AgentSkillsCN

agent-fix-lint-plus

在本地仅运行 lint 步骤，应用确定性的 lint 自动修复，并在需要人工干预时给出清晰的指导（或发出 Codex 修复胶囊）。当用户请求仅进行 lint 修复，或当 lint CI 步骤失败时，此技能将为您提供帮助。

SKILL.md
--- frontmatter
name: agent-fix-lint-plus
description: "Run lint-only CI steps locally, apply deterministic lint auto-fixes, and stop with clear guidance (or emit a Codex repair capsule) when manual intervention is required. Use when lint-only fixing is requested or lint CI steps fail."

Agent fix:lint+

Follow the CI source of truth

  • •Discover lint steps by reading .github/workflows/ci.yml.
  • •Select steps whose name includes "lint" (case-insensitive), in order.
  • •Run the exact run: command with its specified working-directory.

Deterministic lint fix loop

  1. •Run each lint step.
  2. •If a lint step fails, apply deterministic lint auto-fixes only:
    • •pnpm lint:fix if it exists in package.json.
    • •turbo lint -- --fix if turbo lint is already used in repo scripts.
    • •pnpm -r lint -- --fix only when recursive lint is already used.
    • •Fallback: pnpm exec eslint in apps/web and/or packages/shared only when the failing output indicates those packages.
  3. •Re-run the failing lint step.
  4. •Stop after 2 fix attempts total.

Safety guardrails

  • •Never modify .github/workflows/**.
  • •Never change lint rules or configuration.
  • •Never weaken CI.
  • •Never auto-commit changes.
  • •Never delete code to satisfy lint (except clearly unused imports/vars).

Use the bundled scripts

  • •Run node skills/agent-fix-lint-plus/agent-fix-lint-plus.mjs to execute lint-only checks.
  • •Run node skills/agent-fix-lint-plus/agent-fix-lint-plus.mjs --codex to emit a Codex repair capsule when lint cannot be auto-fixed.
  • •Reuse helpers from skills/agent-fix-ci/scripts/* for workflow parsing and capsule output.

Codex repair mode contract

  • •When --codex is provided and lint remains failing, the runner emits:
    • •CODEX_REPAIR_REQUIRED
    • •A single-line JSON capsule describing the failure.
  • •The capsule contains the failing step, failure reason, output tail, guardrails, allowed actions, and a rerun command.

Codex guardrails (automated)

  • •The --codex runner tracks progress in .codex/agent-fix-lint-plus.json.
  • •Guardrails enforced between runs:
    • •Max iterations: 7
    • •Max files changed per iteration: 10
    • •Max total lines changed per iteration: 300
    • •Stop if .github/workflows/** is modified
    • •Stop if the same failing step + reason repeats twice

Expected output

  • •Print each lint step name, command, and working directory.
  • •Summarize what failed, what was fixed, and why it stopped (if it did).
  • •End with:
    • •Lint checks green — ready to commit, or
    • •Stopped: manual intervention required (Lint failed (eslint).)