Connect with ChatGPT rejects Codex subagents and leaves the chat unable to continue #17679

Activity

1. samanyudas commented on Oct 10, 2026

   

   samanyudas

   on Oct 10, 2026 – with ChatGPT Codex Connector

   Last edited by samanyudas

   Author

   More actions

   Additional reproduction from a new screenshot supplied by the reporter:

   1. Send the text-only prompt Calling a subagent.
   2. The assistant says it will call a subagent to confirm delegation is working.
   3. A single child named Delegation Check appears and fails after approximately one second.
   4. The parent displays a provider error with type: invalid_request_error, param: input, and code: subscription_sharing_unsupported_capability.

   The screenshot shows GPT-6.1-Sol, low reasoning effort, and full access. This narrows the observed reproduction to a simple delegation check; it does not identify the exact rejected Responses API input item.

   

   Added by Codex via T3 triage.
2. roelschr commented on Oct 10, 2026

   

   roelschr

   on Oct 10, 2026

   More actions

   Additional reproduction on Windows, plus a working alternative delegation path.

   Environment

   • T3 Code desktop: 0.0.46-nightly.20261009.2873 (reported in the app-server initialize request).
   • Managed Codex provider, using ChatGPT subscription sharing. The Codex session metadata identifies model_provider: openai_token_sharing.
   • Codex app-server: 0.156.1 (session metadata).
   • Model: gpt-6-astra, medium reasoning, full access.

   Observed failure

   I asked the agent to have a subagent review a GitHub PR while the parent helped test the PR's remaining checklist items. T3's injected orchestration instructions prefer native subagent tools for compatible same-provider work.

   Normal local commands and T3 MCP calls succeeded. The agent then used native spawn_agent. The child failed with subscription_sharing_unsupported_capability (param: input); the parent subsequently failed with the same error.

   The local Codex rollouts contain this sequence on 2026-10-10 (UTC):

   15:23:01.544  parent: function_call, name=spawn_agent
   15:23:02.930  child: response_item, type=agent_message,
                        author=/root, recipient=/root/review_pr17
   15:23:04.228  child: subscription_sharing_unsupported_capability
   15:23:11.361  parent: response_item, type=agent_message,
                        author=/root/review_pr17, recipient=/root
                        (contains the child's failure notification)
   15:23:12.633  parent: subscription_sharing_unsupported_capability
   

   This makes agent_message a suspected unsupported input item. These are local rollout/provider logs, not a captured outgoing Responses HTTP body, so the exact rejected wire item is not independently confirmed. I have not tested continuing the failed chat.

   Working workaround

   I tried the suggested prompt explicitly selecting T3-owned delegation, and it works:

   Use the t3-code MCP delegate_task tool for the review, not native spawn_agent or collaboration tools. Delegate a review of the PR using provider codex and model gpt-6-astra, while you help me test its unchecked items. Track the review with task_status.

   The project path and PR URL have been omitted here because they are incidental to the provider failure. This is a user-confirmed working workaround for the requested task, not an independently captured trace of every tool call or a test of #17499.

   It would be useful for the managed subscription-sharing route to select a compatible delegation mechanism automatically, and for a child failure to avoid introducing history that causes the parent to fail as well.
3. maria-rcks

   closed this as completedin #17499on Oct 11, 2026
4. samanyudas commented on Oct 12, 2026

   

   samanyudas

   on Oct 12, 2026

   Author

   More actions

   Issue still persists after update.

   

   👍React with 👍1Reacted by Caleb Mech
5. samanyudas commented on Oct 12, 2026

   

   samanyudas

   on Oct 12, 2026 – with ChatGPT Codex Connector

   Author

   More actions

   The issue still reproduces after updating the Linux background server to 0.0.46-nightly.20261011.2994.

   The reporter tested both the existing failed chat and a new chat. The installed server bundle contains both launch overrides added by #17499: features.multi_agent=false and features.multi_agent_v2=false. The service state identifies .2994 as active, and the failing provider's initialize request reports that same version.

   New-chat reproduction on 2026-10-12, UTC:

   1. At 01:47:07, a new provider thread starts.
   2. Send the text-only prompt try callin' a subagent using GPT-6.1-Sol, low reasoning effort, full access, and openai_token_sharing.
   3. At 01:47:21, native subagent activity appears.
   4. At 01:47:22, the child fails with subscription_sharing_unsupported_capability.
   5. At 01:47:26, the parent fails with the same error.

   The provider log also records a native collaboration wait call. The outgoing thread/start config does not explicitly re-enable multi-agent features. The exact spawned process arguments and rejected Responses request body have not been captured, so the reason native delegation remains available is not yet established. This is evidence that the merged prevention fix is insufficient in this installation, rather than only a retry of contaminated history.

   Environment: Linux x64, kernel 7.2.5-3-omarchy, Node v26.8.2. The affected provider's status cache reports Codex 0.162.1; the running executable version has not been independently confirmed. Current client platform has not been confirmed in this follow-up.

   The reporter confirms that the workaround works in a new chat when explicitly requesting T3's delegate_task and task_status tools and avoiding native spawn_agent and collaboration tools. Native delegation still fails as described above.

   No configuration, credentials, database data, or source was modified during this investigation.

   Prepared by Codex via T3 triage. The primary session's exact model identifier is not exposed.