Transcription accuracy, per-game screen guide, and model selection

[MyNamesEMurray]

Summary

Makes voice awareness meaningfully more accurate without adding a second speech model, teaches the cast to read each game's screen so it can react to game state rather than surface appearance, and opens up model selection. Also fixes a live issue where thinking tokens were sharing a budget sized for a one-line chat message.

Changes

Transcription accuracy (LocalVocal stays the transcriber)

• src/speech.js — new: hallucination filtering, a per-streamer correction map, word-level alignment, WER, and derivation of corrections from a known reference reading. Whisper's silence filler ("Thank you.", "please subscribe", stuck repetition) is matched whole-line only, so a real sentence containing those words survives.
• Mic check (Settings → Voice) — generates a reading script from your ghost names, channel and vocabulary, listens through the live LocalVocal path, and reports your real word error rate plus one-click fixes. It is a measurement, not training: nothing is uploaded and no model is modified. Past scores persist so a LocalVocal change can be judged by a number.
• Click-to-fix — click a 🎙 line in the deck feed, retype it, press Enter; the changed words become corrections and apply to the very next line, no save or restart.
• The cast is told which proper nouns matter, so it recovers a near-miss the map hasn't learned yet.
• Warns once when voice is configured, OBS is connected, and nothing has ever arrived for 10 minutes — a wrong path currently looks identical to the cast ignoring you.

Per-game screen guide

• New ---GAMEINFO--- tail section: the cast learns where the HUD is, what a death/menu screen looks like, and where your overlay sits, then reuses it on later streams of that game. Written on generations that already happen (no extra API calls) and refined across a session, so it self-corrects rather than baking in the first frame.
• Stored per game in game-notes.json, injected into the cached context block, and visible/editable at Settings → Stream — it's treated as truth on every message, so a wrong line has to be fixable.
• Nothing is learned under a category that isn't one game (memory.gameInfoSkip, pre-filled with 20 common Twitch categories): a guide keyed to "Just Chatting" would blend every game played beneath it.

Model selection

• Settings → Brain → Custom model ID… accepts any model the key can reach, so a new release isn't gated behind an app update. The cost hint says plainly when there are no published rates instead of implying a price it can't compute.
• Catalog: added claude-opus-5; kept claude-opus-4-8 so existing configs keep costing correctly.

Cost / correctness

• Prompt cache moved to the 1h TTL. At quietSeconds (8-minute gaps) a 5-minute entry expires between messages, so every call paid the write premium and never read.
• Current-generation models think by default and max_tokens caps thinking and visible text together — the budget for a plain chat message was 200. Added headroom, and pinned effort: low for models known to accept the parameter.

Verification

• Full CI reproduced locally from a clean npm ci --ignore-scripts: syntax check on every JS file, config.example.json validates, 88 tests pass (1 pre-existing Windows-only TTS skip), npm run smoke OK.
• New tests: test/brain.test.js (172 lines) stubs the SDK client so the outgoing request body is asserted directly — budget floor, effort sent only where supported, no sampling/thinking parameters, 1h TTL on both cache breakpoints. Plus test/speech.test.js, and additions to transcript/models/twitch/storage tests.
• End-to-end against a live headless server: click-to-fix corrected the next transcript line automatically; mic check went 8% → 0% projected and carried history between runs; the skip list refused Just Chatting case-insensitively while a real game saved.
• Browser-driven (Chromium): deck click-to-fix — Enter commits, Escape abandons, blur does not silently save a half-typed edit; Settings model picker prefill/validation/round-trip; game-guide picker loads and saves a past game's guide. No JS errors.

Risk & rollout

• Config additions (all defaulted in config.example.json, so existing configs deep-merge cleanly): speech.*, memory.gameInfoEvery, memory.gameInfoSkip. README config-reference rows added for each.
• New data files game-notes.json and miccheck.json — both added to DATA_FILES, or "Move data folder…" would silently leave them behind. A test now asserts dataFilePaths() and DATA_FILES stay in step.
• Cost: should go down. The effort: low cap and the 1h TTL both cut spend; the screen guide adds tokens but sits in the cached prefix.
• Ethics guarantees untouched: public/overlay.html is not modified in this PR — the 🤖 AI badge and the "simulated viewers" watermark are unchanged. Nothing added here posts anywhere; all new writes are local files.
• Not yet measured against a live key: the thinking/max_tokens interaction follows documented behaviour, not an observed truncation. The deck's cost pill is the one-minute check. Likewise the silent-transcript warning requires a connected OBS, which wasn't available here.
• Release-Bump: minor → v3.6.3 would tag v3.7.0.

Checklist

• npm run smoke passes locally (or CI is green)
• Follows the design language in DESIGN.md (if UI changed)
• AI-labeling / "simulated viewers" guarantees intact (if bot output or overlay changed)
• Docs updated (README / config reference) if behavior or config changed
• Targets main

🤖 Generated with Claude Code

https://claude.ai/code/session_01SK871Ay2eiEzxTPorRufLP

---

Generated by Claude Code