agent-dev-workflow（Agent 开发工作流程）

Dev Workflow — Orchestrated Coding Agent Dispatch

Structured 6-phase gated workflow for driving coding agents to implement requirements with quality control.

Prerequisites

• Claude Code (claude CLI) installed with --permission-mode bypassPermissions --print
• cc-plugin installed in Claude Code (provides llmdoc read/write, investigator/scout/recorder sub-agents)
• Global ~/.claude/CLAUDE.md configured with cc-plugin settings

Workflow Overview

Phase 1: Environment Check
    ↓
Phase 2: Spec  ← spec-writer skill
    ↓ [user confirms]
Phase 3: Task Planning
    ↓ [user confirms]
Phase 4: Agent Dispatch
    ↓
Phase 5: Verification
    ↓
Phase 6: Delivery

Each phase must complete before moving to the next. Phase 2 and Phase 3 have explicit user confirmation gates.

Phase 1: Environment Check

Before any work, check the target project:

1. llmdoc/ exists?

- Yes → read llmdoc/index.md + overview files to understand project - No → dispatch Claude Code with /tr:initDoc to generate it first

1. CLAUDE.md in project root? — verify cc-plugin config is present; fix if missing
2. git status — ensure working tree is clean enough to work on

Phase 2: Spec

Generate a structured implementation spec using the spec-writer skill.

1. Activate spec-writer skill — read its SKILL.md and follow its workflow
2. It will gather project context, draft a spec using its template, and present to the user
3. The spec covers: objectives, user stories, technical plan, boundaries, verification criteria, and a preliminary task breakdown
4. 🚫 Gate: Do not proceed until the user confirms the spec.

The confirmed spec replaces the old requirement-doc + verification-doc pair — everything is in one document now.

Fallback: If spec-writer skill is not installed, use the templates in references/ (requirement-template.md + verification-template.md) as before.

Phase 3: Task Planning

Refine the spec's preliminary task breakdown into a precise, executable task list. This is the bridge between "what to build" and "how to tell the agent."

3.1 Extract Tasks from Spec

Start from Section 10 (Task Breakdown) of the confirmed spec. For each task, expand it into a Task Card:

### Task [#]: [Title]

**Objective:** What this task accomplishes (one sentence)
**Input:** What must exist before this task starts (files, tables, APIs from prior tasks)
**Output:** What this task produces (new/modified files, passing tests, working endpoint)
**Affected files:**
- `src/xxx.py` — create / modify
- `tests/test_xxx.py` — create

**Agent instructions:**
- [Specific, actionable instruction 1]
- [Specific, actionable instruction 2]
- Reference: [relevant spec section, e.g., "See Spec §4.2 for design decision"]

**Verification:**
- [ ] [How to confirm this task is done — a command, a test, an observable result]

**Depends on:** Task #X, #Y (or "none")
**Complexity:** Low / Medium / High
**Parallel:** Can run in parallel with Task #Z (or "no")

3.2 Scoping Rules

Each task should be:

• Independently implementable — an agent can complete it without context from other incomplete tasks
• Independently verifiable — there is a concrete way to check it worked (test passes, endpoint responds, file exists)
• Small enough to review — aim for changes reviewable in one pass (rough guide: <300 lines changed)
• Large enough to be meaningful — don't split a 20-line function into 3 tasks

Splitting heuristics:

• If a task touches >3 unrelated modules → split by module
• If a task has both "create infrastructure" and "implement logic" → split
• If a task requires multiple design decisions → split by decision boundary
• If you can describe two independent verification criteria → likely two tasks

3.3 Dependency Graph

After expanding all tasks, produce a dependency summary:

Task 1 (Low)  ─┐
Task 3 (Low)  ─┤─→ Task 5 (Medium) ─→ Task 7 (Medium)
Task 2 (Medium) ─→ Task 4 (Medium) ─┤
Task 6 (High) ────────────────────────┘─→ Task 8 (Medium)

Identify:

• Critical path — longest chain of dependent tasks
• Parallel groups — tasks that can run simultaneously
• Execution order — the sequence to dispatch agents

3.4 Context Budgeting

For each task, determine what context the agent needs:

• Spec sections to include — only the relevant parts, not the full spec (avoid context overload)
• Existing code to reference — specific files, not "the whole src/"
• Boundaries to emphasize — the ✅/⚠️/🚫 rules most relevant to this task

Principle: feed the agent only what it needs for the current task. A well-fed agent is like a well-fed function — give it only the inputs for the job at hand.

3.5 User Confirmation

Present the task list to the user with:

• All task cards (or a summary table + full cards on request)
• Dependency graph
• Estimated execution order (serial vs parallel)
• Highlight the critical path

🚫 Gate: Do not dispatch any agents until the user confirms the task plan.

Common discussion points:

• Is the granularity right? (too fine → overhead; too coarse → risky)
• Are dependencies correct? Can anything be parallelized further?
• Any task missing from the spec's intent?

Phase 4: Agent Dispatch

Execute the confirmed task plan by dispatching coding agents.

4.1 Prompt Construction

For each task, construct the agent prompt from the task card:

claude --permission-mode bypassPermissions --print '<task prompt>' 2>&1

Task prompt structure:

## Task: [Title]

### Objective
[From task card]

### Context
[Relevant spec sections — copy only what's needed]
[Relevant existing code snippets or file references]

### Instructions
[Agent instructions from task card]

### Constraints
- 不要 git commit
- 完成后执行 /update-doc 更新 llmdoc
- [Relevant boundaries from spec §7]

### Verification
When done, confirm:
- [Verification criteria from task card]

Use workdir to scope the agent to the project directory. Use background: true for long-running tasks, monitor with process tool.

4.2 Execution Strategy

• Serial tasks (has dependencies): wait for dependency to complete and verify before dispatching
• Parallel tasks (independent): dispatch simultaneously, use git worktrees if touching overlapping files
• After each task completes:

1. Check agent output for errors 2. Run the task's verification criteria 3. If verification fails → re-prompt the agent with the failure details (up to 2 retries) 4. If still fails → stop and report to user

4.3 Error Handling

Phase 5: Verification

After all tasks complete, verify the whole against the spec's verification criteria (Section 9):

1. Automated checks (§9.1) — run tests, build, lint, type check
2. Functional verification (§9.2) — execute each test scenario step by step
3. Edge cases & error handling (§9.3) — verify each edge case
4. Regression (§9.4) — confirm existing functionality unaffected
5. Code quality — style consistency, no stray files, no hardcoded secrets

If any verification fails:

• Identify which task(s) caused the issue
• Re-dispatch agent to fix, providing the failure context
• Re-verify after fix

Phase 6: Delivery

Present results to the user with:

1. All changed files — every file touched, including llmdoc updates, config changes, everything
2. Suggested commit message — conventional commits format
3. Task completion summary — which tasks passed, any retries needed
4. Verification results — what passed, any caveats
5. Items for human review — what needs the user's attention

Never commit. The user handles all git commits.

Agent Selection

Default: Claude Code (claude --permission-mode bypassPermissions --print)

Key Constraints

• Never commit — code changes only, user commits
• Never run agents in ~/.openclaw/ — agents will read soul/identity files
• Always update llmdoc — every agent run should end with doc update
• Always list all changed files on delivery
• Always provide commit message on delivery
• Interrupt on design issues — don't let agents drift on wrong assumptions
• Feed minimal context — each agent gets only spec sections and files relevant to its task