Joern Slice

Joern Code Slicer

Use Joern after syntax-level inspection has identified a function, API, or behavior worth deeper analysis. It builds a code property graph (CPG), asks Joern for a bounded slice, and emits both the raw graph and a stable Rekit evidence document. This is a deeper layer than Tree-sitter, not a replacement for broad multi-language scanning.

Install

rekit install joern-slice
rekit doctor joern-slice

The explicit install step pulls the official image by immutable multi-platform digest. Analysis never pulls and always runs with Docker networking disabled.

Usage

rekit run joern-slice ./source ./out --language pythonsrc
rekit run joern-slice ./source ./out --language jssrc --mode data-flow \
  --sink-filter '.*(?:eval|exec).*'
rekit run joern-slice ./source ./out --language c --mode usages

behavior-flow is the default. Its declarative profile selects source and sink patterns without accepting executable Joern or Scala supplied by the caller. Pass a versioned profile with --profile; its schema is in assets/profile.schema.json. data-flow exposes Joern's bounded graph directly, while usages records method and type usages. Use --reuse-cpg when iterating on profiles against an already generated CPG from the same target and frontend. Reuse requires its sibling evidence.json and verifies the target manifest, frontend, and immutable image before accepting the CPG.

Every run publishes cpg.bin, raw-slice.json, and evidence.json atomically. The evidence records the target manifest digest, frontend, profile digest, image identity, coverage, graph, paths, findings, resource limits, and phase timing. Joern frontends vary: some graphs contain an explicit edge to the sink, while others end at the value selected by the sink query. Rekit labels the latter slice-selected-by-sink; it does not fabricate a reaching-definition edge.

Scope and safety

One run analyzes one frontend and one CPG. Mixed-language repositories require one run per relevant language; the results do not imply cross-language data flow. The default input bounds are 10,000 regular files and 512 MiB. Symlinks are excluded and reported. The target is copied into a temporary read-only mount. Target code is parsed but never executed.

The container uses a read-only root filesystem, no network, no capabilities, no privilege escalation, and CPU, memory, process, temporary-storage, and wall-clock limits. /tmp is private but executable because Joern's trusted zstd JNI library is loaded from there. Only the temporary analysis output is writable.

Treat an empty, unresolved, or excluded result as bounded evidence, not proof that a behavior is absent. Use the reported coverage and unresolved fields when composing a reverse-engineering conclusion.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: batteryshark
• Source: batteryshark/rekit
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.