收口 RN 移动端系统优化

.agents/skills/check-compiler-errors/SKILL.md

@@ -0,0 +1,48 @@
+---
+name: check-compiler-errors
+description: "运行并修复编译或类型检查错误。用户提到 compiler/type-check/tsc/类型错误/编译失败/Kotlin native build 失败/CI compile failed 时使用。"
+---
+
+# Check Compiler Errors
+
+Use this skill when compile or type-check failures block validation.
+
+## Command Order
+
+Prefer repo scripts over ad hoc commands:
+
+1. TypeScript: `npm run type-check`
+2. Lint when relevant: `npm run lint`
+3. Unit/package tests when type fixes affect behavior: `npm run test:unit` or `npm run test --workspaces --if-present`
+4. Android/native changes: `cd android && ./gradlew assembleDebug --no-daemon`
+
+If a script is missing, inspect `package.json` and use the closest existing command. Do not add a new dependency or script unless the user asks.
+
+## Workflow
+
+1. Run the smallest command that reproduces the compiler failure.
+2. Summarize errors by file and category.
+3. Fix the highest-confidence error first; prefer the canonical type boundary over casts.
+4. Re-run the same command.
+5. If one fix reveals the next compiler error, continue until clean or blocked.
+
+## N.E.K.O.-RN Notes
+
+- JS lip sync smoothing belongs in `LipSyncService`; Kotlin `PCMStreamPlayer` handles playback-head time alignment.
+- After Kotlin/native Android edits, Metro hot reload is insufficient; run the Gradle debug build.
+- For `@project_neko/*` package changes, check whether workspace tests or package sync are needed before declaring clean.
+
+## Guardrails
+
+- Never use `any` or broad casts just to quiet TypeScript unless there is no better local boundary.
+- Do not change runtime behavior while fixing a pure type error unless the compiler exposed a real bug.
+- Do not edit `.env` files.
+
+## Output
+
+```text
+Status: clean | blocked | still failing
+Commands: <commands run>
+Errors fixed: <files/categories>
+Remaining blocker: <only if any>
+```

.agents/skills/check-compiler-errors/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Check Compiler Errors"
+  short_description: "Run type checks and summarize compiler failures"
+  default_prompt: "Run compiler checks and fix high-confidence failures."

.agents/skills/deslop/SKILL.md

@@ -0,0 +1,39 @@
+---
+name: deslop
+description: "清理 AI 生成感明显的代码膨胀。用户要求 deslop、去掉 AI 味、清理冗余、提交前简化 diff、删除多余注释/防御式代码/无意义抽象时使用。"
+---
+
+# Deslop
+
+Use this skill to clean the current diff while preserving behavior.
+
+## Workflow
+
+1. Inspect `git status --short`, `git diff --staged`, and `git diff`.
+2. Limit edits to changed lines or directly adjacent code needed to simplify those lines.
+3. Remove only clear slop:
+   - Comments that restate code or explain obvious behavior.
+   - TODO/FIXME/HACK comments not explicitly requested by the user.
+   - Defensive checks, try/catch blocks, or fallbacks inconsistent with trusted local paths.
+   - `any`, broad casts, or optionality added only to silence TypeScript.
+   - One-off helpers, wrappers, flags, or configuration added for a single use.
+   - Nested conditionals that can be made direct with early returns.
+   - Debug logs or temporary output.
+4. Preserve existing local style even if another style would be preferable.
+5. Run the smallest relevant verification, usually `npm run type-check`, focused tests, or lint.
+
+## Guardrails
+
+- Keep behavior unchanged unless fixing a clear bug found during cleanup.
+- Do not refactor unrelated code.
+- Do not introduce new dependencies.
+- Do not delete files.
+- Do not edit `.env` files.
+- If cleanup would become a rewrite, stop and propose the smaller path first.
+
+## Output
+
+Report only:
+- What slop was removed.
+- Verification run and result.
+- Any remaining risk or skipped check.

.agents/skills/deslop/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Deslop"
+  short_description: "Remove AI-style code slop from a diff"
+  default_prompt: "Clean up AI-generated slop in the current diff."

.agents/skills/make-pr-easy-to-review/SKILL.md

@@ -0,0 +1,53 @@
+---
+name: make-pr-easy-to-review
+description: "让当前 diff 或 PR 更容易审查。用户说 make this easy to review、整理 PR、写 reviewer notes、拆分说明、审查前清理提交信息/描述时使用。"
+---
+
+# Make PR Easy To Review
+
+Use this skill to improve reviewability without changing product behavior.
+
+## Workflow
+
+1. Inspect context:
+   - `git status --short`
+   - `git diff --stat`
+   - `git diff --staged --stat`
+   - PR metadata with `gh pr view` only if a PR exists and the user asks about it.
+2. Identify reviewability issues:
+   - Mixed mechanical and behavioral edits.
+   - Stale or missing PR description.
+   - Generated files mixed with hand-written files.
+   - Large files where the reviewer needs an entry point.
+   - Missing test notes.
+3. Prefer non-code improvements first:
+   - Concise TL;DR.
+   - Changed-file map.
+   - Risk notes.
+   - Verification notes.
+4. If code cleanup is needed, use the `deslop` standard and keep behavior unchanged.
+
+## Guardrails
+
+- Do not rewrite history, squash commits, force-push, push, or open PRs without explicit approval.
+- Do not hide behavior changes under "cleanup".
+- Do not edit unrelated files.
+- If the PR is too large to make reviewable with notes, recommend splitting it instead of polishing around the problem.
+
+## Output
+
+Produce reviewer-ready notes:
+
+```markdown
+## TL;DR
+<one paragraph>
+
+## Reviewer Entry Points
+- <file>: <why it matters>
+
+## Risk
+- <risk or "Low">
+
+## Verification
+- <commands/results>
+```

.agents/skills/make-pr-easy-to-review/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Make PR Easy To Review"
+  short_description: "Prepare a diff for easier review without changing behavior"
+  default_prompt: "Make the current PR easier to review."

.agents/skills/run-smoke-tests/SKILL.md

@@ -0,0 +1,47 @@
+---
+name: run-smoke-tests
+description: "运行冒烟测试或最小端到端验证。用户要求 smoke test、跑一下关键路径、验证改动没有破坏主流程、改完跑测试时使用。"
+---
+
+# Run Smoke Tests
+
+Use this skill to validate the smallest meaningful user path after a change.
+
+## Test Selection
+
+Pick the narrowest check that can catch the likely regression:
+
+- Type-only change: `npm run type-check`
+- Shared TS behavior: `npm run test:unit`
+- Package behavior: `npm run test --workspaces --if-present`
+- Lint-sensitive edit: `npm run lint`
+- Android/native edit: `cd android && ./gradlew assembleDebug --no-daemon`
+- Mobile dev stack behavior: `npm run dev:mobile-stack` only when the user asks for interactive stack validation.
+
+Use `npm test` only when broad validation is warranted; it can be slower than a focused check.
+
+## Workflow
+
+1. Identify the changed behavior and the smoke surface.
+2. Run prerequisites only when required.
+3. Run the focused command.
+4. If failing, inspect the first actionable failure and apply a minimal fix.
+5. Re-run the same command. Broaden only if the fix touches a shared path.
+
+## Guardrails
+
+- Prefer deterministic assertions over waiting or manual inspection.
+- Do not add Playwright, Detox, or other test dependencies without explicit approval.
+- Do not start an emulator unless explicitly requested.
+- Do not edit `.env` files.
+
+## Output
+
+Keep the result short:
+
+```text
+Smoke surface: <why this check was chosen>
+Result: passed | failed | blocked
+Command: <command>
+Failure/fix: <only if relevant>
+```

.agents/skills/run-smoke-tests/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Run Smoke Tests"
+  short_description: "Run focused smoke validation and triage failures"
+  default_prompt: "Run the relevant smoke tests and report failures."

.agents/skills/thermo-nuclear-code-quality-review/SKILL.md

@@ -0,0 +1,53 @@
+---
+name: thermo-nuclear-code-quality-review
+description: "极严格的可维护性审查。用户要求 thermo-nuclear review、严厉代码质量审查、深度 maintainability audit、检查抽象/大文件/复杂度/意大利面式条件分支时使用。"
+---
+
+# Thermo-Nuclear Code Quality Review
+
+Use this skill for a strict maintainability review of a diff or branch. Findings come first.
+
+## Review Standard
+
+Be ambitious about simplification. Do not accept "it works" if the change makes the system harder to reason about.
+
+Check for:
+
+- A simpler structure that removes whole branches, helpers, flags, or modes.
+- Files pushed toward or past 1,000 lines without a strong reason.
+- New one-off conditionals inserted into unrelated flows.
+- Business logic leaking into UI, hooks, or shared plumbing.
+- Service-layer code depending on React lifecycle or hooks.
+- Hook code containing business logic instead of event-to-state mapping.
+- Components importing service instances directly instead of consuming hooks.
+- `any`, `unknown`, broad casts, or optionality that hide unclear boundaries.
+- Sequential orchestration where independent work should be parallel or atomic.
+- Thin abstractions that add indirection without reducing complexity.
+
+## Workflow
+
+1. Collect `git status --short`, `git diff --stat`, and the relevant diff.
+2. Read changed files enough to understand boundaries and local style.
+3. Review by severity, not by file order.
+4. Report only actionable issues grounded in file/line references.
+5. If there are no issues, say so and name remaining test gaps.
+
+## Guardrails
+
+- Do not nitpick formatting while structural issues exist.
+- Do not request broad refactors unless the diff introduced or exposed the complexity.
+- Do not propose new dependencies.
+- Do not rewrite code during review unless the user asks for fixes.
+
+## Output
+
+```text
+Findings
+- [P0/P1/P2/P3] file:line - issue, impact, recommended fix
+
+Open questions
+- <only if blocking>
+
+Test gaps / residual risk
+- <only if relevant>
+```

.agents/skills/thermo-nuclear-code-quality-review/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Thermo Nuclear Review"
+  short_description: "Run a strict maintainability review"
+  default_prompt: "Run a strict maintainability review of the current diff."

.agents/skills/verify-this/SKILL.md

@@ -0,0 +1,44 @@
+---
+name: verify-this
+description: "验证一个具体声明是否成立。用户说“verify this / 证明它好了 / 看看是否修复 / 给证据 / 是否真的变快”等需要 before/after、本地测试、日志、截图或构建产物证据时使用。"
+---
+
+# Verify This
+
+Use this skill to prove or disprove a falsifiable claim with local evidence. Do not turn it into a recap.
+
+## Workflow
+
+1. Restate the claim in measurable form: condition, expected behavior, and pass/fail threshold.
+2. Pick the smallest local surface that can disprove it:
+   - TypeScript behavior: focused unit test, `npm run test:unit`, or a small repro.
+   - Workspace packages: `npm run test --workspaces --if-present`.
+   - Compile safety: `npm run type-check`.
+   - Lint safety: `npm run lint`.
+   - Android/native behavior: `cd android && ./gradlew assembleDebug --no-daemon`.
+   - UI behavior: local screenshot or browser/app interaction evidence when applicable.
+3. Capture baseline and treatment with the same command, input, environment, and threshold.
+4. If a real baseline is unavailable without reverting user changes, say so and use `INCONCLUSIVE`; do not reset or overwrite user work.
+5. Compare raw artifacts, not impressions.
+6. Return exactly one verdict: `VERIFIED`, `NOT VERIFIED`, or `INCONCLUSIVE`.
+
+## Guardrails
+
+- Never modify `.env` files.
+- Never run destructive git commands to create a baseline.
+- Keep artifacts minimal; prefer terminal output snippets unless screenshots/logs are necessary.
+- If writing artifacts, use `/tmp/verify-this/<claim-slug>/`, not the repo, unless the user asks.
+- Do not soften a negative result. A clear `NOT VERIFIED` is useful.
+
+## Output
+
+```text
+VERIFIED | NOT VERIFIED | INCONCLUSIVE
+Claim: <falsifiable claim>
+
+Evidence:
+- <command/artifact>: baseline=<...>, treatment=<...>, delta=<...>, threshold=<...>
+
+Reasoning:
+<one concise paragraph naming evidence and confounds>
+```

.agents/skills/verify-this/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Verify This"
+  short_description: "Verify a claim with before/after evidence"
+  default_prompt: "Verify this claim with local evidence."

.agents/skills/workflow-from-chats/SKILL.md

@@ -0,0 +1,47 @@
+---
+name: workflow-from-chats
+description: "从用户反馈或近期对话中沉淀可复用工作流。用户要求记住偏好、把经验写成 skill/rule/AGENTS.md、以后按这个方式做、从聊天中总结工作习惯时使用。"
+---
+
+# Workflow From Chats
+
+Use this skill to turn repeated user feedback into durable instructions.
+
+## Workflow
+
+1. Define the target surface: communication, review, debugging, validation, shipping, simplification, or repo-specific workflow.
+2. Extract preference atoms from the current conversation and available repo instructions:
+   - trigger
+   - decision rule
+   - required checks
+   - stop condition
+   - examples of what to avoid
+   - confidence
+3. Classify confidence:
+   - strong: explicit user rule or repeated correction
+   - medium: accepted workflow with clear evidence
+   - weak: one-off inference
+   - contradicted: ask before writing
+4. Choose the artifact:
+   - Skill: recurring multi-step workflow.
+   - AGENTS.md rule: broad behavior preference.
+   - Existing skill edit: refinement of a known command.
+   - No file change: situational or low-confidence note.
+5. Write only reusable guidance. Remove anecdotes and private details.
+
+## Guardrails
+
+- Do not expose local transcript paths, secrets, private messages, or credentials.
+- Do not rewrite global preferences from weak evidence.
+- Do not create broad rules from a one-off task.
+- Prefer editing existing project guidance over adding a new file when the behavior is already covered.
+
+## Output
+
+```text
+Target workflow: <surface>
+Adopt: <strong guidance to write>
+Consider: <medium-confidence guidance>
+Dismissed: <weak or task-specific observations>
+Artifact: <skill/rule/doc/no change>
+```

.agents/skills/workflow-from-chats/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Workflow From Chats"
+  short_description: "Turn repeated preferences into durable guidance"
+  default_prompt: "Extract reusable workflow guidance from recent feedback."

.gitignore

@@ -26,6 +26,7 @@ npm-debug.*
 yarn-debug.*
 yarn-error.*
 .npm-cache/
+.devlogs/
 
 # macOS
 .DS_Store