Subagent child sessions export empty (0 messages, outcome null) despite successful completion #59136

Activity

1. opencode-agent commented on Oct 26, 2026

   

   opencode-agentBot

   on Oct 26, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the report and the reproduction steps.

   In v2.1.4, the export endpoint gets the session info from the same record that GET /api/session?parentID=... lists, and its response always includes a tokens object. If the session has no outcome, the outcome field is left out rather than set to null. So a response with "tokens": null and "outcome": null doesn't look like the endpoint's raw output. Your tooling may have reshaped it, or the request may have gone to a different server than the one that ran the subagent. To narrow it down, could you share:

   1. The raw, unedited response of curl -s <base-url>/api/experimental/session/<childID>/export for one affected child, run against the same server (with prompts redacted if needed).
   2. The base URL/port you sent it to, and how that server was started (the desktop app's service, or opencode serve from the CLI). You mention both 2.1.4 and CLI 2.1.2 in Subagent-bound command: prompt to freshly-created child session can be silently dropped (no error, no log) #59135.
   3. The raw entry for the same child from GET /api/session?parentID=<parentID>, and the parent's <subagent sessionID=...> line, so we can confirm the IDs match.

   You can reply here with the details or edit the issue. If you edit it, please also leave a short comment saying you've updated it, so we know when you're done editing.

   These issues are related, ordered by likelihood, though none looks like the same problem:

   • Subagent-bound command: prompt to freshly-created child session can be silently dropped (no error, no log) #59135: your companion report, where the prompt to a new child session is silently dropped, leaving it with 0 messages. That would explain an empty child, but not one that completed.
   • subagent: completed child returns an empty result to the parent (v2, idle_outcome=succeeded with real token usage) #59051: a finished subagent that returns an empty result to the parent, even though the child recorded real token usage.
2. opencode-agent

   added

   triaging

   on Oct 26, 2026
3. megadoom99 commented on Oct 26, 2026

   

   megadoom99

   on Oct 26, 2026

   Author

   More actions

   Follow-up from testing: a related manifestation — when the subagent's final turn produces no text (e.g. it exhausts its step budget mid-tool-call before composing its summary), the job reports outcome: "succeeded" but the parent synthetic says Subagent completed without a text response. The child generated ~66k output tokens of work across its turns, all of which is unreachable: the child export is empty (this issue's main subject) and the synthetic only carries the last text message. Two consequences worth considering upstream: (1) the subagent completion payload could carry the accumulated final text or a transcript reference rather than only the last message; (2) a step-limit exhaustion arguably should not be a silent success.
4. opencodejr commented on Oct 28, 2026

   

   opencodejrBot

   on Oct 28, 2026

   More actions

   Thanks for the report. I checked it against the v2.1.4 code. The export endpoint gets the session info with the same lookup as GET /api/session and reads messages from the same store. It leaves out only messages that haven't finished yet. Also, tokens is always present in session info and outcome is left out rather than set to null. So a response with "tokens": null, "outcome": null shouldn't be possible, and neither should a list entry that has real tokens while the export of the same session has none. That makes me think the JSON in the issue was paraphrased, or the export went to a different session or server.

   To work out what's happening, could you share, briefly and in your own words:

   • The raw, unedited responses, all for the same child session ID and from the same running service:
     • GET /api/session/{childID}
     • the matching entry from GET /api/session?parentID={parentID}
     • GET /api/session/{childID}/message
     • GET /api/experimental/session/{childID}/export
   • Whether the sessionID in the parent's <subagent …> message is exactly the child ID you exported.
   • Your command and agent frontmatter (leave out anything private).
   • Whether you get the same result with opencode serve instead of the desktop app's bundled service.

   You can reply here or edit the issue. If you edit it, please also leave a short comment saying you've updated it, so we know when you're done.

   Related issues, ordered by likelihood (none looks like an exact match):

   • Subagent-bound command: prompt to freshly-created child session can be silently dropped (no error, no log) #59135: your companion report about a prompt to a new child session being dropped, which leaves the child with 0 messages
   • subagent: completed child returns an empty result to the parent (v2, idle_outcome=succeeded with real token usage) #59051: a child finishes as succeeded but the parent gets an empty result (the opposite direction)
   • session: first-turn subtask command leaves parent session empty and untitled #52468: a first-turn subtask command leaves the parent session empty, with an inbox row that is never delivered
5. opencodejr

   closed this as not plannedon Oct 31, 2026
6. opencodejr commented on Oct 31, 2026

   

   opencodejrBot

   on Oct 31, 2026

   More actions

   I'm closing this as not planned because we haven't heard back. In v2.1.4 the export endpoint reads session info and messages from the same place as the session list. So the quoted output ("tokens": null, and real tokens in the list but not in the export) can't come from it. We needed the raw responses for one child session ID to see what's actually going on.

   If this still happens, please reply here with the unedited output of GET /api/session/{childID}, GET /api/session/{childID}/message and GET /api/experimental/session/{childID}/export, all from the same running service, and we'll reopen it. #59135 covers the related problem of a child session left with 0 messages.
7. opencodejr

   reopened this on Nov 3, 2026