AGENT SKILL

taskagent

Load before ANY taskagent operation, or when you read a file referencing taskagent (org/projects, plan files, docs). taskagent is the human-facing backlog and durable cross-session memory — not in-session working memory. For steps you'll finish this session, use TodoWrite + the plan file.

License 许可
MIT
最近更新
2026-09-22
Stars
★ 0
Forks
0

01 前言 Front Matter

name
taskagent
description
Load before ANY taskagent operation, or when you read a file referencing taskagent (org/projects, plan files, docs). taskagent is the human-facing backlog and durable cross-session memory — not in-session working memory. For steps you'll finish this session, use TodoWrite + the plan file.

03 SKILL.md 正文

Taskagent - Shared Work Tracker

What taskagent is for

taskagent serves two purposes:

  1. The human's view — backlog of deliverables, what is active, what is done.
  2. Durable cross-session memory — LEARNING / MISALIGNMENT notes and handoff state that must outlive the conversation.

It is not your in-session working memory. Sessions can carry a large amount of work, so tracking your own place within a task belongs in TodoWrite, and the design/spec of a task belongs in the plan file.

Where things live

Tool Holds Lifetime
TodoWrite This session's tactical steps The session
plan file Design, spec, wiring, current state of a ticket The ticket / worktree
agentmemory Auto-captured, searchable observations Long-term, searchable
taskagent Human-facing backlog + status + durable LEARNING/MISALIGNMENT Long-term, human-curated

When to create a taskagent task

Create a task when it is:

  • A deliverable at PR / ticket / epic granularity the human would track, or
  • Work that will outlive this session and needs pickup later, or
  • Offshoot work surfaced mid-session that is not part of the current deliverable but should be remembered — a bug noticed in passing, a refactor opportunity, a follow-up, a side quest.

That last case is one of taskagent's most valuable uses: it moves the thought out of the volatile session and into the human's backlog without derailing current work. Use discovered_from:<parent_uuid> and/or +side_quest. Do not park such work in TodoWrite — TodoWrite dies with the session.

Otherwise, default to not creating a task. In-session steps go to TodoWrite.

Anti-pattern: pre-decomposing one deliverable

Do not shatter a single deliverable into a tree of taskagent subtasks with implementation specs in the annotations. That structure is not read across sessions, adds no durable memory, and ends up deleted. Keep one task per deliverable; put the steps in TodoWrite and the spec/wiring in the plan file. Reserve annotations for durable LEARNING / MISALIGNMENT and handoff state.

If you are about to create more than a couple of tasks for work that completes in this session, stop — that is TodoWrite plus a plan file.

CRITICAL: Command Name

Always use taskagent, NEVER use task directly.

taskagent is a hardened wrapper that shares the human's taskwarrior database (~/.task) but enforces that the agent may only modify tasks tagged +agent. Reads of any task are allowed. Using task directly is denied by opencode permissions. taskagent is globally available — works from anywhere, no cd needed.

CRITICAL: +agent tag and write access

Every task the agent creates is automatically tagged +agent by an on-add hook. The agent may only modify/done/delete/annotate +agent tasks; modifying a human task is rejected at the taskwarrior layer. The +agent tag is sticky. Default views (list, ready, next) are filtered to +agent. To read a human task, use its UUID directly:

taskagent <uuid> info          # reads any task, including human tasks
taskagent <uuid> agentinfo     # same, minimal format

Forbidden (blocked by the wrapper): undo, config, execute, any rc: or rc.* overrides.

CRITICAL: Use Partial UUIDs, Not Short IDs

Short IDs (1, 2, 3…) renumber when tasks complete/delete. Use the first 8 chars of the UUID — permanent and shown as uuid.short in all reports.

taskagent abc1234d info        # GOOD — stable partial UUID
taskagent 3 info               # BAD — short ID may change

When creating tasks, capture the UUID from the output for future references.

Status

Keep status current so the human's taskagent +ACTIVE list reflects reality:

taskagent <uuid> start   # began this deliverable
taskagent <uuid> stop    # paused it
taskagent <uuid> done    # complete — after review (see below)

Recording Learnings and Misalignments

This is the highest-value content taskagent stores. Annotate as they occur:

LEARNING = new durable info (user preference, codebase pattern, correct approach)

taskagent <uuid> annotate "LEARNING: <fact and where it applies>"

MISALIGNMENT = wrong direction taken and corrected (wasted effort)

taskagent <uuid> annotate "MISALIGNMENT: <what went wrong and the correction>"

A misalignment often produces a learning — record both. These are retrieved during /handoff and written to the plan file's ## Learnings section.

Which store: ticket-specific context → a taskagent annotation on the deliverable (surfaced when that task is read, and via /handoff into the plan file). Reusable insight that applies beyond this ticket — codebase convention, tool preference, recurring pitfall → also call memory_lesson_save, so agentmemory auto-injects it into future sessions. taskagent annotations don't decay; agentmemory lessons do if unused.

Checkpointing (only when state must survive)

Annotate progress only when the state needs to outlive the conversation: approaching compaction, ending a session mid-deliverable, or blocked. Do not annotate at every milestone within a session.

Write for a reader with zero conversation context:

COMPLETED: <deliverables>  IN PROGRESS: <state + next step>  BLOCKERS: <…>  KEY DECISIONS: <…>

Session Start Protocol

  1. taskagent ready — unblocked work.
  2. taskagent +ACTIVE list — anything left active.
  3. For active/relevant work: taskagent <uuid> agentinfo (or info).
  4. If the project is unclear, ask; then check ~/org/projects/<project>.md ## Agent section for taskagent project and plan file.
  5. Read the plan file if it exists.
  6. Report: "X ready, Y active. [summary]. Continue with Z?"

Core Operations

Run basic ops (add, annotate, modify, start, stop, list, info) directly — no subagent. For completion review run @task-reviewer-opus + @task-reviewer-gemini in parallel, then @review-merger (see Completion Review below); use @taskagent-reader for complex queries. For code review, ask the user to run /peerreview.

taskagent ready                         # unblocked tasks
taskagent project:<name> ready
taskagent blocked                       # blocked by deps
taskagent blocking                      # blocking others

taskagent add "Title" project:<name>
taskagent add "Title" project:<name> details:"context / why"
taskagent add "Found issue" project:<name> discovered_from:<parent_uuid>

taskagent <uuid> start | stop | done
taskagent <uuid> annotate "LEARNING: …"
taskagent <uuid> modify depends:<other_uuid>
taskagent <uuid> agentinfo | info
taskagent project:<name> list

Task Fields Reference

Field Purpose When
description Short title At creation
details Initial context, the "why" (not the full spec) At creation
project Links to org/projects At creation
annotations Durable LEARNING/MISALIGNMENT + handoff (not specs) During work
discovered_from Parent task UUID When discovered
+side_quest / +bug / +feature / +task / +chore / +epic Classification At creation

Completion Review (per deliverable, not per micro-task)

Run one review per deliverable-sized task (the PR/ticket) before taskagent <uuid> done. A separate review for every small sub-item is not required — review the thing that ships.

For externally-tracked work (Jira/GitHub):

  1. Before PR/merge: ask the user to run /peerreview <ticket-ID|PR-number> (launches code-reviewer-opus + code-reviewer-gemini as children of root — do not spawn @code-reviewer yourself).
  2. Before done: same parallel pattern, run it yourself: a. In a single response, issue two task calls at once — @task-reviewer-opus and @task-reviewer-gemini — both direct children of this session, with the input below. Never nest one inside the other or inside a third agent: a grandchild session's permission prompts never surface in the TUI and the whole run hangs silently. b. Once both return, call @review-merger with both reports labelled OPUS REVIEW: / GEMINI REVIEW:, and treat its output as the verdict.

Input for both reviewers:

Review task <uuid>
- Project: <name>   - plan file: <path or none>   - org/projects: <path or none>
- Summary: <1-2 sentences>   - Key files: <list or none>

Verdicts: PASS → may done. PASS WITH RESERVATIONS → ask the human. NEEDS WORK → fix and re-review.

What survives, what doesn't

  • Survives: taskagent data, org/projects/*.md, agentmemory, and the project plan file when it lives outside a per-ticket worktree.
  • Does not survive: conversation history and TodoWrite. A plan file that lives inside a throwaway worktree only survives as long as that worktree does.

Keep the plan file at a durable, project-level location (configured in org/projects), not inside a worktree that may be removed. If a decision or lesson must outlive any single worktree, also record it as a taskagent LEARNING or in the org/projects file.

Integration with TodoWrite

  • TodoWrite: this session's tactical steps; your own checklist.
  • taskagent: deliverables, status, durable lessons, offshoot work to remember.

At session start, read the relevant taskagent task and expand it into a TodoWrite checklist. Work, ticking TodoWrite. At a real boundary, record a LEARNING/handoff annotation and update status. Do not mirror every TodoWrite item as a taskagent task.

Plan file & org/projects

The project plan file holds the design, spec, wiring, and current state of the work. Keep it at a durable, project-level path (outside per-ticket worktrees) so it persists across tickets and sessions. The org/projects/<project>.md ## Agent section configures it:

## Agent
- **taskagent project**: `project-name`
- **plan file**: `~/path/to/plan.md`

taskagent project defaults to the filename; plan file defaults to PLAN.md in the repo root. Read it at session start; update its ## Notes / ## Learnings with significant progress (ask before editing ## Current Focus).

04 同仓库更多 Skills