feature-list

name	feature-list
description	Expand the
when_to_use	Reach for this immediately after a plan is approved (ExitPlanMode) and before the generator dispatches — `/tdd-loop` and `/verify` both read the resulting features.json. Do NOT use for free-form work without a plan; feature-list expects a structured `## Contract` section as its source of truth.
disable-model-invocation	true
allowed-tools	["Read","Write","Glob","Grep","Bash"]
logical	.claude/features.json contains a testable requirements array derived from the plan's contract

/feature-list — Contract → Testable Requirements JSON

Convert the latest plan's ## Contract into .claude/features.json. This file becomes the shared source of truth for what "done" means; /tdd-loop consumes it, /verify runs the verify_cmds, and after-subagent.sh flips items to status: done.

Process

1. Locate the latest plan. Glob .claude/plans/*.md; pick the most recently modified. If none, stop: No plan found. Run Plan mode first.

2. Extract ## Contract. Read the plan; find the ## Contract section. If missing, stop: Plan has no ## Contract section. /feature-list expects one.

3. Parse criteria. Each bullet under ## Contract becomes one feature. Trim; skip empty lines and sub-bullets (sub-bullets become part of the parent description).

4. Derive verify_cmd per item. Inspect the criterion text and the repo stack:

   • Mentions a test? → <test-runner> --filter=<hint> (pest/pytest/jest/cargo test/go test).
   • Mentions a file edit? → test -f <path> or grep -q '<expected>' <path>.
   • Mentions a command output? → <cmd> | grep -q '<pattern>'.
   • No clear verification → # manual — flag for review.

5. Emit .claude/features.json (overwrite if present; this is a fresh expansion per plan):

   [
     {
       "id": "F1",
       "description": "<contract bullet>",
       "verify_cmd": "<command or '# manual'>",
       "status": "pending",
       "plan": ".claude/plans/<file>.md"
     }
   ]
   

   IDs are stable: F1, F2, ... in Contract order.

6. Report: Expanded N features from <plan>. Run /tdd-loop or /verify to execute verify_cmds.

Integration

• .claude/features.json is read by:
  • /verify (evaluator plugin) — runs each verify_cmd and writes results.
  • after-subagent.sh (workflow plugin) — flips status when matching work completes.
  • surface-progress.sh (this plugin, SessionStart) — summarizes pending/in_progress/done counts.
  • /rest-audit — Efficiency axis (pending/done ratio).
• /tdd-loop does not read this file; it accepts a feature description as $ARGUMENTS and runs Red/Green/Refactor against the test suite. Use /feature-list first to make the contract machine-readable for /verify, then drive /tdd-loop per item.

Failure Modes

• .claude/features.json exists and has done items → do NOT clobber silently. Read it; preserve completed entries when re-expanding; assign new IDs to new items.
• Contract bullets are ambiguous → emit verify_cmd: "# manual" and report count; the user can refine.
• No recognizable stack for test command → emit # manual.