Skip to content

[Bug]: Astra messages fail in T3 Code but the same prompt/model works in terminal Codex #21061

Closed as not planned

Description

@milewski

Summary

Sending a regular message with Astra in T3 Code repeatedly fails with subscription_sharing_unsupported_capability. Sol works in T3 Code. The reporter confirms that the same prompt with the same Astra model works when running Codex directly from the terminal.

This is occurring on T3 Code 0.1.1-nightly.20261018.3285 on Windows. Local protocol logs show two newly started Astra conversations failing after native subagent activity, followed by a text-only retry failing in the first conversation.

Area

apps/server — managed Codex / Connect with ChatGPT subscription-sharing integration.

Steps to reproduce

Reporter-observed sequence; the exact private prompt is omitted:

  1. Send regular messages using Sol in T3 Code; this works.
  2. Select Astra and send a regular message.
  3. Observe subscription_sharing_unsupported_capability.
  4. Run Codex directly in the terminal with the same prompt and the same Astra model; the reporter confirms this works.

The user-visible trigger is a regular message. The captured T3 Astra runs start native subAgentActivity internally before the rejection. A text-only follow-up in an affected conversation also fails, and a second newly started Astra conversation reproduces the child/parent failure sequence.

Expected behavior

The same prompt and Astra model should work in T3 Code as they do in terminal Codex. Internal subagent execution should use a mechanism compatible with the configured authentication route, and a child failure should not prevent ordinary follow-up messages.

Actual behavior

The Astra child fails, then its parent fails with the same provider rejection. A subsequent text-only follow-up also fails. All five captured error notifications have willRetry: false, type: invalid_request_error, and param: input; the corresponding turns end with status: failed.

Exact error supplied by the reporter:

{
  "error": {
    "message": "User subscription sharing currently supports only text, image, and file messages, item references, additional tools, compaction summaries, reasoning items, Web Search call items, developer function and custom tool call items, and programmatic tool calling items. Remove unsupported input items or use an API key instead.",
    "type": "invalid_request_error",
    "param": "input",
    "code": "subscription_sharing_unsupported_capability"
  }
}

Impact

Frequent failures prevent the reporter from using Astra in T3 Code, even though the same prompt/model works in terminal Codex. An affected conversation also rejects ordinary text follow-ups.

Version or commit

0.1.1-nightly.20261018.3285, confirmed by the live T3 server and the affected Codex initialize request.

Environment

  • Windows x64; Codex initialize response reports Windows 10.0.26300.
  • T3-managed Codex app-server initialize response reports 0.163.0.
  • Affected conversations: gpt-6-astra; thread/start results report modelProvider: openai_token_sharing.
  • Native children use Astra / xhigh; the captured failed follow-up also uses Astra / xhigh.
  • Sol (gpt-6.1-sol) works for the reporter in T3 Code. The investigation session uses Sol on the same subscription-sharing route and can run tools successfully.
  • Terminal Codex success with the same prompt and Astra model is reporter-confirmed. The terminal CLI version and authentication route were not independently captured, so this comparison does not isolate model selection from authentication/integration differences.

Logs or stack traces

Read-only local protocol evidence. All timestamps are UTC on 2026-10-18; identifiers and private message contents are omitted.

01:29:00.922  first Astra conversation starts; openai_token_sharing
01:29:30.284  parent: subAgentActivity started, gpt-6-astra / xhigh
01:29:32.760  child: subscription_sharing_unsupported_capability
01:29:43.631  parent: same error
01:33:04.070  parent: text-only follow-up, Astra / xhigh
01:33:05.929  follow-up: same error

01:35:39.246  second Astra conversation starts; openai_token_sharing
01:36:05.924  parent: subAgentActivity started, gpt-6-astra / xhigh
01:36:08.345  child: subscription_sharing_unsupported_capability
01:36:16.175  parent: same error

The outgoing Responses HTTP body was not captured, so the exact unsupported wire item is unconfirmed.

Related reports and proposed fix

Workaround

Use Sol in T3 Code, or run the same prompt and Astra model directly with terminal Codex, both of which the reporter confirms work. A second new Astra conversation in T3 Code still failed in the captured logs.

No configuration, credentials, installed code, or conversation history was modified during this investigation.

Prepared with Codex assistance (GPT-6.1 Sol) from the reporter's description and read-only local protocol evidence. Raw logs, private prompts, credentials, local paths and conversation identifiers are omitted.

Activity

  1. DannyMarkus commented on Oct 20, 2026

    @DannyMarkus

    Additional independent reproduction on Windows (2026-10-20), with a more specific rejection than the one in this issue.

    Environment

    • T3 Code 0.1.0 stable and 0.1.1-nightly.20261020.3386 (both affected)
    • T3-managed Codex CLI/app-server 0.156.1
    • ChatGPT subscription sharing (modelProvider: openai_token_sharing)
    • Model gpt-6-astra, reasoning medium and xhigh
    • awork MCP configured and working with Sol

    Minimal reproduction: Start a new T3 Code conversation with Astra/medium and send: Antworte ausschließlich mit OK. Verwende keine Werkzeuge. ("Reply only OK. Do not use tools."). Reproduces even with No project selected, no subagent request and no tool use in the user message. The turn fails before any assistant reply, with:

    invalid_request_error (param: input)
    code: subscription_sharing_unsupported_capability
    User subscription sharing currently supports only 'web_search' and developer-defined 'function' and 'custom' tools, including 'namespace' and 'programmatic_tool_calling'; tool 'tool_search' is not supported. Remove it or use an API key instead.
    

    Controls

    • Same managed codex.exe (0.156.1) invoked directly via codex exec --model gpt-6-astra -c 'model_reasoning_effort="medium"' --sandbox read-only --skip-git-repo-check "Antworte ausschließlich mit OK. Verwende keine Werkzeuge." succeeds (OK).
    • T3 Code with Sol (GPT-5.6-Sol and GPT-6-Sol) can access the awork MCP server.
    • Updating T3 stable to the nightly above did not resolve the problem.

    Redacted app-server trace of the failed T3 turn (UTC):

    04:05:31.471 thread/start model=gpt-6-astra, config includes T3's local MCP server
    04:05:33.910 thread created: modelProvider=openai_token_sharing
    04:05:33.972 turn/start model=gpt-6-astra, effort=medium, text-only input
    04:05:34.226 awork MCP ready
    04:05:34.444 codex_apps and t3-code MCP ready
    04:05:36.835 thread status=systemError
    04:05:36.836 error: tool 'tool_search' not supported; willRetry=false
    04:05:36.856 turn completed, status=failed
    

    The T3 turn/start also includes extensive additionalContext orchestration instructions and sandboxPolicy: dangerFullAccess, unlike the direct read-only CLI control. Neither is established as causal. The log doesn't contain the full upstream Responses request, so the component that inserts tool_search is still unconfirmed.

    Potentially useful distinction from the original report: here a brand-new, text-only Astra conversation fails without any observed native subagent activity. Please investigate why the T3-managed Codex app-server request advertises tool_search on the ChatGPT subscription-sharing route and whether that tool can be omitted for this route.

    Only redacted evidence included; no raw logs, internal paths, prompts beyond the minimal reproduction, IDs, credentials or private project information.

  2. marketing-namou commented on Oct 30, 2026

    @marketing-namou

    Removed — posted from the wrong account.

  3. matt2102 commented on Oct 30, 2026

    @matt2102

    Fresh repro on current stable (0.1.5) — and evidence the #17499 fix isn't effective

    Hit the exact same failure today on a brand-new thread, on the current stable release (not a nightly), which suggests this isn't just a lingering nightly regression.

    What happened

    A new top-level Codex thread (model gpt-6-astra, provider openai_token_sharing) started a native Codex subagent (subAgentActivity, not a T3 delegate_task) partway through its first turn. The spawned child thread failed immediately with subscription_sharing_unsupported_capability; ~10s later the parent turn failed with the identical error.

    Repro sequence (UTC, from local protocol logs)

    21:37:21.036  turn/start on parent thread (gpt-6-astra, model_provider=openai_token_sharing)
    21:37:39.582  parent begins a reasoning item
    21:37:42.243  parent emits item/started subAgentActivity → spawns child thread
    21:37:43.782  child thread: thread/status/changed → systemError
    21:37:43.783  child turn/completed, status "failed", error.code = subscription_sharing_unsupported_capability
    21:37:53.305  parent thread: thread/status/changed → systemError (same error.code)
    21:37:53.309  parent turn/completed, status "failed"
    

    Both errors carry "param":"input", "willRetry":false, and the same message as in the issue description ("User subscription sharing currently supports only text, image, and file messages... Remove unsupported input items or use an API key instead.").

    Why I don't think the PR #17499 fix is landing

    CodexManagedRuntime.ts on v0.1.5 still sets the launch overrides from that PR:

    features.multi_agent=false
    features.multi_agent_v2=false
    

    But the trace above shows native subAgentActivity firing anyway for an openai_token_sharing session, on Codex CLI 0.156.1. The PR's own verification notes flagged this gap: "A live managed-provider request has not been verified." That appears to be exactly what's biting here — the config override doesn't seem to actually suppress native subagent dispatch in this Codex CLI build, so the original failure mode is unchanged.

    Environment

    • T3 Code server: 0.1.5 (stable)
    • Codex app-server/CLI: 0.156.1
    • OS: macOS (darwin arm64)
    • Auth route: Connect with ChatGPT (subscription sharing), model gpt-6-astra
    • Desktop app + local server, both version 0.1.5

    Produced by Claude (Opus 4.8) via T3 Code triage (t3 triage).

  4. juliusmarminge commented on Nov 12, 2026

    @juliusmarminge
    Member

    I tried to reproduce this on the current build and could not.

    Commit 3b732f4 (origin/main), real Codex CLI 0.167.2. A full T3 run under Connect with ChatGPT (openai_token_sharing) is not possible on a worker (no first-party ChatGPT OAuth), so the provider rejection itself was not seen. Root-cause check: ran codex exec with an empty CODEX_HOME, --ignore-user-config, --model gpt-6-astra, and main's exact managed launch args from CodexManagedRuntime.ts (model_provider=openai_token_sharing, features.api_key_model_discovery=true, features.multi_agent=false, features.multi_agent_v2=false). The provider base_url pointed at a loopback mock that served Codex's bundled model catalog (gpt-6-astra has multi_agent_version "v2") and recorded requests. The first /v1/responses request carried an additional_tools input item with a collaboration namespace: spawn_agent, wait_agent, send_message, list_agents, followup_task, interrupt_agent. Re-running with -c agents.enabled=false added removed that namespace and nothing else. A run with a dummy MCP server also offered the agent tools but no tool_search, so the tool_search variant in one comment was not reproduced. Failing test: one assertion that the managed launch args include agents.enabled=false, added to the existing case in apps/server/src/provider/CodexManagedRuntime.test.ts. All 4 parameterised cases fail on that assertion only.

    Closing as not planned. If it still happens for you, reply with the exact steps and the version you are on and it will be looked at again.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions