| name | generate-agentdesk-task |
| description | Explicitly specified only: use this skill only when the user writes `$generate-agentdesk-task` or `generate-agentdesk-task` by name. Never infer or proactively trigger it from task content, workflow fit, or task-generation requests alone. When explicitly requested by name, create or prepare an AgentDesk task for a project before subagent execution. |
Generate AgentDesk Task
Explicit Invocation Only
The user must explicitly specify this skill by writing $generate-agentdesk-task or generate-agentdesk-task by name. Never infer this skill from task content, workflow fit, or task-generation requests alone.
Use this skill to turn a user request into an AgentDesk control-plane task.
By default, generate the task title, brief, checklist, and task.md content in Chinese unless the user explicitly asks for another language or the target repository clearly requires another language.
AgentDesk is CLI/MCP-only and headless: use the MCP stdio tools or verunectl for task generation and execution. There is no web UI or HTTP API, so do not add Web app, browser, or npm run dev validation steps.
AgentDesk GitHub Version Check
At the start of every invocation, before reviewing task brief completeness, run the best-effort update helper next to this SKILL.md when it is available:
sh /path/to/generate-agentdesk-task/check-github-version.sh
If the helper reports that AgentDesk or generate-agentdesk-task differs from GitHub agentdesk/next, surface that warning to the user immediately and then continue the normal workflow unless the user asks to stop. If the helper is unavailable, GitHub cannot be reached, or the check is disabled with AGENT_DESK_SKIP_UPDATE_CHECK=1, do not block task generation.
Task Brief Completeness Review
Before creating an AgentDesk task, review whether the user's request is complete enough to produce an executable task.md.
Treat the request as complete when it provides, or the current workspace lets you safely infer, these essentials:
- Target project root or repository.
- Concrete goal or outcome.
- Scope boundaries, including what should not be changed when relevant.
- Acceptance criteria or verification expectations.
- Important constraints such as branch, launcher, concurrency, files, services, or deadlines.
If any essential detail is missing and cannot be safely inferred from local context, ask the user a concise follow-up question and wait for the answer before calling create_agentdesk_task or the CLI fallback.
Do not ask for extra detail when the missing information is non-blocking or the repository context provides a reasonable default.
Workflow
- Confirm the project root from the current workspace unless the user provides one.
- Review task brief completeness using the checklist above; ask the user for missing blocking details before generating the task.
- When the task is intended for subagent execution, include enough structure in the generated brief/task notes for the coordinating model to assess task complexity, file ownership, and likely concurrent-edit conflicts before choosing a per-batch subagent count.
- Keep coordinator duties out of the executable subtask checklist. Concurrency planning, launcher selection, user notification, and final cross-agent report aggregation belong in Context or Acceptance Criteria, not as subagent-owned checkbox items.
- For validation, QA, smoke-test, or health-check tasks, instruct subagents to write only their required
report.json and session artifacts. If the user wants a repository-level report under docs/, make that a main/coordinating-agent aggregation step after the subagent session finishes.
- Prefer MCP:
- Call
create_agentdesk_task with projectRoot, title when known, and a concrete Chinese brief.
- Then call
read_agentdesk_task until the task reaches ready or failed.
- Fallback to CLI when the MCP tool is unavailable:
- Run
./scripts/verunectl.sh tasks create --project <projectRoot> --title "<title>" --brief "<brief>" --json, using a Chinese title and brief by default.
- Use
./scripts/verunectl.sh tasks show <taskId> --project <projectRoot> --json to inspect the generated task.
- Verify the generated task has executable checklist subtasks.
- Report the
taskId, status, task file path, and any failure message.
Guardrails
- Keep AgentDesk centered on
task.md; do not reintroduce PRD JSON, Gemini CLI, or Claude Code compatibility flows.
- Prefer the MCP stdio server and
verunectl as the supported interfaces.
- Generate user-facing task content in Chinese by default, including titles, briefs, checklist items, task notes, and final summaries.
- Prefer explicit session defaults in user-facing notes:
gpt-5.5, xhigh, service tier fast, execution mode auto, batches of up to 6, and the repository-designated working branch as the base for new work.
- If the user or repository designates a non-
master working branch, carry that branch into task notes and do not describe future work as based on or automatically integrated into master. In this AgentDesk workspace, use agentdesk/next as the working branch and leave any merge to master for the user to perform manually.
- Make clear that before subagent execution, the coordinating model should review task complexity and concurrency-conflict risk, decide a recommended per-batch subagent count within the configured cap, notify the user of that recommendation, and respect any later user-selected concurrency.
- Do not generate an executable checklist item whose only job is "assess concurrency before starting"; that work must happen before
start_subagent_session.
- For health-check or validation tasks, do not assign repository-level final report creation to a subagent. Subagents should leave evidence in their AgentDesk
report.json and session artifacts; the main agent should write any final tracked report after reviewing all results.
- If task generation fails, preserve the task directory and summarize
stdout.log / stderr.log rather than deleting artifacts.