Research redteam loop

Research Red-Team Loop Skill

Purpose

Bundles the proven research→adversarial-review→rework→consensus cycle (validated in #2710: 78→92→91→94 ACCEPT) into a single invocation. Fleet/free-cloud only. Cross-family invariant: reviewer MUST NOT be Anthropic/Claude family.

Invocation

/research-redteam-loop #N [--gate 93] [--web] [--cap 5]

Arguments: --gate N = minimum acceptance score (default 93). --web = inject web search evidence (default on; degrades gracefully). --cap N = max iterations.

Loop Protocol

1. Produce/refresh the research deliverable for issue #N (read existing body or linked wiki page). 1a. Corpus-overlap stage (MANDATORY before ACCEPT — #2801). Run node scripts/global/corpus-overlap-scan.js --scope-file <deliverable> --json. It reuses the #2617 tokenizer but scans open AND closed issues (the near-miss #2518 was a closed epic) via overlap-coefficient. Embed buildOverlapBlock (related_tickets: + overlap_decision:) in the deliverable and inject reviewerOverlapFragment into the reviewer prompt so the red-team validates the boundary (overlap vs conflict vs redundancy vs gap). ACCEPT is blocked when shouldBlockAccept(scan, deliverable) is true (overlap present, no overlap_decision:). Distinct from #2617/Epic #2623 (ticket-CREATION stage).
2. Web augmentation: fetch 3+ relevant references via WebSearch; inject into reviewer prompt. On search failure, log web-search-unavailable to ~/.megingjord/incidents.jsonl and continue without web context.
3. Dispatch review (G3 fleet-first) — canonical CLI is cascade-dispatch.js (#2858 / Epic #2926 D1):
   • node scripts/global/cascade-dispatch.js --prompt "<review prompt>" --model qwen2.5-coder:32b [--json] — orchestrates fleet → free-cloud fallback → escalation; timeouts per timeout-policy.json. It already emits [cascade] escalate→<tier> to stderr on escalation.
   • On a fleet availability failure it fails over to the $0 free-cloud chain automatically (#2619/#2621); premium is reached only on a free-cloud capability failure (Epic #2926 D4).
   • fleet-red-team-dispatch.js is a LIBRARY (no CLI) — do NOT invoke it directly; it redirects here.
   • Cross-family invariant enforced at every tier (no Anthropic model).
4. Score check: if score > gate and shouldBlockAccept is false → post ACCEPT and stop. If overlap blocks (no overlap_decision:), record the boundary and re-review. Else → post ITER comment and proceed to rework.
5. Manager rework: revise deliverable using reviewer findings + web context.
6. Repeat steps 2–5 until ACCEPT or cap exhausted → post REJECT.

Verdict Comments (posted to target issue)

Per-iteration:

RESEARCH_REDTEAM_ITER: N | score=<N>/100 | reviewer=<model@host> | web=<yes|no>

Terminal ACCEPT:

RESEARCH_REDTEAM_ACCEPT: score=<N>/100 | reviewer=<model@host> | iterations=<N>

Terminal REJECT (cap hit):

RESEARCH_REDTEAM_REJECT: max_iterations_reached | final_score=<N>/100

Final verdict + per-goal vector recorded to ~/.megingjord/incidents.jsonl.

Infrastructure Used (AC6 — no new plumbing)

• scripts/global/fleet-red-team-dispatch.js — HAMR-wrapped Ollama dispatcher
• scripts/global/free-cloud-dispatch.js — $0 cloud chain (gemini/groq/cerebras)
• scripts/global/corpus-overlap-scan.js — deterministic overlap scan (open+closed; #2801)

Goal Lens

G3 Zero-Cost: fleet/free-cloud only; zero paid API calls during loop. G2 Quality: gated loop with calibrated scoring, bounded iterations, audit trail. G8 Observability: structured per-iteration comments + incidents.jsonl record.

Related

• skills/cross-family-review/SKILL.md — cross-family invariant contract
• skills/role-red-team-critique/SKILL.md — single-shot adversarial review
• config/red-team-model-matrix.yml — model selection matrix
• .claude/commands/research-redteam-loop.md — slash command entry point
• tests/research-redteam-loop.spec.js — eval-harness test suite