[lidge-jun]

Summary

Antigravity Gemini 4 Argon can return no thought summaries when only includeThoughts accompanies its tiered model ID. When a summary is requested, repeat the selected wire tier as thinkingLevel; retain suffix-only requests when summaries are disabled, including the web-search sidecar. Reject inherited effort-map keys rather than selecting a non-string wire ID.

Supersedes #8139

Co-authored-by: Andrew (agent) 59988150+andrew05060414@users.noreply.github.com

Google documents Argon's supported thinking levels in its model documentation. The original PR reported the Cloud Code Assist summary behavior from a live measurement; this carry has not repeated that measurement and does not change the unmeasured level-only request path.

Verification

• On the unmodified base, the three focused regression cases failed (0 pass, 3 fail) for the missing summary level, inherited effort key and discovered-family path.
• env -u GOOGLE_ANTIGRAVITY_USER_AGENT bun test tests/adapters/google/antigravity-discovered-families.test.ts tests/adapters/google/gemini-4-argon-antigravity.test.ts tests/adapters/google/gemini-web-search.test.ts — exit 0; 45 pass, 0 fail. The override is unset only for this test process; the unmodified base with the inherited override set produced 4 unrelated UA expectation failures.
• bun run typecheck — exit 0; 2,317 files.
• bun run structure:check — exit 0 on pre-rebase head 70aead1181.
• bun test tests/ci-workflows/file-size-ratchet.test.ts — exit 0 on pre-rebase head 70aead1181; 9 pass, 0 fail.
• (cd docs-site && bun install --frozen-lockfile && bun run build) — both exit 0 on pre-rebase head 70aead1181; 641 pages and internal links checked.
• env -u GOOGLE_ANTIGRAVITY_USER_AGENT bun run test:changed on pre-rebase head 70aead1181 — exit 143 after 12 reported failures; 990 of 2,249 selected files had not started. Cause of SIGTERM was not established. On rebased head a73784c42d, the same command exited 124 after its 900-second limit; it reported 50 failure events and 478 of 2,249 selected files had not started. The additional failing files have not all been classified against clean base yet; this PR is not ready to merge.
• Re-ran each affected file alone using env -u GOOGLE_ANTIGRAVITY_USER_AGENT bun scripts/test.ts <file> --test-name-pattern "$pattern" on head a73784c42d and clean base 8f43924bba. The exact pattern was retained-sync keeps the reviewer|retained sync and convergence each audit|HTTP fixture npm wrapper returns all metadata|published Node launcher|published CLI packs|a sibling model on the same provider|inbound chat images and paired screenshots|native openai-responses Claude route logs cyber terminals|published check cannot block on a FIFO|plan identity binds the bytes a real npm-global install placed. Results:
  • tests/codex-integration/catalog-auto-review-row.test.ts: head exit 0, base exit 1 (timeout); category b, passed alone on head.
  • tests/codex-integration/codex-catalog-write-audit.test.ts: head/base exit 0; category b.
  • tests/codex-integration/codex-cli-update-apply.test.ts: head/base exit 1, three failures on each; category a, baseline. The fixture reported blocked_active_session while real Codex sessions were running. No user process was terminated.
  • tests/responses/chat-completions-endpoint.test.ts: head/base exit 0; category b.
  • tests/claude-integration/claude-messages-endpoint.test.ts: head/base exit 0; category b.
  • tests/cli/cli-codex-cli-update-boundary.test.ts: head/base exit 0; category b.
• bun run privacy:scan — exit 0 on pre-rebase head 70aead1181.
• Full local bun run test was not run because six lanes share the machines. The ci:full label requests the full Ubuntu shards; cross-platform release coverage remains with the normal batch CI.
• The test:changed wrapper automatically installed missing gitignored GUI dependencies despite this PR changing no GUI files.

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

[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

[github-actions]

✅ Deterministic PR hygiene checks passed.