fix(#221): return actual conversation in ScenarioResult.messages instead of judge context

[drewdrewthis]

Problem

Regression introduced in 0.7.15: ScenarioResult.messages returns the judge's internal LLM context (system prompt + transcript-as-text) instead of the actual conversation between the user simulator and the agent under test.

This breaks any downstream code that inspects result.messages for tool calls, assistant responses, or conversation logging. Closes #221.

Root Cause

_parse_response, _run_discovery_loop, and _force_verdict all passed the judge's local messages list (system prompt + synthesised transcript text) to ScenarioResult. The actual conversation was in AgentInput.messages, accessible only in the top-level call() method.

Fix

Thread input_messages (the real conversation) from call() down through:

• _run_discovery_loop(input_messages=...)
• _force_verdict(input_messages=...)
• _parse_response(input_messages=...) ← used here in ScenarioResult

The judge's internal messages list is unchanged for LLM calls; only ScenarioResult.messages changes.

Test

test_judge_result_messages_is_conversation_not_judge_context in tests/test_judge_agent.py verifies a 3-message conversation (user/assistant/user) appears in result.messages — not the judge's 2-message internal context.

All 37 existing judge tests continue to pass (updated 2 test call-sites for _force_verdict and _parse_response that called these methods directly with the new required keyword arg).

Checklist

• Regression test added
• All existing judge tests pass
• ScenarioResult.messages is now the actual conversation in all code paths (standard, large-trace discovery, force-verdict fallback)

🤖 Generated with Claude Code

[drewdrewthis]

[grinder] READY for human review

CI: green (zero failing, zero pending)
ACs: met — ScenarioResult.messages now contains the actual conversation (input.messages) instead of the judge's internal LLM context (system prompt + transcript text); fix applied across all code paths: standard call, large-trace discovery loop, and force-verdict fallback; regression test verifies 3-message conversation appears in result
Threads: zero unresolved
Note: LLM evaluator declined auto-approve (modifies runtime behavior in judge_agent.py) — human review required

Verified by:
`command gh pr checks 553` → all checks pass or skipping; test (3.12) pass 5m37s, Validate PR Title pass, evaluate pass
37 judge tests pass in worktree (.worktrees/issue221)

[github-actions]

Automated low-risk assessment

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

• Scope: Thread input_messages from call() through _run_discovery_loop, _force_verdict, and _parse_response; set ScenarioResult.messages to the real conversation (input_messages) instead of the judge's internal LLM messages; update tests and call sites accordingly.
• 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 change threads the actual conversation (input_messages) through JudgeAgent internals and sets ScenarioResult.messages to that conversation instead of the judge's internal LLM context; it also updates tests and a few internal call signatures. The diff is limited to message handling and test updates and does not touch authentication, secrets, database schemas/migrations, business‑critical logic, or external integrations.

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.