Repository navigation
test(antigravity): verify Windows hook payload delivery through launcher - #21666
Conversation
|
Independent verification (2026-09-20):
Verdict: approve the test-only change for the Windows hook-delivery contract. Keeping it draft is reasonable if the author wants to pair it with a production launcher fix; no merge performed. |
ELI5 / evidence / trade-offs auditELI5: On Windows, hook data reaches the launcher even when paths contain spaces or Unicode characters. Proof: Windows CI passed 477 tests with both new payload cases; visual proof is not applicable to this test-contract change. User-facing trade-off: This proves payload delivery, not every downstream hook implementation. This note scopes the claim to the evidence above; it does not claim untested remote, SSH, WSL, or vendor combinations. |
Native Windows verification update (2026-09-20)Fetched PR head This is the best low-risk shape for this PR: it changes tests and CI environment setup only, leaves the production launcher unchanged, and exercises the real Verdict: low risk; approve as a test-contract PR. |
|
This run was cancelled 🛑 The workflow was cancelled before completion. Please check the link below for details. |
* fix(cli): accept agy alias for supervised worker starts * test(antigravity): exercise Windows hook payload delivery in CI (#21666)

ELI5
A hook can start successfully and still lose the message it was supposed to deliver. These tests check the actual message reaching Orca from Antigravity's Windows hook scripts.
What Changed
Extend the existing Windows payload-delivery suite with Antigravity PreToolUse for ordinary profile paths and paths containing spaces. The latter exercises the encoded PowerShell fallback. Both require the entire Unicode payload to arrive unchanged at a local listener and the hook to return exactly
{"decision":"ask"}without extra output. The existing Windows CI step already executes this suite; it now explicitly setsORCA_BACKGROUND_LAUNCH=1.Why
PR #19272 proposes removing an execution-policy command from batch hook launchers but has no Windows process evidence. Existing Antigravity tests invoke wrappers directly, bypassing the encoded launcher. Reusing the delivery harness tests the real process chain instead of merely comparing generated strings. This draft starts with the existing launcher so CI can establish a baseline before any production change.
Linked Issue
Refs #17672, #19829; verification follow-up for #19272. No status or lifecycle issue is claimed fixed. Stacked on #21664.
Visual Proof
N/A — automated process-delivery tests only; no product UI change. Windows CI’s test step passed with the unchanged launcher; no successful Antigravity generation is claimed.
Testing
macOS collection passed: one platform-independent check passed, three Windows execution cases skipped. Node typecheck, changed-code quality and staged lint/format checks pass. Windows package job 105895776713 completed successfully. Its log confirms all four delivery-suite tests passed, including Antigravity ordinary path (488 ms) and path with spaces (780 ms). The full Windows boundary step passed 477 tests, with 22 skipped. Both new cases ran; neither was skipped. These launch actual managed scripts with synthetic hook JSON; they do not run the agy binary on Windows or validate antivirus behavior, live hooks or sidebar transitions.
Review
No production launcher change yet. Temporary hook install and loopback listener use the existing isolated test harness. No global authentication or hook settings are changed.
Agent skill upstream boundary
Checklist