wildlifechorus/condenseit v2.6.0 on GitHub

What's Changed

OpenAI-compatible provider — CondenseIt can now summarize via any OpenAI-compatible endpoint (Ollama's /v1 compat layer, LM Studio, vLLM, llama.cpp, real OpenAI, Azure OpenAI). Configured via llm.provider: openai, llm.openai_base_url, and optional llm.openai_api_key.

Raw JSON summaries eliminated — LLM responses truncated at the token limit were previously stored as raw ```json blobs visible in the UI. Fixed by raising max_tokens/num_predict from 900 → 1400, adding partial-field recovery from truncated JSON (rescues tldr/key_takeaways even when summary is cut off), and refusing to store unparseable JSON blobs as summary text. A finish_reason=length warning is now logged when truncation is detected.
Ollama think tokens — Pass think=False so Qwen3-class and other reasoning models emit visible output instead of only internal thinking tokens.
Localized timestamps — Sidebar and admin UI timestamps are now shown in the user's local timezone.
SQLite stability — Prevent sqlite misuse errors and transient key upsert failures under concurrent digest runs.

bfacf99 fix(llm): prevent raw JSON summaries from truncated LLM responses
e9c2d0a Merge pull request #3 from wildlifechorus/feature/openai-api-support-2
aac25e7 feat(llm): add openai-compatible summarizer provider
486cebe Merge pull request #1 from melissamcdougall/fix/disable-think-tokens
26cb0c7 fix(ollama): pass think=False so Qwen3-class thinking models produce visible summaries
e41df7c fix(frontend): localize timestamps in sidebar and admin ui
0b31363 fix(pipeline): prevent sqlite misuse and transient key upsert errors