Multi-Agent Debate

Orchestrates sequential multi-agent debates. Dynamic agents (2-5) with unique characters debate, main session acts as judge, secretary summarizes, judge writes final decisions.

Phase 0: Validation

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/validate.sh" && echo "VALID" || echo "FAILED"

STOP if FAILED — fix missing files before continuing.

Read archetypes into context:

Read file: ${CLAUDE_SKILL_DIR}/agents/archetypes.md

Phase 1: Parse Arguments

Arguments: $ARGUMENTS

Mode not specified (when -m omitted)

If mode is NOT explicitly provided via -m flag or clearly stated in the topic text, do NOT auto-detect. Ask user using AskUserQuestion:

Which debate mode?

1. Challenge — generate/receive variants, debate to select the best one
2. Strategy — each agent proposes independently, then debate to converge
3. Critic — all agents attack the given solution to find weaknesses/risks

Reply with mode name or number.

Only proceed after explicit user choice.

If topic is a file path (exists on disk) — read file content as topic.

Phase 2: Init + Display

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/init-log.sh" && echo "INIT_OK" || echo "INIT_FAILED"

STOP if INIT_FAILED — cannot create report directory.

Capture output — it prints:

REPORT_DIR=<path>
LOG_FILE=<path>

Store REPORT_DIR and LOG_FILE for all subsequent phases.

Display to user:

Debate Setup
  Mode:       {detected_mode}
  Agents:     {n}
  Max rounds: {r}
  Report:     {REPORT_DIR}
  Log:        {LOG_FILE}
  Topic:      {topic_summary}

Phase 3: User Interview

Ask user using AskUserQuestion tool:

Debate configuration:

Mode: {mode} | Agents: {n} | Max rounds: {r} Topic: {topic_first_100_chars}

Options:

1. Proceed with these settings
2. Change mode (challenge/strategy/critic)
3. Change agent count (2-5)
4. Change max rounds
5. Describe custom agent profiles (instead of auto-generated)

Apply any user changes. If user provides custom profiles — skip auto-generation in Phase 4 and use their descriptions.

Phase 4: Agent Profiles

Read reference for agent generation:

Read file: ${CLAUDE_SKILL_DIR}/references/setup-flow.md

Follow setup-flow.md to generate agent profiles. Result: a table of agents with name, role, character archetype, perspective, and WHY chosen.

Display agent table to user. Ask confirmation using AskUserQuestion:

Agent Team:

#	Name	Role	Archetype	Perspective
...	...	...	...	...

Options:

1. Proceed
2. Swap an agent (specify which)
3. Regenerate all

Phase 5: Discovery (Mandatory)

Research phase — gather current, verified information before debate begins.

Read file: ${CLAUDE_SKILL_DIR}/references/discovery-flow.md

Follow discovery-flow.md to spawn parallel research agents:

1. Codebase Explorer — searches project for relevant code, patterns, dependencies
2. Web Researcher — searches internet for current best practices, official docs, recent changes

All findings saved to {REPORT_DIR}/discovery.md with sources.

Every debate argument in Phase 6 MUST reference findings from discovery.md. Unsourced claims are not valid arguments.

Display discovery summary to user before proceeding to debate.

Phase 6: Debate

Load mode-specific flow reference and execute debate.

Read the matching reference file and follow its instructions exactly.

Agent spawning: Use Task tool with subagent_type: "general-purpose". Build each agent's prompt dynamically by combining:

1. Base template from ${CLAUDE_SKILL_DIR}/agents/debater-template.md
2. Role overlay from ${CLAUDE_SKILL_DIR}/agents/{role}-template.md
3. Agent's character traits from archetypes
4. Discovery findings from {REPORT_DIR}/discovery.md (injected as Evidence Base)
5. Current debate context (recent JSONL entries)

After each agent turn, append to log:

EXECUTE using Bash tool:

bash "${CLAUDE_SKILL_DIR}/scripts/append-log.sh" "LOG_FILE_PATH" '{"ts":"...","from":"agent-name","to":["targets"],"what":"<20 words","why":"<40 words (include [Source: #N] refs)","type":"argument","mode":"MODE"}'

Judge interventions (main session): After each round, evaluate if consensus emerging, redirect if stuck, end early if unanimous agreement.

Phase 7: Summary

Read file: ${CLAUDE_SKILL_DIR}/references/summary-flow.md

Follow summary-flow.md:

1. Read full debate log and {REPORT_DIR}/discovery.md
2. Spawn secretary agent to write summary.md in REPORT_DIR

Phase 8: Decision

Judge (main session) writes decisions.md:

• Winning position or synthesized result
• Key arguments that decided the outcome
• Minority opinions worth noting
• Confidence level: high / medium / low
• Recommended next steps

Write to: {REPORT_DIR}/decisions.md

Phase 9: Final Output

Display final status:

Debate Complete
  Mode:      {mode}
  Rounds:    {actual_rounds}/{max_rounds}
  Outcome:   {consensus | partial | no-consensus}
  Agents:    {agent_table_brief}

Decisions (top 3-5):
  - {bullet_1}
  - {bullet_2}
  - {bullet_3}

Artifacts:
  - {REPORT_DIR}/discovery.md
  - {REPORT_DIR}/decisions.md
  - {REPORT_DIR}/summary.md
  - {REPORT_DIR}/debate-log.jsonl

If --review flag was set:

Invoke: Skill(skill="brewcode:standards-review", args="{REPORT_DIR}")