subagent-prompt-review

name

subagent-prompt-review

description

Review a prompt about to be sent to a subagent (Agent tool, sqfan dispatch, squire env, scheduled remote agent, MCP-driven worker) for defects that cause silent failure, wasted runs, or runaway scope. Single-axis review: is this prompt fit-to-dispatch? Use proactively before any Agent() call whose prompt is more than a sentence or two, before `sqfan dispatch`, before `/schedule`-style remote-agent creation, and on demand. Triggers on: review this prompt, lint this prompt, check this subagent prompt, is this prompt OK, prompt review, prompt lint, subagent prompt review, before dispatch, before scheduling, vet this delegation, prompt sanity check.

Subagent Prompt Review

Single axis: is this prompt fit to dispatch? Subagents start cold — no parent context, no shared state unless given, no follow-up questions. Applies to any dispatch surface: Agent tool, sqfan dispatch / sqfan nudge, squire task create, scheduled remote agents.

Skip trivial one-liners ("run make test") — they don't need cwd, branch, or success criteria, and the review costs more than it saves.

BLOCKING — run fails or does harm

• Nonexistent model name. Verify against the canonical source, never training data:

  • sqfan: defaultAllowedModels in /Users/rch/repo/sqfan/pkg/opencode/allowlist.go plus allowed_models in ~/.sqfan/config.yaml.
  • Agent tool model: must be sonnet | opus | haiku | fable; anything else fails with InputValidationError.
  • /schedule-style remote agents: claude-{opus,sonnet,haiku}-{version} (e.g. claude-sonnet-4-6), checked against a current source.