fix(agent): preserve tool_name on sanitizer-injected stub tool results

[soynchux]

What

Use make_tool_result_message() when _sanitize_api_messages() injects a stub tool result for an assistant tool call whose result is missing.

This keeps sanitizer-generated tool messages consistent with normal tool execution messages by including both:

• name for provider/wire-format compatibility
• tool_name for internal session DB and JSON session log persistence

Why

Recent fixes made tool result construction populate tool_name so state.db, JSON session logs, and tool-aware search do not get blank tool names.

The pre-call sanitizer still had one manual stub-message path that only set name, content, and tool_call_id. If that path ran after a missing/dropped tool result, the generated stub could still persist with an empty tool_name.

Testing

Ran targeted tests on Windows with thread-based pytest timeout override:

uv --cache-dir .uv-cache run --extra dev pytest tests/run_agent/test_agent_guardrails.py::TestSanitizeApiMessages -q -n0 --timeout-method=thread --basetemp .pytest-tmp
uv --cache-dir .uv-cache run --extra dev pytest tests/run_agent/test_tool_name_db_persistence.py -q -n0 --timeout-method=thread --basetemp .pytest-tmp

- 7
- 1

PASSED

[teknium1]

Thanks for isolating the remaining manual stub construction.

Problems

• agent/conversation_loop.py:792-905 builds and sanitizes an api_messages copy, while run_agent.py:1692-1695 persists the canonical messages list. The proposed tool_name therefore cannot reach the session DB or JSON log.
• More importantly, agent/chat_completion_helpers.py:1705-1708 strips tool_name before agent/chat_completion_helpers.py:1728 runs this sanitizer, and then sends the result directly at agent/chat_completion_helpers.py:1852. Replacing the stub with make_tool_result_message() would re-add tool_name; agent/transports/chat_completions.py:159-163 documents that strict providers reject it.

Suggested changes

• Keep this API-bound stub schema-only, or strip tool_name again after sanitization on direct-request paths. If persistence is the target, trace the canonical message construction path and add an end-to-end persistence regression instead.

Automated hermes-sweeper review.

[teknium1]

sanitize_api_messages() operates on the per-request API copy, not the persisted history. In the iteration-summary path, tool_name is deliberately stripped before this sanitizer and the request then bypasses ChatCompletionsTransport; this helper would add tool_name back and strict providers reject that schema-foreign field. Keep this stub API-schema-only rather than using make_tool_result_message() here.