run-er

name	run-er
description	Run Expert Review (MoE critics + ScrumMaster) on any architecture that has an existing analysis. Four execution modes — partial_parallel (default/recommended), sequential (max cross-reference), parallel (fastest), auto (complexity-adaptive). Supports full_moe (5 critics + SM) or single-critic re-runs. Streams live critic results with scores, gaps, and final confidence. ~60–120s for full_moe.
allowed-tools	Bash(python3:*) Bash(source:*)
compatibility	Requires API server running (./scripts/api/api_start.sh) and an existing analysis (ground_truth.json) for the target architecture.

run-er — Expert Review Runner

Streams the expert review pipeline live: each critic result prints as it arrives — score, validation status, top gaps, and confidence adjustment. Handles auth, SSE parsing, and cancellation on Ctrl-C automatically.

Run

# 3 core critics — partial_parallel (default, recommended)
cd "$(git rev-parse --show-toplevel)" && source .venv/bin/activate \
  && python3 .claude/skills/run-er/scripts/run-er.py 04_zero_trust

# Full MoE — 5 critics + ScrumMaster (needed for Plan-Actionable TATB score)
cd "$(git rev-parse --show-toplevel)" && source .venv/bin/activate \
  && python3 .claude/skills/run-er/scripts/run-er.py 04_zero_trust --full

# Full MoE with sequential mode (max cross-reference — Red Team reads Tester)
cd "$(git rev-parse --show-toplevel)" && source .venv/bin/activate \
  && python3 .claude/skills/run-er/scripts/run-er.py 04_zero_trust --full --mode sequential

# Run critics in parallel (all 3 core critics fully blind, fastest)
cd "$(git rev-parse --show-toplevel)" && source .venv/bin/activate \
  && python3 .claude/skills/run-er/scripts/run-er.py 04_zero_trust --mode parallel

# Re-run a single critic on an existing MoE result
cd "$(git rev-parse --show-toplevel)" && source .venv/bin/activate \
  && python3 .claude/skills/run-er/scripts/run-er.py 04_zero_trust --critic red_team

# Run ScrumMaster only (needs 07_moe_orchestrator.json from a prior ER run)
cd "$(git rev-parse --show-toplevel)" && source .venv/bin/activate \
  && python3 .claude/skills/run-er/scripts/run-er.py 04_zero_trust --critic scrum_master

Modes

Mode	Critic ordering	SM	Use when
partial_parallel (default)	Architect ∥ Red Team (blind); Tester reads Architect after; Purple+BH sequential	optional	Standard full_moe runs — reduces Red Team anchoring bias
sequential	Architect → Tester (reads Arch) → Red Team (reads Tester) → Purple → BH	optional	Deep review where Red Team catching Tester gaps matters
parallel	All 3 core critics fully blind simultaneously	optional	Speed-only or independent-view comparison
auto	Complexity score ≥60 → sequential; <60 → partial_parallel	optional	Let harness decide based on architecture size

--full adds Purple Team, Blackhat, and ScrumMaster (needed for Plan-Actionable TATB). Without --full, only Architect/Tester/Red Team run.

Decision guidance

• Use partial_parallel (default) for most runs — Architect and Red Team run concurrently so Red Team is not anchored on Tester's framing at LLM time. Tester still reads Architect. Purple Team and Blackhat always have full prior context.
• Use sequential when you want the Red Team to specifically build on gaps the Tester identified — more correlated output but higher anchoring risk.
• Use parallel for speed comparisons or when you want three fully independent perspectives before synthesis.
• Use auto when you're unsure — the harness picks based on architecture complexity (node × edge × path × technique count).

What it shows (live)

▸ Architect
  ⚠️  Architect     score= 78  adj=-5.0%  gaps=5  strengths=6
      ▸ [HIGH] Validation failure blocks assessment...
      ▸ [MEDIUM] WAF and INPUT VALIDATION are web controls...

▸ Tester
  ✅  Tester        score= 98  adj= 0.0%  gaps=0  strengths=6

▸ Red Team
  ⚠️  Red Team      score= 42  adj=-3.0%  gaps=5  strengths=5

──────────────────────────────────────────────────────────────
  Final confidence:  68.4%  (-11.0pp)  ████████████████░░░░░░░░
  Critical items:    5
  High-priority:     6

Prerequisites

1. API server running: ./scripts/api/api_start.sh
2. ground_truth.json exists for the architecture (run analysis first)
3. For --critic scrum_master: 07_moe_orchestrator.json must exist (run ER first)

Failure fixes

Error	Fix
API server not running	./scripts/api/api_start.sh
No analysis found for '<arch>'	Run analysis from dashboard or python3 -m chatbot.main --gen-arch-truth
Prerequisites missing: 07_moe_orchestrator.json	Run run-er <arch> (without --critic) first
TM-API-KEY not found	Add API_KEY=<key> to .env
HTTP 400 / invalid critic_mode	Valid modes: sequential, partial_parallel, parallel, auto
Synthesis hangs at 97%	Bedrock latency — wait up to 60s; if stuck add drop_params=True to litellm call

Notes

• --full uses the --mode flag (default: partial_parallel) — Purple Team and Blackhat always run sequentially with full context regardless of mode; the mode only affects the 3 core critics (Architect/Tester/Red Team). Pass --mode sequential explicitly if you want Red Team to read Tester before running.
• ScrumMaster generates 08_scrum_master.json and unlocks the Plan-Actionable rubric in TATB. Without it, Plan shows N/A.
• Labels are preserved — demo_expert_llm.sh stashes expected_threats.json before cleaning; run-er calls the API directly so labels are never touched.
• Ctrl-C cancels cleanly — sends DELETE to /api/v1/expert-review/cancel.

Related skills

• /tatb-score — check TATB rubric scores after running ER
• /tatb-loop — automated observe→fix→verify loop targeting weak signals
• /aivss-score — AIVSS governance scores
• /quick-test — verify API + deps before a session