multi-agent-delegationlisted

# Multi-Agent Delegation Discipline Discipline for an orchestrator agent that delegates work to subagents, so that delegation does not end in a no-op (a successful-looking completion notice with zero artifacts) or a silent failure (an execution-environment difference that fakes success). ## When To Use - You delegated a task to a subagent and received a completion notice, but the expected files or changes do not exist. - The completion notice text reads as success, yet `git status --porcelain` or an artifact-existence check finds nothing. - You are about to write a delegation prompt and want the mandatory clauses (prevention is cheaper than recovery). - A review pass produced many findings and you want to fan them out to parallel subagents safely (see the ledger-as-spec section). ## Core Rule A subagent's completion notice is a claim, not evidence. Verify artifacts before acting on it or reporting it upstream. ## Tool Portability The principles here are tool-agnostic. Concrete names used below map as follows: - Claude Code: the `Agent` tool spawns subagents; `SendMessage` resumes a previously spawned agent by its id with context intact. - Codex and other agent CLIs: spawning is starting a delegated task or thread; "resume" means re-instructing the same thread or session so its context is retained, instead of starting a new one. - Verification commands are plain Git and shell, and work in any environment. ## Procedure ### 1. Mandatory clauses in every d