Skip to content

v1.55.2

Choose a tag to compare

@github-actions github-actions released this 09 Nov 06:37
· 314 commits to main since this release

Spend fewer tokens per task

This is the 1.55 release. v1.55.0 and v1.55.1 were tagged but never published, because test-only fixes were needed; their contents are all here.

Picking the model

  • Auto (best for the task) is a new choice in Settings → General → Default Models. It shows a What's the task? box above the agents in the sidebar, and what you type there picks Claude's model and effort for that launch. An Auto session's subagents run on Sonnet. A new install starts on Auto; an existing install keeps its setting.
  • Re-route offer. After a session's first prompt, if another model fits the task better, the agent's header offers it once. Ignoring the offer changes nothing.

Spending fewer tokens

  • Late, not early. Token Headroom now leaves a tool result whole until it reaches 40,000 characters. On the author's machine the model has recently asked for most of what was shortened early, and every such retrieval is an extra turn that re-reads the cached prefix. Legacy (Settings → Token Savings → Shorten tool results) keeps the v1.54 behaviour.
  • Fresh session. When Claude Code compacts a session, the header offers Fresh session. It starts a new session and carries the work over through a memory handoff note, so the new session re-reads a short prefix instead of a long one.
  • Part of a result. retrieve_full takes lines (such as 120-180) or grep (literal text, or /a|b/ for any of several), so an agent can bring back only what it needs. Text now comes back as plain text instead of a JSON string.
  • Outline before a whole read. Before Claude Code reads the whole of a file over 1,500 lines, it gets the file's outline from the code graph, so it can read just the part it needs.
  • Cheaper review loops. Swarm review loops use Sonnet reviewers at medium effort, at most 3 rounds, and each round reviews only the changes since the last.
  • Lean output. Output steering asks for a diff or the changed lines rather than a whole file.
  • Smaller context window (opt-in, off by default). Settings → Token Savings → Long conversations can set a 500k, 400k, 300k or 200k window, so Claude Code compacts sooner.
  • One row per lever on the receipt. Settings → Token Savings lists Routing, Late elision, Relay, Focused recall, Graph first and Lean output. A row says measured only after its comparison against a randomised holdout clears a 99% interval. Until then it says collecting (n of N) or observed, and it never claims a saving it hasn't measured.

Memory

  • One memory primer per session, focused on the session's first prompt.
  • Claude Code's saved notes (~/.claude/projects/<project>/memory/*.md) are copied into memory under their project, so Codex and the Antigravity CLI can find them too. This happens only while your agents are connected and the switch in Settings → Memory & Learning is on (on by default).
  • A lesson has to say something to be kept. Narration such as "Root cause found." or "Next:" is no longer stored as a lesson.

Second Opinion

  • The Claude models that the Antigravity CLI offers are listed under Claude, not Gemini. A Claude review runs on Sonnet at medium effort unless you pick otherwise.

Safety fixes from review

  • retrieve_full grep never compiles a regex. The pattern comes from the model, and a backtracking pattern could have frozen the app.
  • Fresh session never restarts a session without a clear handoff mark.

Full Changelog: v1.54.7...v1.55.2