Fix post-compact 400 on SGLang/vLLM: merge leading system messages

[thomwebb]

Summary

Resuming a session after auto-compact on SGLang / vLLM-style OpenAI-compatible endpoints fails with:

400: {'message': 'System message must be at the beginning.', 'type': 'BadRequest'}

Root cause

After SummarizingCompaction runs, the compacted history opens with a summary SystemPromptPart (produced by pydantic-ai-harness). On the next request, pydantic-ai's OpenAIChatModel._map_messages also inserts the agent's per-turn instruction_parts right after leading system messages — so the wire format carries two leading system messages. Strict backends accept only one.

pydantic-ai already ships _merge_leading_system_messages for exactly this case, but it only fires when the model profile sets openai_chat_supports_multiple_system_messages=False — and none of our OpenAI-compatible model types did.

Fix

New _strict_openai_profile() in model_factory.py:

• Defaults openai_chat_supports_multiple_system_messages=False, so leading system messages are merged (joined with \n\n). Harmless for endpoints that accept multiples — content is identical, just in one message.
• Merges the existing thinking_tags profile and an optional extra= profile (preserves the Cerebras strict-tool-def override).
• Per-model opt-out via "supports_multiple_system_messages": true (JSON boolean). A non-bool value raises TypeError instead of silently inverting intent — bool("false") is True in Python, so coercion can't save a string config.
• Applied to custom_openai, openrouter, cerebras, zai_coding, and zai_api. Routing Cerebras through the shared helper also picks up thinking_tags, which the old bare profile skipped.
• Plain openai/azure_openai are intentionally untouched: real OpenAI accepts multiple system messages, and merging would forfeit inter-system-message prompt-cache breakpoints.

README documents the default behavior, the SGLang symptom, and the opt-out.

Testing

• Per-model-type profile assertions: custom_openai (default + opt-out), openrouter, cerebras (both flags), zai_coding, zai_api
• Fail-fast guard: string "false" raises TypeError
• Thinking-tags × extra × merge triple-merge, asserted against the runtime-resolved tags (non-vacuous)
• End-to-end wire-format test: a post-compaction history with two leading system parts + instruction_parts passed through _map_messages yields exactly one merged system message
• Full factory/basics/settings suites: 67 passed, 2 pre-existing env-gated skips; ruff format + ruff check clean