[enowdev]

GOA got a worker's goa report line, then a "finished a turn" line with an older message, sometimes from mid-turn.

Cause

Your reading was right, with one correction: the old message and the two lines are separate causes.

1. Two lines: goaReportSeen (daemon/goa_tasks.go:400) tells at once. goa report is a tool call made while the turn still runs, so goaWorkerChanged (:289) tells again when the turn ends. Nothing joins them. The no-task path has the same split (daemon/goa_told.go:188,195), where the check for "that was its turn's end" only holds when both arrive in one update.
2. Old message: goaResult (:456) took the transcript's last assistant message at that moment, with no check that the transcript had reached the turn's end.

Measured on real Claude Code 2.1.318: in a turn that said INTERIM-ONE, ran a command and said FINAL-TWO, the Stop hook received last_assistant_message: "FINAL-TWO", while its transcript had no readable rows at that moment or 1.5 seconds later. So the transcript does lag. That was one first-turn run.

The session you named: 37a8b7c577f1 is on the Mac. I read its recent chat through the app and it no longer shows that turn, so I couldn't confirm this on that exact case. The waiting queue only delays lines; it doesn't cause them.

Case 5 of #30: not touched there. This is that fix.

Fix

• Real text: the turn-end hook's last_assistant_message is kept as SessionInfo.TurnText and travels with session updates, so GOA on another host gets it too.
• Fallback: otherwise the transcript is re-read until its last entry is a message, up to 6 seconds, then the line says the message isn't recorded yet rather than give an old one.
• One line: a report made while the agent works waits up to 30 seconds for its turn's end to join it. A turn's end still queued takes a later report. Past that, it goes alone and the end gets its own line, so two lines are still possible on a long tail.
• Never lost or merged: questions, failures and corrections stay separate, and they release what was held so order holds. Each real turn keeps its own line. Lines of one session don't pass what it holds; other sessions aren't held up.

Older daemons on another host send no TurnText, and the fallback covers them. Codex's hook field isn't verified, as it isn't installed here; the fallback covers it too.

A worker that ends a turn waiting on a background command: that message really is its last, so it stays, marked "may be mid-work" as before.

Tests

Report then end, end then report, transcript written late, queued while GOA is busy, two real turns, no task, a held report not lost, per-session hold. On enowxlabs: daemon and core/session.

Not run: a real GOA with a real worker; Codex.

Merge order: this conflicts with #30 in goaSeen and rpc_goa.go, both small: #30 moves the same lines into goaWorkerSeen, where the goaTurnStarted call then belongs.

Touches no go.mod, version/ or cloud/, so merging doesn't deploy the cloud.

🤖 Generated with Claude Code

[enowdev]

Its commits are in main through the 0.7.0 release merge (#40, e306b3b).