fix: strip markdown code fences from reflection JSON response

[jwm4]

Problem

When running review_roadmap with self-reflection enabled, the reflection node was logging a warning:

reflection_response_not_json   content='```json\n{"passed": true, "notes": "Self-review: Roadmap is well-structured...

The LLM was returning valid JSON, but wrapped in markdown code fences (json ... ) because the prompt shows the expected format that way. This caused json.loads() to fail, triggering the fallback path.

Solution

• Add regex-based code fence stripping before JSON parsing
• Handle both complete code fences (json...\n) and truncated ones (where the closing ``` is missing)
• The fix is robust and will still work if the LLM returns plain JSON without code fences

Testing

Added two new tests:

• test_reflection_handles_json_in_code_fence - Tests complete code fence handling
• test_reflection_handles_truncated_code_fence - Tests truncated code fence handling

All 91 tests pass with 84.32% coverage.

[jwm4]

🗺️ Auto-Generated Review Roadmap

This roadmap was automatically generated by review-roadmap.
Model: anthropic-vertex/claude-opus-4-5

---

Review Roadmap: Strip Markdown Code Fences from Reflection JSON

High-Level Summary

This PR fixes a parsing issue in the reflection node of the review roadmap agent. When the LLM returns JSON wrapped in markdown code fences (```json ... ```), the json.loads() call was failing because it received the raw fenced content instead of clean JSON. The fix adds regex-based preprocessing to strip these fences before parsing, with handling for both complete and truncated fence patterns.

---

Review Order

This is a small, focused bug fix. I recommend reviewing in this order:

1. Start with the Implementation (nodes.py)

Lines to examine:

• Lines 475-490 - The new code fence stripping logic

Key questions:

• Does the regex correctly handle all edge cases (with/without json language tag, with/without newlines)?
• Is the import placement appropriate (inside the function)?

2. Then Review the Tests (test_agent_nodes.py)

Lines to examine:

• Lines 456-508 - Two new test cases

Key questions:

• Do the test cases cover the scenarios mentioned in the description?
• Are assertions sufficient to verify correct behavior?

---

Watch Outs

1. Regex Edge Cases in nodes.py

• Pattern r'^```(?:json)?\s*\n?(.*?)\n?```$': The ? after \n makes newlines optional. Verify this handles:
  • ```json{"passed": true}``` (no newlines at all)
  • ```\n{"passed": true}\n``` (no language tag)
  • Content with internal newlines in the JSON

2. Double Stripping Logic

• Lines 483-488: The fallback branch uses two separate re.sub() calls. Consider whether a truncated response like ```json\n{"passed": true (no closing fence, no closing brace) would still fail at json.loads() — that's expected, but worth confirming the flow gracefully falls back.

3. Import Placement

• The import re is added at line 475 inside the function alongside import json. This is consistent with the existing pattern but could be moved to module-level for clarity — a minor style consideration.

4. Test Coverage Completeness

• Confirm both tests actually exercise distinct code paths (one hits the main if match: branch, the other hits the else: truncated handling branch)
• Consider: Is there a test for plain JSON with no fences to ensure the fix doesn't break the happy path?

---

Existing Discussions

No comments found on this PR yet. This is a fresh review.

---

Quick Checklist

• Regex handles all fence variations (with/without language tag, newlines)
• Truncated fence handling is robust
• Plain JSON (no fences) still parses correctly
• Tests clearly map to the two code paths (complete fence vs truncated)
• No regressions to existing reflection functionality