fix: handle malformed tool names with XML tag fragments

[juanmichelini]

Summary

When LLMs like qwen3-coder-next emit malformed tool names such as str_replace </parameter or str_replace</function>, extract the first valid identifier and map it to the correct tool.

This addresses issue #2818 which reported a 77.1% error rate for qwen3-coder-next, with the most common error (35.8%) being Tool 'str_replace </parameter' not found.

Changes

• Add _try_fix_malformed_tool_name() function in openhands-sdk/openhands/sdk/agent/utils.py
• Update normalize_tool_call() to use the new function before applying aliases
• Add 6 regression tests for malformed tool name handling

Testing

• All 20 tests in test_tool_call_compatibility.py pass
• Pre-commit hooks pass

Fixes #2818

---

Agent Server images for this PR

• GHCR package: https://github.com/OpenHands/agent-sdk/pkgs/container/agent-server

Variants & Base Images

Variant	Architectures	Base Image	Docs / Tags
java	amd64, arm64	eclipse-temurin:17-jdk	Link
python	amd64, arm64	nikolaik/python-nodejs:python3.13-nodejs22-slim	Link
golang	amd64, arm64	golang:1.21-bookworm	Link

Pull (multi-arch manifest)

# Each variant is a multi-arch manifest supporting both amd64 and arm64
docker pull ghcr.io/openhands/agent-server:ff8d7a6-python

Run

docker run -it --rm \
  -p 8000:8000 \
  --name agent-server-ff8d7a6-python \
  ghcr.io/openhands/agent-server:ff8d7a6-python

All tags pushed for this build

ghcr.io/openhands/agent-server:ff8d7a6-golang-amd64
ghcr.io/openhands/agent-server:ff8d7a6-golang_tag_1.21-bookworm-amd64
ghcr.io/openhands/agent-server:ff8d7a6-golang-arm64
ghcr.io/openhands/agent-server:ff8d7a6-golang_tag_1.21-bookworm-arm64
ghcr.io/openhands/agent-server:ff8d7a6-java-amd64
ghcr.io/openhands/agent-server:ff8d7a6-eclipse-temurin_tag_17-jdk-amd64
ghcr.io/openhands/agent-server:ff8d7a6-java-arm64
ghcr.io/openhands/agent-server:ff8d7a6-eclipse-temurin_tag_17-jdk-arm64
ghcr.io/openhands/agent-server:ff8d7a6-python-amd64
ghcr.io/openhands/agent-server:ff8d7a6-nikolaik_s_python-nodejs_tag_python3.13-nodejs22-slim-amd64
ghcr.io/openhands/agent-server:ff8d7a6-python-arm64
ghcr.io/openhands/agent-server:ff8d7a6-nikolaik_s_python-nodejs_tag_python3.13-nodejs22-slim-arm64
ghcr.io/openhands/agent-server:ff8d7a6-golang
ghcr.io/openhands/agent-server:ff8d7a6-java
ghcr.io/openhands/agent-server:ff8d7a6-python

About Multi-Architecture Support

• Each variant tag (e.g., ff8d7a6-python) is a multi-arch manifest supporting both amd64 and arm64
• Docker automatically pulls the correct architecture for your platform
• Individual architecture tags (e.g., ff8d7a6-python-amd64) are also available if needed

[github-actions]

Python API breakage checks — ✅ PASSED

Result: ✅ PASSED

Action log

[github-actions]

REST API breakage checks (OpenAPI) — ✅ PASSED

Result: ✅ PASSED

Action log

[github-actions]

Coverage Report •

File	Stmts	Miss	Cover	Missing
openhands-sdk/openhands/sdk/agent
utils.py	165	12	92%	105, 125–126, 199, 205, 257, 259, 261, 267, 277, 293, 347
TOTAL	22592	6483	71%

[juanmichelini]

Still some errors but looking better
https://openhands-eval-monitor.vercel.app/?run=swebench%2Flitellm_proxy-openrouter-qwen-qwen3-coder-next%2F24368359316%2F&text=qwen+coder

Will try on SWTbench which is the crucial one

[juanmichelini]

Errors are down, but score in SWT is still low. Might be interested to merge it anyway

See run https://openhands-eval-monitor.vercel.app/?run=swebench%2Flitellm_proxy-openrouter-qwen-qwen3-coder-next%2F24368359316%2F&text=qwen+coder

[juanmichelini]

@OpenHands /codereview-raosted

[openhands-ai]

I'm on it! juanmichelini can track my progress at all-hands.dev

[openhands-ai]

Summary

I performed a code review of PR #2819 which fixes handling of malformed tool names like "str_replace </parameter" or "str_replace</function>" that LLMs like qwen3-coder-next sometimes emit.

Review findings:

✅ All checks passed - No issues found

• Correctness: The implementation correctly extracts valid identifiers and maps them to tools via aliases
• Testing: All 20 tests pass, including 6 new regression tests for malformed name handling
• Security: Pure string manipulation, no injection risks
• Complexity: Minimal, focused implementation

Changes reviewed:

1. New _try_fix_malformed_tool_name() function using regex to extract the first valid identifier
2. Updated normalize_tool_call() to apply the fix before aliases/terminal fallback
3. 6 new tests covering malformed name patterns (XML tags, function tags, aliases, edge cases)

Recommendation: APPROVE - The PR is low-risk and ready to merge.