Debate

Have multiple LLMs debate the best approach, then synthesize and implement.

Load the consult-llm skill before proceeding — it defines the invocation contract (stdin heredoc, flags, output format, multi-turn). Do not call the CLI without loading it first.

Available models

Selectors resolvable in this environment (depends on configured API keys):

!`consult-llm models`

Phase 0: Load consult-llm Skill

Load it now. Follow its invocation contract for all CLI calls in this workflow.

Configuration

Arguments: $ARGUMENTS

Check the arguments for flags:

Model flags: any --<selector> from the Models block above selects a debater (e.g. --gemini, --openai, --deepseek). Repeat for multiple. Need at least two debaters. Translate model flags and defaults according to the loaded consult-llm skill's model-selection rules.

Mode flags:

• --dry-run → debate and plan only, skip implementation
• --skip-final → skip the final review phase
• --rounds N → number of debate rounds (default: 2, max: 3)

Strip all flags from arguments to get the task description.

Throughout this skill, references to "each LLM"/"each debater" mean every selected model. Use the selector name (gemini, openai, etc.) as the label when presenting per-model output.

Phase 1: Understand the Task (No Questions)

1. Explore the codebase - use Glob, Grep, Read to understand:

   • Relevant files and their structure
   • Existing patterns and conventions
   • Dependencies and interfaces

   Before planning or consulting, do enough research to understand how the requested behavior actually works. Before starting, think about what resources would be useful to obtain first: relevant source files, tests, logs, generated files, config, examples, command output, external docs, or authoritative upstream source. Gather the cheapest useful evidence before forming a plan.

   Do not stop at the first plausible file, definition, setting, or example. Follow references, callers, related tests, and runtime usage until you can explain the current behavior and the likely impact of changing it.

2. Ground external semantics before planning - understand the requested behavior in the real system, not just this repo

   • If the task depends on an external product, CLI, API, protocol, file format, or ecosystem convention, verify the relevant behavior using the cheapest authoritative evidence available: local binaries/flags, generated files, official docs, public source, package/library code, or web search.
   • Capture only decision-relevant facts that affect scope, acceptance criteria, compatibility, or implementation constraints.
   • Do not create a separate research artifact unless the evidence materially changes the plan.

3. Make evidence-backed assumptions - do NOT ask clarifying questions

   • Use best judgment based on codebase and external context
   • Prefer simpler solutions when ambiguous
   • Follow existing patterns in the codebase

4. Prepare context summary - create a brief summary of:

   • The task to be implemented
   • Relevant files discovered
   • Key patterns and conventions in the codebase
   • Any constraints or considerations

Phase 2: Opening Arguments

Have both LLMs propose their approach independently (in parallel).

Opening prompt:

I need to implement the following task:

[Task description]

Here's what I found in the codebase:
[Context summary - relevant files, patterns, conventions]

Propose your implementation approach:
1. **Approach**: Describe your recommended approach in 2-3 sentences
2. **Key decisions**: List the main architectural/design decisions
3. **Files**: What files to create or modify
4. **Steps**: High-level implementation steps
5. **Trade-offs**: What are the pros and cons of this approach?

Be specific and opinionated. Defend your choices.

Invoke consult-llm with -f <path> for each relevant source file, sending the opening prompt per the consult-llm invocation contract. If explicit debater flags were supplied, pass one -m <selector> per debater. Otherwise omit -m so consult-llm applies configured defaults. All models are queried in parallel in a single call.

Extract per-model thread IDs from the response — needed for Phase 3 since each model receives the others' rebuttals.

Phase 3: Debate Rounds

For each round (default 2, configurable with --rounds N, max 3):

Have each LLM critique the others' latest arguments (in parallel). Pass each LLM's thread ID via -t <id> to continue its conversation — they already have full context of the task and their own prior arguments, so you only need to send the opponents' latest responses.

Round 1 rebuttal prompt (same template for each debater; embed every other debater's opening argument, labeled by selector):

Your opponent(s) proposed these alternative approaches:
[Opponents' opening arguments, each labeled with the selector name]

Provide a rebuttal:
1. **Critique**: What are the weaknesses in each opponent's approach?
2. **Defense**: Address any weaknesses in your own approach
3. **Concessions**: Are there any good ideas worth adopting?
4. **Updated position**: State your refined recommendation

Be constructive but thorough in your critique.

Subsequent round prompt (same template; embed every other debater's latest rebuttal):

Your opponent(s) have responded to your critique:
[Opponents' latest rebuttals, each labeled with the selector name]

Continue the debate:
1. **Critique**: What weaknesses remain in their updated positions?
2. **Defense**: Address any new points raised against your approach
3. **Concessions**: Any new ideas worth adopting?
4. **Updated position**: State your refined recommendation

Focus on unresolved disagreements. Don't repeat settled points.

Each model receives every other model's latest response. Invoke consult-llm once with one --run per debater, continuing each model's thread.

Present both responses to the user after each round.

Phase 4: Moderator's Verdict

As the moderator, analyze the debate and synthesize the best approach:

1. Score the arguments:

   • Which approach is simpler?
   • Which approach better fits existing patterns?
   • Which critiques were valid?
   • What concessions were made?

2. Identify consensus: Where did all the debaters agree?

3. Resolve disagreements: For each point of contention:

   • Evaluate the arguments from each side
   • Pick the strongest position or find a middle ground
   • Prefer simpler solutions when arguments are equally strong

4. Write the verdict as part of the plan:

# [Feature Name] Implementation Plan

**Goal:** [One sentence describing what this builds]

## Debate Summary

**Positions** (one bullet per debater, labeled with the selector name):
- **<selector>:** [1-2 sentence summary]

**Points of agreement:**
- [Consensus point 1]
- [Consensus point 2]

**Resolved disagreements:**
- [Issue]: <selector-A> said X, <selector-B> said Y → **Verdict:** [Your decision and why]

**Verdict:** [2-3 sentences on the final synthesized approach]

---

### Task 1: [Short description]

**Files:**
- Create: `exact/path/to/file.py`
- Modify: `exact/path/to/existing.py` (lines 123-145)

**Steps:**
1. [Specific action]
2. [Specific action]

**Code:**
```language
// Include actual code, not placeholders
```

---

Guidelines:

• Exact file paths - never "somewhere in src/"
• Complete code - show the actual code
• Small tasks - 2-5 minutes of work each
• DRY, YAGNI - only what's needed

Save the plan to history/plan-<feature-name>.md.

Phase 5: Implement

If --dry-run: Skip to Phase 7 (Summary) - report the debate and plan without implementing.

Implement the plan without further interaction:

1. Follow the plan exactly - implement each task in order
2. Commit after each logical unit - keep commits small and focused
3. If something is unclear - make a reasonable decision and note it in the commit message
4. If a task fails - attempt to fix it before moving on
5. Only stop if there's a blocking error that cannot be resolved

Implementation rules:

• Work through tasks sequentially
• Test changes when possible
• Keep commits atomic and well-documented
• Use commit messages that explain the "why"

Phase 6: Final Review

If --skip-final: Skip to Phase 7 (Summary).

After implementation, have every debater LLM review the result (in parallel). Pass each LLM's thread ID via -t <id> to continue its conversation — they already have full context of the task and the debate, so you only need to send the review prompt and the diff.

Final review prompt:

Forget which side you argued during the debate. Review the implementation purely on its merits:
- Any obvious bugs or edge cases missed?
- Code quality issues (error handling, naming, structure)?
- Deviations from best practices?
- Security concerns?

Be concise. Only flag issues worth fixing.

Invoke consult-llm --task review once with one --run per debater, passing --diff-files and --diff-base as shared context, continuing each model's thread.

Apply fixes if multiple reviewers identify the same issue, or if one raises a clearly valid concern:

• Fix bugs and edge cases
• Commit each fix separately with clear messages

Skip minor style suggestions or conflicting opinions.

Phase 7: Summary

Present a final summary to the user:

## Summary

**Implemented:** [One sentence describing what was built]

**Debate outcome:**
- One bullet per debater, labeled with the selector: `<selector>` advocated: [key position]
- Final verdict: [synthesized approach]

**Key decisions from debate:**
- [Decision 1 and why]
- [Decision 2 and why]

**Post-implementation fixes:**
- [Fix applied after final review, if any]

**Commits:**
- `abc1234` - [commit message]
- `def5678` - [commit message]