semgrep-skill

Local Semgrep security scan

Use only <skill-dir>/scripts/scan.py, where <skill-dir> contains this file. The package capability.json defines the input, result and optional invocation. Read it after the caller verifies the discovered package hash. The generated index selects the package; it does not contain or grant the invocation.

Safety contract

• Require the user's explicit authorized absolute local workspaceRoot and a directory target inside it. Never default to the current project or drive.
• The wrapper accepts only rules/local-security.yml, extracts just python-subprocess-shell-true and javascript-eval into temporary configuration before scanning, and records both hashes. The bundled legacy secret-pattern rule is not executed by this route; credential checks are outside this capability.
• It executes an already-installed trusted host CLI with an argv array and shell:false. Never execute files from the scan target as programs.
• Both the version probe and scan use --metrics off, --disable-version-check, isolated settings/log/cache paths, and a minimal subprocess environment. Caller tokens, proxies and custom Semgrep configuration are not inherited. Windows APPDATA and a computed installed Python user base/site may be retained solely to locate the existing host runtime; caller PYTHONPATH is not inherited.
• Do not use --config auto, registry URLs, Semgrep Cloud, login or uploads.
• Do not use --autofix, write reports, install Semgrep or fetch rules. The installer installs the Skill projection only and requires separate write approval.
• stdout contains rule IDs, workspace-relative paths, spans, severity and bundled rule messages. Raw source, metavariables and raw CLI diagnostics are withheld.

Invoke and interpret

From any working directory, pass one UTF-8 JSON object over stdin:

{"schemaVersion":1,"workspaceRoot":"<authorized-absolute-local-root>","target":"<directory-within-root>"}

python "<skill-dir>/scripts/scan.py" --input-json -