1.9 KiB

Raw Blame History

Read-Only Subagent Review Protocol

Use this protocol when Lead/Orchestrator wants parallel review help for a domain pack, semantic replay, or repair handoff.

Rule

Subagents are tools, not owners.

Lead/Orchestrator owns:

final verdict;
issue_code selection;
repair decision;
merge of findings;
code changes;
autorun save;
acceptance.

Subagents must not:

edit code;
write artifacts except their own returned summary if the caller stores it;
save autoruns;
mutate docs/orchestration/active_domain_contract.json;
mutate contracts, prompt registry, or capability mapping;
mark a run accepted.

Roles

business_answer_reviewer

Reads user-facing output.md style artifacts first.
Judges direct-answer-first behavior, business usefulness, and technical garbage.

route_capability_reviewer

Reads turn.json, debug payloads, capability traces, and route candidate traces.
Judges route family, exact capability, missing axes, and wrong-domain leakage.

evidence_field_truth_reviewer

Reads evidence payloads, step state, scenario state, and output only for claim comparison.
Judges field truth, dates, amounts, selected object continuity, and carryover.

regression_pack_reviewer

Reads issue catalog, rerun matrix, repair targets, pack state, and accepted-pack context.
Suggests reruns and smoke coverage for a proposed fix.

Expected Summary Shape

Each subagent returns JSON only:

{
  "reviewer": "business_answer_reviewer",
  "status": "accepted|partial|blocked|needs_exact_capability",
  "issue_codes": [],
  "root_layers": [],
  "evidence_paths": [],
  "findings": [],
  "minimal_patch_direction": null,
  "required_reruns": []
}

Lead/Orchestrator converts useful findings into:

business_audit.json;
issue_catalog_snapshot.json;
detector_candidates.json;
rerun_matrix.json;
lead_coder_handoff.json.