Worktree-bound Codex starts fail with “app-server process is not running” while unbound starts succeed

[mplibunao]

Summary

Codex-backed Agent Mode sessions fail during provider startup whenever they are bound to an app-managed worktree, while the same Codex model starts successfully without a worktree binding and Claude Code starts successfully in the same worktree.

The session is created and queued, then terminates with:

Codex native start failed: Codex app-server process is not running.

Reproduction

Starting state:

• RepoPrompt CE with a multi-root workspace
• An app-managed Git worktree created through manage_worktree
• Codex CLI installed through npm and launched with a mise-managed Node runtime

Run the following in the same application session:

Start type	Result
Codex, no worktree binding	Starts and completes
Claude Code, worktree-bound	Starts and completes
Codex, bound through worktree_id	Fails during provider startup
Codex, bound through worktree: "@branch:<name>"	Fails during provider startup

The failure reproduced on every worktree-bound Codex attempt.

Expected behavior

A Codex session should start with the worktree as its thread and turn cwd, regardless of whether the app-server process itself is launched from the logical checkout or another stable directory.

If the launcher exits, the resulting error should include a bounded stderr excerpt and termination status rather than only reporting that the process is not running.

Actual behavior

RepoPrompt projects the same resolved worktree path into both:

1. the Codex app-server process working directory; and
2. the Codex thread, turn, and sandbox working directory.

This makes provider bootstrap sensitive to repository-local toolchain configuration before thread/start occurs.

At failure time, another healthy Codex app-server owned by an unbound session remained alive. The worktree-bound attempt did not leave an additional app-server process running.

Source findings

CodexNativeSessionController.startOrResume calls:

await client.updateWorkingDirectory(workspacePath)

and then uses the same workspacePath for thread/start.cwd and thread/resume.cwd. The client passes that directory directly to ProcessLauncher.spawn.

CodexAppServerClient drains stderr but retains it only through optional debug printing. Process termination/reaping results are also discarded. Several transport-loss paths consequently collapse into .processNotRunning.

The precise historical launcher trigger is not recoverable. A cwd-sensitive wrapper, Node/mise shim, or trust/configuration failure is plausible, but not proven.

The relevant Codex startup files are unchanged between v1.0.28 and v1.0.29, and the same coupling remains in v1.0.29.

Regression context

Worktree-bound Codex agents worked in the same workspace around July 9–13, 2026. The failure was first observed on July 15.

Restarting RepoPrompt CE and removing stale debug CLI daemon processes did not change the result.

Related but distinct issues

• Agent Mode worktree sub-agent starts can fail inherited routing or park without actionable startup failure #383 covered routed worktree inheritance and non-actionable startup waits.
• Codex resolution misses nvm-installed CLI (non-interactive shell skips .zshrc) #471 covers Codex/Node executable resolution from shell configuration.
• Codex child agents can start without RepoPromptCE MCP tools when routing/bootstrap failures are swallowed #514 covers Codex children starting without RepoPrompt MCP tools.

This issue occurs after the worktree binding is accepted and specifically affects Codex app-server provider startup under the worktree cwd.

Acceptance criteria

• A worktree-bound Codex session starts successfully while retaining the worktree for thread, turn, and sandbox execution.
• App-server launch cwd and Codex execution cwd are modeled independently.
• Both paths come from the same owning logical-root/worktree binding in multi-root workspaces.
• A controller is replaced when either launch cwd or execution cwd changes.
• Early launcher exit reports executable, launch cwd, termination status or signal, and a bounded stderr tail.
• Regression coverage exercises unbound, existing-worktree, created-worktree, inherited-worktree, and secondary-root projections.
• Explicit shutdown and lifecycle teardown are not mislabeled as observed process exits.

Environment

• RepoPrompt CE debug build; source rechecked against v1.0.29
• macOS, Apple Silicon
• Codex CLI installed npm-global
• Node runtime managed through mise

[baron]

@mplibunao appreciate your amazing work on this and your other contributions. If you open a PR we will definitely fast track the review (without making compromises which are obviously not needed with your work). Thanks for being awesome!