[bryangruneberg]

No description provided.

[jamespember]

Continuing to be impressed with the design here @bryangruneberg. The OpenAICompatibleSummarizer base is exactly the right abstraction — three concrete providers now (OpenAI proper via the existing subclass pattern, AssemblyAI's LLM Gateway, and DeepInfra) all reduce to BASE_URL + ENV_VAR + LABEL + MODELS, and adding a fourth (Together, Groq, OpenRouter's own OpenAI-compatible endpoint, self-hosted vLLM…) is a ~15-line subclass.

Rebased on top of #18 and #19 which just merged — the two "already-applied" commits got skipped cleanly, leaving one commit and a much smaller diff (~360 lines vs the original 1,626). Local tests: 159 passed, ruff clean.

Design notes I want to call out:

• AssemblyAI reuses one key across the pipeline. ASSEMBLYAI_API_KEY authenticates both transcription and (via the LLM Gateway) summarisation. That's a genuinely nice UX win — one paste in Settings, whole pipeline works. And the shared key input in the settings screen (render_assemblyai_key_input) makes sure changing it in one place updates the other.
• Skipping temperature on OpenAICompatibleSummarizer. Claude 5 models on the AssemblyAI gateway reject any temperature value; the SDK default is fine for summaries everywhere else. Good comment explaining why.
• MAX_TOKENS = 16000. Same reasoning as fix(summaries): treat a truncated AI summary as an error #18 — reasoning models spend output tokens before writing text. The concrete example in the comment (5,420 tokens for 1,100 tokens of visible summary) is exactly the sort of context a future reader needs.
• Truncation handling inline rather than via _raise_if_truncated. Slight duplication with the base helper you added in fix(summaries): treat a truncated AI summary as an error #18, but the inline version has a nicer message (includes MAX_TOKENS in the error). Fine either way.
• DeepInfra as a second concrete example — good sanity check that the base class actually is generic; if adding DeepInfra had needed base-class changes, we'd know the abstraction was leaky.

Test coverage:

• test_summarizer_uses_gateway_and_parses — confirms the base URL, model id, prompt content and parsed action items flow correctly.
• test_summarizer_requires_key, test_deepinfra_requires_its_own_key — the DeepInfra one is particularly good because it sets ASSEMBLYAI_API_KEY to prove it doesn't fall back to the wrong provider's key.
• test_validate_accepts_deepinfra_with_assemblyai_transcription — the mix-and-match case (DeepInfra summariser + AssemblyAI transcriber) is a real user config, worth locking down.
• test_note_maker_summarises_speaker_text — verifies that summary_input (the speaker-labelled form) is what actually reaches the summariser, not the raw transcript_text. This is the invariant that makes speaker labels useful for action items.
• test_output_budget_leaves_room_for_reasoning — locks the 16k budget so a future well-meaning cleanup doesn't regress it.
• test_assemblyai_provider_and_transcriber_switching in test_textual_smoke.py — drives the settings screen through eight provider/transcriber toggles and asserts the shared key input mounts exactly once. Great regression guard for the DOM interaction that's easy to get subtly wrong.

Merging locally (same workflow scope issue as #19 — my gh token can't touch .github/workflows/ci.yml even though this PR doesn't actually touch it after rebase; hitting it safely just in case).

[jamespember]

Merged as commit fcdd9be on main (same local-squash workflow as #19). GitHub marks the branch closed rather than merged; contribution is preserved via Co-authored-by: on the squash commit. Thanks again @bryangruneberg — this is a well-shaped abstraction.

[bryangruneberg]

Nice one 🎉