PR #14444: fix(transports): drop malformed tool_call.arguments before outbound request

• Repository: NousResearch/hermes-agent
• Author: aisaacsmitchell
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/14444

Description (problem / solution / changelog)

Problem

A truncated streamed assistant response can save a tool_call with non-JSON function.arguments to conversation history. On the next turn, Anthropic-compatible proxies (LiteLLM in particular) reject the outbound request with 400 Failed to parse tool call arguments, and because the bad message is pinned in history, every retry and the direct-Anthropic fallback fail the same way. See #14443.

Fix

Add a pre-pass to ChatCompletionsTransport.convert_messages and AnthropicTransport.convert_messages that:

1. Scans every assistant message's tool_calls and json.loads() each function.arguments string.
2. Drops any entry that fails to parse. If all are dropped, the tool_calls key is removed entirely (strict providers reject tool_calls: []).
3. Injects a placeholder content if the message would otherwise be empty.
4. Strips orphan tool messages whose tool_call_id no longer has a surviving counterpart.
5. Emits a single logger.warning per call if anything was dropped.
6. Short-circuits with no mutation and no deepcopy when nothing is malformed.

The sanitization is symmetrical across both transports so the behavior is independent of which provider we're talking to. The existing anthropic_adapter.convert_messages_to_anthropic adapter-level guard is left untouched — this adds a transport-level safety net that protects providers we hit via the chat-completions path (LiteLLM, OpenRouter routing to Anthropic, etc.).

Design choice: drop vs. repair

We drop malformed tool_calls rather than repairing arguments to {}. A {}-arg tool call (e.g. a patch with no path/old_string) would execute a garbage action or surface a confusing downstream error. Dropping removes the bad call entirely; the model re-attempts on the next turn with whatever context it still has.

Tests

• tests/agent/transports/test_chat_completions.py::TestMalformedToolCallSanitization — 5 cases
• tests/agent/transports/test_anthropic.py::TestAnthropicMalformedToolCallSanitization — 6 cases (including warning log assertion)

Coverage: mixed valid/malformed, all-malformed with orphan tool result, all-clean identity (verifies no deepcopy), empty-string arguments treated as clean (not malformed), surviving-tool-result preservation, logger.warning emission.

Full suite: pytest tests/agent/transports/ — 130 passed, 0 regressions.

Files changed

• agent/transports/chat_completions.py — +113 lines (pre-pass + module logger)
• agent/transports/anthropic.py — +114 lines (pre-pass + module logger)
• tests/agent/transports/test_chat_completions.py — +163 lines (new test class)
• tests/agent/transports/test_anthropic.py — new file, 244 lines

No changes to run_agent.py, error_classifier.py, or anthropic_adapter.py.

Non-goals

• Not changing the retry loop in run_agent.py — once the bad message is cleaned out of the outbound request, the existing retry logic works.
• Not adding a new FailoverReason for this class of error — prevention at the outbound boundary is simpler and covers more cases than post-hoc classification.

Changed files

• agent/transports/anthropic.py (modified, +111/-3)
• agent/transports/chat_completions.py (modified, +108/-5)
• tests/agent/transports/test_anthropic.py (added, +244/-0)
• tests/agent/transports/test_chat_completions.py (modified, +163/-0)

---

PR #14455: fix: sanitize malformed tool_calls and orphan tool results in outbound messages

• Repository: NousResearch/hermes-agent
• Author: HiddenPuppy
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/14455

Description (problem / solution / changelog)

Summary

When a streamed assistant response is cut off mid input_json_delta, the resulting tool_call.arguments is saved to conversation history as a truncated (non-JSON) string. Every subsequent API request re-sends that history, and Anthropic-compatible proxies (LiteLLM in particular) reject it with HTTP 400. The retry loop cannot recover because the bad message is deterministically pinned in history.

Fix

Add sanitize_tool_calls_in_messages() in transports/base.py and call it from both outbound paths:

• ChatCompletionsTransport.convert_messages — covers OpenAI-compatible providers (LiteLLM, OpenRouter, etc.)
• anthropic_adapter.convert_messages_to_anthropic — covers direct Anthropic path

The sanitizer:

1. Drops assistant messages whose tool_calls contain malformed function.arguments (not valid JSON)
2. Drops orphan tool results whose tool_call_id no longer has a matching assistant tool_call

Dropping (rather than repairing to {}) is intentional — a {}-arg tool call would execute a garbage edit; dropping lets the model re-attempt on the next turn.

Verification

• tests/agent/transports/test_chat_completions.py — 5 new regression tests for malformed/valid/mixed tool_calls and orphan results
• All existing chat_completions tests pass (39 tests)
• All existing anthropic_adapter tests pass (137 tests)
• Other transport tests pass (250 total, 11 codex failures are pre-existing missing openai dependency)

Fixes #14443

Changed files

• agent/anthropic_adapter.py (modified, +4/-0)
• agent/transports/base.py (modified, +110/-1)
• agent/transports/chat_completions.py (modified, +19/-10)
• tests/agent/transports/test_chat_completions.py (modified, +91/-0)