Debate Research

Input Parameters

Collect from user before starting. Only topic is required; all others have defaults.

Param	Required	Default	Description
INLINECODE1	yes	—	Debate subject
INLINECODE2

no | Proponent + Opponent | 2-4 role objects: {name, stance, model?}. Default: Proponent (argue for) and Opponent (argue against). Model inherits from global. |
| goal | no | inferred | What question to answer |
| audience | no | "self" | Who reads the report: self / team / public |
| decision_type | no | "personal-choice" | personal-choice / team-standardization / market-analysis |
| evidence_round | no | "auto" | false / true / auto (enable when topic is fact-dense) |
| confirm_plan | no | true | Show plan and wait for user OK before execution |
| model | no | inherit | Global subagent model; role-level override takes priority |
| output_path | no | null | File path for report; null = return in conversation |

Implicit parameter: language — inferred from the user's topic/conversation language. All subagent prompts output in this language.

Example User Prompt

• - INLINECODE12

Execution Pipeline

Phase 0 — Pre-flight

Step 0a: Model reachability check

Collect all unique models (global + per-role + judge). For each unique model,
probe via sessions_spawn with a minimal one-sentence task (e.g. "Reply OK")
and model: <target>. Do NOT use curl or external HTTP — all models route
through OpenClaw's provider config.

If any probe fails:

• - If user explicitly specified the failed model → abort, report failure, suggest alternatives
• If model was default-assigned → warn user, fall back to session default model, continue

Step 0b: Plan presentation (if confirm_plan: true)

Present to user:

• - Topic
• Role × model assignment table
• Evidence round: on/off/auto (with rationale if auto)
• Estimated subagent call count
• Goal / audience / decision_type interpretation

[STOP — wait for user confirmation]

If confirm_plan: false, skip directly to Phase 1.

Phase 1 — Stance Investigation (parallel)

Spawn one subagent per role, all in parallel.

Each agent receives a prompt built from:

• - Role name + stance
• Topic
• web_search: enabled

Required output format per agent:
CODEBLOCK0

Use sessions_spawn + sessions_yield to wait for all completions.

Error handling:

• - Agent timeout → mark output [INCOMPLETE], continue pipeline

Phase 2 — Cross Rebuttal (parallel)

Spawn one subagent per role, all in parallel.

Each agent receives:

• - Its original stance
• All other roles' Phase 1 output (cross-injected)
• web_search: disabled

Required output format per agent:
CODEBLOCK1

Word limit: 300 × number_of_opponents words per agent.

Error handling:

• - Agent timeout → mark [INCOMPLETE], continue

Phase 2.5 — Evidence Audit (optional)

Triggered when evidence_round: true, or when auto and topic involves
measurable claims. Auto-enable heuristic: topic contains performance benchmarks,
cost comparisons, security assessments, market data, or quantitative metrics.
When in doubt with auto, skip (false positive costs more than false negative).

Spawn 1 subagent as "evidence auditor":

• - Input: all Phase 1 + Phase 2 output
• web_search: disabled
• Task: extract every factual claim, tag each as:

[official-docs] [community-feedback] [personal-blog] [no-source] [exaggerated]

• - Output: concise fact checklist

Phase 3 — Neutral Judgment

Spawn 1 subagent as neutral judge:

• - Input: Phase 1 + Phase 2 + Phase 2.5 (if available)
• web_search: disabled
• Weigh arguments by confidence scores AND source quality tags

Required output structure:

1. 1. Strong arguments per side
2. Exaggerated claims per side
3. Shared limitations (problems neither option solves)
4. Core disagreements (value-level, not just factual)
5. Consensus points
6. Recommendation — explicit directional advice, adapted to INLINECODE26
7. Open Questions — unresolved unknowns that could change the conclusion
8. Scenario selection matrix (table: scenario × recommendation × rationale)
9. One-sentence summary

Phase 4 — Report Assembly

Orchestrator (main conversation) assembles all outputs into Markdown:

CODEBLOCK2

If output_path specified → write file.
Otherwise → return in conversation.

Completion States

State	Condition	Behavior
INLINECODE28	All phases completed normally	Full report
INLINECODE29

1+ agents timed out or returned [INCOMPLETE] | Report with degradation note | | aborted | Model pre-check failed / user cancelled plan | No report; return error summary |

Prompt Templates

See references/prompts.md for the exact prompt templates used in each phase.
Orchestrator builds prompts dynamically from parameters + these templates.