vigiles 6.0.0 → 7.0.0

package/dist/scan-behavioral.js

@@ -17,12 +17,17 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.probePluginTriggersWith = probePluginTriggersWith;
 exports.probePluginTriggers = probePluginTriggers;
 exports.formatBehavioralReport = formatBehavioralReport;
+exports.buildSelectionReport = buildSelectionReport;
+exports.measurePluginSelectionWith = measurePluginSelectionWith;
+exports.measurePluginSelection = measurePluginSelection;
+exports.formatSelectionReport = formatSelectionReport;
 const node_fs_1 = require("node:fs");
 const node_path_1 = require("node:path");
 const scan_js_1 = require("./scan.js");
 const eval_js_1 = require("./eval.js");
 const harness_assert_js_1 = require("./harness-assert.js");
 const harness_test_js_1 = require("./harness-test.js");
+const plugin_loader_js_1 = require("./adapters/claude-code/plugin-loader.js");
 const eval_js_2 = require("./adapters/codex/eval.js");
 const driver_js_1 = require("./adapters/codex/driver.js");
 function buildProbe(dir, harness) {
@@ -56,6 +61,39 @@ function pluginName(dir) {
         return null;
     }
 }
+/**
+ * Does the plugin declare a SessionStart hook? The STUBBED measurement path rebuilds
+ * the plugin to skills-only (`packageSkillsDir`), DROPPING `hooks/` — so a SessionStart
+ * hook that primes skill selection (e.g. superpowers' `using-superpowers` gateway
+ * injection) is silently lost, and a recall collapse to 0 under stubbing is then a
+ * measurement ARTIFACT, not a real miss. Detect it to LABEL honestly (Layer 1) rather
+ * than report a misleading 0%. See `research/plugin-selection-collision.md`.
+ */
+function hasSessionStartHook(dir) {
+    try {
+        const hooks = (0, plugin_loader_js_1.loadPlugin)(dir).settings.hooks;
+        return hooks !== undefined && Object.keys(hooks).includes("SessionStart");
+    }
+    catch {
+        return false;
+    }
+}
+const HOOK_PRIMED_NOTE = "hook-primed — the stubbed run dropped the plugin's SessionStart hook (which can " +
+    "prime skill selection), so 0% recall is likely a measurement artifact; re-run " +
+    "against the full plugin install to measure faithfully";
+/**
+ * Layer-1 honesty: a STUBBED run on a SessionStart-hooked plugin where EVERY measured
+ * skill sits at recall 0 is the dropped-hook artifact — not a real result. The
+ * all-zero gate keeps a genuine single-skill miss reported as real (if siblings fired,
+ * the hook ran or wasn't needed). Applied to both the trigger column and the matrix.
+ */
+function isStubbedHookArtifact(dir, stub, recalls) {
+    if (!stub || recalls.length === 0)
+        return false;
+    if (!recalls.every((r) => r === 0))
+        return false;
+    return hasSessionStartHook(dir);
+}
 /** Probe one skill via the harness probe's eval driver → result, never throwing. */
 async function probeSkill(ctx, name, ps) {
     try {
@@ -108,7 +146,19 @@ async function probePluginTriggersWith(dir, promptSet, probe, opts = {}) {
         }
         results.push(await probeSkill(ctx, s.name, ps));
     }
-    return { available: true, results };
+    return {
+        available: true,
+        results: relabelTriggerArtifact(dir, probe, results),
+    };
+}
+/** Relabel an all-zero-recall stubbed run on a hooked plugin as unmeasured (Layer 1). */
+function relabelTriggerArtifact(dir, probe, results) {
+    const recalls = results.filter((r) => r.measured).map((r) => r.recall ?? 0);
+    if (!isStubbedHookArtifact(dir, probe.stub, recalls))
+        return [...results];
+    return results.map((r) => r.measured && (r.recall ?? 0) === 0
+        ? { skill: r.skill, measured: false, note: HOOK_PRIMED_NOTE }
+        : r);
 }
 /**
  * Probe a plugin's skills against the real harness (default Claude Code; Codex via
@@ -147,4 +197,192 @@ function formatBehavioralReport(b) {
     }
     return lines.join("\n");
 }
+/** Map a namespaced skill id (`ns:name`) to its bare name. */
+function bareSkillName(id) {
+    const i = id.lastIndexOf(":");
+    return i >= 0 ? id.slice(i + 1) : id;
+}
+function emptySelection(skills) {
+    return {
+        available: true,
+        skills,
+        matrix: skills.map(() => []),
+        perSkill: [],
+        collisionRate: 0,
+        n: 0,
+    };
+}
+function buildSkillStat(a) {
+    const { skill, i, skills, row, n, collisions } = a;
+    const collidesWith = skills
+        .map((s, j) => ({ skill: s, rate: n > 0 ? row[j] / n : 0 }))
+        .filter((c) => c.skill !== skill && c.rate > 0)
+        .sort((x, y) => y.rate - x.rate);
+    return {
+        skill,
+        recall: n > 0 ? row[i] / n : 0,
+        collisionRate: n > 0 ? collisions / n : 0,
+        n,
+        collidesWith,
+    };
+}
+/**
+ * Pure aggregation: fold per-run fired-skill sets into the N×N selection matrix +
+ * per-skill recall/collision + the plugin-level collision rate. Separated from the
+ * model-driving so it's unit-testable with synthetic runs (no model).
+ */
+function buildSelectionReport(skills, runs) {
+    const index = new Map(skills.map((s, i) => [s, i]));
+    const n = skills.length;
+    const matrix = skills.map(() => new Array(n).fill(0));
+    const nBy = new Array(n).fill(0);
+    const collisionBy = new Array(n).fill(0);
+    let collisionTotal = 0;
+    for (const run of runs) {
+        const i = index.get(run.intended);
+        if (i === undefined)
+            continue;
+        nBy[i] += 1;
+        let collided = false;
+        for (const fb of run.firedBare) {
+            const j = index.get(fb);
+            if (j === undefined)
+                continue;
+            matrix[i][j] += 1;
+            if (j !== i)
+                collided = true;
+        }
+        if (collided) {
+            collisionBy[i] += 1;
+            collisionTotal += 1;
+        }
+    }
+    const total = nBy.reduce((a, b) => a + b, 0);
+    return {
+        available: true,
+        skills,
+        matrix,
+        perSkill: skills.map((s, i) => buildSkillStat({
+            skill: s,
+            i,
+            skills,
+            row: matrix[i],
+            n: nBy[i],
+            collisions: collisionBy[i],
+        })),
+        collisionRate: total > 0 ? collisionTotal / total : 0,
+        n: total,
+    };
+}
+/** Build the prompts × trials work list across every skill that has prompts. */
+function selectionJobs(candidates, promptSet, trials) {
+    return candidates.flatMap((c) => {
+        const ps = promptSet[c.name];
+        if (!ps || ps.prompts.length === 0)
+            return [];
+        return ps.prompts.flatMap((prompt) => Array.from({ length: trials }, () => ({ intended: c.name, prompt })));
+    });
+}
+/** The injectable core (for tests): drive the matrix via a fake/real probe. */
+async function measurePluginSelectionWith(dir, promptSet, probe, opts = {}) {
+    const candidates = (0, scan_js_1.scanPlugin)(dir).skills.filter((s) => !s.userInvoked && s.hasDescription);
+    const skills = candidates.map((c) => c.name);
+    if (skills.length < 2) {
+        return {
+            ...emptySelection(skills),
+            note: "needs ≥2 model-invocable skills to measure cross-skill collision",
+        };
+    }
+    const own = new Set(skills);
+    const jobs = selectionJobs(candidates, promptSet, Math.max(1, opts.trials ?? 1));
+    if (jobs.length === 0) {
+        return {
+            ...emptySelection(skills),
+            note: "no prompts supplied for any skill",
+        };
+    }
+    // Selection is decided at the frontmatter (the selector picks BEFORE the body
+    // loads), so stub each body to a no-op: the run stops AT selection instead of
+    // executing the whole workflow — the same affordability trick trigger-rate uses.
+    const pluginDir = probe.stub ? (0, eval_js_1.stubbedPluginDir)(dir) : dir;
+    try {
+        const d = probe.evalDriver;
+        const outcomes = await (0, eval_js_1.runPool)(jobs, Math.max(1, opts.concurrency ?? 1), (job) => (0, eval_js_1.runSkillSelectionTrial)({
+            prompt: job.prompt,
+            pluginDir,
+            runner: d.runner,
+            parse: d.parse,
+            runError: d.runError,
+            model: opts.model ?? "sonnet",
+        }));
+        const runs = [];
+        jobs.forEach((job, k) => {
+            if (outcomes[k].errored)
+                return;
+            runs.push({
+                intended: job.intended,
+                firedBare: outcomes[k].fired
+                    .map(bareSkillName)
+                    .filter((b) => own.has(b)),
+            });
+        });
+        const report = buildSelectionReport(skills, runs);
+        // Layer-1 honesty: an all-zero-recall stubbed run on a hooked plugin is the
+        // dropped-hook artifact — flag it instead of presenting a 0%-collision result
+        // computed from a plugin that never fired.
+        const recalls = report.perSkill.filter((s) => s.n > 0).map((s) => s.recall);
+        return isStubbedHookArtifact(dir, probe.stub, recalls)
+            ? { ...report, note: HOOK_PRIMED_NOTE }
+            : report;
+    }
+    finally {
+        if (probe.stub)
+            (0, node_fs_1.rmSync)(pluginDir, { recursive: true, force: true });
+    }
+}
+/**
+ * Measure a plugin's cross-skill selection-collision matrix against the real
+ * harness (Claude Code only — Codex has no skill-selection event). Needs the
+ * `claude` CLI + model auth; degrades to `available: false` otherwise.
+ */
+async function measurePluginSelection(dir, promptSet, opts = {}) {
+    const harness = opts.harness ?? "claude-code";
+    if (harness !== "claude-code") {
+        return {
+            ...emptySelection([]),
+            available: false,
+            note: `selection-collision is Claude Code only (no skill-selection event on ${harness})`,
+        };
+    }
+    const probe = buildProbe(dir, harness);
+    if (!probe.available()) {
+        return {
+            ...emptySelection([]),
+            available: false,
+            note: "needs the claude CLI + model auth",
+        };
+    }
+    return measurePluginSelectionWith(dir, promptSet, probe, opts);
+}
+/** Format the selection-collision matrix as a scan-report section. */
+function formatSelectionReport(r) {
+    if (!r.available)
+        return `Selection-collision: unavailable — ${r.note ?? "n/a"}`;
+    if (r.n === 0)
+        return `Selection-collision: ${r.note ?? "not measured"}`;
+    const lines = [
+        `Selection-collision: ${pct(r.collisionRate)} of ${String(r.n)} runs hit a sibling skill`,
+    ];
+    if (r.note)
+        lines.push(`  ⓘ ${r.note}`);
+    for (const s of r.perSkill) {
+        if (s.n === 0)
+            continue;
+        const mark = s.collisionRate === 0 ? "✓" : "⚠";
+        const top = s.collidesWith[0];
+        const tail = top ? ` — top collider: ${top.skill} ${pct(top.rate)}` : "";
+        lines.push(`  ${mark} ${s.skill} — recall ${pct(s.recall)}, collision ${pct(s.collisionRate)}${tail}`);
+    }
+    return lines.join("\n");
+}
 //# sourceMappingURL=scan-behavioral.js.map

package/dist/scan.d.ts

@@ -18,6 +18,7 @@ import { type HookEventIssue } from "./core/hook-events.js";
 import { type McpIssue } from "./core/mcp-config.js";
 import { type DescriptionOverlap } from "./core/description-overlap.js";
 import { type McpToolIssue } from "./core/mcp-tool.js";
+import { type McpContractToolError } from "./core/mcp.js";
 import { type McpHookIssue } from "./core/mcp-hook.js";
 import { type PurityLevel, type EffectSurface } from "./core/effects.js";
 /** A named writing system. The label `unexpectedScript` reports + the config's expectation parse into this. */
@@ -176,6 +177,19 @@ export interface SurfaceClassifier {
 export declare function unexpectedScript(text: string, expected?: Script): Script | null;
 /** Scan a plugin/repo directory and report its surfaces + structural issues. */
 export declare function scanPlugin(dir: string, layout?: PluginLayout, dialect?: HarnessDialect): ScanReport;
+/**
+ * LIVE MCP tool resolution for a scanned plugin — the opt-in (`scan --verify-mcp`)
+ * dynamic check no static linter can do: it STARTS each declared MCP server and
+ * checks every `mcp__server__tool` the plugin's agents reference actually exists on
+ * it (catching rename/removal rot, e.g. `create_issue`→`issue_write`). Reuses the
+ * already-computed `report` (its agents' tool lists) + the declared server configs;
+ * returns `[]` when the plugin declares no MCP servers (nothing to start). Async +
+ * side-effecting (spawns servers) — which is exactly why it's opt-in, not a default
+ * lint rule. See `verifyMcpContractTools` (core/mcp.ts).
+ */
+export declare function verifyLiveMcpTools(report: ScanReport, layout: PluginLayout, dialect: HarnessDialect, timeoutMs?: number): Promise<McpContractToolError[]>;
+/** Render the live MCP tool-check result (human-readable). */
+export declare function formatMcpContractReport(errors: readonly McpContractToolError[]): string;
 /**
  * A plugin MARKETPLACE (`.claude-plugin/marketplace.json`) decomposed into its
  * members. A marketplace either VENDORS its plugins in-tree (string `source`

package/dist/scan.js

@@ -15,6 +15,8 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.unexpectedScript = unexpectedScript;
 exports.scanPlugin = scanPlugin;
+exports.verifyLiveMcpTools = verifyLiveMcpTools;
+exports.formatMcpContractReport = formatMcpContractReport;
 exports.inspectMarketplace = inspectMarketplace;
 exports.expandMarketplace = expandMarketplace;
 exports.formatScanReport = formatScanReport;
@@ -31,6 +33,7 @@ const linters_js_1 = require("./core/linters.js");
 const frontmatter_read_js_1 = require("./core/frontmatter-read.js");
 const description_overlap_js_1 = require("./core/description-overlap.js");
 const mcp_tool_js_1 = require("./core/mcp-tool.js");
+const mcp_js_1 = require("./core/mcp.js");
 const mcp_hook_js_1 = require("./core/mcp-hook.js");
 const agent_runtime_js_1 = require("./adapters/claude-code/agent-runtime.js");
 const test_coverage_js_1 = require("./test-coverage.js");
@@ -535,6 +538,35 @@ function scanPlugin(dir, layout, dialect = dialect_js_1.claudeCodeDialect) {
         puritySummary,
     };
 }
+/**
+ * LIVE MCP tool resolution for a scanned plugin — the opt-in (`scan --verify-mcp`)
+ * dynamic check no static linter can do: it STARTS each declared MCP server and
+ * checks every `mcp__server__tool` the plugin's agents reference actually exists on
+ * it (catching rename/removal rot, e.g. `create_issue`→`issue_write`). Reuses the
+ * already-computed `report` (its agents' tool lists) + the declared server configs;
+ * returns `[]` when the plugin declares no MCP servers (nothing to start). Async +
+ * side-effecting (spawns servers) — which is exactly why it's opt-in, not a default
+ * lint rule. See `verifyMcpContractTools` (core/mcp.ts).
+ */
+async function verifyLiveMcpTools(report, layout, dialect, timeoutMs = 10000) {
+    // collectMcpServers yields the raw JSON server entries; a malformed one (no
+    // command) just fails to start → server-unreachable (handled), so the cast is safe.
+    const servers = collectMcpServers((0, node_path_1.resolve)(report.dir), layout);
+    if (Object.keys(servers).length === 0)
+        return [];
+    const tools = report.agents.flatMap((a) => a.tools ?? []);
+    return (0, mcp_js_1.verifyMcpContractTools)(tools, servers, dialect, timeoutMs);
+}
+/** Render the live MCP tool-check result (human-readable). */
+function formatMcpContractReport(errors) {
+    if (errors.length === 0) {
+        return "Live MCP tool check: every referenced mcp__server__tool resolves ✓";
+    }
+    const lines = [`Live MCP tool check — ${String(errors.length)} issue(s):`];
+    for (const e of errors)
+        lines.push("  ✗ " + (0, mcp_js_1.mcpContractToolMessage)(e));
+    return lines.join("\n");
+}
 /**
  * Read a `marketplace.json` beside the layout's plugin manifest and classify its
  * members into on-disk vs external. Returns `null` when `dir` is not a
@@ -710,7 +742,7 @@ function formatScanReport(r) {
     // verdict — it points at the behavioral column, it doesn't fail the scan.
     const mismatched = r.skills.filter((s) => s.descriptionScript);
     if (mismatched.length > 0) {
-        out.push(`⚠ ${String(mismatched.length)} skill(s) have descriptions in an unexpected script (cross-language trigger risk) — measure with \`scan --trigger\``, "");
+        out.push(`⚠ ${String(mismatched.length)} skill(s) have descriptions in an unexpected script (cross-language trigger risk) — measure with \`vigiles measure\``, "");
     }
     // Skill-metadata is a RECOMMENDATION, not a structural defect (the skill loads
     // via fallbacks) — reported as a soft note, never counted in the verdict.

package/dist/score-explainer.js

@@ -155,7 +155,7 @@ function explainSurface(report, surface) {
 /** Render explanations for a CLI/report — grouped under the symptom, fix called out. */
 function formatExplanations(exps) {
     if (exps.length === 0) {
-        return "No deterministic cause found — the cause is likely behavioral (measure with `scan --trigger` / an eval).";
+        return "No deterministic cause found — the cause is likely behavioral (measure with `vigiles measure` / an eval).";
     }
     const lines = [];
     for (const e of exps) {

package/dist/self-command-refs.d.ts

@@ -0,0 +1,21 @@
+export interface CommandRefIssue {
+    readonly file: string;
+    readonly line: number;
+    /** The offending invocation, e.g. `vigiles compile-hook`. */
+    readonly ref: string;
+    readonly reason: string;
+}
+export interface KnownCommands {
+    readonly verbs: readonly string[];
+    readonly kinds: readonly string[];
+}
+/**
+ * Find stale/unknown vigiles command references across the given files. Pure —
+ * the caller supplies file contents (so it works over the repo in a test, or any
+ * file set).
+ */
+export declare function findStaleCommandRefs(files: readonly {
+    readonly path: string;
+    readonly content: string;
+}[], known?: KnownCommands): CommandRefIssue[];
+//# sourceMappingURL=self-command-refs.d.ts.map

package/dist/self-command-refs.js

@@ -0,0 +1,125 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.findStaleCommandRefs = findStaleCommandRefs;
+/**
+ * Self-command-reference verification — the cross-reference moat applied to
+ * vigiles's OWN docs. vigiles verifies that a CLAUDE.md's commands resolve; its
+ * own docs cite ITS OWN commands, and nothing checked those until a rename left
+ * `vigiles compile-hook` / `vigiles run-skill` rotting in the docs. This is the
+ * deterministic gate: every `vigiles <cmd>` reference in a doc/comment must
+ * resolve to a real command (a VERB or a `hook-runtime <kind>`).
+ *
+ * HIGH-PRECISION by construction (don't-cry-wolf). A reference is inspected only
+ * when it's an unambiguous COMMAND — i.e. it sits inside an inline `` `span` ``,
+ * inside a ```shell fence```, or is prefixed `npx vigiles` / `Usage: vigiles` /
+ * is a `cli.js` or `${CLI}` invocation (the harness-test convention `node ${CLI}
+ * <cmd>`). Prose ("vigiles compiles the spec", "adds vigiles to devDependencies")
+ * and non-shell fences (```text agent prompts, ```ts) are never matched. A bare unknown VERB is flagged only when hyphenated (every
+ * renamed vigiles command is — `compile-hook`, `run-skill`) or in an explicit
+ * invocation; `hook-runtime <kind>` is checked in any command context. Source of
+ * truth: {@link VERBS} / {@link HOOK_RUNTIME_KINDS}.
+ *
+ * KNOWN LIMITATION (measured, deliberately not "fixed"): this catches the
+ * `vigiles <cmd>` / `cli.js <cmd>` INVOCATION form. It does NOT flag a BARE
+ * command name (`agent-hook` without the `vigiles ` prefix) nor a bare file path
+ * (`src/foo.ts`), because both collide with legitimate non-references: bare
+ * command names double as CONCEPT names (`the refs-hook nudge` appears as a
+ * feature name in ~10 files; `agent-hook` as a rail name + in test labels +
+ * `makeTmpDir("agent-hook-cli")`), and bare paths collide with ILLUSTRATIVE
+ * examples (the README's `src/auth/login.ts`, `docs/foo.md`) and test fixtures
+ * (`session.test.ts` builds fake `src/compile.ts` diffs). A denylist over either
+ * would cry wolf on dozens of valid usages — so terminology/path-accuracy stays
+ * the JUDGMENT half the `doc-consistency` rule assigns to discipline, not a check.
+ */
+const cli_commands_js_1 = require("./cli-commands.js");
+const DEFAULT_KNOWN = {
+    verbs: cli_commands_js_1.VERBS,
+    kinds: cli_commands_js_1.HOOK_RUNTIME_KINDS,
+};
+const TOKEN = "[a-z][a-z0-9-]*";
+// `[npx ][vigiles|cli.js|${CLI}] <verb> [<kind>]` — capture the optional prefix +
+// literal. `${CLI}` is the harness-test convention (`const CLI = new
+// URL("…/dist/cli.js")` → `node ${CLI} <cmd>`); without it a stale runtime ref in
+// a `.mjs`/`.test.ts` invocation slips through (it did: refs-nudge.harness.mjs kept
+// calling `refs-hook` after the hook-runtime rename). `\b` can't anchor `${CLI}`
+// (it starts with `$`, a non-word char), so each word literal carries its own `\b`.
+const INVOKE = new RegExp(String.raw `(npx |Usage: )?(\bvigiles|\bcli\.js|\$\{CLI\})\s+(${TOKEN})(?:\s+(${TOKEN}))?`, "g");
+// A bare `hook-runtime <kind>` (e.g. `node ${CLI} hook-runtime agent` in a test).
+const KIND = new RegExp(String.raw `hook-runtime\s+(${TOKEN})`, "g");
+const SHELL_FENCE = /^\s*```(?:bash|sh|shell|zsh|console|shell-session)\s*$/;
+const FENCE = /^\s*```/;
+/** Inclusive index ranges of inline `code spans` in a line. */
+function codeSpans(line) {
+    const ranges = [];
+    for (const m of line.matchAll(/`[^`\n]+`/g))
+        ranges.push([m.index, m.index + m[0].length]);
+    return ranges;
+}
+const inSpan = (i, spans) => spans.some(([a, b]) => i >= a && i < b);
+function scanLine(line, inShellFence, known) {
+    const spans = inShellFence ? null : codeSpans(line);
+    const ctx = (i) => inShellFence || inSpan(i, spans ?? []);
+    const out = [];
+    const seen = new Set();
+    const push = (ref, reason) => {
+        const key = ref + "\0" + reason;
+        if (!seen.has(key)) {
+            seen.add(key);
+            out.push({ ref, reason });
+        }
+    };
+    for (const m of line.matchAll(INVOKE)) {
+        const explicit = inShellFence || Boolean(m[1]) || m[2] === "cli.js" || m[2] === "${CLI}";
+        if (!explicit && !ctx(m.index))
+            continue;
+        const verb = m[3];
+        if (verb === "hook-runtime") {
+            if (m[4] !== undefined && !known.kinds.includes(m[4]))
+                push(`hook-runtime ${m[4]}`, `unknown hook-runtime kind "${m[4]}"`);
+        }
+        else if (!known.verbs.includes(verb) &&
+            (verb.includes("-") || explicit)) {
+            push(`${m[2]} ${verb}`, `unknown/removed command "${verb}"`);
+        }
+    }
+    // Bare `hook-runtime <kind>` (no vigiles/cli.js literal) inside a command context.
+    for (const m of line.matchAll(KIND)) {
+        if (!ctx(m.index))
+            continue;
+        if (!known.kinds.includes(m[1]))
+            push(`hook-runtime ${m[1]}`, `unknown hook-runtime kind "${m[1]}"`);
+    }
+    return out;
+}
+/**
+ * Find stale/unknown vigiles command references across the given files. Pure —
+ * the caller supplies file contents (so it works over the repo in a test, or any
+ * file set).
+ */
+function findStaleCommandRefs(files, known = DEFAULT_KNOWN) {
+    const issues = [];
+    for (const { path, content } of files) {
+        let inAnyFence = false;
+        let inShellFence = false;
+        const lines = content.split("\n");
+        for (let i = 0; i < lines.length; i++) {
+            const line = lines[i];
+            if (FENCE.test(line)) {
+                if (inAnyFence) {
+                    inAnyFence = inShellFence = false;
+                }
+                else {
+                    inAnyFence = true;
+                    inShellFence = SHELL_FENCE.test(line);
+                }
+                continue;
+            }
+            if (line.includes("vigiles:ignore-cmd"))
+                continue;
+            for (const f of scanLine(line, inShellFence, known))
+                issues.push({ file: path, line: i + 1, ref: f.ref, reason: f.reason });
+        }
+    }
+    return issues;
+}
+//# sourceMappingURL=self-command-refs.js.map

package/dist/testing.d.ts

@@ -10,13 +10,15 @@
  * boundary forbids importing `src/adapters/*` from here. See
  * `research/adapter-api-design.md`.
  */
-export * from "./run-hook.js";
-export * from "./eval.js";
+export { runHook, propertyHook } from "./run-hook.js";
+export type { HookRunResult, RunHookOptions, HookInput, HookOutput, HookPropertyResult, } from "./run-hook.js";
+export { runEval, measure, measureArms, measureTriggerRate, assertRates, assertPromptDiversity, checkPromptDiversity, checkReportToJUnit, formatCheckReport, formatEvalReport, formatTriggerRateReport, claudeEvalDriver, parseClaudeRun, stubSkillBody, } from "./eval.js";
+export type { EvalArm, EvalDriver, EvalSpec, EvalReport, EvalUsage, MeasureSpec, ArmsMeasureSpec, ArmReport, ArmUsage, ArmsCheckReport, CheckRate, CheckReport, MetricStat, Metrics, ModelOutputParser, ParsedModelRun, PromptDiversityIssue, PromptTriggerStat, RunContext, RunOut, SelectionTrialResult, TriggerRateReport, TriggerRateSpec, AgentRunArgs, AgentRunner, } from "./eval.js";
 export * from "./harness-assert.js";
 export * from "./check.js";
 export { hookFired } from "./check.js";
 export { judge } from "./judge.js";
 export * from "./tool-stub.js";
-export { runHarnessTest, runHarness, parseToolCalls, parseSubagents, parseResultEvent, parseOutput, parseHooks, decideSandbox, specTrusted, sandboxAvailable, } from "./harness-test.js";
+export { runHarnessTest, runHarness } from "./harness-test.js";
 export type { HarnessTestSpec, Trace, SubagentTrace, HarnessTestResult, RunHarnessTestOptions, ModelTurn, ModelRequest, ToolCall, HookFire, HarnessTestDriver, SandboxMode, } from "./harness-test.js";
 //# sourceMappingURL=testing.d.ts.map

package/dist/testing.js

@@ -1,4 +1,20 @@
 "use strict";
+/**
+ * `vigiles/testing` — Pillar 2 entry point: the **harness-testing** API. Re-exports
+ * the three tiers — `runHook` (unit), `runHarnessTest` (deterministic), `runEval`
+ * (eval) — plus the runner-agnostic predicates/assertions. Kept deliberately
+ * separate from `vigiles/claude-code` so this surface can stay harness-agnostic as
+ * more harnesses are added. Granular paths (`vigiles/run-hook`, etc.) still work.
+ *
+ * It re-exports the composition-root runner modules (which do the Claude-Code
+ * default-wiring), never an adapter directly — the `agnostic-surface` eslint
+ * boundary forbids importing `src/adapters/*` from here. See
+ * `research/adapter-api-design.md`.
+ */
+// Curated public re-exports (named, not `export *`) so the internal seams —
+// the injectable `*With` runners, low-level `parse*`, pool/aggregate/model-tier
+// helpers — stay out of the public surface, the api reports, and the docs site.
+// (vigiles's own tests import those from the source modules directly.)
 var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
     if (k2 === undefined) k2 = k;
     var desc = Object.getOwnPropertyDescriptor(m, k);
@@ -14,21 +30,27 @@ var __exportStar = (this && this.__exportStar) || function(m, exports) {
     for (var p in m) if (p !== "default" && !Object.prototype.hasOwnProperty.call(exports, p)) __createBinding(exports, m, p);
 };
 Object.defineProperty(exports, "__esModule", { value: true });
-exports.sandboxAvailable = exports.specTrusted = exports.decideSandbox = exports.parseHooks = exports.parseOutput = exports.parseResultEvent = exports.parseSubagents = exports.parseToolCalls = exports.runHarness = exports.runHarnessTest = exports.judge = exports.hookFired = void 0;
-/**
- * `vigiles/testing` — Pillar 2 entry point: the **harness-testing** API. Re-exports
- * the three tiers — `runHook` (unit), `runHarnessTest` (deterministic), `runEval`
- * (eval) — plus the runner-agnostic predicates/assertions. Kept deliberately
- * separate from `vigiles/claude-code` so this surface can stay harness-agnostic as
- * more harnesses are added. Granular paths (`vigiles/run-hook`, etc.) still work.
- *
- * It re-exports the composition-root runner modules (which do the Claude-Code
- * default-wiring), never an adapter directly — the `agnostic-surface` eslint
- * boundary forbids importing `src/adapters/*` from here. See
- * `research/adapter-api-design.md`.
- */
-__exportStar(require("./run-hook.js"), exports);
-__exportStar(require("./eval.js"), exports);
+exports.runHarness = exports.runHarnessTest = exports.judge = exports.hookFired = exports.stubSkillBody = exports.parseClaudeRun = exports.claudeEvalDriver = exports.formatTriggerRateReport = exports.formatEvalReport = exports.formatCheckReport = exports.checkReportToJUnit = exports.checkPromptDiversity = exports.assertPromptDiversity = exports.assertRates = exports.measureTriggerRate = exports.measureArms = exports.measure = exports.runEval = exports.propertyHook = exports.runHook = void 0;
+// --- unit tier: runHook ---
+var run_hook_js_1 = require("./run-hook.js");
+Object.defineProperty(exports, "runHook", { enumerable: true, get: function () { return run_hook_js_1.runHook; } });
+Object.defineProperty(exports, "propertyHook", { enumerable: true, get: function () { return run_hook_js_1.propertyHook; } });
+// --- eval tier: runEval / measure / trigger-rate ---
+var eval_js_1 = require("./eval.js");
+Object.defineProperty(exports, "runEval", { enumerable: true, get: function () { return eval_js_1.runEval; } });
+Object.defineProperty(exports, "measure", { enumerable: true, get: function () { return eval_js_1.measure; } });
+Object.defineProperty(exports, "measureArms", { enumerable: true, get: function () { return eval_js_1.measureArms; } });
+Object.defineProperty(exports, "measureTriggerRate", { enumerable: true, get: function () { return eval_js_1.measureTriggerRate; } });
+Object.defineProperty(exports, "assertRates", { enumerable: true, get: function () { return eval_js_1.assertRates; } });
+Object.defineProperty(exports, "assertPromptDiversity", { enumerable: true, get: function () { return eval_js_1.assertPromptDiversity; } });
+Object.defineProperty(exports, "checkPromptDiversity", { enumerable: true, get: function () { return eval_js_1.checkPromptDiversity; } });
+Object.defineProperty(exports, "checkReportToJUnit", { enumerable: true, get: function () { return eval_js_1.checkReportToJUnit; } });
+Object.defineProperty(exports, "formatCheckReport", { enumerable: true, get: function () { return eval_js_1.formatCheckReport; } });
+Object.defineProperty(exports, "formatEvalReport", { enumerable: true, get: function () { return eval_js_1.formatEvalReport; } });
+Object.defineProperty(exports, "formatTriggerRateReport", { enumerable: true, get: function () { return eval_js_1.formatTriggerRateReport; } });
+Object.defineProperty(exports, "claudeEvalDriver", { enumerable: true, get: function () { return eval_js_1.claudeEvalDriver; } });
+Object.defineProperty(exports, "parseClaudeRun", { enumerable: true, get: function () { return eval_js_1.parseClaudeRun; } });
+Object.defineProperty(exports, "stubSkillBody", { enumerable: true, get: function () { return eval_js_1.stubSkillBody; } });
 __exportStar(require("./harness-assert.js"), exports);
 // The declarative check vocabulary is now first-class at the front door. Its
 // `hookFired` (a `Check<Trace>`) supersedes the legacy boolean predicate of the
@@ -51,12 +73,4 @@ __exportStar(require("./tool-stub.js"), exports);
 var harness_test_js_1 = require("./harness-test.js");
 Object.defineProperty(exports, "runHarnessTest", { enumerable: true, get: function () { return harness_test_js_1.runHarnessTest; } });
 Object.defineProperty(exports, "runHarness", { enumerable: true, get: function () { return harness_test_js_1.runHarness; } });
-Object.defineProperty(exports, "parseToolCalls", { enumerable: true, get: function () { return harness_test_js_1.parseToolCalls; } });
-Object.defineProperty(exports, "parseSubagents", { enumerable: true, get: function () { return harness_test_js_1.parseSubagents; } });
-Object.defineProperty(exports, "parseResultEvent", { enumerable: true, get: function () { return harness_test_js_1.parseResultEvent; } });
-Object.defineProperty(exports, "parseOutput", { enumerable: true, get: function () { return harness_test_js_1.parseOutput; } });
-Object.defineProperty(exports, "parseHooks", { enumerable: true, get: function () { return harness_test_js_1.parseHooks; } });
-Object.defineProperty(exports, "decideSandbox", { enumerable: true, get: function () { return harness_test_js_1.decideSandbox; } });
-Object.defineProperty(exports, "specTrusted", { enumerable: true, get: function () { return harness_test_js_1.specTrusted; } });
-Object.defineProperty(exports, "sandboxAvailable", { enumerable: true, get: function () { return harness_test_js_1.sandboxAvailable; } });
 //# sourceMappingURL=testing.js.map

package/dist/tool-intercept.d.ts

@@ -15,7 +15,7 @@
  *
  * - `decideIntercept` — does this call get intercepted, and with what deny reason;
  * - `buildInterceptSettings` — the PreToolUse hook fragment routing matched tools
- *   through `vigiles intercept-tool-hook`;
+ *   through `vigiles hook-runtime intercept-tool`;
  * - `serializeIntercepts` / `parseIntercepts` — the env round-trip (incl. RegExp
  *   matchers) the hook subprocess reads back.
  *
@@ -32,7 +32,7 @@
  * tools, so deny+reason is the closest available — with that ceiling.
  */
 import { type ArgMatcher } from "./arg-match.js";
-/** Env var the spawned `vigiles intercept-tool-hook` reads its intercept list from. */
+/** Env var the spawned `vigiles hook-runtime intercept-tool` reads its intercept list from. */
 export declare const INTERCEPT_TOOLS_ENV = "VIGILES_INTERCEPT_TOOLS";
 /** Declare a tool to intercept: deny its real execution with a block message. */
 export interface ToolIntercept {
@@ -63,7 +63,7 @@ export type InterceptDecision = {
 /**
  * Decide whether a tool call should be intercepted. Returns the first matching
  * intercept's denial reason (preventing real execution), or `{ intercept: false }`
- * to let the call run for real. Pure — the same logic `vigiles intercept-tool-hook`
+ * to let the call run for real. Pure — the same logic `vigiles hook-runtime intercept-tool`
  * runs.
  */
 export declare function decideIntercept(toolName: string, input: unknown, intercepts: readonly ToolIntercept[]): InterceptDecision;
@@ -81,7 +81,7 @@ export interface InterceptSettings {
 }
 /**
  * Build the PreToolUse hook fragment that routes every intercepted tool through
- * `vigiles intercept-tool-hook`. The `matcher` is a CC tool-name regex over the
+ * `vigiles hook-runtime intercept-tool`. The `matcher` is a CC tool-name regex over the
  * union of intercepted tool names (each escaped), so unrelated tools are never
  * intercepted. Merge the result into an arm's `settings`; the intercept list
  * itself travels in the {@link INTERCEPT_TOOLS_ENV} env var (see