vskill 1.0.0 → 1.0.2

package/agents.json

@@ -1,6 +1,6 @@
 {
   "version": 1,
-  "generatedAt": "2026-04-27T07:19:05.521Z",
+  "generatedAt": "2026-04-27T15:07:05.009Z",
   "agentPrefixes": [
     ".adal",
     ".agent",

package/dist/commands/plugin.d.ts

@@ -0,0 +1,9 @@
+import type { Command } from "commander";
+interface PluginNewOpts {
+    description?: string;
+    withSkill?: string;
+    cwd?: string;
+}
+export declare function pluginNewCommand(pluginName: string, opts: PluginNewOpts): Promise<void>;
+export declare function registerPluginCommand(program: Command): void;
+export {};

package/dist/commands/plugin.js

@@ -0,0 +1,123 @@
+// ---------------------------------------------------------------------------
+// 0793 — `vskill plugin <subcommand>` family.
+//
+// Today: `new <name>` scaffolds a Claude Code plugin (`.claude-plugin/plugin.json`)
+// with optional first skill (`--with-skill <slug>`). We delegate schema
+// validation to `claude plugin validate <path>` so vskill never duplicates
+// Claude Code's plugin schema. When `claude` is not on PATH, validation
+// soft-skips with a warning so vskill stays usable in CI.
+//
+// This command exists because Claude Code's `claude plugin` CLI has no
+// `new`/`create`/`init` subcommand — plugin authoring is otherwise unowned.
+// ---------------------------------------------------------------------------
+import { existsSync, mkdirSync, unlinkSync, writeFileSync } from "node:fs";
+import { resolve, join } from "node:path";
+import { pluginJsonScaffold, skillMdScaffold, validateKebabName, } from "../eval-server/authoring-routes.js";
+import { validateClaudePlugin } from "../core/plugin-validator.js";
+import { bold, cyan, dim, green, red, yellow } from "../utils/output.js";
+function exit(code) {
+    process.exit(code);
+    // satisfies TS in tests where exit is spied
+    throw new Error("process.exit did not terminate");
+}
+export async function pluginNewCommand(pluginName, opts) {
+    const nameErr = validateKebabName(pluginName, "plugin name");
+    if (nameErr) {
+        console.error(red("Error: ") + dim(nameErr));
+        exit(1);
+    }
+    const cwd = opts.cwd ? resolve(opts.cwd) : process.cwd();
+    const pluginDir = join(cwd, pluginName);
+    const manifestDir = join(pluginDir, ".claude-plugin");
+    const manifestPath = join(manifestDir, "plugin.json");
+    if (existsSync(manifestPath)) {
+        console.error(red("Error: ") +
+            dim(`plugin manifest already exists: ${manifestPath}`));
+        console.error(dim("  Pick a different name or delete the existing manifest first."));
+        exit(1);
+    }
+    let withSkill = null;
+    if (opts.withSkill !== undefined) {
+        const skillErr = validateKebabName(opts.withSkill, "--with-skill");
+        if (skillErr) {
+            console.error(red("Error: ") + dim(skillErr));
+            exit(1);
+        }
+        withSkill = opts.withSkill;
+        const skillDir = join(pluginDir, "skills", withSkill);
+        if (existsSync(skillDir)) {
+            console.error(red("Error: ") +
+                dim(`skill directory already exists: ${skillDir}`));
+            exit(1);
+        }
+    }
+    const description = (opts.description ?? "").trim() || `Plugin: ${pluginName}`;
+    // Write the manifest (and optional first skill) as an all-or-nothing block.
+    // If validation fails we roll back the manifest; we leave the (newly
+    // created) skill folder in place because deleting user-visible new files we
+    // just created is more surprising than leaving them.
+    try {
+        mkdirSync(manifestDir, { recursive: true });
+        writeFileSync(manifestPath, pluginJsonScaffold(pluginName, description), "utf8");
+        if (withSkill) {
+            const skillDir = join(pluginDir, "skills", withSkill);
+            mkdirSync(skillDir, { recursive: true });
+            writeFileSync(join(skillDir, "SKILL.md"), skillMdScaffold(withSkill, description), "utf8");
+        }
+    }
+    catch (err) {
+        const message = err instanceof Error ? err.message : String(err);
+        console.error(red("Error: ") + dim(`failed to write plugin files: ${message}`));
+        exit(1);
+    }
+    // Schema validation via `claude plugin validate` — single source of truth
+    // for the plugin schema.
+    const validation = validateClaudePlugin(pluginDir);
+    if (!validation.ok) {
+        try {
+            unlinkSync(manifestPath);
+        }
+        catch {
+            /* best-effort rollback */
+        }
+        console.error(red("Error: ") + dim("claude plugin validate rejected the generated manifest:"));
+        console.error(validation.stderr || "(no stderr captured)");
+        console.error(dim("  The manifest has been removed. Please file an issue if this is unexpected."));
+        exit(1);
+    }
+    if (validation.skipped) {
+        console.warn(yellow("Warning: ") +
+            dim("`claude` CLI not found on PATH; plugin schema was NOT validated."));
+        console.warn(dim("  Install Claude Code (https://claude.com/claude-code) to enable validation."));
+    }
+    // Success summary
+    console.log(green("✔ Plugin scaffolded"));
+    console.log(dim("  manifest: ") + manifestPath);
+    if (withSkill) {
+        console.log(dim("  skill:    ") + join(pluginDir, "skills", withSkill, "SKILL.md"));
+    }
+    console.log("");
+    console.log(bold("Next steps:"));
+    console.log("  " + cyan(`cd ${pluginName}`));
+    if (!withSkill) {
+        console.log("  " +
+            cyan(`vskill skill new --prompt "..."`) +
+            dim("  # add your first skill"));
+    }
+    console.log("  " + cyan(`vskill studio`) + dim("  # open in Skill Studio"));
+}
+export function registerPluginCommand(program) {
+    const plugin = program
+        .command("plugin")
+        .description("Plugin authoring: scaffold a Claude Code plugin (.claude-plugin/plugin.json + skills/)");
+    plugin
+        .command("new <name>")
+        .description("Scaffold a new plugin folder with manifest and (optional) first skill")
+        .option("--description <text>", "Description written into plugin.json")
+        .option("--with-skill <slug>", "Also scaffold <name>/skills/<slug>/SKILL.md as the first skill")
+        .option("--cwd <path>", "Create the plugin under this directory (default: process.cwd())")
+        .action(async (name, opts) => {
+        await pluginNewCommand(name, opts);
+    });
+}
+//# sourceMappingURL=plugin.js.map

package/dist/commands/plugin.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"plugin.js","sourceRoot":"","sources":["../../src/commands/plugin.ts"],"names":[],"mappings":"AAAA,8EAA8E;AAC9E,8CAA8C;AAC9C,EAAE;AACF,oFAAoF;AACpF,wEAAwE;AACxE,2EAA2E;AAC3E,wEAAwE;AACxE,0DAA0D;AAC1D,EAAE;AACF,uEAAuE;AACvE,4EAA4E;AAC5E,8EAA8E;AAE9E,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,UAAU,EAAE,aAAa,EAAE,MAAM,SAAS,CAAC;AAC3E,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AAG1C,OAAO,EACL,kBAAkB,EAClB,eAAe,EACf,iBAAiB,GAClB,MAAM,oCAAoC,CAAC;AAC5C,OAAO,EAAE,oBAAoB,EAAE,MAAM,6BAA6B,CAAC;AACnE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,GAAG,EAAE,KAAK,EAAE,GAAG,EAAE,MAAM,EAAE,MAAM,oBAAoB,CAAC;AAQzE,SAAS,IAAI,CAAC,IAAY;IACxB,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IACnB,4CAA4C;IAC5C,MAAM,IAAI,KAAK,CAAC,gCAAgC,CAAC,CAAC;AACpD,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,gBAAgB,CACpC,UAAkB,EAClB,IAAmB;IAEnB,MAAM,OAAO,GAAG,iBAAiB,CAAC,UAAU,EAAE,aAAa,CAAC,CAAC;IAC7D,IAAI,OAAO,EAAE,CAAC;QACZ,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC;QAC7C,IAAI,CAAC,CAAC,CAAC,CAAC;IACV,CAAC;IAED,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,GAAG,EAAE,CAAC;IACzD,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,UAAU,CAAC,CAAC;IACxC,MAAM,WAAW,GAAG,IAAI,CAAC,SAAS,EAAE,gBAAgB,CAAC,CAAC;IACtD,MAAM,YAAY,GAAG,IAAI,CAAC,WAAW,EAAE,aAAa,CAAC,CAAC;IAEtD,IAAI,UAAU,CAAC,YAAY,CAAC,EAAE,CAAC;QAC7B,OAAO,CAAC,KAAK,CACX,GAAG,CAAC,SAAS,CAAC;YACZ,GAAG,CAAC,mCAAmC,YAAY,EAAE,CAAC,CACzD,CAAC;QACF,OAAO,CAAC,KAAK,CACX,GAAG,CAAC,gEAAgE,CAAC,CACtE,CAAC;QACF,IAAI,CAAC,CAAC,CAAC,CAAC;IACV,CAAC;IAED,IAAI,SAAS,GAAkB,IAAI,CAAC;IACpC,IAAI,IAAI,CAAC,SAAS,KAAK,SAAS,EAAE,CAAC;QACjC,MAAM,QAAQ,GAAG,iBAAiB,CAAC,IAAI,CAAC,SAAS,EAAE,cAAc,CAAC,CAAC;QACnE,IAAI,QAAQ,EAAE,CAAC;YACb,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,GAAG,CAAC,QAAQ,CAAC,CAAC,CAAC;YAC9C,IAAI,CAAC,CAAC,CAAC,CAAC;QACV,CAAC;QACD,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC;QAC3B,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;QACtD,IAAI,UAAU,CAAC,QAAQ,CAAC,EAAE,CAAC;YACzB,OAAO,CAAC,KAAK,CACX,GAAG,CAAC,SAAS,CAAC;gBACZ,GAAG,CAAC,mCAAmC,QAAQ,EAAE,CAAC,CACrD,CAAC;YACF,IAAI,CAAC,CAAC,CAAC,CAAC;QACV,CAAC;IACH,CAAC;IAED,MAAM,WAAW,GAAG,CAAC,IAAI,CAAC,WAAW,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,IAAI,WAAW,UAAU,EAAE,CAAC;IAE/E,4EAA4E;IAC5E,qEAAqE;IACrE,4EAA4E;IAC5E,qDAAqD;IACrD,IAAI,CAAC;QACH,SAAS,CAAC,WAAW,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;QAC5C,aAAa,CAAC,YAAY,EAAE,kBAAkB,CAAC,UAAU,EAAE,WAAW,CAAC,EAAE,MAAM,CAAC,CAAC;QAEjF,IAAI,SAAS,EAAE,CAAC;YACd,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,EAAE,QAAQ,EAAE,SAAS,CAAC,CAAC;YACtD,SAAS,CAAC,QAAQ,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAC;YACzC,aAAa,CACX,IAAI,CAAC,QAAQ,EAAE,UAAU,CAAC,EAC1B,eAAe,CAAC,SAAS,EAAE,WAAW,CAAC,EACvC,MAAM,CACP,CAAC;QACJ,CAAC;IACH,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACb,MAAM,OAAO,GAAG,GAAG,YAAY,KAAK,CAAC,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QACjE,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,GAAG,CAAC,iCAAiC,OAAO,EAAE,CAAC,CAAC,CAAC;QAChF,IAAI,CAAC,CAAC,CAAC,CAAC;IACV,CAAC;IAED,0EAA0E;IAC1E,yBAAyB;IACzB,MAAM,UAAU,GAAG,oBAAoB,CAAC,SAAS,CAAC,CAAC;IACnD,IAAI,CAAC,UAAU,CAAC,EAAE,EAAE,CAAC;QACnB,IAAI,CAAC;YACH,UAAU,CAAC,YAAY,CAAC,CAAC;QAC3B,CAAC;QAAC,MAAM,CAAC;YACP,0BAA0B;QAC5B,CAAC;QACD,OAAO,CAAC,KAAK,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,GAAG,CAAC,yDAAyD,CAAC,CAAC,CAAC;QAC/F,OAAO,CAAC,KAAK,CAAC,UAAU,CAAC,MAAM,IAAI,sBAAsB,CAAC,CAAC;QAC3D,OAAO,CAAC,KAAK,CACX,GAAG,CAAC,8EAA8E,CAAC,CACpF,CAAC;QACF,IAAI,CAAC,CAAC,CAAC,CAAC;IACV,CAAC;IAED,IAAI,UAAU,CAAC,OAAO,EAAE,CAAC;QACvB,OAAO,CAAC,IAAI,CACV,MAAM,CAAC,WAAW,CAAC;YACjB,GAAG,CAAC,kEAAkE,CAAC,CAC1E,CAAC;QACF,OAAO,CAAC,IAAI,CACV,GAAG,CAAC,8EAA8E,CAAC,CACpF,CAAC;IACJ,CAAC;IAED,kBAAkB;IAClB,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,qBAAqB,CAAC,CAAC,CAAC;IAC1C,OAAO,CAAC,GAAG,CAAC,GAAG,CAAC,cAAc,CAAC,GAAG,YAAY,CAAC,CAAC;IAChD,IAAI,SAAS,EAAE,CAAC;QACd,OAAO,CAAC,GAAG,CACT,GAAG,CAAC,cAAc,CAAC,GAAG,IAAI,CAAC,SAAS,EAAE,QAAQ,EAAE,SAAS,EAAE,UAAU,CAAC,CACvE,CAAC;IACJ,CAAC;IACD,OAAO,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;IAChB,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC;IACjC,OAAO,CAAC,GAAG,CAAC,IAAI,GAAG,IAAI,CAAC,MAAM,UAAU,EAAE,CAAC,CAAC,CAAC;IAC7C,IAAI,CAAC,SAAS,EAAE,CAAC;QACf,OAAO,CAAC,GAAG,CACT,IAAI;YACF,IAAI,CAAC,iCAAiC,CAAC;YACvC,GAAG,CAAC,0BAA0B,CAAC,CAClC,CAAC;IACJ,CAAC;IACD,OAAO,CAAC,GAAG,CAAC,IAAI,GAAG,IAAI,CAAC,eAAe,CAAC,GAAG,GAAG,CAAC,0BAA0B,CAAC,CAAC,CAAC;AAC9E,CAAC;AAED,MAAM,UAAU,qBAAqB,CAAC,OAAgB;IACpD,MAAM,MAAM,GAAG,OAAO;SACnB,OAAO,CAAC,QAAQ,CAAC;SACjB,WAAW,CACV,wFAAwF,CACzF,CAAC;IAEJ,MAAM;SACH,OAAO,CAAC,YAAY,CAAC;SACrB,WAAW,CAAC,uEAAuE,CAAC;SACpF,MAAM,CAAC,sBAAsB,EAAE,sCAAsC,CAAC;SACtE,MAAM,CACL,qBAAqB,EACrB,gEAAgE,CACjE;SACA,MAAM,CAAC,cAAc,EAAE,iEAAiE,CAAC;SACzF,MAAM,CAAC,KAAK,EAAE,IAAY,EAAE,IAAmB,EAAE,EAAE;QAClD,MAAM,gBAAgB,CAAC,IAAI,EAAE,IAAI,CAAC,CAAC;IACrC,CAAC,CAAC,CAAC;AACP,CAAC"}

package/dist/core/plugin-validator.d.ts

@@ -0,0 +1,17 @@
+export interface PluginValidationResult {
+    ok: boolean;
+    skipped: boolean;
+    stderr: string;
+}
+export interface ValidateClaudePluginOptions {
+    /**
+     * Override the binary name (tests inject a fake "claude").
+     */
+    bin?: string;
+    /**
+     * Timeout in ms. Defaults to 10s — `claude plugin validate` is local-only
+     * and typically returns in <300ms; 10s is generous headroom.
+     */
+    timeoutMs?: number;
+}
+export declare function validateClaudePlugin(pluginPath: string, opts?: ValidateClaudePluginOptions): PluginValidationResult;

package/dist/core/plugin-validator.js

@@ -0,0 +1,45 @@
+// ---------------------------------------------------------------------------
+// 0793 — `claude plugin validate <path>` wrapper.
+//
+// vskill never duplicates Claude Code's plugin schema. Both `vskill plugin new`
+// (CLI) and `POST /api/authoring/convert-to-plugin` (Studio) call this helper
+// after writing a manifest to confirm the JSON we wrote conforms. When `claude`
+// is not on PATH (e.g. CI without Claude Code installed), validation soft-skips
+// with `{ ok: true, skipped: true }` so vskill stays usable; callers should
+// surface that state to the user.
+// ---------------------------------------------------------------------------
+import { spawnSync } from "node:child_process";
+export function validateClaudePlugin(pluginPath, opts = {}) {
+    const bin = opts.bin ?? "claude";
+    const result = spawnSync(bin, ["plugin", "validate", pluginPath], {
+        encoding: "utf8",
+        timeout: opts.timeoutMs ?? 10_000,
+    });
+    // ENOENT => `claude` not on PATH. Soft-skip so vskill remains usable in
+    // environments without Claude Code (CI, fresh dev boxes, etc).
+    // `result.error` is a NodeJS.ErrnoException when spawn itself failed.
+    const err = result.error;
+    if (err && err.code === "ENOENT") {
+        return { ok: true, skipped: true, stderr: "" };
+    }
+    if (err) {
+        // Any other spawn error (timeout, permission, etc): surface but don't
+        // treat as schema failure — caller decides what to do.
+        return {
+            ok: false,
+            skipped: false,
+            stderr: `failed to invoke ${bin}: ${err.message}`,
+        };
+    }
+    if (result.status === 0) {
+        return { ok: true, skipped: false, stderr: "" };
+    }
+    // `claude plugin validate` writes diagnostics to stderr (or stdout in some
+    // versions). Concatenate both so callers always see the message.
+    const stderr = [result.stderr ?? "", result.stdout ?? ""]
+        .filter((s) => s && s.trim().length > 0)
+        .join("\n")
+        .trim();
+    return { ok: false, skipped: false, stderr };
+}
+//# sourceMappingURL=plugin-validator.js.map

package/dist/core/plugin-validator.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"plugin-validator.js","sourceRoot":"","sources":["../../src/core/plugin-validator.ts"],"names":[],"mappings":"AAAA,8EAA8E;AAC9E,kDAAkD;AAClD,EAAE;AACF,gFAAgF;AAChF,8EAA8E;AAC9E,gFAAgF;AAChF,gFAAgF;AAChF,4EAA4E;AAC5E,kCAAkC;AAClC,8EAA8E;AAE9E,OAAO,EAAE,SAAS,EAAE,MAAM,oBAAoB,CAAC;AAoB/C,MAAM,UAAU,oBAAoB,CAClC,UAAkB,EAClB,OAAoC,EAAE;IAEtC,MAAM,GAAG,GAAG,IAAI,CAAC,GAAG,IAAI,QAAQ,CAAC;IACjC,MAAM,MAAM,GAAG,SAAS,CAAC,GAAG,EAAE,CAAC,QAAQ,EAAE,UAAU,EAAE,UAAU,CAAC,EAAE;QAChE,QAAQ,EAAE,MAAM;QAChB,OAAO,EAAE,IAAI,CAAC,SAAS,IAAI,MAAM;KAClC,CAAC,CAAC;IAEH,wEAAwE;IACxE,+DAA+D;IAC/D,sEAAsE;IACtE,MAAM,GAAG,GAAG,MAAM,CAAC,KAA4C,CAAC;IAChE,IAAI,GAAG,IAAI,GAAG,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;QACjC,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,EAAE,EAAE,CAAC;IACjD,CAAC;IACD,IAAI,GAAG,EAAE,CAAC;QACR,sEAAsE;QACtE,uDAAuD;QACvD,OAAO;YACL,EAAE,EAAE,KAAK;YACT,OAAO,EAAE,KAAK;YACd,MAAM,EAAE,oBAAoB,GAAG,KAAK,GAAG,CAAC,OAAO,EAAE;SAClD,CAAC;IACJ,CAAC;IAED,IAAI,MAAM,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;QACxB,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,KAAK,EAAE,MAAM,EAAE,EAAE,EAAE,CAAC;IAClD,CAAC;IAED,2EAA2E;IAC3E,iEAAiE;IACjE,MAAM,MAAM,GAAG,CAAC,MAAM,CAAC,MAAM,IAAI,EAAE,EAAE,MAAM,CAAC,MAAM,IAAI,EAAE,CAAC;SACtD,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,IAAI,CAAC,CAAC,IAAI,EAAE,CAAC,MAAM,GAAG,CAAC,CAAC;SACvC,IAAI,CAAC,IAAI,CAAC;SACV,IAAI,EAAE,CAAC;IACV,OAAO,EAAE,EAAE,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,EAAE,MAAM,EAAE,CAAC;AAC/C,CAAC"}

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

@@ -951,6 +951,17 @@ function getEffectiveRawModel() {
 function getClient() {
     return createLlmClient(currentOverrides);
 }
+// Per-request client: prefer body-supplied provider/model, then session globals.
+// Lets the frontend send the user's selected model with each request so historical
+// runs reliably reflect the picker value rather than a stale session default.
+function clientFromBody(body) {
+    const reqProvider = typeof body?.provider === "string" ? body.provider : undefined;
+    const reqModel = typeof body?.model === "string" ? body.model : undefined;
+    const provider = (reqProvider || currentOverrides.provider || "claude-cli");
+    const model = reqModel || currentOverrides.model;
+    const client = createLlmClient({ provider, model });
+    return { client, provider };
+}
 /** Derive sidebar badge status from benchmark + current eval IDs. */
 function computeBenchmarkStatus(benchmark, evalIds, hasEvals) {
     if (!benchmark)
@@ -2549,7 +2560,7 @@ export function registerRoutes(router, root, projectName) {
             const evals = loadAndValidateEvals(skillDir);
             const skillMdPath = join(skillDir, "SKILL.md");
             const skillContent = existsSync(skillMdPath) ? readFileSync(skillMdPath, "utf-8") : "";
-            const client = getClient();
+            const { client, provider: effectiveProvider } = clientFromBody(body);
             const systemPrompt = buildEvalSystemPrompt(skillContent);
             // Create separate judge client if judgeModel is specified
             let judgeClient;
@@ -2570,7 +2581,7 @@ export function registerRoutes(router, root, projectName) {
             }
             await runBenchmarkSSE({
                 res, skillDir, skillName: evals.skill_name, systemPrompt,
-                runType: "benchmark", provider: currentOverrides.provider || "claude-cli",
+                runType: "benchmark", provider: effectiveProvider,
                 evalCases: evals.evals, filterIds, client, judgeClient, judgeCache,
                 isAborted: () => aborted, concurrency,
             });
@@ -2592,11 +2603,11 @@ export function registerRoutes(router, root, projectName) {
         initSSE(res, req);
         try {
             const evals = loadAndValidateEvals(skillDir);
-            const client = getClient();
+            const { client, provider: effectiveProvider } = clientFromBody(body);
             await runBenchmarkSSE({
                 res, skillDir, skillName: evals.skill_name,
                 systemPrompt: "You are a helpful AI assistant.",
-                runType: "baseline", provider: currentOverrides.provider || "claude-cli",
+                runType: "baseline", provider: effectiveProvider,
                 evalCases: evals.evals, filterIds, client, isAborted: () => aborted,
             });
         }
@@ -2636,14 +2647,14 @@ export function registerRoutes(router, root, projectName) {
             }
             const skillMdPath = join(skillDir, "SKILL.md");
             const skillContent = existsSync(skillMdPath) ? readFileSync(skillMdPath, "utf-8") : "";
-            const client = getClient();
+            const { client, provider: effectiveProvider } = clientFromBody(body);
             const systemPrompt = isBaseline
                 ? buildBaselineSystemPrompt()
                 : buildEvalSystemPrompt(skillContent);
             await sem.acquire();
             const benchCase = await runSingleCaseSSE({
                 res, evalCase, systemPrompt, client, isAborted: () => aborted,
-                provider: currentOverrides.provider || "claude-cli",
+                provider: effectiveProvider,
             });
             if (!released) {
                 released = true;
@@ -2659,7 +2670,7 @@ export function registerRoutes(router, root, projectName) {
                         cases: [benchCase],
                         overall_pass_rate: benchCase.pass_rate,
                         type: isBaseline ? "baseline" : "benchmark",
-                        provider: currentOverrides.provider || "claude-cli",
+                        provider: effectiveProvider,
                         totalDurationMs: benchCase.durationMs ?? 0,
                         totalInputTokens: benchCase.inputTokens ?? null,
                         totalOutputTokens: benchCase.outputTokens ?? null,