[Mahdi-Hasan]

Description

Prepare Lesson and semantic assessment parsed the model reply with json.loads after stripping only a leading fence. With a reasoning model at high effort, this broke in two ways:

• A <think> block or a prose preface made parsing fail.
• Hidden reasoning used up the 1800/2200-token budget, so the JSON came back truncated or empty.

Changes:

• _json_response parses a complete reply first. If that fails, it:
  • removes closed <think>/<thinking> blocks
  • treats a reply that still opens with reasoning as unanswered
  • returns the first JSON object found by JSONDecoder.raw_decode scanning, so fences and prose anywhere are handled
• Truncated JSON still fails. It is never repaired into a partial object, and a nested fragment of it is never returned.
• When a reply is unusable, the request is retried once at RETRY_REASONING_EFFORT. This is the project rule in services/llm/structured_retry.py, applied here to lesson preparation and semantic assessment. Provider errors are not retried.
• Error types and messages are unchanged.

I did not reuse parse_json_response or extract_json_object:

• parse_json_response runs json_repair, which turns a truncated payload into a partial object.
• extract_json_object doesn't strip <thinking>, so it can return a JSON draft from inside the reasoning.

Here, both outcomes would hide the failure.

Related Issues

• Closes [Bug]: Prepare Lesson fails with "invalid source-selection response" when reasoning_effort is high/max #2367
• Related to fix(mastery): make source preparation demand-driven and preserve partial teaching #1954, PLEASE START HERE [Tracking] Mastery Path reliability: contributions, evidence and remaining teaching-workflow gaps #1940

Module(s) Affected

• Other: learning
• tests

Checklist

• This PR targets dev (not main — see the base-branch selector above).
• I have read and followed the contribution guidelines.
• My code follows the project's coding standards.
• I have run pre-commit run --all-files and fixed any issues.
  • All hooks pass except two that fail only on Windows:
  • deeptutor-repo-hygiene calls python3, which Windows lacks. python scripts/check_repo_hygiene.py passes.
  • mypy checks against win32 and flags fcntl/os.killpg. With --platform linux, it reports 3 errors, all in files this PR does not touch (capability_registry.py, capabilities/mastery/tools.py).
• I have added relevant tests for my changes.
• I have updated the documentation (if necessary). Not needed: no user docs cover this parsing path, and error types and messages are unchanged.
• My changes do not introduce any new security vulnerabilities.

Additional Notes

• 21 new tests in deeptutor/learning/tests/test_lesson_json_response.py:
  • think/thinking blocks, prose prefix, mid-text fence, nested braces in strings
  • truncated and unclosed-think input
  • the retry path
• 12 of them fail on dev. The invalid-input cases pass on both, which confirms the error type is the same.
• Focused learning and mastery suites: 1570 passed. The 1 failure (test_old_submission_retry_cannot_overwrite_newer_answer) also fails on dev on Windows.
• Not verified against a live OpenRouter or DeepSeek endpoint.
• At max effort, the first call is usually wasted and the retry runs at low effort. If you'd rather raise max_tokens as well, I can add that.

🤖 Generated with Claude Code