fix(agent): normalize empty tool-call content for strict chat APIs

[Nigelkil01]

Summary

• Normalize empty/whitespace assistant content to null for pure tool-call turns sent through strict Chat Completions paths
• Apply the same normalization in the strict tool-call sanitizer used before provider replay
• Add regression coverage for empty, whitespace-only, missing, and substantive assistant content with tool calls

Fixes #31583.

Test Plan

• ../hermes-upstream-full-20260513-201914/.venv/bin/pytest -o addopts='' tests/run_agent/test_strict_api_validation.py -q
• ../hermes-upstream-full-20260513-201914/.venv/bin/pytest -o addopts='' tests/run_agent/test_provider_parity.py -q
• ../hermes-upstream-full-20260513-201914/.venv/bin/pytest -o addopts='' tests/agent/transports/test_transport.py -q
• ../hermes-upstream-full-20260513-201914/.venv/bin/python -m py_compile run_agent.py agent/transports/chat_completions.py tests/run_agent/test_strict_api_validation.py
• git diff --check

[alt-glitch]

Related: #31175 (same root cause — empty assistant content sanitization in chat_completions.py). This PR also patches run_agent.py sanitizer and adds regression tests. Likely supersedes #31175.

[Nigelkil01]

Maintainer ping: can you confirm which PR you want to carry this fix — #31615 or #31582?

#31615 was intended to supersede #31175. It fixes the same empty-tool-call-content bug, but also covers the strict run_agent.py sanitizer and adds regression tests. Right now there are multiple overlapping PRs for the same bug, so I don’t want to keep rebasing the wrong one.

If #31615 is the preferred path, I’ll rebase it and clean up the conflicts. If not, I’m happy to fold the work into the PR you want to keep.

[poorel]

I independently reproduced this against a Bedrock-backed OpenAI-compatible Claude/Sonnet endpoint, and this transport-layer normalization matches the observed provider behavior.

Repro shape:

{
  "role": "assistant",
  "content": "",
  "tool_calls": [
    {
      "id": "call_probe_123",
      "type": "function",
      "function": {
        "name": "probe_tool",
        "arguments": "{\"x\":1}"
      }
    }
  ]
}

When replayed through the Bedrock-backed Claude/Sonnet OpenAI-compatible gateway, the request fails with:

ValidationException: messages: text content blocks must be non-empty

Changing only the assistant tool-call turn from content: "" to content: null makes the same request succeed.

I also verified this in a real Hermes tool-call flow using a custom OpenAI-compatible endpoint routing to Claude/Sonnet on Bedrock: the model called the terminal tool, the tool result was replayed, and the follow-up assistant response succeeded after normalizing the empty tool-call content to null.

So this affects custom OpenAI-compatible endpoints that route Claude/Sonnet traffic to AWS Bedrock, not just synthetic strict-schema validators.

[teknium1]

Thanks for the focused strict-provider compatibility fix. The underlying bug is still present on current main: agent/chat_completion_helpers.py:1077 constructs empty assistant content, and :1202-1266 attaches the tool calls; agent/transports/chat_completions.py:196-210 currently returns that otherwise-clean history unchanged.

Problems

• The transport has since moved to copy-on-write sanitization. The proposed direct mutation needs to be ported through needs_sanitize and mutable_msg() in agent/transports/chat_completions.py:180-255; otherwise the identity fast path at :209-210 remains active.
• The submitted tests exercise _build_api_kwargs(), but the max-iteration summary request bypasses the transport and instead calls _sanitize_tool_calls_for_strict_api() at agent/chat_completion_helpers.py:1688-1711.

Suggested changes

• Salvage the normalization into both current outgoing boundaries and add direct regression coverage for each boundary, including empty, whitespace-only, missing, substantive, and no-tool-call content.

Automated hermes-sweeper review.