Skip to content
Source
Sign inSign up
TanStack/schemas

Fill reasoning on chat rows that reason but have no reasoning fact #269

Send to my agent
Closedtombeckenham opened this issue 1 comment
tombeckenhamcommented

Problem

Some chat rows have neither a reasoning object nor reasoning in capabilities, but the model does reason. Measured 2026-10-09 against the models.dev list as a hint (not as a source):

ProviderRowsExamples
openrouter122bytedance-seed/seed-2-1-turbo, poolside/laguna-s-2.1
amazon-bedrock72global.anthropic.claude-haiku-4-5-20251001-v1:0, eu.amazon.nova-2-lite-v1:0, google.gemma-4-31b
huggingface53XiaomiMiMo/MiMo-V2.5-Pro, tencent/Hy3
cloudflare-ai-gateway40anthropic/claude-opus-4.8, anthropic/claude-haiku-4.5
dashscope32qwen3.7-max, qwq-plus, qwen3.5-flash
gemini13gemini-flash-latest, gemma-4-31b-it
nvidia8deepseek-ai/deepseek-v4.1-flash
fireworks7accounts/fireworks/models/glm-5p3
together6deepseek-ai/DeepSeek-V4.1-Flash, zai-org/GLM-5.3-Flash
grok3grok-4.20-0309-reasoning, grok-build-0.1
deepseek2deepseek-flash, deepseek-v4-pro (both had reasoning in capabilities on 2026-10-03)

Also, of rows that do have a reasoning mode of effort or adaptive, efforts is empty on OpenAI 29 / 58 and Mistral 14 / 17 (gap report).

Ask

  1. For each row above, fill reasoning from the provider's own source. A gateway row with a sameAs link may state reasoning in its own listing (OpenRouter supported_parameters, Bedrock model cards).
  2. Find why DeepSeek lost reasoning from capabilities.
  3. Fill efforts for OpenAI and Mistral from the model docs.
  4. Where the provider is silent, add a ledger entry.

Acceptance

  • The gap report reasoning and efforts columns are complete or ledgered for these providers.
  • deepseek/deepseek-v4-pro has a reasoning object.
  1. tombeckenham added labelsbug

  2. Completed through the reviewed native-source sync changes and PR #293. Production verification after the scheduled poll finds zero unaccounted reasoning or effort gaps across the issue's 13 providers. Published controls/effort menus are parsed during sync; exact model scopes record checked omissions, preserving null instead of filling values.

    NVIDIA's Content Safety row now carries its own published toggle and source hash from its exact native Build contract. Guard retains an unknown reasoning control with checked, model-specific evidence; its human API title is not mistaken for an input selector. DeepSeek and the other requested rows retain sourced controls. The new API defaults and web Sources sections expose recorded URLs, hashes, and extraction paths.

    Validated all current ledger scopes, actual upstream captures, Worker sync/poll persistence, independent exact-head reviews, all 1,881 tests, and required CI. The separately discovered stale historical NVIDIA documentation-error status is being corrected in a focused follow-up; current target catalog acceptance is zero gaps, not a claim that all unrelated metadata fields are populated.

  3. tombeckenham closed this

Sign in to join this conversation.