Claude-Code-Workflow/.claude/skills/team-lifecycle-v4/subagents/discuss-subagent.md

Discuss Subagent

Lightweight multi-perspective critique engine. Called inline by produce roles (analyst, writer, reviewer) instead of as a separate team member. Eliminates spawn overhead while preserving multi-CLI analysis quality.

Design Rationale

In v3, discussant was a full team role requiring: agent spawn -> Skill load -> Phase 1 task discovery -> Phase 2-4 work -> Phase 5 report + callback. For what is essentially "run CLI analyses + synthesize", the framework overhead exceeded actual work time.

In v4, discuss is a subagent call from within the producing role, reducing each discuss round from ~60-90s overhead to ~5s overhead.

Invocation

Called by produce roles after artifact creation:

Task({
  subagent_type: "cli-discuss-agent",
  run_in_background: false,
  description: "Discuss <round-id>",
  prompt: `## Multi-Perspective Critique: <round-id>

### Input
- Artifact: <artifact-path>
- Round: <round-id>
- Perspectives: <perspective-list>
- Session: <session-folder>
- Discovery Context: <session-folder>/spec/discovery-context.json (for coverage perspective)

### Perspective Routing

| Perspective | CLI Tool | Role | Focus Areas |
|-------------|----------|------|-------------|
| Product | gemini | Product Manager | Market fit, user value, business viability |
| Technical | codex | Tech Lead | Feasibility, tech debt, performance, security |
| Quality | claude | QA Lead | Completeness, testability, consistency |
| Risk | gemini | Risk Analyst | Risk identification, dependencies, failure modes |
| Coverage | gemini | Requirements Analyst | Requirement completeness vs discovery-context |

### Execution Steps
1. Read artifact from <artifact-path>
2. For each perspective, launch CLI analysis in background:
   Bash(command="ccw cli -p 'PURPOSE: Analyze from <role> perspective for <round-id>
   TASK: <focus-areas>
   MODE: analysis
   CONTEXT: Artifact content below
   EXPECTED: JSON with strengths[], weaknesses[], suggestions[], rating (1-5)
   CONSTRAINTS: Output valid JSON only

   Artifact:
   <artifact-content>' --tool <cli-tool> --mode analysis", run_in_background=true)
3. Wait for all CLI results
4. Divergence detection:
   - Coverage gap: missing_requirements non-empty -> High severity
   - High risk: risk_level is high or critical -> High severity
   - Low rating: any perspective rating <= 2 -> Medium severity
   - Rating spread: max - min >= 3 -> Medium severity
5. Consensus determination:
   - No high-severity divergences AND average rating >= 3.0 -> consensus_reached
   - Otherwise -> consensus_blocked
6. Synthesize:
   - Convergent themes (agreed by 2+ perspectives)
   - Divergent views (conflicting assessments)
   - Coverage gaps
   - Action items from suggestions
7. Write discussion record to: <session-folder>/discussions/<round-id>-discussion.md

### Discussion Record Format
# Discussion Record: <round-id>

**Artifact**: <artifact-path>
**Perspectives**: <list>
**Consensus**: reached / blocked
**Average Rating**: <avg>/5

## Convergent Themes
- <theme>

## Divergent Views
- **<topic>** (<severity>): <description>

## Action Items
1. <item>

## Ratings
| Perspective | Rating |
|-------------|--------|
| <name> | <n>/5 |

### Return Value

**When consensus_reached**:
Return a summary string with:
- Verdict: consensus_reached
- Average rating
- Key action items (top 3)
- Discussion record path

**When consensus_blocked**:
Return a structured summary with:
- Verdict: consensus_blocked
- Severity: HIGH | MEDIUM | LOW
  - HIGH: any rating <= 2, critical risk identified, or missing_requirements non-empty
  - MEDIUM: rating spread >= 3, or single perspective rated <= 2 with others >= 3
  - LOW: minor suggestions only, all ratings >= 3 but divergent views exist
- Average rating
- Divergence summary: top 3 divergent points with perspective attribution
- Action items: prioritized list of required changes
- Recommendation: revise | proceed-with-caution | escalate
- Discussion record path

### Error Handling
- Single CLI fails -> fallback to direct Claude analysis for that perspective
- All CLI fail -> generate basic discussion from direct artifact reading
- Artifact not found -> return error immediately`
})

Round Configuration

Round	Artifact	Perspectives	Calling Role
DISCUSS-001	spec/discovery-context.json	product, risk, coverage	analyst
DISCUSS-002	spec/product-brief.md	product, technical, quality, coverage	writer
DISCUSS-003	spec/requirements/_index.md	quality, product, coverage	writer
DISCUSS-004	spec/architecture/_index.md	technical, risk	writer
DISCUSS-005	spec/epics/_index.md	product, technical, quality, coverage	writer
DISCUSS-006	spec/readiness-report.md	all 5	reviewer

Integration with Calling Role

The calling role is responsible for:

1. Before calling: Complete primary artifact output
2. Calling: Invoke discuss subagent with correct round config
3. After calling:

Verdict	Severity	Role Action
consensus_reached	-	Include action items in Phase 5 report, proceed normally
consensus_blocked	HIGH	Include divergence details in Phase 5 SendMessage with structured format (see below). Do NOT self-revise -- coordinator decides.
consensus_blocked	MEDIUM	Include warning in Phase 5 SendMessage. Proceed to Phase 5 normally.
consensus_blocked	LOW	Treat as consensus_reached with notes. Proceed normally.

SendMessage format for consensus_blocked (HIGH or MEDIUM):

[<role>] <task-id> complete. Discuss <round-id>: consensus_blocked (severity=<severity>)
Divergences: <top-3-divergent-points>
Action items: <prioritized-items>
Recommendation: <revise|proceed-with-caution|escalate>
Artifact: <artifact-path>
Discussion: <discussion-record-path>

The coordinator receives this and routes per severity (see monitor.md Consensus-Blocked Handling):

• HIGH -> creates revision task (max 1) or pauses for user
• MEDIUM -> proceeds with warning logged to wisdom/issues.md
• DISCUSS-006 HIGH -> always pauses for user decision (final sign-off gate)

Comparison with v3

Aspect	v3 (discussant role)	v4 (discuss subagent)
Spawn	Full general-purpose agent	Inline subagent call
Skill load	Reads SKILL.md + role.md	None (prompt contains all logic)
Task discovery	TaskList + TaskGet + TaskUpdate	None (called with context)
Report overhead	team_msg + SendMessage + TaskUpdate	Return value to caller
Total overhead	~25-45s framework	~5s call overhead
Pipeline beat	1 beat per discuss round	0 additional beats