Incremental mode crashes with "ValueError: Call with_srcs first." in WriteCode

[diagraph-de]

Bug description
MetaGPT crashes in incremental mode during the engineer/write-code phase with:

ValueError: Call with_srcs first.

This appears to happen because the incremental code path accesses project_repo.srcs before initializing the source path via with_src_path(...).

Bug solved method
Not solved upstream yet.

Possible root cause:
In metagpt/actions/write_code.py, WriteCode.run() behaves differently for normal vs incremental mode.

Non-incremental branch:

project_repo=self.repo.with_src_path(self.context.src_workspace)

Incremental branch:

project_repo=self.repo

Later, get_codes() calls:

src_file_repo = project_repo.srcs

But in metagpt/utils/project_repo.py, project_repo.srcs raises:

ValueError("Call with_srcs first.")

if _srcs_path was never initialized.

So the incremental path seems to miss a required initialization step before using project_repo.srcs.

Environment information

• LLM type and model name:

  • Ollama
  • thinkbook-coder:latest (based on qwen2.5-coder:7b)

• System version:

  • Windows 11
  • Version 10.0.26100

• Python version:

  • Python 3.11.9

• MetaGPT version or branch:

  • MetaGPT 0.8.2

• packages version:

  • metagpt 0.8.2
  • local backend via Ollama
  • installed in a local Python 3.11 virtual environment

• installation method:

  • pip install metagpt

Screenshots or logs
Relevant traceback:

Traceback (most recent call last):
  File ".../metagpt/utils/common.py", line 664, in wrapper
    return await func(self, *args, **kwargs)
  File ".../metagpt/roles/role.py", line 551, in run
    rsp = await self.react()
  File ".../metagpt/roles/role.py", line 520, in react
    rsp = await self._react()
  File ".../metagpt/roles/role.py", line 475, in _react
    rsp = await self._act()
  File ".../metagpt/roles/engineer.py", line 149, in _act
    return await self._act_write_code()
  File ".../metagpt/roles/engineer.py", line 156, in _act_write_code
    changed_files = await self._act_sp_with_cr(review=self.use_code_review)
  File ".../metagpt/roles/engineer.py", line 112, in _act_sp_with_cr
    coding_context = await todo.run()
  File ".../metagpt/actions/write_code.py", line 109, in run
    code_context = await self.get_codes(
  File ".../metagpt/actions/write_code.py", line 171, in get_codes
    src_file_repo = project_repo.srcs
  File ".../metagpt/utils/project_repo.py", line 128, in srcs
    raise ValueError("Call with_srcs first.")
ValueError: Call with_srcs first.

Additional note:
This does not appear to be an Ollama runtime issue. The local model continues working correctly; the crash occurs inside MetaGPT's incremental write-code flow.

[ztexydt-cqh]

I've submitted PR #2011 to address this issue.

[smqd19]

This issue seems to stem from the fact that the incremental mode isn't properly initializing the source path before attempting to access project_repo.srcs. From your description, it looks like the incremental branch skips the with_src_path(...) call, leading to the ValueError.

A quick fix could be to ensure that the incremental path initializes _srcs_path before accessing srcs. You might consider modifying the incremental branch to include something like:

if not self.repo._srcs_path:
    self.repo = self.repo.with_src_path(self.context.src_workspace)

This should prevent the ValueError from being raised. In terms of production experience, I’ve dealt with similar issues while implementing RAG systems where proper initialization sequences are crucial, especially when handling source paths in repositories. Always ensure that the initialization order is maintained to avoid accessing uninitialized variables.

[ztexydt-cqh]

Thanks for the detailed analysis, @smqd19! Your diagnosis matches exactly what we found — the incremental branch in was skipping the initialization before accessing , which triggered the .

I've submitted PR #2011 to fix this by ensuring is properly initialized in the incremental path. The approach aligns closely with your suggested fix. Would appreciate any feedback you have on the implementation there.