Remedy, Now in Critique Chat
From review to fix in one conversation. Pick a model, press Go, and watch an isolated cloud agent repair your PR — live, in your chat thread.
From review to fix in one conversation
Pick a model, press Go, and watch an isolated cloud agent repair your PR — live, in your chat thread. Your machine never runs a line of this code.
The gap between finding and fixing
Every code review produces the same frustrating pattern. You see the issues. You understand what needs to change. But actually fixing them means context-switching back into the codebase, remembering the exact lines, writing the patches, running the tests, and pushing — for every single finding.
That gap between "I know what's wrong" and "it's actually fixed" is where engineering velocity goes to die. Not in the review. In the repair.
Remedy closes that gap. And now it lives exactly where you already are — inside Critique Chat.
How it works
When a review completes in your chat thread, the AI will offer to start Remedy. You'll see a card with a model picker — choose your preferred model, hit Go, and the entire fix cycle unfolds in real time. Sandbox created. Repo cloned. Fixes written. Tests run. Commit pushed. Every step streams back to you live.
Your machine never runs a line of this code
This is not a convenience. It is a safety principle. The code that fixes your bugs should not run on your laptop. It should run somewhere isolated, disposable, and auditable.
Triggering Remedy from Chat
There are two natural ways Remedy appears in your workflow.
- Review completes with findings
- AI suggests "Want me to start Remedy to fix these?"
- Confirmation card appears in thread
- Pick model, press Go, done
- Ask directly: "start remedy on this review"
- Or: "fix the issues you found"
- Same confirmation card appears
- Same flow from there
The confirmation card shows your repository name, PR number, and how many findings are queued for fixing. It is a single decision point — not a configuration screen, not a wizard. Pick a model and go.
The model menu
Remedy offers a curated set of models, each with different strengths and credit costs. The key insight: you do not need the most expensive model for every fix. A linting error and a security vulnerability deserve different reasoning budgets.
End-to-end cost includes model + managed execution (2cr fixed)
| Model | Provider | Model Cost | Execution | Total From |
|---|---|---|---|---|
| Qwen3.6 Plus | Alibaba | Free | 2cr | 2cr |
| KAT Coder Pro V2 | KwaiPilot | 1cr | 2cr | 3cr |
| MiniMax M2.7 | MiniMax | 2cr | 2cr | 4cr |
| GLM-5 Turbo | Zhipu | 3cr | 2cr | 5cr |
| MiMo V2 Pro | Xiaomi | 3cr | 2cr | 5cr |
| GPT-5.4 Mini | OpenAI | 6cr | 2cr | 8cr |
Re-review loop (if validation fails and retries) adds 3–12cr depending on model. Max 2 loops.
The pricing is transparent and end-to-end. Model cost plus a fixed 2cr for the managed execution layer — sandbox provisioning, repo cloning, validation, git operations, and push. If validation fails and Remedy retries (up to 2 loops), the re-review adds 3–12cr depending on which model you chose.
What Remedy actually does
Inside that sandbox, here is the full cycle:
The validation step is where most autonomous fix tools fail. Remedy runs your actual test suite, linter, and build process. If anything fails, it gets the error output, understands what went wrong, and tries again — up to two loops. If it still cannot pass validation, it stops and hands back to you with a full report.
Safety rails you can see
- Modify only files identified in review findings
- Run validation commands (test, lint, build)
- Commit and push to the PR branch
- Rebase on non-fast-forward conflicts
- Touch files outside the review scope
- Run code on your local machine
- Push without passing validation first
- Run more than 2 autonomous loops
- Access repositories it cannot verify
If Remedy attempts to modify a file that was not part of the review findings, the execution halts immediately. Scope rejection is not a warning — it is a hard stop.
Recommended flows
Review finds linting errors, missing null checks, or straightforward style issues. Use Qwen3.6 Plus (free model cost). Remedy patches the files, runs lint, pushes. Done in under a minute. Cost: 2cr.
model: Qwen3.6 · cost: 2crReview identifies a logic bug, security pattern violation, or architectural drift. Use GPT-5.4 Mini or MiMo V2 Pro. Remedy writes the fix, runs full test suite, validates build, pushes. Cost: 5–8cr.
model: GPT-5.4 Mini · cost: 8crThe dashboard entry point still exists for "queue and forget" workflows — click Queue Remedy on a review run and walk away. But the chat flow is where Remedy feels most natural. You are already in conversation with your codebase. Fixing becomes just the next message.
What you see in real time
The confirmation card does not disappear after you press Go. It transforms into a live progress feed. Each phase streams in:
E2B spins up an isolated cloud VM with your review findings loaded.
Your repository is cloned into the sandbox and checked out to the PR head.
OpenCode reads the findings and writes patches to the affected files.
Test suite, linter, and build process execute against the patched code.
Changes committed with a descriptive message and pushed to your PR branch.
If validation fails on the first attempt, you will see the error output and a retry notice. The AI gets the failure context and tries again. You watch the entire thing — no black boxes, no silent failures.
Who this is for
- 1Do you have review findings that need fixing?If yes, Remedy is designed for exactly this.
- 2Are the fixes deterministic and verifiable?Linting, tests, straightforward bugs — ideal. Architectural rewrites — probably not.
- 3Does your repo have a working test suite or lint config?Remedy uses these for validation. Without them, it can still write fixes but cannot auto-validate.
- 4Are you comfortable with an AI pushing to your PR branch?You stay in control. Remedy only modifies files in scope and only pushes after validation passes.
The teams getting the most value from Remedy are the ones who already have clean CI pipelines, reliable tests, and a structured review culture. Remedy does not replace engineering discipline — it amplifies it.
From "understand" to "fixed" without leaving chat
Critique Chat was built to help you understand your codebase. Ask questions, get answers grounded in real file paths and line numbers. Now Remedy extends that conversation from understanding into action. You do not just learn what is wrong — you watch it get fixed.
The future of code review is not a longer list of findings. It is a shorter distance between finding and fixing. Remedy in Chat is our answer to that distance.
Try Remedy in Chat now.
Connect GitHub, run a review, and let Remedy fix the findings. Start with Qwen3.6 Plus — free model cost, 2cr execution.
Open Chat →