Files
english/.opencode/skills/cook/references/workflow-steps.md
2026-04-12 01:06:31 +07:00

193 lines
8.1 KiB
Markdown

# Unified Workflow Steps
All modes share core steps with mode-specific variations.
**Task Tool Fallback:** `TaskCreate`/`TaskUpdate`/`TaskGet`/`TaskList` are CLI-only — unavailable in VSCode extension. If they error, use `TodoWrite` for progress tracking. All workflow steps remain functional without Task tools.
## Step 0: Intent Detection & Setup
1. Parse input with `intent-detection.md` rules
2. Log detected mode: `✓ Step 0: Mode [X] - [reason]`
3. If mode=code: detect plan path, set active plan
4. Use `TaskCreate` to create workflow step tasks (with dependencies if complex)
**Output:** `✓ Step 0: Mode [interactive|auto|fast|parallel|no-test|code] - [detection reason]`
## Step 1: Research (skip if fast/code mode)
**Interactive/Auto:**
- Spawn multiple `researcher` agents in parallel
- Use `/ck:scout ext` or `scout` agent for codebase search
- Keep reports ≤150 lines
**Parallel:**
- Optional: max 2 researchers if complex
**Output:** `✓ Step 1: Research complete - [N] reports gathered`
### [Review Gate 1] Post-Research (skip if auto mode)
- Present research summary to user
- Use `AskUserQuestion` to ask: "Proceed to planning?" / "Request more research" / "Abort"
- **Auto mode:** Skip this gate
## Step 2: Planning
**Interactive/Auto/No-test:**
- Use `planner` agent with research context
- Create `plan.md` + `phase-XX-*.md` files
**Fast:**
- Use `/ck:plan --fast` with scout results only
- Minimal planning, focus on action
**Parallel:**
- Use `/ck:plan --parallel` for dependency graph + file ownership matrix
**Code:**
- Skip - plan already exists
- Parse existing plan for phases
**Output:** `✓ Step 2: Plan created - [N] phases`
### [Review Gate 2] Post-Plan (skip if auto mode)
- Present plan overview with phases
- Use `AskUserQuestion` to ask: "Validate the plan or approve plan to start implementation?" - "Validate" / "Approve" / "Abort" / "Other" ("Request revisions")
- "Validate": run `/ck:plan validate` skill invocation
- "Approve": continue to implementation
- "Abort": stop the workflow
- "Other": revise the plan based on user's feedback
- **Auto mode:** Skip this gate
## Step 3: Implementation
**IMPORTANT:**
1. `TaskList` first — check for existing tasks (hydrated by planning skill in same session)
2. If tasks exist → pick them up, skip re-creation
3. If no tasks → read plan phases, `TaskCreate` for each unchecked `[ ]` item with priority order and metadata (`phase`, `planDir`, `phaseFile`)
4. Tasks can be blocked by other tasks via `addBlockedBy`
**All modes:**
- Use `TaskUpdate` to mark tasks as `in_progress` immediately.
- Execute phase tasks sequentially (Step 3.1, 3.2, etc.)
- Use `ui-ux-designer` for frontend
- Use `ck:ai-multimodal` for image assets
- Run type checking after each file
**Parallel mode:**
- Utilize all tools of Claude Tasks: `TaskCreate`, `TaskUpdate`, `TaskGet` and `TaskList`
- Launch multiple `fullstack-developer` agents
- When agents pick up a task, use `TaskUpdate` to assign task to agent and mark tasks as `in_progress` immediately.
- Respect file ownership boundaries
- Wait for parallel group before next
**Output:** `✓ Step 3: Implemented [N] files - [X/Y] tasks complete`
### [Review Gate 3] Post-Implementation (skip if auto mode)
- Present implementation summary (files changed, key changes)
- Use `AskUserQuestion` to ask: "Proceed to testing?" / "Request implementation changes" / "Abort"
- **Auto mode:** Skip this gate
## Step 4: Testing (skip if no-test mode)
**All modes (except no-test):**
- Write tests: happy path, edge cases, errors
- **MUST** spawn `tester` subagent: `Task(subagent_type="tester", prompt="Run test suite", description="Run tests")`
- If failures: **MUST** spawn `debugger` subagent → fix → repeat
- **Forbidden:** fake mocks, commented tests, changed assertions, skipping subagent delegation
**Output:** `✓ Step 4: Tests [X/X passed] - tester subagent invoked`
### [Review Gate 4] Post-Testing (skip if auto mode)
- Present test results summary
- Use `AskUserQuestion` to ask: "Proceed to code review?" / "Request test fixes" / "Abort"
- **Auto mode:** Skip this gate
## Step 5: Code Review
**All modes - MANDATORY subagent:**
- **MUST** spawn `code-reviewer` subagent: `Task(subagent_type="code-reviewer", prompt="Review changes. Return score, critical issues, warnings.", description="Code review")`
- **DO NOT** review code yourself - delegate to subagent
**Interactive/Parallel/Code/No-test:**
- Interactive cycle (max 3): see `review-cycle.md`
- Requires user approval
**Auto:**
- Auto-approve if score≥9.5 AND 0 critical
- Auto-fix critical (max 3 cycles)
- Escalate to user after 3 failed cycles
**Fast:**
- Simplified review, no fix loop
- User approves or aborts
**Output:** `✓ Step 5: Review [score]/10 - [Approved|Auto-approved] - code-reviewer subagent invoked`
## Step 6: Finalize
**All modes - MANDATORY subagents (NON-NEGOTIABLE):**
1. **MUST** spawn these subagents in parallel:
- `Task(subagent_type="project-manager", prompt="Run full sync-back for [plan-path]: reconcile all completed Claude Tasks with all phase files, backfill stale completed checkboxes across every phase, then update plan.md frontmatter/table progress. Do NOT only mark current phase.", description="Update plan")`
- `Task(subagent_type="docs-manager", prompt="Update docs for changes.", description="Update docs")`
2. Project-manager sync-back MUST include:
### Status Sync (Finalize)
Use CLI commands for deterministic status updates:
```bash
# Mark completed phases
ck plan check <phase-id>
# Mark in-progress phases
ck plan check <phase-id> --start
# Revert if needed
ck plan uncheck <phase-id>
```
**Fallback:** If `ck` is not available, edit plan.md directly —
only change the Status column cell, preserve table structure.
- Sweep all `phase-XX-*.md` files in the plan directory.
- Mark every completed item `[ ] → [x]` based on completed tasks (including earlier phases finished before current phase).
- Update `plan.md` status/progress (`pending`/`in-progress`/`completed`) from actual checkbox state.
- Return unresolved mappings if any completed task cannot be matched to a phase file.
3. Use `TaskUpdate` to mark Claude Tasks complete after sync-back confirmation.
4. Onboarding check (API keys, env vars)
5. **MUST** spawn git subagent: `Task(subagent_type="git-manager", prompt="Stage and commit changes", description="Commit")`
**CRITICAL:** Step 6 is INCOMPLETE without spawning all 3 subagents. DO NOT skip subagent delegation.
**Auto mode:** Continue to next phase automatically, start from **Step 3**.
**Others:** Ask user before next phase
**Output:** `✓ Step 6: Finalized - 3 subagents invoked - Full-plan sync-back completed - Committed`
## Mode-Specific Flow Summary
Legend: `[R]` = Review Gate (human approval required)
```
interactive: 0 → 1 → [R] → 2 → [R] → 3 → [R] → 4 → [R] → 5(user) → 6
auto: 0 → 1 → 2 → 3 → 4 → 5(auto) → 6 → next phase (NO stops)
fast: 0 → skip → 2(fast) → [R] → 3 → [R] → 4 → [R] → 5(simple) → 6
parallel: 0 → 1? → [R] → 2(parallel) → [R] → 3(multi-agent) → [R] → 4 → [R] → 5(user) → 6
no-test: 0 → 1 → [R] → 2 → [R] → 3 → [R] → skip → 5(user) → 6
code: 0 → skip → skip → 3 → [R] → 4 → [R] → 5(user) → 6
```
**Key difference:** `auto` mode is the ONLY mode that skips all review gates.
## Critical Rules
- Never skip steps without mode justification
- **MANDATORY SUBAGENT DELEGATION:** Steps 4, 5, 6 MUST spawn subagents via Task tool. DO NOT implement directly.
- Step 4: `tester` (and `debugger` if failures)
- Step 5: `code-reviewer`
- Step 6: `project-manager`, `docs-manager`, `git-manager`
- Use `TaskCreate` to create Claude Tasks for each unchecked item with priority order and dependencies (or `TodoWrite` if Task tools unavailable).
- Use `TaskUpdate` to mark Claude Tasks `in_progress` when picking up a task (skip if Task tools unavailable).
- Use `TaskUpdate` to mark Claude Tasks `complete` immediately after finalizing the task (skip if Task tools unavailable).
- All step outputs follow format: `✓ Step [N]: [status] - [metrics]`
- **VALIDATION:** If Task tool calls = 0 at end of workflow, the workflow is INCOMPLETE.