A code reviewer that reads before it comments.

ocra splits a change into focused review tasks. Each agent can only read your repository, has to quote the code it means, and must say what it checked. Most runs end with a handful of findings. Some end with none, and that is fine.

Get started View source on GitHub

Node 22+, any model OpenCode supports

ocra review — example output

$ ocra review --from main
[ocra] 4 file(s) selected, 1 excluded · risk tier: full
[ocra] 2 bundle(s) (grouped)
[ocra] 4 review task(s), 2 reviewer/bundle pair(s) skipped by scope
[ocra] security-1 completed in 9.2s · 2 finding(s)
[ocra] correctness-1 completed in 7.4s · 2 finding(s)
[ocra] Verified 3 finding(s), dropped 1 that the code disproves
[ocra] Verdict: significant concerns
 
src/auth/session.ts
  critical   L42       Every session is treated as expired [verified] #a41c9e03
    isExpired() compares expiresAt in seconds with Date.now() in
    milliseconds, so users are logged out right after signing in.
    Suggestion: return session.expiresAt * 1000 < Date.now();
 
1 finding(s) (1 critical, 0 warning, 0 suggestion) · tokens: 48912 in (31040 cached), 1106 out, 212 reasoning · $0.0412

What happens when you run ocra review

The steps that must not go wrong are plain, tested code. Models are only asked for judgment.

Example run · Changed files

Select

Five files changed. The lock file is set aside with its reason; four are worth reading.

src/auth/session.ts
src/auth/token.ts
src/api/login.ts
docs/auth.md
package-lock.jsonlock file, set aside

Binaries, lock files, vendored and generated code, likely secrets and oversized diffs are set aside, each with a recorded reason. Migrations are always kept. A secret can never be opted back in.

Matrix

Each bundle gets the reviewers whose scope fits. Security and performance skip the docs bundle, and the report lists why.

Bundle 1

src/auth/session.ts
src/auth/token.ts
src/api/login.ts

correctnesssecurityperformance

Bundle 2

docs/auth.md

correctnesssecurity, performance · skipped by scope: docs only

Correctness always runs. Security and performance join from the lite tier and skip documentation and tests. Every skipped pair is listed in the report, so a cheaper plan never hides a gap.

Review

Four isolated agents read the code with read-only tools and report four findings, each quoting the code it means.

#1Every session is treated as expired
correctnesscritical
#2Expiry compares seconds with milliseconds
securitywarning
#3Token stays valid after logout
securitywarning
#4user may be undefined
correctnesswarning

The agent reads the exact revision under review through three tools and reports each issue by quoting code. It has no shell, cannot write, and stops after 20 steps.

Anchor

ocra finds each quote in the diff and pins it to a line. The model never picks the line.

#1Every session is treated as expiredsession.ts:42
correctnesscritical
#2Expiry compares seconds with millisecondssession.ts:42
securitywarning
#3Token stays valid after logouttoken.ts:31
securitywarning
#4user may be undefinedlogin.ts:17
correctnesswarning

The quote is matched in the changed hunks, then the whole file, then the other changed files. If nothing matches, the finding stays on the file instead of disappearing.

Filter

One finding was accepted by the team earlier, so it is dropped before a model is paid to check it.

#1Every session is treated as expiredsession.ts:42
correctnesscritical
#2Expiry compares seconds with millisecondssession.ts:42
securitywarning
#3Token stays valid after logouttoken.ts:31
securitywarning
#4user may be undefinedlogin.ts:17
correctnesswarningaccepted in .ocra/memory.json

Findings the repository's memory accepts, and those a reviewer dismissed on the pull request, are removed before any model is paid to check them. On a pull request, earlier findings are compared too: one counts as fixed only when the code it pointed at is gone.

Verify

The verifier rereads each finding next to the code: one is disproved, two are confirmed.

#1Every session is treated as expiredsession.ts:42
correctnesscriticalverified
#2Expiry compares seconds with millisecondssession.ts:42
securitywarningverified
#3Token stays valid after logouttoken.ts:31
securitywarningdisproved: token.ts:31 revokes the token on logout
#4user may be undefinedlogin.ts:17
correctnesswarningaccepted in .ocra/memory.json

A model rereads each file's findings next to its diff and the lines around them. A finding is dropped only when that code proves it wrong, with the reason kept in the report; the rest are marked confirmed or unconfirmed.

Judge

The judge sees every reviewer's findings together and merges the two that share a root cause.

#1Every session is treated as expiredsession.ts:42
correctnesscriticalverified
#2Expiry compares seconds with millisecondssession.ts:42
securitywarningmerged into #1: same root cause
#3Token stays valid after logouttoken.ts:31
securitywarningdisproved: token.ts:31 revokes the token on logout
#4user may be undefinedlogin.ts:17
correctnesswarningaccepted in .ocra/memory.json

A stronger model reads every reviewer's findings together, merges the same root cause, drops nitpicks and fixes severities. The verdict itself comes from a fixed rubric, so the same findings always get the same verdict, and only a critical finding the verifier confirmed can block.

Decisions we made on purpose

Each of these costs something. We think the trade is worth it.

01
The model never picks the line.
Models are bad at line numbers and good at quoting. So they quote, and ocra finds the lines.
02
Reviewers are told what to leave alone.
Style, speculation, missing tests and untouched code are out of scope in every prompt. You get fewer comments, not weaker ones.
03
Nothing runs with write access.
Agents can read files, read diffs and search. Editing, shell and network tools are switched off.
04
Your machine stays out of the prompt.
Local OpenCode config, installed skills and instruction files are disabled before the first request.
05
Every attempt shows its bill.
Input, output, reasoning and cached tokens, with cost, per model call and per run.
06
When a model falls over, the next one takes the task.
Give each tier a list of models. Overloads move on to the next model, a model that keeps failing is paused for a while, short rate limits are waited out, and a model out of quota is dropped for the rest of the run.

What a finding looks like

An inline comment on the pull request, with enough to decide in a few seconds whether to fix it or dismiss it. What happens next is up to the code and the reviewers.

src/auth/session.tsExample pull request

41 export function isExpired(session: Session) {

42+ return session.expiresAt < Date.now();1

43 }

github-actionsbot

Every session is treated as expiredcriticalverifiedcorrectness2

isExpired() compares expiresAt in seconds with Date.now() in milliseconds, so every session counts as expired and users are logged out right after signing in.3

Suggestion: return session.expiresAt * 1000 < Date.now();4

The finding stays open on the next push as long as its code is unchanged, even if no reviewer reports it again, so the same code keeps the same verdict.

1The line the agent quoted, found in the diff by ocra
2Severity, whether the verifier confirmed it, and which reviewer found it
3Why it is wrong, in plain words
4The smallest fix, when there is one

Your team's rules, as a plugin

The Git and GitHub adapters, the OpenCode runtime and the three reviewers that ship today are plugins too. Yours get the same small contract: register rules, reviewers, tools or listeners, and receive your own settings.

Three lifecycle hooks, run in a fixed order
Settings validated per plugin
Clashing or late registrations fail with the plugin's name

Read the plugin guide

tools/ocra-team-rules.mjs

export default {
  name: "team-rules",
  configure(ctx) {
    ctx.registerRules([{
      path: "services/**",
      rule: `${ctx.settings.team}: require idempotency keys`,
    }]);
  },
};

Where it stands

M1Local reviewshipped
CLI, selection, bundling, anchoring, the correctness reviewer, OpenCode runtime, plugins, benchmark harness.
M2More reviewersshipped
Security and performance reviewers, a review matrix, verification, a judge and a fixed verdict rubric.
M3GitHubshipped
ocra review --pr and a GitHub Action: inline comments, one summary comment, and re-reviews of only what changed since the last push that resolve fixed threads and respect dismissals.
M4Hardeningshipped
Circuit breakers per model, shared configuration over https, a review memory, and --ultra for recall.
NextMeasured qualityplanned
A baseline of precision and recall on AACR-Bench with a stronger model, then tuning prompts and stages against it.
NextDeeper reviewsplanned
Reviewers for docs and AGENTS.md, a planning phase and caller impact analysis for --ultra, model re-location of hard-to-anchor findings, and the judge reassessing disagreements.
Nextnpm releaseplanned
Publishing the @open-cr-agent packages; today ocra installs from source.

Try it on a repository you know

It runs against any Git repository on your machine, or on pull requests through the GitHub Action. Your model provider sees the change under review and the files the agents open, nothing else.

Read the quickstart

zsh

git clone https://github.com/jma49/Open-CR-Agent.git
cd Open-CR-Agent && npm install && npm run build
npm link --workspace @open-cr-agent/cli
export GEMINI_API_KEY=...
cd your-repository && ocra review

A code reviewer that reads before it comments.

What happens when you run ocra review

Select

Triage

Bundle

Matrix

Review

Anchor

Filter

Verify

Judge

Report

Decisions we made on purpose

The model never picks the line.

Reviewers are told what to leave alone.

Nothing runs with write access.

Your machine stays out of the prompt.

Every attempt shows its bill.

When a model falls over, the next one takes the task.

What a finding looks like

Your team's rules, as a plugin

Where it stands

Try it on a repository you know