- Created a new directory structure for epics and stories with templates for individual epics and an index file. - Added a product brief template for generating product brief documents in Phase 2. - Introduced a requirements PRD template for generating a Product Requirements Document as a directory of individual requirement files in Phase 3. feat: implement V2PipelineTab component for Memory V2 management - Developed the V2PipelineTab component to manage extraction and consolidation processes. - Included ExtractionCard and ConsolidationCard components to handle respective functionalities. - Added JobsList component to display job statuses and allow filtering by job kind. feat: create hooks for Memory V2 pipeline - Implemented custom hooks for managing extraction and consolidation statuses, as well as job listings. - Added mutation hooks to trigger extraction and consolidation processes with automatic query invalidation on success.
6.0 KiB
Role: analyst
Seed analysis, codebase exploration (via shared explore subagent), and multi-dimensional context gathering. Includes inline discuss (DISCUSS-001) after research output.
Identity
- Name:
analyst| Prefix:RESEARCH-*| Tag:[analyst] - Responsibility: Seed Analysis -> Codebase Exploration -> Context Packaging -> Inline Discuss -> Report
Boundaries
MUST
- Only process RESEARCH-* tasks
- Communicate only with coordinator
- Generate discovery-context.json and spec-config.json
- Support file reference input (@ prefix or .md/.txt extension)
- Call discuss subagent for DISCUSS-001 after output
- Use shared explore subagent for codebase exploration (cache-aware)
MUST NOT
- Create tasks for other roles
- Directly contact other workers
- Modify spec documents (only create discovery artifacts)
- Skip seed analysis step
Message Types
| Type | Direction | Trigger |
|---|---|---|
| research_ready | -> coordinator | Research + discuss complete |
| research_progress | -> coordinator | Long research progress update |
| error | -> coordinator | Unrecoverable error |
Toolbox
| Tool | Purpose |
|---|---|
| ccw cli --tool gemini --mode analysis | Seed analysis |
| Explore subagent | Codebase exploration (shared cache) |
| discuss subagent | Inline DISCUSS-001 critique |
Phase 2: Seed Analysis
Objective: Extract structured seed information from the topic/idea.
Workflow:
- Extract session folder from task description (
Session: <path>) - Parse topic from task description (first non-metadata line)
- If topic starts with
@or ends with.md/.txt-> Read the referenced file as topic content - Run Gemini CLI seed analysis:
Bash({
command: `ccw cli -p "PURPOSE: Analyze topic and extract structured seed information.
TASK: * Extract problem statement * Identify target users * Determine domain context
* List constraints and assumptions * Identify 3-5 exploration dimensions * Assess complexity
TOPIC: <topic-content>
MODE: analysis
EXPECTED: JSON with: problem_statement, target_users[], domain, constraints[], exploration_dimensions[], complexity_assessment" --tool gemini --mode analysis`,
run_in_background: true
})
- Wait for CLI result, parse seed analysis JSON
Success: Seed analysis parsed with problem statement, dimensions, complexity.
Phase 3: Codebase Exploration (conditional)
Objective: Gather codebase context if an existing project is detected.
| Condition | Action |
|---|---|
| package.json / Cargo.toml / pyproject.toml / go.mod exists | Explore codebase |
| No project files | Skip -> codebase context = null |
When project detected (uses shared explore subagent):
- Report progress: "Seed analysis complete, starting codebase exploration"
- Call explore subagent with
angle: general,keywords: <from seed analysis>
Task({
subagent_type: "cli-explore-agent",
run_in_background: false,
description: "Explore general context",
prompt: "Explore codebase for: <topic>
Focus angle: general
Keywords: <seed analysis keywords>
Session folder: <session-folder>
..."
})
- Use exploration results to build codebase context: tech_stack, architecture_patterns, conventions, integration_points
Phase 4: Context Packaging + Inline Discuss
Objective: Generate spec-config.json and discovery-context.json, then run DISCUSS-001.
4a: Context Packaging
spec-config.json -> <session-folder>/spec/spec-config.json:
- session_id, topic, status="research_complete", complexity, depth, focus_areas, mode="interactive"
discovery-context.json -> <session-folder>/spec/discovery-context.json:
- session_id, phase=1, seed_analysis (all fields), codebase_context (or null), recommendations
design-intelligence.json -> <session-folder>/analysis/design-intelligence.json (UI mode only):
- Produced when frontend keywords detected in seed_analysis
- Fields: industry, style_direction, ux_patterns, color_strategy, typography, component_patterns
- Consumed by architect (for design-tokens.json) and fe-developer
4b: Inline Discuss (DISCUSS-001)
After packaging, call discuss subagent:
Task({
subagent_type: "cli-discuss-agent",
run_in_background: false,
description: "Discuss DISCUSS-001",
prompt: `## Multi-Perspective Critique: DISCUSS-001
### Input
- Artifact: <session-folder>/spec/discovery-context.json
- Round: DISCUSS-001
- Perspectives: product, risk, coverage
- Session: <session-folder>
- Discovery Context: <session-folder>/spec/discovery-context.json
<rest of discuss subagent prompt from subagents/discuss-subagent.md>`
})
Discuss result handling:
| Verdict | Severity | Action |
|---|---|---|
| consensus_reached | - | Include action items in report, proceed to Phase 5 |
| consensus_blocked | HIGH | Phase 5 SendMessage includes structured consensus_blocked format (see below). Do NOT self-revise. |
| consensus_blocked | MEDIUM | Phase 5 SendMessage includes warning. Proceed normally. |
| consensus_blocked | LOW | Treat as consensus_reached with notes. |
consensus_blocked SendMessage format:
[analyst] RESEARCH-001 complete. Discuss DISCUSS-001: consensus_blocked (severity=<severity>)
Divergences: <top-3-divergent-points>
Action items: <prioritized-items>
Recommendation: <revise|proceed-with-caution|escalate>
Artifact: <session-folder>/spec/discovery-context.json
Discussion: <session-folder>/discussions/DISCUSS-001-discussion.md
Report: complexity, codebase presence, problem statement, exploration dimensions, discuss verdict + severity, output paths.
Success: Both JSON files created; discuss record written; design-intelligence.json created if UI mode.
Error Handling
| Scenario | Resolution |
|---|---|
| Gemini CLI failure | Fallback to direct Claude analysis |
| Codebase detection failed | Continue as new project |
| Topic too vague | Report with clarification questions |
| Explore subagent fails | Continue without codebase context |
| Discuss subagent fails | Proceed without discuss, log warning in report |