# Project Orchestrator > Use when routing Codex setup, planning, execution, verification, or workflow repair. - Skill: `cyz26/project-orchestrator` (Agent Skill) - Install (CLI): `npx skillmds@latest add cyz26/project-orchestrator` - Raw SKILL.md: https://api.skillmd.com/api/skills/cyz26/project-orchestrator/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Productivity - Author: cYz26 (https://skillmd.com/u/cyz26) - Updated: 2026-09-21 - Page: https://skillmd.com/skills/cyz26/project-orchestrator --- # Project Orchestrator Use `AGENTS.md`, DevFlow state, OpenSpec, and the active ledger as canonical. ## Route 1. Missing workflow files route to `project-setup`. 2. Classify with `feature-intake`; behavioral and compatibility changes require Full OpenSpec. 3. Diagnose exact capabilities and preview activation before apply. 4. External uncertainty uses `capability-research` and its Capability Evidence Gate. 5. Route plans to `change-plan`, work to `execute-task`, completion to `verify-and-archive`, drift to `workflow-doctor` with `root-cause-diagnosis`, and technical plans to `ai-native-tech-plan`. Before product mutation run: ```bash python3 scripts/implementation_readiness.py check-mutation --repo \ --change-id --ordinary-authority --json ``` When `implementation_readiness.required: true`, Required, NotReady, or stale Ready evidence is technical repair. Ready is evidence, not authority. Only concrete missing authority reaches the Human Gate recorder. ## Continuous Execution After approval use `auto-until-terminal` and keep the enclosing `execute -> evidence -> decide -> continue` loop: 1. Select the active Full OpenSpec tasks, or `TASK_LEDGER.md` only when no OpenSpec task list is active. 2. Send one dependency-ready item to `execute-task`, verify its completion receipt, and update canonical task and state evidence. 3. Derive `CONTINUE_NEXT_ITEM`, `CHECKPOINT_AND_CONTINUE`, `VERIFY_ACTIVE_CHANGE`, `AWAIT_HUMAN`, `READY_FOR_EXTERNAL_EFFECT`, or `COMPLETE`; continue immediately for the first three. A phase label is not a Human Gate. Technical failure is `FAIL_CLOSED_REPAIR`. Only a concrete authority delta may persist `AWAIT_HUMAN` with aligned evidence and state. Recheck readiness before writes and after resume. For `model.*`, the exact stable identity is Standing Goal Execution Authority. A one-use attempt receipt protects replay; it is not one-use permission. Record actual monetary cost. Same-authority repair, refreeze, review, and retry continue without another question; provider, model, account, credential privilege, or acceptance-contract changes are material deltas. Optional work uses `DEFER_AND_CONTINUE` and stays off the critical path. A standing milestone covers only its exact reviewed external chain; archive, PR, merge, force, and undeclared targets remain excluded. ## Generated Artifact Lifecycle After the owning process exits, automatically apply and verify only a fresh `AUTO_CLEAN` plan; direct CLI use still requires `cleanup --apply`. Preserve the cleanup receipt. `WAIT_OWNER` retries, `RETAIN` preserves, and drift is technical repair. Legacy `HUMAN_GATE` is resolver input, not permission to write awaiting state. Hooks, doctors, and validators never clean artifacts. ## Git Transport vs GitHub Control Plane A gh authentication failure is not Git transport failure. Authorized push uses expected-base and fast-forward proof plus `git ls-remote`; PR, release, and settings use `github.control_plane_write`. For immutable-tag release prefer `github_actions`, then `github_cli`, then `human_web`. The workflow exists in the immutable tag target, repository policy allows it, and `GITHUB_TOKEN` has least privilege. Require publication readback before local promotion. On failure, preserve the tag and reviewed identity. For `github_cli`, allow one diagnosis and at most one applicable remediation attempt. Details live in `docs/git_transport_routing.md`. ## Capability Routing Use stable IDs from `workflow_methodology.py`: `decision-resolution`, `implementation-planning`, `test-first-execution`, `root-cause-diagnosis`, `change-review`, `completion-proof`, `execution-orchestration`, `architecture-guidance`, and triggered `domain-language-modeling`. Matt skills are bounded primitives; DevFlow and OpenSpec own canonical plans and evidence. ## Goal Gate Route goal-backed, long-running, multi-slice, migration/release, or cross-context work to `define-goal`. Record outcome, verification, scope, non-goals, success threshold, and stop conditions. Scripts only recommend Goal mode. ## SubAgent Decision Gate Delegate when independent domains and disjoint write sets materially help. First require a validated Agent Task Contract containing Goal, Scope, Constraints, Verification, Evidence, and Human Gate. The main agent owns OpenSpec, `.planning/devflow/`, shared control-plane and release files, and final integration. Delegation uses `execution-orchestration` and reports status, files, commands, risks, and review needs. ## Incidental Finding Lifecycle Route through the central resolver: `CONTINUE_WITH_MINIMAL_GUARD` for a bounded required guard, `DEFER_AND_CONTINUE` for recorded optional work, and `BLOCKED_AWAITING_HUMAN` only for concrete material scope, authority, risk, ownership, or product-decision deltas. Unknowns fail closed. The register does not authorize follow-up, and Completion Contract work cannot be deferred. ## Completion Routing names mode, capabilities, canonical artifacts, blockers, authority, next skill, and validation. The request ends only at `COMPLETE`.