fix(agents): Agent Chat streaming fails on Ollama (400 'invalid message content type: <nil>')

[w7-mgfcode]

Summary

Interactive Agent Chat (Chat page → WebSocket /agents/stream → AgentService.stream_chat → agent.run_stream) fails on a local-Ollama agent. Both the primary (ollama:qwen3:8b) and fallback (ollama:llama3.1:8b) return the same Ollama-side 400, so FallbackModel raises a FallbackExceptionGroup and the UI shows "Stream error: All models from FallbackModel failed (2 sub-exceptions)".

Exact error (both sub-exceptions)

openai.BadRequestError: 400 - {'message': 'invalid message content type: <nil>', 'type': 'invalid_request_error'}
pydantic_ai.exceptions.ModelHTTPError: status_code: 400, model_name: qwen3:8b   (and llama3.1:8b)

Raised from pydantic_ai/models/openai.py:request_stream → _completions_create (stream=True) against Ollama's OpenAI-compatible /v1/chat/completions. Fast fail (~1s), first turn (history_length=0).

Root cause

Streaming-path incompatibility between PydanticAI's OpenAIChatModel.request_stream (OpenAI client + OllamaProvider, base_url …/v1) and Ollama's OpenAI-compat endpoint: a message in the streamed request is serialized with content: null, which Ollama rejects (stricter than the real OpenAI API, which tolerates it).

Key distinction (verified)

• The non-streaming path (agent.run() in AgentService.chat, used by the showcase agent_hitl_flow step) works on Ollama (ran ~15 s, HTTP 200, tokens returned).
• Only the streaming path (agent.run_stream in AgentService.stream_chat) hits the 400.
• Cloud providers stream fine (Gemini streamed earlier; it only failed on quota). So this is Ollama-OpenAI-compat + streaming specific — not a key/quota/model-name issue, and unrelated to the fix(agents): persist pending_action for gated tool calls #336 (pending_action) / fix(api): showcase agent step skips on local Ollama (no API key needed) #340 (key-present) fixes.

Proposed fix

Add a non-streaming fallback for the ollama provider in AgentService.stream_chat: when agent_default_model (and/or fallback) is an ollama: model, run the turn with agent.run() and emit the result through the existing event path — one text_delta with the full text, then the existing deps.pending_action → approval_required handling, then complete. This sidesteps the broken streamed request while preserving the WS contract and the #336 HITL approval flow.

• Cloud providers keep the true streaming path (must remain unaffected).
• Alternative considered: sanitize the null-content message before the streamed request — rejected as brittle (it lives in PydanticAI/openai-client serialization). A PydanticAI version bump is out of scope (stop-and-ask per AGENTS.md).

Tests

• stream_chat with an ollama:* agent_default_model emits text_delta + complete (and approval_required when a gated tool fires) without calling run_stream.
• stream_chat with a cloud agent_default_model still uses run_stream (regression guard).

Notes

Found 2026-06-01 testing the #336 HITL approval card via Chat on a local-Ollama stack. Compounding local limitation: even on the working non-streaming path, qwen3:8b (8B) often doesn't emit tool calls (see related work).

[w7-mgfcode]

Fixed in PR #343 and merged into dev. Closing manually (dev is not the default branch, so GitHub auto-close only fires on the next release merge to main).