Claude Code 2.1.232 changes more than the version number. Forked subagents now inherit the full conversation and prompt cache by default, interactive non-teammate spawns move to the background, and Windows permission bypasses were closed.
That means a successful launch is not enough. Context inheritance, session routing, permissions, repository trust, MCP probes, and Remote Control recovery are separate operating boundaries.
1. Verify inherited context
Spawn one child agent with a harmless task. Confirm that it sees the intended parent context and does not inherit unrelated instructions.
2. Verify background completion
Start one interactive non-teammate child. Check that it moves to the background and that its final status returns to the expected parent session.
3. Verify session-name routing
Resume or message a session by name. Confirm the message reaches the intended session rather than another similarly named run.
4. Test one allowed and one denied write on Windows
Use a disposable fixture. Perform one harmless write inside the approved path, then attempt one write that should be denied. Both outcomes matter.
5. Check nested repository trust
Open a nested repository and verify that the trust boundary is explicit. Do not assume the parent repository's approval automatically covers every nested path.
6. Force a recovery path
Trigger one bounded MCP probe failure and one Remote Control resume. Confirm that tools, session state, and recovery messages remain coherent.
The acceptance rule
Use one symptom, one fixture, and one changed layer. A broad end-to-end run can pass while a permission or routing regression remains hidden.
The full source-linked checklist, including the official release evidence and rollback boundaries, is available here:
https://tgwise.com/intelligence/claude-code-2-1-232-forked-subagents-windows-permission-fixes/
Top comments (0)