Root AGENTS.md is now the single source of truth for coding agents;
CLAUDE.md, GEMINI.md and docs/CLAUDE.md are pointers. The product spec
moves from docs/CLAUDE.md to PRODUCT.md, DESIGN.md records the shipped
visual system, and docs/PERMISSIONS.md defines agent permission tiers
rendered to .claude/, .codex/ and .gemini/. Adds docs/specs/ and
docs/plans/, logs issues noticed (not fixed) in WorkBreakdown.md.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Measured with the ort-web-under-Bun harness: no reliable speed difference on
clean, language-matched audio. On out-of-distribution audio (wrong/undetected
language), the multilingual model's extra forced task token measurably
changes how long its repetition-loop hallucination runs before EOS (434 vs 67
tokens on the da_festuge fixture), which is where the reported slowness comes
from. Both models resolve to identical, merged, KV-cache-enabled decoders at
the same dtype — no catalog/config fix applies here. Documents the mechanism
and flags a future no_repeat_ngram_size/repetition_penalty mitigation as the
actionable follow-up if this recurs.
startLive()'s browser-engine branch flipped liveStatus to "listening" (the
red live-dot, "Listening..." hint) the instant the mic was granted, even if
the Whisper pipeline was still loading — the user saw an active-looking
session while nothing was being transcribed yet. transcribeLiveSegment()
already correctly awaits the model-ready gate before posting a segment, so
this was purely a misleading-status bug.
Adds a "loading-model" LiveStatus between "starting" and "listening";
startLive() now waits on the same browserModelReady() gate the transcribe
path uses before claiming "listening", with race-safe cleanup if the load
fails while waiting. Server-engine live is unaffected.
Adds Decisions.md D-034; updates tests/live.md's manual verification steps.
Transformers.js's progress_callback always fires status:"download" regardless
of whether the file is actually served from the Cache API — there's no
cacheHit signal for the worker to forward. Since App.tsx already tracks which
models are downloaded (for ModelPicker's badges), it now checks that before
starting a load and threads a loadedFromCache flag into ModelLoadProgress, so
the heading/hint reflect reality for the common already-cached case.
Adds Decisions.md D-033.
