tracer-dev（追踪器开发）

Tracer Development

Execute the build-validate-review cycle for tracer bullet development.

Locate State Files

Before any flow, find the plan files:

• Search docs/tracers/**/state.json
• If multiple matches, list them and use the ask question tool to ask questions about which feature to work on
• If none, STOP: "Run spec-planner to create docs/specs//SPEC.md, then run /overseer-plan to create executable tasks"

Two files per feature:

• state.json - Machine state (task passes, commits, current tracer)
• PROGRESS.md - Human-readable plan and learnings

All state updates go to state.json. Learnings are appended to PROGRESS.md.

Decision Tree

Check state.json:
│
├─ No state.json found
│  └─ STOP: "Run spec-planner then /overseer-plan first to create a plan"
│
├─ Current tracer has tasks with passes: false
│  └─ → NEXT-TASK FLOW
│
├─ All tasks pass, tracer status != "complete"
│  └─ → REVIEW-TRACER FLOW
│
├─ Tracer complete, more tracers remain
│  └─ Advance currentTracer → NEXT-TASK FLOW
│
└─ All tracers status: "complete"
   └─ DONE: "Feature complete!"

NEXT-TASK FLOW

Precondition Check

Is lastReviewedTracer == currentTracer - 1?
├─ No → "Complete tracer review first"
└─ Yes → Continue

If feature status is planning, update it to in_progress in state.json before starting work.

Proceed immediately to Step 1 without asking for confirmation once preconditions are satisfied.

Step 1: Select Task

Read state.json and find first task with passes: false in the current tracer.

Also read the context block for patterns and key files to guide exploration.

Step 2: Write Micro-Plan

Before ANY code, produce a brief plan:

## Micro-Plan: Task X.Y - <name>

**Goal**: <one sentence>

**Implementation** (minimal):
1. <step>
2. <step>
...

**Validation**:
- <how we prove this works>

**Not doing** (explicit scope):
- <what we're deferring>

Step 3: Implement

Key principles:

• Use EXISTING abstractions before creating new ones
• Minimal code that satisfies the task
• Production bar: types + error handling
• No speculative features

Step 4: Validate

Run ALL verification steps from the task's steps array. Also run project validation (typecheck, tests, lint).

• Detect project scripts (package.json, Makefile) or use the ask question tool to ask questions if unclear
• If a step cannot run, mark SKIPPED with reason
• All steps must pass before proceeding

Step 5: Code Review Gate (BLOCKING)

afplay /System/Library/Sounds/Ping.aiff

Present to user:

## Code Review: Task X.Y

Files changed:
- <file> (+X, -Y)

Verification steps:
- <step 1>: PASS
- <step 2>: PASS

Project validation:
- Typecheck: PASS or SKIPPED (reason)
- Tests: PASS or SKIPPED (reason)
- Lint: PASS or SKIPPED (reason)

Request approval inline (do not open question mode): reply **"lgtm"** to approve, or provide feedback.

WAIT FOR USER RESPONSE

• "lgtm" (exactly) → Continue to commit
• Any other response → Treat as feedback, revise, re-request review

Loop until user says "lgtm".

After "lgtm", proceed immediately to commit, mark the task complete, and start the next task without asking for confirmation.

Step 6: Commit + State Update (Background Agent)

After approval, always commit the task. This ensures the next task's review only shows new changes.

Run the commit and state.json update using a background agent with openrouter/z-ai/glm-4.7-flash.

git add -A && git commit -m "feat(<scope>): <task description>"

Step 7: Mark Task Complete

Update state.json for the task (handled by the background agent):

{
  "passes": true,
  "commit": "<commit-hash>"
}

Continue to next task, or if all tasks in current tracer have passes: true → REVIEW-TRACER FLOW.

REVIEW-TRACER FLOW

Precondition

All tasks in current tracer have passes: true in state.json.

Step 1: Definition of Done Audit

Check each criterion from the tracer's proves field:

## DoD Audit: Tracer N

proves: "<what this tracer validates>"

Evidence:
- [x] <criterion>: <evidence>
- [x] <criterion>: <evidence>
- [ ] <criterion>: MISSING - <what's needed>

If any criterion unmet → Add tasks to state.json (with passes: false), return to NEXT-TASK FLOW. If tasks are added after a tracer was reviewed, set in state.json:

{
  "lastReviewedTracer": N-1,
  "currentTracer": N
}

Step 2: Demo

afplay /System/Library/Sounds/Hero.aiff

Present demo instructions:

## Demo: Tracer N

Please run:

    <demo command from PROGRESS.md>

Expected result:
<expected output from PROGRESS.md>

Request response inline (do not open question mode): run the demo and reply "pass" to continue, or describe what went wrong.

WAIT FOR USER RESPONSE

Continuation Logic

• "pass" → Continue to learnings
• Any other response → Treat as feedback, discuss adjustments, may add tasks

Step 3: Record Learnings

Use the ask question tool to ask questions:

What did we learn from this tracer?
- Decisions made?
- Surprises?
- Things to remember for future tracers?

Update both files:

state.json - Add to tracer's learnings array:

{
  "learnings": ["<learning>", "<learning>"]
}

PROGRESS.md - Append under the tracer's Learnings section:

### Learnings

- <learning>
- <learning>

Step 4: Update Future Tracers

Review remaining tracers in light of learnings:

## Future Tracer Review

Based on learnings from Tracer N:

Tracer N+1: <name>
- Still makes sense? [yes/adjust]
- Demo still valid? [yes/adjust]
- New tasks needed? [no/add]

Update PROGRESS.md if any changes needed.

Step 5: Mark Tracer Complete

Update state.json:

{
  "lastReviewedTracer": N,
  "currentTracer": N+1,
  "tracers": [{
    "id": N,
    "status": "complete",
    "completed": "<ISO date>"
  }]
}

If this was the final tracer, also set:

{
  "status": "complete"
}

Step 6: Celebrate

afplay /System/Library/Sounds/Funk.aiff

Tracer N complete!

Progress: N/<total> tracers done
Next: Tracer N+1 - <name>

Continuing to the next tracer.