fix: repair broken tool_call arguments at the recovery write site

[guigui0711]

Summary

Follow-up micro-improvement to #15348 (7a192b12), as suggested by @teknium1 in #14518:

Your proposed approach of sanitizing at the write site is slightly cleaner as defense-in-depth. If you'd like to land that as a follow-up micro-improvement on top of the existing fix, a new PR narrowly scoped to the recovery append site at line 11711 would be a good candidate.

The loop-top _sanitize_tool_call_arguments() pass (added in #15348) repairs corrupted entries on the next iteration, but the recovery append site still writes broken JSON into messages[] first. This creates a one-iteration window where corrupted data exists in the list and can be persisted to session state.

This PR repairs arguments at the write site using the existing _repair_tool_call_arguments() helper, so corrupted JSON never enters the message history at all.

Changes

• 1 file changed, 11 insertions, 1 deletion
• Narrowly scoped to the recovery append site at run_agent.py:11911

Ref: #4662, #14518

cc @alt-glitch

🤖 Generated with Claude Code

[guigui0711]

Hi @teknium1 @alt-glitch — this is the follow-up micro-improvement suggested in #14518, narrowly scoped to the recovery append site. Would appreciate a review when you have a chance 🙏

[teknium1]

Thanks for the narrowly scoped defense-in-depth follow-up. The underlying write-time gap still exists on current main: agent/conversation_loop.py:4519-4522 detects invalid arguments, and agent/conversation_loop.py:4571-4572 appends the built assistant message before recovery tool results are added.

Problems

• The live loop no longer resides in run_agent.py. Commit 053025238434cfbf121873977b39888d7f27d1c1 extracted it to agent/conversation_loop.py; this PR's run_agent.py hunk will not affect the active recovery path.
• The diff adds no regression coverage for the third-invalid-response recovery branch.

Suggested changes

• Apply the same repair to recovery_assistant at agent/conversation_loop.py:4571, before messages.append() at line 4572.
• Add an end-to-end loop test that reaches this branch and asserts the recovery assistant's persisted/replayed tool_calls[*].function.arguments are valid JSON while tool-call/result pairing remains intact.

Automated hermes-sweeper review.

[teknium1]

Current main extracted the live run_conversation loop into agent/conversation_loop.py; the active recovery branch is now at agent/conversation_loop.py:4571 (commit 053025238). Salvage this write-site repair there, before the assistant message is appended.