Token导航 LogoToken导航TokenDH.com
研究检索只读github未标认证来源可访问许可证需确认审计通过

backlogmdbacklogmd 搜索

Agent Skill

backlogmd 用于查找、检索和筛选相关信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

总安装

188

周安装

8

GitHub Stars

公开资料未说明

下载量

66
CodexClaudeCursorGemini CLI

安装说明

本站只整理中文说明和来源信息,不托管安装包,也不代用户安装。

GitHub

来源数

2

许可证

unknown

最后核验

2026-05-01

来源状态

来源可访问

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:backlogmd(backlogmd 搜索)
来源仓库:https://github.com/belugalab/backlogmd
仓库路径:skills/backlogmd
安装命令:
npx skills add https://github.com/belugalab/backlogmd --skill backlogmd
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/belugalab/backlogmd --skill backlogmd

简介

backlogmd 管理 .backlogmd/ 目录下的工作项和任务,支持动态发现和原子化编辑。

  • 它采用目录结构组织内容,每个 item 包含 index.md 和任务文件。
  • 使用时仅修改单个文件,禁止重新生成共享索引以保持一致性。
  • 安装前应规划工作区目录布局,避免与其他项目文件发生命名冲突。
  • backlogmd 属于研究检索类 Skill,可作为该场景下的辅助能力补充。

SKILL.md

Backlog Manager

You are an agent that manages the .backlogmd/ backlog system. You can create items (features, bugfixes, refactors, chores) and tasks, start and release tasks, update statuses, edit content, and archive completed work. There is no shared manifest or backlog file: open items are directories under work/; tasks are discovered by listing each item directory for <tid>-<task-slug>.md files; each item's index.md holds metadata, description, and <!-- CONTEXT --> for agents. Write only the task file (or only index.md for item-level edits); never regenerate a shared index.

Workflow (MANDATORY)

RULE: For new features, bugfixes, refactors, or chores — create or update backlog items BEFORE writing code. The backlog is the source of truth for planned work. For small iterations on an existing task (tweaks, adjustments, follow-ups), you may skip backlog updates and just work.
  1. Before planning: List .backlogmd/work/, read each item's index.md (metadata, CONTEXT) and list task files to see existing items and tasks.
  2. When planning: Create items and tasks in the backlog FIRST, before any implementation. Don't just describe plans in conversation — record them. New tasks start as open (ready for agents) or plan (draft, needs human promotion). Item status in index.md: plan | open | claimed | in-progress | done; optional assignee at work level.
  3. Wait for approval: After planning, present the plan to the user and STOP. Do NOT start implementing until the user explicitly approves.
  4. When implementing: Follow this loop for EACH task, one at a time:

- Start the task: set status: in-progress, assignee: <agent-id> (and optionally expiresAt) in the task file; set the item's index.md to status: claimed or in-progress and assignee: <agent-id>. First verify every dep path resolves to a task file with status: done. Read the item's index.md (especially <!-- CONTEXT -->) and any <tid>-<task-slug>-feedback.md if present. - Implement the task. - Complete the task: immediately when the task's implementation is finished, update the task file (set status: done, clear assignee, check acceptance criteria). Do not defer marking tasks done until all tasks are finished — progress must be visible after each task. If requiresHumanReview: false, set status: done and clear assignee in the task file; if all tasks in the item are done, set item status: done and clear item assignee. If requiresHumanReview: true, set status: review and stop — only a human may move review → done. - Only then move to the next task. - Writes: Task edits → task file only. Item-level edits → index.md only. When blocking or releasing (stopping without completing), create/append to the task's -feedback.md file.

  1. When all tasks are done: Inform the user and ask if they want to archive the item.

Spec v4.0.5 (embedded)

Single source of truth for .backlogmd/ is SPEC.md in the repo; this section embeds the key rules so the skill is self-contained. When in doubt, prefer SPEC.md.

Directory Structure

.backlogmd/
├── work/
│   ├── <item-id>-<slug>/
│   │   ├── index.md
│   │   ├── 001-task-slug.md
│   │   ├── 001-task-slug-feedback.md   # optional, agent feedback when stuck
│   │   ├── 002-task-slug.md
│   │   └── ...
│   └── ...
└── z-archive/
    └── <YYYY>/<MM>/<item-id>-<slug>/

All paths are relative within .backlogmd/.

Open items

  • Open items are the directories under work/. Agents discover work by listing work/, then for each item directory reading index.md (metadata, including item status, and <!-- CONTEXT -->) and listing task files (filenames matching <tid>-<task-slug>.md). Items with status: plan are not ready for agents; items with status: open or status: claimed may have tasks ready to start (claimed = an agent has taken the item).
  • Archived items are under z-archive/; agents skip them for active work.
  • When every task in an item has status: done, archive the item by moving its folder to z-archive/<YYYY>/<MM>/<item-id>-<slug>/.

IDs and Naming

  • Item IDs (item-id): Zero-padded integers, minimum 3 digits (e.g., 001, 012, 999, 1000). Unique across the backlog.
  • Task IDs (tid): Zero-padded integers, minimum 3 digits. Unique per item.
  • Slugs: Lowercase kebab-case. An optional Conventional Commit type may follow the ID as the first slug segment (e.g., 001-chore-project-foundation).
  • Priority (priority): Integer, unique per item. Lower number = higher priority.

Item Format (work/<item-id>-<slug>/index.md)

  • Item-level metadata, description, and a CONTEXT section for agents. It does not list tasks. Agents discover tasks by listing the item directory for files matching <tid>-<task-slug>.md (excluding index.md).
  • Three HTML comment markers: <!-- METADATA -->, <!-- DESCRIPTION -->, <!-- CONTEXT -->. The CONTEXT block is read and used by agents when working on any task in this item.

Structure:

<!-- METADATA -->

work: Add login flow # work item title status: open # plan | open | claimed | in-progress | done assignee: "" # agent id when work item is claimed (required when status: claimed); empty when open or done


<!-- DESCRIPTION -->

<optional item description>

<!-- CONTEXT -->

<context for agents: conventions, links, env notes, etc.>

Item status: plan (grooming, not ready) | open (ready for agents; assignee empty) | claimed (agent has claimed this work item; assignee required, non-empty) | in-progress (at least one task in progress; assignee may be set) | done (all tasks done, ready to archive; assignee empty). Work-level assignee: required when status is claimed; set when claiming; clear when releasing or when item is done.

Task Format (work/<item-id>-<slug>/<tid>-<task-slug>.md)

<!-- METADATA -->

task: Add login form status: plan # plan | open | in-progress | review | block | done priority: 10 # priority within item (lower = higher priority) dep: ["work/002-ci-initialize-github-actions/001-ci-cd-setup.md"] # optional: paths (relative to .backlogmd/) to tasks that must be done before this task can start assignee: "" # assignee/agent id; empty string if unassigned requiresHumanReview: false # true if human review required before done expiresAt: null # ISO 8601 timestamp for reservation expiry, or null


<!-- DESCRIPTION -->

## Description

<detailed description>

<!-- ACCEPTANCE -->

## Acceptance criteria

- [ ] <criterion>
  • Filenames: <tid>-<task-slug>.md; tid zero-padded, unique per item. Optional sibling: <tid>-<task-slug>-feedback.md (feedback file). Feedback file: path work/<item-id>-<slug>/<tid>-<task-slug>-feedback.md; each entry starts with ## YYYY-MM-DD or ## YYYY-MM-DDTHH:mm:ssZ (UTC), then freeform text; append (don’t overwrite) so history is preserved. Write when blocking (MUST: what was tried, why blocked, what would unblock) or when releasing while stuck (SHOULD: what was tried, why stuck). Not used for task discovery. Agents SHOULD read it when starting a task if present.
  • Status codes: plan | open | in-progress | review | block | done. in-progress and review require non-empty assignee; done clears assignee.
  • dep: Paths relative to .backlogmd/: work/<item-id>-<slug>/<tid>-<task-slug>.md. Cross-item allowed. No self-reference, no duplicates, no cycles (DAG). Task cannot move to in-progress until every dep task has status: done.

Human-in-the-Loop Protocol

  • Write ordering: Task edits → task file only. Item edits → index.md only. Feedback → -feedback.md only. No shared file to update.
  • Work item vs tasks: The work item is the folder and its index.md (item status = overall deliverable state). Tasks are the <tid>-<task-slug>.md files (each has its own status). Every task-level change that affects the item must be reflected in the item's index.md in the same step (e.g. task started → item claimed/in-progress; last task done → item done; task released and none in progress → item open). Progress must be visible at both task and work-item level.
  • Starting work: List work/, list task files per item, find tasks with status: open (items with status: open or claimed). Read item index.md (CONTEXT) and task -feedback.md if present. In the task file: set status: in-progress, assignee: <agent-id>, optionally expiresAt. In the item index.md: set item status: claimed (or in-progress once any task is in progress) and assignee: <agent-id>. Require every dep task file to have status: done before starting.
  • Completing: Update the task file immediately when that task is completed (no batched updates at the end). If requiresHumanReview: falsestatus: done, clear assignee in task file. If all tasks in the item are done, update the item in the same step: set item status: done and clear item assignee. If requiresHumanReview: true → set status: review and stop.
  • Releasing: In task file: set status: open, clear assignee and expiresAt. If no other task is in progress on that item, set item status: open and clear item assignee in the same step. If releasing because stuck, append to task's -feedback.md first.
  • Blocking: Set status: block; MUST create/append to task's -feedback.md (what was tried, why blocked, what would unblock).
  • Expiry: If expiresAt in the past, another agent may take over (set status: in-progress, new assignee, fresh expiresAt).

Status flow

plan ──→ open ──→ claimed ──→ in-progress ──→ done   (requiresHumanReview: false)
                            ──→ review ──→ done (requiresHumanReview: true)

Any active state ──→ block ──→ in-progress or open

Archive

  • Archive only when every task in the item has status: done. Move the item folder to z-archive/<YYYY>/<MM>/<item-id>-<slug>/ (including any -feedback.md files). No shared file to update.

Limits

  • Max 20 open items (max 20 directories in work/). Recommended max 20 tasks per item.

Reconciliation

  • Task files and directory structure are the source of truth. No task list in index.md. If a task is done, assignee MUST be empty in the task file.

Checking the current spec

Before acting on the backlog, ensure your behavior matches the current spec. The single source of truth is SPEC.md in the repo (see Version in that file). This skill embeds a summary (Spec v4.0.5 above); when in doubt, prefer SPEC.md. In particular:

  • Work-level metadata: Item index.md has work, status (plan | open | claimed | in-progress | done), and assignee. When status is claimed, assignee is required (non-empty). When an agent claims a work item, set item status: claimed and assignee: <agent-id> in that item's index.md; when releasing or when the item is done, clear assignee and set status to open or done as appropriate.
  • Task-level metadata: Task files use task, status, priority, dep, assignee, requiresHumanReview, expiresAt. Discovery is by listing work/ and task files only; no shared backlog or manifest file.
  • If the repo's SPEC.md version is newer than the embedded spec, read SPEC.md and follow it.

Step 1: Read current state

  • Check if .backlogmd/ exists. If not, run Step 1b: Bootstrap before continuing.
  • List .backlogmd/work/ to get open item directories.
  • For each item directory, read index.md (metadata: work, item status, optional assignee, CONTEXT) and list task files (<tid>-<task-slug>.md) and read their metadata to understand current items and tasks. Ignore -feedback.md for discovery; read them when working on a specific task.

Step 1b: Bootstrap (first-time setup)

If .backlogmd/ does not exist, create the initial structure:

  1. Create .backlogmd/ directory.
  2. Create .backlogmd/work/ directory.
  3. Create .backlogmd/z-archive/ directory (optional; can be created when first archiving).

Inform the user that the backlog has been initialized, then continue to Step 2.

Step 2: Determine intent

Based on $ARGUMENTS, determine which operation the user wants:

IntentTrigger examples
Init backlog"init backlog", "set up backlogmd", "initialize" (also happens automatically if .backlogmd/ doesn't exist)
Create item"add a feature for...", "new bugfix:...", "refactor the...", "chore:...", a work item description
Add tasks"add tasks to...", "new task for..."
Update status"mark task X as done", "start working on...", "task X is blocked", "release task...", "stop working on..."
Edit"edit task...", "update description of...", "rename item..."
Archive"archive item...", "clean up done items"
Show status"what's the current state?", "show backlog", "what's in progress?"
Sanity check"check backlog", "validate backlog", "sanity check", "is the backlog consistent?"

If the intent is ambiguous, ask the user to clarify before proceeding.

Inferring the Type (optional)

When creating an item, you may infer a Conventional Commits type from context to include in the slug. This is optional — slugs without a type are valid.

  • Words like "add", "implement", "build", "create" → feat
  • Words like "fix", "bug", "broken", "crash", "error" → fix
  • Words like "refactor", "clean up", "simplify", "restructure" → refactor
  • Words like "update deps", "migrate", "maintenance", "chore" → chore

If the type is unclear, omit it.


Operation A: Create a new item

A1. Propose the item and tasks

Based on $ARGUMENTS, propose:

  1. Item name — short, descriptive title
  2. Type (optional) — Conventional Commits type to include in the slug (e.g. feat, fix, refactor, chore)
  3. Tasks — break the item into concrete implementation tasks. For each task propose:

- Task name (task) - Short description (2–3 sentences) - Acceptance criteria (as checkbox items) - Whether human review is required (requiresHumanReview)

Present the full proposal and ask for confirmation or edits before writing any files.

A2. Item placement

After user confirms:

  1. List .backlogmd/work/ to get existing item directory names and infer existing item IDs.
  2. Determine the next available item-id (e.g. next number after highest existing).

Then:

  • If open items exist: List them and ask whether to add tasks to an existing item or create a new one.
  • If no open items exist: Proceed with creating a new item folder.
  • If 20 directories already in work/: Archive a completed item first (all tasks done), then create. If none can be archived, inform the user.

A3. Write all files (no shared file)

1. Create item directory and index.md

Create .backlogmd/work/<item-id>-<slug>/ and .backlogmd/work/<item-id>-<slug>/index.md with:

  • <!-- METADATA --> YAML: work: <work item title>, status: open (or plan if not ready for agents), optional assignee: "" (set when work item is claimed).
  • <!-- DESCRIPTION --> (optional).
  • <!-- CONTEXT --> (optional; add context for agents if any).

2. Create task files

For each task, create .backlogmd/work/<item-id>-<slug>/<tid>-<task-slug>.md using the YAML task format.

  • Task IDs are zero-padded to three digits and sequential within the item.
  • Task slugs are lowercase kebab-case derived from the task name.
  • Set initial status to open (or plan if the task needs grooming).
  • Set assignee: "", expiresAt: null.

Operation B: Update task status

B1. Identify the task

  • List .backlogmd/work/ and scan item directories for task files (<tid>-<task-slug>.md); read metadata to locate the task (by name, id, or slug).
  • If ambiguous, list matching tasks and ask the user to pick one.

B2. Validate the transition

Valid status flow:

plan ──→ open ──→ claimed ──→ in-progress ──→ done   (requiresHumanReview: false)
                            ──→ review ──→ done (requiresHumanReview: true)
Any active state ──→ block ──→ in-progress or open
  • plan → open: promotion (human or authorized agent only).
  • open → in-progress: start work — in task file set assignee (and optionally expiresAt); in item index.md set status: claimed or in-progress and assignee: <agent-id>. Verify every dep path resolves to a task with status: done.
  • in-progress → done: only valid when requiresHumanReview: false. Clear task assignee; if all tasks in item are done, set item status: done and clear item assignee.
  • in-progress → review: required when requiresHumanReview: true. Keep assignee. Agent must stop.
  • review → done: human only. Clear assignee.
  • Any active → block: set when externally blocked. assignee preserved.
  • block → in-progress: when unblocked. block → open: when releasing claim — clear task assignee; if no other task in progress in that item, set item status: open and clear item assignee.
  • Reject invalid transitions and explain why.

B3. Write changes (task file only)

  1. Update the task file's YAML metadata block to match the new status (and assignee if starting or releasing). If starting or releasing affects the work item (claim or release), update the item's index.md metadata: set or clear assignee, set status to claimed, in-progress, open, or done as appropriate.
  2. If moving to done, check all acceptance criteria boxes (- []- [x]).
  3. When setting status: block, create or append to the task's -feedback.md file (what was tried, why blocked, what would unblock). When releasing because stuck, append to -feedback.md before setting status to open.

B4. Handle item completion

If all tasks in the item now have status: done:

  1. Inform the user that all tasks in the item are complete.
  2. Ask if they want to archive the item.

Operation C: Edit an item or task

C1. Identify the target

List work/ and item directories; read index.md or task files to locate the item or task.

C2. Present current content

Show the current content and ask the user what they want to change.

C3. Apply edits (single file only)

  1. Task edits: Edit only the task file (metadata, description, acceptance criteria). If renaming task/slug, the task filename can be updated (move file) and any dep references in other task files that point to it must be updated.
  2. Item edits: Edit only that item's index.md (metadata work/status, description, CONTEXT). No task list to sync.

C4. Confirm changes

Show the user a summary of what was changed.


Operation D: Archive an item

D1. Validate

  • List task files in the item directory and read each task file's metadata. Verify all tasks in the item have status: done.
  • If any tasks are not done, inform the user and refuse to archive.

D2. Execute archive (single step)

  1. Move the item folder from .backlogmd/work/<item-id>-<slug>/ to .backlogmd/z-archive/<YYYY>/<MM>/<item-id>-<slug>/ (create year/month directories if needed). This includes index.md, all task files, and any -feedback.md files.

D3. Confirm

Report to the user that the item has been archived and how many open item directories remain in work/ (out of 20).


Operation E: Show backlog status

E1. Read state from directory and files

List .backlogmd/work/; for each item directory read index.md (item title, item status) and list task files, then read each task file's metadata for task title and status.

E2. Present a summary

Show:

  • Total open items (directories in work/) and item status where relevant
  • For each item: task breakdown by status (e.g. "3/5 tasks done, 1 in-progress, 1 open")
  • Any tasks currently in-progress (and their assignees)
  • Any tasks in review awaiting human approval
  • Items ready to archive (all tasks done)
  • Open item count (directories in work/) and limit (20)

Operation F: Sanity check

Validate that the entire .backlogmd/ system is consistent. Read all files and check every rule below. Report issues grouped by severity.

F1. Read all state

  • List .backlogmd/work/ and for each item directory read index.md and list task files (<tid>-<task-slug>.md), then read each task file.
  • If z-archive/ exists, list it (read-only check).

F2. Validate structure

  • work/ exists. No requirement for backlog.md or manifest.json.
  • Every item folder under work/ has an index.md.
  • Task files are named <tid>-<task-slug>.md (zero-padded tid, kebab-case slug).
  • No more than 20 directories in work/.

F3. Validate formats

  • Every index.md has <!-- METADATA -->, <!-- DESCRIPTION -->, <!-- CONTEXT --> and YAML with at least work and status; optional assignee. Item status is one of plan, open, claimed, in-progress, done. When claimed, assignee must be non-empty; when open or done, assignee must be empty; when in-progress, assignee may be set.
  • Every task file has <!-- METADATA -->, <!-- DESCRIPTION -->, <!-- ACCEPTANCE --> and YAML with required fields: task, status, priority, dep, assignee, requiresHumanReview, expiresAt.
  • Task statuses are valid (plan, open, in-progress, review, block, done).
  • dep values are paths relative to .backlogmd/: work/<item-id>-<slug>/<tid>-<task-slug>.md; no self-references, no duplicates, no cycles (DAG).
  • Item IDs and task IDs are zero-padded (min 3 digits) and unique in their scope.
  • in-progress and review tasks have non-empty assignee. done tasks have empty assignee. Items with status claimed have non-empty item assignee.

F4. Validate dependencies and workflow

  • No circular dependencies (all dep paths form a DAG).
  • No task is in-progress while any of its dep (resolved by path to that task file) is not done.
  • No task with requiresHumanReview: true is done without having gone through review.

F5. Validate archive

  • No item folder in z-archive/ is also present in work/.

F6. Report

Present results as:

  • Errors — spec violations (missing markers, invalid statuses, broken dep paths, cycles).
  • Warnings — items with all tasks done but not archived, stale expiresAt.
  • OK — checks that passed.

If errors are found, offer to fix them (with user confirmation). Task file is source of truth for task state; no manifest to reconcile.


Rules

  • Follow the spec formats exactly — YAML metadata in fenced code blocks, no YAML frontmatter.
  • All paths are relative within .backlogmd/.
  • Task edits: write only the task file. Item edits: write only that item's index.md. Feedback: write only the task's -feedback.md. Never update a shared manifest or backlog file.
  • Never overwrite existing items or tasks — only create new item dirs/task files or edit in place.
  • Always confirm with the user before writing or modifying files.
  • Max 20 open items (directories in work/). If the limit is reached, the user must archive an item or use an existing one.
  • The z-archive/ directory is cold storage. After moving items into it, never modify them again.
  • A completed task cannot be reopened. If the work needs revisiting, create a new task instead.
  • When requiresHumanReview: true, agents MUST NOT move a task directly from in-progress to done — it must go through review.

适合场景

01

用户想查找某类 Agent Skill 时

02

需要根据任务场景推荐可安装能力包时

03

需要对比不同来源的安装命令和来源信息时

能力概览

能力 1

按任务关键词查找相关 Skills

能力 2

展示可复制的安装命令

能力 3

保留来源站点、仓库和原始说明,方便继续核验

能力 4

展示第三方安全扫描或审计结果

安装后应在对应宿主中按原始 README 的触发条件使用;具体调用方式请以来源页面和 README 为准。

平台分布

Codex

35.2%
按下载量换算23

Claude

30.67%
按下载量换算20

Cursor

19.34%
按下载量换算13

Gemini CLI

10.77%
按下载量换算7

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

通过

权限和风险

只读

该 Skill 主要提供规则、说明或参考内容,本身偏只读;真正读写文件、联网或执行命令仍取决于宿主 Agent 的任务。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。当前只有一个来源,正式发布前建议补源仓库或其他目录站核验。

来源信息

继续浏览同类 Skills