Skip to content

[Bug]: Antigravity supporting turn fails executor startup on t3-code roots notification with Bad Request #36084

Closed

Description

@seankoji

Note

Written by GPT-6.1 Sol on behalf of @seankoji

Before submitting

  • I searched existing issues and did not find an exact duplicate in the searches performed.
  • I included enough detail to investigate the problem.

Area

apps/server

Steps to reproduce

Observed sequence on 13 November 2026, at 18:51 AEDT, UTC+11. This is an existing incident, not a newly run or proven deterministic reproduction.

  1. Start a fresh Antigravity session using gemini-4-hydrogen-high, provider options {}, and runtime mode full-access. The assigned task was read-only.
  2. Send an initial text-only prompt and wait for it to complete.
  3. Send a supporting message in the same session with three small WAV files and two small MP4 files, each below the native attachment size limits.
  4. All five uploads return HTTP 204. Stored-file byte checks match the prepared inputs.
  5. The supporting turn fails while constructing the executor, with the error below.

The observed input was a mixed WAV/MP4 message. A WAV-only synthetic reduction has not been run. The evidence does not establish that media is necessary to trigger the failure.

Expected behavior

After the initial text turn completes, the supporting turn should construct its executor and complete the T3 MCP startup/roots-notification exchange so the provider can process the message.

Actual behavior

Executor construction fails while sending notifications/roots/list_changed to t3-code, reporting HTTP 400 / Bad Request. The ACP request returns successfully carrying the host error, and T3 records the turn as failed.

A retained provider-local conversation receipt contains the complete original WAV bytes before the executor failure. This rules out alleging that these WAVs never reached provider-local storage. It does not prove backend receipt or native hearing. No audio-format or model-capability rejection was established.

Impact

Major degradation or frequent failure.

The affected supporting-media workflow cannot proceed to a usable result despite a successful first text turn and successful uploads. This report does not establish failure across all T3 sessions or providers.

Version or commit

T3 Code server 0.1.10-nightly.20261113.4616, verified as the actual running server during the existing investigation.

Environment

  • Linux x64.
  • Antigravity ACP provider, model gemini-4-hydrogen-high. High is encoded in the model ID; provider options were {}.
  • Runtime mode full-access, with read-only task scope.
  • Managed installation record and archive identify Antigravity 1.3.0. This is installation evidence only. The exact executed binary, per-run handshake version and any binary override are unknown.
  • Browser/client and Node/Bun versions were not retained.

Logs or stack traces

Relevant retained ACP timeline, UTC on 13 November 2026, equivalent to 18:51 AEDT:

07:51:12.365  initial session/prompt started
07:51:17.426  initial session/prompt succeeded
07:51:20.255  supporting session/prompt started
07:51:20.296  ACP response returned carrying the host error

failed to construct executor: MCP load failed for t3-code: sending "notifications/roots/list_changed": Bad Request

Screenshots, recordings, or supporting files

None attached. This report contains only redacted observations.

Workaround

None verified. A later resumed text turn added no fresh media to its new stored input. Earlier backend-history retention is unknown, so that resume does not verify recovery of the supporting-media workflow.

Investigation limits and verification criteria

The exact HTTP 400 cause remains unresolved. Existing static inspection found a T3 roots-change handler and multiple HTTP admission paths that can return 400. No retained rejection headers/body identifies which path rejected this exchange. There is no evidence here for a specific missing-header defect or an unsupported-notification defect.

The installed MP4 path is separately gated by reported Antigravity version, platform and resolved model. This report does not establish native video delivery or a separate video-gate bug.

Useful verification would show a clean executor start for an initial text turn followed by a supporting turn on a verified runtime. If the rejection recurs, redacted method, negotiated protocol, session-header presence, HTTP status and rejection reason would make it actionable without private prompts, credentials or media. Native media perception needs separate verification after startup succeeds. No patch or verified fix is reported.

Related public context

Activity

  1. SrirajBehera commented on Nov 16, 2026

    @SrirajBehera

    Also reproduces on macOS desktop nightly, after a cross-provider handoff (Claude → Codex → Antigravity)

    Same failure as this issue, but the trigger differs from the fresh Antigravity session in the original report, so it may help narrow things down.

    What happened
    The thread's third provider was Antigravity (Gemini flash, high; I don't have the exact model ID). It failed on a turn after it had been handed work from a Codex session. The UI showed:

    The MCP server 't3-code' failed to initialize: sending "notifications/roots/list_changed": Bad Request. Fix the MCP server, or remove it from your configuration and start a new session, to continue.
    Provider error: Antigravity stopped this turn because of an error.

    Sequence

    1. A thread started on Claude (Opus 5.5, medium; Claude Agent SDK).
    2. The user handed off to Codex (Sol), roughly 400 provider events with no error.
    3. The user handed off to Antigravity (Gemini flash, high), the second handoff. The first Antigravity event is at 2026-11-15 21:12:44Z.
    4. The last session/prompt at 21:33:32Z started and returned "succeeded" 12ms later, carrying the host error. That was the failing turn. The local time of the failure was 3:03 am (UTC+5:30, assumed).

    The provider event log records only field counts for the request and result, so I don't have the HTTP response headers or body for the 400. The ACP response comes back as success, as in the original report, and T3 then marks the turn failed.

    Environment

    • macOS (darwin arm64), desktop app, nightly channel, connected to the local server.
    • Server 0.1.10-nightly.20261115.4736. This is newer than the 20261113.4616 build in the original report, so the problem still reproduces on later nightlies.
    • Antigravity version and exact model ID are unknown.

    Differences from the original report

    • macOS rather than Linux.
    • Reached through a handoff from another provider, not a fresh Antigravity session.
    • No media attachments were involved as far as I know, so the failure doesn't depend on a mixed WAV/MP4 message.

    Not tried: whether a brand new Antigravity thread works, or whether a second turn on the same Antigravity session works after a restart.

    Written by Claude Sonnet 5.5 via Claude Code, running the T3 Code triage playbook on the user's behalf.

  2. loumizhu commented on Nov 16, 2026

    @loumizhu

    Maybe it's the same
    my agent recommended I post this

    "Since nightly 20261113, Antigravity (ACP) turns die mid-run with failed to construct executor: MCP load failed for t3-code: sending "notifications/roots/list_changed": Bad Request. The t3-code MCP endpoint returns 400 on the client's roots/list_changed notification. It reproduces on every Antigravity account. 26 runs were killed in 3 days. It never happened before 20261113."

  3. roptant commented on Nov 17, 2026

    @roptant

    Also reproduces on Linux with stable T3 0.1.10 and Antigravity, so this isn’t limited to nightly builds or macOS.

    Affected threads fail during executor startup with:

    failed to construct executor: MCP load failed for t3-code:
    sending "notifications/roots/list_changed": Bad Request
    

    Once a thread reaches this state, retrying or telling it to continue fails at the same initialization step, so the thread cannot recover on its own.

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