Weekly development log 2026‑W39: Give every capability a durable home

01

Full-history review and additions

The complete week contains 64 commits. One commit after the old 63-commit snapshot made bench installation conditional on --bench, completing the Knowledge/media/capability and optional-installation record.

All reachable history of main 95706c4b was rechecked on 2026-11-07. Counts use committer dates from Monday 00:00 through Sunday 23:59:59 KST, including merges, releases and distinct parallel-branch commits. Commit counts are not feature counts. Tests and operational measurements below are historical source records, not evaluations rerun for this refresh.

02

What changed

• Video follow-ups pass duration and aspect ratio through the orchestrator and suppress corrupted captions. Logfare joined the BYOK provider catalog, while local non-chat models for embeddings and music became selectable in capability assignments (#973, #975, #979).
• Music generation arrived through music.generate on the DGX ACE-Step 1.5 gateway, then moved to ACE-Step XL turbo. Chat can now download generated media, with corrections to source music and video duration metadata (#977, #981–#982).
• The Base/Add-on integration defined a Capability Registry, execution approval, artifact ownership, Job Runtime, and a media runtime add-on. A follow-up repaired the administrator add-on toggle's SQL typing and job polling leases, parallel progress, and retry-budget handling (#984, #986, #992).
• Hard-coded prompts, LLM parameters, thresholds, and timers moved into config and prompt sources. A security repair restored all external connectivity after an undici 8 incompatibility in safeFetch (#990, #994).
• Knowledge Space introduced a document-backed, pgvector RAG workspace as a general Base extension. Its first fixes aligned route contracts, restored Space detail views, and addressed profile saving, retries, citation-number conflicts, and reindex handoff races (#996, #998, #1001).
• The latest releases unify role and capability assignments into model slots, recover truncated simple Planner plans, and show the responding model in real time (#1003, #1005).
• Version 1.88.0 consolidates environment models — a single main tag with release tags for online, plus per-environment LiteLLM, default models, and runtime images — and switches the default database image to pgvector PG16 as the vector-extension prerequisite for the Knowledge add-on (#965, #1007).
• Versions 1.88.1–1.88.5 address long multi plans cut off before media execution, lyric handoff and note-length handling, settings overwritten after a rate limit, empty choices in non-stream responses, and local text.reason inference. Three legacy model-assignment tables were also removed (#1009, #1011, #1013, #1016, #1020).
• Version 1.89.0 brings audio, music, and video analysis into capability executors; adds Space instructions, memory, and embeddings tied to the active index model in Knowledge Space; and bounds external provider model-list requests with a prior-cache fallback (#1022–#1023, #1026).
• Version 1.90.0 adds separate one-shot installers for macOS and Linux with omk operational options. Music requests can use pasted lyrics or lyrics from the preceding answer, and a failed Planner plan now gets an explanation. This snapshot also includes the September 26 README update (#1028–#1031).