context-task-planning

分类: 工具与效率 | 上传者: excitedhaha | 下载: 0 | 版本: v1.0（最新）

用于复杂的多步骤工作的任务范围上下文工程。当任务需要澄清、分阶段执行、基于文件的持久状态、上下文丢失后恢复或可选的子代理委派时使用。

更新日志: Source: GitHub https://github.com/excitedhaha/context-task-planning

目录结构

当前层级: tree/main/skill/

📁 claude-hooks/
- 📁 scripts/
  - 📄 __init__.py 0 B
  - 📄 hook_common.py 16.9 KB
  - 📄 pre_tool_use.py 3.1 KB
  - 📄 session_start.py 1.3 KB
  - 📄 statusline.py 8.2 KB
  - 📄 user_prompt_submit.py 2.1 KB
- 📄 README.md 1.4 KB
- 📄 settings.example.json 1.3 KB
📁 opencode-plugin/
- 📄 task-focus-guard.binding-title.smoke.mjs 2.9 KB
- 📄 task-focus-guard.js 24.0 KB
- 📄 task-focus-guard.smoke.mjs 5.5 KB
📁 schemas/
- 📄 context-compact.schema.json 1.5 KB
- 📄 delegate-status.schema.json 1.6 KB
- 📄 state.schema.json 5.1 KB
📁 scripts/
- 📄 archive-task.sh 4.7 KB
- 📄 block-delegate.sh 5.1 KB
- 📄 cancel-delegate.sh 4.9 KB
- 📄 check-complete.sh 2.8 KB
- 📄 check-switch-safety.sh 335 B
- 📄 check-task-drift.sh 320 B
- 📄 compact-context.sh 338 B
- 📄 compact_context.py 38.8 KB
- 📄 complete-delegate.sh 5.9 KB
- 📄 create-delegate.sh 8.8 KB
- 📄 current-task.sh 329 B
- 📄 done-task.sh 5.1 KB
- 📄 ensure-switch-safety.sh 338 B
- 📄 init-task.sh 11.6 KB
- 📄 install-macos.sh 1.6 KB
- 📄 install-opencode-plugin.sh 1.2 KB
- 📄 list-delegates.sh 3.9 KB
- 📄 list-repos.sh 317 B
- 📄 list-tasks.sh 5.2 KB
- 📄 list-worktrees.sh 325 B
- 📄 pause-task.sh 3.4 KB
- 📄 prepare-delegate.sh 6.4 KB
- 📄 prepare-task-worktree.sh 4.3 KB
- 📄 promote-delegate.sh 6.1 KB
- 📄 register-repo.sh 324 B
- 📄 resolve-plan-dir.sh 2.0 KB
- 📄 resolve-workspace-root.sh 795 B
- 📄 resume-delegate.sh 2.4 KB
- 📄 resume-task.sh 5.7 KB
- 📄 set-active-task.sh 3.4 KB
- 📄 set-task-repos.sh 1.1 KB
- 📄 set-task-spec-context.sh 1.3 KB
- 📄 slugify.sh 750 B
- 📄 smoke-test-cli-guidance.sh 23.0 KB
- 📄 smoke-test-task-first-worktrees.sh 4.0 KB
- 📄 start-delegate.sh 5.2 KB
- 📄 subagent-preflight.sh 333 B
- 📄 task_guard.py 126.9 KB
- 📄 validate-task.sh 25.9 KB
📁 templates/
- 📄 delegate_brief.md 458 B
- 📄 delegate_result.md 297 B
- 📄 findings.md 467 B
- 📄 progress.md 1.1 KB
- 📄 task_plan.md 2.7 KB
📄 examples.md 7.9 KB
📄 reference.md 9.9 KB
📄 SKILL.md 10.8 KB

SKILL.md

---
name: context-task-planning
description: Task-scoped context engineering for complex multi-step work. Use when a task needs clarification, phased execution, durable file-based state, recovery after context loss, or optional sub-agent delegation.
license: MIT
metadata:
  version: "0.2.0"
allowed-tools: Read Write Edit Bash Glob Grep WebFetch Task
---

# Context Task Planning

Use persistent planning files as a task-scoped context system.

This skill is for work that is too large or too stateful to trust to the live context window alone.

## When to use

Use this skill when the work involves one or more of the following:

- more than a few steps
- requirements that still need clarification
- long-running implementation or research
- recovery after context loss or agent switching
- optional delegate or sub-agent work on isolated subproblems

## Core operating model

Each task lives in its own directory:

```text
.planning/<slug>/
  task_plan.md
  findings.md
  progress.md
  state.json
  delegates/
```

### File roles

- `state.json` - machine-readable source of truth for current task status
- `task_plan.md` - human-readable plan, with a compact `Hot Context` section at the top
- `findings.md` - durable discoveries, external information, and material to distill later
- `progress.md` - session log, checkpoints, and verification results

## First steps

1. Run `scripts/init-task.sh "<task title>"`
2. Read `.planning/<slug>/state.json`
3. Read `.planning/<slug>/task_plan.md`
4. Fill in missing goal, non-goals, acceptance criteria, constraints, and open questions before implementation

If multiple sessions are active, give the current session a stable session key and bind one task to it:

```bash
export PLAN_SESSION_KEY=manual:<name>
sh scripts/set-active-task.sh <slug>
```

Use `PLAN_TASK` only for a one-off manual override inside the current shell.

If another session is already writing the same task, use `set-active-task.sh --observe <slug>` instead of creating a second writer. Observers may still work inside delegate lanes.

If the workspace is a parent directory that contains multiple git repos, register repos explicitly before you bind them to tasks:

```bash
sh scripts/list-repos.sh --discover
sh scripts/register-repo.sh --id frontend frontend
sh scripts/register-repo.sh --id backend backend
sh scripts/set-task-repos.sh <slug> --repo frontend --repo backend --primary frontend
```

Use auto-discovery only to review candidates. The actual registration step should stay explicit.

Once the parent workspace owns `.planning/`, commands run from that parent
directory, any registered repo path, or a recorded
`.worktrees/<task-slug>/<repo-id>/` checkout should all resolve back to the
same shared workspace. Ignore unrelated ancestor `.planning/` directories
unless the current path actually belongs to that older workspace.

## Workflow

### 1. Clarify first

Do not jump into implementation until the task has:

- a clear goal
- explicit non-goals
- acceptance criteria
- constraints
- open questions or assumptions listed

### 2. Keep the hot context small

Repeated reads should focus on the smallest useful snapshot:

- task status
- goal
- current mode
- current phase
- next action
- blockers
- verification target

Do not repeatedly reload large notes when a compact snapshot will do.

### 3. Treat `state.json` as operational truth

Whenever the task status changes, update `state.json` first and keep markdown aligned with it.

### 4. Distill findings intentionally

External or untrusted content belongs in `findings.md`, not in `Hot Context`.

Promote only the distilled conclusions into the main plan.

### 5. Verify before marking done

Every task should declare its verification targets and record actual results in `progress.md`.

### 6. Archive, do not overwrite

When a task is complete, keep the task directory as history. Switch or remove the session binding or workspace fallback pointer instead of reusing the directory for unrelated work.

### 7. Guard against task drift on hosts without runtime adapters

When the host does not provide native task UI or prompt hooks, use the shared scripts as the fallback policy:

- use `scripts/current-task.sh --compact` when you need to verify which task is currently active
- use `scripts/check-task-drift.sh --prompt "..." --json` before mixing a new complex request into the active task when the match is uncertain
- if the result is `likely-unrelated` or `unclear`, ask whether to continue the current task, switch tasks, or create a new task before changing planning state

This is especially important for Codex-style environments where the file protocol exists but host-native reminder surfaces may not.

When you need more than the prompt-friendly cue, `scripts/current-task.sh` now prints a human-readable summary with the recommended next step, suggested commands, and repo/worktree context. Keep `--compact` for prompts, tmux, or status bars.

## Delegate protocol

Delegates are optional and should be used only for isolated work such as:

- discovery
- spikes
- verification triage
- review
- catchup summarization

Delegate folders live under:

```text
delegates/<delegate-id>/
  brief.md
  result.md
  status.json
```

### Delegate workflow

1. For faster setup, use `scripts/prepare-delegate.sh` to infer and create a lane in one step, or use `scripts/create-delegate.sh` for full manual control
2. Fill `brief.md` with the bounded question, constraints, and expected deliverable
3. Mark the lane `running` with `scripts/start-delegate.sh` when work begins
4. Let the subagent or isolated worker operate only inside that delegate directory
5. If the lane stalls, use `scripts/block-delegate.sh`; if it becomes irrelevant, use `scripts/cancel-delegate.sh`
6. Mark the lane complete with `scripts/complete-delegate.sh`
7. Merge useful results back with `scripts/promote-delegate.sh`

Blocked lanes can be resumed explicitly with `scripts/resume-delegate.sh`.

### Delegate triggers

Create a delegate lane when the main task produces a bounded subproblem such as:

- repository exploration or entry-point mapping
- option comparison or feasibility spike
- test failure triage or validation pass
- diff review or focused code review

If you are about to launch a subagent for one of those questions, create the delegate lane first.

### Single writer rule

Only the coordinator updates:

- `task_plan.md`
- `progress.md`
- `state.json`

Delegates should write only inside their own scratch folders.

In session terms: one task may have one writer session and additional observer sessions. Observers may create or update delegate lanes, but they must not edit `task_plan.md`, `progress.md`, `state.json`, or `findings.md`.

## Recovery order

When resuming a task, recover in this order:

1. `state.json`
2. `task_plan.md` Hot Context
3. latest relevant entries in `progress.md`
4. unresolved delegates

This keeps recovery portable across Claude Code, Codex, and OpenCode.

## Scripts

- `scripts/init-task.sh` - create or resume `.planning/<slug>/`, warning before switching away from dirty git work
- `scripts/resolve-plan-dir.sh` - resolve current task from `PLAN_TASK`, session binding, `.active_task`, or latest plan
- `scripts/current-task.sh` - show the resolved task; default output is an action-oriented summary, `--compact` stays short for prompts and status bars, and `--json` adds recommendation metadata for host adapters
- `scripts/check-task-drift.sh` - classify whether a new prompt still fits the active task
- `scripts/check-switch-safety.sh` - inspect whether a git worktree should be stashed or committed before switching tasks
- `scripts/list-repos.sh --discover` - show registered repos plus direct-child git repos that could be registered under the current workspace
- `scripts/register-repo.sh --id <repo-id> <path>` - explicitly register a repo under a parent workspace
- `scripts/set-task-repos.sh <slug> --repo <repo-id> [--repo ...] [--primary <repo-id>]` - declare which registered repos a task is allowed to touch and print which repos are safe to keep shared versus which now need a worktree
- `scripts/prepare-task-worktree.sh --task <slug> --repo <repo-id>` - create and bind a dedicated checkout for an overlapping writer task; the standard location is `.worktrees/<task-slug>/<repo-id>/`
- `scripts/list-worktrees.sh` - inspect task-scoped worktree bindings grouped by task
- `scripts/install-opencode-plugin.sh` - symlink the bundled OpenCode plugin into the standard plugin directory
- `scripts/set-active-task.sh <slug>` - update the current session binding when available, otherwise the shared fallback pointer; use `--observe` for read-only sessions or `--steal` to take over the writer lease
- `scripts/validate-task.sh` - check task state consistency across `state.json`, markdown files, delegates, and derived compact artifacts; add `--fix-warnings` to resync warning-level snapshot drift from `state.json` and refresh compact artifacts when needed
- `scripts/prepare-delegate.sh` - infer and create a delegate lane, optionally auto-starting it
- `scripts/create-delegate.sh` - create a delegate lane under the current task
- `scripts/list-delegates.sh` - show delegate lanes for the current task
- `scripts/start-delegate.sh <delegate-id>` - mark a delegate lane running when active work begins
- `scripts/resume-delegate.sh <delegate-id>` - resume a blocked delegate lane back to running
- `scripts/block-delegate.sh <delegate-id>` - mark a delegate lane blocked while keeping it open
- `scripts/cancel-delegate.sh <delegate-id>` - cancel a delegate lane and remove it from active delegates
- `scripts/complete-delegate.sh <delegate-id>` - mark a delegate lane complete and remove it from active delegates
- `scripts/promote-delegate.sh <delegate-id>` - append delegate results into `findings.md`
- `scripts/list-tasks.sh` - show all task workspaces and their states
- `scripts/pause-task.sh [slug]` - pause a task without losing its current phase or next action
- `scripts/resume-task.sh [slug]` - reactivate a paused task, bind it as the writer for the current session or workspace fallback, and guard dirty git worktrees before switching
- `scripts/ensure-switch-safety.sh` - enforce the dirty-worktree guard, prompting to stash, commit manually, or continue dirty
- `scripts/done-task.sh [slug]` - mark a task done after verification and clear the shared pointer if needed
- `scripts/archive-task.sh [slug]` - archive a task and clear the shared pointer if needed
- `scripts/check-complete.sh` - summarize current task completion state
- `scripts/install-macos.sh` - install symlinks for Claude Code, Codex, and OpenCode

## Safety rules

- Keep untrusted external content in `findings.md`
- Do not let delegates mutate the main planning files
- Do not mark tasks done without recording real verification results
- Before switching tasks in a git repo, handle dirty worktrees explicitly instead of silently carrying changes into the next task
- Do not reuse an old slug for unrelated work

登录后下载/点赞/收藏 ❤ 8 | ★ 0

context-task-planning

目录结构

SKILL.md

举报内容

提示