[orbitingflea]

Problem

A Stop hook that works in the agy CLI does nothing in T3's Antigravity provider. Google's ACP server (agy_acp_server.par) registers hooks.json Stop hooks as a session-end hook, so they run once at teardown and ignore {"decision":"continue","reason":...}. The server also gives hooks different input from the CLI: terminationReason is UNSPECIFIED instead of NO_TOOL_CALL, and transcriptPath points to a SQLite database instead of JSONL. Google's latest release, 1.4.3, still has the same code.

Fix

• Pin 1.4.3 (antigravityRelease.ts), with new archive hashes and sizes for all six platforms.
• Patched copy for chat only (antigravityChatPatch.ts, AntigravityInstallation.chatExecutable). T3 copies the installed PAR to <managed>/patched/<releaseId>-r<rev>/ and patches four STORED entries at their original length: hooks.py, server.py, and their .pyc files, whose magic is broken so the sources load. It updates both CRC fields and publishes the copy with a rename. Each original entry is checked by sha256 first. On any mismatch, unknown version, non-Linux platform, or I/O error, T3 logs a warning and launches the original. Google's files are never changed.
• Only orchestration chat uses the copy (AntigravityDriver.ts). Sign-in, model refresh, install validation, and text generation keep the original server.
• What the patch does:
  • Stop hooks run at the end of every turn, and continue keeps the agent working inside the same session/prompt.
  • The hook input matches the CLI: NO_TOOL_CALL and the CLI's other reason names, plus a temporary JSONL transcript with the reply. After a continue, the transcript also holds the CLI's Stop hook blocked termination system message. Hooks written for the CLI, such as our agy.py adapter, work unchanged.
  • The continue reason reaches the client as an agent_thought_chunk tagged _meta["t3.stop_hook"], between the judged reply and the continued one.
  • A cancel kills a running hook command, so the cancel no longer waits up to the hook timeout.
• T3 display (adapter.ts, AntigravityProtocol.ts, AntigravityAdapterV2.ts). An Antigravity-only flavor hook turns the tagged chunk into a system_notice item. The user sees the original reply, then a "Stop hook feedback: …" line, then the rewrite as a new assistant message, all in the same turn.

Verification

• Unit tests (141 pass across 5 files):
  • The patcher on a synthetic ZIP: patched result, CRCs, size, untouched original, three fallback cases, cache reuse, unknown version.
  • The driver's chat and original selection.
  • The adapter's notice and segment split.
  • The marker extractor.
• vp run check:changed passes.
• Real run on 1.4.3 linux-x64, with the patched PAR built through chatExecutable and driven by a raw ACP client:
  • initialize reports 1.4.3 and 6 models;
  • a hook that asks to continue once gives HELLO-1, the tagged notice, PINEAPPLE-OK, then end_turn in a single prompt, and the second hook call allows;
  • agy.py's claude_payload() reads the reply and stop_hook_active correctly on both calls;
  • a cancel while a hook sleeps returns cancelled in about 0.1 s, and the next prompt works.

Not yet done:

• Not run in the T3 UI, and no sandbox check yet (sandboxes mount the host runtime read-only).
• linux-arm64 goes through the same byte checks but was not run.

Known limits

• The patch uses private SDK attributes (_processor.step_queue, _client_cancelled). A new agy version fails the hash check and runs unpatched until the table is updated.
• executionNum is the zero-based continue count. The CLI's numbering is undocumented (its docs example shows 1), and agy.py does not read it.
• Old patched copies, about 0.93 GB each, stay until a full remove. Old versions/ directories already behave the same way.

Done with Claude Opus 5.5 in Claude Code (via T3 Code).

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

[orbitingflea]

Closed: the fix moves to a local patch script; T3 stays unchanged.