dsh-code-review

On-demand code review for DeepSeek Harness. /review audits the uncommitted changes of the workspace with a reviewer model and renders the report as a card in the conversation.

• Modes cr reviews the changed lines, arc reviews the shape of the change. You add your own under modes in the settings file: a role, the fields a finding must state, its severity vocabulary, and a preset of run settings.
• A review that survives a bad run The reviewer records each finding the moment it is settled, so a broken stream, a timeout, a cancel or a model that never finishes still hands you everything recorded before it stopped — marked incomplete, never lost.
• Yours alone By default the agent is not told about the review, nothing is read into its context and no code changes. You read the report and decide what to do with it. One exception, and a small one: with projectAccess on, the reviewer reads through the harness's own tools, and the harness counts that as the agent touching those files — which can inform the instruction files the agent later sees. Nothing about the review itself reaches it.

Install

A bundle for a DeepSeek Harness profile. Nothing is built and nothing is installed with it: the package is plain JavaScript with no dependencies.

1. Install it into your profile. In a session, ask for the bundle installer — plugin_manager, action: install_bundle, target: github:SeanTolstoyevski/dsh-code-review Or use web ui → plugin manager, paste the URL and install. Installing from a clone does the same thing: pass the absolute path of the directory you cloned instead.
2. Restart dsh web. A replaced package needs a fresh module generation before its JavaScript is loaded again. That first start also writes DSH_HOME/code-review/config.json, every setting and both built-in modes at their defaults, so there is a file to edit on a machine that never had one.
3. Reload the page: the client half is fetched once per page.
4. Type /review in the session you want reviewed.

Use

/review [full|session] [mode=<id>] [provider=<id>] [model=<id>] [reasoningEffort=<id>]
        [language=<code>] [gitRev=<rev>] [source=auto|git|session] [ignored=<pattern,…>]
        [ignoreDefaults=<bool>] [respectGitIgnore=<bool>] [focus message]

A name that does not resolve — mode=, reasoningEffort= — is refused before any model call, with the names that do resolve listed. The reviewer never reads the conversation.

What you get

A card in the conversation: titled by the mode, chipped with the mode's own word for the verdict, opened by the summary the reviewer recorded, then one block per finding in the fields that mode declared. Copy report takes the whole Markdown report, Copy takes one finding. Every finding quotes the line that proves it, and the quote is checked against the diff and the files the reviewer actually read — a claim that does not hold up is listed as withheld with its reason instead of being published. The card shows and copies; it has no control that sends anything.

On a session whose first message has not been sent yet, the harness draws no transcript at all — and /review never starts a turn — so the card waits above the composer instead, saying that the review is running, then handing over the report. It moves into the transcript the moment the conversation starts, and it reads the newest /review of that session, error or report alike.

Documentation

License

MIT