Add verification stop loop

[OutThisLife]

Summary

• add default-on agent.verify_on_stop / HERMES_VERIFY_ON_STOP behavior for coding turns, with config/env escape hatches to disable
• track concrete resolved file paths for landed write_file / patch edits during a turn
• before accepting a final answer, ask the model to run canonical checks, repair failures, or explain the concrete blocker when edited code lacks fresh passing verification evidence
• keep the loop bounded and preserve message alternation by storing the attempted final answer internally followed by a synthetic user nudge
• avoid surfacing the premature final answer to users while the verification guard is asking for another repair/check turn
• clean retry/empty-response scaffolding before verification-stop nudges, matching normal final-response transcript hygiene

Stack

Stacked on #52285 (bb/verify-ledger). This consumes the passive verification ledger; it does not depend on the desktop RPC in #52286.

Test plan

• scripts/run_tests.sh tests/agent/test_verification_stop.py tests/agent/test_verification_evidence.py tests/agent/test_coding_context.py tests/run_agent/test_file_mutation_verifier.py -- -q

[tonydwb]

Code Review Summary

Verdict: Approved

Well-structured verify-on-stop loop implementation. The feature is well-scoped with proper config/env escape hatches, bounded retry loop, and clean separation of concerns. Good test coverage.

Changes:

• agent/conversation_loop.py: Verification check before accepting final answer
• agent/tool_dispatch_helpers.py: New _extract_landed_file_mutation_paths function
• agent/turn_context.py: Per-turn verification state initialization
• agent/verification_stop.py: New module with build_verify_on_stop_nudge and verify_on_stop_enabled
• hermes_cli/config.py: New verify_on_stop config option
• run_agent.py: Re-export of _extract_landed_file_mutation_paths
• tests/agent/test_verification_stop.py: Comprehensive test suite
• tests/run_agent/test_file_mutation_verifier.py: Updated tests for landed paths

LGTM.

Reviewed by Hermes Agent

[gmaxwell]

This seems to be having pretty bad effects for me with multiple backend llms including qwen 27b. The agent is misinterpreting the prods from it as instructions to go run the code it just wrote even when I keep telling it to not do so I can review it. It's also mistaking it as other commands from me and just confusing thing. I'm not sure if it's unique to me, but between this and a sudden new problem with context cache invalidation I've rolled back to a prior commit I was on. (I see I could just turn this off .. but I doubt this is the cause of the cache problem).

[5tevebaker]

This seems to be having pretty bad effects for me with multiple backend llms including qwen 27b. The agent is misinterpreting the prods from it as instructions to go run the code it just wrote even when I keep telling it to not do so I can review it. It's also mistaking it as other commands from me and just confusing thing. I'm not sure if it's unique to me, but between this and a sudden new problem with context cache invalidation I've rolled back to a prior commit I was on. (I see I could just turn this off .. but I doubt this is the cause of the cache problem).

You aren't the only one. This had all kinds of issues for me. I'm glad I finally happened upon a way to disable it. Running GPT-5.6 sol and this was causing failed verification loops because it insisted the verification script be written and run from somewhere the agent didn't have access to.