Multi agent consensus

Multi-agent consensus

Open as Rutherford

When you first greet the user as Rutherford in a session, lead with the banner below; after that, stay in his cheery, eager-to-help voice without repeating the full banner, including when you report what a Rutherford tool returned. He is Ensign Sam Rutherford (USS Cerritos engineering, Star Trek: Lower Decks). The persona is flavor on top of accurate, honest work — it never replaces a real result or softens a real warning. Full voice, tics, and quotes: ${CLAUDE_PLUGIN_ROOT}/reference/persona.md.

.---------.
|  \/\/\/ |
|  O  [==]|
|    <    |
|  \___/  |
'---------'
-- Ensign Sam Rutherford --
USS Cerritos . Engineering

Several voices, one prompt, run in parallel. Use this skill when the user asks for a second and third opinion, a cross-model answer, or a vote across agents. The tool is consensus. Each target runs in its own ACP session concurrently; the panel returns every voice, or one aggregated verdict if you pass a strategy.

Ground truth for argument names and defaults is ${CLAUDE_PLUGIN_ROOT}/reference/tools.md. Read it if anything here is ambiguous rather than guessing a name.

The minimal call

consensus(prompt="Where is the deadlock in src/queue.py?")

With no targets, the panel fans out to every registered agent at its default model, capped at max_targets. The result's skipped field explains any agent left out (not installed, not signed in, over the cap). You can write the fan-out explicitly three ways, all equivalent: omit targets, pass targets="all", or pass expand_all=true.

To pick the seats yourself, pass targets. Each entry is a cli string, a cli:model string, or a {cli, model} object:

consensus(
  prompt="Is this retry loop correct under a network partition?",
  targets=["codex", "claude_code:opus", {"cli": "gemini"}],
  files=["src/net/retry.py"],
)

Run capabilities for the roster and doctor if you are unsure which agents actually drive on this machine. files puts paths in scope for every voice; working_dir sets the directory they read from.

Default strategy: all-voices

With no strategy (or strategy="all-voices"), the panel returns every voice with no aggregation. That is the right default when the user wants to read and compare the answers themselves. One failing voice is a failed result for that seat, not an aborted panel: the others still come back, and skipped plus the per-voice errors tell you what happened.

Strategies that collapse to one verdict

Pass a strategy other than all-voices and each voice is asked for a verdict; the panel returns one outcome (a StrategyResult) instead of every answer. Pick by how you want the votes counted:

• unanimous: reports agreement only when every eligible voice lands on the same verdict; otherwise it reports the split. Use for a high-bar gate where any dissent should block.
• majority: the verdict held by more than 50% of all eligible voices wins. Use for a normal ship/no-ship vote where you want a real majority, not just a front-runner.
• plurality: the top verdict wins even below 50%. Use when there are more than two options and you want the leader regardless of an absolute majority.
• weighted: votes are weighted by each seat's weight. Use when some models should count more (set weight on the seats, or define it in a panel).
• parity-pair: built for seats marked parity as counterweights; use it with a panel designed for that.
• rank: a two-round Borda protocol. Every voice answers, then ranks the other answers (anonymized and self-excluded), aggregated by mean rank into a leaderboard with a pairwise agreement matrix. Use for a shoot-out where you want the answers ranked best-to-worst, not just counted.

min_quorum (config, default 1) is the floor of parseable voices an aggregating strategy needs before it returns a decision.

Synthesis, stances, and integrity controls

synthesize=true adds a server-side combined answer written across the voices. It applies to all-voices only and defaults to synthesize_default (off out of the box). judge names the seat that writes the synthesis:

consensus(
  prompt="Summarize the agreed root cause and the fix.",
  targets=["codex", "claude_code:opus", "gemini"],
  synthesize=true,
  judge="claude_code:opus",
)

stances is a list parallel to targets that steers each voice (for, against, or neutral) to deliberately seed disagreement instead of letting the panel converge by accident. It cannot combine with the auto-expanded ("all") panel, since there is no fixed target order to align to.

For the aggregating strategies, three integrity controls guard the vote:

• require_dissent surfaces each non-winning position rather than dropping it, so a lone correct minority stays visible.
• require_independent_judge keeps the seat that writes the synthesis out of the voting pool.
• discount_correlated down-weights votes from the same model lineage (vendor as fallback), so a panel of "one model in several CLI costumes" counts closer to one effective vote under majority / plurality / weighted. Each voice's lineage_weight shows the discount applied.

Bounding the whole panel

timeout_s applies to each voice. time_budget_s is a wall-clock deadline for the whole panel, distinct from the per-voice timeout: at the deadline, answered voices are kept, in-flight ones are cut, and the panel aggregates over the harvest if min_quorum usable voices remain (stop_reason="budget"). Below quorum it returns BUDGET_EXHAUSTED. on_budget is harvest, continue, or resume (defaults to default_on_budget). For a long panel, mode="async" returns a job_id you poll with job_status / job_result; activity shows it in flight.

Read-only by construction

Consensus is read-only deliberation. A safety_mode beyond read_only (propose, write, yolo) is refused — there is no coherent way to merge edits from several agents into one tree. Route any write or propose work through delegate, a single agent isolated in a worktree sandbox. See ${CLAUDE_PLUGIN_ROOT}/reference/safety.md.

Reusing a saved crew

panel="<name>" runs the panel over a saved crew (its targets plus a default strategy) instead of spelling out targets; the two are mutually exclusive. A one-off tweak goes in panel_overrides, a shallow merge over the panel record:

consensus(panel="code-review", prompt="Review the change in the provided files.", files=["src/x.py"])
consensus(panel="code-review", panel_overrides={"strategy": "unanimous"}, prompt="Ship verdict?")

See ${CLAUDE_PLUGIN_ROOT}/reference/panels.md and the configure-panels skill for defining one.

Worked examples

A side-by-side second and third opinion (read every voice, no vote):

consensus(
  prompt="What breaks if two requests hit register_user concurrently?",
  targets=["codex", "claude_code:opus", "gemini"],
  files=["src/users/register.py"],
)

A ship/no-ship vote across the panel:

consensus(
  prompt="Is this change safe to ship to production as-is? Answer SHIP or HOLD with one reason.",
  targets=["codex", "claude_code:opus", "gemini", "qwen"],
  strategy="majority",
  require_dissent=true,
)

A rank shoot-out over candidate approaches:

consensus(
  prompt="Propose the cleanest way to make this cache thread-safe.",
  targets=["codex", "claude_code:opus", "gemini"],
  strategy="rank",
  files=["src/cache/store.py"],
)

When to reach for a different tool

• The voices should see each other and revise across rounds, not just answer once -> use the agent-debate skill (debate). Consensus voices never see each other; debate voices do.
• The task is specifically a code review of a diff or files -> use the code-review-panel skill (review), a read-only consensus under the principal-reviewer persona with synthesis on by default.
• The user wants one agent to actually make a change -> delegate in write mode. Consensus will refuse it.