Claude Code plugin that subjects non-trivial in-flight decisions to a fresh-context adversarial review before they stand — catching wrong directions while course-correction is still cheap, instead of only at final review.
Use when correctness matters more than speed: an architectural decision under uncertainty, non-trivial code about to be committed, a non-obvious claim ("this is safe", "this scales"), or work in unfamiliar code.
The cycle:
- CLAIM — name the decision and why it matters, in two or three lines.
- EXTRACT — isolate the smallest reviewable artifact + the contract it must satisfy, stripped of your own reasoning.
- DOUBT — dispatch a genuinely fresh-context reviewer (a non-
forkClaude Code subagent, or an external CLI) with an adversarial prompt that must find issues, not validate. Optionally offers a user-authorized cross-model second opinion (Gemini CLI / Codex CLI, read-only sandbox). - RECONCILE — classify every finding against the artifact text: contract misread, valid + actionable, valid trade-off, or noise.
- STOP — bounded at 3 cycles or trivial findings; escalates to the user rather than looping forever.
No credentials or external services required — cross-model escalation is opt-in per call and only runs a locally installed CLI the user explicitly authorizes.
Adapted from the doubt-driven-development skill in
addyosmani/agent-skills
(MIT licensed). Cross-references to that pack's own agents/ persona
roster and references/orchestration-patterns.md were rewritten to be
self-contained and Claude-Code-native (explicit subagent_type guidance
instead of a persona roster this plugin doesn't ship).
MIT