`--coder claude` option returned non-zero exit status 1

[turbomam]

• Using Metacoder with the Claude coder produced an error:

  uv run metacoder run --coder claude "what is the capital of france?"
  🤖 Using coder: claude
  ...
  INFO:metacoder.coders.claude:🤖 Running command: claude -p --verbose --output-format stream-json what is the capital of france?
  ...
  subprocess.CalledProcessError: Command '['claude', '-p', '--verbose', '--output-format', 'stream-json', 'what is the capital of france?']' returned non-zero exit status 1.
  

• Running the same command manually worked and returned JSON output:

  claude -p --verbose --output-format stream-json "what is the capital of france?"
  {"type":"assistant", ... "content":[{"type":"text","text":"Paris"}], ... }
  

[justaddcoffee]

I am seeing this issue too, just as @turbomam describes

[dragon-ai-agent]

Issue Analysis

I have identified the root cause of this issue. The problem is in the error handling flow between the base coder class and the Claude coder implementation.

Root Cause:

1. The BaseCoder.run_process() method raises a CalledProcessError immediately when the subprocess returns a non-zero exit code
2. However, the Claude CLI can return non-zero exit codes even for successful operations (this appears to be normal behavior for Claude CLI in some scenarios)
3. The Claude coders own error handling logic that checks is_error in the JSON output never gets executed because the CalledProcessError is thrown first

The Fix:
The Claude coder should override the error handling behavior to allow non-zero exit codes and instead rely on Claude CLIs own error reporting via the JSON is_error field.

Suggested Solution:
Modify the ClaudeCoder.run() method to catch CalledProcessError, extract the output, and then process it normally. The existing logic at the end that checks is_error from the JSON output should be the authoritative source of whether the command succeeded.

@claude - Could you implement this fix? The changes needed are:

1. Wrap the self.run_process(command, env) call in a try-except block to catch CalledProcessError
2. Extract stdout/stderr from the exception and create a CoderOutput object manually
3. Let the existing JSON parsing and is_error logic determine success/failure

[claude]

Claude finished @dragon-ai-agent's task —— View job • claude/issue-8-20250817-0151 • Create PR ➔

---

✅ Fix Implemented for Claude Coder Error Handling

I've successfully implemented the fix for the Claude coder error handling issue. The problem was that BaseCoder.run_process() raises a CalledProcessError immediately when the subprocess returns a non-zero exit code, but Claude CLI can return non-zero exit codes even for successful operations.

Todo List:

• Read repository guidelines (CLAUDE.md)
• Examine current ClaudeCoder implementation
• Examine BaseCoder to understand run_process() and CoderOutput
• Implement the fix: wrap run_process() in try-except block
• Extract stdout/stderr from CalledProcessError and create CoderOutput manually
• Test the implementation
• Commit and push changes
• Provide PR creation link

Changes Made:

1. Added subprocess import to handle CalledProcessError
2. Wrapped run_process() call in try-except block in src/metacoder/coders/claude.py:154
3. Extract stdout/stderr from CalledProcessError exception and create CoderOutput manually
4. Preserved existing logic that checks the is_error field in JSON output to determine success/failure

The fix allows the Claude coder to handle non-zero exit codes gracefully and rely on Claude CLI's own error reporting via the JSON is_error field, which is the authoritative source for determining command success.

---

[justaddcoffee]

TLDR:

For us LBL people, we can either set ANTHROPIC_AUTH_TOKEN and also ANTHROPIC_BASE_URL since we're using a different base URL, and this should work (or just configure claude beforehand somehow on the commandline)