fix(agent): preserve verify-on-stop attempted final answer

[DECK6]

Summary

Fix verify-on-stop follow-up handling so the user-facing final response preserves the model's attempted final answer instead of replacing it with only the verification follow-up.

When verify-on-stop triggers, the attempted final answer is kept in conversation history for role alternation, but it was not surfaced if the follow-up produced a verification receipt/blocker. This change stores the pending attempted answer and merges it into the eventual final response unless the follow-up already includes it.

Changes

• Add _merge_verify_on_stop_final_response() helper.
• Preserve the first verify-on-stop attempted final answer while the internal verification nudge runs.
• Merge the attempted final answer with the verification follow-up before appending the final assistant message.
• Add a regression test exercising the real AIAgent.run_conversation() state machine with a stubbed provider.

Verification

• scripts/run_tests.sh tests/agent/test_verification_stop.py tests/agent/test_verification_evidence.py -q
  • 65 passed
• ruff check agent/conversation_loop.py tests/agent/test_verification_stop.py
  • All checks passed
• git diff --check
  • passed
• Independent reviewer
  • Approved

Notes

The static added-line scan flags api_key="test-key" in the new test fixture, but this is a fake test value and not a credential.

[tonydwb]

Code Review Summary

Verdict: Approved

Preserves the attempted final answer when verify-on-stop triggers a verification follow-up. The merge function correctly handles four cases: neither present, only one present, attempted embedded in verification, or both present (concatenated with separator).

✅ Looks Good

• Clean merge logic with good edge case handling
• Good test coverage: full conversation loop with two model responses
• Verifies both final_response and messages contain both answers
• Maintains role alternation invariant

---

Reviewed by Hermes Agent

[tonydwb]

Code Review Summary

Verdict: LGTM

Clean 2-file fix preserving verify-on-stop attempted final answer. When verification stops the agent, the attempted final answer should not be lost.

Looks Good

• Well-targeted fix for answer preservation
• Good merge logic (deduplication, fallback to verification)
• Proper handling of empty/None cases

---

Reviewed by Hermes Agent

[teknium1]

Thanks for isolating the lost-answer case and adding an end-to-end loop regression.

Problems

• Current main deliberately requires a genuine later verified report to supersede the pending premature report: tests/run_agent/test_verification_continuation_budget.py:126-145 asserts exactly "verified final report". The proposed helper concatenates every non-embedded pair, so it would change that established behavior as well as cover the receipt/blocker case.
• The surrounding state machine has since moved to _pending_verification_response (agent/conversation_loop.py:5207-5212) plus provenance-aware fallback handling (agent/turn_finalizer.py:64-82). The new independent pending variable does not integrate with those guards.

Suggested changes

• Distinguish a verification receipt/blocker from a substantive replacement answer, retaining the existing supersession behavior for the latter; test both cases.
• Adapt the change to the current pending-response/finalizer flow rather than adding parallel state.

Automated hermes-sweeper review.