maestro

<required_reading> @/.maestro/workflows/run-mode.md @/.maestro/workflows/orchestrator-run-loop.md @/.maestro/prepare/maestro.md @/.maestro/workflows/codex-run-mode.md </required_reading>

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 the state machine.

<deferred_reading>

• maestro.md — read before initial intent classification
• ralph-amend-goal.md — read only for --amend </deferred_reading>

<purpose> Turn a user intent into the initial Skill chain, create one canonical topic Session through `maestro session open --chain <commands...>` (no chain-file), then execute the shared Run loop. Static versus dynamic is not a Session or command mode: each Skill contract decides whether it emits a typed chain proposal. For new intents, use this command. For policy-driven execution over existing Sessions, use `/maestro-ralph`. </purpose> <interface> Only these user flags are accepted:

• -y — skip all confirmation/clarification interactions, use default choices. Does NOT change data semantics (no auto-deferred decisions). Never bypasses: high-risk classification, confidence <60, ambiguity requiring user input, failed gates, or drift escalation.
• -c — continue the unique live compatible Session.
• --amend — amend that Session's goal; remaining text is the change request.

Execution always dispatches run-executor (the default behavior); this never changes Session type or chain semantics.

All other text is intent. Unknown flags are not silently reinterpreted. Platform, roadmap, quality, template reuse, parallelism and adversarial depth are inferred. </interface>

<invariants> 1. **One chain** — every task uses the same Session/Run protocol; no static/dynamic, Maestro/Ralph, or executor-specific Session type. 2. **Session before execution** — open via `maestro session open "<objective>" --id <slug> --chain <commands...>` before allocating a step Run. 3. **Creator owns decomposition** — Maestro creates `boundary_contract` and outcome-oriented goals; later orchestrators consume rather than overwrite them. 4. **Runtime owns mutation** — prompt never writes session.json/run.json and never auto-uses admin chain commands. 5. **Skill owns domain adaptation** — optional chain changes come only from the current Skill's validated `chain-proposal/1.0`. 6. **Verdict advances** — execution steps advance only through fenced `maestro run complete ... --advance --verdict done|done_with_concerns`; decision steps only through fenced `maestro run decide`. 7. **Historical similarity remains read-only evidence** — it never selects a Session or binds outputs. 8. **Compatibility commands are out of band** — normal orchestration calls only `maestro run ...`. 9. **Auto is bounded** — `-y` never bypasses high risk, low confidence, ambiguity, failed gates or drift escalation. 10. **Router is not a step** — `/maestro-next` may route here but never appears inside the chain. 11. **Running means continue** — while canonical continuation authority is `automatic`, execute it and re-read the receipt in the same turn; `suggest_only` is Runtime passivity, not a reason to end the turn. </invariants>

<state_machine>

<states> S_PARSE — parse intent and flags S_CONTINUE — locate the unique live Session S_AMEND — audited goal amendment S_CLASSIFY — select the smallest sufficient initial chain S_DECOMPOSE — derive boundary, criteria and observable goals S_CREATE — open via `session open --chain` S_CONFIRM — confirm classification unless `-y` S_RUN_LOOP — execute `orchestrator-run-loop.md` S_FALLBACK — request missing intent or disambiguation </states> <transitions> S_PARSE: → S_AMEND WHEN: `--amend` → S_CONTINUE WHEN: `-c` → S_CLASSIFY WHEN: intent present → S_FALLBACK OTHERWISE

S_CONTINUE: → S_RUN_LOOP WHEN: exactly one live compatible Session → S_FALLBACK WHEN: Session has an open decision gate or a stuck Run (suggest /maestro-ralph -c for audited recovery) → S_FALLBACK WHEN: none or multiple

S_AMEND: → S_RUN_LOOP WHEN: shared amend protocol committed → END WHEN: cancelled or blocked

S_CLASSIFY: → S_RUN_LOOP WHEN: existing compatible Session found (do not rebuild) → S_DECOMPOSE WHEN: multi-step chain → S_CREATE WHEN: narrow/single-step chain → S_FALLBACK WHEN: confidence < 60

S_DECOMPOSE → S_CREATE S_CREATE → S_RUN_LOOP WHEN: -y AND risk ≠ high AND confidence ≥ 60 S_CREATE → S_CONFIRM WHEN: -y AND (risk == high OR confidence < 60) S_CREATE → S_CONFIRM OTHERWISE S_CREATE → S_FALLBACK WHEN: creation fails (delete temp file, report error) S_CONFIRM → S_RUN_LOOP WHEN: confirmed S_CONFIRM → S_CLASSIFY WHEN: revised (maestro re-classifies the revised intent from scratch because a changed intent may reshape the chain; ralph returns to S_BUILD instead since its chain shape is already fixed) S_CONFIRM → END WHEN: cancelled </transitions>

<actions>

A_CLASSIFY

Read deferred maestro.md. Record matched evidence, excluded alternatives and confidence before creation.

Minimum chain rules:

Intent evidence	Initial chain
narrow fix/change	analyze → plan → execute → review/test as required
broad rewrite/migration	analyze-macro → scope decision → plan/roadmap path
brainstorm/explore	brainstorm, then only Skill-proposed continuation
stress/grill	grill, then only Skill-proposed continuation
formal specification	blueprint → plan path
existing compatible Session	do not rebuild; enter shared loop

Roadmap is inferred only for multi-release evidence. Quality depth follows project specs, UI evidence needs frontend verification, and every executable command is resolved by Run Runtime.

A_DECOMPOSE

For broad intent, ask at most 3 questions covering scope, constraints and observable done criteria; broad ambiguity is not skipped by -y. (broad = affects ≥3 modules OR requires cross-package interface changes OR ≥2 of 3 decomposition questions remain unanswered.) Produce:

{
  "boundary_contract": { "in_scope": [], "out_of_scope": [], "constraints": [], "definition_of_done": "" },
  "decomposition": {
    "execution_criteria": [],
    "goals": [{ "id": "G1", "goal": "", "boundary": "", "done_when": "", "evidence": "", "lifecycle": [], "status": "pending" }],
    "changelog": []
  }
}

Goals describe outcomes, not lifecycle stages.

A_CREATE

Assemble and create per prepare/maestro.md §1–§4 (specs precheck, Skill-name prevalidation, chain assembly, creation). Maestro-specific policy:

• The chain assembly protocol (§3 template) lives in prepare/maestro.md (required_reading); if it is not in context, Read that file directly. In v3 the same prepare guidance is injected into the run next / run create birth packet (guidance-snapshot/1.0) — prepare is embedded in the Run, not a standalone run prepare step.
• Maestro does not emit formal decision nodes; new chains express quality/goal/scope checks as Skill steps that own a Run and may return a proposal. (The closed-loop policy that mandates decision nodes before seal belongs to /maestro-ralph; route there when the work needs it.)
• For narrow/single-step chains, generate a minimal implicit boundary_contract: in_scope = [intent], out_of_scope = [], constraints = [], definition_of_done = 'step completed with passing gates'.
• Do not inline unescaped JSON.

A_CONTINUE

Use read-only maestro run recall plus maestro session status --session {session_id} --json. A Session with an open decision gate or a stuck Run is out of scope here — report it and route to /maestro-ralph -c for audited recovery (S_FALLBACK); completed/archived Sessions are terminal. Multiple live candidates require explicit selection.

A_AMEND

Read ralph-amend-goal.md, use maestro session status --session {session_id} --json for the snapshot, perform read-only impact analysis, confirm, then apply the amendment through a chain-aware typed proposal: goal/decomposition metadata is committed with fenced maestro session chain insert|replace (per-step --goal-ref / --stage / --decision-ref). session meta update is session/1.x/2.0 compatibility-only and must not be used in the canonical branch. Any pending-tail change must come from a planning Skill proposal.

</actions>

</state_machine>

<task_tracking>

时机与操作（plan 是 session 权威状态的 UI 镜像，不替代 session 状态）：

时机	操作	示例
Session 创建后	update_plan 初始化步骤清单	update_plan({ plan: [{ step: "Step {index}: {step.skill}", status: "pending" }, ...] })
Step 派发时	update_plan 标记当前 step	update_plan({ plan: [..., 当前 step status: "in_progress"] })
Step 完成时	update_plan 标记完成	update_plan({ plan: [..., 该 step status: "completed"] })
Step 失败时	update_plan + explanation 说明	update_plan({ explanation: "Step {index} failed: {reason}", plan: [...] })

</task_tracking>

<success_criteria>

• Public flags are -y, -c, --amend.
• Initial classification is auditable and the Session exists before step execution.
• Every step follows run next → brief → execute → run check → run complete --advance; decision nodes use run decide.
• Chain adaptation is Skill-proposed and atomically applied by the producing Run (session chain insert|replace|skip).
• Normal output and recommendations contain only maestro run ... lifecycle commands. </success_criteria>

Legacy session/1.x/2.x Compatibility Branch

The v2 command surface is deprecated / compatibility-only for explicitly selected old CLI/schema (see run-mode.md Legacy session/1.x/2.x Compatibility Branch): maestro session create --chain-file, maestro session done --verdict, maestro session decide, maestro session meta update, and the standalone run prepare dispatcher are never used by the canonical flow above. Normal orchestration calls only the v3 surface: session open/status/list/chain insert|replace|skip/complete, run next/create/brief/check/complete --advance/decide/transition/cancel, and knowledge stage/review/promote.