Cursor rule

1---
2description: "Delegate one complex task to a single subagent, review its work in two stages before merging back. Sequential — one agent at a time, with oversight. Use when a task is complex and requires careful review before the result is accepted. Distinct from dispatch-agents (no parallelism here; reviewer sees full diff before proceeding)."
3alwaysApply: false
4---
5 
6# story: e45s30
7 
8# Delegate Task
9> **HARD GATE** — **HARD GATE** — Delegated work must have clear success criteria and verification commands. The delegate must be able to verify completion independently.
10 
11 
12Delegate a single complex task to a subagent with a two-stage review gate before accepting the result. Use when oversight of a single task matters more than speed.
13 
14**Distinct from `dispatch-agents`:** This skill runs one subagent sequentially with a mandatory review. `dispatch-agents` runs multiple subagents in parallel without inter-task review gates.
15 
16## Subagent depth tiers (e45s30)
17 
18Select brief depth from task `risk:` and skill `effort:` before spawning:
19 
20| Tier | When | Brief includes |
21|------|------|----------------|
22| `full_maturity` | P0 stories, multi-file refactors, security work | Full template + CONVENTIONS excerpts + threat model if present |
23| `standard` | Default implementation tasks | Goal, scope, out-of-bounds, constraints, verify, prior decisions |
24| `minimal_decisive` | Light probes, read-only audits | Goal, verify, explicit file list (≤15 lines total) |
25 
26State `depth: <tier>` in the Agent tool description field.
27 
28## Process
29 
30### 1. Define the task
31 
32Before spawning the agent, read `specs/state.yaml` if it exists. Then write a minimal self-contained brief using this template (brief size directly controls token cost and hallucination risk — do not pad):
33 
34```
35Goal: [one sentence — specific, measurable outcome]
36In scope: [explicit file or module list]
37Out of bounds: [what NOT to do]
38Constraints: [relevant CONVENTIONS.md rules, existing patterns, test requirements]
39Verify: [runnable command]
40Prior decisions: [relevant entries from specs/state.yaml — omit section if none apply]
41```
42 
43Do not include full file contents, full conversation history, or decisions unrelated to this task.
44 
45### 2. Spawn the subagent (iterative retrieval, max 3 cycles)
46 
47Use the Agent tool with a **fresh context** per spawn. Pass prior decisions only via `specs/state.yaml`.
48 
49**Cycle:** dispatch → evaluate output vs goal → refine brief → re-spawn if needed (max 3 cycles).
50 
51Include in each brief:
52- All context the agent needs (it starts cold — no shared state)
53- Reference to CONVENTIONS.md constraints
54- The verify command it must run before reporting done
55 
56### 3. Stage 1 review — output inspection
57 
58When the subagent returns, review its report before looking at the diff:
59- Did it run the verify command? Did it pass?
60- Does it explain what it changed and why?
61- Are there any concerns raised by the agent?
62 
63If the report raises red flags, ask the subagent for clarification or re-run with adjusted instructions.
64 
65### 4. Stage 2 review — diff inspection
66 
67Inspect the actual diff:
68```bash
69git diff main...HEAD
70```
71 
72Check:
73- [ ] Changes are scoped to what was asked — nothing extra
74- [ ] No `any`, no `@ts-ignore`, no disabled lint rules
75- [ ] Tests added for new behavior
76- [ ] CONVENTIONS.md compliance (naming, structure, no gh issue creation)
77- [ ] Boy Scout Rule: touched areas are cleaner than before
78 
79### 5. Decision
80 
81- **Accept**: merge the result into the main working branch
82- **Revise**: send back to the subagent with specific feedback
83- **Reject**: discard and re-approach differently
84 
85**After accepting**, append to `specs/state.yaml` under `## Active Decisions`:
86```
87**[task short name]**: [what approach the agent chose and why — one sentence]
88```
89 
90Report the decision and rationale to the user.
91