orca/docs/reference/antigravity-native-chat.md

Antigravity terminal-backed Chat UI

Antigravity uses the existing experimental Chat UI over its terminal and saved transcript. It is not a structured-session provider. The opt-in default-chat setting applies, and users can return to the terminal.

Transcript contract

The execution host reads .gemini/antigravity-cli/brain/<conversation-id>/.system_generated/logs/transcript.jsonl. A hook-reported transcript path takes precedence. Existing WSL exact-path and host-isolation rules apply; a missing guest transcript must not fall back to a native host's same-named conversation.

The sanitized fixture in src/main/native-chat/__fixtures__/antigravity/ records these observed shapes:

• User input is USER_EXPLICIT/USER_INPUT with a USER_REQUEST wrapper.
• Planner responses carry text, thinking, or tool calls. Tool arguments are in args.
• Tool output is a MODEL record with a tool-specific type such as RUN_COMMAND or VIEW_FILE. It is not a TOOL_RESULT record.
• A planner step marked DONE can precede more tool work in the same user turn. The decoder therefore emits no turn-completion verdict. The existing host-owned hook/status system remains authoritative.

Full reads, incremental tails, and legacy journal imports reuse the same decoder. Unknown record types are skipped. Tool failures retain a nonzero exit code or ERROR status as an error result.

Mixed versions

New hosts advertise native-chat.antigravity.v1. A new client's chat transport checks the owning host before reading or subscribing to Antigravity. An older host gets the existing update-runtime error; a failed capability probe gets the existing read error. Neither becomes a perpetual first-transcript wait. No stream opcode or message-block type was added.

Verification limits

On macOS with installed agy 1.2.7, a resumed test conversation rendered from its real saved transcript, and a message submitted in the chat composer appeared in that transcript. Generation returned 401 ACCESS_TOKEN_TYPE_UNSUPPORTED. Hooks were disabled in the isolated app profile, so pane/session metadata was supplied as a fixture; this did not validate actual hook delivery or live status transitions. Tool rendering was checked separately with the sanitized fixture.

Five bounded real transcript samples supplied 309 tool/activity records; the corrected decoder preserved all 309. This is offline format evidence, not proof of a successful new model turn. Windows, Linux, WSL, SSH, paired-runtime execution, permission questions, and successful tool-generating turns still need live QA.