flake: chatd subscribe/promote queued (multiple variants)

[flake-investigator]

CI Run: https://github.com/coder/coder/actions/runs/23692597425
Failed Jobs: test-go-pg-17, test-go-pg (ubuntu-latest)
Commit: 386b44927385e74b71c5f8491699c66a7c8f3c11 (author: Kyle Carberry)

Failures (from job logs):

=== FAIL: coderd/x/chatd TestSubscribeNoPubsubNoDuplicateMessageParts (0.46s)
    chatd_test.go:1731:
        Error:      Condition satisfied
        Messages:   expected no duplicate events after snapshot

=== FAIL: coderd/x/chatd TestPromoteQueuedAllowsAlreadyQueuedMessageWhenUsageLimitReached (0.48s)
    chatd_test.go:884:
        Error:      Not equal:
                    expected: "pending"
                    actual  : "running"

(test-go-pg-17 also failed TestPromoteQueuedAllowsAlreadyQueuedMessageWhenUsageLimitReached with the same pending vs running mismatch.)

Error analysis:

• Both failures are in coderd/x/chatd/chatd_test.go.
• The tests assert timing-sensitive chat status/stream behavior (pending status and no duplicate events after snapshot), but the chat processor appears to advance to running and emit events earlier than expected.
• No panic, OOM, or data race warnings observed in the logs.

Root cause classification: Flaky test (timing-sensitive async behavior in chatd).

Assignment analysis:

• Unable to run git blame here; used git log --oneline -10 --follow coderd/x/chatd/chatd_test.go (via API).
• The most recent non-trivial change touching chatd tests is commit 386b4492 ("reduce chat streaming latency with event-driven acquisition") by Kyle Carberry, which also notes test stabilization adjustments.
• Assigning to the most recent substantive modifier of these tests and chatd timing behavior: @kylecarbs.

Related issues search (coder/internal):

• "TestPromoteQueuedAllowsAlreadyQueuedMessageWhenUsageLimitReached"
• "TestSubscribeNoPubsubNoDuplicateMessageParts"
• "duplicate events" + "chatd"
• "pending" + "running" + "chatd"
  No direct duplicates found.

Reproduction (if needed):

go test ./coderd/x/chatd -run TestSubscribeNoPubsubNoDuplicateMessageParts -count=1
GOFLAGS=-count=1 go test ./coderd/x/chatd -run TestPromoteQueuedAllowsAlreadyQueuedMessageWhenUsageLimitReached -count=1

[flake-investigator]

New occurrence on main (2026-03-30):

• CI Run: https://github.com/coder/coder/actions/runs/23731085276
• Job: test-go-pg-17
• Commit: 78986efed8a72341dac57feaa6b4382e69fb6aef (TJ)

Failure snippet:

=== FAIL: coderd/x/chatd TestPromoteQueuedAllowsAlreadyQueuedMessageWhenUsageLimitReached (0.28s)
    chatd_test.go:884:
        Error:      Not equal:
                    expected: "pending"
                    actual  : "running"

Notes:

• No panic/OOM/data race indicators in the job logs.
• Matches the same pending vs running mismatch tracked in this issue.

[flake-investigator]

New occurrence (2026-03-30):

• CI Run: https://github.com/coder/coder/actions/runs/23743007886
• Commit: 73f6cd81699bb11bf633db32955ff8f3b1854325 (Michael Suchacz)
• Failed Job: test-go-pg (ubuntu-latest)
• Failing test: coderd/x/chatd TestSendMessageInterruptBehaviorQueuesAndInterruptsWhenBusy
• Error:

=== FAIL: coderd/x/chatd TestSendMessageInterruptBehaviorQueuesAndInterruptsWhenBusy (0.31s)
    chatd_test.go:515:
        Error: "[...]" should have 1 item(s), but has 2

Log context shows the interrupt message was queued, but a second chat message was inserted into chat_messages before the assertion, likely due to wake-triggered immediate processing (same timing sensitivity as other variants in this issue).