feat(aixy): add curated 50-model catalog

[oscarcpozas]

Summary

• Add 49 curated routes, bringing the Aixy catalog to exactly 50 models including the existing GPT-4.1 Mini entry.
• Cover recent non-deprecated models from 13 labs without mirroring complete upstream inventories.
• Prefer first-party Aixy routes; use a small OpenRouter selection for current models without a direct route.
• Keep availability tenant-specific and BYOK. This PR adds catalog TOMLs only and does not add bulk sync automation.

Selection

• OpenAI: 11
• Anthropic: 8
• Google: 6
• Zhipu AI / GLM: 2
• Alibaba / Qwen: 5
• DeepSeek: 2
• Mistral: 5
• Meta: 3
• xAI: 4
• ByteDance Seed, MiniMax, Moonshot AI, and NVIDIA: 1 each

Source mapping

• Model identity, capabilities, and default limits inherit through base_model from the canonical lab entries already reviewed in models.dev; the Aixy files keep only host-specific deltas.
• Direct routes mirror the established first-party provider surfaces: OpenAI, Anthropic, Gemini, Alibaba Model Studio, DeepSeek, Mistral, and xAI.
• The five OpenRouter-prefixed routes use OpenRouter's authoritative model API for IDs, pricing, limits, and supported parameters, plus its reasoning controls.
• The zero-cost Meta Llama routes reflect Meta's officially announced limited free Llama API preview; availability remains credential- and project-scoped.
• Aixy's integration contract and authenticated catalog behavior are documented in the Aixy integration guide.

Validation

• bun validate
• 50 resolved Aixy models
• 0 deprecated or beta routes
• All 44 reasoning models define provider-appropriate reasoning_options

This replaces the earlier bulk-generated approach. #5507 remains closed.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/aixy.ts:1283 - Check: Every toggle reasoning control must have a leading top-of-file wire-path comment. Why: toggleHeader() only writes “forwards… / mirrors providers/…”, not the exact request field (e.g. enable_thinking, thinking.type, thinking_config.thinking_budget). Many source TOMLs also lack a leading wire comment (e.g. Google Gemini 2.5, Anthropic Sonnet 5, OpenRouter Claude), so generated Aixy routes such as providers/aixy/models/gemini/gemini-2.5-flash.toml and providers/aixy/models/anthropic/claude-sonnet-5.toml ship type = "toggle" without a usable wire path. Action: Emit (or copy) a real wire-path header for every toggle route—source leading comments when they document the control, otherwise a concrete Aixy/upstream field path—and fix already generated toggle files.
• [high] [violation] packages/core/src/sync/providers/aixy.ts:1139 - Check: Sync must not treat a deliberately incomplete on-disk catalog as the full desired set under default delete/create behavior. Why: aixy is registered in hourly aggregators and builds the full 18-provider desired set, with default deleteMissing (not false). Part 1 keeps only OpenRouter/OpenAI/Anthropic/Gemini routes (~249). The next scheduled bun models:sync aixy will create the withheld ~257 routes (and any later filter change would delete local-only rows), so the intentional 1/2 split is not stable and bypasses part-2 review. Action: Keep the full generated catalog in one mergeable state, or gate automation until both parts land (e.g. omit aggregators/provider registration until complete, or set deleteMissing: false plus an explicit create policy that matches the split), and document the intended automation behavior in sync.md.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/aixy.ts:1229 - Check: Provider-authored experimental (and related request-shape) fields must be preserved on override-only gateway routes. Why: buildAixyModel factors name/capabilities/cost/limits/reasoning but never copies model.experimental (or top-level provider). Part 1 routes that inherit from sources with fast/pro modes therefore lose those modes and mode-specific costs—e.g. providers/aixy/models/openai/gpt-5.4.toml, gpt-5.5.toml, gpt-5.4-mini.toml, gpt-5.6-*.toml, and providers/aixy/models/anthropic/claude-opus-5.toml / claude-opus-4-8.toml vs their OpenAI/Anthropic sources. Consumers of the Aixy catalog get incomplete pricing and request controls even though Aixy claims to forward the upstream payload. Action: Pass experimental (and any source provider overrides) into factorBaseModel, regenerate the affected Aixy TOMLs, and cover at least one OpenAI and one Anthropic experimental-mode route in packages/core/test/aixy.test.ts.

[github-actions]

No actionable findings.

[rekram1-node]

too big

[rekram1-node]

also ur including like every model ever I think rather than ones that are actively served (i think, i ddnt loook too closely)

[oscarcpozas]

Fair point — I over-expanded the catalog from upstream provider inventories. I have reduced the work from 253 files to 49 curated routes, giving Aixy 50 models total including the existing entry. The bulk sync is removed, #5507 is closed, and every selected route is current and non-deprecated. Availability remains tenant-specific. Thanks for the callout.

[github-actions]

No actionable findings.

[github-actions]

Action items

• [high] [violation] providers/aixy/models/openrouter/minimax/minimax-m3.toml:4 - Check: Relay reasoning_options must match the lab/same-surface peer control set; [] means no caller control, not uncertainty. Why: This Aixy OpenRouter route sets reasoning_options = [] while providers/openrouter/models/minimax/minimax-m3.toml and first-party providers/minimax/models/MiniMax-M3.toml both expose a real toggle (reasoning.enabled). Catalog consumers will treat MiniMax-M3 as uncontrollable on Aixy even though the upstream surface supports on/off. Action: Replace [] with the OpenRouter peer shape (toggle plus leading # Toggle: reasoning.enabled = true|false wire comment), or cite Aixy-specific docs proving this route strips that control.
• [medium] [violation] providers/aixy/models/openrouter/minimax/minimax-m3.toml:11 - Check: After base_model, keep only provider-specific fields and real deltas. Why: [limit] context = 1_048_576 / output = 512_000 restates the identical values already on models/minimax/MiniMax-M3.toml, so the override-only rule is broken. Action: Drop the redundant [limit] block (or keep only fields that truly differ).
• [low] [possible mistake] .pr-review/pull-request.json (PR body) - Check: Data-changing PRs should cite first-party pricing/docs/API and map claims; PR summary should match the delivered catalog. Why: The body still describes 11 new routes / 12 total models, but the final tip expands to ~50 curated routes with many cost, limit, modality, and reasoning deltas and few top-level source comments. Reviewers cannot verify material pricing/capability claims from the PR text alone. Action: Update the PR body to list the final model set and add direct citations mapped to pricing, limits, and reasoning controls (especially Meta zero-cost Llama routes and OpenRouter-prefixed entries).

[github-actions]

No actionable findings.

[oscarcpozas]

@rekram1-node would it be better now?

[oscarcpozas]

Closing this in favor of a fresh resubmission of the same focused, one-commit catalog change. The replacement uses this exact validated commit and diff; I will link it here once created so the review context remains traceable.

[oscarcpozas]

Replacement: #6089. It points to the same validated commit (a890eed52) and preserves the same 49-file catalog diff.