Trailmark

Parses source code into a directed graph of functions, classes, calls, and semantic metadata for security analysis.

When to Use

• Mapping call paths from user input to sensitive functions
• Finding complexity hotspots for audit prioritization
• Identifying attack surface and entrypoints
• Understanding call relationships in unfamiliar codebases
• Security review or audit preparation across polyglot projects
• Adding LLM-inferred annotations (assumptions, preconditions) to code units
• Pre-analysis before mutation testing (genotoxic skill) or diagramming

When NOT to Use

• Single-file scripts where call graph adds no value (read the file directly)
• Architecture diagrams not derived from code (use the diagramming-code skill or draw by hand)
• Mutation testing triage (use the genotoxic skill, which calls trailmark internally)
• Runtime behavior analysis (trailmark is static, not dynamic)

Rationalizations to Reject

Installation

MANDATORY: If uv run trailmark fails (command not found, import error, ModuleNotFoundError), install trailmark before doing anything else:

uv pip install trailmark

DO NOT fall back to "manual verification", "manual analysis", or reading source files by hand as a substitute for running trailmark. The tool must be installed and used programmatically. If installation fails, report the error to the user instead of silently switching to manual code reading.

Quick Start

# Auto-detect and merge every supported language under the tree
uv run trailmark analyze --language auto --summary {targetDir}

# Explicit languages (single language or comma-separated list)
uv run trailmark analyze --language rust {targetDir}
uv run trailmark analyze --language python,rust {targetDir}

# Complexity hotspots
uv run trailmark analyze --language auto --complexity 10 {targetDir}

Programmatic API

from trailmark.parse import detect_languages, supported_languages
from trailmark.query.api import QueryEngine

# Ask the installed Trailmark build what it supports
supported_languages()
detect_languages("{targetDir}")

# Prefer auto for unknown or polyglot trees; use explicit lists when needed
engine = QueryEngine.from_directory("{targetDir}", language="auto")
engine = QueryEngine.from_directory("{targetDir}", language="python,rust")

engine.callers_of("function_name")
engine.callees_of("function_name")
engine.paths_between("entry_func", "db_query")
engine.complexity_hotspots(threshold=10)
engine.attack_surface()
engine.summary()
engine.to_json()

# Run pre-analysis (blast radius, entrypoints, privilege
# boundaries, taint propagation)
result = engine.preanalysis()

# Query subgraphs created by pre-analysis
engine.subgraph_names()
engine.subgraph("tainted")
engine.subgraph("high_blast_radius")
engine.subgraph("privilege_boundary")
engine.subgraph("entrypoint_reachable")

# Add LLM-inferred annotations
from trailmark.models import AnnotationKind

engine.annotate("function_name", AnnotationKind.ASSUMPTION,
                "input is URL-encoded", source="llm")

# Query annotations (including pre-analysis results)
engine.annotations_of("function_name")
engine.annotations_of("function_name",
                       kind=AnnotationKind.BLAST_RADIUS)
engine.annotations_of("function_name",
                       kind=AnnotationKind.TAINT_PROPAGATION)

Pre-Analysis Passes

Always run `engine.preanalysis()` before handing off to genotoxic or `diagramming-code` skills. Pre-analysis enriches the graph with four passes:

1. Blast radius estimation — counts downstream and upstream nodes per

function, identifies critical high-complexity descendants

1. Entry point enumeration — maps entrypoints by trust level, computes

reachable node sets

1. Privilege boundary detection — finds call edges where trust levels

change (untrusted -> trusted)

1. Taint propagation — marks all nodes reachable from untrusted

entrypoints

Results are stored as annotations and named subgraphs on the graph.

For detailed documentation, see references/preanalysis-passes.md.

Language Selection

Do not hardcode a stale language table in downstream workflows. Ask the installed Trailmark build what it supports:

from trailmark.parse import detect_languages, supported_languages

supported_languages()
detect_languages("{targetDir}")

CLI patterns:

# Auto-detect and merge
uv run trailmark analyze --language auto {targetDir}

# Explicit list for a known polyglot target
uv run trailmark analyze --language python,rust {targetDir}

Graph Model

Node kinds: function, method, class, module, struct, interface, trait, enum, namespace, contract, library, template

Edge kinds: calls, inherits, implements, contains, imports

Edge confidence: certain (direct call, self.method()), inferred (attribute access on non-self object), uncertain (dynamic dispatch)

Per Code Unit

• Parameters with types, return types, exception types
• Cyclomatic complexity and branch metadata
• Docstrings
• Annotations: assumption, precondition, postcondition, invariant,

blast_radius, privilege_boundary, taint_propagation, finding, audit_note (last two set by augment_sarif / augment_weaudit)

Per Edge

• Source/target node IDs, edge kind, confidence level

Project Level

• Dependencies (imported packages)
• Entrypoints with trust levels and asset values
• Named subgraphs (populated by pre-analysis)

Key Concepts

Declared contract vs. effective input domain: Trailmark separates what a function declares it accepts from what can actually reach it via call paths. Mismatches are where vulnerabilities hide:

• Widening: Unconstrained data reaches a function that assumes validation
• Safe by coincidence: No validation, but only safe callers exist today

Edge confidence: Dynamic dispatch produces uncertain edges. Account for confidence when making security claims.

Subgraphs: Named collections of node IDs produced by pre-analysis. Query with engine.subgraph("name"). Available after engine.preanalysis().

Query Patterns

See references/query-patterns.md for common security analysis patterns.

See references/preanalysis-passes.md for pre-analysis pass documentation.