feat(channel): answer follow-ups after a mention subscribes

[BenTaylorDev]

What

onMention ran the agent but never subscribed the conversation, and no onMessage handler was registered at all — so OpenTag answered only when explicitly addressed, on every single turn.

• A mention now subscribes the conversation and runs the agent.
• Later messages in that conversation run the agent too, guarded by isSubscribed() so the bot stays silent in conversations it was never invited into.
• The existing error handling (user-facing reply + reportRecoverableError, AggregateError when the reply itself fails) moves into a shared runTurn handler instead of being duplicated across both entry points.

Dispatch is exclusive in channels-core — a mentioned turn goes to onMention whenever any handler is registered, and to onMessage otherwise — so the mention that subscribes still runs the agent exactly once.

runAgent(managedRunInput(message)) is kept rather than a bare runAgent(): managedRunInput carries the multimodal content parts and the per-platform tools/context that a managed turn needs.

Testing

Two new tests in app/channel.test.ts, both driven through FakeAdapter's ingress sink with explicit operation.mentioned (turn dedup keys off logicalMessageId/revisionId, so each turn gets its own id):

• stays quiet in a conversation that never addressed it — one non-mention turn, zero agent runs.
• answers follow-ups once a mention subscribes that conversation — mention in c1, follow-up in c1, unrelated non-mention in c2; exactly 2 runs.

Red-checked — with app/channel.tsx stashed, the second test fails, so it genuinely covers the new wiring:

=== WITHOUT the onMessage wiring ===
AssertionError: expected [ 'run' ] to have a length of 2 but got 1
      Tests  1 failed | 12 skipped (13)

With the change:

 Test Files  1 passed (1)
      Tests  1 passed | 12 skipped (13)

Runs are counted from inside the FakeAgent script, not from a field on the agent: every turn runs on a clone, and FakeAgent.clone() copies the remaining script but resets its own counters, so CapturingAgent.calls reads 0 even when the agent did run.

Pre-existing red, unchanged by this PR

pnpm check-types and pnpm test are already failing on main from drift against @copilotkit/channels@0.6.0 (PlatformUser no longer exported, ProviderActor.kind now required, FakeAgent.clone() resetting counters). Measured before and after:

	tsc errors	Test files	Tests
baseline (main)	20	3 failed / 17 passed	9 failed / 130 passed (139)
this branch	20	3 failed / 17 passed	9 failed / 132 passed (141)

The failing test names are byte-identical before and after (diff of the sorted failure list is empty), and no error is reported in app/channel.tsx. Fixing that drift is out of scope here.

🤖 Generated with Claude Code

[BenTaylorDev]

Closing — redundant. main already subscribes on mention and gates onMessage on isSubscribed() (and additionally drops bot/app-authored messages, which this branch did not), and it already carries equivalent tests. I based this on the commit the channels-sdk submodule pins, which is 22 commits behind main; the wiring was added upstream in 44a8a49 and d67f23a.