Codex runtime: session stops dispatching after stream compact disconnect

Summary

After a Codex runtime stream disconnect during compact (chatgpt.com), the affected session stops dispatching new messages entirely — inbox accumulates pending items but no new Codex process is spawned.

Environment

• duoduo version: 0.5.4-pre.1
• channel: feishu (channel-feishu 0.5.2-pre.6)
• runtime: codex

Reproduction

1. A Codex job or session is running and hits a network disconnect during compact:

``
Error running remote compact task: stream disconnected before completion:
error sending request for url (https://chatgpt.com/backend-api/codex/responses/compact)
``

1. After disconnect, the session state shows last_event_at updated on new ingress events, but no new streaming-card started log appears.
2. Session inbox accumulates .pending files indefinitely.
3. Subsequent messages from the user get no response.

Observed behavior

From feishu channel logs:

[INFO] [bot] ingress {"msgId":"om_x100b6e44275d2ca8c21994025c002d5", "sessionKey":"lark:oc_473ab07146f5702cfcefea0d63f1a9cd:...", "textLen":6}

— no streaming-card started follows. Three consecutive messages (file attachment, @mention, text) all ingressed but none dispatched.

Session inbox:

2026-05-27T14-57-09-050Z_i6wnj7yu6aj.pending
2026-05-27T14-58-31-457Z_dw84zcdc25c.pending

Root cause (observed)

The stale codex app-server processes (spawned by duoduo at 4:20PM, 5:30PM, 5:52PM) appear to hold dead connections. Killing all three duoduo-managed codex app-server PIDs (22579/22580, 30557/30558, 37887/37888) clears the block. However, duoduo does not auto-restart these processes after kill — next dispatch requires a new message trigger.

Also affected

Same root cause caused daily-tod-noon-scan-1230 job failure:

last_error: "Error running remote compact task: stream disconnected before completion: 
error sending request for url (https://chatgpt.com/backend-api/codex/responses/compact)"

Expected behavior

1. After a compact stream disconnect, the session should retry or spawn a fresh Codex process for pending inbox items.
2. Alternatively, the dispatcher should detect stale Codex app-server processes and restart them automatically.