Skip to content

docs(dispatch): note that Grok lanes need an unsandboxed Codex parent - #82

Open
smashru wants to merge 1 commit into
ericlitman:mainfrom
smashru:docs/codex-parent-grok-sandbox
Open

smashru wants to merge 1 commit into
ericlitman:mainfrom
smashru:docs/codex-parent-grok-sandbox

Conversation

@smashru

@smashru smashru commented Sep 23, 2026

Copy link
Copy Markdown

What changed

provider-dispatch.md already says the parent's tool sandbox controls whether a child CLI can reach its credentials and network. This adds the Grok-specific case, which fails in two ways that don't look like a sandbox problem from the receipt alone. On macOS, a Codex parent running any sandbox cannot host a Grok lane:

  • read-only, or workspace-write without network. grok models cannot reach the API and falls back to its built-in model list. The preflight then records unavailable-model for a current model such as grok-4.7, which reads like an account or model problem.
  • workspace-write with network. Grok exits with "could not apply the 'workspace' sandbox profile ... Refusing to start with its protections missing", because macOS sandboxes do not nest. The receipt records child-failed.

The note says a Codex parent that routes roles to grok:* must run with danger-full-access, chosen deliberately in the user's profile. Grok's own sandbox still confines the lane. The note keeps the existing rule that a blocked CLI is a dropout, never a reason for the parent to escalate itself.

Docs only. No runner or skill behavior changes. Found while verifying #81.

Verification

  • Bun tests, strict typecheck, static invariants, and plugin validation pass. Static invariants and git diff --check pass. No code or manifest changed.
  • The exact candidate is installed in every affected harness. Not applicable to a docs-only change. The behavior it describes was observed with Open Pstack 1.4.1 plus fix(runner): let headless Grok writers run shell commands #81 installed in Codex.
  • The changed behavior passes from each real user surface.
  • The installed version, action, and observed result appear below.

Live evidence (macOS, codex-cli 0.155.1, grok CLI 1.0.34, grok:grok-4.7@xhigh isolated-write lane launched by a codex exec session through the installed pstack-runner):

Codex sandbox Receipt Grok evidence
workspace-write, no network exit 69, unavailable-model preflight listed only grok-4.6, grok-4.5, plus "HTTP export failed: network error"
workspace-write, network on, ~/.grok writable exit 70, child-failed "sandbox initialization failed: Operation not permitted", "could not apply the 'workspace' sandbox profile"
danger-full-access exit 0, complete, grok-4.7-build verified lane output GIT_STATUS: ## HEAD (no branch)

🤖 Generated with Claude Code

Under a sandboxed Codex parent on macOS, Grok either cannot reach its
API and reports only built-in models, or refuses to start because its
own sandbox profile cannot nest inside Codex's. Name both failure
signatures and the parent profile a Grok-routing Codex setup needs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@mergify

mergify Bot commented Sep 23, 2026

Copy link
Copy Markdown

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

  • Queue this pull request

@greptile-apps

greptile-apps Bot commented Sep 23, 2026

Copy link
Copy Markdown

RetriggerConfidence Score: 5/5

The documentation-only change appears safe to merge.

Summary

Documents the macOS sandbox constraints for Grok lanes launched by a Codex parent.

  • Explains the distinct unavailable-model and child-failed outcomes produced by restricted parent profiles.
  • Requires users to deliberately start the Codex parent with danger-full-access while retaining Grok's child sandbox.
  • Preserves the rule that blocked provider CLIs remain dropouts rather than triggering automatic privilege escalation.

Reviews (1) · Last reviewed commit: "docs(dispatch): note that Grok lanes nee..."

@openswebot openswebot Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

✅ Open SWE Review: No issues found

Open SWE reviewed this PR and found no potential bugs to report.

Open in Web

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.

1 participant