euthyna
用于安全审计的确定性事实生成器：哪些被删除的代码源自安全修复，以及哪些发生变更的符号从未被任何测试执行。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add euthyna@0.5.0Euthyna 以 euthyna 发布，当前版本为 0.5.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
εὔθυνα — in classical Athens, the audit every outgoing official had to submit. You did not get to simply walk away from office. You handed over your accounts and they were examined. Pass, and you left with your standing intact. Fail, and you faced trial.
euthyna is a code security audit framework for AI coding agents. It is not another scanner. It does two things: it produces the facts an agent cannot compute by reading code, and it forces every security claim the agent makes through gates before it counts as a finding.
When an AI coding agent touches security, it fails in two specific ways:
Neither is fixed by telling the agent to be more careful. euthyna changes the handshake between you and the agent:
The agent saying "I'm done" does not count. The accounts get handed over, and the gates decide.
Three fact producers — a zero-dependency Node CLI:
history — for every line a change deletes, it attributes the line to a commit
and classifies that commit from its message, its own diff, and the deleted line's
content (a deleted if (!authorized) is flagged even under a "tweaks" subject). By
default the attribution traces the commit that first introduced the content with
git log -S (--no-origins falls back to blame's "last touched", which is faster
but mis-attributes a security line later moved by a formatting commit). If the
deleted code came from a security fix, that is flagged. This is git archaeology no
model can do from reading a diff.coverage — was this symbol ever actually invoked by a test? It has exactly two
answers: never invoked (established), or entered but that proves nothing about any
specific call site (unknown). It never reports "executed" — V8 coverage marks
unreachable code as covered, and "line covered → call ran" is wrong in exactly the
direction an audit cannot afford. The reasoning is in
docs/fact-contract.md §6.2.deps — what version is a dependency actually pinned to? It reads the lockfile
(package-lock.json / Cargo.lock / go.mod) and reports the resolved versions, or that a
dependency is absent from the tree entirely. This is the fact that resolves
supply-chain claims ("the app uses a vulnerable version of X"): the version-to-CVE
mapping is left to the adjudication layer, exactly as the fact contract requires. See
docs/fact-contract.md §6.3.The skill is the audit discipline itself, as loadable Markdown. Every security claim must pass six gates — reachability, trust boundary, real impact, and their counterparts. A claim that cannot produce evidence is downgraded to an observation, not reported as a finding. "I'm done" becomes a package: claims, evidence, and the commands that reproduce both.
The gates are not only prose. euthyna gate <report> reads an adjudication report and
mechanically checks every finding against its verdict — evidence down to path:L123, a
reproduce command, an impact statement, consistent gate statuses — and downgrades
whatever does not measure up. --verify re-runs the reproduce commands: git commands
by default, interpreter commands (node/npm/python) only with the explicit
--allow-exec, because an interpreter command from a report is arbitrary code and the
flag is the caller vouching for that report.
Prerequisite for everything: Node >= 20 and git. There is nothing else to install — the project is deliberately zero-dependency.
| Host | The skill (audit discipline) | The CLI (fact producers) |
|---|---|---|
| DSH | dsh plugin --profile web add euthyna — the npm package mounts its own skill; or copy .agents/skills/euthyna/ into ~/.agents/skills/ (user-wide) or <project>/.agents/skills/. Markdown hot-reloads; no restart needed. | npm install -g euthyna — runs in any terminal |
| Claude Code | /plugin marketplace add slow-stack/euthyna then /plugin install euthyna@euthyna, or copy .agents/skills/euthyna/ into ~/.claude/skills/ | Same |
| Codex | The same Markdown layer works; packaging goes through Codex's plugin/marketplace format | Same |
| Hermes | Copy the same folder into ~/.hermes/skills/ under a category folder, or install from a repo with hermes skills install — and for the slash command, install the repo as a plugin (see below) | Same |
| OpenCode | Copy the same folder into ~/.agents/skills/ or ~/.config/opencode/skills/ (OpenCode loads both; unknown frontmatter fields are ignored). For a /euthyna slash command, copy .opencode/commands/euthyna.md into ~/.config/opencode/commands/ — it is a prompt template that tells the model to run the CLI | Same |
| OpenClaw | Copy the same folder into ~/.agents/skills/ (or <workspace>/.agents/skills/) — OpenClaw exposes every user-invocable skill as a slash command natively, so /euthyna works with no extra file. Published on ClawHub as @modusensus/euthyna (zh) and @modusensus/euthyna-en (en): clawhub install modusensus/euthyna | Same |
| GitHub Copilot | Copy the same folder into .github/skills/ (project), ~/.copilot/skills/ or ~/.agents/skills/ (personal) — Copilot reads all three. For a prompt, copy into the same-named folder of your repo or profile. can also install skills from this repository |
The skill ships in two languages. The directory above is the Chinese original (euthyna);
an English mirror lives at .agents/skills/euthyna-en/ (skill name euthyna-en) — install
that one for an English-speaking agent, or translate the folder yourself. Both define the
same discipline, and euthyna gate accepts report markers in either language.
Two honest notes:
npm install -g euthyna) or
keep this repository checked out; on a host without the CLI, the skill requires the
unmeasurable criteria to be recorded as not evaluated rather than guessed at — that
fallback is the design, not a gap.--lang en (any command) or set EUTHYNA_LANG=en; the default never changes for
existing users. The skill text and the CLI's reports exist in both languages — the
English reader should expect Chinese only from a run that did not ask for English.euthyna is user-invoked, not auto-triggering. The skill declares
disable-model-invocation: true, so the model will not load it on its own —
you have to ask for the audit explicitly. That is also how you tell it apart
from other security skills you may have installed: if you did not name it,
it did not run.
Say it plainly in conversation:
「用 euthyna 审计一下
D:\some-project，审完再交付」
The agent then loads the skill and runs the audit discipline: it produces the
deterministic facts (history / coverage / deps) first, walks every
suspect through the six gates, and writes a report file instead of only
replying in chat.
| Host | Command | Setup |
|---|---|---|
| Claude Code | /euthyna <path> | copy .claude/commands/euthyna.md from the package into ~/.claude/commands/ (user-wide) or .claude/commands/ (project) |
| Codex | — | Codex's custom-prompt slash commands (~/.codex/prompts/) are deprecated upstream; the same Markdown can still be placed there manually |
| DSH | /euthyna | installed with the plugin; after a restart the command prints the usage manual (DSH slash commands run without reaching the model) |
| Hermes | /euthyna <subcommand> <args…> | install this repository as a Hermes plugin: hermes plugins install slow-stack/euthyna then hermes plugins enable euthyna (needs Node >= 20 on PATH; the repo root carries plugin.yaml + __init__.py). The command forwards to the CLI and prints its output without reaching the model. The skill itself can still be installed separately with hermes skills install slow-stack/euthyna/.agents/skills/euthyna |
| OpenCode | /euthyna <args…> | copy .opencode/commands/euthyna.md into ~/.config/opencode/commands/ or .opencode/commands/ |
| GitHub Copilot | /euthyna <target> | copy .github/prompts/euthyna.prompt.md into .github/prompts/ (project) or the prompt folder of your profile; available in VS Code and Copilot CLI |
| Cursor | /euthyna <target> | copy .cursor/commands/euthyna.md into .cursor/commands/ (project) or ~/.cursor/commands/ (global); plain Markdown, no frontmatter |
<PROJECT>_EUTHYNA_AUDIT_<YYYY-MM-DD>.mdBUG #N TRUE POSITIVE / FALSE POSITIVE / INCONCLUSIVEpath:L123 evidence and a reproduce commandnpm install -g euthyna
euthyna audit --repo <path> --base main --head HEAD # history + deps in one run
euthyna coverage --coverage coverage/coverage-final.json --symbol <name>
euthyna deps --repo <path> --all # or --dep <name> for a single dep
euthyna gate <adjudication-report.md> --verify --cwd <repo> # mechanically check the six gates
Or without a global install: npx euthyna audit --repo <path> --base main.
From a checkout instead (development):
git clone https://github.com/slow-stack/euthyna
cd euthyna && npm test # 207 tests; no install step exists
node bin/euthyna.js audit --repo <path> --base main --head HEAD
What history reports looks like this:
已确证 (7)
• 本次变更删除了 4 行来自提交 ab877f9d70 的代码，分布在 2 个文件。
提交信息："fix(link-in-text-block): don't match style or script text (#3775)"，分类：fix
证据: lib/checks/color/link-in-text-block-evaluate.js (ab877f9d70)
复现: git blame --porcelain -L 104,104 -L 114,114 <base> -- lib/checks/.../evaluate.js
Every deleted line is blamed back to a commit, and the 复现
(Reproduce) command lets you re-derive the claim yourself without trusting the report.
Add --json for the structured fact report and --pickaxe to detect lines that were
removed and are now being added back. Origin tracing is on by default (deleted lines
are attributed to the commit that first introduced their content); --no-origins
falls back to blame's last-modifier attribution, which is faster.
Surveying eight measurement plugins in this ecosystem found none of them publishes a process exit code, which makes their output unusable as a CI gate. This one does:
| Code | Meaning |
|---|---|
0 | Measured; nothing security-classified found — or a gate report fully passes |
10 | Measured; at least one security-classified fact exists — or a gate report has findings downgraded to observations |
1 | Usage error |
2 | Could not measure at all — must not be read as clean (also: a gate report that cannot be read or has no findings) |
2 being distinct from 0 is the whole point: failing to measure and measuring and
finding nothing are different things.
| Piece | What it is | Status |
|---|---|---|
| Fact producers | A zero-dependency Node CLI that answers three questions deterministically | Working, tested |
| The skill | The audit discipline itself, as loadable Markdown | Working, loadable |
| The benchmark | A blind recall measurement for the adjudication layer | Four rounds complete; the loop is one command (bench/adjudicate.js), with a deterministic golden round on CI |
This project tries to be explicit about the difference. Current state:
history attribution against real repositories — in two languages. Run against
axe-core (JavaScript) and
crewAI (Python); deleted lines attributed to the
commits that introduced them, then checked by hand against git blame. The checks
developed for that comparison now run as regression tests in the suite.coverage on real output in three formats — c8/V8 JSON, classic istanbul (jest/nyc, same
fnMap/f shape) and coverage.py JSON (format 3) — distinguishing all three states correctly, and
refusing anything that is not a recognizable coverage report instead of answering "symbol not
located" against it.deps against a real lockfile — resolved versions reported with a line-level evidence
pointer into the lockfile, and absent dependencies reported as established absences rather
than silent skips. Verified against a populated npm v3 lockfile and fixture lockfiles for
Cargo.lock and go.mod.bench/RESULTS.md,
bench/RESULTS-round2.md,
bench/RESULTS-round3.md,
and .bench/README.md documents the command) rather than a CI gate that could flake.--verify as a sandbox. It is not one, and does not claim to be: it executes the
report's commands with your privileges, git only unless --allow-exec is passed.deps beyond three formats and version facts only. pnpm/yarn/poetry lockfiles are
detected but not parsed (reported as not evaluated, never guessed at); go.mod reports the
declared requirement, not the resolved build version; and the producer never maps a version
to a CVE — that mapping is deliberately left to the adjudication layer.docs/case-study-crewai.md and
docs/case-study-axe-core.md.euthyna/
├── bin/ src/ test/ Fact producers (zero dependencies, Node >= 20)
├── .agents/skills/euthyna/ The audit discipline, as a portable skill
├── plugin.yaml __init__.py Hermes plugin: the /euthyna slash command
├── bench/ Blind recall benchmark + results
├── docs/ Design notes and case studies (English and Chinese)
├── tools/ Research and verification scripts
└── data/ DSH plugin catalog snapshot
tools/fetch-references.js downloads the upstream sources this project reads into .refs/
(gitignored). No third-party files are distributed in this repository — see
NOTICE.md for why, and for what is owed to whom.
Early, and honest about it. Working: the fact producers (euthyna audit runs history
and the full dependency surface in one command — the skill documents it as the first
command of every audit), the skill, the benchmark harness. Not yet built: the
git-history / coverage work needed to close the remaining recall gap.
The design notes are available in English and Chinese; the -zh files are the originals.
Read CONTRIBUTING.md. The short version: no AI attribution in commits,
one change per commit, and do not claim something works until you have run it.
The fact producers execute git and parse coverage output; their integrity is the product.
What is in scope, and how to report privately: SECURITY.md. Note that the
benchmark fixtures in bench/cases/ are vulnerable by construction and are not vulnerabilities.
Apache License 2.0 — see LICENSE.
/euthyna.github/prompts/euthyna.prompt.mdgh skill| Same |
| Cursor | Copy the same folder into .cursor/skills/ (project), ~/.cursor/skills/ (global) — Cursor also reads .claude/skills/ for compatibility. For a /euthyna command, copy .cursor/commands/euthyna.md into the matching commands/ folder (plain Markdown, no frontmatter) | Same |
| Any terminal | — | npm install -g euthyna, then euthyna … |
docs/dsh-stop-gate.md.euthyna gate is not a claim in prose: the suite
pins the contract per verdict (a TRUE POSITIVE without path:L123 evidence, a reproduce
command or all six gates passing is downgraded; a FALSE POSITIVE needs a failing gate with
a reason), and --verify is tested to not execute an interpreter command without
--allow-exec. test/skill.test.js fails CI if the skill text stops declaring the six
gates, the three verdicts, or the command that enforces them.bench/adjudicate.js runs a whole round —
blind tree, one adjudicator process per case, machine-validated reports, scoring — and a
deterministic golden round runs on every push, so the pipeline (and score.js going red
on a wrong verdict) is checked without a model.