ZeroClaw
ZeroClaw proposes Antigravity delegation that treats denied actions as failure
ZeroClaw's proposed Antigravity coding tool checks structured output instead of trusting exit code zero. The contributor reports that headless agy can call a run successful even when it denies an action. The adapter stays off by default and does not turn a validated starting directory into a sandbox.
A ZeroClaw contributor has proposed an opt-in coding-delegation tool for Antigravity CLI. The request is about letting the existing agent hand work to another coding program, not adding a new model provider. Its most useful finding concerns the definition of success: in the author's Linux probes, agy could exit zero and return a SUCCESS status while also reporting that a requested action had been denied.
The proposed adapter therefore reads structured output and includes the denial list in its result decision. It is a small but important example of why an agent harness cannot treat every successful process exit as completed work. The task-level outcome has to survive the trip across the external program's own semantics.
The facts
- PR 11076 remains an open proposal and introduces an optional agy_cli section that defaults to disabled. - Success requires exit code zero, structured status SUCCESS and an empty denied_actions list. - Malformed or status-less output fails closed rather than being reported as a successful coding task. - The adapter reuses shared authorization, timeout, output bounds and working-directory validation. - Tool-owned prompt and model arguments follow operator extra arguments, and a dash-prefixed prompt is passed as text rather than a new flag. - The author tested real agy 1.2.9 on Linux but did not verify macOS, Windows or other agy versions.
Why it matters
Delegating to a coding CLI introduces another permission system. ZeroClaw can decide whether to launch it and where it starts, but the external program decides what happens after launch under its own configuration. Reporting a denied action as success hides precisely the information an operator needs. The proposal makes that boundary visible without claiming to replace the external tool's sandbox or silently changing the existing Gemini CLI integration.
Current
Checked September 24, 2026. The inspected release baseline is v0.8.5. The primary source is open proposal. Package availability and development-source state are separate: a tag does not establish that an open proposal is usable, and a merge does not establish that every installation has received it.
Evidence
Primary: zeroclaw-labs/zeroclaw PR #11076 (https://github.com/zeroclaw-labs/zeroclaw/pull/11076). Supporting context: zeroclaw-labs/zeroclaw issue #11075 (https://github.com/zeroclaw-labs/zeroclaw/issues/11075); zeroclaw-labs/zeroclaw PR #11076 — implementation diff (https://github.com/zeroclaw-labs/zeroclaw/pull/11076/files). Source pages and available diffs were inspected; test and live-account results remain attributed to their authors, not independently repeated measurements.
Operator take
Review the external CLI's effective permissions as well as ZeroClaw's launch policy before any trial. A workspace-contained starting directory is not proof that all subsequent writes stay inside that directory. Check a genuinely denied operation, a plain response, malformed output and a prompt beginning with a dash. Preserve the reported test revisions: the live exercise ran at an earlier commit, while the follow-up only adds stderr handling and focused tests. Do not flatten those into one universal platform qualification.
Caveat
The Google product-transition assertions in the contributor's motivation were not independently validated and are not treated here as a new Google announcement. The evidence supports the proposed ZeroClaw adapter and the author's bounded agy experiment. CI lacks the real binary, the full workspace suite was not run locally, and broad permission flags remain an operator responsibility rather than being blocked by warning messages.
The Google product-transition assertions in the contributor's motivation were not independently validated and are not treated here as a new Google announcement. The evidence supports the proposed ZeroClaw adapter and the author's bounded agy experiment. CI lacks the real binary, the full workspace suite was not run locally, and broad permission flags remain an operator responsibility rather than being blocked by warning messages.