subagent-delegation · git:20261103.530d686 · 2026-11-03 · sha256 5a6af8b813a314ab
subagent-delegation git:20261103.530d686A
Immutable. This exact content never changes and is served at /api/v1/blob/5a6af8b813a314ab.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_ei4mzweks7pxlcyf/5a6af8b813a314abd9f979c4177a6910e860073cf0c28b5e3939f92a85ba7289.json
Then /plugin install subagent-delegation@markdownregistry-ei4mzweks7pxlcyf in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: subagent-delegation
description: Use when the user invokes the subagent delegation playbook, or the user's instructions direct its use when delegating work or deciding whether to delegate.
---
# Subagent delegation
## Core principle
Match each bounded assignment to the capability and judgment it needs, while
keeping architecture, integration, authorization, and final acceptance with
the lead. Installed agent definitions are authoritative for pinned settings.
## Procedure
1. Apply this playbook when the user invokes it or the user's instructions
direct its use for delegation; either covers delegation and reassignment
throughout the current task. Without either, a generic request to implement
work or use subagents does not apply it. Keep required constraints in agent
briefs so they apply without child agents loading this skill. Loading the
skill does not authorize implementation.
2. Prefer solo work for small or tightly coupled tasks; use delegation or
worktrees when independent work or isolation justifies them. Leave task-graph
execution, worktree lifecycle, integration sequencing, and implementation
gates to the calling workflow, such as `implement-with-subagents`.
3. Choose the capability before the runtime role name: specific discovery,
predetermined mechanical checking, implementation with self-review and
repair ownership, or independent read-only analysis. For analysis, state
the mode: review against requirements, or investigation of a concrete
unresolved question. Establish concrete evidence before an investigation.
4. Load only the active runtime's reference:
- [Codex](references/codex.md): exposed roles, installed TOML pins, model
and effort routing, and fresh-context calls.
- [Claude Code](references/claude-code.md): agent definitions, model and
effort routing, invocation overrides, and owner/reviewer lifecycle.
- Other runtimes: inspect their exposed capabilities and installed policy;
preserve configured choices rather than translating either provider's
model identifiers or effort values. Resolve missing required settings
before dispatch.
5. Reassess capability and judgment for every new assignment, including
reassignment, independently of the parent task's size or model. Preserve
pinned settings and access restrictions. If required capability or settings
are unavailable, report and resolve the mismatch before dispatch. Name the
assignment category in the brief and justify routing departures. For
investigations, name the unresolved question; for reviews, identify the
correctness checks; for synthesis, identify the evidence or deliverable
demands. If work stalls, diagnose missing inputs, tool failures, or routine
test repairs before considering a higher tier.
6. Give each agent a self-contained brief that fills every field of the
[handoff brief](references/subagent-selection.md#handoff-brief). Tell
workers they share the codebase, must preserve others' edits, and must
accommodate concurrent changes. Further delegation
requires the lead's explicit authority. Keep remote mutations with the
root task holding the user's authorization, with one mutation owner.
7. Launch independent assignments in parallel. Queue work behind another agent
only for a stated dependency, such as shared files, a contract, or ordering;
name that dependency. Separate worktrees alone do not remove dependencies.
8. Choose independent audits proportionally to risk. Use a fresh read-only
reviewer without the implementation conversation. Supply original
requirements, the plan when available, and exact review scope; require a
`ship`, `fix-first`, or `rethink` verdict with evidence. Keep repairs with
the original worker and context where possible; the lead may take ownership
or explicitly reassign them.
9. Inspect delegated diffs and verification evidence before accepting the work.
Repeat checks when evidence is incomplete or affected code changes, without
duplicating adequate validation. Return failures to the owner for repair
and reassess affected evidence.
## Finish gate
Finish when the lead has inspected the assigned result and its evidence,
resolved failures or reported the specific blocker, and made the acceptance
decision under the calling workflow. For advice-only requests, finish with the
assignment choices and their reasons; do not spawn agents or start execution.