feat(models): update Gemini model context windows and output limits

[CleoMenezesJr]

Summary

• What changed: Added context window and max output token entries for native Google Gemini models in openaiContextWindows.ts — covering gemini-2.0-flash, gemini-2.5-flash, gemini-2.5-pro, gemini-3-flash, and gemini-3.1-pro (used via CLAUDE_CODE_USE_GEMINI).
• Why it changed: Without these entries, the system falls back to an 8k default context window. This causes calculateTokenWarningState to signal a blocking limit at session start, triggering auto-compaction before any real context is consumed.

Impact

• user-facing: Gemini sessions no longer compact prematurely. Token warning thresholds now reflect the actual model limits (e.g., 1M tokens for Flash, 2M for Gemini 3.1 Pro).
• developer/maintainer: Pure data addition — no logic changed. Entries follow the existing ascending-version ordering used for other provider families in the same file.

Testing

• bun run build
• bun run smoke
• focused tests: launch with bun run dev:gemini using gemini-2.5-pro or gemini-3-flash, send a message, and confirm that the token indicator reflects the correct context window instead of triggering an immediate compaction prompt.

Notes

• provider/model path tested: CLAUDE_CODE_USE_GEMINI=1 with gemini-2.5-pro and gemini-3-flash via native endpoint.
• screenshots attached (if UI changed): N/A
• follow-up work or known limitations: Output token limits for gemini-3-flash and gemini-2.0-flash are set conservatively at 8k. These should be updated if Google raises the limits in a future release.

[Vasanthdev2004]

Reviewed on head d5f5ce7 — adds context window and max output token entries for native Gemini models. CI green ✅

The intent is right — without these entries, native Gemini sessions fall back to 8k context and trigger premature compaction. But two of the five values need correction based on Google's official specs:

🔴 Incorrect values:

1. gemini-3.1-pro context window: 2,097,152 (2M) — Google's official model card states: "token context window of up to 1M". Should be 1,048,576.

2. gemini-3-flash max output tokens: 8,192 — Per sim.ai model tracking and Google's API docs, Gemini 3 Flash supports up to 65,536 output tokens (same as 2.5 Pro and 2.5 Flash). Setting this to 8k would unnecessarily limit responses from the model.

✅ Correct values:

Model	Context	Max Output	Source
gemini-2.0-flash	1,048,576	8,192	Correct per Google specs
gemini-2.5-flash	1,048,576	65,536	Correct
gemini-2.5-pro	1,048,576	65,536	Correct (was already in file, just reordered)
gemini-3-flash	1,048,576	8,192 → 65,536	Needs fix
gemini-3.1-pro	2,097,152 → 1,048,576	65,536	Needs fix

Verdict: Needs changes — two values don't match Google's published specs. Once corrected, this is approve-ready.

[gnanam1990]

I can’t approve this yet. The direction is right, but some of the Gemini context-window and output-limit values still need correction before merge. Since this data feeds runtime limit handling, I’d want the numbers fixed first, then I’m happy to recheck.

[Vasanthdev2004]

Re-reviewed on head 7def95e — both blockers from the previous review are fixed ✅

Model	Context	Max Output	Status
gemini-2.0-flash	1,048,576	8,192	✅
gemini-2.5-flash	1,048_576	65,536	✅
gemini-2.5-pro	1,048,576	65,536	✅
gemini-3-flash	1,048,576	65,536	✅ (was 8,192)
gemini-3.1-pro	1,048,576	65,536	✅ (was 2,097,152)

All values now match Google's published model cards. CI green. Pure data addition, no logic changes. Approved ✅

[kevincodex1]

@gnanam1990 kindly have a look again bro

[gnanam1990]

.

[gnanam1990]

Still request changes.

bun run build passes on the current head, but I’m still not comfortable approving the exact Gemini metadata values being added here. This PR changes runtime context-window and max-output tables, and those numbers directly feed warning thresholds, blocking behavior, and auto-compact decisions. Before merging, I’d want the added Gemini limits verified against a clear source of truth or tightened so we’re not shipping incorrect runtime metadata.

[auriti]

Thanks for adding the missing Gemini native context windows — this is a real problem (the 8k fallback triggers premature compaction).

However, most of this PR is already covered by current main and by #783 (which adds the same entries plus OpenRouter variants and output token fallbacks). Here's the overlap:

Model	main (v0.5.2)	PR #783	This PR
gemini-2.0-flash	✅	✅	✅
gemini-2.5-flash	✅	✅	✅
gemini-2.5-pro	✅	✅	✅
gemini-3.1-pro	✅	✅	✅
gemini-3-flash	❌	❌	✅
gemini-3-flash-preview	❌	✅	❌
gemini-3.1-pro-preview	❌	✅	❌
google/gemini-3* (OpenRouter)	❌	✅	❌

The only entry unique to this PR is gemini-3-flash (without -preview). A couple of notes on that:

1. Naming: Google's current API uses gemini-3-flash-preview — the -preview suffix is required until the model reaches GA. gemini-3-flash without the suffix will likely not match any model ID today and would be a no-op entry. When Google promotes it to GA, the name may change — at which point we'd add it.

2. Context window values look correct: 1M for both Flash and Pro, 65k output tokens. These match the official specs.

Suggestion: If #783 lands first, this PR would be fully superseded. If you'd like to contribute the gemini-3-flash (GA name) entry as a forward-looking addition, I'd suggest rebasing on top of #783 and adding just that one entry — but only after confirming the model ID works against the native Gemini endpoint.