Reply only when addressed, plus a beta release line

[MyNamesEMurray]

Summary

Cuts the dominant cost driver — the cast was treating almost everything the streamer said as a question aimed at it — and adds a beta release line so changes like this can be tested on a real stream without merging or compiling anything.

Changes

Voice replies only when the cast is actually addressed

A 2.5h stream on v3.7.0 cost $4 on Opus, ~5x the README's estimate. The per-message price wasn't the story; the message count was. onSpeech treated any speech within 120s of a cast message as an answer to it — but a streamer talks almost continuously, so an extra generation fired after nearly every cast message.

• A voice reply now needs the speech to be addressed: a ghost named out loud, or a prompt answer to a question one just asked.
• Nothing is lost when it isn't — the transcript window is read by the next scheduled generation either way, so undirected talk still reaches the cast, it just no longer buys its own API call.
• The question rule is on a short fuse (voiceAnswerWindowSeconds, 30s): the "curious" archetype ends a lot of messages with "?", so a generous window would match nearly everything and undo the point.
• Name matching is whole-word, so "wispy" doesn't summon Wisp. This signal is only reliable because the correction map already keeps ghost names intact through transcription.
• cadence.voiceReplyRequiresAddress: false restores the old behaviour. voiceReplyDelaySeconds / voiceReplyWindowSeconds were hardcoded and are now config. All hot-appliable.

Cost visibility — hover the cost pill for the share of input served from cache and the average output tokens per message (a one-line message is ~60; far above that means thinking is doing the spending). cache_read_input_tokens was already collected and then merged into a single total, which hid exactly the number needed to diagnose the above.

Beta release line — Actions → Beta Release → pick a branch. Tags vX.Y.Z-beta.N off the newest stable tag, builds the Windows installer, publishes a GitHub prerelease. Stable installs can't see it: the -beta.N suffix makes electron-builder write beta.yml, and a stable install only polls latest.yml.

app.updateChannel (auto | stable | beta) — because a beta install does not find its own way back. electron-updater follows a channel file, not a version comparison, so a stable release (which publishes only latest.yml) is never offered to a beta install. stable is the deliberate way home, and the only case that sets allowDowngrade.

Verification

• Full CI reproduced locally from a clean npm ci --ignore-scripts: syntax check, config validation, 95 tests pass (1 pre-existing Windows-only skip), npm run smoke OK.
• 8 new loop tests pin the voice gating: named ghost replies, undirected narration doesn't, a prompt answer counts but a late one doesn't, a statement isn't a question, "wispy" doesn't match "Wisp", the opt-out still works, and being addressed can't bypass the rate floor or staleness window.
• Beta version maths tested against a scratch repo seeded with the real tag history: correct base from the newest stable tag, beta.10 follows beta.9 (numeric sort), and a beta tag never becomes the base for the next one.
• Update-channel decision matrix checked across every setting × build combination — allowDowngrade is true in exactly one (explicitly chose stable while on a prerelease).
• All three workflow files parse as YAML.

Risk & rollout

• Two pre-existing release paths had to be fixed first, or the first beta tag would have broken stable releases. auto-release.yml picked the highest v* tag to bump from; a beta tag sorts above the release it precedes, and splitting 3.8.0-beta.1 on . puts 0-beta.1 into $patch. Verified to fail with invalid arithmetic operator, which under Actions' bash -e fails the job — every stable release blocked until someone deleted the tag. Now filtered to ^v[0-9]+\.[0-9]+\.[0-9]+$. Separately, release.yml's v* tag trigger also matches beta tags, and publishing one there would emit latest.yml — pushing a test build to every install. Its job is now guarded.
• Behaviour change worth watching: the cast will speak noticeably less often in response to talking. That's the intent, but if it now feels unresponsive, voiceReplyRequiresAddress: false restores the old behaviour immediately (no restart).
• Config additions, all defaulted: cadence.voiceReply*, cadence.voiceAnswerWindowSeconds, app.updateChannel. README rows added.
• Not verifiable here: the updater channel logic needs a packaged Windows build to exercise electron-updater for real — which is exactly what the beta line now enables. The first beta build is the test of the beta machinery.
• Ethics guarantees untouched: public/overlay.html is not modified; the 🤖 AI badge and "simulated viewers" watermark are unchanged.
• Release-Bump: minor → v3.7.0 would tag v3.8.0.

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