Ask several agents at once

Sometimes one agent's answer is not enough. /poll asks two to five agents the same question independently and shows how they vote. /council gives a roster of two to five agents the same task and can have a judge write one answer from theirs. Both are read-only: the members read your project, argue and vote, but never change a file.

Ask a quick question with /poll

/poll takes the number of participants, then the question. --options limits each vote to one of the listed answers, and --rationale asks for a short reason with each vote:

/poll --options node:test|vitest|jest --rationale 3 Which test runner should a small dependency-free ES module package like this one use?

Without --models, every participant runs on your chat model. --models litellm,big-box/qwen3-coder-next seats them in turn on the routes you list, so the answers can come from different models. Each participant answers alone: no one sees another's answer.

Approve the plan

A poll is a small fleet, so at the default autonomy level CLIO Coder shows the plan before anything starts. It names each participant, its agent and its model:

Plan ────────────────────────────── Alt+V shows every field
p1    scout on litellm · qwen3.8-27b
p2    scout on litellm · qwen3.8-27b
p3    scout on litellm · qwen3.8-27b
cost  counts against this session's $5.00 limit

Enter allows it, and Esc denies it without starting anyone. At yolo the plan runs without this stop. Participants run as the scout agent, and every member of a poll or council gets only read-only tools: read, grep, find, ls, git, code_nav and context.

Read the vote

Each participant appears with its time and the start of its answer. When all have voted, CLIO Coder prints the distribution:

poll       3 of 3 voted; majority: node:test
node:test  ### 3 (100%)

The headline reads majority when one answer has more than half the votes, plurality when it leads without a majority, and tie when the top two are level. Here all three read package.json, saw a node --test script and kept it. /view opens each full answer. A participant whose vote names none of the options is asked again, with the options quoted.

Declare a council roster

A council needs a roster. Without one, /council stops and says where to add it:

/council needs a roster. Declare one under fleet.rosters in settings.yaml, for example fleet.rosters.default with two to five members, each with a label and a target, then run /council <task> or name another roster with --roster <name>.

Each member has a label and a target, and can name a model and an agent. This roster seats one plain reader and one tester:

fleet:
  rosters:
    default:
      members:
        - { label: reader, target: litellm }
        - { label: tester, target: litellm, agent: tester }

The roster is read on the next command, without a restart. --roster <name> picks another roster. A member whose target runs on an SSH fleet node, or on an agent runtime that cannot enforce read-only tools, such as an ACP peer, is refused before you approve anything.

Ask the council, with a judge

/council --synthesis judge Is add() in index.js covered by a test, and what is the smallest test that would cover it?

The plan card listed reader as scout, tester as tester, and a synthesis run as oracle. The tester member kept its own way of working, but the council's read-only tools still applied, so it argued for a test without writing one. After both answered, the judge read their answers, checked them against the files and wrote one answer. It opened:

Both council members agree on the facts, and my own inspection confirms them:

It went on to say that index.test.js already imports add and checks add(2,2) against 4 under node:test, so the function is covered.

--synthesis vote tallies the members' votes instead, and none, the default, returns each answer as it is. --rounds 2 or 3 shows each member the others' previous answers before it answers again. --deadline 90s stops a member that runs too long, and the council returns the rest marked partial.

The answers stay out of your main conversation. /fleet share <runId> brings one member's answer into it; shared from the synthesis run, it brings the synthesis with every member's labelled answer.