weighted-consensus

name	weighted-consensus
description	Multi-model council — fan a question out to several models in parallel, then synthesize one weighted verdict with a minority report. Use with /weighted-consensus to decide, review, or compare options across diverse models.
argument-hint	<question> [--models a,b,c] [--context @file]

Weighted Consensus — multi-model council

/weighted-consensus runs an independent model council: it asks the same question to several models in parallel (each with its own role), then a separate synthesizer model reasons over their answers — using each member's weight as a prior, not a vote tally — to produce one final verdict plus a minority report.

When to use

• You want a higher-confidence decision than a single model gives
• Reviewing a change, comparing approaches, or making an architecture call
• You want dissent surfaced instead of averaged away

How it works

1. Fan-out — each member runs as an independent copilot --model <X> process (parallel, capped concurrency). Members never see each other's answers.
2. Structured output — each member returns JSON wrapped in <<<JSON>>> … <<<END>>> sentinels: verdict, confidence, rationale, optional risks/dissent.
3. Synthesis — a low-context synthesizer model merges survivors into a final verdict, confidence, rationale, and minority_report. Evidence quality can override weight.
4. Graceful degradation — errored, unavailable, or unparseable members are dropped; timed-out members that produced valid JSON are recovered as survivors. The council synthesizes from all survivors and only fails if fewer than min-survivors (default 2) remain.

Agent execution steps (FOLLOW EXACTLY)

When /weighted-consensus is invoked, you (the agent) MUST:

Step 1 — Build the question

Take the user's text as the <question>. If they referenced a file, a diff, or prior conversation, capture it as shared context (write it to a temp file and pass --context @<file> so large context isn't mangled on the command line).

Step 2 — Run the council command

Run the engine via the shell. It manages the parallel model calls, parsing, degradation, and synthesis for you — do NOT try to call models yourself:

omp council "<question>" [--models a,b,c | model:role:weight,...] \
  [--context <text|@file>] [--rubric <text|@file>] \
  [--synth <model>] [--probe] [--timeout <ms>] [--synth-timeout <ms>] \
  [--min-survivors <n>] [--max-concurrency <n>] [--tmp-dir <dir>] [--json]

• --models — inline roster override. Bare tokens (a,b,c) get round-robin default roles (critic, architect, pragmatist); long-form model:role:weight sets all three. If omitted, the default roster from .omp/config.json (or the built-in default) is used.
• --context @file / --rubric @file — read context or an evaluation rubric from a file (prefer this over inline for anything multi-line).
• --probe — opt-in preflight that prunes models your plan cannot access (off by default; the normal run already reports unavailable models inline).
• --json — emit the full structured result for programmatic use; omit it for a readable verdict + dropped-member summary.
• --synth-timeout <ms> — synthesizer timeout (default: 2× --timeout). The synth processes all member outputs so may need longer than individual members. The synth prompt does NOT include the shared context (members already digested it), so this is mainly needed for councils with many members or long rationale.
• Timed-out members that produced valid JSON before the kill signal are automatically recovered as survivors — they are not dropped.

If omp council is not found (the published omp predates this feature), update the global CLI: npm i -g @damian87/omp@latest, then re-run omp council. (The node dist/src/cli.js council build only resolves from inside the oh-my-copilot repo, so it is not a portable fallback for an arbitrary project.)

Step 3 — Present the result

Report the synthesizer's verdict, confidence, and rationale, then ALWAYS surface the minority_report if non-empty (it exists to stop dangerous edge cases being voted away). Note any dropped members and why. Point to the artifacts dir for the raw per-member JSON if the user wants to inspect individual answers.

Example invocations

/weighted-consensus "Should this service use gRPC or REST?"
/weighted-consensus "Is this migration safe to ship?" --models gpt-5-mini,claude-sonnet-4.6
/weighted-consensus "Review the staged diff for blockers" --context @/tmp/diff.txt --rubric @/tmp/rubric.md

Configuration

The default roster and synthesizer live in .omp/config.json under a council block; inline --models overrides it. If no config is present, a built-in default roster is used (all gpt-5-mini — the included model every plan can run).

Run omp models first to see which slugs your plan actually supports, then list those (the example below mixes the included GPT tier with a Claude member for cross-provider diversity):

{
  "council": {
    "synthesizer": "gpt-5-mini",
    "minSurvivors": 2,
    "maxConcurrency": 4,
    "synthTimeoutMs": 240000,
    "probe": false,
    "members": [
      { "model": "gpt-5-mini", "role": "critic", "weight": 0.4 },
      { "model": "claude-sonnet-4.6", "role": "architect", "weight": 0.35 },
      { "model": "gpt-5-mini", "role": "pragmatist", "weight": 0.25 }
    ]
  }
}

Notes

• Diversity comes from the model mix and per-role prompts — the Copilot CLI headless surface exposes no temperature/sampling control.
• Per-member JSON artifacts are written under a temp directory for debugging and are not auto-cleaned.
• For genuine independence, prefer distinct models. If your plan only includes the cheap GPT tier (e.g. gpt-5-mini), distinct ROLES still add diversity; add a different-provider model (e.g. a Claude or Gemini model) if omp models shows your plan supports one.

Composition

Pair with /code-review (pass a review rubric + the diff as context) for a multi-model review verdict, or with /ralplan to weigh competing plans.

Cost/token note

This skill can drive multiple tool calls or long-running output. Use omp cost [--today] [--session <id>] for local hook-ledger estimates only; it is not provider billing. Keep injected summaries concise and prefer bounded output when rerunning noisy commands.