the
+   * a2a target must match the held message). Runs in addition to the
+   * approvalAction + live-row checks.
+   */
+  grantMatches?: (grant: PendingApproval, input: GuardInput) => boolean;
+}
+
+declare const guardedActionBrand: unique symbol;
+/**
+ * A defined guarded action — only defineGuardedAction can mint one. The
+ * brand makes the type nominal: a hand-rolled { action, baseline } object
+ * does not typecheck at a consult site, and fails the runtime check too.
+ */
+export type GuardedAction = Readonly<GuardedActionSpec> & { readonly [guardedActionBrand]: true };
+
+const defined = new Map<string, GuardedAction>();
+const minted = new WeakSet<object>();
+
+export function defineGuardedAction(spec: GuardedActionSpec): GuardedAction {
+  if (defined.has(spec.action)) {
+    throw new Error(`guarded action "${spec.action}" is already defined — action names are the catalog key`);
+  }
+  const def = Object.freeze({ ...spec }) as GuardedAction;
+  minted.add(def);
+  defined.set(spec.action, def);
+  return def;
+}
+
+/**
+ * Runtime backstop for callers outside the type system (plain JS, casts):
+ * only values minted by defineGuardedAction pass — guard() denies the rest.
+ */
+export function isGuardedAction(value: unknown): value is GuardedAction {
+  return typeof value === 'object' && value !== null && minted.has(value);
+}
+
+export function listGuardedActions(): GuardedAction[] {
+  return [...defined.values()].sort((a, b) => a.action.localeCompare(b.action));
+}

src/guard/guard.test.ts

@@ -0,0 +1,160 @@
+/**
+ * Guard decision-function unit tests: the baseline is the decision (allow /
+ * hold / deny returned as-is), grant semantics (satisfies holds, never
+ * denies; invalid → refuse), the runtime backstop against forged action
+ * values, and the fail-closed posture on a throwing baseline.
+ *
+ * Uses synthetic actions defined per test — the catalog is per-worker module
+ * state with no reset, so action names are unique.
+ */
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { guard } from './guard.js';
+import { defineGuardedAction, type GuardedAction } from './guard-actions.js';
+import { ALLOW, DENY, HOLD, type GuardInput } from './types.js';
+
+const mockGetPendingApproval = vi.fn();
+vi.mock('../db/sessions.js', () => ({
+  getPendingApproval: (...args: unknown[]) => mockGetPendingApproval(...args),
+}));
+vi.mock('../log.js', () => ({
+  log: { info: vi.fn(), warn: vi.fn(), error: vi.fn(), debug: vi.fn() },
+}));
+
+const AGENT = { kind: 'agent', agentGroupId: 'ag-1', sessionId: 'sess-1' } as const;
+
+function input(extra: Partial<GuardInput> = {}): GuardInput {
+  return { actor: AGENT, payload: {}, ...extra };
+}
+
+beforeEach(() => {
+  mockGetPendingApproval.mockReset();
+});
+
+afterEach(() => {
+  vi.clearAllMocks();
+});
+
+describe('the baseline is the decision', () => {
+  it('baseline allow → allow', () => {
+    const action = defineGuardedAction({ action: 't.allow1', baseline: () => ALLOW('ok') });
+    expect(guard(action, input()).effect).toBe('allow');
+  });
+
+  it('baseline hold → hold, default approver chain', () => {
+    const action = defineGuardedAction({ action: 't.hold1', baseline: () => HOLD('needs approval') });
+    const d = guard(action, input());
+    expect(d.effect).toBe('hold');
+    if (d.effect === 'hold') {
+      expect(d.reason).toBe('needs approval');
+      expect(d.approverUserId).toBeUndefined();
+    }
+  });
+
+  it('baseline hold → hold, carrying a named approver', () => {
+    const action = defineGuardedAction({ action: 't.hold2', baseline: () => HOLD('policy row', 'telegram:dana') });
+    const d = guard(action, input());
+    expect(d.effect).toBe('hold');
+    if (d.effect === 'hold') expect(d.approverUserId).toBe('telegram:dana');
+  });
+
+  it('baseline deny → deny, carrying the reason', () => {
+    const action = defineGuardedAction({ action: 't.deny1', baseline: () => DENY('structurally unauthorized') });
+    const d = guard(action, input());
+    expect(d.effect).toBe('deny');
+    if (d.effect === 'deny') expect(d.reason).toBe('structurally unauthorized');
+  });
+
+  it('a forged action value (not from defineGuardedAction) is denied', () => {
+    const forged = { action: 't.forged', baseline: () => ALLOW('never vetted') } as unknown as GuardedAction;
+    const d = guard(forged, input());
+    expect(d.effect).toBe('deny');
+    if (d.effect === 'deny') expect(d.reason).toContain('undefined action');
+  });
+});
+
+describe('grants', () => {
+  const grantRow = (action: string) =>
+    ({ approval_id: 'appr-1', action, payload: '{}' }) as unknown as NonNullable<GuardInput['grant']>;
+
+  it('a valid live grant satisfies a hold', () => {
+    const action = defineGuardedAction({
+      action: 't.g1',
+      approvalAction: 'g1_approved',
+      baseline: () => HOLD('b'),
+    });
+    const grant = grantRow('g1_approved');
+    mockGetPendingApproval.mockReturnValue(grant);
+    expect(guard(action, input({ grant })).effect).toBe('allow');
+  });
+
+  it('a grant never satisfies a deny — the baseline is re-checked live', () => {
+    const action = defineGuardedAction({
+      action: 't.g2',
+      approvalAction: 'g2_approved',
+      baseline: () => DENY('revoked since'),
+    });
+    const grant = grantRow('g2_approved');
+    mockGetPendingApproval.mockReturnValue(grant);
+    const d = guard(action, input({ grant }));
+    expect(d.effect).toBe('deny');
+    if (d.effect === 'deny') expect(d.reason).toBe('revoked since');
+  });
+
+  it('a dead grant (row deleted) refuses instead of re-holding', () => {
+    const action = defineGuardedAction({
+      action: 't.g3',
+      approvalAction: 'g3_approved',
+      baseline: () => HOLD('b'),
+    });
+    mockGetPendingApproval.mockReturnValue(undefined);
+    const d = guard(action, input({ grant: grantRow('g3_approved') }));
+    expect(d.effect).toBe('deny');
+  });
+
+  it("a grant for a different action doesn't transfer", () => {
+    const action = defineGuardedAction({
+      action: 't.g4',
+      approvalAction: 'g4_approved',
+      baseline: () => HOLD('b'),
+    });
+    const grant = grantRow('other_action');
+    mockGetPendingApproval.mockReturnValue(grant);
+    expect(guard(action, input({ grant })).effect).toBe('deny');
+  });
+
+  it('a domain grantMatches binding can refuse a payload mismatch', () => {
+    const action = defineGuardedAction({
+      action: 't.g5',
+      approvalAction: 'g5_approved',
+      grantMatches: () => false,
+      baseline: () => HOLD('b'),
+    });
+    const grant = grantRow('g5_approved');
+    mockGetPendingApproval.mockReturnValue(grant);
+    expect(guard(action, input({ grant })).effect).toBe('deny');
+  });
+
+  it('a grant on an already-allowed action is a no-op', () => {
+    const action = defineGuardedAction({
+      action: 't.g6',
+      approvalAction: 'g6_approved',
+      baseline: () => ALLOW('ok'),
+    });
+    const grant = grantRow('g6_approved');
+    mockGetPendingApproval.mockReturnValue(grant);
+    expect(guard(action, input({ grant })).effect).toBe('allow');
+  });
+});
+
+describe('fail-closed posture', () => {
+  it('a throwing baseline denies', () => {
+    const action = defineGuardedAction({
+      action: 't.f1',
+      baseline: () => {
+        throw new Error('boom');
+      },
+    });
+    expect(guard(action, input()).effect).toBe('deny');
+  });
+});

src/guard/guard.ts

@@ -0,0 +1,72 @@
+/**
+ * guard() — the one decision function every privileged action consults.
+ *
+ * The decision is the action's structural baseline — today's code checks,
+ * defined per action at the module edges. The consult site holds the
+ * GuardedAction value itself (defineGuardedAction), so there is no name
+ * lookup and no fail-open path for an unknown action: an unwired consult is
+ * a compile error, and a value that didn't come from defineGuardedAction is
+ * denied at runtime. Policy-as-data (tighten-only rule sources composing
+ * with the baseline) is deliberately deferred to phase 3 of the
+ * guarded-actions design, where the generalized rules table arrives with its
+ * first operator-visible consumer; until then the one policy table
+ * (agent_message_policies) is consulted inside the a2a.send baseline.
+ *
+ * Grants: an approved replay carries the verified approval row. A valid
+ * grant (live pending row whose action matches the entry's approval action,
+ * plus any domain binding) satisfies a hold — the human already decided —
+ * but NEVER a deny: the baseline is re-checked live, so approve-then-revoke
+ * no longer executes. A grant that is present but invalid fails closed to
+ * deny (no second card).
+ *
+ * The guard itself fails closed: a throwing baseline denies.
+ */
+import { getPendingApproval } from '../db/sessions.js';
+import { log } from '../log.js';
+import { isGuardedAction, type GuardedAction } from './guard-actions.js';
+import { ALLOW, DENY, type GuardDecision, type GuardInput } from './types.js';
+
+export function guard(action: GuardedAction, input: GuardInput): GuardDecision {
+  if (!isGuardedAction(action)) {
+    // JS-level backstop — the branded type already forbids this. A
+    // hand-rolled object must not carry a baseline never vetted at
+    // definition time.
+    log.error('Guard consulted with an undefined action — failing closed', {
+      action: (action as { action?: unknown } | null)?.action,
+    });
+    return DENY('guard consulted with an undefined action (failing closed)');
+  }
+
+  let decision: GuardDecision;
+  try {
+    decision = action.baseline(input);
+  } catch (err) {
+    log.error('Guard evaluation threw — failing closed', { action: action.action, err });
+    return DENY('guard failure (failing closed)');
+  }
+
+  if (!input.grant || decision.effect !== 'hold') {
+    // A grant never loosens a deny (the baseline re-check is live), and a
+    // grant on an already-allowed action is a no-op.
+    return decision;
+  }
+
+  // An invalid grant on a replay is a refusal, not a fresh hold — approved
+  // replays must execute exactly once.
+  if (grantSatisfies(action, input)) {
+    return ALLOW(`hold satisfied by approval ${input.grant.approval_id}`);
+  }
+  return DENY('replay carried an invalid or mismatched grant');
+}
+
+function grantSatisfies(action: GuardedAction, input: GuardInput): boolean {
+  const grant = input.grant;
+  if (!grant || !action.approvalAction) return false;
+  if (grant.action !== action.approvalAction) return false;
+  // The row must still be live — resolution deletes it, so a grant can only
+  // execute once and a fabricated row object doesn't pass.
+  const live = getPendingApproval(grant.approval_id);
+  if (!live || live.action !== action.approvalAction) return false;
+  if (action.grantMatches && !action.grantMatches(grant, input)) return false;
+  return true;
+}

src/guard/index.ts

@@ -0,0 +1,28 @@
+/**
+ * Guard — the privileged-action decision seam (guarded-actions phase 2).
+ *
+ * See the guarded-actions decisions doc on the team hub. One decision
+ * function (guard.ts) and a definition-derived action catalog
+ * (guard-actions.ts). Consults carry the GuardedAction value returned by
+ * defineGuardedAction — never a name to look up — so mis-wiring is a build
+ * error, not a runtime fail-open.
+ * Domain-free leaf: domain baselines are defined at the domain modules' edges.
+ */
+export { guard } from './guard.js';
+export {
+  defineGuardedAction,
+  isGuardedAction,
+  listGuardedActions,
+  type GuardedAction,
+  type GuardedActionSpec,
+} from './guard-actions.js';
+export {
+  ALLOW,
+  DENY,
+  HOLD,
+  unguarded,
+  type GuardActor,
+  type GuardDecision,
+  type GuardInput,
+  type Unguarded,
+} from './types.js';

src/guard/types.ts

@@ -0,0 +1,64 @@
+/**
+ * Guard vocabulary — the decision seam every privileged action passes.
+ *
+ * The guard is a domain-free leaf: this module may import the DB read layer,
+ * config, log, and shared types — never src/cli/* or src/modules/*. Domain
+ * knowledge (what an action's structural baseline checks) arrives via
+ * definition: domain modules call defineGuardedAction (guard-actions.ts) at
+ * their module edges and pass the returned value to every consult and
+ * registration site — the wiring is a symbol reference the compiler checks.
+ */
+import type { PendingApproval } from '../types.js';
+
+/** Who is attempting the action. Mirrors the CLI CallerContext + click identities. */
+export type GuardActor =
+  | { kind: 'host' }
+  | { kind: 'agent'; agentGroupId: string; sessionId?: string }
+  | { kind: 'human'; userId: string }
+  | { kind: 'system' };
+
+export interface GuardInput {
+  actor: GuardActor;
+  /** Domain resource reference, e.g. { from, to } for a2a.send. */
+  resource?: Record<string, string>;
+  /** Action arguments — what the card summarizes and rules may later match on. */
+  payload: Record<string, unknown>;
+  /**
+   * Verified approval row carried by an approved replay. A valid grant
+   * satisfies a hold (the human already decided) but never a deny — the
+   * structural baseline is re-checked live on every replay.
+   */
+  grant?: PendingApproval | null;
+}
+
+declare const unguardedBrand: unique symbol;
+/**
+ * A registration that deliberately carries no guard. Omission is not
+ * representable — every registry requires either a guard spec or this
+ * marker, so the decision to run unguarded is visible, and justified, in
+ * the diff that registers the handler. The reason travels with the
+ * registration; `grep "unguarded("` is the complete inventory.
+ */
+export type Unguarded = { readonly reason: string; readonly [unguardedBrand]: true };
+
+export function unguarded(reason: string): Unguarded {
+  return Object.freeze({ reason }) as Unguarded;
+}
+
+export type GuardDecision =
+  | { effect: 'allow'; reason: string }
+  | { effect: 'hold'; reason: string; approverUserId?: string }
+  | { effect: 'deny'; reason: string };
+
+export const ALLOW = (reason: string): GuardDecision => ({ effect: 'allow', reason });
+export const DENY = (reason: string): GuardDecision => ({ effect: 'deny', reason });
+/**
+ * approverUserId names an exclusive approver for the hold (the a2a policy
+ * row's named approver). Absent, the hold goes to the approvals primitive's
+ * default chain (scoped admins → global admins → owners).
+ */
+export const HOLD = (reason: string, approverUserId?: string): GuardDecision => ({
+  effect: 'hold',
+  reason,
+  approverUserId,
+});

src/index.ts

@@ -14,6 +14,7 @@ import { initDb } from './db/connection.js';
 import { runMigrations } from './db/migrations/index.js';
 import { ensureContainerRuntimeRunning, cleanupOrphans } from './container-runtime.js';
 import { startActiveDeliveryPoll, startSweepDeliveryPoll, setDeliveryAdapter, stopDeliveryPolls } from './delivery.js';
+import { enforceGuardConformance } from './guard-conformance.js';
 import { startHostSweep, stopHostSweep } from './host-sweep.js';
 import { routeInbound } from './router.js';
 import { log } from './log.js';
@@ -24,16 +25,7 @@ import { enforceUpgradeTripwire } from './upgrade-state.js';
 // effects, and the modules call registerResponseHandler/onShutdown at top
 // level — which would hit a TDZ error if the arrays lived here. Re-exported
 // here so existing callers see the same surface.
-import {
-  registerResponseHandler,
-  getResponseHandlers,
-  onShutdown,
-  getShutdownCallbacks,
-  type ResponsePayload,
-  type ResponseHandler,
-} from './response-registry.js';
-export { registerResponseHandler, onShutdown };
-export type { ResponsePayload, ResponseHandler };
+import { getResponseHandlers, getShutdownCallbacks, type ResponsePayload } from './response-registry.js';
 
 async function dispatchResponse(payload: ResponsePayload): Promise<void> {
   for (const handler of getResponseHandlers()) {
@@ -78,6 +70,12 @@ async function main(): Promise<void> {
   // outside the sanctioned path (raw `git pull` instead of /update-nanoclaw).
   enforceUpgradeTripwire();
 
+  // 0.6 Guard conformance — every import-time registration has already run;
+  // refuse to start if any privileged command / delivery action is unmapped
+  // (CI can't see skill-installed code — this makes the invariant hold at
+  // the boundary where third-party registrations enter).
+  enforceGuardConformance();
+
   // 1. Init central DB
   const dbPath = path.join(DATA_DIR, 'v2.db');
   const db = initDb(dbPath);

src/modules/agent-to-agent/agent-route.ts

@@ -27,14 +27,15 @@ import { getAgentGroup } from '../../db/agent-groups.js';
 import { getInboundSourceSessionId, getMostRecentPeerSourceSessionId } from '../../db/session-db.js';
 import { getSession } from '../../db/sessions.js';
 import { wakeContainer } from '../../container-runner.js';
+import { guard } from '../../guard/index.js';
 import { log } from '../../log.js';
 import { openInboundDb, resolveSession, sessionDir, writeSessionMessage } from '../../session-manager.js';
-import type { Session } from '../../types.js';
+import type { PendingApproval, Session } from '../../types.js';
 import { requestApproval } from '../approvals/index.js';
-import { hasDestination } from './db/agent-destinations.js';
-import { getMessagePolicy } from './db/agent-message-policies.js';
+import { A2A_MESSAGE_GATE_ACTION, a2aSend } from './guard.js';
 
 export { isSafeAttachmentName };
+export { A2A_MESSAGE_GATE_ACTION } from './guard.js';
 
 export interface ForwardedAttachment {
   name: string;
@@ -230,56 +231,64 @@ function resolveTargetSession(msg: RoutableAgentMessage, sourceSession: Session,
   return resolveSession(targetAgentGroupId, null, null, 'agent-shared').session;
 }
 
-export async function routeAgentMessage(msg: RoutableAgentMessage, session: Session): Promise<void> {
+export async function routeAgentMessage(
+  msg: RoutableAgentMessage,
+  session: Session,
+  opts: { grant?: PendingApproval } = {},
+): Promise<void> {
   const sourceAgentGroupId = session.agent_group_id;
   const targetAgentGroupId = msg.platform_id;
   if (!targetAgentGroupId) {
     throw new Error(`agent-to-agent message ${msg.id} is missing a target agent group id`);
   }
-  const isSelf = targetAgentGroupId === sourceAgentGroupId;
-  if (!isSelf && !hasDestination(sourceAgentGroupId, 'agent', targetAgentGroupId)) {
-    throw new Error(`unauthorized agent-to-agent: ${sourceAgentGroupId} has no destination for ${targetAgentGroupId}`);
-  }
-  if (!getAgentGroup(targetAgentGroupId)) {
-    throw new Error(`target agent group ${targetAgentGroupId} not found for message ${msg.id}`);
+
+  // The a2a.send baseline (guard.ts) carries the checks verbatim in their
+  // original order: destination ACL deny, target-exists deny, self-send
+  // allow, agent_message_policies hold. An approved replay carries the
+  // grant — the hold is satisfied but the structure is re-checked live, so
+  // revoking a destination between hold and approve blocks delivery.
+  const decision = guard(a2aSend, {
+    actor: { kind: 'agent', agentGroupId: sourceAgentGroupId, sessionId: session.id },
+    resource: { from: sourceAgentGroupId, to: targetAgentGroupId },
+    payload: { id: msg.id, platform_id: targetAgentGroupId, content: msg.content, in_reply_to: msg.in_reply_to },
+    grant: opts.grant ?? null,
+  });
+
+  if (decision.effect === 'deny') {
+    throw new Error(decision.reason);
   }
 
   // Gated edge: hold the message and return (not throw) so the delivery loop
-  // consumes the outbound row; `applyA2aMessageGate` re-routes it on approve.
-  if (!isSelf) {
-    const policy = getMessagePolicy(sourceAgentGroupId, targetAgentGroupId);
-    if (policy) {
-      const { approver } = policy;
-      const sourceName = getAgentGroup(sourceAgentGroupId)?.name ?? sourceAgentGroupId;
-      const targetName = getAgentGroup(targetAgentGroupId)?.name ?? targetAgentGroupId;
-      await requestApproval({
-        session,
-        agentName: sourceName,
-        action: A2A_MESSAGE_GATE_ACTION,
-        approverUserId: approver,
-        title: 'Message approval',
-        question: buildGateQuestion(sourceName, targetName, msg.content),
-        payload: {
-          id: msg.id,
-          platform_id: targetAgentGroupId,
-          content: msg.content,
-          in_reply_to: msg.in_reply_to,
-        },
-      });
-      log.info('Agent message held for approval', {
-        from: sourceAgentGroupId,
-        to: targetAgentGroupId,
-        msgId: msg.id,
-      });
-      return;
-    }
+  // consumes the outbound row; `applyA2aMessageGate` re-enters here with the
+  // grant on approve.
+  if (decision.effect === 'hold') {
+    const sourceName = getAgentGroup(sourceAgentGroupId)?.name ?? sourceAgentGroupId;
+    const targetName = getAgentGroup(targetAgentGroupId)?.name ?? targetAgentGroupId;
+    await requestApproval({
+      session,
+      agentName: sourceName,
+      action: A2A_MESSAGE_GATE_ACTION,
+      approverUserId: decision.approverUserId,
+      title: 'Message approval',
+      question: buildGateQuestion(sourceName, targetName, msg.content),
+      payload: {
+        id: msg.id,
+        platform_id: targetAgentGroupId,
+        content: msg.content,
+        in_reply_to: msg.in_reply_to,
+      },
+    });
+    log.info('Agent message held for approval', {
+      from: sourceAgentGroupId,
+      to: targetAgentGroupId,
+      msgId: msg.id,
+    });
+    return;
   }
 
   await performAgentRoute(msg, session, targetAgentGroupId);
 }
 
-export const A2A_MESSAGE_GATE_ACTION = 'a2a_message_gate';
-
 const GATE_CARD_BODY_MAX = 1500;
 
 function parseMessageContent(contentStr: string): { text: string; files: string[] } {
@@ -308,9 +317,11 @@ function buildGateQuestion(sourceName: string, targetName: string, contentStr: s
 
 /**
  * Cross-session route: pick the target session, forward files, write to its
- * inbound DB, wake it. Authorization is the caller's responsibility.
+ * inbound DB, wake it. Module-private — the only door is routeAgentMessage's
+ * guard decision (the approve continuation re-enters with a grant rather
+ * than calling this directly).
  */
-export async function performAgentRoute(
+async function performAgentRoute(
   msg: RoutableAgentMessage,
   session: Session,
   targetAgentGroupId: string,

src/modules/agent-to-agent/create-agent.test.ts

@@ -2,26 +2,48 @@
  * Tests for create_agent host-side authorization.
  *
  * Regression guard for the audit finding: `create_agent` is a privileged
- * central-DB write with no host-side authz. The fix authorizes by CLI scope —
- * trusted owner agent groups ('global') create directly; confined groups
- * ('group', the default and the prompt-injection victim) must get admin
- * approval. These tests pin that branch decision.
+ * central-DB write with no host-side authz. Authorization is the guard's
+ * `agents.create` baseline — trusted owner agent groups ('global') create
+ * directly; confined groups ('group', the default and the prompt-injection
+ * victim) hold for admin approval. These tests drive the REAL wrapped
+ * delivery action (the only reachable path) and the approve continuation's
+ * grant-carrying re-entry.
  */
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 
-import type { Session } from '../../types.js';
+import type { PendingApproval, Session } from '../../types.js';
 
 // Mocks for the collaborators the branch decides between / depends on.
-const mockRequestApproval = vi.fn().mockResolvedValue(undefined);
-const mockGetContainerConfig = vi.fn();
-const mockCreateAgentGroup = vi.fn();
-const mockInitGroupFilesystem = vi.fn();
-const mockUpdateScalars = vi.fn();
-const mockWriteDestinations = vi.fn();
-const mockNotifyWrite = vi.fn();
+// vi.hoisted: the module barrel import below runs before this file's const
+// initializers, and the mock factories close over this state.
+const {
+  mockRequestApproval,
+  mockGetContainerConfig,
+  mockCreateAgentGroup,
+  mockInitGroupFilesystem,
+  mockUpdateScalars,
+  mockWriteDestinations,
+  mockNotifyWrite,
+  liveApprovals,
+  approvalHandlers,
+} = vi.hoisted(() => ({
+  mockRequestApproval: vi.fn().mockResolvedValue(undefined),
+  mockGetContainerConfig: vi.fn(),
+  mockCreateAgentGroup: vi.fn(),
+  mockInitGroupFilesystem: vi.fn(),
+  mockUpdateScalars: vi.fn(),
+  mockWriteDestinations: vi.fn(),
+  mockNotifyWrite: vi.fn(),
+  liveApprovals: new Map<string, import('../../types.js').PendingApproval>(),
+  approvalHandlers: new Map<string, (ctx: Record<string, unknown>) => Promise<void>>(),
+}));
 
 vi.mock('../approvals/index.js', () => ({
   requestApproval: (...a: unknown[]) => mockRequestApproval(...a),
+  notifyAgent: vi.fn(),
+  registerApprovalHandler: (action: string, handler: (ctx: Record<string, unknown>) => Promise<void>) => {
+    approvalHandlers.set(action, handler);
+  },
 }));
 vi.mock('../../db/container-configs.js', () => ({
   getContainerConfig: (...a: unknown[]) => mockGetContainerConfig(...a),
@@ -42,36 +64,81 @@ vi.mock('./write-destinations.js', () => ({
 vi.mock('./db/agent-destinations.js', () => ({
   getDestinationByName: () => undefined,
   createDestination: vi.fn(),
+  hasDestination: () => true,
   normalizeName: (s: string) => s.toLowerCase().replace(/[^a-z0-9]+/g, '-'),
 }));
 // notifyAgent writes to the session inbound.db + wakes the container; stub both.
+// delivery.ts and agent-route.ts pull more session-manager exports at import time.
 vi.mock('../../session-manager.js', () => ({
   writeSessionMessage: (...a: unknown[]) => mockNotifyWrite(...a),
+  openInboundDb: vi.fn(),
+  openOutboundDb: vi.fn(),
+  clearOutbox: vi.fn(),
+  readOutboxFiles: vi.fn().mockReturnValue([]),
+  resolveSession: vi.fn(),
+  sessionDir: vi.fn().mockReturnValue('/tmp/nowhere'),
+  inboundDbPath: vi.fn().mockReturnValue('/tmp/nowhere/inbound.db'),
 }));
 vi.mock('../../container-runner.js', () => ({
   wakeContainer: vi.fn().mockResolvedValue(undefined),
 }));
 vi.mock('../../db/sessions.js', () => ({
   getSession: (id: string) => ({ id, agent_group_id: 'ag-1' }),
+  getPendingApproval: (id: string) => liveApprovals.get(id),
+  getRunningSessions: () => [],
+  getActiveSessions: () => [],
+  createPendingQuestion: vi.fn(),
 }));
 
-import { handleCreateAgent } from './create-agent.js';
+// The a2a module barrel registers ./guard.js (catalog entries) and the
+// guard-wrapped create_agent delivery action — the path under test.
+import './index.js';
+import { getDeliveryAction } from '../../delivery.js';
 
 const SESSION = { id: 'sess-1', agent_group_id: 'ag-1' } as Session;
 
+async function runCreateAgent(content: Record<string, unknown>): Promise<void> {
+  const wrapped = getDeliveryAction('create_agent');
+  expect(wrapped).toBeDefined();
+  await wrapped!(content, SESSION, undefined as never);
+}
+
+function liveGrant(approvalId: string, payload: Record<string, unknown>): PendingApproval {
+  const row = {
+    approval_id: approvalId,
+    session_id: SESSION.id,
+    request_id: approvalId,
+    action: 'create_agent',
+    payload: JSON.stringify(payload),
+    created_at: new Date().toISOString(),
+    agent_group_id: 'ag-1',
+    channel_type: null,
+    platform_id: null,
+    platform_message_id: null,
+    expires_at: null,
+    status: 'pending',
+    title: '',
+    options_json: '[]',
+    approver_user_id: null,
+  } as PendingApproval;
+  liveApprovals.set(approvalId, row);
+  return row;
+}
+
 beforeEach(() => {
   vi.clearAllMocks();
+  liveApprovals.clear();
 });
 
 afterEach(() => {
   vi.restoreAllMocks();
 });
 
-describe('handleCreateAgent — scope-based authorization', () => {
+describe('create_agent — guard-based authorization (wrapped delivery action)', () => {
   it('global scope: creates directly, no approval requested', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'global' });
 
-    await handleCreateAgent({ name: 'Scout', instructions: 'help' }, SESSION);
+    await runCreateAgent({ name: 'Scout', instructions: 'help' });
 
     expect(mockRequestApproval).not.toHaveBeenCalled();
     expect(mockCreateAgentGroup).toHaveBeenCalledTimes(1);
@@ -84,7 +151,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
     // dropping the inheritance leaves the child provider-less (→ claude).
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'global', provider: 'codex' });
 
-    await handleCreateAgent({ name: 'Scout', instructions: 'help' }, SESSION);
+    await runCreateAgent({ name: 'Scout', instructions: 'help' });
 
     expect(mockInitGroupFilesystem).toHaveBeenCalledWith(
       expect.anything(),
@@ -96,7 +163,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('claude creator leaves the child provider unset (built-in default)', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'global' }); // no provider
 
-    await handleCreateAgent({ name: 'Scout', instructions: 'help' }, SESSION);
+    await runCreateAgent({ name: 'Scout', instructions: 'help' });
 
     expect(mockUpdateScalars).not.toHaveBeenCalled();
   });
@@ -104,7 +171,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('group scope (default): requires approval, does NOT create directly', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
 
-    await handleCreateAgent({ name: 'Scout', instructions: 'help' }, SESSION);
+    await runCreateAgent({ name: 'Scout', instructions: 'help' });
 
     expect(mockRequestApproval).toHaveBeenCalledTimes(1);
     expect(mockRequestApproval.mock.calls[0][0]).toMatchObject({ action: 'create_agent' });
@@ -115,7 +182,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('missing config: fails closed to approval (no direct create)', async () => {
     mockGetContainerConfig.mockReturnValue(undefined);
 
-    await handleCreateAgent({ name: 'Scout' }, SESSION);
+    await runCreateAgent({ name: 'Scout' });
 
     expect(mockRequestApproval).toHaveBeenCalledTimes(1);
     expect(mockCreateAgentGroup).not.toHaveBeenCalled();
@@ -124,7 +191,7 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('disabled/other scope: requires approval', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'disabled' });
 
-    await handleCreateAgent({ name: 'Scout' }, SESSION);
+    await runCreateAgent({ name: 'Scout' });
 
     expect(mockRequestApproval).toHaveBeenCalledTimes(1);
     expect(mockCreateAgentGroup).not.toHaveBeenCalled();
@@ -133,9 +200,58 @@ describe('handleCreateAgent — scope-based authorization', () => {
   it('empty name: neither creates nor requests approval', async () => {
     mockGetContainerConfig.mockReturnValue({ cli_scope: 'global' });
 
-    await handleCreateAgent({ name: '' }, SESSION);
+    await runCreateAgent({ name: '' });
 
     expect(mockRequestApproval).not.toHaveBeenCalled();
     expect(mockCreateAgentGroup).not.toHaveBeenCalled();
   });
 });
+
+describe('create_agent — approved replay (grant-carrying re-entry)', () => {
+  it('valid grant executes exactly once — baseline hold is satisfied, create runs', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    const payload = { name: 'Scout', instructions: 'help' };
+    const approval = liveGrant('appr-ca-1', payload);
+
+    const continuation = approvalHandlers.get('create_agent');
+    expect(continuation).toBeDefined();
+    await continuation!({ session: SESSION, payload, approval, userId: 'telegram:admin', notify: vi.fn() });
+
+    expect(mockCreateAgentGroup).toHaveBeenCalledTimes(1);
+    expect(mockRequestApproval).not.toHaveBeenCalled(); // no second card
+  });
+
+  it('dead grant (row already resolved) refuses the replay', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    const payload = { name: 'Scout', instructions: 'help' };
+    const approval = liveGrant('appr-ca-2', payload);
+    liveApprovals.delete('appr-ca-2'); // resolution consumed the row
+
+    await approvalHandlers.get('create_agent')!({
+      session: SESSION,
+      payload,
+      approval,
+      userId: 'telegram:admin',
+      notify: vi.fn(),
+    });
+
+    expect(mockCreateAgentGroup).not.toHaveBeenCalled();
+    expect(mockRequestApproval).not.toHaveBeenCalled(); // refused, not re-held
+  });
+
+  it('mismatched grant (approved for a different name) refuses the replay', async () => {
+    mockGetContainerConfig.mockReturnValue({ cli_scope: 'group' });
+    const approval = liveGrant('appr-ca-3', { name: 'OtherAgent' });
+
+    await approvalHandlers.get('create_agent')!({
+      session: SESSION,
+      payload: { name: 'Scout' },
+      approval,
+      userId: 'telegram:admin',
+      notify: vi.fn(),
+    });
+
+    expect(mockCreateAgentGroup).not.toHaveBeenCalled();
+    expect(mockRequestApproval).not.toHaveBeenCalled();
+  });
+});