unity-orchestrate

Orchestrate one docs/<feature>/ directory. Use fresh subagents to discover completed work, implement the remainder, and review the resulting change set. Stop at the first failed gate and report the blocker.

1. Preflight

1. Require one feature directory containing spec.md and at least one file under tasks/.
2. Read the complete spec and every task.
3. Capture git rev-parse HEAD as the fixed point. A dirty worktree is valid; scouting uses the current repository state, and review covers every change since the fixed point.

2. Discover completed tasks

Call one fresh scout subagent with the feature directory, fixed point, and these instructions:

• Read the complete spec and every task, then inspect the current implementation, tests, assets, configuration, validation records, and relevant Git history.
• Judge every acceptance criterion and required validation from repository evidence. Names, checkboxes, commits, test source, and human claims do not prove completion or a successful validation run.
• Use HUMAN-PENDING only for work that inherently requires human action or judgment. Missing evidence for an automatable check is UNPROVEN.
• Mark a task COMPLETE only when every criterion is evidenced or HUMAN-PENDING, every validation is proven or HUMAN-PENDING, every blocker is complete, and no implementation or automatable validation remains. Mark every other task RUN.
• Begin with Status: complete or Status: blocked. On success, list every task once in numeric order with its COMPLETE or RUN decision, concise acceptance and validation evidence, any HUMAN-PENDING items, and blocker status.

Continue only when scouting returns Status: complete, covers every task, and supports every COMPLETE decision. Otherwise report the scouting blocker; do not default to running all tasks.

3. Implement

Process each RUN task sequentially in numeric order. Call one fresh worker subagent per task with:

/skill:unity-implement <task-path>

Continue only when the response begins Status: complete. Record changed files and deferred or unavailable validation. On failure, stop before starting another task and report progress and the blocker. Tasks classified COMPLETE need no worker and retain that status.

4. Review

Inspect tracked, staged, committed, and untracked changes against the fixed point. If none exist, record not run (no changes) and finish.

Otherwise call one fresh reviewer subagent with:

/skill:unity-code-review <fixed-point> <feature-directory>

A response of No findings. is clean. Treat a failed or incomplete review as a blocker.

5. Remediate once

When the first review has findings, pass them unchanged to one fresh worker subagent:

/skill:unity-implement Review remediation for <feature-directory>.

Review findings:
<verbatim findings>

Continue only when the response begins Status: complete; otherwise report the blocker, findings, and validation status. Record remediation changes and finding dispositions.

After successful remediation, run one fresh full review with the original fixed point and feature directory. Report its result without another remediation cycle.

6. Report

Return a compact report with:

• each task's complete or blocked status;
• implementation changed files;
• deferred or unavailable validation, including HUMAN-PENDING items;
• first-review result and findings;
• remediation changed files and finding dispositions, when applicable;
• final-review result and remaining findings, when applicable.

Do not run git add, git commit, or git stash.