vigiles 5.2.0 → 7.0.0

package/dist/scaffold-test.js

@@ -45,21 +45,35 @@ function header(title, run) {
 function hookScaffold(input) {
     const cmd = input.hookCommand ?? `bash ${input.path}`;
     return `${header(`Starter unit test for the \`${input.name}\` hook.`, `npx vigiles test ${suggestedPath(input)}`)}
-import { runHook, assertHookAllowed } from "vigiles/unit";
+import {
+  runHook,
+  assertHookAllowed,
+  verifyGuardrail,
+  formatGuardrailReport,
+  // assertBlocksDisasters, // uncomment to gate CI on the battery (see below)
+} from "vigiles/unit";
+
+const cmd = ${JSON.stringify(cmd)};
 
+// 1) A benign event should pass through.
 // TODO: set the event + input your hook actually inspects (PreToolUse/Bash shown).
-const event = {
+const benign = {
   hook_event_name: "PreToolUse",
   tool_name: "Bash",
   tool_input: { command: "echo hello" },
 };
+assertHookAllowed(runHook(cmd, benign));
+console.log("✓ ${input.name}: allowed the benign event");
 
-const r = runHook(${JSON.stringify(cmd)}, event);
+// 2) SAFETY: if this is a guard, PROVE it blocks the dangerous battery (the #1 hook
+//    pain is a guard that silently doesn't — exit 1 instead of 2, wrong jq path, …).
+//    This prints a coverage map; it does NOT fail by default (your hook may not be
+//    meant to block all of these).
+console.log(formatGuardrailReport(cmd, verifyGuardrail(cmd)));
 
-// TODO: assert the decision you expect. Use assertHookBlocked(r) for the deny case
-// (and a second runHook with the input that SHOULD be blocked).
-assertHookAllowed(r);
-console.log("✓ ${input.name}: hook allowed the benign event");
+// 3) To GATE CI: declare what this guard MUST block, then assert it. Uncomment +
+//    pick the categories your hook is responsible for:
+// assertBlocksDisasters(cmd, { categories: ["destructive-git"] });
 `;
 }
 /** A skill → the eval tier (`measureTriggerRate`): does its description FIRE? */
@@ -100,18 +114,68 @@ console.log(formatTriggerRateReport(report));
 assertTriggerRate(report, { min: 0.8, maxFalsePositive: 0.3 });
 `;
 }
-/** A subagent → the deterministic harness tier; point at the railway/Result path. */
-function agentScaffold(input) {
+/** A JSON value placeholder for an `OutputFieldType`, for the `vigiles:ok` block. */
+function placeholderFor(type) {
+    switch (type) {
+        case "number":
+            return 1;
+        case "boolean":
+            return true;
+        case "string[]":
+            return ["example"];
+        default:
+            return "example";
+    }
+}
+/** Render a `result(ok, err)` builder call reconstructed from the parsed contract. */
+function renderContractBuilder(contract) {
+    const shape = (fields) => `{ ${fields.map((f) => `${f.name}: ${JSON.stringify(f.type)}`).join(", ")} }`;
+    return `result(\n  ${shape(contract.ok)},\n  ${shape(contract.err)},\n)`;
+}
+/**
+ * The OUTCOME test, GENERATED FROM the subagent's `result()` contract: reconstruct
+ * the contract, build a matching `vigiles:ok` block, and `assertAgentOk` it —
+ * deterministic, no LLM judge. This is the typed-spec payoff a markdown
+ * `description:` cannot give you: a parseable, typed outcome a test reads directly.
+ */
+function outcomeSection(input, contract) {
+    const okValue = Object.fromEntries(contract.ok.map((f) => [f.name, placeholderFor(f.type)]));
+    const firstField = contract.ok[0]?.name;
+    const fieldAssertion = firstField
+        ? `// TODO: assert the VALUES you expect (the shape is already validated above), e.g.:\n//   assert.ok(value.${firstField}, "expected a ${firstField}");`
+        : "";
+    return `import assert from "node:assert/strict";
+import { result } from "vigiles/spec";
+import { assertAgentOk } from "vigiles/testing";
+
+// Reconstructed from ${input.name}'s ## Output contract (its compiled .md) — the
+// typed result() the spec wrote. assertAgentOk parses + validates the outcome
+// with NO model judge; swap \`okOutput\` for a real \`runHarness\` turn (Part B in
+// examples/harness/railway-result.harness.mjs) to assert REAL behaviour.
+const contract = ${renderContractBuilder(contract)};
+
+const okOutput = [
+  "${input.name} finished its task.",
+  "\`\`\`vigiles:ok",
+  ${JSON.stringify(JSON.stringify(okValue))},
+  "\`\`\`",
+].join("\\n");
+
+const value = assertAgentOk(okOutput, contract); // deterministic — no LLM judge
+${fieldAssertion}
+console.log("✓ ${input.name}: result() outcome parses + validates against its typed contract");
+`;
+}
+/** The fallback when the subagent has no `result()` contract — assert a tool use. */
+function fallbackSection(input) {
     const toolHint = input.tools && input.tools.length > 0
         ? `assertToolUsed(r, ${JSON.stringify(input.tools[0])}); // its declared contract: ${input.tools.join(", ")}`
         : `assertToolUsed(r, "Task"); // TODO: assert what the subagent should do`;
-    return `${header(`Starter harness test for the \`${input.name}\` subagent.`, `npx vigiles test ${suggestedPath(input)}`)}
-import { runHarnessTest, assertToolUsed } from "vigiles/testing";
-
-// A subagent's OUTCOME is best asserted via a result() contract — deterministic,
-// no LLM judge. If ${input.name} has one, use assertAgentOk(r.output, contract)
-// instead; see the railway-result example in the vigiles docs.
+    return `import { runHarnessTest, assertToolUsed } from "vigiles/testing";
 
+// ${input.name} has no result() contract, so its outcome can't be asserted
+// deterministically — add one (result() on its agent() spec) for a no-judge
+// outcome test. For now, assert it reaches for the right tool.
 const r = await runHarnessTest({
   plugin: ".", // TODO: the plugin dir holding this subagent
   // TODO: a prompt that dispatches ${input.name} (via the Task tool).
@@ -124,6 +188,61 @@ ${toolHint}
 console.log("✓ ${input.name}: subagent test ran");
 `;
 }
+/**
+ * The SAFETY check, GENERATED FROM the subagent's side-effecting `tools`: assert it
+ * stays inside its declared write surface and never reaches for a destructive op.
+ * The `tools` allowlist + effectSurface identify the "hole"; the check asserts it —
+ * a test the typed contract writes for you (markdown can declare the tools, not test them).
+ */
+function safetySection(input, sideEffecting) {
+    const checks = [];
+    if (sideEffecting.includes("Bash")) {
+        checks.push(`  notTool("Bash", { command: /git push|rm -rf/ }), // never a destructive op`);
+    }
+    if (sideEffecting.some((t) => t === "Write" || t === "Edit")) {
+        checks.push(`  didNotWrite("secrets.env"), // TODO: the path(s) it must NOT write outside its surface`);
+    }
+    if (checks.length === 0) {
+        checks.push(`  // TODO: a notTool()/didNotWrite() per side-effecting tool: ${sideEffecting.join(", ")}`);
+    }
+    return `
+// --- Safety (deterministic) — generated from ${input.name}'s side-effecting tools: ${sideEffecting.join(", ")} ---
+// In a real run, replace this constructed Trace with a real \`runHarness\` /
+// \`measure\` turn (use interceptTools so a real model's attempt is DENIED, never
+// executed — see docs/eval-architecture.md). The checks below are derived from the
+// declared tools contract — the agent's "hole" asserted to stay in its lane.
+{
+  const trace = {
+    output: "done",
+    turns: 1,
+    hooks: [],
+    toolCalls: [
+      // TODO: the tool calls a benign run of ${input.name} makes.
+      { name: "Bash", input: { command: "git status" } },
+    ],
+    file: () => null,
+  };
+  assertChecks(trace, [
+${checks.join("\n")}
+  ]);
+  console.log("✓ ${input.name}: stayed inside its declared side-effect surface");
+}
+`;
+}
+/** A subagent → deterministic outcome + safety tests, generated from its typed contract. */
+function agentScaffold(input) {
+    const head = header(`Starter harness test for the \`${input.name}\` subagent.`, `npx vigiles test ${suggestedPath(input)}`);
+    const safetyImport = input.sideEffectingTools && input.sideEffectingTools.length > 0
+        ? `import { notTool, didNotWrite, assertChecks } from "vigiles/testing";\n`
+        : "";
+    const body = input.resultContract
+        ? outcomeSection(input, input.resultContract)
+        : fallbackSection(input);
+    const safety = input.sideEffectingTools && input.sideEffectingTools.length > 0
+        ? safetySection(input, input.sideEffectingTools)
+        : "";
+    return `${head}\n${safetyImport}${body}${safety}`;
+}
 const TIER = {
     hook: "unit",
     skill: "eval",

package/dist/scan-behavioral.d.ts

@@ -70,4 +70,64 @@ export declare function probePluginTriggersWith(dir: string, promptSet: TriggerP
 export declare function probePluginTriggers(dir: string, promptSet: TriggerPromptSet, opts?: ProbeOptions): Promise<BehavioralReport>;
 /** Format the behavioral column as a scan-report section. */
 export declare function formatBehavioralReport(b: BehavioralReport): string;
+/** Knobs for the selection-collision measurement. */
+export interface SelectionOptions {
+    /** Repeats per prompt (default 1). */
+    readonly trials?: number;
+    /** Selector model — defaults to Sonnet (a weaker model under-selects). */
+    readonly model?: string;
+    /** Parallel runs across the prompts × trials grid (default 1). */
+    readonly concurrency?: number;
+    /** Which harness drives it (default `"claude-code"`; others report n/a). */
+    readonly harness?: ProbeHarness;
+}
+/** One run's outcome for the matrix: which of the plugin's OWN skills fired. */
+interface SelectionRun {
+    readonly intended: string;
+    readonly firedBare: readonly string[];
+}
+export interface SkillSelectionStat {
+    readonly skill: string;
+    /** Fraction of its own prompts on which it fired (matrix diagonal). */
+    readonly recall: number;
+    /** Fraction of its own prompts on which a SIBLING skill also/instead fired. */
+    readonly collisionRate: number;
+    /** Non-errored runs measured for this skill. */
+    readonly n: number;
+    /** Sibling skills that fired on this skill's prompts, by rate (desc, rate>0). */
+    readonly collidesWith: readonly {
+        readonly skill: string;
+        readonly rate: number;
+    }[];
+}
+export interface SelectionReport {
+    /** False when the harness CLI / auth is absent, or the harness has no selector. */
+    readonly available: boolean;
+    /** Matrix axes — the plugin's model-invocable skill names (bare). */
+    readonly skills: readonly string[];
+    /** matrix[i][j] = times skill j fired when skill i's prompt was given. */
+    readonly matrix: readonly (readonly number[])[];
+    readonly perSkill: readonly SkillSelectionStat[];
+    /** Plugin-level: fraction of all runs where a non-intended skill fired. */
+    readonly collisionRate: number;
+    readonly n: number;
+    readonly note?: string;
+}
+/**
+ * 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).
+ */
+export declare function buildSelectionReport(skills: readonly string[], runs: readonly SelectionRun[]): SelectionReport;
+/** The injectable core (for tests): drive the matrix via a fake/real probe. */
+export declare function measurePluginSelectionWith(dir: string, promptSet: TriggerPromptSet, probe: HarnessProbe, opts?: SelectionOptions): Promise<SelectionReport>;
+/**
+ * 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.
+ */
+export declare function measurePluginSelection(dir: string, promptSet: TriggerPromptSet, opts?: SelectionOptions): Promise<SelectionReport>;
+/** Format the selection-collision matrix as a scan-report section. */
+export declare function formatSelectionReport(r: SelectionReport): string;
+export {};
 //# sourceMappingURL=scan-behavioral.d.ts.map

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