| name | scientific-python-debugging |
| description | Debug failing scientific Python inside isanagent execution_run—tracebacks, truncation, seeds, provider mismatch. |
| requires | {"bins":[],"env":[]} |
| always | false |
Scientific Python debugging (isanagent execution harness)
Use this when execution_run returns non‑zero exit, empty stdout, traceback text, or results that look nondeterministic or environment‑wrong.
1. Read the last tool output carefully
- Stderr usually carries the traceback. If stderr mentions truncation (
... (truncated)), you hit max_output_bytes—rerun with less logging, write logs to a file under the sandbox, or tail-style smaller prints.
- Exit code on local Python REPL reflects the worker’s status flag (0 = clean, 1 = exception caught in the worker). On subprocess / shell it is the real process exit code.
2. Reproduce smaller
Shrink data (fewer rows, smaller tensor, subset of files) until the failure is obvious. Set explicit random seeds (random, numpy.random.seed / Generator, framework-specific seed args) so “flakey” runs become stable.
3. Confirm environment with execution_env_info
- Local: packages = host interpreter you run under.
- Jupyter: packages = server kernel environment—imports that work locally may fail there. Adjust install instructions for the user on the server, not in the agent sandbox.
4. Separate “code bug” from “resource bug”
- Hang / no output: try
execution_cancel, then shorter timeout_secs, smaller input, or progress prints to a file.
NameError / ImportError after an earlier success: if local REPL mode is on, you may have shadowed a name or partially executed cells—reread prior runs in the same session or restart by execution_session_close + new session.FileNotFoundError: paths must be valid inside the sandbox (or server cwd for Jupyter)—use glob_files / list_directory before assuming paths.
5. When to stop iterating
After two substantially different fixes still fail, summarize: hypothesis, evidence (stderr snippet, exit code), and what you already tried. Call ask_user only if you need a decision you cannot infer (e.g. which remote dataset, credentials, or kernel image).
References (for you, not the end user)
Python — errors and exceptions, traceback.
