Leon Ward 32b47e847c fix(llm): sanitize chat messages to OpenAI wire shape for strict gateways ... A real-LLM agent run against a strict OpenAI gateway (Manifold) failed on the follow-up turn with `Unrecognized key: "model"`: the planner's working assistant messages carry internal fields (model, reasoning_content) that lenient servers (DeepSeek) ignore but strict ones reject. Add toOpenAiWireMessages() — keeps only spec fields (role, content, name, tool_calls, tool_call_id) and expands image attachments — and use it in all OpenAI-compatible providers (openai/custom, deepseek, openrouter, ollama). Stripping reasoning_content is also correct per DeepSeek's own guidance (must not be echoed back). Also add an env-gated real-LLM path to the mission harness (MISSION_LLM_*, defaults to the mock so deterministic/flake-soak runs are unaffected) and a skipped-by-default real-llm-mission test. Verified end-to-end: deepseek-v4-pro via Manifold reads a vault file through read_file and returns the sentinel.		2026-06-24 15:30:11 +08:00
..
e2e	fix(test): raise long-task wdio timeout 30m->50m (v4-pro pace can exceed 30m)	2026-06-14 21:13:42 +08:00
harness	fix(llm): sanitize chat messages to OpenAI wire shape for strict gateways	2026-06-24 15:30:11 +08:00
missions	fix(agent): retry empty/truncated turns instead of silently completing	2026-06-14 18:42:44 +08:00
mutation	test: widen mutation gate to {adapter, json-utils} at 90.77%	2026-06-14 20:23:54 +08:00
perf	test: add memory-leak guard and non-blocking timing metrics	2026-06-13 00:49:01 +08:00