fix(provider): add Gemini 3.8, 3.7, 3.6 Flash and 3.5 Flash-Lite models

[hz-xiaxz]

Summary

Add the Gemini Flash models Google released between July and September 2026 so they can be selected with :model on the vertex_ai and google_ai_studio providers:

• gemini-3.8-flash (released 2026-09-02)
• gemini-3.7-flash (released 2026-08-13)
• gemini-3.6-flash (released 2026-07-21)
• gemini-3.5-flash-lite (released 2026-07-21)

This is additive only. No existing model entries are removed or changed, per the feedback on #3748.

Changes

• crates/forge_repo/src/provider/provider.json: add the four models to the vertex_ai and google_ai_studio model lists, placed next to the existing gemini-3.5-flash entry.
• vertex.json: add the same four models with the google/ prefix.

All entries use the same shape as gemini-3.5-flash: context_length 1048576, tools and parallel tool calls supported, reasoning supported, text + image input. Token limits and capabilities were taken from the Gemini API model pages:

Model	Input tokens	Output tokens	Function calling	Thinking
gemini-3.8-flash	1,048,576	65,536	yes	yes
gemini-3.7-flash	1,048,576	65,536	yes	yes
gemini-3.6-flash	1,048,576	65,536	yes	yes
gemini-3.5-flash-lite	1,048,576	65,536	yes	yes

Not included

gemini-3.5-flash-cyber and gemini-3.8-flash-cyber are only available through Google's limited-access Fairwind program and are not listed in the public Gemini API model catalog, so they are left out.

Testing

cargo test -p forge_repo

Both JSON files were also parsed with a script to confirm they are valid and contain no duplicate model ids within a provider.

Related

Supersedes #3780 (closed by the stale bot). Fixes #3779.

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[github-actions]

Action required: PR inactive for 5 days.
Status update or closure in 10 days.

[amitksingh1490]

Independent review for #3779 (2026-09-16), against current PR head 5bfe035bc and authored model commit f4c4fbf16. I also reviewed related #3748 (including the maintainer's request not to remove non-EOL models) and closed/unmerged #3780. No duplicate PR or contributor-branch changes are needed for the public-model additions.

Model verification

Current authoritative catalogs and individual model pages confirm the exact public IDs gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3.7-flash, and gemini-3.8-flash on Gemini API and Google Cloud:

• https://ai.google.dev/gemini-api/docs/models
• https://ai.google.dev/gemini-api/docs/models/gemini-3.6-flash
• https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash-lite
• https://ai.google.dev/gemini-api/docs/models/gemini-3.7-flash
• https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
• https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/google-models
• https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-6-flash
• https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-5-flash-lite
• https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-7-flash
• https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/gemini/3-8-flash

The pages confirm 1,048,576 input/context tokens, 65,536 maximum output tokens, function calling and thinking support. The API changelog confirms July 21 / August 13 / September 2 release dates: https://ai.google.dev/gemini-api/docs/changelog . Catalog availability is not proof of access for a particular account/location.

Flash Cyber remains a scope/access question, not an entry to invent. Neither public catalog lists the requested gemini-3.5-flash-cyber. I could not independently substantiate the PR's Fairwind-specific claim from the official pages inspected. Please provide an authoritative endpoint/access document (and an appropriately authorized test account if restricted), or explicitly defer Cyber from #3779. This is not a claim that no restricted Cyber model exists.

Bounded checks

Read-only Python fixture checks passed on the latest head: valid JSON; no duplicate model IDs within static provider lists/root Vertex list; exactly four additions per target catalog; expected capability metadata; consistent google/ prefix entries; all pre-existing entries/provider configuration unchanged; authored model catalogs unchanged by the autofix commit. git diff --check origin/main...HEAD passed.

cargo fmt --all -- --check exited 1 on installed rustfmt 1.9.0-stable (Rust 1.97): differences in info.rs and fd.rs, plus warnings that the repository's nightly-only formatting options are unsupported. This is not a passing repository lint result; no nightly toolchain was downloaded.

Not ready yet

1. The authored change is 2 JSON files, but the current head includes an autofix.ci commit touching 139 additional Rust files (141 total). Please restore/review the intended scope without duplicating this PR. I have not modified automation or contributor branches.
2. The latest-head check API reports only successful Release Drafter, not build/test CI; merge state is BLOCKED. Older PRs' CI results do not validate this head.
3. Local storage is 95% full (30GB volume, only 1.6GB free). No Cargo compile/test/clippy or new debug build was launched, and no other worker's artifacts were removed. Relevant Rust tests and supported-toolchain linting remain required when a safe build budget/CI is available.
4. Debug-cli guidance was read. An existing debug binary from a different issue was used ONLY for baseline smoke checks: --help and list model --help succeeded; isolated-home list model --porcelain exited 0 with no models; isolated noninteractive -p 'Reply with OK without using tools.' printed No such device or address (os error 6) during credential migration despite exit 0. That is NOT successful inference, not PR-head catalog verification, and not a mock/live model pass. A rebuilt PR-head CLI with authorized Gemini API and Vertex credentials is still needed for selection and live prompt/tool-call checks.

Co-Authored-By: ForgeCode noreply@forgecode.dev

[hz-xiaxz]

Thanks for the review. Updated:

• Dropped the autofix.ci commit and rebased onto current main. The PR is back to the single authored commit touching only crates/forge_repo/src/provider/provider.json and vertex.json (+120 lines, nothing removed). Merge conflicts are gone.
• gemini-3.5-flash-cyber is deferred from [Feature]: Add support for Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber #3779. It is not in the public Gemini API or Vertex catalogs, so I have nothing authoritative to add for it. Happy to follow up if a public model page appears.
• cargo test -p forge_repo passes on the new head (333 passed, 0 failed). JSON re-checked: valid, four additions per catalog, no duplicate ids.

[github-actions]

Action required: PR inactive for 5 days.
Status update or closure in 10 days.