What happened?
On Claude Code 2.1.199 and later, clicking Approve Plan in the PlanBridge browser UI no longer hands control back to the agent. The review round-trips correctly — the browser window shows the approval confirmation and closes — but the Claude Code terminal session stays blocked on its native plan-approval prompt ("Claude has written up a plan and is ready to execute. Would you like to proceed?") and implementation never starts until the user manually answers that prompt in the terminal.
The cause is an upstream Claude Code regression, filed as anthropics/claude-code#74256: as of v2.1.199, Claude Code runs the PermissionRequest hook for ExitPlanMode and reads its response, but discards the allow decision and shows its own approval prompt anyway. The changes-requested path is unaffected — deny decisions with feedback still round-trip correctly, so the annotate → refine loop keeps working. Only the approve handoff is broken.
contextbridge version
0.9.2-alpha.1 (also reproduces on the released 0.9.2 plugin)
Operating system
macOS
Steps to reproduce
- Install the
planbridge plugin in Claude Code 2.1.199 or later (verified on 2.1.199 and 2.1.201)
- Start
claude in plan mode and ask for any plan
- When Claude calls
ExitPlanMode, the PlanBridge review UI opens in the browser
- Click Approve Plan without adding comments
What did you expect to happen?
The behavior on Claude Code <= 2.1.198: the hook's allow decision dismisses Claude Code's plan-approval prompt, the session switches to accept-edits mode, and implementation starts immediately.
What actually happened?
The terminal stays blocked on Claude Code's own plan-approval chooser. Implementation only starts after the user additionally selects "Yes" in the terminal, which reads as PlanBridge hanging after approval.
Relevant logs or stack trace
# contextbridge emits the documented allow envelope on stdout; Claude Code >=2.1.199 reads and discards it:
{"hookSpecificOutput":{"hookEventName":"PermissionRequest","decision":{"behavior":"allow","updatedPermissions":[{"type":"setMode","mode":"acceptEdits","destination":"session"}]}}}Anything else
Version bracketing (same plugin build, driven against each Claude Code TUI):
| Claude Code |
Approve behavior |
| 2.1.198 |
works — implementation starts immediately |
| 2.1.199 |
broken — terminal blocks on native prompt |
| 2.1.201 |
broken — terminal blocks on native prompt |
The regression reproduces with a minimal one-line auto-approve hook and no PlanBridge involvement, so it is not caused by anything in this codebase — full standalone reproduction, variant testing (bare allow, allow + setMode, PreToolUse allow), and the working-deny-path evidence are documented in anthropics/claude-code#74256. Nothing in the Claude Code 2.1.199–2.1.201 changelog announces a behavior change for hook-approved plan exits, and the current hooks documentation and @anthropic-ai/claude-agent-sdk 0.3.201 still describe the envelope above as the supported auto-approval mechanism.
What happened?
On Claude Code 2.1.199 and later, clicking Approve Plan in the PlanBridge browser UI no longer hands control back to the agent. The review round-trips correctly — the browser window shows the approval confirmation and closes — but the Claude Code terminal session stays blocked on its native plan-approval prompt ("Claude has written up a plan and is ready to execute. Would you like to proceed?") and implementation never starts until the user manually answers that prompt in the terminal.
The cause is an upstream Claude Code regression, filed as anthropics/claude-code#74256: as of v2.1.199, Claude Code runs the
PermissionRequesthook forExitPlanModeand reads its response, but discards theallowdecision and shows its own approval prompt anyway. The changes-requested path is unaffected —denydecisions with feedback still round-trip correctly, so the annotate → refine loop keeps working. Only the approve handoff is broken.contextbridge version
0.9.2-alpha.1 (also reproduces on the released 0.9.2 plugin)
Operating system
macOS
Steps to reproduce
planbridgeplugin in Claude Code 2.1.199 or later (verified on 2.1.199 and 2.1.201)claudein plan mode and ask for any planExitPlanMode, the PlanBridge review UI opens in the browserWhat did you expect to happen?
The behavior on Claude Code <= 2.1.198: the hook's allow decision dismisses Claude Code's plan-approval prompt, the session switches to accept-edits mode, and implementation starts immediately.
What actually happened?
The terminal stays blocked on Claude Code's own plan-approval chooser. Implementation only starts after the user additionally selects "Yes" in the terminal, which reads as PlanBridge hanging after approval.
Relevant logs or stack trace
Anything else
Version bracketing (same plugin build, driven against each Claude Code TUI):
The regression reproduces with a minimal one-line auto-approve hook and no PlanBridge involvement, so it is not caused by anything in this codebase — full standalone reproduction, variant testing (bare allow, allow +
setMode,PreToolUseallow), and the working-deny-path evidence are documented in anthropics/claude-code#74256. Nothing in the Claude Code 2.1.199–2.1.201 changelog announces a behavior change for hook-approved plan exits, and the current hooks documentation and@anthropic-ai/claude-agent-sdk0.3.201 still describe the envelope above as the supported auto-approval mechanism.