fix: ring the bell when an opencode session asks a question

[omartelo]

An opencode session blocked on an interactive question looked exactly like one that was working: spinner, and the tool line reading question. No bell, no toast, no desktop notification.

Only permissions were reported. A question rides an event of its own, question.asked, and the module listened for permission.asked by name — so the last thing the card heard was the busy from the tool call that asked.

What a real run shows

opencode serve 1.18.16, driven through its own HTTP API, SSE on /event:

ask     +  130ms session.status {"type":"busy"}
        + 1665ms question.asked {"id":"que_…","sessionID":"ses_…"}
        …45s of silence — no status at all while the question is pending

reply   question.replied  → session.status busy   (+129ms)
reject  question.rejected → session.status idle   (+81ms)

So the spinner is not stale for a moment; it is stale for as long as the user takes. And the recovery needs nothing from us: answering and dismissing each produce a status of their own, which is why question.replied / question.rejected are not registered here.

Why by suffix, not by name

opencode asks in four spellings today — permission.asked, question.asked, and a .v2. variant of each — and its catalogue is not exhaustive: the same run emits server.heartbeat, which appears in none of the 94 event schemas its /doc publishes. Enumerating names is precisely how this bug happened.

The failure modes are not symmetric. A surplus bell is cleared by the next status report, which a reply produces within ~100ms. A name nobody enumerated is a session waiting behind a card that says it is busy.

Tests

tests/opencode.test.mjs covers all four .asked types, an unknown one (elicitation.asked — the point of the rule), the bound (question.asked.dismissed must report nothing), and a question raised inside a sub-session. Checked against the pre-fix module: the four new cases fail there and pass here, and permission.asked passes in both.

# tests 126
# pass 126
# fail 0

Contract

docs/session-state.md and docs/providers.md here now describe the rule instead of the name. The canonical contract in the lich repository (docs/hooks/session-state.md:42) still lists permission.asked in its opencode column — a follow-up there, no lich code change: waiting is an existing state and the payload is unchanged.