feat(sync): add safe NanoGPT model catalog sync

[huggix]

Summary

This supersedes #2891 with the sync implementation separated from generated catalog data, addressing the automated review feedback.

• add NanoGPT as an automated aggregator sync source
• consume verified reasoning_efforts rather than inferring controls from IDs
• preserve existing prices when source pricing is null and omit modality-dependent token prices
• refresh limit.input from the source context length
• factor known variants through canonical base_model metadata
• use the normal stale-entry deletion behavior
• never propagate NanoGPT's internal providers field
• leave generated model files to the existing hourly automation

Safety and compatibility

• NanoGPT's source change is additive and limited to the detailed catalog response
• no provider identifiers are written to models.dev output
• missing pricing is never converted to zero
• fixed-reasoning models do not receive invented toggles
• open-weight metadata is sourced explicitly or inherited from canonical metadata

Evidence

• Models endpoint documents the public /api/v1/models?detailed=true catalog and its detailed pricing, limits, and capability fields.
• Live detailed catalog is the authoritative per-model source for the optional reasoning_efforts arrays consumed by this sync.
• Chat Completions reasoning effort documents the request values represented by those arrays: none, minimal, low, medium, high, and xhigh.
• Model suffixes documents thinking and budget aliases normalized before canonical base_model lookup.

Verification

• bun test packages/core/test/sync.test.ts — 50 passing
• bun validate
• live dry run on 2026-07-20 — 57 created, 544 updated, 59 stale removed, 14 unchanged
• generated-output audit — no provider fields, no invalid input limits, and no inferred toggle controls

[github-actions]

Action items

• [medium] [possible mistake] packages/core/src/sync/providers/nano-gpt.ts:272,288 - Check: resolveNanoGptBaseModel only strips the :thinking suffix and only resolves canonical base models for IDs that carry an org prefix (plus the zai-org/, TEE/, and explicit alias cases). Why: providers/nano-gpt/provider.toml documents three NanoGPT model-ID variants — :thinking, legacy -thinking, and Anthropic :<number> budgets — but the code only strips :thinking. Existing catalog entries such as claude-haiku-4-5-20251001-thinking and claude-sonnet-4-5-20250929-thinking (legacy -thinking) have matching models/anthropic/*.toml metadata, as do unprefixed IDs like gemini-2.5-pro (models/google/gemini-2.5-pro.toml) and qwen3.5-27b (models/alibaba/qwen3.5-27b.toml). For every one of these, resolveCanonicalBaseModel returns undefined because there is no / or the suffix is not stripped, so the sync writes a full inline definition instead of using base_model. That conflicts with the AGENTS.md "New models" blocker: "Must use base_model when a models/ metadata entry exists." Action: Either extend resolveNanoGptBaseModel to strip legacy -thinking and :<number> suffixes and to discover canonical metadata for unprefixed IDs (e.g., by probing known provider directories), or confirm that the live ?detailed=true API always returns org-prefixed, :thinking-only IDs for models with matching models/ entries so the legacy/unprefixed existing files are removed as stale rather than rewritten inline.

• [low] [possible mistake] PR body — Check: Citations for the NanoGPT catalog API and the reasoning_efforts field. Why: The sync consumes reasoning_efforts as the verified source for reasoning_options and depends on the ?detailed=true endpoint shape, but the PR body does not link to the API reference or model documentation that proves the field exists and maps to the documented reasoning_effort request values (none, minimal, low, medium, high, xhigh). providers/nano-gpt/provider.toml already cites the docs, but the PR body itself states nothing about what each citation supports, so the reasoning-option mapping cannot be reviewed from the PR alone. Action: Add an "Evidence" section to the PR body citing the NanoGPT API reference for the detailed=true catalog response and the reasoning_efforts field, stating exactly what each link proves.

[rekram1-node]

plz address review

[huggix]

Addressed both review items in 4df632f:

• canonical lookup now normalizes legacy -thinking and Anthropic :<number> budget suffixes, then deterministically probes canonical metadata for unprefixed IDs
• regression coverage includes the cited Claude, Gemini, and Qwen examples
• generated-output verification confirms those entries now serialize with base_model
• the PR body now includes an Evidence section linking the detailed catalog, live reasoning_efforts source, reasoning-effort semantics, and suffix documentation

The validation/SDK workflow is green; the refreshed automated review is still running.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/nano-gpt.ts:237 - Check: Provider models must use base_model when matching models/ metadata exists. Why: TEE/ handling only rewrites deepseek/qwen/glm. Other TEE IDs (e.g. TEE/kimi-k2.5, TEE/gpt-oss-120b, TEE/gemma-4-31b-it) fall through still prefixed with TEE/, and if (direct !== undefined || normalized.includes("/")) return direct then skips the unprefixed resolver. Canonical entries such as moonshotai/kimi-k2.5, openai/gpt-oss-120b, and google/gemma-4-31b-it are never linked, so sync keeps full duplicated provider-agnostic fields. Action: After stripping TEE/, resolve the remainder through the same alias/canonical/unprefixed path used for bare IDs (not only deepseek/qwen/glm), and add tests for those TEE models.
• [high] [violation] packages/core/src/sync/providers/nano-gpt.ts:96 - Check: base_model must point at the matching models/<provider>/<model>.toml entry. Why: "xiaomi/mimo-v2.5-pro-ultraspeed": "xiaomi/mimo-v2.5-pro" forces the wrong canonical model even though models/xiaomi/mimo-v2.5-pro-ultraspeed.toml already exists (different name, dates, weights). Without the alias, resolveCanonicalBaseModel("xiaomi/mimo-v2.5-pro-ultraspeed") would select the correct metadata. Action: Remove this alias (or retarget it to xiaomi/mimo-v2.5-pro-ultraspeed) and cover it in the base-model tests.
• [medium] [violation] packages/core/src/sync/providers/nano-gpt.ts:226 - Check: Variant model IDs should factor against canonical metadata after stripping NanoGPT suffixes. Why: Suffix cleanup only removes a trailing :\d+ and one final :thinking/-thinking. Catalog IDs such as anthropic/claude-opus-4.6:thinking:low (and :medium/:high/:max) keep the extra effort suffix, so canonical resolution misses anthropic/claude-opus-4-6 and emits full inline copies. Action: Strip chained NanoGPT variant suffixes (at least :thinking:<effort> and repeated thinking markers) before alias/canonical lookup, with tests for those IDs.

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/nano-gpt.ts:146 - Check: Sync must preserve hand-authored provider fields when the API is not authoritative (sync.md). Why: preserveExistingOverrides is only true when existing.base_model === baseModel. On first factoring of the many full NanoGPT TOMLs (no base_model yet), API-absent fields fall through to undefined and inherit canonical metadata instead of keeping local overrides. That drops provider-specific values such as reasoning, tool_call, structured_output, limits, and modalities whenever the detailed payload omits them—e.g. providers/nano-gpt/models/anthropic/claude-sonnet-4.6.toml (reasoning = false, limit.output = 128_000, structured_output = true) against models/anthropic/claude-sonnet-4-6.toml (reasoning = true, output = 64_000, no structured_output). Action: When the source does not authoritatively set a field, fall back to existing for any local model (not only already-factored ones)—same pattern as DeepInfra/LLM Gateway. Keep the stricter “don’t copy full-definition junk into a new factor” behavior only for fields you intentionally want to re-inherit (e.g. description), and add a first-factor test that API-silent reasoning/limit.output/structured_output retain the existing provider overrides.

[github-actions]

No actionable findings.

[github-actions]

No actionable findings.

[huggix]

@rekram1-node Thanks for the ping. conflicts should be closed now.

[huggix]

@rekram1-node Could you let me know if there is anything else we can do from our side?

[jonaspm]

wohoo!! congratulations and AWESOME work! @huggix