@uinaf/skillcheck 1.5.0 → 1.5.2

package/dist/cli.js

@@ -61,12 +61,13 @@ function parseArgs(argv) {
 function resolveRoot(flags) {
 	return path.resolve(flags.get("--root") ?? process.cwd());
 }
-function stateDirs(root) {
+function stateDirs(root, uid = process.getuid?.()) {
 	const base = path.join(root, ".skillcheck");
 	const id = createHash("sha256").update(path.resolve(root)).digest("hex").slice(0, 12);
+	const owner = uid === void 0 ? "" : `${uid}-`;
 	return {
 		results: path.join(base, "results"),
-		scratch: path.join(fs.realpathSync(os.tmpdir()), `skillcheck-${id}`),
+		scratch: path.join(fs.realpathSync(os.tmpdir()), `skillcheck-${owner}${id}`),
 		scorecards: path.join(base, "scorecards")
 	};
 }
@@ -137,7 +138,7 @@ function assertUniformConfig(entries, allowMixed) {
 	for (const [harness, configs] of byHarness) if (configs.size > 1) throw new Error(`${harness} results span multiple run configurations (${[...configs.values()].map(describeConfig).join("; ")}); rerun them to match or pass --allow-mixed`);
 }
 function ensureEvalPackages(opts) {
-	const missing = requiredEvalPackages(opts, process.env.ANTHROPIC_API_KEY !== void 0).filter((pkg) => resolvePackageDir(pkg) === void 0);
+	const missing = requiredEvalPackages(opts, Boolean(process.env.ANTHROPIC_API_KEY)).filter((pkg) => resolvePackageDir(pkg) === void 0);
 	if (missing.length === 0) return;
 	const peers = JSON.parse(fs.readFileSync(path.join(packageDir, "package.json"), "utf8")).peerDependencies ?? {};
 	const specs = missing.map((pkg) => `"${pkg}@${peers[pkg] ?? "latest"}"`).join(" ");

package/dist/grok-provider.js

@@ -89,8 +89,10 @@ var GrokProvider = class {
 					cost = event.total_cost_usd;
 				}
 			};
+			child.stdout?.setEncoding("utf8");
+			child.stderr?.setEncoding("utf8");
 			child.stdout?.on("data", (chunk) => {
-				stdout += chunk.toString("utf8");
+				stdout += chunk;
 				if (stdout.length > OUTPUT_CAP) {
 					stop("grok event stream exceeded limit");
 					return;
@@ -103,7 +105,7 @@ var GrokProvider = class {
 				}
 			});
 			child.stderr?.on("data", (chunk) => {
-				stderr = (stderr + chunk.toString("utf8")).slice(-4e3);
+				stderr = (stderr + chunk).slice(-4e3);
 			});
 			child.on("error", (cause) => {
 				error ??= `grok could not start: ${cause.message}`;

package/dist/scenario.js

@@ -1,3 +1,4 @@
+import { SKIPPED_DIRS } from "./transform.js";
 import { createRequire } from "node:module";
 import { createHash } from "node:crypto";
 import fs from "node:fs";
@@ -74,6 +75,41 @@ function stripHiddenFlag(skillMd) {
 	if (!range) return skillMd;
 	return skillMd.split("\n").filter((l, i) => !(i >= range[0] && i < range[1] && l.startsWith("disable-model-invocation:"))).join("\n");
 }
+const MARKDOWN_LINK = /\]\(\s*<([^>]+)>|\]\(\s*([^)\s]+)|^ {0,3}\[[^\]]+\]:\s*<?([^\s>]+)/gm;
+function linkedSiblings(skillDir) {
+	const skillsRoot = path.dirname(skillDir);
+	const realRoot = fs.realpathSync(skillsRoot);
+	const self = path.basename(skillDir);
+	const found = /* @__PURE__ */ new Set([self]);
+	const pending = [skillDir];
+	for (let dir = pending.pop(); dir !== void 0; dir = pending.pop()) for (const file of skillMarkdown(dir)) for (const m of fs.readFileSync(file, "utf8").matchAll(MARKDOWN_LINK)) {
+		const href = m[1] ?? m[2] ?? m[3];
+		if (href.startsWith("#") || /^[a-z][a-z0-9+.-]*:/i.test(href)) continue;
+		const target = path.resolve(path.dirname(file), href.split("#")[0]);
+		const rel = path.relative(skillsRoot, target);
+		const name = rel.split(path.sep)[0];
+		if (name === "" || name === ".." || path.isAbsolute(rel) || found.has(name)) continue;
+		if (!fs.existsSync(path.join(skillsRoot, name, "SKILL.md"))) continue;
+		if (path.dirname(fs.realpathSync(path.join(skillsRoot, name))) !== realRoot) continue;
+		found.add(name);
+		pending.push(path.join(skillsRoot, name));
+	}
+	found.delete(self);
+	return [...found].sort();
+}
+function skillMarkdown(dir) {
+	const out = [];
+	const walk = (d) => {
+		for (const e of fs.readdirSync(d, { withFileTypes: true })) {
+			const p = path.join(d, e.name);
+			if (e.isDirectory()) {
+				if (!(d === dir && e.name === "evals")) walk(p);
+			} else if (e.isFile() && e.name.endsWith(".md")) out.push(p);
+		}
+	};
+	walk(dir);
+	return out;
+}
 const RESERVED = /* @__PURE__ */ new Set([
 	".claude",
 	".agents",
@@ -105,10 +141,15 @@ function materialize(s, runDir, harness, control = false) {
 		fs.writeFileSync(dest, content);
 	}
 	const roots = control ? [] : harness === "codex" ? [".claude", ".agents"] : harness === "grok" ? [".grok"] : [".claude"];
-	for (const root of roots) fs.cpSync(s.skillDir, path.join(workdir, root, "skills", s.skill), {
-		recursive: true,
-		filter: (src) => path.basename(src) !== "evals"
-	});
+	const skillsRoot = path.dirname(s.skillDir);
+	const installed = control ? [] : [s.skill, ...linkedSiblings(s.skillDir)];
+	for (const root of roots) for (const name of installed) {
+		const from = path.join(skillsRoot, name);
+		fs.cpSync(from, path.join(workdir, root, "skills", name), {
+			recursive: true,
+			filter: (src) => src !== path.join(from, "evals")
+		});
+	}
 	for (const root of roots) {
 		const skillMd = path.join(workdir, root, "skills", s.skill, "SKILL.md");
 		const text = fs.readFileSync(skillMd, "utf8");
@@ -120,7 +161,7 @@ function materialize(s, runDir, harness, control = false) {
 		for (const e of fs.readdirSync(dir, { withFileTypes: true })) {
 			const p = path.join(dir, e.name);
 			if (e.isDirectory()) {
-				if (e.name !== ".claude" && e.name !== ".grok" && e.name !== "node_modules") walk(p);
+				if (!SKIPPED_DIRS.has(e.name)) walk(p);
 			} else manifest[path.relative(workdir, p)] = createHash("sha256").update(fs.readFileSync(p)).digest("hex");
 		}
 	};
@@ -151,7 +192,10 @@ function agentProvider(opts, workdir, skill, paths) {
 			sandbox_mode: "workspace-write",
 			network_access_enabled: true,
 			web_search_enabled: true,
-			cli_env: { CODEX_HOME: path.join(workdir, "..", "..", "codex-home") }
+			cli_env: {
+				CODEX_HOME: path.join(workdir, "..", "..", "codex-home"),
+				HOME: path.join(workdir, "..", "..", "home")
+			}
 		}
 	};
 	return {
@@ -293,6 +337,26 @@ function privateCodexHome(dir) {
 		if (fs.existsSync(from)) fs.symlinkSync(from, path.join(dir, file));
 	}
 }
+const HIDDEN_HOME_ENTRIES = /* @__PURE__ */ new Set([
+	".agents",
+	".claude",
+	".codex"
+]);
+function privateHome(dir) {
+	const source = os.homedir();
+	fs.rmSync(dir, {
+		recursive: true,
+		force: true
+	});
+	fs.mkdirSync(dir, {
+		recursive: true,
+		mode: 448
+	});
+	for (const name of fs.readdirSync(source)) {
+		if (HIDDEN_HOME_ENTRIES.has(name)) continue;
+		fs.symlinkSync(path.join(source, name), path.join(dir, name));
+	}
+}
 function generateRun(scenarioDir, opts, paths) {
 	const s = loadScenario(scenarioDir);
 	const name = runNameFor(scenarioDir, opts.harness, opts.control);
@@ -302,7 +366,10 @@ function generateRun(scenarioDir, opts, paths) {
 		force: true
 	});
 	const trials = Array.from({ length: opts.trials ?? 1 }, (_, i) => materialize(s, path.join(runDir, trialLabel(i)), opts.harness, opts.control));
-	if (opts.harness === "codex") privateCodexHome(path.join(runDir, "codex-home"));
+	if (opts.harness === "codex") {
+		privateCodexHome(path.join(runDir, "codex-home"));
+		privateHome(path.join(runDir, "home"));
+	}
 	const sdkDir = sdkNodeModulesDir();
 	if (sdkDir !== void 0) {
 		const link = path.join(runDir, "node_modules");
@@ -323,4 +390,4 @@ function generateRun(scenarioDir, opts, paths) {
 	};
 }
 //#endregion
-export { DEFAULT_CLAUDE_AGENT, FILE_BLOCK, buildConfig, encodeRunNamePart, generateRun, isHiddenSkill, loadScenario, materialize, privateCodexHome, requiredEvalPackages, resolvePackageDir, runNameFor, sdkNodeModulesDir, stripHiddenFlag, trialLabel };
+export { DEFAULT_CLAUDE_AGENT, FILE_BLOCK, buildConfig, encodeRunNamePart, generateRun, isHiddenSkill, linkedSiblings, loadScenario, materialize, privateCodexHome, privateHome, requiredEvalPackages, resolvePackageDir, runNameFor, sdkNodeModulesDir, stripHiddenFlag, trialLabel };

package/dist/transform.js

@@ -14,6 +14,14 @@ function capped(rel, text) {
 	if (text.length <= PER_FILE_CAP) return `=== OUTPUT FILE: ${rel} ===\n${text}\n=== END OUTPUT FILE ===`;
 	return `=== OUTPUT FILE: ${rel} (truncated: showing ${PER_FILE_CAP} of ${text.length} chars) ===\n${safeSlice(text, PER_FILE_CAP)}\n=== END OUTPUT FILE ===`;
 }
+const SKIPPED_DIRS = /* @__PURE__ */ new Set([
+	".claude",
+	".grok",
+	"node_modules",
+	".cache",
+	".npm",
+	".pnpm-store"
+]);
 function transform(output, context) {
 	const workdir = context.vars.workdir;
 	const manifest = JSON.parse(fs.readFileSync(context.vars.manifest, "utf8"));
@@ -23,7 +31,7 @@ function transform(output, context) {
 		for (const e of fs.readdirSync(dir, { withFileTypes: true })) {
 			const p = path.join(dir, e.name);
 			if (e.isDirectory()) {
-				if (e.name !== ".claude" && e.name !== ".grok" && e.name !== "node_modules") walk(p);
+				if (!SKIPPED_DIRS.has(e.name)) walk(p);
 			} else if (e.isFile()) {
 				const rel = path.relative(workdir, p);
 				visited.add(rel);
@@ -64,4 +72,4 @@ function transform(output, context) {
 	return `${output}\n\n${appended}`;
 }
 //#endregion
-export { transform as default };
+export { SKIPPED_DIRS, transform as default };

package/docs/adoption.md

@@ -10,7 +10,9 @@ pnpm add -D @uinaf/skillcheck
 ```
 
 Runners that install with `--ignore-scripts` are fine: the package ships
-compiled ESM and has no install, prepare, or postinstall script.
+compiled ESM and has no install or postinstall script. Its `prepare` script
+only wires this repository's own commit hook; package managers do not run it
+for a registry install.
 
 ```json
 { "scripts": { "skills:lint": "skillcheck lint" } }
@@ -25,6 +27,7 @@ jobs:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
+      - uses: pnpm/action-setup@ea17c68df8912ef543352723c149a84f56e3d413 # v6.1.0
       - uses: actions/setup-node@249970729cb0ef3589644e2896645e5dc5ba9c38 # v6.5.0
         with:
           node-version: "24"
@@ -33,7 +36,8 @@ jobs:
 ```
 
 The job needs no secrets and no network beyond the install. Node 24 is the
-floor.
+floor. `pnpm/action-setup` installs the pnpm version pinned in the repo's
+`packageManager` field; hosted runners do not ship pnpm.
 
 Run it through the script rather than a bare `npx skillcheck`: the script
 resolves the version the repo pinned, and `npx` would resolve the latest one on

package/docs/releasing.md

@@ -5,15 +5,13 @@
 A push to `main` runs one workflow, `.github/workflows/release.yml`:
 
 ```text
-verify ──┐
-         ├──> release   npm publish, OIDC + uinaf-ci  (release environment)
-scan ────┘
+verify ──> release   npm publish, OIDC + uinaf-ci  (release environment)
 ```
 
-`verify` and `scan` are the shared gate: `verify` is called from `verify.yml`,
-and `scan` calls the shared scan in `uinaf/.github`, the same one `scan.yml`
-runs for pull requests. Keep it that way: a second copy of the gate on a push-to-`main` workflow
-races this one over the same commit.
+`verify` is the one gate, called from `verify.yml`; its last step runs the
+shared push-time scan from `uinaf/.github` over the pushed range. Keep it that
+way: a second copy of the gate on a push-to-`main` workflow races this one over
+the same commit.
 
 The file name `release.yml` is load-bearing. See below.
 

package/docs/scenarios.md

@@ -109,12 +109,23 @@ frontmatter block; body text mentioning the key does not count.
 ## The workdir
 
 Per run, under a scratch directory in the system temp dir
-(`skillcheck-<hash of the root>/<name>/trial-<n>/`), rebuilt from scratch each
+(`skillcheck-<uid>-<hash of the root>/<name>/trial-<n>/`), rebuilt from scratch each
 time. It stays outside the root so the agent cannot reach the skill's source,
 its evals, or the repository's own agent guidance. The skill under test is installed where the harness discovers skills:
 `.claude/skills/<skill>/`, plus `.agents/skills/<skill>/` on codex, or
 `.grok/skills/<skill>/` on Grok, with its `evals/` directory excluded, so
-criteria never leak into the agent's context.
+criteria never leak into the agent's context. Sibling skills under the same root that the skill links to (inline, angle-bracket, or reference links, followed transitively; a symlink out of the root is skipped) are installed beside it
+the same way, so links such as `../other/SKILL.md` resolve as they do when a
+plugin installs the set; a control run installs none of them.
+
+A Codex run gets a private `CODEX_HOME` holding only the operator's
+`config.toml` and `auth.json`, and a private `HOME` that links every entry of
+the operator's home except `.agents`, `.claude`, and `.codex`. Auth helpers that read
+`$HOME` keep working; the operator's user-level skills and global guidance are
+not visible.
+
+Graded output skips agent config roots, `node_modules`, and tool caches
+(`.cache`, `.npm`, `.pnpm-store`), so a cache cannot crowd deliverables out.
 
 Scenario quality is behavioral proof; [authoring](authoring.md) covers the
 judgment layer lint and evals cannot grade.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@uinaf/skillcheck",
-  "version": "1.5.0",
+  "version": "1.5.2",
   "description": "Lint and eval harness for agent skills",
   "homepage": "https://github.com/uinaf/skillcheck#readme",
   "bugs": {
@@ -66,5 +66,5 @@
   "engines": {
     "node": ">=24"
   },
-  "packageManager": "pnpm@12.4.2"
+  "packageManager": "pnpm@12.6.0"
 }