Dialectic Debate

Stress-test ideas by spawning competing perspectives, then synthesize and ground-truth against code.

$ARGUMENTS is the question or topic to debate.

If no argument provided, ask the user what they want to evaluate.

Use TaskCreate / TaskUpdate to track these 4 phases:

1. Frame the debate
2. Spawn thesis + antithesis agents
3. Synthesize positions
4. Verify claims against code

---

Phase 1: Frame the Debate

Parse the topic from $ARGUMENTS. Identify:

• The core tension — what are the two opposing positions?
• Stakes — what depends on getting this right?
• Scope — bound the debate to avoid sprawl

Frame as a clear binary or spectrum question. Examples:

• "Should we use microservices or a monolith for X?"
• "Is caching worth the complexity here?"
• "Should this be a library or an inline implementation?"

---

Phase 2: Spawn Thesis + Antithesis (Parallel)

Spawn two Explore agents in a single message:

Task(
  subagent_type="Explore",
  run_in_background=true,
  description="Thesis: argue FOR",
  prompt="You are arguing FOR: {position A}.
  Topic: {framed question}

  Build the strongest possible case:
  1. Search the codebase for evidence supporting this position
  2. Identify concrete benefits with file:line references
  3. Anticipate and preempt counterarguments
  4. Rate your confidence (high/medium/low) with reasoning

  Be specific — cite code, patterns, and constraints you find.
  DO NOT hedge or present both sides. Argue your position fully."
)

Task(
  subagent_type="Explore",
  run_in_background=true,
  description="Antithesis: argue AGAINST",
  prompt="You are arguing AGAINST: {position A} (i.e., FOR {position B}).
  Topic: {framed question}

  Build the strongest possible case:
  1. Search the codebase for evidence supporting this position
  2. Identify concrete risks/costs of the opposing view with file:line references
  3. Anticipate and preempt counterarguments
  4. Rate your confidence (high/medium/low) with reasoning

  Be specific — cite code, patterns, and constraints you find.
  DO NOT hedge or present both sides. Argue your position fully."
)

---

Phase 3: Synthesize

Collect both results:

TaskOutput(task_id=<thesis_id>, block=true)
TaskOutput(task_id=<antithesis_id>, block=true)

Synthesize into a structured verdict:

1. Where they agree — shared facts, common ground
2. Where they conflict — genuine disagreements with evidence
3. Weak arguments — flag claims that lack code evidence or rely on generalities
4. Verdict — which position has stronger grounding, and why
5. Conditions — when would the other position win instead?

---

Phase 4: Verify Claims

For any factual claims made by either side (file references, pattern assertions, dependency claims):

• Read the cited files to confirm accuracy
• Flag any misrepresentations or hallucinated evidence
• Adjust verdict if verification changes the balance

---

Output

DEBATE: {topic}
==============

Thesis: {position A}
  Confidence: {high/medium/low}
  Key evidence: {1-2 strongest points with file refs}

Antithesis: {position B}
  Confidence: {high/medium/low}
  Key evidence: {1-2 strongest points with file refs}

Verdict: {position} wins because {reason}
Caveat: {position} would win if {conditions}

Verified claims: {N}/{M} checked, {K} corrected

---

Examples

/debating-ideas Should we split the API into microservices?
/debating-ideas Is it worth adding Redis caching to the auth flow?
/debating-ideas Monorepo vs polyrepo for our frontend packages

---

If the debate reaches no clear conclusion, present both positions with evidence and let the user decide.

Frame the debate now.