Critique Code
Critique Code writes the change, runs the checks, and keeps going until the evidence says stop. Same class as Codex and Claude Code, not a comment bot on top of them. It happens to be lethal at review because it is a full coding loop, not a reviewer pretending to code.
$
Then critique-code or critique-code web in the repo you want to edit. Separate from @critiquedotsh/cli.
Point Critique Code at the keys you already pay for. Or route through Critique Inference and skip the provider wrangling. The harness does not care which logo sits on the model.
The test spine
Critique Code is a general coding agent. It implements, refactors, and patches. Cheap checks spend first. Focused tests try to break the change. Review is not a separate product. It is what happens when the same harness refuses to call the job done on vibes.
trust.preflightsecrets, digest, redaction maptypecheckzero agent costlintdeterministic deltasecrets.deltano new credential materialcontract.breakpublic surface still holdstest.focusedCheckRequest, not a model shellreproducecandidate state still breaksnearby.testschanged lines executedadjudicatefresh session, no specialist chatfindingor it stays HYPOTHESISCodex-class loops wrap a model and keep writing. Critique Code still writes. It also owns what may run, which repository state was examined, whether an observation is evidence, and whether the job is actually done. The agent kernel is replaceable. The controller is not.
Critique decides what may be executed, which commit was examined, how tools are provisioned, and which CPU, memory, process, network, and time limits apply. A model turn cannot widen that envelope.
The model names a capability, a target, and a purpose. The harness resolves the binary, arguments, working directory, timeout, network, and write policy. Shell operators, path escapes, and unknown capabilities are rejected.
A tool fact is an observation. Evidence is that fact bound to repository state and tool identity. A hypothesis is a falsifiable claim. A finding is promoted only after a fresh adjudicator tries to kill it. Agreement is never proof.
The loop can inspect, patch, and re-verify. Specialists do not debate the adjudicator. Candidates are selected by evidence. A blind verifier sees the defect, the patch, and the workspace, not the repair chat.
Critique Code is @critiquedotsh/harness, binary critique-code. Critique CLI is @critiquedotsh/cli, binary critique. The CLI will not become Code. Other agents can call the sidecar, or you can run critique-code yourself.
Critique CLI as a teammate
Wire Critique CLI into Codex, Claude Code, Cursor, or OpenCode. The managed runtime other agents already know. Local working tree, optional cloud, finish.v1.
$ critique integrate --agent all
critique finish --intent "add retries and cover them"
Use it like Claude Code or Codex
critique-code is the first-class agent command. Same kernel in the TTY or in a local browser via critique-code web. It writes, tests, repairs, and reviews inside the same harness.
$ critique-code
critique-code web
Bring your own API providers, or use Critique Inference.
Local browser
critique-code web serves the same author session in a browser on 127.0.0.1. Composer, transcript, Review, Repair, Ship. The process still owns CAS writes, user-gated runs, and the /review gate. The tab can close. The kernel cannot move to Critique Cloud or to someone else's web app.
Interactive commands
In a critique-code session, type / or /help for a selectable picker. critique-code web exposes the same lines as buttons. Lines that start with / never go to the model. Spoken review, repair, ship, and exit match the typed commands when voice is on. An empty clip after /voice on returns to the keyboard.
Session
Voice
Proof
Config
End
A general coding harness. The split is who owns the shell, what counts as done, and whether a test log is evidence. Snapshot numbers are the supplied DeepSeek V4 Flash slice, not a Critique evaluation.
Controller-owned harness
Coding agent harness
Coding agent harness
Hosted Critique Code, free to measure. Fair-use and abuse safeguards still apply. Model availability can change as the beta evolves.
UnlimitedDeepSeek V4 Flash on Verify for the CritiqueCode beta.
100B+DeepSeek V4 Pro tokens to test the latest Pro route on real jobs.
On the supplied DeepSeek V4 Flash snapshot, CritiqueCode passed 20 of 30 tasks. Claude Code and Codex passed 16. Cost per success was $0.028 against $0.195 for Claude Code and $0.081 for Codex. Median time was 132.2s, between Claude Code at 122.7s and Codex at 245.0s. Latency is mid-pack, not a wall-clock miracle.
DeepSeek V4 Flash, 30-task slice
DeepSeek V4 Flash, 30-task slice
Lower bar is the spend. Longer fill is cheaper.
DeepSeek V4 Flash. CritiqueCode is not the fastest.
This is an externally supplied snapshot, recreated from the introducing-CritiqueCode post. Pass rates for Claude Code, Codex, and OpenCode are the 16/30, 16/30, and 14/30 task counts from that slice. It is useful context for why harness design changes cost and pass rate. It is not a Critique evaluation, and it is not a guarantee on your repository. CritiqueCode will be promoted on our own corpus: same task, same model, same plan, same verifier.
An agentic coding harness. Same class as Codex and Claude Code: it writes code, runs tools, tests the change, and patches until the evidence says stop. It is also unusually good at review, because review is the same loop, not a sidecar. It is not a model, not a comment bot, and not Critique CLI.
Critique CLI is @critiquedotsh/cli, binary critique. Critique Code is @critiquedotsh/harness, binary critique-code. Other agents can call the CLI, or you can run critique-code yourself. The CLI will not become Code.
npm install --global @critiquedotsh/harness, then run critique-code or critique-code web in the repository you want to edit. The web UI binds 127.0.0.1 only. Run critique-code login and approve on the website to use Critique Inference with no pasted key. Interactive sessions keep stdout quiet; pass --json for the machine envelope. Keep @critiquedotsh/cli if other agents should call critique integrate --agent all or critique finish. OpenRouter Activity attributes Critique Code to critique.sh.
A local browser client of the CritiqueCode author process. Run critique-code web in the repo you mean to edit. It binds 127.0.0.1 only. Same /review, /repair, and /ship as the TTY. It is not OpenCode's web UI, not the Critique CLI sidecar, and not a hosted session.
Default is Critique Inference (critique/auto) after website approval. You can still paste OpenAI, OpenRouter, Anthropic, Google, Qwen, and other keys. The harness is not locked to one lab.
Same category: coding agent harnesses. On the supplied DeepSeek V4 Flash 30-task snapshot, CritiqueCode passed 20 of 30. Claude Code and Codex passed 16 of 30. Cost per success was $0.028 against $0.195 for Claude Code and $0.081 for Codex. Median time was 132.2s, between Claude Code at 122.7s and Codex at 245.0s. That snapshot is not a Critique evaluation. Critique Code is a general coding loop that also happens to be lethal at review.
You can run it instead of those loops. You can also keep them and have them call Critique CLI. Those are the two local options. Critique Code itself is not a PR comment bot sitting on top of someone else's agent.
Open Verify, pick CritiqueCode (Beta), and run a real job. For the beta, DeepSeek V4 Flash is free and unlimited. We are also offering over 100 billion DeepSeek V4 Pro tokens. Fair-use and abuse safeguards still apply.
Yes. In a critique-code session, /voice records one prompt and transcribes it with Qwen3 ASR 0.6B. /voice on keeps listening after each reply. Spoken review, repair, ship, and exit still run as commands. critique-code --voice starts already listening. Transcription bills $0.000003 per second of audio on the Inference API. The recorder is bundled ffmpeg; you do not need a system sox install.
Type / or /help for a picker. Slash lines never go to the model. /login connects Critique Inference. /voice, /voice on, and /voice off handle speech. /review, /review all, /repair, and /ship run the evidence loop. /settings, /models, /keys, and /skills configure the session. /done ends after a review. /exit leaves. Unknown /foo is rejected, not forwarded as chat.