PR review
kcap’s review tools connect a pull request to the recorded sessions behind it. Use them alongside the diff to understand decisions, rejected approaches, and what verification actually ran.
Start a local review
Section titled “Start a local review”kcap review https://github.com/owner/repo/pull/123kcap review 'owner/repo#123'The command checks that review context exists on your kcap server, then launches Claude Code with a PR-scoped review MCP server. Claude must be installed and available on your path, and you must be signed in to kcap.
The parser also accepts GitLab merge-request URLs. Available context depends on the source-host integration and captured sessions.
If the command reports No review context found, confirm that sessions were recorded and linked to that PR. It does not launch a context-free review on a 404.
Start a review from the web UI
Section titled “Start a review from the web UI”On a session linked to an open pull request, choose Review this PR. The dialog launches the review on a connected daemon:
- Agent review runs one hosted agent that reviews the PR once, in an isolated worktree, with the review tools below. It appears under Agents.
- Review flow runs a Claude driver with a separate unattended reviewer, round after round until sign-off. It appears under Flows.
Checkout lists each daemon that has the repository, once per repository; a linked worktree is shown as its main checkout. For an agent review, Harness offers Claude Code or Codex, whichever that daemon has installed; a daemon on a CLI older than v1.2 offers Claude Code only. For a review flow, Reviewer offers the vendors the daemon can host. Leave Model blank for the vendor default.
If the selected daemon has neither Claude Code nor Codex, the dialog says so and does not start the review. Install one of them on that machine and restart its daemon.
Ask about the implementation
Section titled “Ask about the implementation”Useful questions include:
- “Why did this change remove the cache invalidation?”
- “What alternatives were tried before this approach?”
- “Which tests actually ran, and did any failures remain?”
- “Was the original plan completed?”
- “Show me the session that changed this file.”
The six review tools are:
| Tool | Purpose |
|---|---|
get_pr_summary | Summarize sessions associated with the PR. |
list_pr_files | List changed files and related session context. |
get_file_context | Read the relevant context for a file. |
search_context | Search the linked session transcripts. |
list_sessions | List sessions linked to the PR. |
get_transcript | Read a paged session transcript. |
The tools return captured evidence. An absence of recorded test output is not proof that nobody ran a test outside the session.
Use review tools from an existing agent
Section titled “Use review tools from an existing agent”Setup registers the review MCP server for supported harnesses. Ask the agent to inspect a PR’s implementation context and supply the PR URL when the session is not already scoped to it.
This lets a hosted or local agent review the same evidence without starting a nested kcap review process. The local command above always launches Claude; the web UI’s agent review runs on Claude Code or Codex; the MCP evidence tools are usable by other supported agents.
Visibility
Section titled “Visibility”Review tools enforce session visibility and detail limits. Give a reviewer access through Visibility & sharing if they need to read a private implementation session.
A PR can include sessions from different harnesses. The reviewer does not need to use the harness that produced them.
Request an independent review loop
Section titled “Request an independent review loop”For an independent reviewer that returns findings and checks subsequent revisions, use Review flows. Those flows launch a separate participant on a connected daemon and maintain the rounds until the review concludes.
Session recap and recall searches past work beyond one PR.