메뉴
Require structured candidate-result JSON from swarm leaves so critique_candidates.py and synthesis run on evidence, not prose self-reports.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Keep coordinator and leaf context lean-batch reads, cap logs, use execute_code to filter-so swarms stay fast and within token limits.
Use at the start of non-trivial software tasks. Gather repo ground truth with batched reads and searches before planning or delegate_task.
Obey Plan 1A hook session phases so mutating tools, delegate_task sizing, and re-planning do not waste turns or trigger blocks.
Use when spawning delegate_task leaves for coding work. Pack self-contained goals and context so parallel Composer workers ship verifiable patches without coordinator chat history.
Enrich slim hook-injected delegate_tasks with touch map, failure excerpts, and verification commands so parallel leaves produce verifiable patches without coordinator chat history.
After failed candidates or a red ship gate, spawn a narrow focused delegate_task wave with failure evidence instead of re-running a full-width council.
| name | heavy-leaf-candidate-output |
| description | Require structured candidate-result JSON from swarm leaves so critique_candidates.py and synthesis run on evidence, not prose self-reports. |
| version | 0.1.0 |
| author | CodeGraphTheory |
| license | MIT |
| metadata | {"hermes":{"tags":["heavy-coder","delegation","schemas","effectiveness"],"related_skills":["heavy-leaf-brief","heavy-synthesize-winner","heavy-team-default"]}} |
Parallel leaves return competing implementations. Prose summaries force the coordinator to eyeball eight narratives. Structured JSON enables deterministic ranking (scripts/critique_candidates.py) and honest synthesis (heavy-synthesize-winner).
delegate_task entries (coordinator)context (paste Return format section).heavy-coder/evidence/<child>.json from subagent_stop hookEach leaf must end with JSON matching schemas/candidate-result.schema.json:
{
"candidate_id": "c1",
"role": "minimal-fix",
"commit_sha": null,
"changed_files": ["path/to/file.py"],
"tests": [
{"command": "pytest tests/test_x.py -q", "exit_code": 0, "summary": "1 passed"}
],
"assumptions": [],
"residual_risks": ["edge case Y not covered"],
"confidence": 0.75
}
Rules for leaves (put in every context):
tests[].exit_code must be the real exit code from terminal-never omit or invent.changed_files lists every path touched.confidence is 0.0-1.0; lower when tests were skipped.residual_risks and set confidence ≤ 0.3.Gather evidence files under .heavy-coder/evidence/ when present.
Validate:
python skills/heavy-issue-to-merge/scripts/validate_candidate.py path/to/candidate.json
Rank:
python scripts/critique_candidates.py path/to/c1.json path/to/c2.json ...
Proceed to heavy-synthesize-winner using rank + schema errors, not eloquence.
Done when: winner choice cites candidate_id, test exit codes, and file list.
Append to each leaf context:
Return format (mandatory): Final message must include a fenced JSON block matching
schemas/candidate-result.schema.json with candidate_id, role, changed_files, tests
(with exit_code), residual_risks, confidence. No "tests passed" without a command and exit code.
| Symptom | Fix |
|---|---|
| All candidates score low in critique | Missing or invalid JSON |
| Coordinator trusts "green" summary | Re-run ship gate on merged tree |
| Schema validation fails | Fix fields; do not discard ranking entirely |