Claude-Code-Workflow/.codex/skills/team-ultra-analyze/instructions/agent-instruction.md

TASK ASSIGNMENT

MANDATORY FIRST STEPS

1. Read shared discoveries: .workflow/.csv-wave/{session-id}/discoveries.ndjson (if exists, skip if not)
2. Read project context: .workflow/project-tech.json (if exists)

---

Your Task

Task ID: {id} Title: {title} Role: {role} Description: {description} Perspective: {perspective} Dimensions: {dimensions} Discussion Round: {discussion_round} Discussion Type: {discussion_type}

Previous Tasks' Findings (Context)

{prev_context}

---

Execution Protocol

1. Read discoveries: Load shared discoveries from the session's discoveries.ndjson for cross-task context
2. Use context: Apply previous tasks' findings from prev_context above
3. Execute by role:

Role: explorer (EXPLORE-* tasks)

Explore codebase structure from the assigned perspective, collecting structured context for downstream analysis.

• Determine exploration strategy by perspective:

Perspective	Focus	Search Depth
general	Overall codebase structure and patterns	broad
technical	Implementation details, code patterns, feasibility	medium
architectural	System design, module boundaries, interactions	broad
business	Business logic, domain models, value flows	medium
domain_expert	Domain patterns, standards, best practices	deep

• Use available tools (Read, Glob, Grep, Bash) to search the codebase
• Collect: relevant files (path, relevance, summary), code patterns, key findings, module relationships
• Generate questions for downstream analysis
• Focus exploration on the dimensions listed in the Dimensions field

Role: analyst (ANALYZE-* tasks)

Perform deep analysis on exploration results from the assigned perspective.

• Load exploration results from prev_context

• Detect if this is a direction-fix task (description mentions "adjusted focus"):

  • Normal: analyze from assigned perspective using corresponding exploration results
  • Direction-fix: re-analyze from adjusted perspective using all available explorations

• Select analysis approach by perspective:

Perspective	CLI Tool	Focus
technical	gemini	Implementation patterns, code quality, feasibility
architectural	gemini	System design, scalability, component interactions
business	gemini	Value, ROI, stakeholder impact
domain_expert	gemini	Domain-specific patterns, best practices

• Use ccw cli for deep analysis:

  ccw cli -p "PURPOSE: Deep analysis of '<topic>' from <perspective> perspective
  TASK: • Analyze patterns found in exploration • Generate insights with confidence levels • Identify discussion points
  MODE: analysis
  CONTEXT: @**/* | Memory: Exploration findings
  EXPECTED: Structured insights with confidence levels and evidence" --tool gemini --mode analysis
  

• Generate structured output:

  • key_insights: [{insight, confidence (high/medium/low), evidence (file:line)}]
  • key_findings: [{finding, file_ref, impact}]
  • discussion_points: [questions needing cross-perspective discussion]
  • open_questions: [areas needing further exploration]
  • recommendations: [{action, rationale, priority}]

Role: discussant (DISCUSS-* tasks)

Process analysis results and generate discussion summary. Strategy depends on discussion type.

• initial: Cross-perspective aggregation

  • Aggregate all analysis results from prev_context
  • Identify convergent themes across perspectives
  • Identify conflicting views between perspectives
  • Generate top 5 discussion points and open questions
  • Produce structured round summary

• deepen: Deep investigation of open questions

  • Use CLI tool to investigate uncertain insights:

    ccw cli -p "PURPOSE: Investigate open questions and uncertain insights
    TASK: • Focus on questions from previous round • Find supporting evidence • Validate uncertain insights
    MODE: analysis
    CONTEXT: @**/*
    EXPECTED: Evidence-based findings" --tool gemini --mode analysis
    

• direction-adjusted: Re-analysis from adjusted focus

  • Use CLI to re-analyze from adjusted perspective based on user feedback

• specific-questions: Targeted Q&A

  • Use CLI for targeted investigation of user-specified questions

• For all types, produce round summary:

  • updated_understanding: {confirmed[], corrected[], new_insights[]}
  • convergent themes, conflicting views
  • remaining open questions

Role: synthesizer (SYNTH-* tasks)

Integrate all explorations, analyses, and discussions into final conclusions.

• Read all available artifacts from prev_context (explorations, analyses, discussions)

• Execute synthesis in four steps:

  1. Theme Extraction: Identify convergent themes across perspectives, rank by cross-perspective confirmation
  2. Conflict Resolution: Identify contradictions, present trade-off analysis
  3. Evidence Consolidation: Deduplicate findings, aggregate by file reference, assign confidence levels
  4. Recommendation Prioritization: Sort by priority, deduplicate, cap at 10

• Confidence levels:

Level	Criteria
High	Multiple sources confirm, strong evidence
Medium	Single source or partial evidence
Low	Speculative, needs verification

• Produce final conclusions:
  • Executive summary
  • Key conclusions with evidence and confidence
  • Prioritized recommendations
  • Open questions
  • Cross-perspective synthesis (convergent themes, conflicts resolved, unique contributions)

4. Share discoveries: Append exploration findings to shared board:

   echo '{"ts":"<ISO8601>","worker":"{id}","type":"<type>","data":{...}}' >> .workflow/.csv-wave/{session-id}/discoveries.ndjson
   

   Discovery types to share:

   • exploration: {perspective, file, relevance, summary, patterns[]} — explored file/module
   • analysis: {perspective, insight, confidence, evidence, file_ref} — analysis insight
   • pattern: {name, file, description, type} — code/architecture pattern
   • discussion_point: {topic, perspectives[], convergence, open_questions[]} — discussion point
   • recommendation: {action, rationale, priority, confidence} — recommendation
   • conclusion: {point, evidence, confidence, perspectives_supporting[]} — final conclusion

5. Report result: Return JSON via report_agent_job_result

---

Output (report_agent_job_result)

Return JSON: { "id": "{id}", "status": "completed" | "failed", "findings": "Key discoveries and implementation notes (max 500 chars)", "error": "" }

Role-specific findings guidance:

• explorer: List file count, key files, patterns found. Example: "Found 12 files related to auth. Key: src/auth/index.ts (entry), src/auth/strategies/*.ts (providers). Patterns: strategy, middleware chain."
• analyst: List insight count, top insights with confidence. Example: "3 insights: (1) Strategy pattern for providers [high], (2) Missing token rotation [medium], (3) No rate limiting [high]. 2 discussion points."
• discussant: List themes, conflicts, question count. Example: "Convergent: JWT security (2 perspectives). Conflict: middleware approach. 3 open questions on refresh tokens."
• synthesizer: List conclusion count, top recommendations. Example: "5 conclusions, 4 recommendations. Top: Implement refresh token rotation [high priority, high confidence]."