[Bug]: Claude Code startup reports exit 1 with completed-looking JSON stdout

[kokisanai]

What happened?

A Discord user reported that Open Design could not start Claude Code even though standalone Claude Code works. The Open Design error shows exit 1 while stdout contains what looks like a completed Claude Code JSON result:

Could not start Claude Code: exit 1 · stdout: ":0},"inference_geo":"","iterations":[],"speed":"standard"},"modelUsage":{},"permission_denials":[],"terminal_reason":"completed","fast_mode_state":"off","uuid":"3e593d5f-c14a-4935-bc03-7c521091c145"}.

The user confirmed:

• Standalone Claude Code works: yes.
• Open Design version: 0.9.0.
• Diagnostics were requested from the user in Discord.

Discord signal

• Channel: #help
• User: Great Big Ben London☆Star
• Initial report: https://discord.com/channels/1479002485040480266/1479045739882483753/1514139611608125452
• Screenshot message: https://discord.com/channels/1479002485040480266/1479045739882483753/1514142233660624898
• Follow-up confirming standalone Claude Code works: https://discord.com/channels/1479002485040480266/1479045739882483753/1514168378284314695
• Version reply: https://discord.com/channels/1479002485040480266/1479045739882483753/1514168995723477114

Expected behavior

If Claude Code completes a request or emits a structured failure, Open Design should classify the outcome correctly and show an actionable error. It should not collapse the run into a generic "could not start" message when stdout already contains useful structured data.

Actual behavior

Open Design reports that Claude Code could not start with exit 1, while stdout appears to contain completed run metadata (terminal_reason":"completed").

Related issues

Possibly related to #4102, but that issue was closed after the reporter traced their case to a local proxy/VPN setup. This Discord report is from another user and still needs diagnostics, so tracking it separately unless logs prove it is the same root cause.

[lefarcen]

Hi @kokisanai! Thanks for writing this up with the Discord links — the exit 1 + terminal_reason:"completed" combo is exactly the useful clue here.

I checked current main before replying. There are two daemon-side follow-ups after the open-design-v0.9.0 tag that touch this seam in apps/daemon/src/server.ts: one closes Claude's stdin on a clean turn_end / usage, and another stops marking the run failed when the agent exits non-zero after a clean completion or produced artifact. So one working hypothesis is that this Discord report is hitting an already-fixed post-completion teardown path in 0.9.0, rather than the proxy/auth case from #4102 — but I don't want to call that confirmed until we see the diagnostics.

Could you ask the reporter for four concrete bits of data?

• OS
• claude --version
• the Diagnostics export / daemon log for one failing run
• whether the same prompt still reproduces on a build newer than 0.9.0

If they can also share the raw stderr tail, that will help separate "completed turn followed by teardown exit" from a real startup/auth failure. Keeping this tracked separately from #4102 for now.