council-roster

Show the project's current model-council roster. This is a read of .sdlc/model-council/, not a new assessment — no adapter is called, no tokens are spent, unless rebuild is given (still token-free, see step 3).

Arguments: $ARGUMENTS — optional literal rebuild.

Do this:

1. No roster yet? If .sdlc/model-council/roster.json doesn't exist, say so plainly and stop: point at /sdlc-model-council:council-commission (full onboarding: discover, characterize, audition, roster, policy) or /sdlc-model-council:council-assess (just the assessment) as the two ways to produce one. Don't fabricate a roster from priors here — that's what commission's skip audition choice is for.

2. Roster exists — present it. Show .sdlc/model-council/roster.md as-is (it already has the per-model table: grades, per-dimension posterior/n/ci95, roles, cost/item, p50 latency). Then read .sdlc/model-council/diversity.json and summarise its pairs array:

   • Most decorrelated pair — lowest both_wrong_rate among pairs where insufficient is false — this is the best casting complement.
   • Most correlated pair — highest both_wrong_rate (or any pair marked insufficient, which cast.py treats as maximally correlated by design — call these out separately, they're a data gap not a diversity finding). Reference the council-policy skill if the user asks what any of this means for routing — role assignment and casting rules live there, not in this command.

3. rebuild — re-derive the roster and diversity map from the existing .sdlc/model-council/results.jsonl without dispatching any model. This is for after a manual edit to results.jsonl, a pricing refresh, or a priors update — cheap and token-free because it only re-scores rows already on disk. Run:

   ${CLAUDE_PLUGIN_ROOT}/scripts/council/roster.py \
     --results .sdlc/model-council/results.jsonl \
     --priors-dir ${CLAUDE_PLUGIN_ROOT}/scripts/council/priors \
     --pricing ${CLAUDE_PLUGIN_ROOT}/scripts/council/pricing.json \
     --out-json .sdlc/model-council/roster.json \
     --out-md .sdlc/model-council/roster.md
   
   ${CLAUDE_PLUGIN_ROOT}/scripts/council/diversity.py \
     --results .sdlc/model-council/results.jsonl \
     --stack ${CLAUDE_PLUGIN_ROOT}/assessment/stack/v1/stack.json \
     --out .sdlc/model-council/diversity.json
   

   Then present the rebuilt roster per step 2. If .sdlc/model-council/results.jsonl doesn't exist either, there's nothing to rebuild from — say so and point at council-assess/council-commission instead.

4. Staleness check. Compare the roster's generated_ts (.sdlc/model-council/roster.json) against the timestamps in .sdlc/model-council/assessment-log.jsonl — if any log entry's ts is newer than roster.json's generated_ts, the on-disk roster predates the latest assessment run. Flag this plainly and suggest rebuild (if results.jsonl already reflects the newer run) or a fresh /sdlc-model-council:council-assess (if it doesn't).