mohdel v0.100.0

Added

• reasoningContentPlaceholder model-spec field (chat-completions
  adapter family). When a model spec carries a string value
  (including ''), every assistant message in the request that has
  no extractable reasoning_content gets the placeholder string
  attached as reasoning_content. This unblocks resuming a
  multi-turn session on a thinking model when the prior assistant
  turns came from a non-thinking model (or from a storage path that
  didn't preserve thinking metadata) and the provider's API rejects
  the request with "reasoning_content in thinking mode must be
  passed back".
  • Verified: deepseek/deepseek-v4-pro and
    deepseek/deepseek-v4-flash accept '' and the resumed
    transcript goes through.
  • Default behaviour unchanged when the field is absent: no
    reasoning_content is synthesized, matching pre-0.100.0 wire
    output.

Scope

• Chat-completions adapter only (_chat_completions.js,
  i.e. providers using the OpenAI Chat Completions wire format:
  Groq, Cerebras, DeepSeek, Mistral, OpenRouter, Fireworks). The
  field is silently ignored on Gemini and Anthropic specs — those
  adapters have their own thinking-mode roundtrip rules
  (thoughtSignature and thinking content blocks respectively).