fix(#161): re-parse criteria when LLM returns stringified JSON dict

[drewdrewthis]

Problem

Some LLMs serialise nested objects (like the criteria map) as a quoted JSON string instead of an inline dict. When the JudgeAgent calls .values() on that string it blows up:

AttributeError: 'str' object has no attribute 'values'

Closes #161.

Fix

After json.loads(tool_call.function.arguments), check if criteria_verdicts is a str. If so, try a second json.loads() pass. On decode failure fall back to {} so the judge still produces a result rather than crashing the run.

Test

test_judge_agent_handles_criteria_as_json_string in tests/test_judge_agent.py exercises the exact payload shape that triggered the bug — criteria returned as "{\"test_criterion\": true}" (a JSON string, not a dict). The test was failing with AttributeError before the fix and passes after.

Checklist

• Regression test added
• No behaviour change when LLM returns a normal dict
• Graceful fallback ({}) when the string is not valid JSON

🤖 Generated with Claude Code

[drewdrewthis]

[grinder] READY for human review

CI: green (zero failing, zero pending)
ACs: met — JudgeAgent now re-parses criteria when LLM returns a JSON string instead of a dict; graceful fallback to {} on parse failure; regression test test_judge_agent_handles_criteria_as_json_string verifies the exact payload shape that triggered the bug
Threads: zero unresolved
Note: LLM evaluator declined auto-approve (modifies runtime parsing logic in JudgeAgent) — human review required

Verified by:
`command gh pr checks 552` → all checks pass or skipping; test (3.12) pass 7m31s, evaluate pass, javascript-complete pass
regression test in worktree: `PASSED` (.worktrees/issue161)

[github-actions]

Automated low-risk assessment

This PR was evaluated against the repository's Low-Risk Pull Requests procedure.

• Scope: Re-parse 'criteria' if it's returned as a JSON string in JudgeAgent (fallback to {} on decode failure) and add a regression test exercising this case.
• Exclusions confirmed: no changes to auth, security settings, database schema, business-critical logic, or external integrations.
• Classification: low-risk-change under the documented policy.

The PR only adjusts how the JudgeAgent parses LLM-returned arguments (re-parsing a stringified JSON 'criteria' and falling back to an empty dict) and adds a regression test. It does not touch authentication/authorization, secrets/encryption, database schemas/migrations, business-critical logic, or external integrations, so it fits the low-risk criteria and is easy to revert if needed.

An approving review has been submitted by automation. The PR may merge once required CI checks pass.

[github-actions]

Approved by automation: PR qualifies as low-risk-change under the documented policy.