[LinuxDevil]

What

• createAgent() now runs every agent with an ApprovalStore: a new InMemoryApprovalStore per agent (added in src/execution/InMemoryApprovalStore.ts, exported from the root) or the new approvalStore option. A needsApproval tool pauses the run instead of throwing "requires approval but no approvalStore".
• SimpleAgent.approvals:
  • list() returns the PendingApprovals this agent is paused on.
  • resolve({ id, approved, note? }) takes the existing ApprovalDecision shape, calls resumeAfterApproval() with the agent's store, registry, provider and run options (skills, subagents, maxSteps, toolConcurrency), and returns the continued run's ExecutionResult.
• approve?: (call: PendingApproval) => boolean | Promise<boolean> option (open-harness style). It is a thin loop on top of the same pause/resume path, with no new executor code path. When a run pauses, the callback decides the call and the run resumes with that answer: true runs the tool, false sends the model the existing structured rejection. It applies to send(), sessions and approvals.resolve().
• Sessions: a turn that pauses inside agent.session() is continued in that session when its approval is resolved. A new protected AgentSession.continueTurn() queues the continuation as the session's next turn and records its transcript.
• Paused result contract is unchanged: finishReason: 'awaiting-approval' plus approvalId.
• Docs:
  • README approval-gates section: new type-checked createAgent snippet.
  • API list, docs/api-overview.md (new Approvals section), docs/sessions.md and docs/sub-agents.md updated.
  • llms files regenerated; CHANGELOG Added entry.

Why

createAgent() is the zero-config path, but any needsApproval tool made its runs fail because no approval store was ever passed. eve pauses durably on needsApproval and open-harness has an approve callback. This brings the simple API to parity.

Acceptance criteria

• createAgent({ approvalStore?: ApprovalStore }). When it is omitted, each agent gets its own InMemoryApprovalStore; none existed before, so it was added in the execution/approvals module, not in createAgent.ts.
• approve callback auto-decides with no pause, built on top of the store/resume path.
• A paused send() resolves with the existing awaiting-approval / approvalId shape.
• agent.approvals.list() / agent.approvals.resolve({ id, approved, note? }) work, mirroring ApprovalDecision and delegating to resumeAfterApproval.
• Sessions keep working: resolving continues the same session thread.
• Tests in src/createAgentApprovals.test.ts (mockModel) cover:
  • pause then approve
  • pause then reject (structured rejection)
  • approve auto-approve and auto-deny
  • each agent's default store is isolated
  • a custom approvalStore
  • continuing a session
  • a sub-agent pause resumed through the lead
  • stream() pausing
• Docs (README, api-overview, sessions, sub-agents) updated with one type-checked snippet; llms files regenerated; CHANGELOG entry.

Notes:

• src/execution/AgentExecutor.ts, resume.ts, toolCallExecution.ts and ApprovalGate.ts are not touched.
• serveMcp.test.ts had one test that relied on the old throw. It now expects the MCP server's existing "needs human approval ... cannot be approved over MCP" error.
• stream() pauses but does not apply approve. A paused stream is continued with agent.approvals.resolve(), which does not stream.
• list() covers pauses made by this agent in this process. With a durable store, resolve() by id still works after a restart, but it cannot re-attach the result to a session.

How verified

Ran after merging the latest origin/main (including #68):

• npx tsc --noEmit: no errors
• npm run lint: ✖ 433 problems (0 errors, 433 warnings). Same count as before the change; the changed files have 0 warnings.
• npm run test:types: Type Errors no errors
• npx vitest run: Test Files 134 passed (134), Tests 1986 passed | 3 skipped (1989)
• npm run build && npm run build --workspace=packages/create-loushy-agent: exit 0
• npm run docs:verify-snippets -- --skip-build: all 89 snippet(s) type-check against src/; 6 also run cleanly
• npm run docs:llms:check: exit 0
• npm run test:coverage: exit 0 (134 files, 1986 passed)
• npm run fallow: exit 0. Dead code: ✓ No issues found; 0 above threshold. The only duplication reported is the existing Anthropic/OpenAI provider clone.

🤖 Generated with Claude Code

https://claude.ai/code/session_016bY29ibQgv3MDdjAbUT4YR

---

Generated by Claude Code