[nkeating-cmd]

What

Kōrero 1.45.1: dictation clean-up no longer fails when the local Ollama model has to load first. Decision D-010 (docs/DECISIONS.md), lesson L-014.

• A local Ollama clean-up is never cancelled for running late. It runs as its own task. A dictation still waits at most 30 s at the cursor (actions::CURSOR_WAIT, unchanged).
  • If the clean-up isn't back by then, the words are pasted as spoken and the pill says Pasted as spoken · clean-up continues.
  • The clean-up finishes in the background and is filed on that dictation's History entry, but only while the entry still holds the same words and has no clean-up. A correction or re-transcription made in the meantime wins.
  • Cloud and custom providers keep their own 30 s client limit and are waited for as before (can_finish_late).
  • History retry and Notes wait for the answer.
• The model loads while you speak. When a clean-up dictation starts and the Ollama model isn't loaded at all, Kōrero loads it (prewarm_cleanup_model → warm_if_cold). No transcript text is sent. The context leaves room for about 4,000 characters of dictation, so the request reuses that load.
• Ollama dictation requests may run 300 s (DICTATION_LIMIT, was 30 s). A pre-warm may run 300 s (was 120 s).
• Logs keep four earlier files (KeepSome(4), about 2.5 MB) instead of deleting the log at each 500 KB rotation.
• History stores the per-app routed prompt that was actually used.
• Version 1.45.1; README.

Why

On 23 Oct, two clean-ups on the main laptop failed at 30.1 s (13:47 and 14:17). Ollama's log shows each request gave up while the model was still loading. Kōrero's log shows loaded=None, then "The local model didn't finish its answer within 30 s".

Probe (23 Oct, main laptop)	Result
Cold load of gemma4:12b, about 3 GB RAM free	108 s (14–19 s on 19–20 Oct)
A load is in progress and the request waiting on it disconnects	Ollama aborts the load, even with another request still waiting, then starts again from zero
A request arrives while a pre-warm is loading at the same context	It joins the load: one load, answered 17.6 s after sending
Transcribing a 30 s clip while a cold load runs (two runs each)	2.55–2.86 s alone, 2.77–3.17 s with the load (+0.2–0.3 s)

No fixed limit at the cursor covers every load, and each give-up also threw away the half-loaded model.

Independent review

A separate agent reviewed f5b87a7, briefed to assume it was wrong. It found no Critical, 3 Major and 8 Minor issues. All were fixed in b7000f9 or recorded as accepted costs in D-010.

• M1: the late path and its pill text also applied to providers that can never finish late. Now Ollama only, and the pill text is true when it shows.
• M2: a pre-warm sized for the prompt alone made long dictations load the model twice. Now it warms only a cold model, with room for the dictation.
• M3: the overlap with transcription was unmeasured. It is now measured (above) and accepted.
• Minor:
  • the pill mentions History only when there is an entry;
  • re-transcribed entries are guarded;
  • whitespace counts as empty, as on the History page;
  • the routed prompt is stored;
  • a misplaced doc comment is fixed;
  • the docs match the code.
• Accepted, recorded in D-010:
  • late clean-ups are not capped;
  • a cancelled recording leaves the model loaded for its keep-alive;
  • the new pill text is English only, like the other pill notices.

Evidence (main laptop, Windows)

• cargo check --locked: pass. The two warnings are already on main.
• cargo test --locked --lib: 292 passed, 0 failed, 9 ignored (live tests). New tests:
  • cleanup_wait_tests (7), including giving_up_at_the_cursor_does_not_cancel_the_clean_up, only_ollama_clean_up_may_finish_late and the_prewarm_leaves_room_for_a_dictation;
  • late_clean_up_fills_only_an_empty_entry;
  • dictation_and_warm_outlive_a_slow_cold_load.
• Each new check seen failing. Four deliberate breaks, made one at a time in the test run, were each caught by their own test (4 failed, 288 passed). The tree was restored afterwards. The breaks were:
  • letting any provider finish late;
  • dropping the words guard;
  • aborting the task when the cursor wait runs out (the original bug);
  • a pre-warm with no headroom.
• Live live_cold_dictation_survives_the_cursor_wait (gemma4:12b, starting cold): passed twice, answered 35.0 s and 16.2 s after sending. Each time Ollama loaded once and aborted nothing; the old 30 s limit fails the first.
• Frontend: bun test unit-tests 104 pass; tsc 0 errors; ESLint 0 errors (17 warnings already on main); Vite build OK.
• Safe-publish:
  • gitleaks, all refs: 932 commits, no leaks;
  • personal-data scan of 597 added lines and the commit messages: 0 hits;
  • both commits use the noreply address;
  • no dependency change (cargo deny advisories identical to main);
  • line endings: numstat identical with and without --ignore-cr-at-eol.
• Installed on the main laptop as an unsigned local build of this branch, after a backup. Data files were unchanged by hash, and there were no errors since start.

Bet: fix (not a product bet, so the docs/PRODUCT.md gate doesn't apply).

Not in this PR

• The custom provider (LM Studio and similar) has neither the late path nor a pre-warm.
• Slow loads on the main laptop come from memory pressure on the machine: about 3 GB free of 32 GB, and Ollama loads the whole model into RAM because it doesn't memory-map it under CUDA on Windows. That is the machine, not the app.
• Releasing 1.45.1 (signed build, tag, GitHub release) is a separate ship step after merge.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EXpf5AYj4BYHJqpHxqLNp2