Troubleshooting¶
ModuleNotFoundError: No module named 'trace_engine'¶
The package isn't installed in the active interpreter. Either install it
(pip install -e .) or set PYTHONPATH=src. The repo launchers
(trace.bat, trace.cmd, trace.ps1) set the fallback automatically, and
pytest reads it from pyproject.toml.
Colorama reset_all noise after green tests¶
Exception ignored in atexit callback ... ValueError: I/O operation on closed
file comes from colorama touching pytest's closed capture file during
interpreter shutdown. It happens after the pass — check %ERRORLEVEL%
(cmd.exe) or $? (PowerShell/bash). 0 means green.
Model download stalls first run¶
SecureBERT/Laya resolve ~/.trace/models first, then download once (~1 GB
total). On metered machines, stay on lite or pre-seed the directory. Never
commit model binaries; they travel via Git LFS.
A scan is slow or seems to stall¶
- Pass
--no-modelsto skip neural model provisioning. - Add noisy directories (datasets, generated code) to
.traceignore, or use--scope appto leave out tests and fixtures. Non-code formats and common cache/build directories are already skipped; see What TRACE scans. - Without a reachable
--targetthe scan is static-only and finishes fast. A dead target is probed once and then skipped.
verify keeps saying INCONCLUSIVE or STILL_PRESENT¶
Restart the target after patching so it runs the new code, pass the same
--target you scanned, and run verify before rescanning (a rescan
overwrites the stored findings and can renumber IDs).
Skills do not show up in my agent¶
Run npx trace-sec install-skill --list to see what is available, then
install: npx skills add VK-Amogh/TRACE --skill trace (any harness) or
npx trace-sec install-skill. Restart the agent session so it rescans its
skills directory. Harnesses other than Claude Code read .agents/skills/ or
their own directory; the universal command above handles those.
'#' is not recognized in cmd.exe¶
# comments are PowerShell/bash syntax. In cmd.exe, use REM lines or
plain commands without comment lines.