delegate

Prepare a task for delegation to another agent (Codex, Gemini, human) with full context and guardrails.

SKILL.md

GitHub

Delegate

Prepare a task package for delegation to another agent or human, with context, constraints, and verification criteria.

Why Not Just [dispatch]?

[dispatch] spawns sub-agents within the same session. [delegate] prepares work for EXTERNAL agents (Codex, Gemini, the team lead, a contractor) who work asynchronously via the bus.

Execution

0. Emit SKILL_INVOKE

Post SKILL_INVOKE to the bus before any stateful action.

Type: SKILL_INVOKE
To: all
Message: [skill=delegate] [mode=side_effecting] [args_hash=<sha256>] [session=<session_id>]

(The skill invocation runtime injects the caller's canonical identity as from_id.)

1. Define the task

TASK: <clear, specific deliverable>
ASSIGNED TO: <agent identity>
DEADLINE: <when>
PRIORITY: <P0/P1/P2>

2. Provide context

What does the delegate need to know?

  • Relevant files and their purpose
  • Current state of the work
  • Decisions already made (link to CLP entries)
  • Constraints (stdlib-only, no --no-verify, etc.)

3. Set guardrails

Based on the delegate's trust level:

  • Scope: Which files/directories can they modify?
  • Commit limits: Max LOC / files per commit
  • Prohibited actions: What they must NOT do
  • Review requirement: Does their work need Claude review before merge?

4. Define done

  • Specific deliverable (code, doc, analysis)
  • Tests passing
  • PR created (or bus RECEIPT posted)
  • No guardrail violations

5. Post to bus

Post TASK_REQUEST to the bus for the target agent.

Type: TASK_REQUEST
To: <agent>
Message: TASK: <description>. SCOPE: <files>. DEADLINE: <when>. VERIFY: <criteria>.

(The skill invocation runtime injects the caller's canonical identity as from_id.)

6. Monitor

Use [agent-audit] <agent> to check progress, or watch the bus:

grep "<agent>" _state/coordination/messages.tsv | tail -10

Output Format

Delegation Package | <task> -> <agent>
══════════════════════════════════════

## Task
<clear deliverable>

## Context
<what they need to know>

## Guardrails
- Scope: <approved directories>
- Limits: <LOC/file caps>
- Prohibited: <what not to do>
- Review: <required/optional>

## Done Criteria
- [ ] <deliverable>
- [ ] <tests>
- [ ] <bus receipt>

## Bus Message Posted
<timestamp and content>

Base120 Context

  • Primary: SY11 (Governance Patterns)
  • Related: CO17 (Orchestration vs Choreography), SY13 (Incentive Architecture)

Skill Chains

  • For delegation tasks needing LLM inference -> [reasoning-router] (python ~/bin/reasoning_router.py route)

Mandatory

None — generates a delegation packet; the delegated task itself has its own chains enforced at execution time.

Advisory

  • [agent-audit] — monitor delegated task progress after packet is posted
  • [dispatch] — for same-session sub-agent spawning (contrast with external delegation)
  • [cross-runtime-bridge] — for executing Devin→opencode delegation via python ~/bin/cross_runtime_bridge.py delegate "<message>"

Authority

  • T1 (TRUSTED): May run (prepare delegation packets for any agent/human)
  • T2 (Active/High): May run (prepare delegation packets for any agent/human)
  • T3 (Medium): May run (prepare delegation packets for any agent/human)
  • T4 (Probationary): May run (generates packet only, does not execute the delegated task)
  • Operator: Override any restriction