[khushalsonawat]

Description

The credential check asked for ELEVENLABS_API_KEY for LiveKit's ElevenLabs plugin, but the plugin reads ELEVEN_API_KEY, so an honest voice agent's key never reached it and the run failed. The check now asks for the name the plugin reads, and a constructor that passes its own api_key= is judged by what it passes.

Paired with future-agi/future-agi#3959, whose preflight probe accepts either name. Merge order: future-agi#3959 first, or both together.

Type of Change

• Bug fix (non-breaking change which fixes an issue)

Changes

• The plugin row asks for ELEVEN_API_KEY, and both ELEVEN_ and ELEVENLABS_ names are labelled ElevenLabs text-to-speech.
• api_key= on a LiveKit plugin constructor (_plugin_key_names, read with ast):
  • none: the plugin's own name, as today;
  • a non-empty literal: nothing;
  • os.environ["NAME"]: the code check already requires NAME;
  • os.getenv("NAME", d) or os.environ.get(...): NAME is required when d is absent, None, empty or a placeholder, and nothing when d is a real literal;
  • anything the check can't read (a variable, a settings object, **opts, a bare getenv, a lowercase name): the plugin's own name, and for ElevenLabs both names. That over-asks, never under-asks, for the plugin's names.
• A name an SDK constructor requires keeps required even where another line gives it a default.
• A reserved name passed as the key (getenv("ALK_X"), getenv("HOST")): the customer can never supply it, so the check asks for the plugin's own name instead of crashing or calling the agent ready. Cartesia, Deepgram TTS and OpenAI realtime then read their own variable. OpenAI LLM, Deepgram STT and ElevenLabs refuse an explicit None, so for them the run fails at the plugin either way; asking at least says so before the build.
• oss/simulation-acceptance/.env.example lists ELEVEN_API_KEY.

Testing

• tests/test_harness_credentials.py: 112 passed (26 new); tests/harness/test_runtime_repair.py: 213 passed. Python 3.12.3.
• Planted bugs: 13, each caught by its test.
• Reviews: design in four rounds (one reviewer each; round 4 not clear, its fix taken without a fifth round on the Coordinator's call). Change review: one reviewer, cleared. Its Medium (the reserved-name crash) and test gaps were fixed in 62f81e2d. Fix-only checks since:
  • 62f81e2d, two reviewers: both clear. B's Medium (a reserved key made the agent "ready") was answered in 5ad07d83, whose one-reviewer check was clear but showed four plugins fall back to their own variable.
  • f6e9660f asks for the plugin's own name. Two reviewers: both clear, 0 High, 0 Medium. Results are in the comments.

Known limits

• An unreadable api_key= that reads a name other than the plugin's own (STT(api_key=getenv("MY_DG").strip()), **{"api_key": ...}, from os import environ as e) asks for the plugin's name, not the name the code reads. This was true before this change too.
• A real literal default, getenv("DG_KEY", "dummy"), counts as a key, so nothing is asked. Behind a reserved name, getenv("ALK_X", "sk-..."), it doesn't: the plugin's own name is asked for, though the plugin would use the literal.
• ElevenLabs is asked for both names when its key can't be read, though the installed plugin reads only ELEVEN_API_KEY.
• os.environ["ALK_X"] as the key still asks for nothing: the agent fails at that lookup, so no name would help.
• Over-asking by design: Cartesia and Deepgram's TTS fall back to their own variable on an empty key, so api_key=getenv("CART") is asked for CART even with CARTESIA_API_KEY supplied.
• Labels: a plugin-required name like MY_DEEPGRAM shows as agent runtime configuration, and ElevenLabs' STT is labelled text-to-speech.
• A customer who saved the key as ELEVENLABS_API_KEY adds it again as ELEVEN_API_KEY.
• Same file: fix(harness): review follow-ups for #269 and #276 (TH-8952, TH-9490) #297 edits credentials.py elsewhere; no function overlaps.

Related Issues

Refs: TH-9519

[khushalsonawat]

Fix-only checks since the change review:

• 62f81e2 (skip a reserved name passed as the key), two reviewers:
  • A: clear, 0 High, 0 Medium, 2 Low.
  • B: clear, 0 High, 1 Medium, 1 Low. B's Medium: the agent was then marked ready with nothing asked.
• 5ad07d8 (pin "nothing asked" in the test), one reviewer: clear, 0 High, 1 Medium, 1 Low. The Medium: the comment was wrong for four constructors. Cartesia STT and TTS, Deepgram TTS and OpenAI realtime fall back to their own variable, so skipping hid a name that would make the run work.
• f6e9660 (ask for the plugin's own name), two reviewers:
  • A: clear, 0 High, 0 Medium, 2 Low. A checked 11 constructors against the installed 1.8.3 plugins.
  • B: clear, 0 High, 0 Medium, 2 Low. B ran 80 constructor cases.
  • Planting back the old skip fails 4 tests. Narrowing the check to ALK_ names fails the HOST case.
  • Lows kept as known limits: a usable literal default behind a reserved name is now asked for; ElevenLabs is asked for two names when its key can't be read; the guard after record() is now an unused backstop.