[bug] llm-deepseek: user text serializes before role:tool messages, breaking the tool_calls adjacency requirement

[kharkilirov1]

Affected: master @ 0.1.0-rc.7 — packages/llm/llm-deepseek/src/serialize.ts, serializeMessages user-role branch (~L129)

Symptom

The OpenAI/DeepSeek chat API requires every role:"tool" message to immediately follow the assistant tool_calls turn it answers (the documented tool-call flow is assistant(tool_calls) → tool → next request on every tool sub-turn). When one harness user message carries both tool-result blocks and trailing text, the serializer emits the user text message first and the role:"tool" messages after.

Reproduction

Unit-level and deterministic: call serializeMessages with a conversation whose last user message carries one tool-result block followed by a text block (i.e. every turn where tool results accompany a user reply).

Expected wire:

assistant(tool_calls) → tool → tool → user

Actual wire:

assistant(tool_calls) → user → tool → tool

Root cause

In the user-role branch, the flattened text is pushed onto the wire before the tool-result blocks are expanded into role:"tool" messages.

Fix

Emit the role:"tool" blocks first and the text as the trailing user message. The harness user-message structure and the session log are unchanged — only the wire emission order moves, so this is a pure adapter fix.

A tested patch (ordering regression test in serialize.spec.ts) is available in my fork:

• commit: kharkilirov1@f081047
• full Windows-focused fix batch for comparison: master...kharkilirov1:deepseek-harness:fix/windows-snapshot-infra-and-tool-robustness

Happy to adapt it to the maintainers' preferred format.