YAML Parsing Error from LLM Output in Flow Execution

[lazarkanelov]

I have observed a crash during YAML parsing, due to malformed YAML generated by the LLM - Gemini 2.5 Pro. The following traceback is observed:

yaml.scanner.ScannerError: mapping values are not allowed here
  in "<unicode string>", line 2, column 158:
     ... all the details about your order: what dish you want (service/op ...

This error occurs when the YAML loader (yaml.safe_load) tries to parse a response from the LLM that contains a colon (:) in the value field, but the value is not quoted. This causes PyYAML to interpret it as an invalid mapping.

Full Traceback

File "C:\Dev\PocketFlow-Tutorial-Codebase-Knowledge\main.py", line 97, in <module>
  main()
File "C:\Dev\PocketFlow-Tutorial-Codebase-Knowledge\main.py", line 94, in main
  tutorial_flow.run(shared)
File "...\pocketflow\__init__.py", line 16, in run
  return self._run(shared)
...
File "...\nodes.py", line 181, in exec
  abstractions = yaml.safe_load(yaml_str)
...
yaml.scanner.ScannerError: mapping values are not allowed here
  in "<unicode string>", line 2, column 158:
     ... all the details about your order: what dish you want (service/op ...

Root Cause
• The LLM returns a YAML snippet with unquoted colons inside text values which breaks YAML’s expectation of key: value pairs.
• yaml.safe_load() fails because it interprets the second colon as the start of a new mapping.

Impact
• The flow crashes during execution, and retries do not recover.

Environment
• Windows environment
• Python environment under .venv
• PyYAML version from standard pip install

[zachary62]

This is likely because LLM messes up the output structure, and should be solved by simple retry.

I have disabled cache for retry in the latest commit. Could you retry?

You can also disable the cache with this flag:

--no-cache - Disable LLM response caching (default: caching enabled)