$council — Consensus Review (Codex Tailoring)

This override captures the Codex-native execution model for council-style judging.

Modes

• --quick: inline review, no sub-agents
• default: 2 judges via spawn_agent(...)
• --deep: 3 judges via spawn_agent(...)
• --mixed: 3 runtime-native judges via spawn_agent(...) plus 3 Codex CLI judges via codex exec

--mixed is a strict cross-vendor contract. If codex is missing, codex --version fails, or a requested COUNCIL_CODEX_MODEL cannot run, stop before spawning any judges and tell the operator to install/fix Codex CLI or drop --mixed. Never silently convert --mixed into runtime-native-only judging.

Codex-Native Flow

Step 1: Build the review packet

Collect the target files, diff, plan, or report the judges need to review. Keep the packet concise and give each judge one perspective such as correctness, completeness, or edge cases.

Step 2: Spawn judges

Use one spawn_agent(...) call per judge. Use agent_type="default" unless the task is purely exploratory, in which case agent_type="explorer" is acceptable.

Each judge prompt should include:

• the review target
• the assigned perspective
• the expected verdict vocabulary
• where to write a detailed report if the task is large

For --mixed, also start 3 Codex CLI judge processes after the strict pre-flight succeeds:

mkdir -p .agents/council
codex exec -s read-only -C "$(pwd)" --output-schema skills/council/schemas/verdict.json -o .agents/council/codex-1.json "$PACKET"
codex exec -s read-only -C "$(pwd)" --output-schema skills/council/schemas/verdict.json -o .agents/council/codex-2.json "$PACKET"
codex exec -s read-only -C "$(pwd)" --output-schema skills/council/schemas/verdict.json -o .agents/council/codex-3.json "$PACKET"

Only add -m "$COUNCIL_CODEX_MODEL" when that environment variable is explicitly set.

Step 3: Wait and follow up

Use wait_agent(...) to collect judge results. If one judge needs clarification, use send_input(...) with a narrow follow-up question rather than restarting the council.

Step 4: Consolidate inline

The lead agent reads the judge outputs and produces the final consensus:

• PASS when no judge finds a ship blocker
• WARN when concerns exist but no hard blocker exists
• FAIL when any judge finds a blocking defect

Core consensus rules: All PASS -> PASS; Any FAIL -> FAIL; Mixed PASS/WARN -> WARN; cross-vendor disagreement -> DISAGREE.

Step 5: Close sub-agents

Use close_agent(...) on finished judges once their outputs are integrated.

Constraints

1. Do not assume batch-spawn or worker-report APIs that are not present in this runtime.
2. Keep judge outputs small in the lead context; large analysis should live in files.
3. Prefer stable judge perspectives over decorative personas.