Fix Slack native stream rollover

[tdietert]

Fixes #671

What problems was I solving

Slack native streamed messages can stop accepting chat.appendStream / chat.stopStream calls while a long-running agent turn is still producing output. Slack reports that state as message_not_in_streaming_state; previously the adapter rejected and dropped later output.

This PR makes the Slack adapter treat that documented error as the recovery boundary: when the current native streamed message is no longer streamable, the adapter continues the turn in a fresh streamed message in the same thread and retries only the pending chunk.

What user-facing changes did I ship

• packages/adapter-slack/src/index.ts — Slack native streams now recover reactively from message_not_in_streaming_state by starting a fresh stream and retrying only the pending markdown delta.
• packages/adapter-slack/src/index.ts — The adapter now calls chat.startStream, chat.appendStream, and chat.stopStream directly instead of going through ChatStreamer, so it owns stream lifecycle and confirmed-offset bookkeeping explicitly.
• packages/adapter-slack/src/index.ts — Latest structured task/plan chunks are re-seeded on rollover; fallback updates only the current message segment to avoid duplicating text already emitted in earlier rolled messages.
• packages/adapter-slack/package.json — Keeps the Slack Web API floor at ^7.17.0 for the native stream chunk API surface used by the adapter.

How I implemented it

• Removed the ChatStreamer/buffer_size: 1 path and the open-code-fence-prefix workaround from the PR.
• Added direct native stream state tracking:
  • active Slack stream ts
  • last known message ts
  • current message start offset
  • last confirmed markdown offset
  • latest structured chunk snapshots
  • bounded rollover count
• On append expiry, the adapter starts a new stream in the same thread, optionally re-seeds latest structured chunks, and retries exactly the pending delta.
• On stop expiry, the adapter treats an already-flushed stream as complete; if final blocks or unconfirmed tail need recovery, it uses chat.update on the latest stream segment rather than duplicating the full cumulative answer.
• Updated adapter and integration tests to mock/assert direct Web API streaming methods instead of chatStream().
• Added a patch changeset for @chat-adapter/slack.

How to verify it

Manual testing

• Start a Slack native stream, let Slack report message_not_in_streaming_state, and confirm later deltas continue in a fresh streamed message.
• Confirm task/plan progress chunks still appear after a rollover.
• Confirm stop-time expiry does not truncate an already-flushed response.

Automated tests

pnpm --filter @chat-adapter/slack... build
pnpm --filter @chat-adapter/slack... typecheck
pnpm --filter @chat-adapter/slack test
pnpm check
pnpm validate

All of the above passed locally after the refactor.

Changelog

@chat-adapter/slack now recovers expired Slack native streams by reactively rolling long-running replies onto fresh streamed messages without dropping pending text.

[vercel]

@tdietert is attempting to deploy a commit to the Vercel Team on Vercel.

A member of the Team first needs to authorize it.

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	npm/​@​slack/​web-api@​7.18.0

View full report