fix(v2): refresh connected providers after late host registration

[doublepi123]

Summary

On OpenCode v2, auto-capture and profile learning can report a configured provider as "not connected" for the lifetime of the server process.

The plugin reads the provider directory once during setup. At that point OpenCode has only registered its built-in opencode provider; user-configured providers appear about two seconds later, together with provider.updated and model.updated events. The snapshot therefore stays at ["opencode"], so a provider configured via opencodeProvider (for example a custom OpenAI-compatible provider) fails the connection gate before any generation request is made.

This was reproduced with a probe plugin against an isolated OpenCode 2.0.21 server: ctx.provider.list() returned only opencode during setup and opencode plus the configured provider two seconds later. HTTP /api/provider showed the configured provider as enabled while the plugin kept reporting it as disconnected.

Changes

• Refresh the connected-provider snapshot on provider.updated and model.updated. Bursts are coalesced into at most one trailing refresh; refreshes stop after dispose.
• Before failing the connectivity gate in auto-capture and profile learning, refresh once and re-check (ensureProviderConnected). The error message is unchanged.
• Refresh errors are logged and keep the previous snapshot. A disposing plugin instance only clears its own refresher.
• The activation filter and provider-directory semantics from fix: harden provider discovery, capture scheduling and learning locks #369 are unchanged.

Test plan

• New tests/v2-provider-refresh.test.ts uses the real plugin, V2 adapter event path, provider module state and auto-capture gate. Covers late registration, refresh-on-miss, unknown providers, refresh errors, event bursts and dispose.
• The new tests fail on the current main and pass with this change.
• Existing loader mocks gained ensureProviderConnected.

Checklist

• Branch is based on current main (a24d79e, v2.28.3).
• bun test passes (672 pass / 0 fail / 4 platform skips).
• bun run typecheck passes.
• bun run check passes.
• Docs: no configuration or contribution-process change.

[EyJunge1]

Review summary

Approve — the miss-refresh path (ensureProviderConnected) correctly fixes the late provider-registration bug for auto-capture and profile learning. CI is green and the regression coverage for the gate path looks solid.

Not fully optimal: the event-driven refresh is defense-in-depth but lightly guaranteed, and a couple of comments/wiring details don't match the code. None of that blocks the bugfix.

What works well

• Refresh-on-miss before failing the connectivity gate (auto-capture + profile LLM)
• Snapshot retained on refresh errors; existing error strings unchanged
• Dispose clears the refresher by function identity (multi-instance safe)
• Activation / provider.list() semantics from #369 left alone
• Subprocess fixture covers miss, ghost, capture-error, refresh-error, dispose

Nits (non-blocking follow-ups)

1. Event path vs eventBelongsToLocation
   V2 adapter drops events with neither location.directory nor a resolvable sessionID (src/v2/legacy-client.ts). provider.updated / model.updated are ephemeral with optional location. If the host publishes without location, the event handler never runs and only miss-refresh saves you. Worth confirming real host payloads include location, and giving burst-fixture events a directory so the coalescing test actually exercises the adapter path.

2. Comment vs code — gate bypasses coalescing
   Comment says ensureProviderConnected reuses runCoalescedRefresh, but providerRefresher calls refreshConnectedProviders() directly → parallel list() possible under event + miss. Either wire through coalescing or fix the comment.

3. Init refresh outside the mutex
   void refreshConnectedProviders() doesn't set refreshInFlight. Prefer runCoalescedRefresh() for the initial call so early events don't race a second list.

4. Minor

   • Log still says "Failed to initialize…" on later refreshes
   • Burst assertion listCalls < 8 is soft; a tighter coalesce bound (or location-bearing events) would be stronger
   • Optional unit tests for ensureProviderConnected / clear-by-identity

Verdict

Merge as-is is fine. The gate miss-refresh is the load-bearing fix; the nits above are polish / hardening for a follow-up if you want them.

[EyJunge1]

Nit: This init call bypasses the coalesce mutex below (refreshInFlight is still unset). Prefer runCoalescedRefresh() so a concurrent provider.updated during init cannot start a second parallel provider.list().