Skip to content

Managed Codex sandbox: deterministic runner fails across ds/cx/ag backends #171

Description

@rbinar

Environment: Codex managed workspace-write sandbox on macOS; cli-dispatch 4.27.0; target is already a linked worktree (in-place mode). Reproduction used cli-dispatch-run with an explicit --cwd and --verify commands.\n\nObserved failures before any worker product edit:\n\n1. Default session root: mkdir under ~/.cache/cli-dispatch/sessions fails with Operation not permitted. Setting XDG_CACHE_HOME to /private/tmp works around this.\n2. ds backend: configured/default model deepseek-v4-pro is rejected with claude-code:unrecognized_model. Session f4d0ac14-a3f5-4298-a135-1e47e8464541.\n3. cx backend: cx-stream line 382 tries /dev/fd/63 and fails with Operation not permitted under the managed sandbox. Session dir cx-1787914515-72240.\n4. ag backend model propagation/config: cli-dispatch-run --model claude-sonnet-4-6 initially reached ag-stream as gemini-3.6-flash-high because source_config sources config after environment assignment even though its comment says env wins. With CLI_DISPATCH_CONFIG pointed to a nonexistent file, Sonnet remained selected, but model validation still warned it was unlisted because agy models emits tab-separated slug plus display name and model_key compares the entire line. The run then failed conversation discovery (discover-fail-20260828135753-2704), alongside denied writes under ~/.gemini.\n\nExpected: deterministic runner should either work in this managed sandbox or fail once with a clear actionable compatibility verdict; explicit --model must win over config and a valid tab-listed model slug must validate.\n\nNo product files were changed by these failed runner attempts. A direct claude -p --model sonnet fallback completed the task.

Activity

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

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions