[Tericerice]

Summary

The Antigravity adapter spawned agy for every turn.start and, on every #completeTurn including successes, closed its stdin, waited for exit, and signalled a lingering process. Continuing a conversation therefore repaid the CLI, Language Server, and customization startup cost on each Turn. The official headless protocol keeps one stream-json process and reads the next user message from the same stdin, which this PR adopts.

Fixes #535

Behaviour

• A successful Turn keeps its driver. The process, its Question bridge, and its log move to the Session; the next Turn writes to the same stdin. --conversation is not needed on that path, and --add-dir/--log-file keep pointing at the same directory and log.
• The Question bridge is rebound, not rebuilt. rebind moves the Turn identity, the owning-Turn check, the emit callback, and the duplicate-question guard to the new Turn, so a question from an earlier Turn can neither be attributed to the current Turn nor answered. A disposed or busy bridge refuses to rebind.
• Replayed native output is filtered. agy re-emits the previous successful result and earlier steps when a new stdin line arrives; only a result whose num_turns advanced (or whose payload differs) and steps at or beyond the new Turn's first index reach the Host, so history records each Turn once.
• Every other outcome releases the driver — failure, cancellation, Session close, a different model/effort/permission selection, or a process that exited or lost writable stdin. Release stops the bridge, ends stdin, and escalates SIGTERM → SIGKILL within the existing grace and cleanup-wait bounds.
• Real errors stay errors. A native failure, including a num_turns that did not advance, still fails the Turn; nothing is reported as success merely because text arrived.

Tests

• The adapter test stand-ins now behave like the real CLI: they answer each stdin line instead of exiting after one Turn. This is what exposed the replayed-result filtering.
• New test: three Turns run on one PID, and model.select followed by a fourth Turn starts a new PID, with all four Turns present in history.

Limits (unchanged from the issue)

• No statistical benchmark is claimed. The issue's measurements are a small number of observations with differing tool calls and upstream latency.
• Long idle periods, background subagents, and every native protocol version are not covered here; the driver is released on every non-successful outcome and on Session close, which bounds the exposure.

🤖 Generated with Claude Code

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}