fix(deepseek): preserve reasoning_content in thinking mode for tool calls

[tlotli]

Summary

• DeepSeek's thinking mode requires reasoning_content to be passed back in all subsequent requests when tool calls are involved. Without this, the API returns 400: The reasoning_content in the thinking mode must be passed back to the API.
• This PR fixes multi-turn conversations with tools in deepseek-v4-flash (and other DeepSeek thinking-mode models).

Problem

When using DeepSeek models in thinking mode with tools, Prism discards the reasoning_content from the API response:

1. Text.php handler creates AssistantMessage with empty additionalContent — the reasoning_content from the raw response is lost
2. MessageMap.php does not merge additionalContent into the API payload — even if reasoning_content was in the message object, it was never sent back to the API

When Prism recursively calls handle() for tool-use continuation, the request includes an AssistantMessage without reasoning_content → DeepSeek returns 400.

Changes

src/Providers/DeepSeek/Handlers/Text.php

• Extract reasoning_content from raw response (choices.0.message.reasoning_content)
• Include it in AssistantMessage.additionalContent for both:
  • Tool call intermediate steps (line 66-73)
  • Final response step via addStep() (line 128)

src/Providers/DeepSeek/Maps/MessageMap.php

• Merge $message->additionalContent into the API payload via array_merge() (line 102-106)
• This ensures reasoning_content is sent back to the API in multi-turn conversations

Testing

Verified with 4-turn conversation using tools:

• Turn 1 (no history) ✅
• Turn 2 (with history) ✅
• Turn 3 (deeper history) ✅
• Turn 4 (4-turn conversation) ✅

References

• DeepSeek Thinking Mode Docs — "Between two user messages, if the model performed a tool call, the intermediate assistant's reasoning_content must participate in the context concatenation and must be passed back to the API in all subsequent user interaction turns."

[Lintume]

Hit this exact bug independently while migrating our diary agent (Calorize, ~200 paid users) to direct DeepSeek API in thinking mode. Confirmed deepseek-v4-flash (thinking) 400s on the second iteration of any multi-step tool flow without this fix.

Been running a local patch with the same approach (extract reasoning_content into additionalContent, merge it back in MessageMap) for several days — works as expected, no regressions. Would love to see this merged so we can drop the override.

Separately: DeepSeek\Handlers\Text::addStep() and Stream::extractUsage() also drop prompt_cache_hit_tokens and completion_tokens_details.reasoning_tokens from the usage payload. That's a different concern from this PR — I'll open a focused PR for it, non-overlapping with this change.