DSH | dsh-codex-review | /review on your ChatGPT subscription, one reviewer session per chat

[vitas]

Unofficial community project — independently built and maintained by a community member. Not reviewed, endorsed, or supported by DeepSeek. Please read the source and assess the risk before use.

Repository: https://github.com/vitas/dsh-codex-review ｜ npm: dsh-codex-review@0.4.0 ｜ License: MIT ｜ DSH: tested on 0.2.0-rc.2

---

/review — and the reviewer is not the model that wrote the code

One line: a slash command that hands the current change set to a reviewer subagent pinned to the model you choose — on a ChatGPT subscription it runs through the Codex CLI login instead of an API key, it keeps one session per chat so it remembers the last review, and /review reset makes it forget.

The problem it solves

Router-style plugins inject a recommendation that the main agent may ignore, and the model that wrote the change is the worst judge of it. This is the opposite: a human asks, the subagents service is called directly, and the child is pinned to a route you picked — so the review cannot accidentally be run by the same model that wrote the code, and it does not have to go through your chat route or your OpenRouter balance.

What it does

• A bare /review is a complete request. The reviewer arrives with a standing brief — the stance (report, never edit a file, never run a command that changes the repository), the order to hunt in (correctness, failure handling, contracts, security and data, concurrency and state), and an evidence bar (an unverified claim is a question, not a finding) — and works out for itself what to judge: the change set as it stands, untracked files included; the newest commit when the tree is clean; and, when the workspace is not a repository at all, the repositories inside it that actually report changes.
• Focus narrows what to judge, not what to read: /review focus on the cache invalidation.
• One reviewer session per chat. The second /review continues the same reviewer, which opens with what moved since its last review — fixed, still open, newly introduced — instead of repeating itself. The session survives a host restart.
• /review reset retires that session and stops a review that is still running, because that answer is the one nobody is waiting for. The next /review starts with no memory of the earlier ones.
• Settings card: provider and model from your subagent allowlist, the command name, reasoning effort, the brief itself, and conversation: shared (one reviewer per chat) or fresh (a one-shot reviewer per call, rendered as the command result).

Install

dsh plugin --profile web add dsh-codex-review

Restart, then type /review.

About the route. The plugin owns no LLM client and no credentials: it calls the host subagents service, so the pinned route is whatever authenticates it. The shipped default is the openai-codex route, which needs its own sign-in — on a ChatGPT subscription that is the Codex CLI login (~/.codex/auth.json), registered by the third-party dsh-codex-auth plugin, which reads that file. Nothing in the review plugin touches a credential. Point the route at any other provider you already have and it behaves the same.

Limits, honestly

• The subagent service has no way to delete a session, so /review reset writes the old reviewer off instead of erasing it. A reset followed by a host restart before the next review re-adopts the earlier session.
• A reviewer belongs to the chat that started it. Another chat gets its own reviewer, with no memory of the first — that is deliberate, and it is why running two reviews in two chats looks like two unrelated reviewers.
• A change set spanning several repositories is reviewed repository by repository, not as one diff.

What I checked

35 tests and tsc --noEmit clean. End to end: I planted an off-by-one in a scratch repository and ran a bare /review with no focus — the reviewer read the change set, reproduced the wrong result in Node, and reported it as a blocker with file:line and a concrete failure case. Also verified live on DSH 0.2.0-rc.2 that a second /review in the same chat continues the same child session rather than opening a new one.

Issues and PRs are welcome in the repository.