Qwen3 tool-use: relative file paths in Write/Edit resolve to /tmp instead of the system-prompt working directory

Qwen3 tool-use: relative file paths in Write/Edit resolve to /tmp/… instead of the system-prompt working directory

Summary

When Qwen3 is driven as a coding agent through an Anthropic-compatible endpoint (Alibaba Model Studio "Token Plan") with an Anthropic-style Write tool (which requires an absolute file_path), and asked to create a file with a bare/relative filename, Qwen constructs the absolute path as /tmp/<name> rather than joining it to the working directory provided in the system prompt.

The tool call itself is well-formed and executes successfully — the file just lands in /tmp instead of the project directory, so from the user's perspective the edit "didn't happen." Passing an explicit absolute path works reliably.

This is a tool-argument grounding issue (weak use of the stated cwd), not a tool-calling failure.

Environment

• Model: qwen3.8-max-preview (Token Plan), Anthropic-compatible endpoint (/apps/anthropic/v1/messages).
• Harness: Claude Code pointed at the endpoint via ANTHROPIC_BASE_URL/ANTHROPIC_AUTH_TOKEN/ANTHROPIC_MODEL. (Full, reproducible config: https://github.com/Gerry9000/qwen-on-claude-code)
• Reproduces with the raw Messages API too (below), so it is independent of the harness.

Expected vs. actual

• Expected: given a stated cwd and a bare filename notes.txt, produce file_path = <cwd>/notes.txt.
• Actual: produces file_path = /tmp/notes.txt (or a bare notes.txt), ignoring the stated cwd.

Minimal reproduction (raw Anthropic Messages API)

System prompt states the cwd; a Write tool is offered; the user asks for a bare filename:

{
  "model": "qwen3.8-max-preview",
  "max_tokens": 512,
  "system": "You are a coding agent. The current working directory is /home/user/project. Write files relative to this directory unless told otherwise.",
  "tools": [{"name":"Write","description":"Write content to a file at an absolute path",
             "input_schema":{"type":"object","properties":{"file_path":{"type":"string"},"content":{"type":"string"}},"required":["file_path","content"]}}],
  "messages": [{"role":"user","content":"Create a file named notes.txt containing hello. Use the Write tool."}]
}

Observed file_path across 3 runs: /home/user/project/notes.txt, /home/user/project/notes.txt, notes.txt. Without the explicit "write relative to this directory" instruction (just the tool + a system-prompt cwd), it consistently emits /tmp/notes.txt.

Evidence

• The tool_use block is valid: correct id, and the streamed input_json reconstructs to valid JSON with both file_path and content. So the model is not failing to call the tool.
• In a real harness run (Create dbgW.txt containing hi), the transcript shows: Write(file_path="/tmp/dbgW.txt", content="hi") → File created successfully at: /tmp/dbgW.txt. The file was created — in /tmp, not the cwd.
• Explicit absolute path (Write to <cwd>/file.txt): 3/3 correct.
• Emphasizing the cwd in the system prompt improves but does not eliminate it (2/3 above).

Impact

Agentic coding harnesses (Claude Code, and any tool loop that gives the model an absolute-path Write/Edit tool) will silently drop edits into /tmp when the model uses a relative name — a confusing, hard-to-notice failure mode for anyone running Qwen3 as a coding agent.

Suggested resolution

• Strengthen grounding so relative/bare filenames in tool arguments are joined to the system-prompt cwd; or
• If this is expected, document it (e.g. an FAQ for configuring third-party agent harnesses) with the workaround: always pass absolute paths, or have the harness resolve relative paths to the cwd before dispatch.

Reproducible harness config and a writeup of this + related integration gotchas: https://github.com/Gerry9000/qwen-on-claude-code