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](~/.maestro/workflows/maestro.md) — read before initial intent classification
- [ralph-amend-goal.md](~/.maestro/workflows/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