feat(slack-agent): gate thread follow-ups behind a relevance classifier

[JeremyFunk]

Problem

Thread follow-up promotion (agent/lib/thread-follow-up.ts) only answers a mechanical question — is the bot part of this thread. Once it is, every human reply in the thread dispatched a full agent turn: in an incident thread where two engineers are talking to each other, the bot answered each of their messages. Cost amplifier, and a bot that butts into conversations never addressed to it.

Change

A confirmed follow-up now also passes a semantic gate (agent/lib/follow-up-relevance.ts): one classifier call — recent thread tail + the new reply in, a single RESPOND/PASS word out — asking whether the reply is actually directed at the bot or expects it to act. Orders of magnitude cheaper than the agent turn it gates.

• Real @mentions and DMs skip the gate. An explicit address answers the question itself.
• A PASS is silent by design: it runs before the :eyes: ack and typing indicator, so a message the bot stays out of gets no reaction at all, and drops via the same throw-to-un-dispatch escape hatch as the disengagement drop. The bot stays engaged — the next reply that is for it gets answered. The only trace is a structured follow_up_not_relevant log (ids only, never message text).
• It fails open, same asymmetry as confirmThreadFollowUp: classifier error, timeout (10s), or an unparseable verdict all answer the follow-up. A wrong drop loses a user's message with nothing on screen to explain it; a wrong answer costs one turn. The prompt's tiebreak leans the same way.
• Gate model is separately configurable (OPENROUTER_GATE_MODEL) since the job wants a small, fast model; unset, it follows OPENROUTER_MODEL rather than silently using a third one.

Runs on the dispatch path (post-200, inside eve's waitUntil), so it spends nothing from Slack's ~3s webhook budget.

Trade-off

Every un-mentioned reply in an engaged thread now costs one tiny classifier call. That is the mechanism working; set OPENROUTER_GATE_MODEL to something small to make it cheaper.

Testing

• bun test in apps/slack-agent: 281 pass (12 new — verdict parsing, prompt contents, fail-open on error/unparseable, deps-injected classifier)
• tsc clean

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}