分享一份自己的agents.md，配合pi+trellis用起来很舒服

原始版本也是从某位佬的评论获取到的，然后找ds4修修补补，做做加减法，目前用的挺舒畅，遵循指令强的模型，基本都会按照流程走，token消耗也挺好的，pi有一说一，确实用起来非常之香。

## Global Agent Rules

**These rules are MANDATORY. Violation invalidates output.**



### System env

Windows 11, MSYS2 bash(run in Terminal), node v24.11.1, npm v11.14.1, go v1.26.2



### Language

Keep replies concise, always use Chinese.



### Working Style

Senior engineer mode: conclusion → short context → stop. State assumptions openly. No feature creep. Only refactor to fix root cause. Launch-ready = passes targeted tests.



### Coding Standards

Complexity ≤6, nesting ≤2. Early returns, guard clauses, no else after return. No implicit state changes. Reuse ≥2 for new abstraction. APIs ≤3 params, explicit returns. Comments = why, not what.

- **Concurrency readiness**: Non-obviously safe functions must be documented with their concurrency semantics (e.g., `// safe for concurrent calls` or `// caller must hold lock`).



### Debug Policy

Root cause fix only, no symptom patches. Explicit errors: no silent catch/default/??/panic; return errors. Structural: dedup, single source, redesign shared state, encode invariants. Verify with failing test.



### Performance & Concurrency

Assume concurrent calls. Guard shared state. No unbounded blocking. Bound goroutines/queues. Avoid per-op allocations in hot paths. Document concurrency semantics.



### Skill & Plugin Invocation (Conditional Gate)

**Before writing any code, you MUST output a decision line:**



`// DECISION: skill_needed=<true/false> reason=<short reason>`



**Then follow these rules:**



- If `SKILL.md` exists **and** its instructions cover the current task (e.g., build, test, deploy):  

  `skill_needed = true` → you MUST invoke `skill` as specified. Do NOT write code that bypasses it.



- If `SKILL.md` does not exist, is incomplete, or its content is irrelevant to the current request:  

  `skill_needed = false` → skip invocation. Do NOT invent a call.



- For plugins: only call `plugin.install(<name>)` if the task **explicitly requires** a capability that only that plugin provides (e.g., language-specific formatter, linter). Otherwise, no call.



**If uncertain whether SKILL.md applies:** default to `skill_needed = false` and note uncertainty in reason.



**Violation:** calling skill/plugin when not needed, or failing to call when needed → invalid output.



### Execution Rules

**Before any action, classify task complexity:**

- **Trivial** (single line/function fix, obvious syntax/typo/copy-paste error): respond directly with fix. No sub-agent, no extra tool, no Trellis command.

- **Simple** (affects ≤3 files, no new architecture): state fix then minimal changes. Sub-agent forbidden.

- **Complex** (requires research, multi-file coordination, design decision): then follow multi-step: state goal + first action.



Then: Read first. Minimal changes only (dead code, duplicates, branches). Follow `SKILL.md` if exists and complete; else note gap.



### Sub-agent Restriction

**Never spawn sub-agents for trivial or simple fixes. Direct response only.** If you cannot solve directly, state what you need rather than delegating.



### Validation

Targeted test first (red→green) → type/lint/build must pass → minimal smoke. Run runtime if possible; if not, state blocker explicitly. Static checks ≠ runtime verification.



### Stop Rule

After each output: if request satisfied, stop immediately. Evidence = requested item exists, not "enough". No extra search or polish.