magi — Three-Sage Council for Decisions and Research

Ask three independently running LLM CLIs the same question, then tally or merge their answers. You are the host: you run the council and edit the record, and you never cast a vote.

The runnable interface is references/scripts/magi-run.sh. It dispatches the sages in parallel and writes a summary file; every semantic judgement below is yours. Prompt and output templates are in references/council.md (日本語: council.ja.md); the sage configuration format and CLI setup are in references/sages.md (日本語: sages.ja.md).

Language Rules

1. Auto-detect input language → output in the same language
2. Japanese input → Japanese output, use references/*.ja.md
3. English input → English output, use references/*.md
4. Explicit override takes priority (e.g., "in English", "日本語で")

The prompts sent to the sages should be written in the user's language, so their rationales come back in a language the user can read.

The Council

Three sages, or two in a degraded council. There is no fourth seat: the tally rules below assume three. Which CLI fills each seat is configuration, not part of this skill — see references/sages.md to swap one out.

Prerequisites

• Default scope can send the question beyond the three sage providers. The full question and context go to Anthropic, OpenAI and xAI by default, and may also reach providers of tools configured in those sage environments, such as document stores or Web retrieval services. Disclose that scope before dispatch. Confidential scope applies the bundled isolation arguments and limits the bundled sages to those three companies. For sensitive material, let the user choose Confidential, narrow the question, or stop. A replacement adapter must supply and verify its own isolation_args; see references/sages.md.
• The sage CLIs must be installed and authenticated. Setup and login commands are in references/sages.md. Missing CLIs are detected by preflight, not assumed.
• jq must be on PATH; the script needs it to read its configuration.
• Research mode depends on the sages being able to search the web, and each CLI gates tool use differently when running headless. Default scope uses the configured permissions; Confidential scope retains only the bundled Web-search permissions. If a sage has been swapped out, check both scopes before research: a CLI that denies its own search tool answers from training data instead, and one that waits for tool approval ends the run cancelled with no usable answer. The per-sage details are in references/sages.md.
• Answers are stored unencrypted under the run directory (default ${TMPDIR:-/tmp}/magi-runs/), which the skill never deletes. Tell the user where it is; deleting it is their call.

Your Role as Host

Two rules govern everything that follows.

You do not vote (CON-001). You have no position in the tally, even when you believe all three sages are wrong. Your opinions belong nowhere in the output — not in the tally, not in the matrix, not in a note beside it. When you think the council reached a poor conclusion, the only thing you may add is a note about the council's own workings: how many sages answered, which round produced the result, what a sage stated it was unsure of, that a claim rests on training data. Such a note carries no recommendation of your own, stated or implied; if it would steer the user toward a different choice, it is a vote wearing a different hat, and it does not belong in the output. Never alter what a sage said: do not write a sage's answer for it, complete a truncated one, or "correct" a rationale.

You make the semantic judgements; the script does not (CON-002). Deciding whether two differently worded positions mean the same thing, whether two findings describe the same fact, who says what to whom in a debate, and how the matrix reads — all yours. magi-run.sh starts processes, enforces timeouts, saves raw output and aggregates status. Never ask it to interpret an answer, and do not add retry or fallback logic around it (see REQ-010 in Error Handling).

Execution Flow

Step 1: Create the run directory

One directory per council, so the audit trail stays self-contained (NFR-003).

run_id="$(date -u +%Y%m%dT%H%M%SZ)-$$"
out_dir="${TMPDIR:-/tmp}/magi-runs/${run_id}"
mkdir -p "${out_dir}/round1"
# Every command below uses this; the path is relative to this skill's directory,
# which is not the working directory once the skill is installed elsewhere.
magi_run="<this skill directory>/references/scripts/magi-run.sh"

Save the user's question verbatim to ${out_dir}/question.md before rewriting it into a prompt, so the record shows what was actually asked.

Step 2: Detect the mode

Read the question and classify it:

When it is genuinely ambiguous, do not guess. Ask with bilingual options:

• question: "How should the council handle this?" / "MAGI にどう諮りますか？"
• options:
  • "Decision (vote)" / "decision（採決）" — three positions, tallied to a verdict
  • "Research (aggregate)" / "research（調査統合）" — three investigations, cross-checked

After selecting research, use one AskUserQuestion step containing both bilingual selection questions:

• question: "How deeply should the sages research?" / "賢者にどの深さで調査させますか？"
  • "Standard research" / "標準リサーチ" — "Use the current prompt and normal search effort." / "現行プロンプトで通常の検索量を各賢者に任せます。"
  • "Deep research" / "ディープリサーチ" — "Require five searches, contrary evidence, and URL-backed findings; three sages take longer and consume more provider quota." / "各賢者に5回以上の検索、反証検索、URL付き発見を義務付けます。時間と提供元クォータの消費が増えます。"
• question: "Which tool scope should the sages use?" / "賢者にどのツール範囲を許可しますか？"
  • "Default scope" / "既定のツール範囲" — "Keep configured tools; the question may reach their providers." / "設定済みツールを保ちます。問いがその提供元へ届く場合があります。"
  • "Confidential scope" / "機密モード" — "Limit the bundled sages to Anthropic, OpenAI and xAI." / "同梱の賢者を Anthropic、OpenAI、xAI の3社に限定します。"

For a decision, use Default scope unless the question may be confidential; then ask the same bilingual tool-scope question before dispatch. Record both choices in the audit directory. Deep research uses the Deep template in references/council.md; Standard uses the existing template.

Step 3: Preflight

Append the selected --timeout <sec> and --confidential options to preflight and every dispatched round. This keeps the effective timeout in preflight.json and prevents later rounds from silently changing tool scope.

preflight="$("$magi_run" --out-dir "$out_dir" --preflight-only | tail -1)"

Read the file it names and branch on available_count, not on missing. A config that lists only two sages leaves missing empty and still gives you a two-sage council, so an empty missing is not evidence that the council is complete. Fewer than three available sages means Degraded Council: confirm with the user before dispatching anything.

Step 4: Round 1 — ask all three, independently

Write one prompt from the templates in references/council.md to ${out_dir}/round1/prompt.md. It must contain the question, the required answer format, and nothing about the other sages: round 1 must not reveal that other sages exist or what they might say. Independence is what makes agreement meaningful.

Required answer shapes (full templates in references/council.md):

Deliberation always uses the batched shape above, because one prompt normally carries several findings and the finding_id is what ties a verdict back to the claim it judges. The single-object form {"verdict": "...", "rationale": "..."} is acceptable only when that sage's prompt holds exactly one finding. A schema passed with --schema-file must be the batched one for the same reason.

Optionally write the matching JSON schema to ${out_dir}/schema.json and pass --schema-file. Sages that declare schema_args (CASPER by default) are then constrained to the shape; the others rely on the prompt instruction alone.

summary="$("$magi_run" \
  --prompt-file "${out_dir}/round1/prompt.md" \
  --out-dir "$out_dir" --round round1 | tail -1)"
# Add --schema-file "${out_dir}/schema.json" to constrain the sages that support
# it, --timeout 1200 when Deep needs longer, --confidential for Confidential
# scope, and --sages MELCHIOR,BALTHASAR to dispatch to a subset.

The last line of stdout is the path of summary.json. Exit 2 means a precondition failed and nothing was dispatched.

Step 5: Read the answers

summary.json holds one entry per sage: name, cli, status (ok / timeout / error), exit_code, duration_ms, answer, stdout_file, stderr_file, prompt_file.

For each sage with status: "ok", parse answer as JSON. Sages wrap JSON in code fences or add a sentence of prose more often than they should, so strip fences and take the first complete JSON object before deciding an answer is broken. If nothing parses, or a required field is missing, treat that sage as invalid answer — the same degraded case as a timeout (REQ-010). Never reconstruct the missing content yourself.

Then continue with Tally or Research Aggregation.

Tally (decision mode)

1. Take each sage's position.
2. Group positions by meaning, not by string. "TOML", "I'd go with TOML" and "TOML, for the schema stability" are one position. A position that merely overlaps — same choice but a materially different scope or condition — is a different position.
3. When you cannot tell whether two positions are the same, treat them as different. The cost of being wrong that way is one debate round; the cost of the reverse is silently discarding a disagreement the user needed to see.
4. Write the grouping down as a position → supporting sages table. This table feeds the matrix, so keep it even when the answer is obvious.
5. Any position held by two or more sages passes. Go straight to Result Matrix — no debate, no confirmation round.
6. All three positions different → Debate.

Confidence never overrides the count: a low-confidence pair beats a high-confidence single. Report the confidences in the matrix and let the user weigh them.

Debate (decision mode)

Only when all sages hold different positions.

1. Create ${out_dir}/debate1/ and write one prompt per sage to prompt-<SAGE>.md (e.g. prompt-MELCHIOR.md). Each contains: the original question, that sage's own previous answer, and the other two positions with rationales, attributed to Sage A and Sage B.
2. Anonymize the others. Never name the model or CLI behind Sage A or Sage B. A sage that learns which vendor it is arguing with weights the brand instead of the argument. Keep the mapping (which real sage is A and which is B, per recipient) in your own notes so you can attribute positions in the matrix, and keep it out of every prompt.
3. Ask for keep (hold the position), switch (adopt another) or compromise (propose a third), each with a rationale and a confidence.

4. Dispatch once. All sages have per-sage prompt files, so --prompt-file is not needed:

   summary="$("$magi_run" --out-dir "$out_dir" --round debate1 | tail -1)"
   # In a degraded council, add --sages with the two available names.

5. Re-tally with the same rules. A compromise is a new position, compared by meaning against everything else — two sages proposing compatible compromises is a pass.
6. Still three different positions → run debate2 the same way, showing each sage the other two post-debate positions.
7. After debate2, stop. At most two debate rounds. If nothing has two supporters, the council reaches no consensus: report the final positions plus a breakdown of what they actually disagree about, and hand the decision back to the user. Do not break the tie yourself, and do not run a third round.

Research Aggregation (research mode)

1. Split each sage's findings into individual claims. In Deep research, retain search_queries in the audit record and reject as unverified every finding whose source is not an http:// or https:// URL. Keep rejected findings visible in the final report, but do not send them to deliberation.
2. Cluster claims that assert the same fact, even when the wording, framing or granularity differs. Two claims that agree on the fact but disagree on a detail (a date, a number, a version) are one cluster with a noted conflict, not two findings.
3. Two or more sages → adopted. Record the support count and keep each sage's evidence and source.

4. One sage only → deliberation. Assign each solo finding an id, then write one prompt per sage to ${out_dir}/deliberation1/prompt-<SAGE>.md containing only the findings that sage did not report, each with its claim, evidence, source and id. Ask for agree / conditional / reject with a rationale per finding.

   Per-sage prompts are what keep a sage from reviewing its own claim. A single shared prompt cannot: as soon as two sages each have a solo finding, that prompt either asks each of them to judge its own finding or omits a review that was needed. Dispatch all of them in one parallel run — with per-sage prompt files, --prompt-file is not used:

   summary="$("$magi_run" --out-dir "$out_dir" --round deliberation1 | tail -1)"
   # A sage with nothing to review has no prompt file, so exclude it with
   # --sages; the script refuses to dispatch a sage it has no prompt for.

   Batch every finding a sage must review into that sage's prompt. There is one deliberation pass, so a second circulation is not available.

5. Classify each solo finding from the verdicts:

   Verdicts	Result
   both agree	adopted
   agree + conditional, or both conditional	conditionally adopted, note the conditions
   both reject	rejected
   agree + reject	conditionally adopted, note that the council split and quote both sides
   conditional + reject	conditionally adopted, note the condition and the rejection side by side

   A sage that fails to answer the deliberation does not veto: classify on the verdicts you have, and record the missing one. With a single verdict (degraded council), agree → adopted, conditional → conditionally adopted, reject → rejected.

6. Assemble the report: adopted facts with support counts, conditionally adopted with notes, rejected with reasons, then a per-sage summary. A rejected claim stays in the report — the fact that one sage believed it is a signal.

Result Matrix

Every answer ends with a matrix. Full templates are in references/council.md; the required elements are:

• A verdict headline first: the position that passed and the count (passed — TOML (2 to 1)), or no consensus — returned to the user, or for research a one-paragraph summary of the merged report.
• One row per sage, in configuration order, showing the display name with its actual CLI (MELCHIOR (claude)), the position, the gist of the rationale, the confidence, and what the debate changed. The debate column is part of the matrix even when the vote passed in round 1: fill it with not held rather than removing it, so the reader can tell a council that never debated from one whose debate you did not report.
• The minority opinion, always. Name the sage, state its position and rationale, and say plainly that it lost the vote but is recorded. A council whose dissent is not written down is a single model with extra steps.
• A row for every sage that could not answer, reading no answer (timeout), no answer (error) or no answer (invalid answer). Never drop the row — a silent sage is different from an absent one, and the user cannot see the difference if you delete it.
• The audit log line: the run directory path, noting that it holds every prompt and every raw answer.

For no consensus, add a breakdown of the disagreement: for each contested point, what each sage claimed. That breakdown is what makes the failed vote useful.

Degraded Council

Fewer than three sages, found before dispatch

preflight.json reports available_count (sages whose CLI was found), missing (CLIs that were not) and sages_file (the config actually read). Branch on available_count alone. A council can be short for two different reasons, and only one of them shows up in missing: a CLI is not installed, or the config lists fewer than three sages in the first place.

available_count ≥ 3 — proceed normally.

available_count == 2 — confirm before dispatching, and say which reason applies: name the missing CLI with its install command from references/sages.md, or name sages_file as a config with only two sages. Then ask:

• question: "Only two sages are available. Continue with a two-sage council?" / "利用できる賢者が2体だけです。2体で合議を続けますか？"
• options:
  • "Continue with two" / "2体で続行" — two-sage rules below: 2-0 passes, a 1-1 split gets one debate round
  • "Abort" / "中止" — stop now; install the missing CLI or fix the config, then re-run

Ask this question on both paths. A two-sage roster produces a two-sage verdict whether or not anything was missing, and the user is entitled to know the vote they are about to receive comes from two models.

When a CLI is missing, pass --sages listing exactly the available sages: the script refuses to dispatch a round while any selected sage's CLI is missing, so narrowing the selection is required, not optional. When the roster itself holds two sages, the default selection is already those two and --sages is unnecessary.

available_count ≤ 1 — do not dispatch. A single model is not a council. Report which CLIs are missing or how the config is short, give the install and login commands, and stop.

Two-sage rules

These apply however the council became two — a missing CLI, a roster of two, or a sage lost mid-round.

• decision: agreement 2-0 passes. A 1-1 split gets one debate round (debate1), never two; if the positions still differ after it, the result is no consensus and goes back to the user.
• research: two sages reporting the same fact → adopted. A solo finding circulates to the one remaining sage (still a single pass), and that single verdict decides it: agree → adopted, conditional → conditionally adopted, reject → rejected.

Failure during the round

A sage with status timeout or error, or an answer you could not parse, is out of this round. Continue with the sages that did answer, under the two-sage rules, and record the reason in the matrix.

Do not re-send the prompt to a failed sage within the same round (REQ-010). One sage, one round, one answer — that is what makes the run directory readable as a record, and a retry also spends the timeout budget twice.

If fewer than two sages produced a usable answer, there is nothing to tally. Show the answers you did get, name the failures and their reasons, point to the audit log, and stop.

Error Handling

Usage Examples

# Decision: three positions, tallied
run_id="$(date -u +%Y%m%dT%H%M%SZ)-$$"; out_dir="${TMPDIR:-/tmp}/magi-runs/${run_id}"
mkdir -p "${out_dir}/round1"
# ... write question.md, round1/prompt.md, schema.json ...
"$magi_run" --out-dir "$out_dir" --preflight-only
"$magi_run" --prompt-file "${out_dir}/round1/prompt.md" \
  --out-dir "$out_dir" --round round1 --schema-file "${out_dir}/schema.json"

# Debate: one anonymised prompt per sage, single dispatch
# ... write debate1/prompt-MELCHIOR.md, prompt-BALTHASAR.md, prompt-CASPER.md ...
"$magi_run" --out-dir "$out_dir" --round debate1

# Deliberation: per-sage prompts holding only the findings that sage did not report
# ... write deliberation1/prompt-BALTHASAR.md, prompt-CASPER.md ...
"$magi_run" --out-dir "$out_dir" --round deliberation1 --sages BALTHASAR,CASPER

# Degraded council: dispatch to the sages that exist
"$magi_run" --prompt-file "${out_dir}/round1/prompt.md" \
  --out-dir "$out_dir" --round round1 --sages BALTHASAR,CASPER

Run the script with --help for the full argument contract.

Notes

• Agreement between three models is evidence, not proof. Shared training data produces shared errors, and the matrix exists so the user can see how thin or broad the agreement was.
• A council is slow and costs three inferences. Use it where a wrong answer is expensive — architectural choices, "is this still true" checks on fast-moving facts — not for questions one agent can answer.
• The run directory is the deliverable when the council disappoints. Point the user at it rather than paraphrasing what a sage "basically said".