Problem#
requestMap.replayReasoningContent is true on only 20 rows (DeepSeek and Fireworks). These models also need reasoning_content on earlier assistant turns, or the second turn fails. Measured 2026-10-09; models.dev interleaved.field: "reasoning_content" used as a hint:
| Provider | Rows | Examples |
|---|
| baseten | 11 | deepseek-ai/DeepSeek-V4-Pro-0813, moonshotai/Kimi-K3, zai-org/GLM-5.2 |
| huggingface | 11 | deepseek-ai/DeepSeek-V4-Pro, MiniMaxAI/MiniMax-M2.7, moonshotai/Kimi-K2.6 |
| dashscope | 9 | qwen3.8-max, kimi-k3, deepseek-v4-flash-0731 |
| openrouter | 7 | deepseek/deepseek-v4-pro, z-ai/glm-5.2 |
| together | 7 | deepseek-ai/DeepSeek-V4.1-Flash, zai-org/GLM-5.2 |
| zai | 7 | glm-5.2, glm-5.3-flash, glm-4.7 |
| nvidia | 6 | z-ai/glm-5.3, moonshotai/kimi-k3 |
| moonshotai-cn | 4 | kimi-k2.6, kimi-k3 |
| moonshot | 2 | kimi-k2.6, kimi-k2.7-code |
| cloudflare-workers-ai | 2 | @cf/moonshotai/kimi-k2.6 |
| zhipuai-coding-plan | 2 | glm-5.3 |
| cloudflare-ai-gateway | 1 | alibaba/qwen3.8-max |
Ask#
Fill replayReasoningContent from each provider's docs. The model makers document it (DeepSeek, Moonshot, Z.AI thinking docs). A host that serves the same model on Chat Completions usually needs the same thing; confirm it from the host's docs before setting it.
Acceptance#
- Every row in the table has
replayReasoningContent set to true or false, or a ledger entry.
moonshot/kimi-k2.6 and zai/glm-5.2 are true.
Completed through reviewed provider-native extraction and PR #293. Production verification after the scheduled poll finds zero unaccounted replayReasoningContent gaps across the issue's original 12 providers. Both required rows are sourced true: moonshot/kimi-k2.6 and zai/glm-5.2.
Replay booleans are emitted only from the serving provider's published contract or operational guidance. Where an exact model's healthy native source omits a replay rule, its value stays null with scoped checked-source evidence. Provider-wide guessed recipes were removed. NVIDIA's refreshed contracts preserve exact selectors and published endpoint paths; malformed/unavailable sources remain errors.
Recorded source URLs, hashes, and model-scoped omission evidence are now visible in the public API by default and on the web pages. Upstream captures, all ledger scopes, Worker ingestion/repoll tests, independent exact-head reviews, all 1,881 tests, and required CI passed.