/code-review in the desktop app runs inline in the authoring session (no context isolation), and the docs don't say so

Summary

In the Claude desktop app, /code-review runs inside the current conversation instead of as a background forked subagent. The agent that reviews the PR is the same agent that wrote it, with the full authoring context. As far as I can tell this is intentional, not a bug, but the docs don't mention it. Anyone relying on /code-review for an independent review in the desktop app gets a self-review without knowing it.

What the docs say

From https://code.claude.com/docs/en/code-review:

The review runs as a background subagent with its own context window, so it doesn't fill your conversation.

The review runs in the background by default; before v2.1.218, it ran inside your conversation. It runs in the foreground instead in cases like these: [re-run while a review is in progress / -p or Agent SDK / CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1]

In a host application that requests the findings list, such as the desktop app, Claude reports the review's findings through the ReportFindings tool instead.

Nothing on that page says the desktop app skips the background subagent, or that the review there loses context isolation.

What I observed

• Environment: Claude desktop app 2.110.1, bundled Claude Code 2.1.260 to 2.1.271, macOS 27.0.
• 8 of 8 /code-review runs started from the desktop app (session entrypoint claude-desktop) ran inline:
  • The skill prompt was injected as a meta user message into the main transcript (isSidechain: false).
  • The same agent then ran git diff and called ReportFindings.
  • No subagent transcript was created for the review.
• Every run's banner read minimal prompt → single careful diff pass → ≤15 findings, and no finder/verifier agents were dispatched. By contrast, a terminal CLI report (#86724) describes /code-review forking and fanning out 8 finder agents.
• In one run, the review's first line was literally "Reviewing my own PR — …". The reviewer knew it was the author.
• None of the documented foreground triggers applied: CLAUDE_CODE_DISABLE_BACKGROUND_TASKS not set, not -p, no other review in progress.
• The desktop app sets CLAUDE_CODE_REPORT_FINDINGS=1 for the CLI it spawns. The variable shows up in the spawned process environment, and CLAUDE_CODE_REPORT_FINDINGS="1" appears in app.asar.

Likely reason (my reading, not confirmed)

In ReportFindings mode, the prompt asks the agent to call ReportFindings again later in the same session, marking each finding fixed / skipped / no_change_needed, so the findings card stays up to date. A forked subagent is gone by then, so the review has to run in the long-lived session. That is a reasonable tradeoff (interactive findings card vs. reviewer independence), but it is invisible to users.

Requests

1. Docs: state that in the desktop app (or whenever ReportFindings mode is active), /code-review runs in the current conversation and does not isolate context.
2. Option: let desktop users choose an isolated review. For example, run the review in a forked subagent and have the main session only relay and update the ReportFindings card, or add a setting/flag that trades the live card for isolation.

Related

• #95136: background /code-review subagent (CLI, 2.1.274)
• #86724: terminal /code-review orchestrator fanning out finder agents
• #60720, #49559, #17283: context: fork skills running inline or behaving differently in desktop