6.6 KiB
Deep Workflow
Full pipeline with research, brainstorming, and planning for complex issues. Uses native Claude Tasks with dependency chains.
Task Setup (Before Starting)
Create all phase tasks upfront. Steps 1+2+3 run in parallel (scout + diagnose + research).
T1 = TaskCreate(subject="Scout codebase", activeForm="Scouting codebase", metadata={phase: "investigate"})
T2 = TaskCreate(subject="Diagnose root cause", activeForm="Diagnosing root cause", metadata={phase: "investigate"})
T3 = TaskCreate(subject="Research solutions", activeForm="Researching solutions", metadata={phase: "investigate"})
T4 = TaskCreate(subject="Brainstorm approaches", activeForm="Brainstorming", metadata={phase: "design"}, addBlockedBy=[T1, T2, T3])
T5 = TaskCreate(subject="Create implementation plan", activeForm="Planning implementation", metadata={phase: "design"}, addBlockedBy=[T4])
T6 = TaskCreate(subject="Implement fix", activeForm="Implementing fix", metadata={phase: "implement"}, addBlockedBy=[T5])
T7 = TaskCreate(subject="Verify + prevent", activeForm="Verifying fix", metadata={phase: "verify"}, addBlockedBy=[T6])
T8 = TaskCreate(subject="Code review", activeForm="Reviewing code", metadata={phase: "verify"}, addBlockedBy=[T7])
T9 = TaskCreate(subject="Finalize & docs", activeForm="Finalizing", metadata={phase: "finalize"}, addBlockedBy=[T8])
Steps
Step 1: Scout Codebase (parallel with Steps 2+3)
TaskUpdate(T1, status="in_progress")
Mandatory: Activate ck:scout skill or launch 2-3 Explore subagents in parallel:
Task("Explore", "Find error origin and affected components", "Trace error")
Task("Explore", "Find module boundaries and dependencies", "Map deps")
Task("Explore", "Find related tests and similar patterns", "Find patterns")
Map: all affected files, module boundaries, call chains, test coverage gaps.
See references/parallel-exploration.md for patterns.
TaskUpdate(T1, status="completed")
Output: ✓ Step 1: Scouted - [N] files, system impact: [scope]
Step 2: Diagnose Root Cause (parallel with Steps 1+3)
TaskUpdate(T2, status="in_progress")
Mandatory skill chain:
- Capture pre-fix state: Record ALL error messages, failing tests, stack traces, logs.
- Activate
ck:debugskill (systematic-debugging + root-cause-tracing). - Activate
ck:sequential-thinking— structured hypothesis formation. - Spawn parallel
Exploresubagents to test each hypothesis. - If 2+ hypotheses fail → auto-activate
ck:problem-solving. - Trace backward through call chain to ROOT CAUSE origin.
See references/diagnosis-protocol.md for full methodology.
TaskUpdate(T2, status="completed")
Output: ✓ Step 2: Diagnosed - Root cause: [summary], Evidence: [chain]
Step 3: Research (parallel with Steps 1+2)
TaskUpdate(T3, status="in_progress")
Use researcher subagent for external knowledge.
- Search latest docs, best practices
- Find similar issues/solutions
- Gather security advisories if relevant
TaskUpdate(T3, status="completed")
Output: ✓ Step 3: Research complete - [key findings]
Step 4: Brainstorm
TaskUpdate(T4, status="in_progress") — auto-unblocks when T1 + T2 + T3 complete.
Activate ck:brainstorm skill.
- Evaluate multiple approaches using scout + diagnosis + research findings
- Consider trade-offs
- Get user input on preferred direction
TaskUpdate(T4, status="completed")
Output: ✓ Step 4: Approach selected - [chosen approach]
Step 5: Plan
TaskUpdate(T5, status="in_progress")
Use planner subagent to create implementation plan.
- Break down into phases
- Identify dependencies
- Define success criteria
- Include prevention measures in plan
TaskUpdate(T5, status="completed")
Output: ✓ Step 5: Plan created - [N] phases
Step 6: Implement
TaskUpdate(T6, status="in_progress")
Implement per plan. Use ck:context-engineering, ck:sequential-thinking, ck:problem-solving.
- Fix ROOT CAUSE per diagnosis — not symptoms
- Follow plan phases
- Minimal changes per phase
TaskUpdate(T6, status="completed")
Output: ✓ Step 6: Implemented - [N] files, [M] phases
Step 7: Verify + Prevent
TaskUpdate(T7, status="in_progress")
Mandatory skill chain:
- Iron-law verify: Re-run EXACT commands from pre-fix state. Compare before/after.
- Regression test: Add comprehensive tests. Tests MUST fail without fix, pass with fix.
- Defense-in-depth: Apply all relevant prevention layers (see
references/prevention-gate.md). - Parallel verification: Launch
Bashagents: typecheck + lint + build + test. - Edge cases: Test boundary conditions, security implications, performance impact.
If verification fails: Loop back to Step 2 (re-diagnose). Max 3 attempts → question architecture.
See references/prevention-gate.md for prevention requirements.
TaskUpdate(T7, status="completed")
Output: ✓ Step 7: Verified + Prevented - [before/after], [N] tests, [M] guards
Step 8: Code Review
TaskUpdate(T8, status="in_progress")
Use code-reviewer subagent.
See references/review-cycle.md for mode-specific handling.
TaskUpdate(T8, status="completed")
Output: ✓ Step 8: Review [score]/10 - [status]
Step 9: Finalize
TaskUpdate(T9, status="in_progress")
- Report summary: root cause, evidence chain, changes, prevention measures, confidence score
- Activate
ck:project-managementfor task sync-back, plan status updates, and progress tracking - Use
docs-managersubagent for documentation - Use
git-managersubagent for commit - Run
/ck:journal
TaskUpdate(T9, status="completed")
Output: ✓ Step 9: Complete - [actions taken]
Skills/Subagents Activated
| Step | Skills/Subagents |
|---|---|
| 1 | ck:scout OR parallel Explore subagents |
| 2 | ck:debug, ck:sequential-thinking, parallel Explore, (ck:problem-solving auto) |
| 3 | researcher (runs parallel with steps 1+2) |
| 4 | ck:brainstorm |
| 5 | planner |
| 6 | ck:problem-solving, ck:sequential-thinking, ck:context-engineering |
| 7 | tester, parallel Bash verification |
| 8 | code-reviewer |
| 9 | ck:project-management, docs-manager, git-manager |
Rules: Don't skip steps. Validate before proceeding. One phase at a time.
Frontend: Use chrome, ck:chrome-devtools or any relevant skills/tools to verify.
Visual Assets: Use ck:ai-multimodal for visual assets generation, analysis and verification.