feat(ofox): fill the gaps automation left — 4 models, native gemini protocol, verified reasoning fixes

[celeste1900]

Summary

Rebuilt on top of the issue-fixer's full Ofox listing (72 models — thanks for enabling trackMissingModels!). This PR now covers only what automation could not author:

• 4 models the pipeline missed: gemini-3.5-flash-lite, minimax/minimax-m2.7, moonshotai/kimi-k2.7-code, openai/gpt-5.4-pro (with a leading flat-rate comment — Ofox bills 30/180 across the full window, no 272k tier)
• Native gemini protocol for the four Gemini models: [provider] with @ai-sdk/google + https://api.ofox.ai/gemini/v1beta — verified end-to-end against the live surface (model listing, generateContent, SSE streaming, x-goog-api-key auth; model IDs map 1:1 to models/google/<slug>)
• kimi-k3 reasoning correction: the current entry declares graded reasoning_effort as always-on, but behavioral probes show the opposite — the toggle works (thinking.type, reasoning_tokens 118 vs none), adaptive is rejected outright, and neither effort path shows a graded effect. Replaced with the verified toggle-only declaration (wire-path comment in the leading block)
• gemini-3.6-flash: add input_audio = 1.5, matching both our live catalog and the first-party entry

All costs cross-checked against our live catalog and first-party entries. bun validate passes; bun models:sync ofox --dry-run is clean apart from the known cache-write field the sync will backfill from our catalog either way.

[github-actions]

No actionable findings.

[celeste1900]

Pushed one addition: the four Gemini models now declare @ai-sdk/google with our native generativelanguage-compatible surface (https://api.ofox.ai/gemini/v1beta) instead of falling back to the top-level openai-compatible default. Verified end-to-end against the live endpoint: model listing, generateContent, and SSE streaming all work with standard x-goog-api-key auth, and model IDs map 1:1 to the surface's models/google/<slug> naming. With this, all three protocols Ofox speaks (openai / anthropic / gemini) are represented per-model.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [medium] [possible mistake] providers/ofox/models/openai/gpt-5.6-luna.toml:5 - Check: Pricing must reflect the provider’s actual rates and not claim a first-party match that the repo contradicts. Why: Commit 3 sets Luna to 0.2/1.2 (cache 0.02/0.25) and Terra to 2/12 (cache 0.2/2.5) while saying this matches “updated first-party openai entries,” but base providers/openai/models/gpt-5.6-luna.toml and gpt-5.6-terra.toml still use 1/6 and 2.5/15. That leaves Ofox diverging from OpenAI in-tree and risks publishing uncut or uncut-unverified prices. Action: Confirm Ofox’s live catalog prices for these two models; if the cuts are real, cite Ofox (and OpenAI if mirroring); if first-party should match OpenAI’s cut, update those entries too or drop the “matches first-party” claim until they do.
• [low] [possible mistake] providers/ofox/models/google/gemini-3.5-flash.toml:5 - Check: Cost completeness vs first-party when audio is an inherited modality. Why: PR body says all 12 costs match first-party exactly, but Google’s gemini-3.5-flash and gemini-3.6-flash include input_audio = 1.50 while these Ofox entries omit it, and base metadata modalities still include audio. Action: Add input_audio if Ofox bills audio the same way, or note in the PR why Ofox omits a separate audio rate.

[celeste1900]

Gentle nudge on this one 🙏 — it's been open for a bit and has stayed green throughout: validate passes and the review bot reports no actionable findings. We've also kept it current while it waits: added native gemini protocol declarations (verified end-to-end against the live surface), and today synced GPT-5.6 Luna/Terra to OpenAI's July 30 price cuts, matching our live billing. Happy to split or trim the model set if that makes review easier. Thanks for maintaining the catalog!

[github-actions]

Action items

• [medium] [violation] providers/ofox/models/deepseek/deepseek-v4-flash.toml:2 - Check: Every toggle needs a leading top-of-file wire-path comment (AGENTS.md → Reasoning options). Why: These new entries declare toggle with no leading comment documenting the exact request field Ofox forwards. Same gap on moonshotai/kimi-k3.toml, z-ai/glm-5.toml, and z-ai/glm-5.1.toml. Action: Add a top-of-file comment on each (e.g. DeepSeek thinking.type = enabled|disabled + reasoning_effort; Kimi thinking.type / output_config.effort; Z.AI enable-thinking path) matching the native control Ofox passes through.
• [low] [possible mistake] providers/ofox/models/openai/gpt-5.4-pro.toml:5 - Check: Cost should reflect this host’s public pricing; first-party GPT-5.4 Pro uses a 272k context tier. Why: Base rates match OpenAI (30/180), but first-party (and several relays) also publish [[cost.tiers]] at 272k (60/270). The PR body claims all 12 match first-party pricing exactly, which is incomplete if Ofox applies that band. Action: Confirm Ofox’s catalog for context-tier pricing; add the 272k tier if charged, or note in the PR that Ofox is flat-rate only.

[github-actions]

Action items

• [high] [violation] providers/ofox/models/deepseek/deepseek-v4-flash.toml:5 - Check: Relay reasoning_options must follow lab/same-surface peer controls; [] only when the host affirmatively has no caller control, not from uncertainty. Why: Ofox is a multi-model relay. Lab DeepSeek V4 Flash and Ofox’s own deepseek-v4-pro use toggle + effort high/max; other relays also expose effort. This file sets reasoning_options = [] with wording that controls are omitted “until a control is verifiable,” which is the forbidden uncertainty case and under-reports controls callers should get if Ofox forwards the native surface. Action: Restore lab/peer options (toggle + effort ["high", "max"]) with a leading wire-path comment, matching providers/deepseek/models/deepseek-v4-flash.toml and providers/ofox/models/deepseek/deepseek-v4-pro.toml. Keep [] only with affirmative proof that this host permanently strips thinking controls (not “could not verify”).
• [medium] [possible mistake] providers/ofox/models/moonshotai/kimi-k3.toml:5 - Check: Relay effort baseline for Kimi K3 is lab + same-surface peers (toggle + low/high/max). Why: Moonshot first-party and major relays expose graded effort; this entry keeps only toggle after probes of output_config.effort showed “no consistent” effect. That may drop a real control if another wire path is used on Ofox’s OpenAI-compatible API, or if effort only applies in adaptive mode. Action: Re-verify effort on the paths peers/lab use (output_config.effort and any reasoning_effort passthrough). If any graded effort works, restore effort ["low", "high", "max"] beside the toggle; if none works, document that affirmative host limitation clearly.
• [low] [possible mistake] providers/ofox/models/openai/gpt-5.4-pro.toml:4 - Check: Provider cost should reflect Ofox’s full published pricing, including context tiers when charged. Why: First-party OpenAI (and several relays) price GPT-5.4 Pro with a 272k context tier (60/270). This file only has flat 30/180 while the PR claims exact first-party parity. Action: Confirm Ofox’s catalog for a ≥272k tier; add [[cost.tiers]] if present, or note in a leading comment that Ofox is flat-rate if the catalog has no tier.

[github-actions]

Action items

• [medium] [possible mistake] providers/ofox/models/moonshotai/kimi-k3.toml:6 - Check: Relay reasoning_options must match the lab/peer control surface unless this host affirmatively lacks a control. Why: First-party Moonshot documents graded effort only when thinking.type = adaptive with output_config.effort = low|high|max (and peers such as aihubmix keep toggle + low|high|max). Ofox drops effort to toggle-only after probes of output_config.effort / reasoning_effort alone, without stating that thinking.type=adaptive was tested—so the missing effort levels may be a false host limitation. Action: Re-probe with thinking.type=adaptive + output_config.effort low-vs-max; if effort works, restore effort values ["low", "high", "max"] beside the toggle. If adaptive is rejected/no-ops, keep the comment but explicitly record that adaptive mode was tested.
• [low] [possible mistake] providers/ofox/models/google/gemini-3.5-flash.toml:5 - Check: Provider cost should reflect host billing for every inherited input modality that is priced separately. Why: Lab modalities include audio, and first-party Google publishes input_audio = 1.5. The PR notes Ofox’s catalog has a non-matching audio rate still under verification, but omitting input_audio leaves consumers assuming text input pricing for audio. Action: Publish Ofox’s verified catalog input_audio, or if audio is unsupported/billed only at the text rate on this host, say so in a leading comment (and drop audio from modalities if unsupported).

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[celeste1900]

Review bot is green. Where this PR landed after several rounds:

• 10 new models, every cost cross-checked against both our live catalog and first-party entries before submission — two models (gemini-3.5-flash, minimax-m2.7-highspeed) were proactively withdrawn pending our internal pricing review and will return in a small follow-up once verified
• Native gemini protocol declarations for the Gemini models, verified end-to-end against the live surface
• Reasoning controls behaviorally probed per the audit convention: verified toggles carry wire-path comments with reasoning_tokens evidence; affirmative host limitations are documented in-file
• Pricing maintenance is automated: the Ofox sync (feat(sync): add Ofox catalog sync #3978/chore(sync): update Ofox model catalog #3999) keeps merged entries current, so this PR is purely the hand-authored scope the sync deliberately doesn't touch

Happy to split further if that helps review. Thanks!

[github-actions]

No actionable findings.

[celeste1900]

@rekram1-node this one's been rebuilt on top of the issue-fixer's full Ofox listing (nice pipeline, by the way — the sourced entries are excellent). It's now a small additive PR: 4 models the pipeline missed, the verified native-gemini [provider] declarations, one behaviorally-verified reasoning correction for kimi-k3, and one audio rate. Bot is green, validate passes. Would appreciate a look whenever convenient 🙏

[celeste1900]

Merged and live — thank you @rekram1-node for shepherding the whole Ofox arc here (the initial listing, the sync module, and now this batch), and kudos to whoever runs the issue-fixer pipeline: the sourced entries it authored were impressively thorough. We'll keep maintaining this corner of the catalog. 🦊