forge-council — LLM Council (FULL protocol)

Use ONLY when councilTrigger() (forge-quality.cjs) says FULL: an explicit owner request for pushback, or high impact + high uncertainty. Never by default, never on simple tasks — the expected information gain must outweigh latency, cost and context load. LIGHT exists as an outcome of the trigger but deliberately has NO protocol here yet: first a benchmark that proves its value (anti-overengineering); until then LIGHT falls back to one targeted second opinion.

Protocol (FULL)

Step 0 — neutral framing. Write the decision question NEUTRALLY (no preferred direction, no "we lean towards X"): context, options, criteria, hard limits. Compute context_hash = sha256 of that framing text (forge-quality.cjs::sha256). The framing is a run artifact.

Step 1 — 5 fresh, independent advisors, in parallel. Dispatch via the Agent tool (real subagents, one message with 5 parallel calls; NO simulated agents). Each advisor gets ONLY the neutral framing — no visibility of the others. The five fixed lenses:

• Contrarian — attacks the most popular option; looks for what everyone is missing.
• First Principles — reduces to fundamentals; ignores convention.
• Expansionist — widens the option space; what if the question itself is too narrow?
• Outsider — looks at it as an outsider/end user would; jargon-free.
• Executor — judges feasibility, cost, risk, rollback.

Step 2 — anonymous peer review. Anonymize the five pieces of advice as A-E and send them back to each advisor (a second dispatch round): rank the other four and name the strongest critique of each. Anonymity prevents authority bias and self-preference.

Step 3 — chairman synthesis. The orchestrator (or a separate chairman dispatch) weighs the advice + peer rankings and writes the decision WITH an explicit minority report: dissenting opinions don't disappear, they're recorded alongside the reason.

Record. Fill in a CouncilDecisionRecord: council_id, context_hash, participants (role + runtime + dispatch_id from the REAL Agent dispatches + response_ref), responses, quorum, verdict, minority_report, status. Validate with validateCouncilRecord and persist via node .claude/forge-bin/forge-quality.cjs council-save <run_id> <record.json> — an invalid record is NOT written (the refusal with a reason is the honest result), and a council_id is never overwritten.

LIGHT candidate (recorded, NOT active)

The four-voice council of the pinned ECC toolkit (council skill, commit e2b8a51cea22) is recorded as the candidate LIGHT protocol. It stays inactive until a benchmark shows it changes decisions; until then LIGHT falls back to one targeted second opinion, as above. No step list is kept here on purpose.

Limits (permanent)

• Validation is shape_only: runtime/dispatch_id are not cryptographically bound; real provenance requires the owner-gated gateway dispatch receipt (OWNER-GATED.md).
• Consensus is never proof. Every council decision that touches code still requires ordinary tests/measurements afterward; the record points to them, never the other way round.
• An INCOMPLETE quorum is honestly marked INCOMPLETE — never padded out.