council-run

Run a council play — a cross-model fan-out — over the given task. v1 ships one play, Diff+Synthesis: the same task goes to a decorrelated panel of models, and the council-judge agent synthesises their responses into one attributed verdict with a baseline delta (did the panel beat the single best model?). This is the novel work no per-vendor plugin does.

Arguments: $ARGUMENTS — parse task-type=, input= (a file path or inline text), optional play= (default diff-synthesis), cast= (pin explicit model addresses, comma-separated), budget-usd=, k= (default 3).

Scripts live under ${CLAUDE_PLUGIN_ROOT}/scripts/council/. Do this:

1. Resolve policy. Load the council-policy skill. If the project is commissioned (.sdlc/model-council/policy.json exists), read the entry for task-type to get the play, the cast (pinned addresses or a cast_rule), and the budget guardrail. If NOT commissioned, say so plainly ("this project isn't commissioned; using heuristics") and fall back: play = diff-synthesis, cast = built live from .sdlc/model-council/roster.json if present, else ask the user for cast=. An explicit cast=/play=/budget-usd= argument always overrides policy — note that you did.

2. Materialise the task. Write the task to a file (if input= is a path, use it; if inline text, write it to a scratch file under ./tmp/).

3. Set up the play (deterministic spine, no models yet):

   scripts/council/play.sh setup --dimension <task-type> --k <k> \
     --roster .sdlc/model-council/roster.json \
     --diversity .sdlc/model-council/diversity.json \
     --task-file <task-file> [--cast <pinned>] [--budget-usd N]
   

   It prints PLAY_DIR=… and a dispatch table of label<TAB>model<TAB>result-file. The manifest records the baseline_member (the roster's best single model for this task) — that's what the synthesis is measured against.

4. Fan out — one delegation-runner per member, in a single message (respect the cap of 5). For each dispatch-table row, parse the model address adapter:model[@effort] into backend / model / effort and dispatch the delegation-runner agent with that backend, the same task (from task.md), and posture=read-only. Each runner returns its unified block with an Answer file path. Do NOT read the answers into your context.

5. Collect on disk (no context inheritance). For each member, copy its runner's Answer file to <PLAY_DIR>/<label>.result.md with cp — the answer text goes to disk under its blind label, never through your context. A member that ERROR/TIMEOUT/NO_OUTPUT'd has no usable answer — leave its result file absent (it becomes a non-survivor).

6. Check quorum:

   scripts/council/play.sh combine-check --play-dir <PLAY_DIR>
   

   If verdict is PLAY-DEGRADED (survivors < quorum), report that honestly and stop — a panel that couldn't reach quorum must not be dressed up as a confident synthesis.

7. Synthesise (blind). Dispatch the council-judge agent (Sonnet) with the PLAY_DIR and the manifest's baseline_label. It reads combine/blind-bundle.md + task.md (models anonymised as Model A/B/…) and writes combine/synthesis.md with Convergent / Divergent(attributed) / Adjudication / Confidence / Baseline delta sections.

8. Un-blind and report.

   scripts/council/play.sh unblind --play-dir <PLAY_DIR> --in <PLAY_DIR>/combine/synthesis.md
   

   re-attaches the real model addresses (provenance restored). Present that synthesis to the user, then append a spend line — Council spend: $X across N models — summing each member's cost via scripts/council/usage.py <handle-dir>. Finally append one line to .sdlc/model-council/outcomes.jsonl: {"play":"diff-synthesis","dimension":"<task-type>","cast":[…],"baseline_member":"…","ts":"…"} (the measurability ledger; whether the user accepted or reverted the result is recorded later, out of scope for v1).