Skip to content

[Bug]: Orchestration works with omp but stalls when worker is Codex #12253

Closed

Description

@AmethystLiang

Operating system

Other

Orca version

Unknown

Details

Short summary: Orca orchestration progresses when the worker agent is omp, but stalls when the worker is Codex.

What happened?
Orchestration with an omp worker continues successfully; the same (or similar) flow with a Codex worker gets stuck and cannot continue.

Steps to reproduce

  1. Run an orchestration flow with an omp worker — completes / continues.
  2. Run a similar flow with a Codex worker.
  3. Observe stall / inability to continue.

Expected behavior
Orchestration should progress for Codex workers similarly to omp (dispatch, inject, wait for completion).

Actual behavior
Codex worker path stalls; omp path works.

Additional context
Possibly related:

If this is a duplicate of the inject/Enter path, close as duplicate and link the root cause.

Please comment with: Orca version, OS, orchestration skill/version, and where it stalls (before dispatch / after inject / waiting forever).

Activity

  1. AmethystLiang commented on Aug 3, 2026

    @AmethystLiang
    ContributorAuthor

    Can you share your Orca version & OS version?

  2. pablontiv commented on Aug 3, 2026

    @pablontiv

    Worth ruling out a permissions cause before this is triaged as a Codex inject problem.

    At v1.4.167 a supervised worker inherits its permission flags from one global setting, settings.agentDefaultArgs. When that setting is empty for the agent, which is what Manual mode writes (src/shared/tui-agent-permissions.ts:152), the worker launches with no permission flag and stops at the agent's own approval prompt on its first tool call. Codex under Manual mode gets no --dangerously-bypass-approvals-and-sandbox and will sit on an approval.

    This is easy to misread as an inject failure, because every dispatch-side signal still looks correct: the terminal exists, the agent is recognized, the preamble was accepted. The input did land. The agent is just waiting on a question.

    Two checks that separate it from the --inject path:

    1. Read the worker terminal buffer. An approval prompt, or a literal manual mode on line, points at permissions rather than at inject.
    2. Check Settings > Agents. If Agent Permissions is on Manual, every orchestration worker on this machine is launching interactive.

    If the buffer shows an approval prompt, this is #12360 with a different agent. If the buffer is empty, #11343 is the better lead.

  3. nwparker commented on Sep 19, 2026

    @nwparker
    Contributor

    The available report does not establish an OMP failure: it compares an OMP worker with a Codex worker, and the only technical reproduction comment points to Codex permission/approval mode or the separate Enter/injection path. Current Orca worker-start tests explicitly cover launch preference preservation and permission-gated terminal states.

    I am leaving this open for a concrete current-main repro (Orca version, host, worker permission mode, and the worker terminal transcript). Without that, changing OMP integration would risk hiding a Codex approval prompt or masking a separate inject bug.

  4. nwparker commented on Sep 19, 2026

    @nwparker
    Contributor

    Current-main audit is complete. This report does not establish an OMP failure:

    • worker-start correctly forwards the configured agent launch args.
    • OMP proceeds because its default approval mode is yolo.
    • Codex in Orca Manual mode launches without the bypass flag and waits at its own approval prompt; the worker is alive, but no human is attached.
    • Five focused suites (55 tests) cover permission resolution, Codex turnStart=permission, prompt submission, and worker settlement.

    This is a Codex permission-scope variant of #12360, with #11343/#10666 covering separate Codex input/trust races. Silently forcing Codex to yolo would violate Manual mode, so there is no safe OMP change to make here.

  5. nwparker commented on Sep 19, 2026

    @nwparker
    Contributor

    Closing as diagnosed Codex permission behavior / variant of #12360, not an OMP defect.

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

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions