subagent-protocol

Build a self-contained briefing prompt with a task description, testable acceptance criteria, a scope fence of allowed files, a personality role card, an explicit list of context files to read, and a model hint. Never assume the subagent inherits parent session context.

Dispatch subagents only when the task splits into independent subproblems, specialized roles materially improve reliability, context would overflow one session, or parallelism justifies coordination cost. Tightly coupled work should stay single-agent with sequential role emulation.

Subagents end with DONE, DONE_WITH_CONCERNS, NEEDS_CONTEXT, or BLOCKED, plus a structured report block with summary, concerns, blocking items, and files changed. Specialized agents may add domain verdicts but must keep the shared top-level status line.

Review in two stages: first verify spec compliance against the brief's acceptance criteria, then assess code quality. Reviewers should read the actual diff rather than trusting the implementer's self-report, and a spec failure sends work back before quality review.

For BLOCKED, the parent triages by adding context, splitting the task, or escalating. For failed reviews, run at most three fix rounds with the verbatim findings list, then adjudicate remaining findings with a recorded ruling instead of re-dispatching identical prompts.

Use the fastest model class for exploration and file lookup, fast models for mechanical implementation with a detailed plan, default class for multi-file integration, and the strongest class for architecture, complex review, or a third fix round after two failures.