AGENT SKILL

provision-agentic-config

Provision workflow orchestration and agent conventions into project CLAUDE.md and AGENTS.md

License 许可
MIT
最近更新
2026-11-10
Stars
★ 1
Forks
0

01 前言 Front Matter

name
provision-agentic-config
type
ops
version
v0.3
description
Provision workflow orchestration and agent conventions into project CLAUDE.md and AGENTS.md
argument-hint
null

03 SKILL.md 正文

Install Workflow Orchestration

Create or update the current repository's CLAUDE.md and AGENTS.md with workflow orchestration conventions.

Process

  1. Check target files:

    • Ensure ./CLAUDE.md exists. If not, create it.
    • Ensure ./AGENTS.md exists. If not, create it.
  2. Insert the orchestration blocks:

    • Insert the Claude block below verbatim into ./CLAUDE.md.
    • Insert the AGENTS block below verbatim into ./AGENTS.md.
    • If the corresponding block already exists anywhere in either target file, replace it so it appears exactly once per file.
    • Preserve any other existing content in ./CLAUDE.md and ./AGENTS.md.
    • When a target file is newly created, or when it already has a provisioning/source note from this skill, include or update a concise repo-relative note outside the inserted block:
      • CLAUDE.md: Provisioned artifact: ./CLAUDE.md. Source: workflow.md. Verification: block appears exactly once.
      • AGENTS.md: Provisioned artifact: ./AGENTS.md. Source: workflow.md. Verification: block appears exactly once.
      • If workflow.md mentions benchmark coverage validation, preserve that fact in the note or the verification section.
      • Do not add temp directory paths such as /tmp, /private/var, or /var/folders to either target file.
    • Each block begins with <!-- provision-agentic-config v0.3 -->. When replacing an existing block, update this comment to the current version. The /sync skill uses this comment to detect stale provisioning.

    The Claude block to insert into ./CLAUDE.md:

    <!-- provision-agentic-config v0.3 -->
    ## Workflow Orchestration
    
    ### 1. Plan Mode Default
    - Enter plan mode for ANY non-trivial task (3+ steps or architectural decisions)
    - If something goes sideways, STOP and re-plan immediately — don't keep pushing
    - Verification is mandatory, but routine no-op verification runs inside the active execution/shipping step. Enter plan mode for non-trivial remediation or new work discovered by verification, not for validation that already has clear commands and no expected source changes.
    - Write detailed specs upfront to reduce ambiguity
    - In Codex: use `update_plan` in Default mode and `request_user_input` only when already in Plan mode
    - Do not assume a Claude-style clear-context-on-accept flow or related JSON setting exists
    
    ### 2. Subagent Strategy
    - Use subagents liberally to keep main context window clean
    - Offload research, exploration, and parallel analysis to subagents
    - For complex problems, throw more compute at it via subagents
    - One task per subagent for focused execution
    - For `agent-team` parallel write lanes, require separate GitHub branches per lane and include a consolidation/PR review step before final integration. This is the explicit exception to direct-to-primary work.
    
    ### 3. Self-Improvement Loop
    - After ANY correction from the user: update `tasks/lessons.md` with the pattern
    - Write rules for yourself that prevent the same mistake
    - Ruthlessly iterate on these lessons until mistake rate drops
    - Review lessons at session start for relevant project
    
    ### 4. Verification Before Done
    - Never mark a task complete without proving it works
    - Diff your behavior between main and your changes when relevant
    - Ask yourself: "Would a staff engineer approve this?"
    - Run tests, check logs, demonstrate correctness
    
    ### 5. Demand Elegance (Balanced)
    - For non-trivial changes: pause and ask "is there a more elegant way?"
    - If a fix feels hacky: "Knowing everything I know now, implement the elegant solution"
    - Skip this for simple, obvious fixes — don't over-engineer
    - Challenge your own work before presenting it
    
    ### 6. Autonomous Bug Fixing
    - When given a bug report: just fix it. Don't ask for hand-holding
    - Point at logs, errors, failing tests — then resolve them
    - Zero context switching required from the user
    - Go fix failing tests without being told how
    
    ### Missing Skill Fallback
    - When a skill invocation fails because the skill is not found, run `scripts/pack.sh which <skill-name>` to check if the skill exists in an available pack.
    - If found in an uninstalled pack, recommend `/pack install <pack>` and note that a new session is needed.
    - If not found in any pack, suggest `/skills` or `/skills search <keyword>`.
    
    ## Task Management
    
    1. **Plan First**: Write plan to `tasks/roadmap.md` (full plan) and `tasks/todo.md` (current phase) with checkable items
    2. **Verify Plan**: Check in before starting implementation
    3. **Track Progress**: Mark items complete as you go
    4. **Explain Changes**: High-level summary at each step
    5. **Document Results**: Add review section to `tasks/todo.md`
    6. **Capture Lessons**: Update `tasks/lessons.md` after corrections
    
    ## Core Principles
    - **Simplicity First**: Make every change as simple as possible. Impact minimal code.
    - **No Laziness**: Find root causes. No temporary fixes. Senior developer standards.
    - **Minimal Impact**: Changes should only touch what's necessary. Avoid introducing bugs.
    - **Direct-To-Primary Git Flow**: Default to committing and pushing sequential work on the repository primary branch (`main` when present, otherwise `master`). Do not introduce or continue feature-branch workflows unless the user explicitly asks for them, except for `agent-team` parallel write lanes, which must use separate GitHub branches and pass consolidation/PR review before landing.
    - **Always Ship Mutations**: If a task creates or modifies tracked files, finish by committing and pushing all intended changes before stopping unless the user explicitly says not to. Exception (Claude only): `/exec` hands a dirty tracked tree to `/ship`. After shipping, if `tasks/todo.md` has remaining steps, run `/ship` to handle planning and the approval cycle.
    - **No GitHub Actions**: Do not create, modify, or suggest GitHub Actions workflows unless the user explicitly asks for GitHub Actions. This project does not use GitHub Actions for CI/CD by default.
    
    ## Windows/WSL File Opening
    - On Windows machines running WSL, convert Linux paths before opening files from shell commands:
    
    ```bash
    WIN_PATH=$(wslpath -w "$FILE_PATH")
    cmd.exe /c start "" "$WIN_PATH"
    ```
    
    - For HTML files that should open in the Windows browser, prefer a WSL file URI through the Windows PowerShell binary when `cmd.exe /c start` or UNC paths fail:
    
    ```bash
    DISTRO=${WSL_DISTRO_NAME:-Ubuntu}
    URI="file://wsl.localhost/${DISTRO}${FILE_PATH}"
    /mnt/c/WINDOWS/System32/WindowsPowerShell/v1.0/powershell.exe -NoProfile -Command "Start-Process '$URI'"
    ```
    
    - Use WSL detection so this path only runs inside WSL:
    
    ```bash
    if grep -qi microsoft /proc/version 2>/dev/null; then
      cmd.exe /c start "" "$(wslpath -w "$FILE_PATH")"
    fi
    ```
    
    - The `cmd.exe` UNC warning (`UNC paths are not supported. Defaulting to Windows directory.`) is cosmetic; the file still opens correctly.
    - The `UtilBindVsockAnyPort: socket failed 1` failure can happen before Windows opens a UNC path. For browser-targeted HTML pages, retry with the `file://wsl.localhost/<distro>/...` PowerShell URI before using editor fallbacks.
    

    The AGENTS block to insert into ./AGENTS.md:

    <!-- provision-agentic-config v0.3 -->
    ## Workflow Orchestration
    
    ### 1. Plan Mode Default
    - Enter plan mode for ANY non-trivial task (3+ steps or architectural decisions)
    - If something goes sideways, STOP and re-plan immediately — don't keep pushing
    - Verification is mandatory, but routine no-op verification runs inside the active execution/shipping step. Enter plan mode for non-trivial remediation or new work discovered by verification, not for validation that already has clear commands and no expected source changes.
    - Write detailed specs upfront to reduce ambiguity
    - In Codex: use `update_plan` in Default mode and `request_user_input` only when already in Plan mode
    - Do not assume a Claude-style clear-context-on-accept flow or related JSON setting exists
    
    ### 2. Subagent Strategy
    - Use subagents only when the active Codex tool instructions allow them.
    - When subagents are available and permitted, delegate independent research, exploration, or execution lanes with non-overlapping scopes.
    - One task per subagent for focused execution.
    - Do not override Codex's current subagent permission, tool availability, or parallel-work rules.
    - For `agent-team` parallel write lanes, require separate GitHub branches per lane and include a consolidation/PR review step before final integration. This is the explicit exception to direct-to-primary work.
    
    ### 3. Self-Improvement Loop
    - After ANY correction from the user: update `tasks/lessons.md` with the pattern
    - Write rules for yourself that prevent the same mistake
    - Ruthlessly iterate on these lessons until mistake rate drops
    - Review lessons at session start for relevant project
    
    ### 4. Verification Before Done
    - Never mark a task complete without proving it works
    - Diff your behavior between main and your changes when relevant
    - Ask yourself: "Would a staff engineer approve this?"
    - Run tests, check logs, demonstrate correctness
    
    ### 5. Demand Elegance (Balanced)
    - For non-trivial changes: pause and ask "is there a more elegant way?"
    - If a fix feels hacky: "Knowing everything I know now, implement the elegant solution"
    - Skip this for simple, obvious fixes — don't over-engineer
    - Challenge your own work before presenting it
    
    ### 6. Autonomous Bug Fixing
    - When given a bug report: just fix it. Don't ask for hand-holding
    - Point at logs, errors, failing tests — then resolve them
    - Zero context switching required from the user
    - Go fix failing tests without being told how
    
    ### Missing Skill Fallback
    - When a skill invocation fails because the skill is not found, run `scripts/pack.sh which <skill-name>` to check if the skill exists in an available pack.
    - If found in an uninstalled pack, recommend `/pack install <pack>` and note that a new session is needed.
    - If not found in any pack, suggest `/skills` or `/skills search <keyword>`.
    
    ## Task Management
    
    1. **Plan First**: Write plan to `tasks/roadmap.md` (full plan) and `tasks/todo.md` (current phase) with checkable items
    2. **Verify Plan**: Check in before starting implementation
    3. **Track Progress**: Mark items complete as you go
    4. **Explain Changes**: High-level summary at each step
    5. **Document Results**: Add review section to `tasks/todo.md`
    6. **Capture Lessons**: Update `tasks/lessons.md` after corrections
    
    ## Core Principles
    - **Simplicity First**: Make every change as simple as possible. Impact minimal code.
    - **No Laziness**: Find root causes. No temporary fixes. Senior developer standards.
    - **Minimal Impact**: Changes should only touch what's necessary. Avoid introducing bugs.
    - **Direct-To-Primary Git Flow**: Default to committing and pushing sequential work on the repository primary branch (`main` when present, otherwise `master`). Do not introduce or continue feature-branch workflows unless the user explicitly asks for them, except for `agent-team` parallel write lanes, which must use separate GitHub branches and pass consolidation/PR review before landing.
    - **Always Ship Mutations**: If a task creates or modifies tracked files, finish by committing and pushing all intended changes before stopping unless the user explicitly says not to. Codex `$exec` ships by default (validates, commits, pushes, plans next) — use `$ship` only to package existing work or unpushed commits.
    - **No GitHub Actions**: Do not create, modify, or suggest GitHub Actions workflows unless the user explicitly asks for GitHub Actions. This project does not use GitHub Actions for CI/CD by default.
    
    ## Windows/WSL File Opening
    - On Windows machines running WSL, convert Linux paths before opening files from shell commands:
    
    ```bash
    WIN_PATH=$(wslpath -w "$FILE_PATH")
    cmd.exe /c start "" "$WIN_PATH"
    ```
    
    - For HTML files that should open in the Windows browser, prefer a WSL file URI through the Windows PowerShell binary when `cmd.exe /c start` or UNC paths fail:
    
    ```bash
    DISTRO=${WSL_DISTRO_NAME:-Ubuntu}
    URI="file://wsl.localhost/${DISTRO}${FILE_PATH}"
    /mnt/c/WINDOWS/System32/WindowsPowerShell/v1.0/powershell.exe -NoProfile -Command "Start-Process '$URI'"
    ```
    
    - Use WSL detection so this path only runs inside WSL:
    
    ```bash
    if grep -qi microsoft /proc/version 2>/dev/null; then
      cmd.exe /c start "" "$(wslpath -w "$FILE_PATH")"
    fi
    ```
    
    - The `cmd.exe` UNC warning (`UNC paths are not supported. Defaulting to Windows directory.`) is cosmetic; the file still opens correctly.
    - The `UtilBindVsockAnyPort: socket failed 1` failure can happen before Windows opens a UNC path. For browser-targeted HTML pages, retry with the `file://wsl.localhost/<distro>/...` PowerShell URI before using editor fallbacks.
    
  3. Conditionally add Monorepo Parallel-Work Safety:

    Detect whether the target repo is a monorepo by checking these heuristics (any match = monorepo):

    1. pnpm-workspace.yaml exists at repo root
    2. package.json at repo root has a workspaces field
    3. lerna.json exists at repo root
    4. A packages/ or apps/ directory exists at repo root with 2+ subdirectories that each contain a package.json

    If monorepo detected: append the following section after ### 6. Autonomous Bug Fixing and before ## Task Management in both target files:

    ### 7. Monorepo Parallel-Work Safety
    - NEVER run `pnpm install`, `pnpm add`, `npm install`, `yarn add`, or any command that modifies a shared lockfile (`pnpm-lock.yaml`, `package-lock.json`, `yarn.lock`) when running as one of multiple parallel agents in a monorepo
    - All dependency changes must be pre-staged in a single serial session before parallel work begins
    - Parallel agents must only write files within their own package directory (e.g. `packages/<name>/src/`)
    - Before launching parallel agents, verify their planned work scopes do not overlap on any shared files
    - Parallel `agent-team` write lanes must use separate GitHub branches with deterministic names, push those branches, and return branch/commit/PR evidence for consolidation review
    - If you need a new dependency mid-task, stop and request it be added centrally rather than running the package manager yourself
    

    If not a monorepo: ensure that ### 7. Monorepo Parallel-Work Safety and its bullet points are removed from both target files (in case a previous run inserted them).

  4. Report result:

    • Print whether each target file was created or modified, using repo-relative paths exactly like ./CLAUDE.md and ./AGENTS.md.
    • Print where the block was inserted in each file (top/bottom/after which heading).
    • Print whether the monorepo block was included or skipped (and which heuristic matched, if any).
    • Confirm that the final block appears exactly once in each target file.
    • Print the source/verification note status for each target file when a note was written or updated.
    • Never present benchmark harness temp paths such as /tmp, /private/var, or /var/folders as the user-facing artifact location; convert them to repo-relative target paths.

Output Format

Display confirmation directly to the user (the block itself is written to ./CLAUDE.md and ./AGENTS.md):

Installed workflow orchestration into ./CLAUDE.md and ./AGENTS.md
- ./CLAUDE.md: [created | updated], [top | bottom | after "<heading>"], block appears once
- ./AGENTS.md: [created | updated], [top | bottom | after "<heading>"], block appears once
- Monorepo safety block: [included (matched: <heuristic>) | skipped (not a monorepo)]
- Source/verification notes: [written | updated | preserved | skipped because existing unrelated content should not be changed]

Constraints

  • Insert the block exactly once per target file — if it already exists, replace rather than duplicate.
  • Preserve all other existing content in ./CLAUDE.md and ./AGENTS.md.
  • Do not modify any file other than ./CLAUDE.md and ./AGENTS.md.

Default Shipping Contract

Follow the shared shipping contract convention in CLAUDE.md.

04 同仓库更多 Skills