Navin review agent: /inspect, evidence and a numbered remediation plan

Navin's review agent is not a linter with a chat: it is an expert reviewer that reads your diff or your whole project, runs lint, typecheck and tests to prove its findings, then delivers an HTML report with a numbered remediation plan. You pick #1, the agent implements. Nothing is modified without your go.

How to start a review

Three entry points, one engine:

• Review mode in the composer + free text (automatically prefixed as /inspect).
• /inspect [path|diff|scope] to target a folder, a file or the current diff.
• Actions → Code review for the one-click audit.

The default scope is smart: git status, staged and unstaged diff, recent changes. On a large repository, the agent leans on the metagraph to prioritize hot paths.

Eight analysis layers, not a checklist

The rule that changes everything: evidence

Every finding must include the severity (Critical to Info), the file:line, the impact, a concrete fix and a real example: vulnerable code excerpt, failing test output, or a small PoC. No generic blurb. The agent is allowed to run ruff, eslint, tsc, mypy or pytest to verify what it claims.

The close: HTML report + numbered choices

1. The review-report-[date].html report is written and opens automatically in File Preview: executive summary with severity counters, finding cards, numbered remediation plan (S/M/L effort, risk if delayed, concrete first step).
2. In chat: a short summary and one question - "Which number do we start with?".
3. The review is read-only by default. You answer Start with #1 and the agent switches to build mode (/forge) to implement that precise item.
4. On an open GitHub PR, pr_comments can post the findings as review comments (through gh).

A verdict like a real reviewer

The review ends with a verdict: Approve or Request changes, plus the numbered plan. It is the same contract as a serious human review - available on demand, on every diff.

Review + autonomous board: the full loop

Combined with board autonomy, the loop becomes: the agent codes on an isolated branch, opens a PR, you run /inspect on it, pick the remediations, and the agent handles them as tracked tasks. Review stops being a bottleneck and becomes a pipeline stage.

FAQ

Does the review modify my code?

No, read-only by default. It only touches code if you ask for auto-fix or pick a plan number.

Which model is used?

The review routing role (Settings → Models → Task routing): you choose which model reviews, independently from the one that codes.

Does it work on a diff or the whole project?

Both: current diff by default, or any named path.

Compared with a SaaS review bot?

Everything runs locally, on your machine, with your keys. The report is a self-contained HTML file inside your project, not a page on someone else's cloud.

Download Navin · Features · Documentation

Recommended next reads

• The deep security audit → AppSec security agent
• Proven root cause → debug agent
• The full module cycle → Code module
• Tasks that execute themselves → autonomous board

Conclusion

A review that proves what it claims, delivers a readable report and waits for your green light before touching code: that is the difference between an AI gadget and a reviewer you let into your pipeline.