PR #19826: fix(agent): handle non-string results in _detect_tool_failure

• Repository: NousResearch/hermes-agent
• Author: liuhao1024
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/19826

Description (problem / solution / changelog)

Summary

_detect_tool_failure() assumes its result argument is always a str and unconditionally slices it with result[:500].lower(). When custom-tools plugin handlers return a dict instead of a string, Python evaluates dict[:500] as dict.__getitem__(slice(None, 500, None)), raising KeyError.

Root Cause

In agent/display.py:830, the generic heuristic path does:

lower = result[:500].lower()

with no type guard. The _invoke_tool → registry.dispatch chain can return non-string values when plugin tool handlers return dicts.

Additionally, run_agent.py has multiple function_result[:200] slicing calls in logging paths that would also crash on dict results.

Fix

1. agent/display.py — Add isinstance(result, str) guard at the top of _detect_tool_failure(). Non-string results are converted to JSON string via json.dumps() before the existing heuristics run.

2. run_agent.py — Guard all function_result[:200] and function_result[:self.log_prefix_chars] slicing with isinstance checks, converting to str() first.

Regression Coverage

12 tests in tests/agent/test_detect_tool_failure.py:

• String results (success, error, "failed" keyword) — existing behavior preserved
• None result — existing behavior preserved
• dict result (success, error, empty) — the regression case
• list and int results — edge cases
• Terminal and memory tool dict results — tool-specific paths
• get_cute_tool_message with dict result — integration test

Fixes [Bug]: custom-tools plugin tool results cause KeyError: slice(None, 500, None) in _detect_tool_failure #19814

Changed files

• agent/display.py (modified, +10/-0)
• gateway/platforms/whatsapp.py (modified, +4/-1)
• run_agent.py (modified, +13/-8)
• tests/agent/test_detect_tool_failure.py (added, +98/-0)
• tests/tools/test_file_tools.py (modified, +37/-0)
• tools/file_tools.py (modified, +6/-6)

PR #19877: fix(agent): handle non-string tool results from custom-tools plugins

• Repository: NousResearch/hermes-agent
• Author: konsisumer
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/19877

Description (problem / solution / changelog)

Coerce non-string tool results in three layers so custom-tools plugin tools that return dicts no longer crash the agent or get rejected by OpenAI-compatible providers.

What changed and why

• agent/display.py:_detect_tool_failure — result[:500] on a dict raised KeyError(slice(None, 500, None)), which run_agent then surfaced as the misleading message "Error during OpenAI-compatible API call #N: slice(None, 500, None)". Now coerces dict/list results via json.dumps so the "error"/"failed" heuristic still matches structured payloads.
• tools/delegate_tool.py — same content[:80] slice in the sub-agent tool_trace builder (_run_single_child). Coerces non-string content before slicing and len().
• agent/transports/chat_completions.py:convert_messages — non-string tool message content was forwarded as-is to the API, returning HTTP 400 messages.X.content: Invalid input from OpenRouter / DeepSeek / Ollama / NVIDIA NIM / xAI / Kimi etc. Now serializes dict/list content with json.dumps, mirroring the existing handling in agent/anthropic_adapter.py:1525 and agent/bedrock_adapter.py.

All three paths shared the same assumption ("tool result is always str") and the same fix shape, so they are bundled.

How to test

• pytest tests/agent/test_display.py tests/agent/transports/test_chat_completions.py tests/tools/test_delegate.py -q (all green; new tests cover dict/list/None tool results in each layer).
• New unit tests:
  • TestDetectToolFailure in tests/agent/test_display.py — dict/list/None tool results no longer crash; dicts containing {"error": ...} are still detected as failures.
  • test_convert_messages_serializes_dict_tool_content in tests/agent/transports/test_chat_completions.py — dict content is JSON-serialized; string content is left untouched (no copy).
  • test_tool_trace_handles_non_string_content in tests/tools/test_delegate.py — sub-agent observability builder no longer crashes on dict tool content.

What platforms tested on

• macOS on darwin-arm64 (local)

Fixes #19814

Changed files

• agent/display.py (modified, +23/-5)
• agent/transports/chat_completions.py (modified, +18/-0)
• tests/agent/test_display.py (modified, +38/-0)
• tests/agent/transports/test_chat_completions.py (modified, +28/-0)
• tests/tools/test_delegate.py (modified, +34/-0)
• tools/delegate_tool.py (modified, +5/-2)

PR #20024: fix: normalize non-string tool results to prevent KeyError on slicing

• Repository: NousResearch/hermes-agent
• Author: Kailigithub
• State: open | merged: False
• Link: https://github.com/NousResearch/hermes-agent/pull/20024

Description (problem / solution / changelog)

Summary

Custom-tools plugins can return dict results, but handle_function_call() and downstream consumers (_detect_tool_failure, delegate tracing) assume str. When a dict reaches result[:500], Python evaluates dict.__getitem__(slice(...)) and raises KeyError.

Changes

• model_tools.py — Normalize non-string results to JSON at the return point of handle_function_call(), matching the function's -> str type annotation.
• agent/display.py — Add defensive isinstance(result, str) guard in _detect_tool_failure() before string slicing.
• tools/delegate_tool.py — Add defensive isinstance(content, str) guard in the sub-agent trace builder before slicing.

Verification

• All three files pass py_compile
• No new ruff lint issues introduced
• 166 relevant tests pass (tests/agent/, tests/run_agent/ filtered by tool keywords)
• Pre-existing test failures in test_delegate.py (credential resolution) confirmed identical on main

Closes #19814

Changed files

• agent/display.py (modified, +2/-0)
• model_tools.py (modified, +2/-0)
• tools/delegate_tool.py (modified, +2/-0)