Skip to content
Critique/docs
Reference

Remedy

Bounded autonomous fixes after a review — scope, validation, and safety limits.

Remedy is Critique’s managed execution path: after a review identifies issues, Remedy can apply a small, scoped patch in an isolated environment, run validation commands, and optionally push a commit — without you leaving the product.

This differs from BYOA, where you queue Cursor Cloud Agents, Claude Managed Agents, or OpenAI Codex with your own API keys — or paste a Fix prompt into Copilot and other agents.

When to use Remedy

  • Several findings map to concrete file edits inside the PR’s scope.
  • You want validation (lint, test, build) run automatically after the patch.
  • You accept bounded automation with a hard stop when scope is unclear.

Use BYOA (queue Cursor, Claude, or Codex from the review run) or Fix prompt when you prefer your own agent subscription or manual control. See BYOA setup and export endpoints.

What Remedy receives

Remedy builds an internal blueprint from the review run:

  • Repository and PR identity (branch, SHAs, clone context)
  • Review verdict, summary, and structured findings
  • Optional operator instructions you provide at start time

Goals are derived from findings (typically up to a handful of top issues) or cautious “confirm before editing” goals when no findings exist.

Write scope

Remedy may only modify files on an allow list derived from the review:

  • Paths mentioned in findings, plus related test files when inferable
  • If no paths exist, a capped set of evidence files from the review

If a fix needs files outside that list, Remedy stops and returns a human handoff instead of widening scope silently.

Validation

After patching, Remedy runs repository-appropriate checks when it can infer them (for example lint, test, or build scripts common in JS/TS repos). Results appear in the run timeline.

Safety limits

ConstraintWhy
Maximum execution loopsPrevents endless fix-and-retry cycles (hard cap of two loops)
Scope enforcementNo drive-by refactors in unrelated files
Human handoffRequired when validation fails twice or the fix needs product judgment

Remedy emits a timeline of phases (sandbox ready, patch produced, validation passed/failed, push result) visible in the dashboard and stream UI.

Credits

Remedy consumes more credits than chat Q&A because it runs agent loops and sandboxes. See Billing & credits.