Goal continuation repeatedly fails remote compaction after ~370 seconds, leaving a long session unable to continue (0.162.1) #54029

What issue are you seeing?

A long-running goal-backed Codex session repeatedly fails remote context compaction and cannot continue useful work. Reactivating the goal re-enters the same failure, rather than recovering the conversation. This report adds a locally verified JSONL timeline and repeat durations to related reports #38969 and #53111; it does not establish a shared root cause.

Exact error (10 recorded occurrences):

Error running remote compact task: stream disconnected before completion: Transport error: network error: error decoding response body

The terminal events classify this as codex_error_info: other. In the final six completed continuation attempts, last_agent_message is null and each attempt lasts approximately 369–371 seconds. The user ultimately paused the goal and requested reconstruction in a fresh session.

Environment

• Session metadata: cli_version: 0.162.1, originator: codex-tui, source: vscode. These are recorded metadata, not independent confirmation of the visible client UI or extension version.
• Final model/settings: gpt-6.1-sol, reasoning effort high, provider openai.
• Host checked during investigation: macOS 26.6 (25G72).
• Subscription tier and exact IDE extension version: not established.
• Session started 2026-10-13; failures below occurred 2026-10-14.
• Retained rollout: 19,856 JSONL records, 63,745,364 bytes, 18 recorded successful compactions. File size is on-disk history, not a claim about request payload size or active context length.

What steps can reproduce the bug?

Observed pattern (not a minimal deterministic reproduction):

1. Run a substantial coding/operations task with an active persistent goal over a long session.
2. Continue until remote compaction is needed.
3. Observe the decode/transport error above.
4. Reactivate the same goal to continue. Later attempts repeatedly fail after approximately six minutes without an assistant message.
5. Repeating continuation does not restore useful execution; work must be reconstructed in a fresh session.

Sanitized evidence timeline (UTC)

• Last recorded successful compaction: 2026-10-14 06:46:49.
• Nine Selected model is at capacity failures were recorded between 07:00 and 07:23. These are a separate error family; causality is unknown.
• Remote-compaction decode failures: 07:50:49, 07:59:49, 08:15:16, 10:13:01, 10:22:00, 10:35:23, 10:47:01, 11:20:45, 12:04:45, 12:27:56.
• Useful tool work did resume between some earlier failures. The final six completed attempts after 10:13 failed without further useful work.
• Final six continuation durations in milliseconds: 370979, 368874, 370352, 369684, 369740, 369958.
• Another attempt began at 16:27:10 and was manually interrupted when the goal was paused at 16:31:13; it is not counted as a completed compaction failure.

Representative terminal event with identifiers and task content omitted:

{
  "timestamp": "2026-10-14T12:27:56.212Z",
  "type": "event_msg",
  "payload": {
    "type": "task_complete",
    "last_agent_message": null,
    "error": {
      "message": "Error running remote compact task: stream disconnected before completion: Transport error: network error: error decoding response body",
      "codex_error_info": "other"
    },
    "duration_ms": 369958
  }
}

What is the expected behavior?

Compaction should either recover and allow the original goal to continue, or provide an actionable supported recovery/export path preserving task state. Repeated goal continuation should not silently spend roughly six minutes on the same failing compaction with no useful progress. Distinguishing response decoding, SSE transport, timeout, and backend errors would help diagnosis.

Additional information

The local history remains parseable. A separate new session can read it, reconstruct pending work, and create an active replacement goal. This is a workaround, not a repair of the original session. No history truncation, database modification, or removal of compaction records was performed.

The evidence does not distinguish a service problem, network problem, or client bug. No request IDs or HTTP status/body were available in the inspected terminal error events. Private conversation history, credentials, customer data, infrastructure details, and raw rollout attachments have deliberately not been posted. Please advise a supported private diagnostic channel if backend correlation requires the original thread identifier.