bonsai-ninja
Project maturity: bonsai-ninja is an ambitious early-stage project. Compiler-backed analysis and security modeling across 20 languages leave a lot of room for parser gaps, unresolved dynamic behavior, incorrect findings, performance problems, and ordinary bugs. The current local release gates pass and the tool is ready for people to use, but it is not perfect and should not be the sole basis for a security decision. We are publishing it now to gather real-world feedback, failing examples, rule contributions, and engineering help from the community.
bonsai-ninja is a local code-intelligence and static-analysis engine. It maps repositories, resolves symbols and calls, traces behavior across files, inspects dataflow, exports graph facts, and reports source-to-sink security findings from one compiler-style pipeline.
The project is MIT licensed. It does not require a hosted service, upload source code, or reserve analysis features for a paid tier.
Use cases
Code intelligence for agents and developers
Give agents facts, not file dumps. With tight symbol, file, and kind selectors, bonsai-ninja lets an agent ask for the smallest useful slice of a repository: the definition, callers, references, arguments, path, backward slice, raw dataflow, or source file around one symbol. That means less prompt waste, less repeated reading, and answers tied to compiler evidence.
Use it to map an unfamiliar repository, find the code behind a symptom, follow behavior across files, review change impact, or debug the analyzer itself with AST, HIR, CFG, resolver, call-edge, and taint diagnostics.
Security review
Use the same compiler facts to inventory sources, sinks, sanitizers, and dependencies, then prove modeled source-to-sink paths with the sparse IDG fixed point. Findings can be reviewed in the terminal, emitted as JSON or SARIF for automation, or shared as a standalone HTML report. The analysis is evidence for human review—not a guarantee that code is safe—and completion metadata makes unresolved static behavior visible.
Structured export for model and agent research
Native export exposes compiler, symbol, callgraph, control-flow, dataflow, and compiled-IDG facts in a machine-readable form. GraphML, Cypher, and NetworkX views support graph tooling. These artifacts can be inputs to retrieval, training-data construction, evaluations, code-reasoning experiments, or tool-using agents; bonsai-ninja produces the evidence and does not train or validate a model by itself. The versioned native contract is published as JSON Schema v7.
Our small exploratory tests produced encouraging results, but they are not a general model-quality claim. We would love to see independent teams take the idea further, publish reproducible evaluations, and tell us where it fails— whether that is OpenAI, Anthropic, Google, Poolside/Laguna, Qwen, DeepSeek, academic and independent labs, or local-model hobbyists.
What the pipeline provides
| Capability | Practical benefit |
|---|---|
| Tree-sitter compiler frontends | Parse 20 languages into typed declarations, calls, imports, values, control flow, and dataflow facts |
Focused search, refs, calls, and read-file |
Retrieve a small, source-backed context slice before asking for heavier semantics |
Bounded symbol-summary and inspect --graph-flow packets |
Retrieve source, signature, imports, direct resolved neighbors, and unresolved-call evidence without recursively expanding a call tree |
Compiler-resolved trace, compressed path, and slice |
Follow compiler-evidenced behavior across files and report unresolved dynamic edges instead of inventing them |
| AST, HIR, CFG, resolver, edge, and taint diagnostics | Inspect both the target program and the analyzer's reasoning instead of guessing from text |
| Sparse IDG taint fixed point | Complete source-to-sink reachability over the admitted static graph without a hidden depth, file, iteration, or result cap |
| Stable IDs, explicit page cursors, JSON, and the Rust SDK | Let agents cite evidence, detect when coverage continues, and automate repeatable review workflows |
| JSON, GraphML, Cypher, and NetworkX export | Feed structured compiler evidence and explicit completeness metadata into downstream research and tooling |
| Local execution and external caches | Keep source on the machine while reusing validated compiler work across queries |
Scale, measured
30,055 source files. Completed modeled analysis. Warm navigation in seconds. The current Elasticsearch measurements use a 3 GiB semantic-worker scheduling budget and separate the first explicit semantic index from commands run after it exists:
| Cache state and operation | Measured time | Result |
|---|---|---|
Empty cache: index --semantic |
10m 00.6s | 7.11 GB validated reusable cache under the 3 GiB scheduling profile |
| Empty analysis cache: complete production taint analysis | 30.0s | Requested analysis completed without requiring the whole semantic prewarm |
| After index: semantic generation reopen | 2.3s | Existing compiler objects, linkage, callgraph, retrieval, and IDG validated and reused |
| After index: search | 3.9s | Exact requested matches |
| After index: call lookup | 3.6s | Compiler-resolved call rows |
| After index: default inspect | 7.8s | Structural evidence for the requested target |
| After index: complete production taint analysis | 10.6s | Requested fixed point completed without a semantic cap |
| After index: default native export | 4m 05s | 4.54 GB compiler, callgraph, flow, and compiled-IDG facts |
After index: --full-propagations export |
7m 36s | 6.42 GB with the same exact propagation relation materialized as individual rows |
The measured cold operation is specifically index --semantic, which users
request when they want every reusable semantic sidecar prepared up front.
Ordinary index is the lighter syntax/declaration warm-up and does not force a
whole-workspace callgraph or IDG build; ordinary commands can also compute
their requested facts on demand.
The cold row is an intentional one-time whole-workspace prewarm, not normal command startup. The 3 GiB setting schedules semantic workers rather than limiting operating-system RSS, and compressed export changes representation, not graph meaning. Full methodology, component sizes, memory measurements, and the optimization history live in Release Readiness.
Supported languages
The release includes 20 Tree-sitter frontends:
C, C++, C#, Dart, Elixir, Erlang, Go, Java, JavaScript, Kotlin, Lua, Objective-C, Perl, PHP, Python, Ruby, Rust, Scala, Swift, and TypeScript.
Each adapter owns its grammar and language syntax; shared analysis consumes
typed compiler facts, while framework and security meaning stays in
security-patterns/. See Language Support for the
frontend contract and known dynamic limits.
The release binary embeds that YAML rulepack and uses the same loader and
validator after materializing a content-addressed OS-cache generation. Security
commands therefore work outside the source checkout; --rules-dir remains the
deterministic custom/editable-pack override.
Accuracy contract
bonsai-ninja treats analysis as a compiler pipeline:
source -> Tree-sitter adapter -> typed compiler IR -> resolver/callgraph
-> IDG fixed point -> query, security, SDK, and export views
Production taint reachability is a sparse monotone IDG fixed point. It has no BFS name search, call-depth ceiling, iteration limit, file limit, or result cap. Memory budgets can change worker concurrency, cache retention, and spill behavior; they do not change semantic scope.
Throughout this documentation, exact or exhaustive describes completion over the static facts admitted by the frontends and resolver: the engine does not silently stop that modeled work at a product-imposed cap. It does not mean that the static model recovers every possible runtime behavior, that every adapter or rule is bug-free, or that an empty result proves a program safe.
Static analysis cannot resolve every runtime-generated call. Reflection,
unexpanded macros, computed imports, dynamic dispatch, and metaprogramming can
lack enough source evidence. In those cases the tool reports diagnostics or
analysis_incomplete_reasons; it does not invent a resolved edge. Always
check completion metadata before treating an empty result as proof that no
path exists.
Pagination is presentation only. If output reports another page, follow the
printed page or cursor before claiming complete coverage. --all removes
paging for an intentional exhaustive artifact; it does not enable a different
analysis mode.
Install
Install the published Rust package from crates.io:
# Requires Rust 1.88 or newer.
To replace an older Cargo-installed release with the current one:
Or build the release binary from a checkout:
The tag workflow publishes the same version to crates.io and builds release archives for Linux, macOS, and Windows on x64 and arm64. See Platform And Architecture Support for source-build requirements and parser delivery constraints on other targets. When a tagged archive is published, verify its SHA-256 checksum and GitHub/Sigstore provenance before installing:
Quickstart
Use the release binary. Add --no-color --no-progress for scripts and agent
workflows.
# Explain workspace roots and language coverage.
# Find an anchor before requesting heavier semantic work.
# Inspect the target and request raw dataflow only when needed.
# Get one bounded compiler packet without expanding transitive paths.
# Run production-oriented security analysis.
# Or write an exhaustive SARIF artifact for CI.
The complete walkthrough is in Getting Started.
For any unfamiliar option, use the binary's --help and the
CLI Reference.
Choose the smallest command
| Need | Command |
|---|---|
| Files and directories | tree |
| Workspace and language summary | context |
| Text or symbol anchor | search |
| Declarations, classes, imports, entry points | defs, classes, imports, entrypoints |
| Calls, arguments, references | calls, args, refs |
| Local facts around one target | inspect |
| Bounded compiler packet for one callable | symbol-summary |
| Exact compressed source-to-target corridor | path |
| Execution trace from an entry | trace |
| Backward influence around a symbol | slice |
| One source file and connected context | read-file |
| Reopen a stable result ID | show |
| Parser or semantic internals | dump-*, diagnostics |
| Security model or findings | security |
| Downstream graph artifact | export |
tree is a direct filesystem walk. It does not initialize the compiler,
rulepack, callgraph, IDG, or security engine. Syntax inventory commands also
avoid whole-workspace semantic work unless their requested result requires it.
Go deeper
- Documentation home
- Getting Started
- Concepts
- CLI Reference
- Language Support
- Output Formats
- Configuration
- CI Integration
- Rule Authoring Tutorial
- Rust SDK
- Contributing
- Release Readiness
Generated and executable coverage evidence lives in Taint Coverage Matrix, Coverage Baseline, and mega_flow Coverage. Current validation evidence and release commands live only in Release Readiness.
Contributing and license
Contributions should preserve the compiler/rule boundary and include the smallest positive and negative tests that prove the behavior. Start with Contributing and the PR Review Checklist. Participation is governed by the Code of Conduct.
Use the guided GitHub issue forms for reproducible product defects, analysis quality reports, and feature proposals. Do not attach private source code, credentials, generated analysis caches, or proprietary findings to a public issue.
Report exploitable vulnerabilities through the private process in Security Policy, not through a public issue.
bonsai-ninja is licensed under the MIT License. Dependency license policy is documented in Third-Party Licenses.