Self-hosted features broken on fresh Debian/.deb install (pikaOS): voice server disabled, Codex not routing, Ollama not detected

[Al629176]

Summary

On a fresh pikaOS (Debian-based) install via .deb, all self-hosted features are non-functional: Codex LLM does not respond, Whisper stops working after restart, Piper stopped working after restart, and the agent was not working. User migrated from CachyOS to pikaOS on a pre-formatted system.

User report

"I'm trying to use my Codex token, everything is set up, and the LLM just doesn't answer me. Whisper is not working when I restarted my PC. Piper worked after install but stopped after restart. The agent didn't work before restart either. I migrated from CachyOS to pikaOS (Debian-based distro) and installed the .deb version on a fresh system."

Log analysis (message.txt — provided by user)

Three root causes identified from the startup logs:

1. Voice server auto-start disabled — Whisper and Piper broken after restart

[voice_server] auto-start disabled in config, skipping embedded voice server
[dictation_listener] dictation disabled in config, skipping hotkey listener

The voice server is not configured to auto-start. Whisper and Piper work immediately after install (before restart) because the session config is live, but on restart the voice server is skipped. On Debian/.deb this config does not persist correctly across restarts.

2. Codex token configured but LLM routes to OpenAI instead

[providers][chat-factory] role=chat slug=openai model=gpt-5.6-terra endpoint_host=https://api.openai.com

Despite the user having a Codex token configured, chat is routing to api.openai.com with gpt-5.6-terra — not the Codex subscription endpoint. Same root cause as #5353, may be a regression or not yet effective on this install.

3. Ollama not running — local AI bootstrap degraded

[local_ai] bootstrap degraded: external runtime connectivity check failed:
OpenHuman no longer starts or installs Ollama automatically.
Start your inference runtime yourself and make sure it is reachable at http://localhost:11434.

Ollama is not installed on the fresh pikaOS system. The app logs this as a warning but does not surface a clear user-facing error. The user likely doesn't know they need to install Ollama separately.

Additional observations

• CEF cache lock on first launch (another instance was running) — resolved by user with pkill openhuman
• Config file was world-readable (mode 664) — auto-fixed to 600 by app
• Subconscious disabled: heartbeat.enabled = false
• App version: 0.63.9 (up to date)
• Keyring using file backend (dev keychain) — may indicate install path issue

Steps to reproduce

1. Fresh install on pikaOS (Debian-based) via .deb
2. Configure Codex token in settings
3. Configure Whisper and Piper for voice
4. Restart the system
5. Launch OpenHuman — voice server skipped, Codex routes to OpenAI, Ollama not found

Solution

1. Voice server persistence: Ensure voice server auto-start config persists across restarts on Debian/.deb installs. Investigate why voice_server.auto_start is not being written or read correctly after restart.
2. Codex routing: Verify Codex CLI OAuth import routes chat through regular OpenAI API instead of Codex subscription #5353 fix is effective on v0.63.9 Debian builds. If regression, re-investigate Codex credential routing.
3. Ollama onboarding: Surface a clear setup prompt when Ollama is not detected on first launch, rather than a silent warning in logs.

Affected areas: Voice server config persistence, Codex inference routing, Ollama onboarding, Debian/.deb packaging

Acceptance criteria

• Voice server (Whisper/Piper) auto-start config persists across restarts on Debian/.deb.
• Codex token correctly routes to Codex subscription endpoint, not api.openai.com.
• When Ollama is not installed, user sees an actionable in-app prompt, not a silent log warning.
• All self-hosted features work correctly after a restart on pikaOS/.deb.

Related

• openhuman Codex CLI OAuth import routes chat through regular OpenAI API instead of Codex subscription #5353 (Codex CLI routing — may be same regression)
• openhuman Memory embedder calls deprecated Ollama endpoint — broken since Ollama v0.32 #5354 (Ollama deprecated endpoint)
• openhuman Discord channel in restart loop with 100% CPU on Linux (built from source) #5350 (Linux restart loop)