Skip to content

advisor-board gemini seat DEGRADED 'Gemini broker native exit without an accepted review' with no exit code or cause #1929

Open

Description

@ViperJuice

Summary

On dev0 on 2026-11-08 (phase-loop 0.7.27, agy 1.5.3), the gemini:gemini-4-argon:medium:alternative-approach seat of a plan-tier advisor board returned DEGRADED with detail Gemini broker native exit without an accepted review and empty text. The seat-mode line reported it as jailed -- full tools inside its per-seat jail. In the same run, staged from the same tree, the grok seat produced a usable review and the native Claude fill bound.

Reproduction shape

  • phase-loop advisor-board <~113 KB plan bundle> --landing-tier plan --native-leg claude=<fill-dir> --json
  • Run from a full shallow clone of main at c4a9abc, after a partial-clone staging failure (agent-harness#1419).
  • Bounded monitoring, with no timeouts_by_leg. The whole board took about 20 minutes.

What is missing

The result carries no exit code, signal, stall marker or --print-timeout marker for the agy process. So it is not possible to tell whether this was a broker refusal, a "timeout waiting for response" exit at the input-scaled agy budget (the third clock described in the advisor-board skill), or a native crash. The run-scoped sandbox was cleaned up, which left no leg log behind. Related: agent-harness#1823 (degraded seats keep no per-provider exec argv or exit code).

Ask

  • Record the agy exit status and the last broker state in the leg's detail, or in a retained private leg log.
  • If this was the agy --print-timeout budget, name it as that, not as a generic native exit.

Found during the plan review of agent-harness#1925 and agent-harness#1926. That review was completed with a T3 Antigravity gemini-4-argon delegate in place of the seat.

Activity

  1. ViperJuice commented on Nov 8, 2026

    @ViperJuice
    CollaboratorAuthor

    Second occurrence on dev0 on 2026-11-08, about an hour later, during the plan-tier board for agent-harness#1927 (head 3309dd4, installed phase-loop runtime, bounded monitoring, no timeouts_by_leg).

    • Seat: gemini:gemini-4-argon:medium:alternative-approach. Seat mode: jailed -- full tools inside its per-seat jail.
    • Result: DEGRADED, detail Gemini broker native exit without an accepted review, with empty text. The president input recorded it only as F031: … unusable (DEGRADED).
    • New data point: the board's stderr also showed this notice in the same run:
      seat gemini [gemini_credential_awaiting_refresh]: the agy login has under 10 minutes left and has not expired; agy renews a login that has expired, so the seat waits until this login's expiry and then renews it once (418 s left; until 15:51:02 UTC).
      The seat's verdict file was written at 15:51 UTC, which matches the login's expiry. The native exit therefore looks tied to the wait-then-renew path (agent-harness#1407), not to the review itself. The result does not say so, and still carries no exit code or broker state.
    • Same pattern as before: the grok seat and a native Claude fill produced usable reviews from the same staged tree. The board ran from a full shallow clone, because the shared checkout is a partial clone (agent-harness#1419).

    The lens was covered outside the board by a T3 Grok 4.7 alternative-approach delegate. No credential, config or agy state was changed.

  2. ViperJuice commented on Nov 8, 2026

    @ViperJuice
    CollaboratorAuthor

    Third occurrence on dev0 on 2026-11-08, during the plan-tier board for agent-harness#1931 at head 2642b30.

    • Versions: phase-loop 0.7.27, agy 1.5.3. Bounded monitoring, no timeouts_by_leg. Run from a standalone full-blob shallow clone, inside a T3 Claude Code session.
    • Seat: gemini:gemini-4-argon:medium:alternative-approach. Seat mode: jailed -- full tools inside its per-seat jail.
    • Result: DEGRADED, "Gemini broker native exit without an accepted review", empty text. It failed about 10 s after the seat-modes receipt, so it was fast, not a stall. As before, there was no exit code or cause.
    • Replacement: the seat was filled by a T3 delegate for that review.
  3. ViperJuice commented on Nov 8, 2026

    @ViperJuice
    CollaboratorAuthor

    Fourth occurrence on dev0 on 2026-11-08, during the plan-tier delta re-review of agent-harness#1931 at head 72c5b04.

    • Versions: phase-loop 0.7.27, agy 1.5.3. Bounded monitoring, no timeouts_by_leg. Run from a standalone full-blob shallow clone of the PR head, inside a T3 Claude Code session, with a ~126 KB bundle and --native-leg claude=<fill-dir>.
    • Seat: gemini:gemini-4-argon:medium:alternative-approach. Seat mode: jailed -- full tools inside its per-seat jail.
    • Result: DEGRADED, "Gemini broker native exit without an accepted review", empty text. The leg's verdict record was written in the same minute as the seat-modes receipt, so it failed fast, not at the agy --print-timeout budget. As before, there was no exit code or cause. In the same run, Grok produced a usable review and the native Claude fill bound.
    • Replacement: a T3 Claude Opus 5.5 delegate on a separate account filled the seat for this review.
  4. ViperJuice commented on Nov 9, 2026

    @ViperJuice
    CollaboratorAuthor

    Fifth occurrence on dev0 at 2026-11-09T00:04Z, during the plan-tier delta re-review of agent-harness#1931 at head 93632e9.

    • Versions: phase-loop 0.7.27, agy 1.5.3. Bounded monitoring, no timeouts_by_leg. Run from a standalone full-blob shallow clone of the PR head, inside a T3 Claude Code session, with a ~76 KB bundle and --native-leg claude=<fill-dir>.
    • Seat: gemini:gemini-4-argon:medium:alternative-approach. Seat mode: jailed -- full tools inside its per-seat jail.
    • Result: DEGRADED, "Gemini broker native exit without an accepted review", empty text. The verdict record was written about 11 s after the seat-modes receipt, so it failed fast, not at the agy --print-timeout budget. As before, there was no exit code or cause. In the same run, Grok produced a usable review and the native Claude fill bound.
    • Replacement: this time a T3 Antigravity delegate on the same model (gemini-4-argon, thinking high) filled the seat and returned a usable, source-grounded review. So the model and the account are healthy; the failure is specific to the board's jailed agy route.
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