Analyze the triggers field from SKILL.md manifest for abuse patterns.

Assert all 5 method references point to originals.

Build component_metadata list and has_executable_scripts from paths.

Build a human-readable finding message from YARA match metadata.

Build the metadata section shared by all output formats.

Build a {namespace: filepath} dict from rule files, deduplicating namespace names.

Detect unpinned MCP server command references in skill files.

Detect manifest permission patterns that suggest pre-staging for future abuse.

Detect when skill version is unpinned or uses broad constraints.

Detect hidden instructions in metadata text. Checks for: HTML comments, markdown comments, zero-width chars, base64 blobs, and data URIs.

Detect Unicode-based deception in metadata text.

Detect injection patterns in parameter definitions.

TP4: LLM-based description-behavior mismatch detection. Returns ``(findings, record)`` where *record* is the LLM-call telemetry for ``llm_cal

Collect all YARA rule files under one or more directories, sorted for determinism.

Compile YARA rules from a namespace map. Falls back to per-file compilation on error. Returns (compiled_rules, skipped_count).

Hash over rule file paths and content for cache invalidation. Uses actual file content (not just size) so that edits which preserve file leng

Count lines in a file, handling binary and errors gracefully.

Build a cross-file deduplication key from rule_id and normalized matched_text.

Keep the highest-confidence finding per (file, line, rule_id).

Return set of capability categories found in *content*.

Return the imported module name for an ``importlib.import_module('mod')`` call. Recognizes both ``importlib.import_module('os')`` and the bare-im

Compute Levenshtein edit distance between two strings.

Reconfigure stdout/stderr to UTF-8 so Unicode report output does not crash. On Windows the default console encoding (e.g. cp1252) cannot encode t

Token overhead beyond the base prompt (e.g. formatted findings). Override in subclasses that add findings text to the prompt.

Extract the first match offset and a joined matched-text snippet from a YARA match.

Extract (text, source_field, is_identifier) tuples from a manifest. Returns a list of: - (skill_name, "name", True) - (description, "

Extract (package_name, version_or_None, line_number) from package.json content.

Extract (package_name, version_or_None, line_number) from pyproject.toml. Reads PEP 621 ``[project]`` ``dependencies`` / ``optional-dependencies`

Extract (package_name, version_or_None, line_number) from requirements.txt format.

Return the host from an scp-style Git URL, or None if not scp form.

Extract skill name from SKILL.md frontmatter, falling back to directory name.

Fetch full vulnerability details for a list of IDs.

Walk children to find source calls nested inside a sink call.

Find a source call anywhere in an expression tree (handles chained calls). Handles patterns like ``open("f").read()``, ``requests.get(url).text``

Return the first tainted variable referenced in *node*, or None. Handles Name references, container literals (dict, list, tuple, set), and f-

Find references to tainted variables in a call's arguments and keywords.

Return the char offset of the first Unicode Tag character that is *not* part of a well-formed emoji tag sequence, or ``None`` if there is none.

Generate JSON report string.

Generate Markdown report string.

Format manifest for the LLM prompt.

Generate Rich terminal output and export as string.

Build a human-readable summary string from OSV results.

Get the Unicode script prefix from a character's name. Uses unicodedata.name() to get script information. Returns a short script label (e.g.

Infer file type from path (extension).

Infer file type from path (extension).

Sync retry loop — acquire slot, call LLM, release, retry on 429.

Return True for authored eval datasets that contain test-case prose.

Check if path is a direct file URL.

Return True for standard Docker/Podman commands that are not parameter abuse.

Return True when the matched text is a known-safe install or fetch pattern.

Return True if *path* looks like a test file (test_* or *_test.*).

Return the popular package name if pkg_name is a close-but-not-exact match.

Compile YARA rules from built-in and optional user-supplied directories. Results are cached at module level and reused if directory contents have

Language Reliability indicator for the LR column.

Body of main(), wrapped by deepseek_compat context manager.

Map Agent Skills ``allowed-tools`` tool names to capability category names.

Map declared permission strings to capability category names.

Run compiled YARA rules against *content* and return AnalyzerFindings.

Coerce a manifest ``allowed-tools`` value into a list of tool names. Accepts the list form (``[Bash, Read]``) and the comma-separated string

Extract rule_id, severity, confidence, and description from a YARA match's meta.

Parse YAML frontmatter from SKILL.md content. Returns raw dict (no normalization).

Parse YAML frontmatter from SKILL.md content.

Build file_cache: relative path -> file contents. Uses utf-8 with replace for errors.

Return the default chat model for the endpoint that will be used.

Determine the language for a skill directory. When *cli_lang* is ``"auto"``, reads files and runs heuristic detection. Otherwise returns *cl

Resolve a call to its canonical sink name, including dynamic-import chains. Wraps :func:`resolve_call_name_typed` (type-/alias-aware resolution)

Resolve state skill_path to an existing directory Path.

Resolve the model for *slot* with per-slot env var override support. Precedence: ``SKILLSPECTOR_MODEL_{SLOT}`` env var > provider ``resolve_m

Build a deduplication key for same-file matches.

Emit SC4 findings from the static fallback list (offline mode).

Query OSV.dev and emit SC4 findings for vulnerable packages. Returns: A tuple of (findings, covered_packages) where *covered_packages* is

Return the SSL verification setting from ``SKILLSPECTOR_SSL_VERIFY``.

Warn about models not found in the provider's model registry. When ``SKILLSPECTOR_STRICT_MODEL_VALIDATION=true``, raises ``ValueError`` inste

Simple version comparison: True if v1 < v2 (numeric tuple comparison).

Walk skill directory and return sorted relative path strings. Skips _SKIP_DIRS and hidden files except those starting with .claude.

Write report_body to file or stdout. Uses sarif_report if report_body missing.

Async invoke with automatic key switching on rate-limit.