vigiles 9.1.0 → 10.0.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +47 -63
- package/dist/audit-html.d.ts +15 -4
- package/dist/audit-html.js +15 -6
- package/dist/audit-report.d.ts +57 -1
- package/dist/audit-report.js +29 -0
- package/dist/audit-report.template.html +34 -24
- package/dist/audit-score.js +16 -6
- package/dist/audit-serve.d.ts +109 -0
- package/dist/audit-serve.js +257 -0
- package/dist/cli.js +186 -20
- package/dist/core/compile.d.ts +5 -1
- package/dist/core/compile.js +19 -10
- package/dist/leaderboard.js +24 -13
- package/dist/scan-behavioral.d.ts +85 -0
- package/dist/scan-behavioral.js +225 -0
- package/dist/scan.js +18 -8
- package/package.json +1 -1
package/dist/audit-score.js
CHANGED
|
@@ -120,13 +120,23 @@ function structure(r) {
|
|
|
120
120
|
weight: leaderboard_js_1.W_NO_CONTRACT,
|
|
121
121
|
label: "disallowedTools typo(s) that block nothing",
|
|
122
122
|
},
|
|
123
|
-
{
|
|
124
|
-
n: noContract,
|
|
125
|
-
weight: leaderboard_js_1.W_NO_CONTRACT,
|
|
126
|
-
label: "agent(s) inherit all tools (no contract)",
|
|
127
|
-
},
|
|
128
123
|
]);
|
|
129
|
-
|
|
124
|
+
// inherit-all (no `tools:` line) is ADVISORY, not graded: it's surfaced as a
|
|
125
|
+
// least-privilege NUDGE but never lowers the Structure ring. WHY: omitting the
|
|
126
|
+
// tool contract is a near-universal, legitimate authoring style (a measured OSS
|
|
127
|
+
// sweep of 122 real plugins found 109 whose only finding was this), so grading
|
|
128
|
+
// it would make `audit` cry wolf on idiomatic subagents. See reportDeductions.
|
|
129
|
+
const advisory = noContract > 0
|
|
130
|
+
? [
|
|
131
|
+
`${String(noContract)} agent(s) inherit all tools (no contract) (advisory)`,
|
|
132
|
+
]
|
|
133
|
+
: [];
|
|
134
|
+
return {
|
|
135
|
+
key: "Structure",
|
|
136
|
+
score,
|
|
137
|
+
weight: 1,
|
|
138
|
+
findings: [...findings, ...advisory],
|
|
139
|
+
};
|
|
130
140
|
}
|
|
131
141
|
function tested(r) {
|
|
132
142
|
const { score, findings } = scoreFrom([
|
|
@@ -0,0 +1,109 @@
|
|
|
1
|
+
/** A per-run server session: the secret token + the adopt allowlist. */
|
|
2
|
+
export interface ServeSession {
|
|
3
|
+
/** Crypto-random hex; embedded in the HTML, required on every mutating POST. */
|
|
4
|
+
readonly token: string;
|
|
5
|
+
/** The loopback port the server is bound to (for the Origin check). */
|
|
6
|
+
readonly port: number;
|
|
7
|
+
/**
|
|
8
|
+
* The adoptable surfaces, keyed by their repo-relative path. A POST names a
|
|
9
|
+
* path; we resolve it HERE against this set, so an off-list path is refused.
|
|
10
|
+
*/
|
|
11
|
+
readonly surfaces: ReadonlySet<string>;
|
|
12
|
+
}
|
|
13
|
+
/** The salient, transport-agnostic fields of an incoming request. */
|
|
14
|
+
export interface RequestView {
|
|
15
|
+
readonly method: string;
|
|
16
|
+
/** The URL path (no query string). */
|
|
17
|
+
readonly path: string;
|
|
18
|
+
/** The `X-Vigiles-Token` header, if any. */
|
|
19
|
+
readonly token: string | null;
|
|
20
|
+
/** The `Origin` header, if any. */
|
|
21
|
+
readonly origin: string | null;
|
|
22
|
+
/** `body.target` for an adopt POST, if any. */
|
|
23
|
+
readonly target: string | null;
|
|
24
|
+
}
|
|
25
|
+
/** What the server should do with a request — a pure, testable verdict. */
|
|
26
|
+
export type ServeDecision = {
|
|
27
|
+
readonly kind: "report";
|
|
28
|
+
} | {
|
|
29
|
+
readonly kind: "adopt";
|
|
30
|
+
readonly target: string;
|
|
31
|
+
} | {
|
|
32
|
+
readonly kind: "adopt-all";
|
|
33
|
+
} | {
|
|
34
|
+
readonly kind: "shutdown";
|
|
35
|
+
} | {
|
|
36
|
+
readonly kind: "reject";
|
|
37
|
+
readonly status: number;
|
|
38
|
+
readonly reason: string;
|
|
39
|
+
};
|
|
40
|
+
/**
|
|
41
|
+
* Constant-time token comparison (avoids a timing side-channel). Returns false
|
|
42
|
+
* for a missing/short token rather than throwing.
|
|
43
|
+
*/
|
|
44
|
+
export declare function tokenOk(provided: string | null, expected: string): boolean;
|
|
45
|
+
/**
|
|
46
|
+
* A mutating POST's Origin must be the loopback server itself (or absent — some
|
|
47
|
+
* same-origin fetches omit it, and the token already guards those). A foreign
|
|
48
|
+
* site's Origin never matches, so a cross-site POST is refused even before the
|
|
49
|
+
* token check.
|
|
50
|
+
*/
|
|
51
|
+
export declare function originOk(origin: string | null, port: number): boolean;
|
|
52
|
+
/**
|
|
53
|
+
* Resolve a client-supplied surface path against the allowlist. Returns the path
|
|
54
|
+
* only if it's a known adoptable surface — never trusts a raw path (no traversal).
|
|
55
|
+
*/
|
|
56
|
+
export declare function resolveSurface(target: string | null, surfaces: ReadonlySet<string>): string | null;
|
|
57
|
+
/**
|
|
58
|
+
* The pure router: given a request and the session, decide what to do. Every
|
|
59
|
+
* MUTATING route (adopt / adopt-all / shutdown) requires POST + a valid Origin +
|
|
60
|
+
* a valid token; GET / serves the report page (its body is CORS-protected, so a
|
|
61
|
+
* foreign site can't read it even if it requests it).
|
|
62
|
+
*/
|
|
63
|
+
export declare function decideServe(req: RequestView, session: ServeSession): ServeDecision;
|
|
64
|
+
/** A fresh crypto-random session token (32 hex chars = 16 bytes). */
|
|
65
|
+
export declare function newToken(): string;
|
|
66
|
+
/** Whether `audit` should start the live adoption server. */
|
|
67
|
+
export type ServeGate = "serve" | "skip" | "ask";
|
|
68
|
+
/**
|
|
69
|
+
* The pure serve-gate decision (option B). A plain `audit` stays a terminating,
|
|
70
|
+
* headless-safe read; the live server is only ever offered/started INTERACTIVELY
|
|
71
|
+
* and OWN-REPO (it writes specs — never into a stranger's dir):
|
|
72
|
+
* - `--no-serve`, a foreign repo, or `--json`/headless → skip (never serve).
|
|
73
|
+
* - `--serve` → serve (force, skip the prompt).
|
|
74
|
+
* - a TTY with adoptable surfaces → ask once ("open the live report?").
|
|
75
|
+
* - a TTY with nothing to adopt → skip (no point).
|
|
76
|
+
*/
|
|
77
|
+
export declare function decideServeGate(o: {
|
|
78
|
+
serveFlag: boolean;
|
|
79
|
+
noServeFlag: boolean;
|
|
80
|
+
json: boolean;
|
|
81
|
+
isTTY: boolean;
|
|
82
|
+
ownRepo: boolean;
|
|
83
|
+
adoptableCount: number;
|
|
84
|
+
}): ServeGate;
|
|
85
|
+
/** The outcome of running an adopt action, reported back to the report UI. */
|
|
86
|
+
export interface AdoptOutcome {
|
|
87
|
+
readonly ok: boolean;
|
|
88
|
+
readonly message: string;
|
|
89
|
+
}
|
|
90
|
+
export interface ServeOptions {
|
|
91
|
+
/** The per-run secret token (already injected into `html`). */
|
|
92
|
+
readonly token: string;
|
|
93
|
+
/** The adopt allowlist (repo-relative surface paths). */
|
|
94
|
+
readonly surfaces: ReadonlySet<string>;
|
|
95
|
+
/** The rendered report HTML (with the token already injected). */
|
|
96
|
+
readonly html: string;
|
|
97
|
+
/** Adopt ONE surface (the CLI passes a closure over `init --target=`). */
|
|
98
|
+
readonly runAdopt: (target: string) => Promise<AdoptOutcome>;
|
|
99
|
+
/** Adopt every surface (bare `init`). */
|
|
100
|
+
readonly runAdoptAll: () => Promise<AdoptOutcome>;
|
|
101
|
+
/** Called once the server is listening, with the URL to open. */
|
|
102
|
+
readonly onListening?: (url: string) => void;
|
|
103
|
+
}
|
|
104
|
+
/**
|
|
105
|
+
* Start the loopback adoption server. Resolves when the server shuts down (via
|
|
106
|
+
* the /shutdown route or SIGINT). Bound to 127.0.0.1 only.
|
|
107
|
+
*/
|
|
108
|
+
export declare function serveAudit(opts: ServeOptions): Promise<void>;
|
|
109
|
+
//# sourceMappingURL=audit-serve.d.ts.map
|
|
@@ -0,0 +1,257 @@
|
|
|
1
|
+
"use strict";
|
|
2
|
+
Object.defineProperty(exports, "__esModule", { value: true });
|
|
3
|
+
exports.tokenOk = tokenOk;
|
|
4
|
+
exports.originOk = originOk;
|
|
5
|
+
exports.resolveSurface = resolveSurface;
|
|
6
|
+
exports.decideServe = decideServe;
|
|
7
|
+
exports.newToken = newToken;
|
|
8
|
+
exports.decideServeGate = decideServeGate;
|
|
9
|
+
exports.serveAudit = serveAudit;
|
|
10
|
+
/**
|
|
11
|
+
* `audit --serve` — the optional one-click-local adoption server.
|
|
12
|
+
*
|
|
13
|
+
* The HTML audit report is a STATIC file: a browser can't write your repo, so by
|
|
14
|
+
* default its "Create spec" buttons just COPY the `npx vigiles init …` command.
|
|
15
|
+
* `--serve` (or the TTY prompt) instead starts a tiny LOCAL server the report
|
|
16
|
+
* POSTs to, so a button click actually runs `init` for you — without ever leaving
|
|
17
|
+
* your machine.
|
|
18
|
+
*
|
|
19
|
+
* This is the only path where `audit` WRITES via a button, so it carries the
|
|
20
|
+
* standard localhost-server hardening (the Jupyter recipe — see
|
|
21
|
+
* research/audit-serve-design.md):
|
|
22
|
+
*
|
|
23
|
+
* 1. BIND 127.0.0.1 only (never 0.0.0.0) — unreachable off the machine.
|
|
24
|
+
* 2. A per-run SECRET TOKEN (crypto-random), embedded in the served HTML and
|
|
25
|
+
* REQUIRED on every mutating POST. A foreign website can't read the token
|
|
26
|
+
* (CORS blocks reading a cross-origin GET body), so it can't forge a POST —
|
|
27
|
+
* this is the primary CSRF defense.
|
|
28
|
+
* 3. ORIGIN/Host check as belt-and-suspenders: a POST's Origin must be the
|
|
29
|
+
* loopback server itself.
|
|
30
|
+
* 4. The adopt endpoint takes a surface ID from the pre-computed ALLOWLIST (the
|
|
31
|
+
* surfaces audit already discovered), never a client-supplied path — so a
|
|
32
|
+
* forged request can't traverse outside the repo.
|
|
33
|
+
* 5. The action calls `init` IN-PROCESS (an injected runner), never a shell, so
|
|
34
|
+
* there's no command injection.
|
|
35
|
+
* 6. Worst-case blast radius is tiny: `init` writes a reversible local
|
|
36
|
+
* `.spec.ts` (no exec, no network, no model); `eject` undoes it.
|
|
37
|
+
*
|
|
38
|
+
* The pure decision logic (`decideServe`, `tokenOk`, `originOk`,
|
|
39
|
+
* `resolveSurface`) is unit-tested; the http/IO shell (`serveAudit`) is the thin
|
|
40
|
+
* v8-ignored wrapper.
|
|
41
|
+
*/
|
|
42
|
+
const node_http_1 = require("node:http");
|
|
43
|
+
const node_crypto_1 = require("node:crypto");
|
|
44
|
+
/**
|
|
45
|
+
* Constant-time token comparison (avoids a timing side-channel). Returns false
|
|
46
|
+
* for a missing/short token rather than throwing.
|
|
47
|
+
*/
|
|
48
|
+
function tokenOk(provided, expected) {
|
|
49
|
+
if (!provided || provided.length !== expected.length)
|
|
50
|
+
return false;
|
|
51
|
+
const a = Buffer.from(provided);
|
|
52
|
+
const b = Buffer.from(expected);
|
|
53
|
+
// Lengths are equal here, so timingSafeEqual is safe to call.
|
|
54
|
+
return (0, node_crypto_1.timingSafeEqual)(a, b);
|
|
55
|
+
}
|
|
56
|
+
/**
|
|
57
|
+
* A mutating POST's Origin must be the loopback server itself (or absent — some
|
|
58
|
+
* same-origin fetches omit it, and the token already guards those). A foreign
|
|
59
|
+
* site's Origin never matches, so a cross-site POST is refused even before the
|
|
60
|
+
* token check.
|
|
61
|
+
*/
|
|
62
|
+
function originOk(origin, port) {
|
|
63
|
+
if (origin === null)
|
|
64
|
+
return true; // rely on the token (a foreign site can't have it)
|
|
65
|
+
return (origin === `http://127.0.0.1:${String(port)}` ||
|
|
66
|
+
origin === `http://localhost:${String(port)}`);
|
|
67
|
+
}
|
|
68
|
+
/**
|
|
69
|
+
* Resolve a client-supplied surface path against the allowlist. Returns the path
|
|
70
|
+
* only if it's a known adoptable surface — never trusts a raw path (no traversal).
|
|
71
|
+
*/
|
|
72
|
+
function resolveSurface(target, surfaces) {
|
|
73
|
+
if (!target)
|
|
74
|
+
return null;
|
|
75
|
+
return surfaces.has(target) ? target : null;
|
|
76
|
+
}
|
|
77
|
+
/**
|
|
78
|
+
* The pure router: given a request and the session, decide what to do. Every
|
|
79
|
+
* MUTATING route (adopt / adopt-all / shutdown) requires POST + a valid Origin +
|
|
80
|
+
* a valid token; GET / serves the report page (its body is CORS-protected, so a
|
|
81
|
+
* foreign site can't read it even if it requests it).
|
|
82
|
+
*/
|
|
83
|
+
function decideServe(req, session) {
|
|
84
|
+
if (req.method === "GET" &&
|
|
85
|
+
(req.path === "/" || req.path === "/index.html")) {
|
|
86
|
+
return { kind: "report" };
|
|
87
|
+
}
|
|
88
|
+
const mutating = req.path === "/adopt" ||
|
|
89
|
+
req.path === "/adopt-all" ||
|
|
90
|
+
req.path === "/shutdown";
|
|
91
|
+
if (!mutating) {
|
|
92
|
+
return { kind: "reject", status: 404, reason: "not found" };
|
|
93
|
+
}
|
|
94
|
+
if (req.method !== "POST") {
|
|
95
|
+
return { kind: "reject", status: 405, reason: "method not allowed" };
|
|
96
|
+
}
|
|
97
|
+
if (!originOk(req.origin, session.port)) {
|
|
98
|
+
return { kind: "reject", status: 403, reason: "bad origin" };
|
|
99
|
+
}
|
|
100
|
+
if (!tokenOk(req.token, session.token)) {
|
|
101
|
+
return { kind: "reject", status: 403, reason: "bad or missing token" };
|
|
102
|
+
}
|
|
103
|
+
if (req.path === "/shutdown")
|
|
104
|
+
return { kind: "shutdown" };
|
|
105
|
+
if (req.path === "/adopt-all")
|
|
106
|
+
return { kind: "adopt-all" };
|
|
107
|
+
const target = resolveSurface(req.target, session.surfaces);
|
|
108
|
+
if (!target) {
|
|
109
|
+
return { kind: "reject", status: 400, reason: "unknown surface" };
|
|
110
|
+
}
|
|
111
|
+
return { kind: "adopt", target };
|
|
112
|
+
}
|
|
113
|
+
/** A fresh crypto-random session token (32 hex chars = 16 bytes). */
|
|
114
|
+
function newToken() {
|
|
115
|
+
return (0, node_crypto_1.randomBytes)(16).toString("hex");
|
|
116
|
+
}
|
|
117
|
+
/**
|
|
118
|
+
* The pure serve-gate decision (option B). A plain `audit` stays a terminating,
|
|
119
|
+
* headless-safe read; the live server is only ever offered/started INTERACTIVELY
|
|
120
|
+
* and OWN-REPO (it writes specs — never into a stranger's dir):
|
|
121
|
+
* - `--no-serve`, a foreign repo, or `--json`/headless → skip (never serve).
|
|
122
|
+
* - `--serve` → serve (force, skip the prompt).
|
|
123
|
+
* - a TTY with adoptable surfaces → ask once ("open the live report?").
|
|
124
|
+
* - a TTY with nothing to adopt → skip (no point).
|
|
125
|
+
*/
|
|
126
|
+
function decideServeGate(o) {
|
|
127
|
+
if (o.noServeFlag)
|
|
128
|
+
return "skip";
|
|
129
|
+
if (!o.ownRepo)
|
|
130
|
+
return "skip"; // serve writes specs → own repo only
|
|
131
|
+
if (o.serveFlag)
|
|
132
|
+
return "serve";
|
|
133
|
+
if (o.json || !o.isTTY)
|
|
134
|
+
return "skip"; // headless never serves
|
|
135
|
+
if (o.adoptableCount === 0)
|
|
136
|
+
return "skip"; // nothing to adopt
|
|
137
|
+
return "ask";
|
|
138
|
+
}
|
|
139
|
+
/* v8 ignore start — the http/IO shell; the decision logic above is unit-tested. */
|
|
140
|
+
/** Read a request body to a string, capped to avoid an unbounded read. */
|
|
141
|
+
async function readBody(req) {
|
|
142
|
+
const chunks = [];
|
|
143
|
+
let size = 0;
|
|
144
|
+
for await (const chunk of req) {
|
|
145
|
+
size += chunk.length;
|
|
146
|
+
if (size > 64 * 1024)
|
|
147
|
+
break; // an adopt POST is tiny; cap defensively
|
|
148
|
+
chunks.push(chunk);
|
|
149
|
+
}
|
|
150
|
+
return Buffer.concat(chunks).toString("utf-8");
|
|
151
|
+
}
|
|
152
|
+
function viewOf(req, body) {
|
|
153
|
+
const path = (req.url ?? "/").split("?")[0];
|
|
154
|
+
let target = null;
|
|
155
|
+
try {
|
|
156
|
+
if (body)
|
|
157
|
+
target = JSON.parse(body).target ?? null;
|
|
158
|
+
}
|
|
159
|
+
catch {
|
|
160
|
+
target = null;
|
|
161
|
+
}
|
|
162
|
+
const header = (n) => {
|
|
163
|
+
const v = req.headers[n];
|
|
164
|
+
return typeof v === "string" ? v : null;
|
|
165
|
+
};
|
|
166
|
+
return {
|
|
167
|
+
method: req.method ?? "GET",
|
|
168
|
+
path,
|
|
169
|
+
token: header("x-vigiles-token"),
|
|
170
|
+
origin: header("origin"),
|
|
171
|
+
target,
|
|
172
|
+
};
|
|
173
|
+
}
|
|
174
|
+
function sendJson(res, status, body) {
|
|
175
|
+
const payload = JSON.stringify(body);
|
|
176
|
+
res.writeHead(status, {
|
|
177
|
+
"content-type": "application/json",
|
|
178
|
+
// No CORS headers: same-origin only. A cross-origin site can fire a request
|
|
179
|
+
// but cannot read this response — and can't forge the token anyway.
|
|
180
|
+
"x-content-type-options": "nosniff",
|
|
181
|
+
});
|
|
182
|
+
res.end(payload);
|
|
183
|
+
}
|
|
184
|
+
/**
|
|
185
|
+
* Start the loopback adoption server. Resolves when the server shuts down (via
|
|
186
|
+
* the /shutdown route or SIGINT). Bound to 127.0.0.1 only.
|
|
187
|
+
*/
|
|
188
|
+
async function serveAudit(opts) {
|
|
189
|
+
const { token, surfaces, html, runAdopt, runAdoptAll, onListening } = opts;
|
|
190
|
+
// The bound port is known only after listen(); the request handler reads it via
|
|
191
|
+
// this closure. No request can arrive before the server is listening, so the
|
|
192
|
+
// late assignment is race-free.
|
|
193
|
+
let session = { token, port: 0, surfaces };
|
|
194
|
+
await new Promise((resolveServer) => {
|
|
195
|
+
const server = (0, node_http_1.createServer)((req, res) => {
|
|
196
|
+
void (async () => {
|
|
197
|
+
const body = req.method === "POST" ? await readBody(req) : "";
|
|
198
|
+
const decision = decideServe(viewOf(req, body), session);
|
|
199
|
+
switch (decision.kind) {
|
|
200
|
+
case "report":
|
|
201
|
+
res.writeHead(200, { "content-type": "text/html; charset=utf-8" });
|
|
202
|
+
res.end(html);
|
|
203
|
+
return;
|
|
204
|
+
case "adopt": {
|
|
205
|
+
const out = await runAdopt(decision.target);
|
|
206
|
+
sendJson(res, out.ok ? 200 : 500, out);
|
|
207
|
+
return;
|
|
208
|
+
}
|
|
209
|
+
case "adopt-all": {
|
|
210
|
+
const out = await runAdoptAll();
|
|
211
|
+
sendJson(res, out.ok ? 200 : 500, out);
|
|
212
|
+
return;
|
|
213
|
+
}
|
|
214
|
+
case "shutdown":
|
|
215
|
+
sendJson(res, 200, { ok: true, message: "shutting down" });
|
|
216
|
+
server.close(() => {
|
|
217
|
+
resolveServer();
|
|
218
|
+
});
|
|
219
|
+
return;
|
|
220
|
+
case "reject":
|
|
221
|
+
sendJson(res, decision.status, {
|
|
222
|
+
ok: false,
|
|
223
|
+
message: decision.reason,
|
|
224
|
+
});
|
|
225
|
+
return;
|
|
226
|
+
}
|
|
227
|
+
})().catch(() => {
|
|
228
|
+
try {
|
|
229
|
+
sendJson(res, 500, { ok: false, message: "internal error" });
|
|
230
|
+
}
|
|
231
|
+
catch {
|
|
232
|
+
/* response already sent */
|
|
233
|
+
}
|
|
234
|
+
});
|
|
235
|
+
});
|
|
236
|
+
server.on("error", () => {
|
|
237
|
+
resolveServer();
|
|
238
|
+
});
|
|
239
|
+
// 127.0.0.1 ONLY — never 0.0.0.0; the server is unreachable off the machine.
|
|
240
|
+
// Port 0 → the OS assigns an ephemeral port; we learn it after binding.
|
|
241
|
+
server.listen(0, "127.0.0.1", () => {
|
|
242
|
+
const addr = server.address();
|
|
243
|
+
const port = addr && typeof addr === "object" ? addr.port : 0;
|
|
244
|
+
session = { token, port, surfaces };
|
|
245
|
+
onListening?.(`http://127.0.0.1:${String(port)}/?token=${token}`);
|
|
246
|
+
});
|
|
247
|
+
const stop = () => {
|
|
248
|
+
server.close(() => {
|
|
249
|
+
resolveServer();
|
|
250
|
+
});
|
|
251
|
+
};
|
|
252
|
+
process.once("SIGINT", stop);
|
|
253
|
+
process.once("SIGTERM", stop);
|
|
254
|
+
});
|
|
255
|
+
}
|
|
256
|
+
/* v8 ignore stop */
|
|
257
|
+
//# sourceMappingURL=audit-serve.js.map
|