fix: one shared response reader, so feedback can parse a real reply

[mohammadp1001]

What happened

The feedback agent ran for the first time ever today, in the live run following #176.
It failed all three retries with:

json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)

The position was filed EVALUATION_FAILED. But the model had produced a complete,
correct evaluation - Alphoryn threw it away. That is precisely the outcome FR-016a
exists to prevent.

Root cause

Captured from the live run's OTel logs, the response was two parts:

Part	thought	Content
[0]	True	4077-char reasoning summary
[1]	None	```json\n{ "outcome_judgment": "CORRECT" ... }

feedback_agent skipped the thought correctly (that was #174's fix, and it works),
then handed the fenced string straight to json.loads, which fails at character 0.

main_agent has stripped fences since it was written, and also skips parts whose text
is empty. feedback_agent did neither:

# main_agent.py                        # feedback_agent.py
if is_thought_part(part): continue     if is_thought_part(part): continue
text = getattr(part, "text", None)
if text and text.strip():
    raw_json = _strip_fences(...)      raw_json = part.text
    break                              break

The same logic had been written twice and the two copies drifted.

The fix

Rather than paste the two missing guards into the second copy, both agents now call
alphoryn/agents/responses.py:extract_response_json, which does all three things -
skip thoughts, skip empty parts, strip fences - in one place. _strip_fences moved out
of main_agent into that module.

Two copies of this logic is what caused the bug. One copy cannot drift.

Verified against the bytes that actually failed

Replaying the exact response captured from the failed run:

parts replayed: 2
  [0] thought=True starts="Alright, let's break d"
  [1] thought=None starts='```json\n{\n  "outcome_j'

--- old behaviour (first non-thought part, verbatim) ---
JSONDecodeError: Expecting value: line 1 column 1 (char 0)

--- new behaviour (extract_response_json) ---
parsed OK
outcome_judgment = 'CORRECT'

The old path reproduces the production error exactly; the new path recovers the
evaluation that was discarded.

Tests

700 passed, 100% coverage, ruff check alphoryn/ tests/ clean, verified under CI's
credential-less environment.

tests/unit/test_responses.py is new and covers the shared reader directly, including
test_a_fenced_answer_after_a_thought_is_the_live_failure, which encodes this exact
production response shape. The four _strip_fences tests moved out of
test_main_agent.py into it.

Note

This is the third defect in a row that only a live run could find (#176 was two). All
three were invisible to a green 637-test suite because every test stubs the LLM and the
broker. Worth remembering when judging how much a passing suite proves here.

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (5 files)

• alphoryn/agents/feedback_agent.py
• alphoryn/agents/main_agent.py
• alphoryn/agents/responses.py
• tests/unit/test_main_agent.py
• tests/unit/test_responses.py

---

_{Reviewed by step-3.7-flash · Input: 82.7K · Output: 7.5K · Cached: 84.7K}