Maestro Companion

Agent timeout: spawn_agent 异步执行且无内置超时 — 除明确短任务外一律 spawn_agent 后立即 wait_agent({ timeout_ms: 3600000 })（上限 1 小时）阻塞等待，绝不依赖 30000 默认值；timed_out: true 且 Agent 未完成时再次 wait_agent 续等，不丢弃。批量场景使用 spawn_agents_on_csv({ max_runtime_seconds: 3600, ... })。

If any required file above was not expanded into context by the host, or its content is no longer in context, Read it explicitly before executing any step.

Use when:

• Intent is mechanically clear (no design decisions needed; file count irrelevant)
• No typed artifact consumed by downstream steps
• No gate/verdict needed for lifecycle tracking

Lightweight self-check (all must hold):

• Intent specifies a concrete, bounded action with named target (file, function, error message)
• No typed artifact consumed by downstream steps
• No gate/verdict for lifecycle tracking
• Single concern, no multi-phase span If self-check fails mid-execution, stop and suggest /maestro-next for re-routing.

Mode detection: intent → execute | empty → request_user_input: request intent text; if still empty → display usage hint and exit

Knowledge utilities (note/log/promote) are available via /maestro-knowledge.

Execute (default)

Linear: resolve Session identity -> dispatch Run -> explore -> confirm -> do -> check -> complete Run -> complete Session when the chain is terminal.

1. Create

Follow the self-start flow in run-mode.md. Negotiate capabilities, then execute the three receipt-chained mutations below. {open_request_id}, {insert_request_id}, and {next_request_id} are distinct stable IDs; use the exact session_id and orchestration_revision returned by each preceding receipt. Participant and actor are the same authorized identity.

maestro session open "<intent>" --id <slug> --participant {actor_id} --actor {actor_id} --request-id {open_request_id} --reason "open self-started Companion Session" --json
maestro session chain insert --session {session_id} --step-id {step_id} --command companion --arg "<intent>" --participant {actor_id} --actor {actor_id} --request-id {insert_request_id} --reason "add Companion task" --expected-orchestration-revision {open_orchestration_revision} --json
maestro run next --session {session_id} --participant {actor_id} --actor {actor_id} --request-id {next_request_id} --reason "dispatch Companion task" --expected-orchestration-revision {insert_orchestration_revision} --json

The Session objective is metadata; session chain insert --arg "<intent>" supplies Companion's positional domain text. Never pass task prose through --input; that option accepts only sealed same-Session Artifact IDs. Consume the run next birth packet's task and structured continuation, and retain its exact Run locator and revisions.

Init {run_dir}/evidence/companion-log.md:

# Companion Log: {intent}
> run_id: {run_id} | session: {session_id}

## Evidence

2. Explore

Locate targets and gather evidence before touching anything. Methods (pick what fits):

• maestro explore "FIND: ...\nSCOPE: ..." — codebase search
• maestro search "<keywords>" --type spec --type knowhow — knowledge recall
• spawn_agent(subagent) — multi-file analysis, cross-reference, pattern discovery
• Direct Read/Grep/Glob — known targets, quick lookups

Record findings under ## Evidence:

## Evidence
- {file:line — what was found}
- {spec/knowhow entries loaded, or "none"}
- {subagent conclusions if used}

3. Confirm

Before executing, verify evidence is sufficient:

• Target files/locations identified?
• Change scope clear (what to modify, what to leave alone)?
• No ambiguity requiring design decisions?

If insufficient → continue exploring or ask user. If -y → skip user confirmation interaction, but still perform evidence sufficiency self-check. If critical targets are unlocated, continue exploring (without asking user); only the 'ask user' branch is skipped.

4. Do

Execute the task. After each meaningful action, append under ## Work Log:

### {HH:MM} — {summary}
{outcome, files touched if any}

Rules: batch trivial reads; 1-5 lines per entry; focus on outcome not process.

5. Seal

Append outcome:

## Outcome
**Status:** done | partial
**Summary:** {1-2 sentences}
**Files:** {modified/created, or "none"}

Before completion, put accepted decisions/locked constraints in report.md. If a reusable recipe or pitfall emerged, stage it now:

maestro knowledge stage knowhow "<title>" --content-file <path|-> --run <run_id>
# Then use the complete fenced `maestro run complete ... --advance` and, when the
# chain is terminal, `maestro session complete` from run-mode.md with the current
# locator, orchestration_revision, and identity.

Display: Companion done. Run: {run_id} | Evidence: {path}

If the completion receipt contains candidate IDs, display its review_command. Do not persist the same insight again through /maestro-spec or /maestro-knowhow.

If execution revealed the task requires multi-phase audit/diagnosis (e.g., root cause unknown, >3 files need coordinated changes), suggest: /maestro-odyssey "<scope>" --mode debug|improve for re-planning.