Smart Contract Sentry: an RLM harness that hunts vulnerabilities in Solidity and cites the exact lines it found them at.
Give it a .sol file and the classes to look for. For each class, code first finds the functions where the class could be, and a model then decides which of them are the real thing. A class with no candidate in the file costs nothing. Every finding comes back with the whole lines it cites, a confidence, and whether the class's own definition lets it stand. The lines are always the ones the code chose, never the model's. Output is JSON, or SARIF 2.1.0 for anything that reads static-analysis results, GitHub code scanning among them.
$ sc-sentry audit Vault.sol --class slippage --format sarif --out vault.sarif
auditing Vault.sol
slippage 1 candidate(s), 1 finding(s) standing of 1
wrote vault.sarifpython -m venv .venv && . .venv/bin/activate
pip install sc-sentry
$EDITOR .env # where the model lives: the variables are belowTo work on it, install from a clone instead. The clone also holds .env.example, which names every variable:
git clone https://github.com/qazbnm456/sc-sentry
python -m venv .venv && . .venv/bin/activate
pip install -e "./sc-sentry[dev]"
cp sc-sentry/.env.example .env && $EDITOR .envBoth commands read .env from the directory you run them in, and a variable already exported wins over the file, so RLM_MAIN_MODEL=x sc-sentry audit ... overrides it for one run. RLM_MAIN_MODEL, RLM_SUB_MODEL, RLM_BASE_URL, RLM_API_KEY and RLM_ADAPTER say where the model is. RLM_MAX_TOKENS and SC_SENTRY_TEMPERATURE are optional (see the --max-tokens option below and Three defaults that are measurements). --mode static needs none of them.
sc-sentry classes # the eight classes it hunts
sc-sentry audit contracts/*.sol # every class, JSON to stdout
sc-sentry audit Pool.sol --class reentrancy --class oracle_integrity \
--format sarif --out pool.sarif
sc-sentry audit Pool.sol --mode static # no model, no .env| option | default | |
|---|---|---|
--class |
every class | one class, or several comma-separated; repeatable |
--format |
json |
or sarif |
--out |
stdout | |
--mode |
triage |
triage, explore or static; see How it works |
--limit |
9 | findings kept per class in triage and static, which is the benchmark's cap |
--candidates |
20 | candidate functions sent to the model per class in triage |
--budget-seconds |
none | triage only: what a file may take from the command's start; see Running under a budget |
--budget-calls |
none | triage only: the most model calls for a file |
--budget-tokens |
none | triage only: the most tokens for a file |
--budget-parallel |
off | triage only: with a budget, send the classes' calls side by side |
--withdraw |
off | triage only: mark what the model ruled out as rejected, which a SARIF reader suppresses |
--max-tokens |
RLM_MAX_TOKENS, else 32768 |
per-call generation cap |
--max-calls |
4 | model calls per class |
--max-iterations |
6 | planner turns per class, in explore |
--cache |
off | see Three defaults |
A finding the class definition rejects (say, a slippage citation in a function that makes no exchange call) is kept and marked: rejected in JSON, and a suppressed result with the reason in SARIF. Whoever reads the report decides whether the screen was right, which they could not do if it silently dropped the finding. JSON also says where each finding came from: origin is model when a model decided it, and static when the harness cited it and no model reviewed it.
The harness has two halves. For each class, code (candidates/<class>.py, built on tree-sitter, so a single file works whether or not its project builds) finds the functions where the class could be and ranks them; a model then decides which are the real thing. The lines a finding cites are always the ones the code chose, never the model's.
It runs in one of three modes, chosen with --mode:
triage(the default) shows the model the candidate functions for one class, each with its code and the class's definition, and takes a verdict on each. The model can move a function up the report but cannot hide one.staticasks no model and needs no.env. The candidates are the findings, likeliest first. It is the control group for everything the model adds.exploregives a planner the file, the class and a set of tools and lets it investigate. It is kept only as a comparison withtriageand costs far more.
docs/how-it-works.md has the full behaviour of each mode, running under a budget, and what the package holds.
- How it works: the three modes in full, running under a budget, and what the package holds.
- What it scores: recall on the public corpus, why that figure is a fit, what the model adds, and how to choose a cap. The figures here are fits to a public corpus, and the document says so.
- Development: editing a class, and the defaults that are measurements rather than taste.
- Scoring it with contract-assay: the optional adapter that lets the contract-assay benchmark measure this harness.
MIT. See LICENSE.