[codex] Drop duplicate tool results during message repair

[enzo-adami]

Summary

• Track tool results already accepted for the current assistant tool-call group.
• Drop duplicate tool messages with the same live tool_call_id.
• Preserve distinct consecutive tool results from parallel tool calls.

Why

repair_message_sequence already drops tool messages whose tool_call_id does not belong to the most recent assistant tool-call group. It did not distinguish a valid first result from a duplicated result for the same call id, so a retry/replay artifact could leave two tool outputs for one tool call in history. Providers expect at most one tool result per tool call id.

Validation

• uv run --extra dev pytest tests/run_agent/test_message_sequence_repair.py
• uv run --extra dev ruff check agent/agent_runtime_helpers.py tests/run_agent/test_message_sequence_repair.py
• git diff --check -- agent/agent_runtime_helpers.py tests/run_agent/test_message_sequence_repair.py

[teknium1]

Thanks for the focused regression fix. This is an automated hermes-sweeper review; current main already provides this exact guarantee and broader protection.

• dba585c17 (fix(agent): deduplicate tool_call_id across the pre-API sanitizers (#58327)) landed the repair-path behavior: agent/agent_runtime_helpers.py:507-516 consumes a matched tool_call_id, causing a duplicate result to be dropped while distinct parallel IDs remain accepted.
• The same commit adds the final send-path guard in sanitize_api_messages (agent/agent_runtime_helpers.py:2614-2657) for duplicate tool results and duplicate assistant tool-call IDs.
• Regression coverage already exists in tests/run_agent/test_message_sequence_repair.py:631-671, including the duplicate-result repair case and final sanitizer case.
• The linked follow-up discussions on fix: drop duplicate tool results during message repair #58024 and fix(agent): dedup duplicate tool_call_id after compaction (#58327) #58362 explicitly identify this work as superseded by dba585c17 / fix(agent): deduplicate tool_call_id across pre-API sanitizers (#58327) #58350.

Closing as implemented on main.