Review

Surfaces: CLI (kcap review) · web app (Review this PR on a session) · your agent (the review tools) · Flows for review loops

The web app is at https://<your-workspace>.kcap.ai for a cloud workspace; on the local edition the same pages are served from your machine. The desktop app opens it from Help and support → Open dashboard, and kcap local open opens the local one from a terminal

In kcap, reviewing a pull request means reading the code change together with the sessions that produced it: what they tried first, what tests actually ran, why a line ended up the way it did. The diff still matters; the sessions explain it

What you can do

Ask your agent to review with the evidence. Setup registers the review tools in every supported coding agent, so any agent, hosted or local, can read a PR’s stored record. Give it the PR when the session is not already about one:

Review acme/api#123 with the sessions behind it. Which tests actually ran, and did any failures remain?

Why did the session that made this PR remove the cache invalidation?

Start a code-review flow on this branch with Codex as the reviewer, and iterate until it is clean

The first two read the record; the third launches a separate reviewer on a daemon and runs rounds until sign-off, see Flows

Review a PR from the terminal.

kcap review https://github.com/acme/api/pull/123
kcap review 'acme/api#123'

The command checks that review context exists for the PR on your workspace, then opens Claude Code with a review tool server scoped to that PR. GitLab merge-request URLs work too. If it reports No review context found, no stored session is linked to that PR yet; it does not start a context-free review

Start a review from the browser. On a session linked to an open pull request, choose Review this PR. The dialog launches on a connected daemon:

• Agent review runs one hosted agent that reviews the PR once in an isolated worktree, with the review tools. It appears under Agents
• Review flow runs a working agent with a separate reviewer, round after round until sign-off. It appears under Flows

Checkout lists each daemon that has the repository. Harness offers Claude Code or Codex for an agent review: those are the two harnesses that can run an unattended PR review today, so the menu shows whichever of them the daemon has installed. Reviewer offers every vendor the daemon can host for a flow. Leave Model blank for the default

Know what the evidence can tell you. The review tools return what was stored: a summary of the sessions behind the PR, the files changed with their session context, a search over those transcripts, and the transcripts themselves. Good questions:

Why did this change remove the cache invalidation?

Which tests actually ran, and did any failures remain?

Was the original plan completed?

An absence of recorded test output does not prove no test ran: it may have run outside a stored session, or its record may have been dropped as too large, which the session shows as a capture-loss warning. Treat missing output as unconfirmed, not as absent. The tools respect session visibility: a private implementation session must be shared with the reviewer before it can be read

Choose and compare reviewers. For a review loop, name the reviewer’s vendor and model when you ask, or set a default with kcap config set flows.reviewer_vendor codex. A reviewer from a different vendor than the implementer tends to find different things. The Reviewers tab under Flows shows, per reviewer model, how many rounds it took to clean and how often it found something

Help me get unstuck

No review context found. The PR’s sessions are not linked to it. Sessions link to a PR through the commits they made and the PR they opened; if the work was done before kcap was installed, import the history first

The dialog says the daemon has neither Claude Code nor Codex. Install one on that machine and restart its daemon

The reviewer cannot see a session. It is private. Share it with the reviewer’s identity, or with the organisation

I want a review without a daemon. kcap review runs locally; it opens Claude Code with the review tools, so Claude Code must be on your path. From any other coding agent, ask it to review the PR with the review tools instead