[stamp-build]

^{Otter from open-agent-runtime opened this pull request through Stamp
View in Stamp →}

---

Grok 1.0.70 starts configured stdio MCP children in the ACP process HOME when sharing, whereas isolated sessions start them in the session directory. Both session-supplied and native user-configuration controls reproduce the difference. Session options cannot rule out configured servers, so the public Grok runtime remains isolated under auto and refuses explicit shared with option placement. No native runtime is enabled for sharing by this change.

The candidate adapter remains available only to explicit experiment compositions. It requires exact native 1.0.70 installation/initialize versions and fully qualified startup paths, rejects stdio MCP session options, and keeps custom inputs on the independent path. HTTP-only controls use disposable configuration. Windows drive-relative and root-relative paths are refused; pnpm relative PATH entries also prevent candidate sharing.

The shared ACP router accepts Grok's native subagent_spawned parent/child pair as ownership evidence. Completed subagent results only check that established relationship. Parent ownership is checked against the carrying root; accepted links are removed on owner disposal. Child requests have a one-second ingress deadline independent of any in-flight session/new; the frame count/byte limits remain bounded. Default OpenCode routing keeps its existing request pause. Regressions cover cross-root forged spawn/completion, grandchildren, completion before spawn and the OpenCode pending-new path. Twelve Linux native samples observed spawn 68-83 ms before child output, including five-second children, so this allowance does not wait for tool completion.

Real candidate sharing behavior and lifecycle controls move to grok-sharing-experiment.yml, outside default CI. Its steps depend on successful installation and upload only the Grok artifacts. Native lifecycle controls inject the actual diagnostic callback and require no conflict/expired diagnostics. No native initialization deadline is increased. The earlier Windows lifecycle timeouts remain missing successful evidence, not a passing result.

The manual shared-runtime evidence workflow gains actual child cwd/environment controls for Grok and OpenCode, including session/user stdio MCP, user/project hooks/plugins and terminal tools. Missing observations or differences fail the step and preserve evidence. All 24 Linux cohorts completed; each runtime/version comparison fails for the observed cwd mismatch. OpenCode's separate production withdrawal is already in PR #587. The evidence report, normalized selection and rebuild script retain all observations; no macOS/Windows result is inferred. The workflow can select exact OpenCode 2.2.3 or 2.2.5 and child-environment-only runs. It does not grant admission.

Validation:

• Full CI=1 pnpm run check on base ddec7e5: 2,948 tests and 18 mock behaviors passed. Package builds passed before rebasing the documentation fixes and release version bump from 2c67f52; runtime, test, workflow and evidence files are unchanged by those rebases.
• Before rebasing onto the OpenCode withdrawal, the unchanged shared machinery revision passed all 14 ordinary behavior backends. CodeBuddy was rerun with its correct pinned executable after a local path error caused an unavailable skip in the first batch; that skip was not counted as coverage.
• Grok candidate behavior: 17 passed; native shared/isolated lifecycle: 2 passed; MCP/hook environment controls: 8 passed. New shared timing assertions use the real router's diagnostics.
• Removing the production refusal causes all three public placement regressions to fail with actual shared placement. Four routing mutations from the original implementation were caught, and the requested additional adversarial cases pass.
• Ten Python evidence-selection checks passed. The selected 24 cohorts rebuild byte-for-byte from the raw observations, which have been shared with the reviewer. Both workflow files parse successfully.

Three-platform validation of the new native controls remains manual and pending. Public refusal and synthetic regressions remain in the required CI gate. Cindy owns the release; this change does not bump package versions.