Skip to content

Give Codex a usable sandbox in the consolidated loop - #96

Merged
iamkayleb merged 1 commit into
mainfrom
fix/codex-sandbox-in-consolidated-loop
Sep 11, 2026
Merged

iamkayleb merged 1 commit into
mainfrom
fix/codex-sandbox-in-consolidated-loop

Conversation

@iamkayleb

Copy link
Copy Markdown
Owner

Codex defaults to workspace-write, which makes it sandbox with bubblewrap. bwrap cannot create a loopback interface on a GitHub-hosted runner:

bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted

Every shell command then fails before it runs. Codex reported this exactly — "the workspace shell fails before any command runs ... I therefore can't inspect the existing availability API, run tests, or create the required verified commit safely" — and its session recorded 0 commands executed and 0 files changed.

From outside that reads as the agent producing nothing: the job is green, the ledger closes the task, and the only commit is a telemetry timestamp. Two evaluation rounds were scored against Codex on this basis before the session transcript was read.

The runner is an ephemeral single-tenant VM and is already the isolation boundary, so Codex's inner sandbox buys nothing here. Pass danger-full-access on both Codex paths in the consolidated loop, matching what the runner already documents for CI use.

Claude-Session: https://claude.ai/code/session_01FC8XoyssN5v6hQCTtcjoB5

Workflow Source

Started from:

  • GitHub issue: #
  • Direct PR / remote GitHub work
  • Local Codex/user request
  • Automation run
  • Review follow-up from PR #
  • Sync / maintenance campaign
  • Dependabot or dependency update
  • Do not automate

Automation intent:

  • Verifier should review this
  • Keepalive may manage this PR
  • Human-only unless checks fail

Notes:

Summary

Testing

Codex defaults to `workspace-write`, which makes it sandbox with bubblewrap.
bwrap cannot create a loopback interface on a GitHub-hosted runner:

    bwrap: loopback: Failed RTM_NEWADDR: Operation not permitted

Every shell command then fails before it runs. Codex reported this exactly —
"the workspace shell fails before any command runs ... I therefore can't
inspect the existing availability API, run tests, or create the required
verified commit safely" — and its session recorded 0 commands executed and 0
files changed.

From outside that reads as the agent producing nothing: the job is green, the
ledger closes the task, and the only commit is a telemetry timestamp. Two
evaluation rounds were scored against Codex on this basis before the session
transcript was read.

The runner is an ephemeral single-tenant VM and is already the isolation
boundary, so Codex's inner sandbox buys nothing here. Pass
`danger-full-access` on both Codex paths in the consolidated loop, matching
what the runner already documents for CI use.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FC8XoyssN5v6hQCTtcjoB5
@iamkayleb
iamkayleb merged commit 9be71e7 into main Sep 11, 2026
20 of 23 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants