vskill 0.5.105 → 0.5.107

package/agents.json

@@ -1,6 +1,6 @@
 {
   "version": 1,
-  "generatedAt": "2026-04-24T22:27:42.739Z",
+  "generatedAt": "2026-04-25T05:54:03.282Z",
   "agentPrefixes": [
     ".adal",
     ".agent",

package/dist/eval-server/api-routes.d.ts

@@ -99,6 +99,13 @@ export declare function deriveSourceAgent(skillDir: string, root: string, origin
  * response never fails because of a single bad skill.
  */
 export declare function buildSkillMetadata(skillDir: string, origin: "source" | "installed", root: string): SkillMetadataFields;
+/**
+ * 0682 F-001 — Test helper. Resets `currentOverrides` to the default and
+ * clears the `studioLoaded` flag so subsequent /api/config calls re-attempt
+ * loadStudioSelection. Production code never needs this; it exists solely so
+ * vitest can simulate a fresh server boot per-case.
+ */
+export declare function resetStudioRestoreState(): void;
 interface ModelOption {
     id: string;
     label: string;
@@ -106,6 +113,7 @@ interface ModelOption {
         prompt: number;
         completion: number;
     };
+    resolvedId?: string;
 }
 export declare const PROVIDER_MODELS: Record<ProviderName, ModelOption[]>;
 type OpenRouterCacheEntry = {

package/dist/eval-server/api-routes.js

@@ -1,7 +1,7 @@
 // ---------------------------------------------------------------------------
 // api-routes.ts -- REST API route handlers for the eval UI
 // ---------------------------------------------------------------------------
-import { readFileSync, writeFileSync, mkdirSync, existsSync, rmSync, readdirSync, statSync } from "node:fs";
+import { readFileSync, writeFileSync, mkdirSync, existsSync, readdirSync, statSync } from "node:fs";
 import { execSync } from "node:child_process";
 import { join, resolve, dirname } from "node:path";
 import { homedir } from "node:os";
@@ -10,7 +10,7 @@ import { initSSE, sendSSE, sendSSEDone, withHeartbeat, startDynamicHeartbeat } f
 import { dataEventBus, emitDataEvent } from "./data-events.js";
 import { classifyError } from "./error-classifier.js";
 import { readLockfile } from "../lockfile/lockfile.js";
-import { parseSource } from "../resolvers/source-resolver.js";
+import { resolveSkillApiName as resolveSkillApiNameImpl } from "./skill-name-resolver.js";
 import { runBenchmarkSSE, runSingleCaseSSE } from "./benchmark-runner.js";
 import { getSkillSemaphore } from "./concurrency.js";
 import { resolveSkillDir } from "./skill-resolver.js";
@@ -18,7 +18,7 @@ import { classifyOrigin, scanSkillsTriScope } from "../eval/skill-scanner.js";
 import { scanInstalledPluginSkills, scanAuthoredPluginSkills, } from "../eval/plugin-scanner.js";
 import { resolveGlobalSkillsDir } from "../eval/path-utils.js";
 import { loadAndValidateEvals, EvalValidationError } from "../eval/schema.js";
-import { ANTHROPIC_CATALOG_SNAPSHOT } from "../eval/anthropic-catalog.js";
+import { ANTHROPIC_CATALOG_SNAPSHOT, findAnthropicModel } from "../eval/anthropic-catalog.js";
 import { readBenchmark } from "../eval/benchmark.js";
 import { writeHistoryEntry, listHistory, readHistoryEntry, computeRegressions, deleteHistoryEntry, getCaseHistory, computeStats } from "../eval/benchmark-history.js";
 import { judgeAssertion } from "../eval/judge.js";
@@ -455,6 +455,25 @@ export function buildSkillMetadata(skillDir, origin, root) {
 // (which only matters for the `vskill eval run` CLI command).
 // ---------------------------------------------------------------------------
 let currentOverrides = { provider: "claude-cli" };
+// 0682 F-001 — Tracks whether the persistent studio.json selection has been
+// loaded for this process lifetime. Initialized to false; set to true once
+// loadStudioSelection(root) has been attempted (regardless of whether a file
+// existed). Without this flag, the prior gate `if (!currentOverrides.provider)`
+// was permanently false (the default `{ provider: "claude-cli" }` already
+// populates `provider`), so the persisted selection at .vskill/studio.json
+// was silently discarded on every server boot — a CRITICAL regression of
+// AC-US1-03 / FR-005.
+let studioLoaded = false;
+/**
+ * 0682 F-001 — Test helper. Resets `currentOverrides` to the default and
+ * clears the `studioLoaded` flag so subsequent /api/config calls re-attempt
+ * loadStudioSelection. Production code never needs this; it exists solely so
+ * vitest can simulate a fresh server boot per-case.
+ */
+export function resetStudioRestoreState() {
+    currentOverrides = { provider: "claude-cli" };
+    studioLoaded = false;
+}
 /** Return the effective raw model ID (suitable for round-tripping via the API). */
 function getEffectiveRawModel() {
     if (currentOverrides.model)
@@ -496,11 +515,21 @@ function buildAnthropicProviderModels() {
         },
     }));
 }
+function aliasInfo(alias, fallbackLabel) {
+    const entry = findAnthropicModel(alias);
+    if (!entry)
+        return { label: fallbackLabel };
+    return { label: entry.displayName, resolvedId: entry.id };
+}
 export const PROVIDER_MODELS = {
+    // Opus first so it is the default when no override is set
+    // (getEffectiveRawModel returns models[0]). Labels come from the catalog so
+    // the picker shows the exact dated version (e.g. "Claude Opus 4.7"), not the
+    // bare family name — keeps the Studio truthful when a model is bumped.
     "claude-cli": [
-        { id: "sonnet", label: "Claude Sonnet" },
-        { id: "opus", label: "Claude Opus" },
-        { id: "haiku", label: "Claude Haiku" },
+        { id: "opus", ...aliasInfo("opus", "Claude Opus") },
+        { id: "sonnet", ...aliasInfo("sonnet", "Claude Sonnet") },
+        { id: "haiku", ...aliasInfo("haiku", "Claude Haiku") },
     ],
     "anthropic": buildAnthropicProviderModels(),
     "ollama": [
@@ -959,8 +988,13 @@ export function registerRoutes(router, root, projectName) {
     // (e.g. "claude-sonnet"). The frontend round-trips config.model back to
     // generate-evals and other endpoints, so it must be a valid CLI model ID.
     router.get("/api/config", async (_req, res) => {
-        // On first load (no currentOverrides), try to restore from .vskill/studio.json.
-        if (!currentOverrides.provider) {
+        // 0682 F-001 — Boot-time restoration of .vskill/studio.json selection.
+        // Use a dedicated `studioLoaded` flag rather than checking
+        // `!currentOverrides.provider` because the module default
+        // `{ provider: "claude-cli" }` would make the prior guard permanently
+        // false, silently discarding the persisted selection on every boot.
+        if (!studioLoaded) {
+            studioLoaded = true;
             const stored = loadStudioSelection(root);
             if (stored) {
                 currentOverrides.provider = stored.activeAgent;
@@ -1154,19 +1188,13 @@ export function registerRoutes(router, root, projectName) {
     // MUST be registered BEFORE the /:plugin/:skill catch-all below.
     // -------------------------------------------------------------------------
     const PLATFORM_BASE = "https://verified-skill.com";
-    /** Resolve plugin/skill to full hierarchical API name using lockfile source. */
+    /**
+     * Resolve plugin/skill to full hierarchical API name. Lockfile path first
+     * (installed skills); falls back to authored-skill discovery on disk +
+     * git remote parse for skills authored in this repo. See skill-name-resolver.ts.
+     */
     function resolveSkillApiName(skill) {
-        const lock = readLockfile();
-        if (!lock)
-            return skill;
-        const entry = lock.skills[skill];
-        if (!entry?.source)
-            return skill;
-        const parsed = parseSource(entry.source);
-        if (parsed.type === "github" || parsed.type === "github-plugin" || parsed.type === "marketplace") {
-            return `${parsed.owner}/${parsed.repo}/${skill}`;
-        }
-        return skill;
+        return resolveSkillApiNameImpl(skill, root);
     }
     // T-009 (proxy) + 0707 T-021 (harden): Versions endpoint
     //
@@ -1180,7 +1208,7 @@ export function registerRoutes(router, root, projectName) {
     //
     // Never returns 5xx for the "no VCS surface" case — that is normal empty state.
     router.get("/api/skills/:plugin/:skill/versions", async (req, res, params) => {
-        const fullName = resolveSkillApiName(params.skill);
+        const fullName = await resolveSkillApiName(params.skill);
         const parts = fullName.split("/");
         const apiPath = parts.length === 3
             ? `/api/v1/skills/${parts.map(encodeURIComponent).join("/")}/versions`
@@ -1231,11 +1259,11 @@ export function registerRoutes(router, root, projectName) {
             sendJson(res, { error: "Missing required query params: from and to" }, 400, req);
             return;
         }
-        const fullName = resolveSkillApiName(params.skill);
+        const fullName = await resolveSkillApiName(params.skill);
         const parts = fullName.split("/");
         const basePath = parts.length === 3
-            ? `/api/v1/skills/${parts.map(encodeURIComponent).join("/")}/versions`
-            : `/api/v1/skills/${encodeURIComponent(fullName)}/versions`;
+            ? `/api/v1/skills/${parts.map(encodeURIComponent).join("/")}/versions/diff`
+            : `/api/v1/skills/${encodeURIComponent(fullName)}/versions/diff`;
         try {
             const resp = await fetch(`${PLATFORM_BASE}${basePath}?from=${encodeURIComponent(from)}&to=${encodeURIComponent(to)}`, { signal: AbortSignal.timeout(10_000) });
             if (!resp.ok) {
@@ -1475,7 +1503,10 @@ export function registerRoutes(router, root, projectName) {
             return;
         }
         try {
-            rmSync(skillDir, { recursive: true, force: true });
+            // 0722: route to OS trash (Trash / Recycle Bin / XDG) instead of hard delete
+            // so accidental deletes are recoverable from the user's native Trash app.
+            const trash = (await import("trash")).default;
+            await trash([skillDir]);
             sendJson(res, { ok: true, deleted: `${params.plugin}/${params.skill}` }, 200, req);
         }
         catch (err) {