/council-stats

Display Council performance metrics from the persistent pattern memory file (council-memory.json in the workspace folder).

How to read the data

1. Use the Read tool to load council-memory.json from the workspace directory
2. If the file doesn't exist, report: "No deliberation history yet. Run /deliberate to get started."
3. Parse the JSON and compute the requested metrics

Subcommands

If no arguments or summary:

Show the quick dashboard. Compute from the memory file:

• Total deliberations: length of deliberations array
• Total cost: sum of all cost_usd values
• Average cost: mean cost_usd
• Average latency: mean latency_ms, displayed in seconds (e.g., "12.4s")
• Average confidence: mean confidence, displayed as percentage (e.g., "82%")
• Top categories: count deliberations per query_features.category, show top 3
• Rating breakdown: count each user_rating value (positive, negative, neutral, null)
• Avg deliberation value added: mean of all non-null deliberation_value_added scores, displayed as percentage
• Router status: check routing_model:
  • If last_rebuilt is null and sample_size < 20: "Learning ({N}/20 samples)"
  • If last_rebuilt is not null: "Active ({N} rules, accuracy: {X}%, last rebuilt {date})"
  • If sample_size is 0: "Not started — run deliberations to train"
• Routing breakdown: count deliberations by routing_decision (deliberate vs single_model) and routing_source

Format as a clean text dashboard — no raw JSON.

If models:

Show the model performance leaderboard from model_performance. For each model, display:

• Average rank (sorted best to worst)
• Win rate as percentage
• Chairman selections count
• Self-preference rate as percentage (flag models with rate > 15% with a warning marker)
• Best categories (avg rank <= 1.5 in that category)
• Worst categories (avg rank >= 3.5 in that category)

Format as a table. Highlight the top performer.

If history:

Show the last 10 deliberations from the deliberations array (most recent first). For each, display:

• Record ID
• Relative time ("2m ago", "1h ago", "yesterday")
• Query summary
• Chairman model (short name, e.g., "claude-sonnet-4" not the full path)
• Confidence as stars (★★★★★ = 0.9-1.0, ★★★★ = 0.7-0.89, ★★★ = 0.5-0.69, ★★ = 0.3-0.49, ★ = 0-0.29)
• Cost
• User rating indicator: (+) for positive, (-) for negative, (~) for neutral, blank for unrated

If cost:

Show cost breakdown:

• Total spend: sum of all cost_usd
• This session: sum of deliberations from today (match date in timestamp)
• By model count: group deliberations by length of models_used, show avg cost and count for each group (2, 4, 6 models)
• Estimated stage split: Stage 1 ~42%, Stage 2 ~38%, Stage 3 ~20% (these are estimates based on the cost model in deliberation-engine; individual stage costs aren't tracked separately)
• Monthly spend: sum costs for the current calendar month
• Budget remaining: if a budget cap is configured in /council-config, show remaining. If not configured, show "No budget cap set"

If optimization:

Show optimization history from the optimization_log array:

• Last optimization run: date, experiment count, baseline → final pass rate
• Total optimization spend: sum of all experiment cost_usd values
• Changes kept / reverted / skipped: counts by decision
• Recent experiments (last 5): experiment_id, variable_modified, modification_type, pass rate delta, decision
• Current prompt version: note that stage-prompts.md reflects all kept changes

If no optimization history exists, report: "No optimization runs yet. Rate 3+ deliberations, then run /council-optimize to start improving Council's prompts."