lucky-build-protocol（幸运构建协议）

Build Execution Protocol

Repeatable, enforced process for executing a build queue across sessions. Same discipline as the heartbeat — mandatory checklists, verification gates, no shortcuts.

Auto-Trigger

If a project has NEXT_TASKS.md, follow this protocol automatically. Ray only needs to say "work on [project]" or "keep going."

Phase 0: Session Setup (MANDATORY — COMPLETE IN ORDER, NO EXCEPTIONS)

□ 0. Context Check (HARD STOP)

• Run session_status FIRST
• If context > 70%: Wrap up, update files, request fresh session. STOP.
• If context > 85%: CRITICAL. Save state immediately, do NOT start new work.
• Log context % for reference

□ 1. Read Project State (BEFORE model selection — you need to see the work first)

• Read NEXT_TASKS.md — find the first unchecked [ ] item
• Read STATUS.md (if exists) — understand current build state
• Read today's + yesterday's memory/YYYY-MM-DD.md — know what happened recently
• Skim the last session's work log for any warnings or half-finished items

□ 2. Model Selection (NOW that you've seen the tasks)

• Review the upcoming tasks and select model based on what they need:

- Haiku: Simple doc updates, formatting, config changes, trivial fixes - Sonnet: Standard coding, implementation, routine bug fixes, test writing - Opus: Architecture decisions, complex algorithms, hard debugging, security-critical work

• If the session has a mix: Start with Sonnet, switch to Opus for complex tasks, back to Sonnet after
• Switch model now if needed: session_status(model="...")

□ 3. Verify Previous Session's Work (DO THIS BEFORE NEW WORK)

• Look at what the last session claimed to complete (from memory files / NEXT_TASKS.md checked items)
• Actually verify it: Run the code, check the output, read the files
• If anything was marked done but doesn't work → uncheck it, fix it first
• This exists because sessions lie to themselves. Trust but verify.

□ 4. Baseline Test Run (NON-NEGOTIABLE)

• Run the full fast test suite BEFORE touching any code:

  python3 -m pytest -x -q --ignore=tests/integration

• Record the count: N passed, 0 failed
• If tests are already failing: Fix them FIRST using test-suite-repair skill
• You must end the session with ≥ this test count

□ 5. Git Checkpoint

• If the project uses git: git add -A && git commit -m "checkpoint: session start — N tests passing"
• If no git: copy critical files to a backup location
• Why: If task 3 breaks something silently, you can diff back to this known-good state
• Create a new checkpoint after each completed task

□ 6. Check Jinx Status

• ssh -i ~/.ssh/lucky_to_mac luckyai@100.90.7.148 "curl -s http://localhost:3001/status"
• If Jinx has completed results → review and verify them first (apply same verification standard)
• If Jinx is idle → good, we'll queue tasks as we go

□ 7. Plan the Session

• Look at remaining unchecked tasks in NEXT_TASKS.md
• Pick 2-4 tasks you can realistically complete this session
• WIP limit: 1 task active at a time. Finish or park before starting another.
• Size check: If a task will take >45 min of coding, consider:

- Breaking it into subtasks - Spawning a coding agent via coding-agent skill for the implementation - Queuing parts to Jinx

Phase 1: Task Execution Loop

For EACH task, follow this loop without skipping steps:

Step 1: Understand (2-5 minutes)

□ Read the task description in NEXT_TASKS.md □ Read the relevant source files (use the quick reference table at bottom of NEXT_TASKS.md) □ Read the spec section — find the actual requirement in the spec and quote it:

  SPEC REQUIREMENT: "The Judge worker evaluates 5-10 variants and selects 
  the best with a decision trace including confidence score." 
  — master_spec_full.txt, Section 4.2

Why: This is how we avoid building the wrong thing. If you can't find a spec reference, ask Ray before implementing. □ Define "done" — one sentence: "This task is done when ___" □ Jinx check: "Can Jinx do this?" If it doesn't need live coding/testing → queue for Jinx, skip to next task

Step 2: Implement

□ Jinx split check: Before coding, break the task into parts. Can any part run on Jinx? - Lucky does: Live coding, file system changes, testing, anything needing imports/execution - Jinx does: Writing docs from existing code, analyzing patterns, generating test data, reviewing logic, drafting boilerplate - Split pattern: Lucky writes the core logic → queues Jinx to write tests/docs/edge-case analysis for it - Queue Jinx parts NOW so they run in parallel while you code your parts □ Write the code □ Stay focused — one task only. Don't drift into adjacent work. □ If you discover something broken, note it in NEXT_TASKS.md as a new item but don't fix it now unless it blocks you □ Check skills — if this task type matches a skill (judge-worker-pattern, pipeline-debugging, ui-truth-mapping, etc.), read and follow that skill first

Step 3: Test (MANDATORY)

□ Write tests for the new code (if the task warrants tests) □ Run the full fast test suite:

  python3 -m pytest -x -q --ignore=tests/integration

□ Must pass with 0 failures before proceeding □ Test count must be ≥ baseline — you cannot lose tests

Step 4: Verify (NOT THE SAME AS "TESTS PASS" — USE OPUS)

This is the step that gets skipped. Don't skip it.

□ Switch to Opus model for verification: session_status(model="opus") - Opus catches things Sonnet misses — logic gaps, spec mismatches, subtle bugs - This is the quality gate. Use the strongest model here. □ Read back the code you wrote — does it do what the task asked? Not "does it compile" — does it actually implement the requirement? □ Check the output — call the function/endpoint/CLI and look at what it returns. Print it. Read it.

  # Actually run it and look at the output
  result = the_thing_you_built(test_input)
  print(json.dumps(result, indent=2))

□ Compare to spec — re-read the spec requirement you quoted in Step 1. Does the output match? □ Edge cases — empty input, null, unexpected types, boundary values □ If UI copy: Read the copy. Does it make sense to a human who isn't a developer? □ Switch back to Sonnet after verification (don't burn Opus tokens on the next task's implementation)

Verification failures:

• If gaps found → fix them before marking done (stay on Opus for the fix if it's complex)
• If can't fix quickly → mark task [partial] with notes on what's missing, add fix as new task

Step 5: Checkpoint + Mark Done (ENFORCEMENT GATE — CANNOT SKIP)

□ Git checkpoint: git add -A && git commit -m "done: [task description] — N tests passing" □ Queue Jinx follow-up BEFORE marking done: After EVERY completed task, ask "What can Jinx do now?" - Tests? Docs? Code review? Edge case analysis? Performance audit? - Jinx should NEVER be idle when follow-up work exists - Actually submit the Jinx task via curl NOW — not "later" □ Update NEXT_TASKS.md: Check the box [ ] → [x], add notes if needed □ RUN THE ENFORCEMENT GATE (MANDATORY — NO EXCEPTIONS):

  ~/.openclaw/workspace/scripts/enforce.sh "<project_dir>" "<task_name>" "<memory_file>"

- If it exits non-zero: you are NOT DONE. Fix every failure before proceeding. - Do NOT mark the task done, do NOT move to the next task, do NOT pass Go. - This gate checks: memory log exists with spec quote + verification + Jinx follow-up, tests pass, git clean, NEXT_TASKS.md updated - If the task revealed new work → add it to the appropriate priority section

Step 6: Log (Structured Format)

Update memory/YYYY-MM-DD.md using this format:

### Task: [name from NEXT_TASKS.md]
- **Status:** ✅ Done / ⚠️ Partial / ❌ Blocked
- **What was built:** [1-2 sentences]
- **Files changed:** [list]
- **Tests:** [before] → [after] (added N)
- **Verified:** [how — what you checked beyond tests]
- **Spec match:** [yes/no — quote spec if relevant]
- **Jinx queued:** [what follow-up work]
- **Surprises:** [anything unexpected, lessons learned]

Update STATUS.md if this was a significant milestone. Log to Mission Control if available.

Step 7: Skill Extraction Check

□ Did this task produce a reusable pattern? - New algorithm that could apply to other projects? - Debugging technique worth capturing? - Architecture pattern worth documenting? □ If yes → note it in the Learning Capture section of today's memory □ If it's substantial → create a skill (check existing skills first to avoid duplicates)

Step 8: Context Check + Continue

□ Run session_status — check context usage □ If context > 60%: 1-2 more tasks max. Be selective. □ If context > 70%: Finish current task, go to Phase 3 (wrap-up). STOP new work. □ If context OK: Move to next unchecked task, repeat from Step 1.

Phase 2: Blocker Handling

When you hit a blocker:

□ Classify: - Needs Ray: Decision, approval, info only Ray has - Technical: Bug, missing dependency, unclear spec - Dependency: Task X depends on unfinished Task Y

□ Handle by type: - Needs Ray: Mark [blocked: waiting on Ray — reason] in NEXT_TASKS.md, message Ray, skip to next task - Technical: Attempt to solve (switch to Opus if complex). If stuck >15 min, mark blocked, move on. - Dependency: Skip to the dependency task first, or skip both and note the chain

□ Never let one blocker stop all progress.

□ If ALL tasks blocked: - Log blockers clearly - Check for supporting work (tests, docs, refactoring, skills) - Message Ray with the full blocker list and what you did instead

Phase 3: Session Wrap-Up (MANDATORY)

□ NEXT_TASKS.md current: - All completed items checked [x] - New discovered items added - Blockers noted with reasons - "Next session starts here" marker on the first unchecked task

□ Final test run:

  python3 -m pytest -x -q --ignore=tests/integration

- 0 failures confirmed - Test count ≥ starting baseline

□ Final git checkpoint:

  git add -A && git commit -m "session end: [N] tasks done, [M] tests passing"

□ Structured handoff note in memory/YYYY-MM-DD.md:

  ## Session Wrap-Up [time]
  - **Tasks completed:** N (list them)
  - **Tests:** start count → end count
  - **Next task:** [first unchecked item]
  - **Blockers:** [list or "none"]  
  - **Warnings for next session:** [anything tricky, gotchas, or "none"]
  - **Jinx tasks queued:** [list or "none"]
  - **Context at wrap:** X%

□ Update STATUS.md with current test count and completion %

□ Queue Jinx overnight work if applicable

□ DO NOT leave work half-done. Finish or revert.

Progress Reporting to Ray

When to message Ray:

• After completing 3+ tasks in a session — batch summary, not per-task spam
• When hitting a blocker that needs Ray — immediately, with specific question
• When a milestone completes (e.g., entire Priority section done) — celebration-worthy
• When something unexpected happens — broken assumptions, spec contradictions, major discoveries
• At session wrap if significant progress was made

Don't message Ray for: routine task completion, passing tests, minor progress. Save it for batches.

Quality Gates (ENFORCEMENT)

A task is NOT done unless ALL of these are true:

• [ ] Code written and saved
• [ ] Tests written (if applicable)
• [ ] Full fast suite passes with 0 failures
• [ ] Output verified with Opus model (not just "tests pass")
• [ ] Spec requirement quoted and matched
• [ ] Git checkpoint committed
• [ ] NEXT_TASKS.md checkbox marked
• [ ] Jinx follow-up queued (if applicable)
• [ ] No regressions (test count ≥ baseline)

Red Lines (Never Cross):

• Never mark a task done without verification
• Never decrease test count
• Never start a new task with failing tests
• Never continue past 70% context without wrapping up
• Never leave NEXT_TASKS.md out of date
• Never implement without reading the spec requirement first
• Never skip verifying the previous session's work

Learning Capture

After completing any task, write one line:

• What worked — technique, pattern, or approach worth reusing
• What didn't — mistake, false start, or wasted time
• Skill candidate? — if this pattern is reusable, note it for potential skill extraction

If it's worth keeping long-term → update MEMORY.md at end of day.