ACP "redirect" fabricates empty ~1ms turn_completes when the agent advertises steering=false, wedging the conversation (cancelled/pending_review) and swallowing the send

[lumberjackchef]

Summary

When a user sends a prompt to an ACP agent that does not support mid-turn steering, Codeg's "Redirected the active turn with your correction" flow fabricates an empty turn_complete ~1ms later, moves the conversation to cancelled or pending_review (the status is not deterministic), and leaves the row unable to send. The prompt is either lost entirely, or — worse — it does land in the agent's store and gets answered, but the reply never streams back because the turn was already closed client-side.

Observed with: codeg-server v0.28.1 (native binary), agent = Hermes 0.20.5 via hermes acp, which advertises steering: advertised=false. Four incidents on our fleet in 48 hours (JAR-1103, JAR-1133, JAR-990, and another today), all on the same production server.

Wire evidence (server log, connection d7f4cd4d-…, conversation "Tackle App Launch review batch")

Real turns take seconds to minutes:

2026-08-28T00:01:41.494228Z  status_changed session=...0278c6f6... Connected -> Prompting
2026-08-28T00:05:18.239865Z  turn_complete session=...0278c6f6... stop_reason=end_turn   # 3m37s — real

Two sends while the agent had a live turn that was invisible to Codeg's session state (an out-of-band background-process notification had just been injected into the session at 00:30:00Z):

2026-08-28T00:31:26.905952Z  status_changed session=...0278c6f6... Connected -> Prompting
2026-08-28T00:31:26.907866Z  turn_complete session=...0278c6f6... stop_reason=end_turn   # 1.9 ms — synthetic, empty

2026-08-28T00:33:29.680022Z  status_changed session=...0278c6f6... Connected -> Prompting
2026-08-28T00:33:29.680733Z  turn_complete session=...0278c6f6... stop_reason=end_turn   # 0.7 ms — synthetic, empty

After that: the conversation row stayed pending_review (an earlier incident flipped to cancelled instead — non-deterministic), every further send failed with no surfaced error, and the agent's real turn_complete never arrived (row still stuck 10+ minutes later; auto-reconcile is not guaranteed).

The insidious part — the first send (00:31:26) was lost for good (never reached the agent store), but the second (00:33:29.680Z, same instant as the synthetic completion) landed and was answered: the agent persisted its reply at 00:34:10 (finish_reason=stop), and it never streamed to the tab because the turn had already been closed client-side.

Diagnostic signature for anyone reproducing: grep the server log for the session uuid and look for Prompting → turn_complete pairs with sub-2ms deltas (an earlier incident had three swallowed sends within 8s).

Trigger hypothesis

The redirect path appears to be attempted unconditionally when a send races an agent turn, even when the agent advertised steering: advertised=false. Two candidate triggers, both observed today:

1. Invisible live turn — a background-process notification was injected out-of-band into the session (user-role message, 00:30:00Z), so the agent was mid-turn while Codeg's session state said idle. The send became a "redirect" that could not be delivered.
2. Stale session identity after compression rotation — Hermes rotated its session id mid-conversation (compression, 11:48Z); Codeg's row still maps the old id. Lineage resolution makes load_session work, but the conversational redirect/status logic may key on stale state.

Either way the agent cannot accept the correction, so an empty completion is synthesized and the row's status is flipped — which is the destructive part.

Suggested fixes (any one of these materially helps)

1. When steering: advertised=false, never attempt the redirect flow for mid-turn sends. Queue the prompt for the next turn, or surface a clear "agent is busy" state, instead of fabricating a completion.
2. Don't flip a regular conversation's status (cancelled/pending_review) from an empty synthetic completion — keep the row sendable and the client-stream in sync.
3. If the prompt was passed to the agent anyway (as observed), match its eventual real turn_complete to the conversation instead of discarding it (same direction as Complete grok turns when the session/prompt response is lost #553 for grok).
4. #539 (journal sent prompts) would at least make recovery lossless.

Environment

• codeg-server v0.28.1 native (--supervise), agent Hermes 0.20.5 via hermes acp (steering: advertised=false, native=false)
• Related: 任务会话把「截断的 tool call」写入历史后，每次点「继续」都被上游 400 拒绝并静默回到待审阅 #538 (same silent pending_review end-state, different trigger), fix(acp): journal sent prompts so an interrupted chat cannot reopen blank #539, Complete grok turns when the session/prompt response is lost #553, [Bug] 旧 generation 的 cancelled 事件会清空新 generation 的待处理请求 #564/fix(tasks): preserve current waits after stale generation completion #567 (fixed)