AGENT SKILL

long-running-task

Drive a long-running task through workspace progress files (.cool/task/progress.md + features.json) that survive context compaction and new runs

License 许可
MIT
最近更新
2026-11-03
Stars
★ 0
Forks
0

01 前言 Front Matter

name
long-running-task
metadata
{ "cool.tags": "long-running task progress checkpoint resume compaction durable", "cool.version": "2.0" }
description
Drive a long-running task through workspace progress files (.cool/task/progress.md + features.json) that survive context compaction and new runs
allowed-tools
read_file write_file edit_file list_files shell git

03 SKILL.md 正文

Long-Running Task

You are now operating in Long-Running Task mode. A task too big for one context window is tracked on disk so progress survives compaction, crashes and fresh runs.

Progress files

Maintain two files in the workspace:

  • .cool/task/progress.md — human-readable task journal with three sections:
    • Done — completed work, one bullet per verified step.
    • Next — the immediate next actions, ordered.
    • Acceptance criteria — how the finished task is verified.
  • .cool/task/features.json — machine-readable checklist: {"features": [{"id": "f1", "title": "...", "status": "pending|in_progress|done", "notes": "..."}]}.

Keep both files consistent with each other.

Process

  1. Start: if .cool/task/progress.md exists, read it and features.json, then resume from Next. Otherwise create both files: break the request into a feature checklist and record acceptance criteria.
  2. Work: do the next pending item only. Prefer edit_file for targeted changes so progress-file updates stay small.
  3. Update before answering: before every reply to the user, update progress.md (Done/Next) and features.json statuses. The file is the source of truth for the next run — a new session must be able to resume from it alone.
  4. Finish: when every feature is done and the acceptance criteria are met, say so explicitly and leave the final state in the files.

Principles

  • File over memory: never keep task state only in the conversation — compaction may drop it. Write progress down as it happens.
  • Small verifiable steps: each feature should be independently checkable (a test, a command, a file to inspect).
  • State the state: every reply briefly notes what is done and what is next, matching the files.

04 同仓库更多 Skills