Skip to content
Skip to content
Product7 min read

Introducing CritiqueCode: Verify’s code-repair harness

CritiqueCode (Beta) is the Critique-owned harness for bounded repair runs, with explicit task contracts, controlled tools, durable artifacts, and verification.

Repath Khan

CritiqueCode

A repair runtime—not a prompt wrapper

critique.sh/verify

Now in beta

CritiqueCode is Verify’s new code-repair harness. Critique owns the task contract, tool policy, budgets, event protocol, candidate artifacts, and verdict. Choose OpenCode when you want the established option; choose CritiqueCode (Beta) when you want to help us measure our own harness.

284B
total MoE parameters
13B
active parameters per token
1M
context window
Opus-level
coding target for the trial

DeepSeek V4 Flash is a mixture-of-experts model with 284B total parameters and roughly 13B active per token, a 1M-token context window, and hybrid attention designed for long repository sessions. We describe it as an Opus-level coding alternative in this trial because it is aimed at frontier reasoning and software work—not because we are claiming universal parity with every Opus benchmark. The specs are provider-reported and may change; check the live model page for the current route.

A coding agent is not only a model. The harness decides what the model receives, which tools it can use, how activity becomes evidence, when the loop stops, and who gets to call a patch correct. If those decisions live in an executable we merely spawn, the product is still a wrapper around someone else’s runtime.

CritiqueCode is not a model or a bundle of repository skills. It is the harness that turns a repair request into a bounded, observable run: a canonical task resolved against an immutable repository SHA, a stable event protocol, a controlled tool environment, run budgets, durable artifacts, and a verdict that distinguishes a candidate patch from a verified repair.

CritiqueCode beta run
Resolve the repository and exact base commitStart a bounded CritiqueCode session in an isolated sandboxCapture canonical events and a frozen candidate patchApply the patch in a clean worktree at the recorded SHAPublish a verdict and evidence—not agent confidence

Skills are useful instructions. They do not, by themselves, make a repair reproducible or safe. A production harness needs to know which commit it touched, enforce a ceiling before a tool loop becomes a bill, prevent repository-controlled extensions from becoming executable authority, preserve the patch that was tested, and tell the truth when it has only checked that a patch applies.

What CritiqueCode owns
The harness is the product contract—not a prompt wrapper.
LayerCritiqueCodeWhy it matters
TaskServer-resolved issue context and immutable base SHA.A run can be explained and replayed against the source it actually saw.
ToolsRead, search, edit, and guarded shell tools in an allowlisted environment.Repository instructions cannot become system authority or recover provider credentials from a child shell.
EvidenceCanonical events, a patch hash, changed paths, numstat, and clean-worktree checks.A PR is not mistaken for proof that a bug is fixed.
VerdictStructural success is PARTIAL and refunded until behavioural evidence exists.Verify earns trust by being explicit about what it has and has not proven.

You may have seen recent harness-comparison graphics showing that the same model can produce very different cost, speed, and pass-rate outcomes when the harness changes. That is directionally important, but it is not a CritiqueCode benchmark. We do not know the exact task set, provider route, versions, budgets, policy, or grading method behind the externally supplied charts, so we will not borrow their rankings as a product claim.

Pass rate

DeepSeek V4 Flash · 4 harnesses

Pass rate
HarnessValue
CritiqueCode66.7%
Prime Agent62.5%
Deep Agents53.3%
Hermes Agent50.0%

Tasks passed out of 30

DeepSeek V4 Flash · 8 harnesses

Tasks passed out of 30
HarnessValue
CritiqueCode20 passed
Oh My Pi17 passed
Claude Code16 passed
Codex16 passed
Deep Agents16 passed
Prime Agent15 passed (+6 not graded)
Hermes Agent15 passed
OpenCode14 passed

Cost per successful task

DeepSeek V4 Flash · external comparison

Cost per successful task
HarnessValue
CritiqueCode$0.028
Deep Agents$0.045
Hermes Agent$0.056
OpenCode$0.073
Codex$0.081
Oh My Pi$0.103
Prime Agent$0.131
Claude Code$0.195

Median time per task

DeepSeek V4 Flash · external comparison

Median time per task
HarnessValue
Claude Code122.7s
OpenCode129.7s
CritiqueCode132.2s
Hermes Agent175.5s
Deep Agents187.1s
Prime Agent242.1s
Codex245.0s
Oh My Pi272.4s

External harness comparison attributed to Composio agentic eval. The first series is shown here as CritiqueCode for product naming; these are not Critique-run results and do not establish a production performance guarantee.

Both choices use the same Verify plan, repository flow, draft-PR behaviour, and refund promise. OpenCode remains the established option. CritiqueCode is an opt-in Critique-owned harness running in a pinned E2B runtime. It uses the existing OpenRouter model path, records harness provenance on the run, and is deliberately not the automatic default while we complete smoke coverage and compare real jobs.

The important boundary is unchanged: an agent saying “done” means a candidate is ready. It does not mean a customer should be charged or a repair should be called fixed. That decision belongs to the evidence and verdict pipeline around the agent.

Try CritiqueCode Beta
Run the same repair with the harness you want to evaluate. We will keep the evidence and refund semantics honest while the beta earns its default.