Simon Gardling 318373c09c pi: patch omp to require reasoning_content for OpenRouter reasoning models ...

DeepSeek V4 Pro (and similar reasoning models reached via OpenRouter) reject
multi-turn requests in thinking mode with:

  400 The `reasoning_content` in the thinking mode must be passed back
  to the API.

omp's existing kimi placeholder injection (`requiresReasoningContentForToolCalls`)
covered this requirement only for `thinkingFormat == "openai"`. OpenRouter
sets `thinkingFormat == "openrouter"`, so the gate never fired even though
the underlying providers behind OpenRouter (DeepSeek, Kimi, etc.) all enforce
the same invariant.

This patch:

1. Extends `requiresReasoningContentForToolCalls` detection: any
   reasoning-capable model fronted by OpenRouter now sets the flag.
2. Extends the placeholder gate in `convertMessages` to accept
   `thinkingFormat == "openrouter"` alongside `"openai"`.

Cross-provider continuations are the dominant trigger: a conversation
warmed up by Anthropic Claude (whose reasoning is redacted/encrypted on
the wire) followed by a switch to DeepSeek V4 Pro via OpenRouter. omp
cannot synthesize plaintext `reasoning_content` from Anthropic's
encrypted blocks, so the placeholder satisfies DeepSeek's validator
without fabricating a reasoning trace. Real captured reasoning, when
present, short-circuits the placeholder via `hasReasoningField` and
survives intact.

Side benefit: also closes a latent gap where Kimi-via-OpenRouter
(`thinkingFormat == "openrouter"`) had the compat flag set but the
placeholder gate silently rejected it.

Applies cleanly on top of patch 0001.

2026-04-25 19:20:05 -04:00

7.1 KiB

Raw Blame History

View Raw