council-review
Multi-model validation council — auto-validate plans, architecture changes, and PRs via validate-plan/review before executing
SKILL.mdskills/council-review/SKILL.md
GitHubCouncil of Experts — Multi-Model Validation
⚠️ Status (2026-07-15, ADR-0008 FIX). Command paths corrected
~/bin/→bin/(bin/validate-planis real and on PATH;~/bin/is empty.bin/reviewwas CUT 2026-07-27, ADR-0014 — review is Claude-only by decision, so this skill covers PLAN validation only.) Not yet provisioned here, and deliberately not rewritten:~/.claude/council.yaml(absent — the binaries do not read it), aclaude-fable-5wrapper (absent — Claude is the host), and the Codex reviewer (codexabsent). The real roster + config mechanism is a conclave/council design decision (ADR-0008, handoff item 1), not a prune edit — until then, treat the roster andcouncil.yamlreferences below as illustrative, and rely onbin/validate-plan's own honest exit codes (voted / unavailable / broken → exit 2, no verdict).Where that decision now lives (2026-07-17):
docs/contracts/three-project-cohesion.md— the canonical map of the Conclave·Tessera·pr-arbiter system. The roster/config question is Open decision D1 (routing home: do thesebin/wrappers collapse into conclave gateway calls?). Tension to carry into that decision — this skill is NOT shielded by anti-conflation guard (a). Guard (a) protects pr-arbiter's union-recall code review from conclave's "route, don't judge" null. But council-review's plan-VALIDATION path (a panel judging one plan → approve/revise) is select-best, not union-recall — the exact regime where conclave's null is a genuine caution, not a false blocker. So: a multi-model plan panel must justify itself against the select-best null (does fan-out+judge beat routing the plan to the one strongest reviewer?), whereas the multi-engine PR-review path inherits pr-arbiter's union-recall win. Two paths, two verdicts — do not let the code-review win launder the plan-validation panel.
When to Auto-Trigger
Plans (auto_validate_plans)
When you write a plan to ~/.claude/plans/, automatically validate it:
bin/validate-plan --threshold 2 ~/.claude/plans/<plan-file>.md
- 2+ of 3 approve → execute immediately
- 1 of 3 → surface reviewer feedback to user before proceeding
- 0 of 3 → revise plan, re-validate
Architecture Decisions (auto_review_architecture)
When making architectural changes (new services, API redesigns, database schema changes), run:
# Architecture review: use Claude Code's native /code-review (ADR-0014).
# bin/review was cut — it had no working backend and had never run.
PR Review (auto_review_prs)
Before marking a PR as done, run:
# Changed-file review: native /code-review (ADR-0014).
Configuration
Council behavior is intended to be configured in ~/.claude/council.yaml — not present here, and not read by the binaries yet (see Status banner). The config mechanism is pending the conclave design.
Chief of the Council
chief: claude-fable-5 — Claude Fable 5 (Anthropic's most capable widely-released
model, GA 2026-06-09) leads every panel as the chief: it reviews first and casts
the deciding synthesis. Claude is the host model (no bin/ wrapper). Chief-override config is pending the conclave design.
Reviewer Contexts
The chief leads each context, followed by the panel:
| Context | Default Reviewers | When |
|---|---|---|
plan |
Claude Fable 5 (chief), DeepSeek Pro, Codex, Gemini Pro | Before executing any plan |
review |
Claude Fable 5 (chief), DeepSeek Pro, Kimi | Code review, PR review |
architecture |
Claude Fable 5 (chief), DeepSeek Pro, Gemini Pro, Grok | System design, schema changes |
Threshold Rules
The threshold setting controls how many approvals are needed:
threshold: 2with 3 reviewers → need 2/3 to auto-execute- Clamped to [1, reviewer_count] — can't be 0 or exceed available reviewers
Model Inventory
All 13 tiers are listed in ~/.claude/council.yaml under models:. Each has:
id— unique identifiercmd— CLI command to invoke (null for Claude models, which are the host)tier— routing priority (0=cheapest, 12=most capable)label— human-readable name
Use POST /api/models/health to verify all models are responding.
How This Skill is Used
This skill is loaded by Claude Code on session start. It provides the behavioral rules for when to invoke multi-model validation. The actual execution happens via bin/validate-plan (real, on PATH). Code review is NOT
in scope here — ADR-0014 declared it Claude-only and cut bin/review; use the native
/code-review. What survives is the multi-model PLAN validation council, kept deliberately
per ADR-0007 pending the conclave design.
Do not skip council validation for CLAUDE-tier tasks. The whole point is that architecture and security decisions get independent verification before execution.