Root cause

Interactive Agent Chat (WebSocket /agents/stream → AgentService.stream_chat → agent.run_stream) fails on a local-Ollama agent: Ollama's OpenAI-compatible endpoint rejects PydanticAI's streamed request with 400 - {'message': 'invalid message content type: <nil>'} for both the primary (ollama:qwen3:8b) and fallback (ollama:llama3.1:8b), so FallbackModel raises a FallbackExceptionGroup and the UI shows "Stream error: All models from FallbackModel failed."

The non-streaming path (agent.run() in chat()) works fine on Ollama (the showcase agent_hitl_flow uses it). Cloud providers stream fine. So it's an Ollama OpenAI-compat + streaming incompatibility (a streamed-request message is serialized with content: null).

Fix

stream_chat now detects the provider from agent_default_model. For the ollama provider it uses the non-streaming agent.run() and emits the reply as a single text_delta + the existing approval_required (#336) + complete events. Cloud providers keep the true run_stream token-streaming path unchanged. The post-run logic (session persistence, deps.pending_action → awaiting_approval → approval_required, completion) is now shared by both paths.

Tests

• test_stream_chat_ollama_uses_nonstreaming_path — with agent_default_model=ollama:*, stream_chat calls agent.run() (asserts run_stream is not called), emits text_delta + approval_required (from a gated deps.pending_action) + complete, and flips the session to awaiting_approval.
• test_stream_chat_cloud_keeps_streaming_path — with a cloud model, run_stream is used and run is not (regression guard).
• Hardened 3 existing streaming tests to pin a cloud agent_default_model so they exercise the run_stream path deterministically regardless of the local .env (these were .env-coupled — they pass in CI where the default is anthropic:..., but the fix makes them robust).

Validation

• ruff check ✅ · ruff format --check ✅
• mypy app/ ✅ — only pre-existing xgboost optional-dep import errors in untouched files
• pyright app/features/agents ✅ 0 errors
• pytest -m "not integration" ✅ 1655 passed, 12 skipped

Notes

• Found 2026-06-01 testing the fix(agents): persist pending_action for gated tool calls #336 HITL approval card via Chat on a local-Ollama stack.
• Compounding local limitation (out of scope): even via the working non-streaming path, small (8B) local models like qwen3:8b often don't emit the gated tool call, so the HITL card may not appear; a frontier model triggers it reliably.

Closes #342