Discuss Subagent

Multi-perspective discussion for evaluating optimization strategies and reviewing code change quality. Used by strategist (DISCUSS-OPT) and reviewer (DISCUSS-REVIEW) when complex trade-offs require multi-angle analysis.

Design Rationale

Complex optimization decisions (e.g., choosing between algorithmic change vs caching layer) and nuanced code review findings (e.g., evaluating whether a side effect is acceptable) benefit from structured multi-perspective analysis. This subagent provides that analysis inline without spawning additional team members.

Invocation

Called by strategist, reviewer after their primary analysis when complexity warrants multi-perspective evaluation:

Task({
  subagent_type: "cli-discuss-agent",
  run_in_background: false,
  description: "Discuss <round-id>: <topic> for performance optimization",
  prompt: `Conduct a multi-perspective discussion on the following topic.

Round: <round-id>
Topic: <discussion-topic>
Session: <session-folder>

Context:
<relevant-context-from-calling-role>

Perspectives to consider:
- Performance impact: Will this actually improve the target metric?
- Risk assessment: What could go wrong? Side effects? Regressions?
- Maintainability: Is the optimized code still understandable and maintainable?
- Alternative approaches: Are there simpler or safer ways to achieve the same goal?

Evaluate trade-offs and provide a structured recommendation with:
- Consensus verdict: proceed / revise / escalate
- Confidence level: high / medium / low
- Key trade-offs identified
- Recommended approach with rationale
- Dissenting perspectives (if any)`
})

Round Configuration

Round	Artifact	Parameters	Calling Role
DISCUSS-OPT	/discussions/DISCUSS-OPT.md	Optimization strategy trade-offs	strategist
DISCUSS-REVIEW	/discussions/DISCUSS-REVIEW.md	Code review finding validation	reviewer

Integration with Calling Role

The calling role is responsible for:

Before calling: Complete primary analysis, identify the specific trade-off or finding needing discussion
Calling: Invoke subagent with round ID, topic, and relevant context
After calling:

Result	Action
consensus_reached (proceed)	Incorporate recommendation into output, continue
consensus_reached (revise)	Adjust findings/strategy based on discussion insights
consensus_blocked (HIGH)	Report to coordinator via message with severity
consensus_blocked (MEDIUM)	Include in output with recommendation for revision
consensus_blocked (LOW)	Note in output, proceed with original assessment

Output Schema

{
  "round_id": "<DISCUSS-OPT|DISCUSS-REVIEW>",
  "topic": "<discussion-topic>",
  "verdict": "<proceed|revise|escalate>",
  "confidence": "<high|medium|low>",
  "trade_offs": [
    { "dimension": "<performance|risk|maintainability>", "pro": "<benefit>", "con": "<cost>" }
  ],
  "recommendation": "<recommended-approach>",
  "rationale": "<reasoning>",
  "dissenting_views": ["<alternative-perspective>"]
}

Error Handling

Scenario	Resolution
Single perspective analysis fails	Continue with partial perspectives
All analyses fail	Return basic recommendation from calling role's primary analysis
Artifact not found	Return error immediately
Discussion inconclusive	Return "revise" verdict with low confidence

3.4 KiB Raw Blame History