Gateway crashes on non-voice bootstrap when OPENAI_API_KEY is in the environment

[kenadams1990]

Summary

inkbox-claude bootstrap without --voice-ai produces a gateway that crashes on startup on any machine that has OPENAI_API_KEY exported in the environment (very common — LiteLLM proxies, other AI tooling, shell profiles). The daemon exits and bootstrap/start report Claude Code gateway did not become ready.

Environment

• inkbox-claude installed via the curl | bash installer, app at ~/.inkbox-claude/app (venv-isolated), Python 3.13, macOS.
• OPENAI_API_KEY present in the environment the detached daemon inherits (in our case exported by an unrelated LiteLLM rotation proxy in a shell profile). No voice/realtime config in ~/.inkbox-claude/.env.

Steps to reproduce

1. Have OPENAI_API_KEY exported in the environment.
2. inkbox-claude bootstrap --identity <handle> --start-gateway without --voice-ai.
3. Gateway fails to become ready; check ~/.inkbox-claude/gateway.log.

Expected

With no voice flags and no realtime config, the gateway should start with a non-voice/inkbox_tts_stt stack and ignore an unrelated ambient OPENAI_API_KEY.

Actual

• bootstrap returns {"status":"error","error":"Claude Code gateway did not become ready"}.
• gateway.log shows: RuntimeError: INKBOX_VOICE_STACK=openai_realtime requires INKBOX_REALTIME_API_KEY raised in gateway.py run().
• inkbox-claude doctor confusingly reports phone voice stack openai_realtime and OpenAI Realtime key set even though nothing voice-related is configured in .env.

Root cause (from reading the app source)

• bootstrap only writes INKBOX_REALTIME_ENABLED=false inside _configure_voice(), which is gated behind --voice-ai (bootstrap.py, ~L137 / ~L211). Without --voice-ai, that line is never written.
• With the flag absent, resolve_voice_stack() in config.py falls through to a heuristic: if any OPENAI_API_KEY is in the environment, assume voice_stack=openai_realtime.
• _read_realtime_config() then reads INKBOX_REALTIME_API_KEY or OPENAI_API_KEY, so the ambient key satisfies the key check — but realtime.enabled is still False (needs INKBOX_REALTIME_ENABLED=true), so the gateway's startup invariant raises and the daemon exits. doctor sees the same ambient key, hence its misleading output.

Workaround (config only, survives plugin updates)

printf '%s\n' 'INKBOX_REALTIME_ENABLED=false' >> ~/.inkbox-claude/.env
inkbox-claude restart

resolve_voice_stack then returns inkbox_tts_stt before it ever consults OPENAI_API_KEY.

Suggested fix

Don't infer openai_realtime from a bare ambient OPENAI_API_KEY. Gate the realtime stack on an explicit opt-in (INKBOX_REALTIME_ENABLED=true or an Inkbox-specific INKBOX_REALTIME_API_KEY), and/or write INKBOX_REALTIME_ENABLED=false during a non-voice bootstrap so the default install is deterministic regardless of unrelated env vars.