Models Consensus

Models Consensus

Moderate a council of native seats and local runner skills. Do not act as one of the voting seats. The moderator's job is to run a deterministic preflight, launch only real seats, preserve independence accounting, surface disagreements that actually change scope or behavior, and enforce cost and recovery guardrails.

Seat Discovery

Detect available seats from host-native tools and installed CLI runners. See references/repo-configuration.md for the repository-specific seat catalog, default models, and CLI prerequisites. Adapt RUNNER_BASE_PATH when using this skill in another repository.

Do not claim a seat participated unless the preflight confirms a real execution path for it.

Host Tool Compatibility

Council logic is host-agnostic, but concrete tools differ by platform. See references/runner-invocations.md for the full host capability mapping and native vs. runner launch patterns. Branch all seat launch instructions by host capability.

For any user-facing question, use the Interactive Questions protocol.

Mandatory Preflight

Run preflight before building any round prompt.

0. Resolve seat selection mode

Before the full preflight, determine whether seat selection is automatic or user-directed. Always resolve startup seat selection before running smoke tests or launching any round.

• Accept auto: true in a structured payload.
• Accept --auto in a free-form user request as shorthand for auto: true.
• If auto or --auto is present, skip the startup seat-selection question and target every available seat. Only target all seats when --auto or auto: true is explicitly set; do not auto-select seats just because interactive startup questioning is unavailable.
• Otherwise, ask the user which models or CLIs to use before running smoke tests or launching any round. Never skip the startup seat-selection question in non---auto runs.
• In non---auto runs, if startup selection is still unresolved after the available question channels, stop with awaiting_user_decision.

Selection workflow:

• Do a lightweight capability discovery first: detect host-native seat support and runner binary presence only.
• Build a candidate_seats list from that lightweight discovery.
• Ask one startup selection question following the Interactive Questions protocol.
• Prefer one multi-select question when the host supports it, rather than a series of yes/no prompts. Include All available (Recommended) plus the detected candidate seats.
• Multi-select startup prompt template:

Which models to use? [ ] All available (Recommended) [ ] Claude Opus 4.8 [ ] Claude Sonnet 4.6 [ ] Codex [ ] Gemini [ ] Kimi [ ] Gemma [ ] GLM [ ] Minimax Omit seats that are not present in candidate_seats.

• If the host tool only supports a small single-choice menu, ask a preset question with:

1. All available (Recommended) 2. Core seats only 3. Specify seats manually

• If the user chooses manual selection, ask one concise follow-up question listing the detected seat IDs and accept a comma-separated subset. Use an interactive follow-up question when the host supports one; use a plain-text seat-picking follow-up only when needed.
• Plain-text startup prompt template when no interactive selection tool is available:

Which models to use? Detected seats: <comma-separated seat IDs> Reply with a comma-separated subset or 'all'. Set selection_source=plain_text_manual when this fallback branch is used, and wait for user input.

• Treat Core seats only as the highest-diversity non-duplicate set available. See references/repo-configuration.md for the core-seats definition.
• Persist selected_seats and selection_source in state, and emit the selected and omitted seats, before continuing to smoke tests.

1. Detect host-native seats

• If Agent exists, Claude Opus 4.8 and Claude Sonnet 4.6 use native seats.
• If spawn_agent and wait_agent exist, Codex uses a native seat.
• Otherwise, those seats may use runner scripts if the local CLI exists.

2. Check local runner prerequisites

Use shell commands to verify binaries, auth, and one cheap headless smoke test for every runner-backed seat you intend to launch. Only run smoke tests for seats included by selected_seats or by --auto.

Minimum checks:

• claude in PATH
• gemini in PATH
• kimi-cli in PATH when the Kimi seat is under consideration
• qwen in PATH when any Gemma, GLM, or Minimax seat is under consideration

Mandatory seat smoke tests:

• Run a minimal non-interactive invocation for every runner-backed seat with the exact model you plan to use.
• Run every runner-backed seat with --disable-fallback. Councils must fail a seat explicitly instead of silently borrowing another provider.
• See references/runner-invocations.md for per-seat auth rules, including the critical Claude --bare rule and Qwen transport rule.

Treat missing binaries, missing credentials, or failing smoke tests as seat blockers, not soft warnings.

3. Determine artifact mode

Prefer persisted artifacts in .ai-workflow/consensus/ only when writes are safe.

Set artifact mode to:

• persisted when .ai-workflow/consensus/ exists or its parent is writable
• inline when writes are blocked, risky, or unavailable in the current mode

Apply the per-mode behaviors from references/operations.md#artifact-policy.

4. Build a seat table

For each potential seat, record:

• seat
• selection_status: selected, omitted_by_user, or unavailable
• requested_runner
• execution_path: native, runner, or unavailable
• effective_provider
• effective_model when known
• blocked_reason when unavailable

Do not launch any seat until this table is complete.

Deterministic Seat Selection

Use these rules:

1. Prefer native Claude seats over claude-runner. 2. Prefer native Codex seats over any CLI path. 3. On a Codex host, the Codex seat must run as a native subagent. 4. Use runner scripts only when the native seat path is unavailable. 5. Treat runner fallback to another provider as loss of seat independence. 6. If Gemini CLI is missing, skip Gemini entirely. 7. If Kimi CLI is missing, skip Kimi entirely. 8. If Qwen CLI is missing, skip Gemma, GLM, and Minimax entirely. 9. Use GLM Critic only after all unique-provider seats have been considered. 10. Treat Gemma, GLM, and Minimax as separate providers for diversity accounting unless the smoke test proves they resolve to the same effective provider and model. 11. Continue with the remaining seats and lower confidence; never fabricate a missing seat.

Cost Governance and Crash Recovery

Both procedures live in references/operations.md; read them on demand:

• When more than 4 seats are selected, read references/operations.md#cost-governance before launching seats and apply its cost-transparency rules.
• At preflight, if state_path exists with status != complete, read references/operations.md#crash-recovery-and-state-resumption and resume from state instead of restarting.

Response Schema Validation

Seat outputs ARE schema-validated before they enter the moderator digest: each is checked against the per-round required-field list, retried once on failure, then marked malformed_output and excluded if still invalid. The field lists, validation behavior, and the retry schema-reminder template live in references/operations.md#response-schema-validation.

Interactive Questions

Every user-facing question must use the best available interactive host tool first, on both CLI and app surfaces. Treat plain-text questioning as a last resort only when no interactive tool is available in the current host or mode.

Questioning workflow: 1. Detect which interactive question/input tools are exposed by the current host and current mode. 2. Prefer a tool that can present structured choices or multiple short questions in one call. 3. If the first tool is unavailable, unsupported in the current mode, or errors immediately, try the next interactive tool. 4. Use concise plain-text questions only after exhausting interactive tool options.

Preferred host mappings:

• Claude Code: AskUserQuestion
• Codex: request_user_input when the current mode exposes it
• any other host: the equivalent native interactive question/input tool, if available
• only if none of the above are available: a concise plain-text question with 2-3 options

Additional rules:

• Never choose plain text just because an interactive tool is less convenient.
• If only one question fits per interactive call, use repeated interactive calls rather than switching to plain text.
• When the host supports multiple questions in one call, batch related questions together to reduce back-and-forth.

Never tell the user to reply with numbered choices when an interactive question tool exists.

Startup seat-selection question: governed by preflight step 0 — that is the single normative rule for when and how to ask.

Degrade Gracefully

Apply these rules when the council cannot run at full strength:

• Missing seat: continue and lower overall confidence one level.
• Fallback seat: mark the original seat as unavailable for independence accounting.
• Shared provider, different models: keep both seats but note the shared-provider caveat.
• Same effective provider and same model: collapse to one independent source and call out the lost diversity.
• Malformed output: retry once, then mark seat unavailable and continue.
• Inline artifact mode: return report_path=null and state_path=null, plus a note that persistence was skipped because writes were unavailable.

Use confidence bands:

• high: 4+ independent seats with no critical blockers
• medium: 2-3 independent seats, or meaningful fallback duplication
• low: 1 independent seat, or unresolved blockers on key disagreement points

Moderator time budget:

• Prefer a finite moderation window for opening statements. Recommended defaults: 2-3 minutes for native seats, 10-15 minutes for runner seats.
• Once at least 3 independent providers have completed and the major disagreements are already clear, stop waiting unless another seat is likely to change the decision materially.
• After moderation cutoff, close native seats you no longer need and kill unfinished runner processes so councils do not leave background work behind.

Input Format and Mode Behavior

The input payload schema (question, context_files, max_iterations, session_id, auto, mode, plus the --auto shortcut) and the per-mode pausing behavior are in references/operations.md#input-format and references/operations.md#mode-behavior. Startup seat selection is governed by preflight step 0 in all modes.

Council Model

Assign one stance per active seat per round. See references/stance-rotation-schedule.md for the explicit per-seat, per-round stance mapping and fallback rules.

Available stances:

• supportive_with_integrity
• critical_with_responsibility
• balanced_synthesis
• devils_advocate
• pragmatic_engineering

The stance-to-runner-role mapping lives in the same reference.

If fewer seats are available than unique stances, drop duplicate coverage before dropping unique stances.

Discussion Protocol

1. Build the shared brief

Always include:

• the question
• context file paths only
• the current objective
• previous round digest and user decisions when iteration > 1
• the required response schema

In persisted mode, write prompt files only for runners that need them. In inline mode, keep the brief in memory and pass it directly.

2. Run blinded opening statements

Round 1 inputs:

• shared brief
• stance overlay
• no peer outputs
• no moderator conclusion

Required response schema: the Round 1 fields from Response Schema Validation.

3. Moderate the round

Normalize outputs into:

• agreement_points
• disagreement_points
• decision_options
• evidence_gaps
• follow_up_questions

Do not paste long verbatim runner output into later rounds. Pass only a compact moderator digest.

4. Run rebuttal and refinement rounds

Later rounds get:

• shared brief
• current stance overlay
• prior moderator digest
• explicit instructions to rebut, concede, refine, or integrate

Required response schema: the later-rounds fields from Response Schema Validation.

5. Classify convergence

Use one of:

• full_agreement
• converging
• material_disagreement
• blocked_on_context

Heuristic definitions:

• full_agreement: all selected seats share the same recommended_direction and no seat raises a blocking objection.
• converging: all seats agree on recommended_direction but differ on implementation details, risk weighting, or acceptance criteria.
• material_disagreement: ≥2 seats hold opposing recommended_direction values, and at least one objection affects scope, behavior, or architecture (not style or preference).
• blocked_on_context: seats agree that the question cannot be answered without additional information, file contents, or external validation.

Capture:

• agreement_points
• disagreement_points
• leading_option
• minority_concerns
• open_questions

6. Handle disagreement

Trigger when:

• disagreement remains material after a round
• the decision depends on priorities or preferences
• more context is required
• any recommendation would lead to code or document changes

In interactive mode, ask focused questions with: 1. recommended option first 2. minority alternative(s) 3. Run another round to refine positions when another round could help

Ask these questions using the Interactive Questions protocol.

If a required answer is still missing after available question channels, stop with awaiting_user_decision.

In autonomous mode:

• record each disagreement
• include recommended_resolution
• include resolution_reasoning
• continue to stop conditions

7. Update state

When artifact mode is persisted, update state after every iteration with:

• assigned stances
• seat outputs
• moderator digest
• convergence classification
• user decisions
• recommendation log
• runner execution metadata
• independence accounting

When artifact mode is inline, keep the same structure in memory and return it inline instead of writing it.

8. Stop condition

Stop when:

• full agreement is reached
• the user selects a direction in interactive mode
• all rounds complete in autonomous mode
• max_iterations is reached

9. Handoff after approval

Only after the user approves a direction:

• build a compact handoff brief
• include the final recommendation, minority concerns, and acceptance criteria
• use a native seat when available
• otherwise invoke the selected runner with --role implementer or --role codereviewer

Do not hand off automatically when the user asked only for analysis.

Final Report

When artifact mode is persisted, write .ai-workflow/consensus/{session_id}.md with: 1. question and status 2. preflight summary, including seat selection source and selected seats 3. iteration summary 4. agreement points 5. divergence points 6. user decision summary for any non-unanimous result 7. final recommendation 8. recommended next runner and handoff goal 9. confidence assessment

When artifact mode is inline, return the same sections inline.

Output Contract

Return: 1. report_path or null 2. state_path or null 3. concise inline summary 4. preflight seat table 5. selected seats and selection source

Important Rules

1. Preflight first. Never launch seats from assumptions. 2. Resolve seat selection per preflight step 0 before running smoke tests or rounds. 3. Use the same shared brief for every participant in a round; only the stance overlay changes. 4. Keep opening statements blinded from peer outputs. 5. Treat the moderator as separate from all council seats. 6. Rotate stances across iterations using references/stance-rotation-schedule.md to reduce model-position bias. 7. Continue gracefully when one runner fails; never fabricate missing outputs. 8. In interactive mode, never apply recommendation-derived edits without explicit user approval. 9. Prefer continuation over repetition when moving from consensus to implementation.

Operations Reference

Operational detail for models-consensus: cost governance, crash recovery, response schema validation, artifact policy, runner launch policy, input format, and mode behavior. Read the relevant section when its trigger in SKILL.md fires.

Cost Governance

Multi-model councils can be expensive. Enforce cost transparency before launching seats.

Pre-flight cost estimate:

• After seat selection, estimate total token volume: (selected_seats.count) × (avg_input_tokens + avg_output_tokens).
• Use rough defaults: ~4k input tokens per seat (brief + context), ~2k output tokens per seat.
• Warn the user when >4 seats are selected in non-auto mode: "Council will invoke N models, roughly X tokens total across providers. Proceed?"

Token budgets:

• Cap per-seat output at ~4k tokens for moderation feasibility.
• If shared brief + context files exceed ~8k tokens per seat, truncate or summarize context files before the round.
• Prefer concise briefs over verbatim file dumps.

Cost-conscious defaults:

• In non---auto runs, make "Core seats only" the recommended startup choice rather than "all available."
• When --auto selects 5+ seats, emit a cost warning before the first round.
• Track cumulative token usage across rounds in state for post-council reporting.

Crash Recovery and State Resumption

Council state is resumable. Treat .ai-workflow/consensus/{session_id}.json as the source of truth for progress.

Resume handshake:

• At preflight, check if state_path exists and status != complete.
• If resuming, load prior round outputs, seat assignments, and moderator digests from state.
• Set resumed_from to the previous state's last_completed_round.
• Skip to the next uncompleted round; do not re-run completed rounds.

Orphaned process cleanup:

• When resuming, identify any runner PIDs or background tasks from the prior session and terminate them before launching new seats.
• In inline mode, recovery is limited to what fits in the current context; persist key digests to state when possible.

State update cadence:

• In persisted mode, write state after every iteration.
• In inline mode, still build the same state structure in memory so it can be returned or persisted if the mode changes.

Response Schema Validation

Validate seat outputs before accepting them into the moderator digest. Seat outputs ARE schema-validated against the field lists below.

Round 1 required fields:

• stance
• position_summary
• key_arguments
• risks_or_limits
• recommended_direction
• confidence
• questions_for_the_council

Later rounds required fields:

• updated_position
• what_changed
• points_conceded
• remaining_objections
• best_next_step
• confidence

Validation behavior:

• If a response is missing required fields, retry once with a compact schema reminder prepended to the prompt.
• If the retry also fails, mark the seat as malformed_output, exclude it from the digest, and degrade gracefully.
• Do not fabricate missing fields from the seat's partial output.

Schema reminder template (prepend to prompts when retrying):

Respond in JSON with exactly these top-level keys: [list keys].
No markdown fencing. No extra commentary outside the JSON object.

Artifact Policy

Determine artifact mode in preflight, then apply the per-mode behaviors here.

When artifact mode is persisted, use:

• state: .ai-workflow/consensus/{session_id}.json
• report: .ai-workflow/consensus/{session_id}.md
• per-round outputs: .ai-workflow/consensus/{session_id}-round-{n}-{seat}-output.json
• optional prompt files only when the selected runner requires them
• prefer runner-level --output-file writes over shell redirection so incomplete seats do not leave misleading zero-byte artifacts

When artifact mode is inline:

• do not create temp prompt files; build prompts in memory
• do not require .ai-workflow/consensus/
• return the final report and state inline, with report_path and state_path set to null
• keep round digests and moderator digests in memory

Runner Launch Policy

Launch seats using native host tools when available; fall back to runner scripts only when native paths are unavailable. See runner-invocations.md for complete invocation patterns, auth rules, and the runner output contract.

Runner seats invoke local CLIs and may send prompt context, selected files, and runner metadata to their configured providers. Prefer --restrict-tools for review and planning seats. Do not pass permission bypass or full auto flags unless the user has explicitly approved unattended write capable execution for that run.

Key flags for every runner-backed seat:

• --disable-fallback (mandatory)
• --timeout 900
• --json for wrapper envelope
• --output-file for persisted artifacts

In inline mode, combine stance and brief into a single positional prompt. In persisted mode, use --prompt-file flags when the runner supports them.

Input Format

Expected input payload:

• question: required
• context_files: optional list of repo-relative or absolute file paths
• max_iterations: optional, default 4
• session_id: required unique identifier
• auto: optional boolean, default false
• mode: optional, interactive (default) or autonomous

Free-form shortcut:

• --auto is equivalent to auto=true

Mode Behavior

Startup seat selection is governed by preflight step 0 in SKILL.md (the single normative rule). Mode only affects later pausing:

• interactive: pause later only when disagreement is material or preference-sensitive.
• autonomous: run the rounds without further pauses and return recommendations and reasoning.

Repository Configuration

This reference documents the default seat-to-runner mapping for a repository. Adapt paths and model defaults if using this skill in another project.

Runner Base Path

In this repo, all runner scripts live under:

.agents/skills/{runner-name}/scripts/run_{runner-name}.py

When adapting this skill to another project, update the base path or set a RUNNER_BASE_PATH variable.

Available Seats

CLI Prerequisites

Core Seats Definition

"Core seats" refers to the highest-diversity non-duplicate set available: 1. Native Codex (when on Codex host) 2. Native Claude seats (when on Claude Code host) 3. Gemini 4. Kimi 5. One of Gemma or GLM

Duplicate-coverage seats (e.g., GLM Critic) are only considered after core seats are exhausted.

Artifact Directory

Persisted council artifacts write to:

.ai-workflow/consensus/

In this repo, the .ai-workflow/ directory is guaranteed writable.

Runner Invocation Reference

Complete invocation patterns for every council seat, organized by host capability.

Table of Contents

1. Host Tool Mapping 2. Native Seat Patterns 3. Runner Fallback Patterns 4. Auth and Transport Rules 5. Runner Output Contract

---

Host Tool Mapping

Treat host tooling as a compatibility layer. Council logic stays the same; concrete tools differ by platform.

Never refer to Claude-only tool names as if they are universal. Branch all seat launch instructions by host capability. When the workflow needs user input, use the Interactive Questions protocol in SKILL.md.

---

Native Seat Patterns

Claude Opus 4.8 (Claude Code host)

Agent(
  subagent_type="general-purpose",
  description="Claude Opus 4.8 council seat — round {n}",
  model="claude-opus-4-8",
  prompt="<stance overlay>\n\n---\n\n<shared brief>",
  run_in_background=true
)

Use model="claude-sonnet-4-6" for the Sonnet 4.6 seat.

Codex (Codex host)

spawn_agent(
  fork_context=false,
  model="gpt-5.5",
  reasoning_effort="medium",
  message="<stance overlay>\n\n---\n\n<shared brief>"
)

For adversarial or research-heavy rounds, raise reasoning_effort to high.

If full-history context inheritance is needed, either spawn without explicit model and reasoning_effort overrides, or keep fork_context=false and pass the task-local brief directly.

Critical: On a Codex host, do not invoke codex-runner or codex exec for the Codex seat.

---

Runner Fallback Patterns

Use runner scripts only when the native seat path is unavailable. Pass --disable-fallback so councils fail a seat explicitly instead of silently borrowing another provider.

Claude Opus 4.8 / Sonnet 4.6 (runner fallback)

python3 .agents/skills/claude-runner/scripts/run_claude.py \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-stance-claude-opus.md \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-brief.md \
  --timeout 900 \
  --role planner \
  --model claude-opus-4-8 \
  --output-format json \
  --json \
  --no-session-persistence \
  --restrict-tools \
  --disable-fallback \
  --output-file .ai-workflow/consensus/{session_id}-round-{n}-claude-opus-output.json \
  --metadata-json '{"session":"{session_id}","round":{n},"seat":"claude-opus","stance":"supportive_with_integrity"}'

Use --model claude-sonnet-4-6 for the Sonnet 4.6 seat.

In inline artifact mode, combine the prompt and pass it as a single positional prompt instead of --prompt-file flags.

Codex (runner fallback — non-Codex hosts only)

python3 .agents/skills/codex-runner/scripts/run_codex.py \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-codex.md \
  --timeout 900 \
  --role challenger \
  --model gpt-5.5 \
  --effort high \
  --json \
  --ephemeral \
  --restrict-tools \
  --disable-fallback \
  --output-file .ai-workflow/consensus/{session_id}-round-{n}-codex-output.json \
  --metadata-json '{"session":"{session_id}","round":{n},"seat":"codex","stance":"devils_advocate"}'

codex-runner supports --effort none|minimal|low|medium|high|xhigh. Use high for adversarial or research-heavy rounds, mirroring the native Codex seat guidance.

Gemini

python3 .agents/skills/gemini-runner/scripts/run_gemini.py \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-gemini.md \
  --timeout 900 \
  --role synthesizer \
  --json \
  --output-format json \
  --disable-fallback \
  --output-file .ai-workflow/consensus/{session_id}-round-{n}-gemini-output.json \
  --metadata-json '{"session":"{session_id}","round":{n},"seat":"gemini","stance":"balanced_synthesis"}'

Do not depend on speculative Gemini-only flags such as --thinking-budget or a read-only convenience mode.

Kimi

python3 .agents/skills/kimi-runner/scripts/run_kimi.py \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-kimi.md \
  --timeout 900 \
  --role implementer \
  --model kimi-code/kimi-for-coding \
  --output-format stream-json \
  --json \
  --no-session-persistence \
  --restrict-tools \
  --disable-fallback \
  --output-file .ai-workflow/consensus/{session_id}-round-{n}-kimi-output.json \
  --metadata-json '{"session":"{session_id}","round":{n},"seat":"kimi","stance":"pragmatic_engineering"}'

Gemma

python3 .agents/skills/gemma-runner/scripts/run_gemma.py \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-gemma.md \
  --timeout 900 \
  --role planner \
  --model google/gemma-4-31b-it \
  --output-format stream-json \
  --json \
  --no-session-persistence \
  --restrict-tools \
  --disable-fallback \
  --output-file .ai-workflow/consensus/{session_id}-round-{n}-gemma-output.json \
  --metadata-json '{"session":"{session_id}","round":{n},"seat":"gemma","stance":"supportive_with_integrity"}'

GLM Pragmatic

python3 .agents/skills/glm-runner/scripts/run_glm.py \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-glm.md \
  --timeout 900 \
  --role implementer \
  --model z-ai/glm-5.2 \
  --output-format stream-json \
  --json \
  --no-session-persistence \
  --restrict-tools \
  --disable-fallback \
  --output-file .ai-workflow/consensus/{session_id}-round-{n}-glm-output.json \
  --metadata-json '{"session":"{session_id}","round":{n},"seat":"glm","stance":"pragmatic_engineering"}'

GLM Critic

Use only as a duplicate-coverage seat after unique-provider seats are accounted for.

python3 .agents/skills/glm-runner/scripts/run_glm.py \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-glm-critical.md \
  --timeout 900 \
  --role codereviewer \
  --model z-ai/glm-5.2 \
  --output-format stream-json \
  --json \
  --no-session-persistence \
  --restrict-tools \
  --disable-fallback \
  --output-file .ai-workflow/consensus/{session_id}-round-{n}-glm-critical-output.json \
  --metadata-json '{"session":"{session_id}","round":{n},"seat":"glm-critical","stance":"critical_with_responsibility"}'

Minimax

python3 .agents/skills/minimax-runner/scripts/run_minimax.py \
  --prompt-file .ai-workflow/consensus/{session_id}-round-{n}-minimax.md \
  --timeout 900 \
  --role challenger \
  --model minimax/minimax-m2.7 \
  --output-format stream-json \
  --json \
  --no-session-persistence \
  --restrict-tools \
  --disable-fallback \
  --output-file .ai-workflow/consensus/{session_id}-round-{n}-minimax-output.json \
  --metadata-json '{"session":"{session_id}","round":{n},"seat":"minimax","stance":"devils_advocate"}'

---

Auth and Transport Rules

--disable-fallback

Always pass --disable-fallback to runner-backed seats. Councils must fail a seat explicitly instead of silently borrowing another provider.

Claude --bare rule

Do not use --bare for Claude runner seats when relying on Claude OAuth or keychain-backed login. Claude's own help states that --bare disables OAuth and keychain auth, so a logged-in terminal can still fail with Not logged in in headless mode if --bare is passed. Only use --bare when ANTHROPIC_API_KEY or an explicit apiKeyHelper-based configuration is the intended auth path.

Qwen transport rule

Treat the shared qwen CLI as transport, not independence. Gemma, GLM, and Minimax remain distinct seats when their effective providers or models differ, even though they share the same local wrapper binary.

---

Runner Output Contract

All runner-script paths return a wrapper envelope with fields such as:

• success
• stdout
• stderr
• return_code
• runner
• effective_runner
• role
• prompt_file or prompt_files when applicable
• session_file when applicable

Important:

• --json controls the wrapper envelope.
• Native CLI JSON or JSONL output stays in stdout.
• Every runner skill in this repo emits agent_message (the clean final answer) when it can extract one, and session_id when the underlying CLI reports it — prefer agent_message over parsing stdout. For the Claude seat this requires --output-format json or stream-json. Do not assume usage or token-cost fields exist unless the specific runner emitted them.
• When --output-file is used, treat the file as the source of truth. Stdout may be only a small acknowledgment payload.

Normalize native-seat output into the same envelope shape before comparing seats.

Stance Rotation Schedule

Rotate stances across iterations to reduce model-position bias. The moderator applies this schedule per seat based on the round number.

Round 1: Natural Stances

Each model starts in its natural role:

The Runner Role column above is the canonical stance-to-runner-role mapping (supportive_with_integrity -> planner or implementer, critical_with_responsibility -> codereviewer or adversarial, balanced_synthesis -> synthesizer, devils_advocate -> challenger, pragmatic_engineering -> implementer). For blocked_on_context investigation rounds, use the researcher runner role.

Round 2: Cross-Stance Pressure

Each seat adopts a stance that challenges its Round 1 position:

Round 3: Convergence / Integration

Final round focuses on integration or decisive critique:

Fallback Rules

• If fewer than 3 seats are available, skip Round 3 and run only 2 rounds.
• If a seat's output in Round 1 already matches the Round 3 stance for that seat, keep the Round 2 assignment for Round 2 but shift Round 3 to balanced_synthesis to ensure the seat contributes to convergence.
• If the user explicitly pauses the council to provide direction, resume from the next round using the scheduled stance for that round number — do not restart the rotation.