Home / Guides
Guides
Setup, evaluation and engineering guides for an agentic PR reviewer you control — with a direct model API, gateway, CLI subscription, or optional local-agent companion. Released paths include exact configuration; private-beta guides state the trust boundaries and current limits.
Where evidence gets lost in AI code review
What J-Bot Review learned from Jev ranking, grouped context, uncapped investigation and verifier handoffs—and which fixes shipped.
Published Oct 4, 2026
DeepSeek V4.1 Flash Code Review: Optimization Experiments
DeepSeek V4.1 Flash experiments on tool calls, context, parallel reviews and slim compliance. Compare review speed, cost and source-assessed finding quality.
Published Oct 3, 2026
How to Evaluate AI Code Review: Accuracy, Speed, and Cost
A practical AI code review evaluation method: freeze PRs, assess false positives and missed bugs, repeat comparisons, and measure full-review time and cost.
Published Oct 3, 2026
Why AI code review is slow
Two-thirds of the reviewer’s turns were lookups. On the backend that measures it, running the tools took 1.8% of the time.
Published Sep 24, 2026
23 ways we tried to make AI code review faster
Removing whole sessions and predictable lookups worked. Fifteen of the 23 ideas didn’t make it.
Published Sep 24, 2026
Context packs: how we made AI code review faster
The code a reviewer would look up anyway, handed over before its first turn.
Published Sep 24, 2026
AI code review that reads your team’s coding rules
Guideline files used to stop at 24 KB. Now the reviewer reads them whole and starts with the sections about the files you changed.
Published Sep 24, 2026
Faster AI code review: the before-and-after numbers
Turns fell 30 to 45%, and caught issues stayed within run-to-run noise. Two production reviews took 73 and 98 seconds.
Published Sep 24, 2026
How we added Jev, a relevance model, to AI code review
One HTTPS call decides which callers the reviewer sees first. It cut estimated review cost 16% on the test PR with the most candidates, and on a small one plain ordering was faster on average.
Published Sep 24, 2026
OpenCode V2 migration: lessons from an AI code reviewer
SDK choice, plugin isolation, timeout recovery, and the usage bug that hid most of a review’s tokens.
Published Sep 18, 2026
Does a repository map help an AI code reviewer?
A 12.5 KB map of every tracked file cut no tool calls on two free models. Most reads went to files the pull request had already changed.
Published Sep 26, 2026
How to get better code reviews from free AI models
Reasoning effort took a free model from 2.5 to 6.5 of 27 known issues, and a second pass doubled recall. Prompts, context and filters didn’t.
Published Sep 26, 2026
Do fewer, prioritized rules make an AI code reviewer faster?
We cut the reviewer’s rules by 40% and put the critical ones first. Review time stayed within noise, and the main review caught fewer known bugs.
Published Sep 26, 2026
Which free AI model is best for code review?
Space Bunny, Muse Spark 1.3, DeepSeek V4.1 Flash, MiMo V2.6 Flash and Gemini 3.8 Flash, with $0 routes, measured speed and the quotas you’ll hit.
Published Sep 26, 2026
PR review with the CLI subscription you already pay for
Codex (ChatGPT Plus/Pro), Cursor, Devin, Cline, Kilo, Command Code, Grok Build, Qoder — the per-CLI provider values, secrets, and where each credential comes from.
Updated Jul 11, 2026
Codex code review in GitHub Actions
Reuse your ChatGPT Plus/Pro seat: codex login once, one CODEX_AUTH_JSON secret, and Codex reviews every PR read-only.
Updated Jul 4, 2026
Claude code review in GitHub Actions
One workflow file plus your ANTHROPIC_API_KEY: Claude reviews every PR on your own runner, with findings verified before they post.
Updated Oct 7, 2026
Claude Haiku 5.5 code review: benchmark and pricing
Five reasoning effort levels on one pull request. High found two real bugs for $0.046; max found three for $1.25. Setup and exact model ids.
Published Oct 7, 2026
Cursor code review in GitHub Actions
One crsr_… API key from your Cursor dashboard, and the Cursor CLI reviews every PR in read-only plan mode.
Updated Jul 4, 2026
Cline code review in GitHub Actions
cline auth once, one secret — works on your subscription (cline-pass) or pay-as-you-go credits.
Updated Jul 4, 2026
Kilo code review in GitHub Actions
Defaults to Kilo's free gateway model — the cheapest way to trial the whole pipeline end to end.
Updated Jul 4, 2026
Devin code review in GitHub Actions
Copy the windsurf_api_key value from devin auth login into one secret, and your Devin seat reviews every PR.
Updated Jul 4, 2026
Command Code review in GitHub Actions
One user_… access key from ~/.commandcode/auth.json, and your Command Code seat reviews every PR.
Updated Jul 4, 2026
Grok Build code review in GitHub Actions
grok login --device-auth once, one GROK_AUTH_JSON secret — your Grok account reviews every PR, with an xAI API key as the fallback.
Updated Jul 10, 2026
Kimi K3 code review in GitHub Actions
Moonshot's K3 was reviewing PRs here the day it launched — OpenCode Go or OpenRouter, with the honest quota math.
Updated Jul 17, 2026
Laguna S 2.1 code review in GitHub Actions
Verified direct Poolside setup, plus the free OpenCode and OpenRouter routes—with the exact provider, model, and secret names.
Updated Jul 22, 2026
Ling 3.0 Flash code review in GitHub Actions
Four free routes—OpenCode, Kilo, OpenRouter, and Command Code—with the exact model and secret names.
Published Jul 23, 2026
DeepSeek V4 Flash code review in GitHub Actions
Six advertised $0 routes—including Cline and OpenRouter—with exact setup, snapshot and availability caveats, and official 0731 benchmarks.
Updated Aug 1, 2026
Any OpenAI-compatible API, no CLI required
A LiteLLM proxy, self-hosted vLLM or Ollama, or a hosted API with no CLI — three inputs: base URL, key, explicit model.
Updated Jul 18, 2026
Keep agent credentials on your machine
Route Codex, Cursor, Devin, or Kilo through J-Bot’s ACP gateway to a companion you control. See where credentials, repository clones, and journal data live.
Published Jul 27, 2026
Muse Spark 1.3 code review in GitHub Actions
A $0 contributor route on OpenCode, about 20% fewer tool calls than 1.2, what Meta's contributor terms mean for private code, and how to pin review egress to an IP you control.
Updated Sep 3, 2026
Space Bunny code review in GitHub Actions
$0 on four gateways, included in OpenCode Go and Command Code plans, 6.5 of 27 known issues at high effort, and exact model ids.
Published Sep 26, 2026
MiMo V2.6 Flash code review in GitHub Actions
Free on OpenCode Zen and Cline, open weights, and slow on big pull requests. Exact model ids for seven routes.
Published Sep 26, 2026
DeepSeek V4.1 Flash code review in GitHub Actions
Free through Cline, from $0.14 per million input tokens elsewhere, and review times by pull request size.
Published Sep 26, 2026
Step 5 Preview benchmark: free, ties DeepSeek V4.1 Flash
StepFun’s free 1M-context model on OpenCode Zen, against DeepSeek, MiMo, GLM and other models. It tied the best on small PRs and failed both large ones.
Published Oct 8, 2026More resources
All repeated runs, exploratory screens and comparison limits.
DocsAction referenceEvery input, provider id, and tuning knob.
ExampleSee a real reviewA verified blocking finding, on a real diff.
CompareCodeRabbit alternativePer-seat SaaS vs $0/seat in your CI, side by side.
CompareGreptile alternativeCredits and indexing vs your own runner, side by side.
CompareQodo alternativePlatform and credits vs your own CI — and where PR-Agent fits.
CompareCubic alternativeReviewed-line metering vs no metering at all, side by side.