AgentSkillsCN

delegation-create

独立的调度审批——扫描#needs-X标签，呈现分组审批菜单，执行标签转换。可通过“审查委派”“调度审批”“哪些需要调度”“委派审查”等语句触发。

SKILL.md
--- frontmatter
name: delegation-create
description: "Routes tagged work items to workers via async, blocking, or silent delegation. Triggers: \"delegate this\", \"route this work\", \"hand off this task\", \"delegation create\"."
version: 3.0
tier: lightweight

Routes tagged work items to workers via async, blocking, or silent delegation. Note: This is a utility skill — no session activation. Operates within the caller's session.

GATE CHECK — Do NOT proceed until ALL are filled in:

Output this block in chat with every blank filled:

Boot proof:

  • •COMMANDS.md — §CMD spotted: ________
  • •INVARIANTS.md — ¶INV spotted: ________
  • •TAGS.md — §FEED spotted: ________

[!!!] If ANY blank above is empty: STOP. Go back to step 1 and load the missing file. Do NOT proceed until every blank is filled.

Delegation Create Protocol (The Router)

Lightweight utility for routing tagged work items to other agents or workers. Can be invoked standalone or mid-skill. No session activation, no log file, no debrief.

1. Detect Context

Determine how this skill was invoked:

  • •Standalone: User typed /delegation-create directly. Look for a #needs-X tag and summary in the user's message or arguments.
  • •Mid-skill: Called from within another skill's protocol (e.g., during §CMD_WALK_THROUGH_RESULTS triage or §CMD_CAPTURE_SIDE_DISCOVERIES). The calling skill passes the #needs-X tag and surrounding context as arguments.

Extract:

  • •Tag: The #needs-X tag (e.g., #needs-implementation, #needs-research, #needs-brainstorm)
  • •Summary context: What the work item is about (from the tag's surrounding context, the user's message, or the calling skill's triage output)

If no tag is provided, ask: Execute AskUserQuestion (multiSelect: false):

"What type of work should be delegated?"

  • •"Implementation" — Code changes, feature building (#needs-implementation)
  • •"Research" — Deep investigation, domain exploration (#needs-research)
  • •"Brainstorm" — Ideation, trade-off analysis (#needs-brainstorm)
  • •"Documentation" — Docs updates, README writing (#needs-documentation)
  • •"Review" — Code/output review and validation (#needs-review)
  • •"Chores" — Routine maintenance, cleanup (#needs-chores)

2. Resolve Template

Use engine session request-template to find the REQUEST template for the detected tag:

bash
engine session request-template '#needs-X'

This scans ~/.claude/skills/*/assets/TEMPLATE_*_REQUEST.md for a template matching the tag. If no template is found, report the error and list available templates.

3. Pre-fill REQUEST

Populate the REQUEST template from current context:

  • •Topic: Derived from the tag's surrounding context or user description
  • •Relevant files: From the session's contextPaths, recently-read files, or the calling skill's loaded files
  • •Expectations: Derived from the tagged item — what the worker should deliver
  • •Requesting session: Current session directory (if operating within a session)

The REQUEST file must be self-contained (¶INV_REQUEST_IS_SELF_CONTAINED). A worker picking up this file should be able to start work without additional context from the delegating agent.

4. Show Summary + Confirm

Present a 2-3 line summary of what will be delegated. Do NOT show the full template.

Format:

Delegation summary:

  • •Tag: #needs-X
  • •Target skill: /skill-name
  • •Topic: [1-line description of what will be delegated]
  • •Key expectations: [What the worker should deliver]

5. Mode Selection

Execute AskUserQuestion (multiSelect: false):

"[Tag] detected: [Summary]. How should I handle it?"

  • •"Delegate — worker will notify" (Recommended) — Write REQUEST, tag #needs-X, pool worker picks it up. You continue working.
  • •"Await result from worker now" — Write REQUEST, start await-tag.sh for #done-X. Blocks until worker completes.
  • •"Spawn sub-agent to do it silently" — Pass REQUEST content to Task tool with appropriate subagent_type. Synchronous, in-process.

6. Execute by Mode

Mode A: Async (Delegate — worker will notify)

  1. •

    Write REQUEST file to current session directory via §CMD_DELEGATE.

    • •File naming: [SKILL_UPPER]_REQUEST_[TOPIC].md (e.g., IMPLEMENTATION_REQUEST_AUTH_VALIDATION.md)
    • •Topic portion is derived from the summary, uppercased, spaces replaced with underscores.
  2. •

    Apply tag to REQUEST file:

    bash
    engine tag add '[session_dir]/[REQUEST_FILE]' '#needs-X'
  3. •

    Report:

    "REQUEST filed: [REQUEST_FILE]. Worker will notify via #done-X tag."

  4. •

    Return control to the calling skill. If invoked standalone, the skill is done.

Mode B: Blocking (Await result from worker now)

  1. •

    Write REQUEST file to current session directory via §CMD_DELEGATE.

    • •Same naming convention as Mode A.
  2. •

    Apply tag to REQUEST file:

    bash
    engine tag add '[session_dir]/[REQUEST_FILE]' '#needs-X'
  3. •

    Start background watcher for completion:

    bash
    engine await-tag '[session_dir]/[REQUEST_FILE]' '#done-X'

    Run with run_in_background=true. The watcher polls until the tag transitions from #needs-X to #done-X.

  4. •

    Report:

    "REQUEST filed: [REQUEST_FILE]. Awaiting #done-X..."

  5. •

    Return control. The background watcher will notify when the worker completes.

Mode C: Silent (Spawn sub-agent)

  1. •

    Write REQUEST file to current session directory via §CMD_DELEGATE (audit trail — even silent work gets a REQUEST breadcrumb).

    • •Same naming convention as Mode A.
  2. •

    Determine subagent_type from tag:

    Tagsubagent_type
    #needs-implementationbuilder
    #needs-researchresearcher
    #needs-brainstormgeneral-purpose
    #needs-documentationwriter
    #needs-reviewreviewer
    #needs-choresgeneral-purpose
  3. •

    Launch Task tool with subagent_type and a directive containing the REQUEST expectations. The directive should include:

    • •The full REQUEST content
    • •The session directory path (for writing outputs)
    • •Clear success criteria from the Expectations section
  4. •

    On completion:

    • •Write a RESPONSE breadcrumb file: [SKILL_UPPER]_RESPONSE_[TOPIC].md in the session directory
    • •Swap tag: #needs-X to #done-X on the REQUEST file:
      bash
      engine tag remove '[session_dir]/[REQUEST_FILE]' '#needs-X'
engine tag add '[session_dir]/[REQUEST_FILE]' '#done-X'
  5. •

    Report result summary to user (2-3 lines from the sub-agent's output).

7. Return

Control returns to the calling skill protocol. If invoked standalone (not mid-skill), the skill is done.

No debrief. No session deactivation. No next-skill menu. This is a utility — it does its job and gets out of the way.

Constraints

  • •No session activation (¶INV_DELEGATE_IS_NESTABLE): This skill does not call engine session activate. It operates within the caller's session or standalone without a session.
  • •REQUEST files are self-contained (¶INV_REQUEST_IS_SELF_CONTAINED): A worker must be able to start work from the REQUEST file alone, without context from the delegating agent.
  • •Graceful degradation (¶INV_GRACEFUL_DEGRADATION):
    • •Async without fleet: REQUEST file still gets written and tagged. User can run the resolving skill manually later.
    • •Blocking without fleet: Falls back to async (REQUEST filed, no watcher — user checks manually).
  • •Silent mode is for fully-specified work only: If the work is ambiguous, complex, or requires interrogation, route to async or blocking instead. Silent mode skips interrogation — the REQUEST must be complete.
  • •REQUEST file naming: [SKILL_UPPER]_REQUEST_[TOPIC].md (e.g., IMPLEMENTATION_REQUEST_AUTH_VALIDATION.md, RESEARCH_REQUEST_XACTIMATE_CODES.md)

Auto-Degradation

  • •Blocking mode to async: If the session dies or the watcher fails, the tag persists on disk. A pool worker (or manual user) can still pick up the REQUEST via the #needs-X tag.
  • •Async mode to manual: If no fleet is running (no pool workers monitoring tags), the REQUEST file is still written and tagged. The user can invoke the resolving skill manually later by reading the REQUEST file.