feat: add Gemini 3.1 Flash Lite model

[iceteaSA]

Adds gemini-3.1-flash-lite — the newest Flash Lite the Antigravity API serves.

Verified against the live API

fetchAvailableModels lists it, and a real streamGenerateContent returns HTTP 200 echoing modelVersion: "gemini-3.1-flash-lite" — it's a working model, not a placeholder. Metadata from the API: 1,048,576 context / 65,535 max output, API_PROVIDER_GOOGLE_GEMINI, non-thinking text model. (The API serves no 3.5/3.6 Flash Lite yet; 3.1 is the current Lite.)

Changes

Mirrors the existing gemini-3.6-flash wiring (commit a8e4e28) for a lite text model:

• model-registry.ts — model definition, gemini flash quota group, id list
• transform/model-resolver.ts — resolves as non-thinking (isThinkingModel: false), excluded from the tier-suffix header path
• plugin/request.ts — 65,535 max output tokens
• config + Pi catalog — selectable as google/antigravity-gemini-3.1-flash-lite; exposed in the Pi provider

No thinking-tier or image handling (not applicable to a lite text model).

Verification

unit       1832 pass, 0 fail (+5 new)
e2e        28 pass, 0 fail
typecheck  clean
lint       clean, 0 warnings

Note: branched off current main, so the model lands in the pre-existing gemini flash quota group. If #8 (the two-pool quota rework) merges first, this collapses into the single gemini pool on rebase — a one-line classification adjustment.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

---

Summary by cubic

Adds gemini-3.1-flash-lite as a text-only, non-thinking model (1,048,576 context / 65,535 max output). Fully wired across the opencode stack; resolves correctly on Antigravity and gemini-cli without tiers or preview, and is verified against the live API.

• New Features

  • Registered antigravity-gemini-3.1-flash-lite with text-only limits; added to Antigravity model IDs and the gemini-flash quota group.
  • Resolver maps to gemini-3.1-flash-lite as non-thinking on Antigravity and gemini-cli, skipping -preview.
  • Request builder sets maxOutputTokens to 65,535 with no thinkingConfig.
  • Exposed as google/antigravity-gemini-3.1-flash-lite in the pi provider; pi now derives input modalities from the registry to keep Flash Lite text-only.

• Bug Fixes

  • Drops any thinkingLevel/tier fields for Flash Lite on all paths, even when a tier suffix is present.

^{Written for commit 7115635. Summary will update on new commits.}

[cubic-dev-ai]

All reported issues were addressed across 10 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[iceteaSA]

Both valid — fixed in 726e690.

• Pi image input: the Pi catalog was hardcoding ['text', 'image'] for every model. It now derives input modalities from the registry's modalities.input, so flash-lite (and any future text-only model) exposes input: ['text'].
• gemini-CLI thinking classification: the CLI header path was appending the -preview suffix (no CLI preview variant exists for flash-lite) and the generic gemini-3 logic then enabled thinking. The CLI path now passes the plain gemini-3.1-flash-lite id through and the non-thinking special case applies on both quota paths.

Tests added for both (input: ['text'] assertion; non-thinking resolution on the gemini-cli path). 1833/0 unit, 28/0 e2e.

[cubic-dev-ai]

All reported issues were addressed across 4 files (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[iceteaSA]

Valid — fixed in 7115635.

The leak was specific to the gemini-CLI path: that branch spread requestedTier onto thinkingLevel/tier unconditionally after resolving, so a tier-suffixed flash-lite request came back non-thinking but still carrying tier fields. It now returns before that spread for flash-lite.

Worth noting for the record: the antigravity path and resolveModelWithTier itself were already correct (the latter early-returns for flash-lite before any tier handling), so the leak was one path, not all of them. Tests cover all three anyway as regression guards — the gemini-CLI one fails on the pre-fix code (Received: "high").

1836/0 unit, 28/0 e2e, gates clean.