fix(worker): fall back to thinking field; strip JSON fences before reflection parse (#38)

[linhongyu510]

Problem

Two related reflection-pipeline bugs reported in #38 (unassigned, no competing PR):

Bug 1 — thinking fallback missing. ollama_chat() in docker/worker/services/llm.py only fell back to the reasoning response key when response was empty. glm-5.x-family models (e.g. glm-5.3-flash on Ollama Cloud) return their chain-of-thought under thinking. With a reasoning model and a low num_predict, response stays empty and every token goes to thinking — ollama_chat() returns "" and the reflection job fails even though the model completed fine.

Bug 2 — reflection degrades to a raw blob. reflect_on_memories() in docker/worker/tasks/reflection.py calls json.loads(response) directly. Most current chat models wrap JSON in markdown fences even when the prompt asks for raw JSON, so json.loads fails on the opening backticks, the JSONDecodeError branch fires, and the point is stored as {"raw": ...} — the structured fields (patterns/connections/insights/actions) are lost.

Fix

• _extract_response(data) helper: response → reasoning → thinking (each only if non-empty), used by ollama_chat().
• _strip_json_fences(response) helper: strips one markdown code fence (language-tagged or bare, tolerating trailing whitespace) before json.loads in reflect_on_memories().

Both are minimal, additive, and match the reporter's suggested fixes.

Tests

New pytest module docker/worker/tests/test_llm_reflection.py (the worker had no pytest infra — added deliberately small, scoped to the changed code):

• _extract_response: prefers response; falls back to reasoning; falls back to thinking (the llm.py 'thinking' fallback + reflection JSON fence-stripping — reflection points silently degrade to raw blobs #38 case); returns "" when nothing useful.
• _strip_json_fences: plain JSON passthrough; ```json fenced; bare fence + trailing whitespace.
• Regression: reflect_on_memories() with a fenced LLM response (qdrant/embedding mocked) now upserts a point whose payload text contains patterns — before the fix it stored {"raw": ...}.

Result: 8 passed (7 unit + 1 regression). Baseline check: with the old code the new tests fail to collect / the regression fails (raw-blob path), confirming red → green.

Notes

• No lint config exists in the repo; touched files pass python -m py_compile.
• Verification is local/mocked — no live Ollama or Qdrant was used.