agent-orchestra

Perform pass confirmation bookkeeping after a task validates successfully. Use after a PASS evidence line has been written. Handles checkbox toggle, feature_list.json update, progress.txt append, and optional checkpoint commit.

Supervise one OpenSpec task at a time through any configured dispatcher while preserving strict role boundaries and auditable evidence. Use when a change is tracked in an OpenSpec tasks.md and a worker must implement exactly one eligible task, produce a validation bundle, and leave final validation and bookkeeping to the supervisor.

Validate an OpenSpec validation bundle and produce auditable evidence. Use when a worker has produced a BUNDLE and the supervisor needs to run validation and record a formal EVIDENCE line. Can be used standalone (after manual implementation) or as part of the supervisor workflow.

Dispatch a task via any generic CLI tool such as aider, cursor CLI, or other AI coding assistants. Use when the supervisor needs to delegate implementation to a non-Codex, non-Claude CLI tool.

Dispatch a task to Codex CLI for execution. Use when the supervisor needs to delegate implementation to Codex running in full-auto mode. Codex will run in a sandboxed environment and apply file changes via patch.

Hand off a task to a human for manual implementation. Use when no automated dispatcher is available, the task requires human judgment, or the supervisor wants explicit human confirmation before proceeding.

Dispatch a task via an MCP tool call. Use when an MCP server provides a code generation or task execution tool that can receive a task prompt and produce implementation output.

Dispatch a task by launching a Claude Code subagent via the Agent tool. Use when the supervisor is running inside Claude Code and wants to delegate implementation to an isolated subagent with optional worktree isolation.

Codex entrypoint for fully automated execution of all tasks in an OpenSpec change. Use when Codex itself should run the whole change loop until completion or a terminal stop condition.

Codex entrypoint for a single supervised OpenSpec task attempt. Use when Codex itself should run one task attempt for step-by-step orchestration or debugging.

Build, repair, or review feature_list.json under an OpenSpec change folder. Use when a change needs durable ref-based feature tracking, when refs in tasks.md and feature entries drift, when multiple tasks reference the same ref, or when a supervisor needs a clean pass/fail checklist per stable ref tag.

Implement exactly one OpenSpec task and produce a validation bundle. Use when a CLI or tool is assigned as the "code writer" role in an orchestra workflow. Defines strict permission boundaries so that validation, bookkeeping, and commits remain owned by the supervisor.

Clarify OpenSpec change requests before proposal or implementation. Use when a user has a vague feature idea, a rough proposal, or an existing change-id and needs to turn ambiguous intent into concrete scope, constraints, acceptance criteria, edge cases, and implementation-ready notes for OpenSpec.

Research blockers in an OpenSpec implementation workflow and turn them into actionable unblock guidance. Use when a task is blocked, validation keeps failing, a dependency or library issue is suspected, a worker crashed or timed out, or the supervisor needs evidence-backed next steps instead of ad hoc debugging.