@wrongstack/cli 0.306.4 → 0.307.0

package/dist/{modeldiag-6VS5KDP5.js → modeldiag-ICMPHT5L.js}

@@ -1,208 +1,26 @@
 import {
   discoverAndMergeProviders
 } from "./chunk-52C7PD2H.js";
+import {
+  activeProfileConfigPath
+} from "./chunk-YMXXOOFN.js";
 import {
   mutateConfigProviders,
   normalizeKeys,
   writeKeysBack
 } from "./chunk-SZ42FYPT.js";
-import {
-  activeProfileConfigPath
-} from "./chunk-YMXXOOFN.js";
 import "./chunk-7OCVIDC7.js";
 
 // src/subcommands/handlers/modeldiag.ts
 import * as fs from "node:fs/promises";
-import { color as color2, toErrorMessage as toErrorMessage2 } from "@wrongstack/core/utils";
-import { makeProviderFromConfig, setOAuthTokenPersister } from "@wrongstack/providers";
+import { color as color5 } from "@wrongstack/core/utils";
 
-// src/subcommands/handlers/model-smoke-test.ts
-import { ProviderError } from "@wrongstack/core/types";
-import { toErrorMessage } from "@wrongstack/core/utils";
-var DEFAULT_TIMEOUT_MS = 45e3;
-var DEFAULT_MAX_TOKENS = 32;
-function csv(value) {
-  if (typeof value !== "string") return void 0;
-  const values = value.split(",").map((part) => part.trim()).filter(Boolean);
-  return values.length > 0 ? values : void 0;
-}
-function positiveInteger(value, fallback, label) {
-  if (value === void 0) return fallback;
-  const parsed = typeof value === "string" ? Number(value) : Number.NaN;
-  if (!Number.isSafeInteger(parsed) || parsed <= 0) {
-    throw new Error(`${label} must be a positive integer.`);
-  }
-  return parsed;
-}
-function parseModelSmokeOptions(args, flags = {}) {
-  const local = /* @__PURE__ */ new Map();
-  for (let i = 0; i < args.length; i++) {
-    const arg = args[i];
-    if (!arg?.startsWith("--")) continue;
-    const eq = arg.indexOf("=");
-    if (eq >= 0) {
-      local.set(arg.slice(2, eq), arg.slice(eq + 1));
-      continue;
-    }
-    const name = arg.slice(2);
-    const next = args[i + 1];
-    if (next && !next.startsWith("--")) {
-      local.set(name, next);
-      i++;
-    } else {
-      local.set(name, true);
-    }
-  }
-  const get = (name) => local.get(name) ?? flags[name];
-  const providerFilter = csv(get("providers") ?? get("provider"));
-  const modelFilter = csv(get("models") ?? get("model"));
-  return {
-    allModels: get("all-models") === true || get("all-models") === "true",
-    planOnly: get("plan") === true || get("plan") === "true",
-    json: get("json") === true || get("json") === "true",
-    timeoutMs: positiveInteger(get("timeout"), DEFAULT_TIMEOUT_MS, "--timeout"),
-    maxTokens: positiveInteger(get("max-tokens"), DEFAULT_MAX_TOKENS, "--max-tokens"),
-    ...providerFilter ? { providerFilter } : {},
-    ...modelFilter ? { modelFilter } : {}
-  };
-}
-function pushUnique(target, values) {
-  const seen = new Set(target);
-  for (const value of values) {
-    const normalized = value?.trim();
-    if (!normalized || seen.has(normalized)) continue;
-    seen.add(normalized);
-    target.push(normalized);
-  }
-}
-async function catalogModelsFor(providerId, providerConfig, registry) {
-  const direct = await registry.getProvider(providerId).catch(() => void 0);
-  const inherited = direct ?? (providerConfig?.type && providerConfig.type !== providerId ? await registry.getProvider(providerConfig.type).catch(() => void 0) : void 0);
-  return inherited?.models ?? [];
-}
-function supportsTextCompletion(model) {
-  const outputs = model?.modalities?.output;
-  return !outputs?.length || outputs.includes("text");
-}
-async function buildModelSmokeTargets(config, registry, options) {
-  const providerIds = Object.keys(config.providers ?? {});
-  if (config.provider && !providerIds.includes(config.provider)) providerIds.push(config.provider);
-  const providerFilter = options.providerFilter ? new Set(options.providerFilter.map((id) => id.toLowerCase())) : void 0;
-  const modelFilter = options.modelFilter ? new Set(options.modelFilter) : void 0;
-  const targets = [];
-  for (const providerId of providerIds) {
-    if (providerFilter && !providerFilter.has(providerId.toLowerCase())) continue;
-    const saved = config.providers?.[providerId];
-    const catalogModels = await catalogModelsFor(providerId, saved, registry);
-    const catalogById = new Map(catalogModels.map((model) => [model.id, model]));
-    const textCatalogModels = catalogModels.filter(supportsTextCompletion);
-    const candidates = [];
-    if (options.allModels) {
-      pushUnique(candidates, saved?.models ?? []);
-      pushUnique(candidates, Object.keys(saved?.customModels ?? {}));
-      pushUnique(
-        candidates,
-        textCatalogModels.map((model) => model.id)
-      );
-      if (providerId === config.provider) pushUnique(candidates, [config.model]);
-      pushUnique(candidates, [saved?.model]);
-    } else {
-      const representative = (providerId === config.provider ? config.model : void 0) ?? saved?.model ?? saved?.models?.find(Boolean) ?? Object.keys(saved?.customModels ?? {})[0] ?? await registry.suggestModel(providerId).catch(() => void 0) ?? textCatalogModels[0]?.id;
-      pushUnique(candidates, [representative]);
-    }
-    for (const modelId of candidates) {
-      if (modelFilter && !modelFilter.has(modelId)) continue;
-      if (!supportsTextCompletion(catalogById.get(modelId))) continue;
-      targets.push({ providerId, modelId });
-    }
-  }
-  return targets;
-}
-function classifyFailure(err) {
-  if (err instanceof ProviderError || ProviderError.isProviderError(err)) {
-    const providerError = err;
-    return {
-      error: providerError.describe(),
-      errorKind: providerError.kind,
-      httpStatus: providerError.status
-    };
-  }
-  const isTimeout = err instanceof Error && (err.name === "TimeoutError" || err.name === "AbortError") || /timed?\s*out/i.test(toErrorMessage(err));
-  return {
-    error: toErrorMessage(err),
-    errorKind: isTimeout ? "timeout" : "unknown"
-  };
-}
-async function runModelSmokeTests(params) {
-  const results = [];
-  let providerId;
-  let provider;
-  let providerError;
-  for (let index = 0; index < params.targets.length; index++) {
-    const target = params.targets[index];
-    params.onTargetStart?.(target, index, params.targets.length);
-    const started = Date.now();
-    if (target.providerId !== providerId) {
-      providerId = target.providerId;
-      provider = void 0;
-      providerError = void 0;
-      try {
-        provider = await params.createProvider(target.providerId);
-      } catch (err) {
-        providerError = err;
-      }
-    }
-    if (!provider) {
-      const failure = classifyFailure(providerError ?? new Error("Provider could not be created."));
-      const result = {
-        providerId: target.providerId,
-        modelId: target.modelId,
-        status: "failed",
-        latencyMs: Date.now() - started,
-        ...failure
-      };
-      results.push(result);
-      params.onTargetComplete?.(result, index, params.targets.length);
-      continue;
-    }
-    try {
-      const response = await provider.complete(
-        {
-          model: target.modelId,
-          messages: [
-            {
-              role: "user",
-              content: [{ type: "text", text: "Reply with exactly: OK" }]
-            }
-          ],
-          maxTokens: params.options.maxTokens
-        },
-        { signal: AbortSignal.timeout(params.options.timeoutMs) }
-      );
-      const result = {
-        providerId: target.providerId,
-        modelId: target.modelId,
-        status: "passed",
-        latencyMs: Date.now() - started,
-        stopReason: response.stopReason,
-        usage: response.usage
-      };
-      results.push(result);
-      params.onTargetComplete?.(result, index, params.targets.length);
-    } catch (err) {
-      const result = {
-        providerId: target.providerId,
-        modelId: target.modelId,
-        status: "failed",
-        latencyMs: Date.now() - started,
-        ...classifyFailure(err)
-      };
-      results.push(result);
-      params.onTargetComplete?.(result, index, params.targets.length);
-    }
-  }
-  return results;
-}
+// src/subcommands/handlers/modeldiag-bench.ts
+import { color as color3, toErrorMessage } from "@wrongstack/core/utils";
+
+// src/subcommands/handlers/modeldiag-eval.ts
+import { makeProviderFromConfig } from "@wrongstack/providers";
+import { color as color2 } from "@wrongstack/core/utils";
 
 // src/subcommands/handlers/modeldiag-profiles.ts
 import { color } from "@wrongstack/core/utils";
@@ -447,7 +265,7 @@ function roleCat(role) {
   return ROLE_CATEGORY[role] ?? "general";
 }
 
-// src/subcommands/handlers/modeldiag.ts
+// src/subcommands/handlers/modeldiag-eval.ts
 function createProviderForId(providerId, cfg) {
   const savedCfg = cfg.providers?.[providerId];
   const cfgWithType = Object.assign(
@@ -511,79 +329,836 @@ Output ONLY a ranked list, one per line:
     return responses.map((_, i) => i);
   }
 }
-async function readProviders(cachePath) {
-  if (!cachePath) {
-    return `${color2.red("Models cache not available")}.`;
-  }
-  try {
-    const raw = await fs.readFile(cachePath, "utf8");
-    const parsed = JSON.parse(raw);
-    const payload = parsed.payload ?? parsed;
-    return Object.entries(payload).map(([id, p]) => ({
-      id: p.id ?? id,
-      name: p.name ?? id,
-      family: p.npm ?? id,
-      models: Object.values(p.models ?? {}).map(
-        (m) => ({
-          id: m.id,
-          name: m.name,
-          capabilities: {
-            contextWindow: m.limit?.context,
-            maxOutputTokens: m.limit?.output
-          },
-          pricing: m.cost
-        })
-      )
-    }));
-  } catch {
-    return `${color2.amber("Models cache not available")}. Run wstack sync-models.`;
-  }
-}
-function checkHasKey(pid, config) {
-  if (pid === config.provider && config.provider) return true;
-  const pc = config.providers?.[pid];
-  if (!pc) return false;
-  if (typeof pc.apiKey === "string" && pc.apiKey.length > 0) return true;
-  if (Array.isArray(pc.apiKeys) && pc.apiKeys.some((k) => k?.apiKey)) return true;
-  return false;
-}
-var modeldiagCmd = async (args, deps) => {
-  const sub = args[0]?.toLowerCase() || "full";
-  const cacheResult = await readProviders(deps.paths.modelsCache);
-  if (typeof cacheResult === "string") {
-    deps.renderer.write(`${cacheResult}
-`);
-    return cacheResult.includes(color2.red("")) ? 1 : 0;
-  }
-  const providers = cacheResult;
-  const config = deps.config;
-  const modelMatrix = config.modelMatrix ?? {};
-  function hasKey(pid) {
-    return checkHasKey(pid, config);
-  }
-  function writeLine(line = "") {
+async function runModeldiagEval(args, deps, providers, config, hasKey) {
+  const writeLine = (line = "") => {
     deps.renderer.write(`${line}
 `);
+  };
+  const evalArgs = args.slice(1);
+  const providersEq = evalArgs.find((a) => a.startsWith("--providers="));
+  const providersValue = providersEq?.replace("--providers=", "") ?? (typeof deps.flags?.["providers"] === "string" ? deps.flags["providers"] : typeof deps.flags?.["provider"] === "string" ? deps.flags["provider"] : void 0);
+  const providerFilter = providersValue ? providersValue.split(",").map((s) => s.trim()).filter(Boolean) : void 0;
+  const maxEq = evalArgs.find((a) => a.startsWith("--max="));
+  const maxValue = maxEq?.replace("--max=", "") ?? (typeof deps.flags?.["max"] === "string" ? deps.flags["max"] : void 0);
+  const maxModels = maxValue ? Math.max(1, parseInt(maxValue, 10) || 2) : 2;
+  const quick = evalArgs.includes("--quick") || deps.flags?.["quick"] === true || deps.flags?.["quick"] === "true";
+  const modelsPerCat = quick ? 1 : maxModels;
+  const roleFilter = evalArgs.find((a) => !a.startsWith("--"));
+  const targetCategories = roleFilter ? EVAL_CATEGORIES.includes(roleCat(roleFilter)) ? [roleCat(roleFilter)] : [] : EVAL_CATEGORIES;
+  if (targetCategories.length === 0 && roleFilter) {
+    writeLine(
+      `${color2.amber("Unknown role/category")}: "${roleFilter}". Try: ${EVAL_CATEGORIES.join(", ")}`
+    );
+    return 1;
   }
-  if (sub === "keys") {
-    writeLine(`${color2.bold("API Key Status")}`);
-    writeLine();
-    for (const prov of providers) {
-      const k = hasKey(prov.id);
-      writeLine(`  ${checkMark(k)} ${color2.bold(prov.id.padEnd(18))} ${color2.dim(prov.name)}`);
+  const keyedProviderIds = providers.filter((p) => hasKey(p.id)).map((p) => p.id);
+  let targetProviderIds;
+  if (providerFilter && providerFilter.length > 0) {
+    const unknown = providerFilter.filter((pid) => !keyedProviderIds.includes(pid));
+    targetProviderIds = providerFilter.filter((pid) => keyedProviderIds.includes(pid));
+    if (targetProviderIds.length === 0) {
+      const noKeyMsg = unknown.length > 0 ? `None of the specified providers (${unknown.join(", ")}) have API keys. Add keys with wstack auth.` : "None of the specified providers have API keys configured.";
+      writeLine(`${color2.amber(noKeyMsg)}`);
+      return 0;
     }
-    writeLine();
-    writeLine(`${color2.dim(`Leader: ${config.provider}/${config.model}`)}`);
+  } else if (keyedProviderIds.length === 0) {
+    writeLine(`${color2.amber("No providers have API keys. Add keys with wstack auth.")}`);
     return 0;
-  }
-  if (sub === "caps") {
-    writeLine(`${color2.bold("Model Capabilities")} ${color2.dim("\u2014 matched to known profiles")}`);
-    writeLine();
-    for (const prov of providers) {
-      if (!hasKey(prov.id)) continue;
-      writeLine(`  ${color2.bold(prov.id)} ${color2.dim(`(${prov.name})`)}`);
-      const tiers = {
-        premium: [],
+  } else if (keyedProviderIds.length === 1) {
+    targetProviderIds = keyedProviderIds;
+  } else {
+    const providerList = keyedProviderIds.map((pid, i) => {
+      const info = providers.find((p) => p.id === pid);
+      return `  ${color2.cyan(String(i + 1))}) ${color2.bold(pid.padEnd(16))} ${color2.dim(info?.name ?? "")}`;
+    }).join("\n");
+    deps.renderer.write(
+      `
+${color2.bold("Select providers to evaluate")}
+
+${providerList}
+
+${color2.dim('Enter numbers or provider IDs (comma-separated, or "all"):')}
+`
+    );
+    const input = await deps.reader.readLine("  > ");
+    const selected = input.trim().toLowerCase();
+    if (selected === "" || selected === "all") {
+      targetProviderIds = keyedProviderIds;
+    } else {
+      targetProviderIds = [];
+      for (const part of selected.split(",").map((s) => s.trim())) {
+        const idx = parseInt(part, 10);
+        if (idx >= 1 && idx <= keyedProviderIds.length) {
+          const pid = keyedProviderIds[idx - 1];
+          if (!targetProviderIds.includes(pid)) targetProviderIds.push(pid);
+        } else if (keyedProviderIds.includes(part)) {
+          if (!targetProviderIds.includes(part)) targetProviderIds.push(part);
+        }
+      }
+    }
+    if (targetProviderIds.length === 0) {
+      writeLine(color2.dim("No providers selected."));
+      return 0;
+    }
+  }
+  const leaderModel = config.model ?? "unknown";
+  const unknownProviders = providerFilter ? providerFilter.filter((pid) => !keyedProviderIds.includes(pid)) : [];
+  const warningLine = unknownProviders.length > 0 ? `  ${color2.amber("\u26A0 skipped (no key):")} ${unknownProviders.join(", ")}
+` : "";
+  writeLine(`${color2.bold("Model Competency Evaluation")}`);
+  writeLine(
+    color2.dim(
+      `Providers: ${targetProviderIds.join(", ")}  |  ${targetCategories.length} cats  |  ${modelsPerCat} model(s)/cat/provider`
+    )
+  );
+  writeLine(warningLine);
+  writeLine(color2.dim(`Leader (ranker): ${config.provider}/${leaderModel}`));
+  writeLine();
+  const collected = /* @__PURE__ */ new Map();
+  let total = 0;
+  let ok = 0;
+  for (const pid of targetProviderIds) {
+    const prov = createProviderForId(pid, config);
+    if (!prov) {
+      writeLine(color2.dim(`  \u2298 ${pid}: provider unavailable, skipping`));
+      continue;
+    }
+    for (const cat of targetCategories) {
+      const task = EVAL_TASKS[cat];
+      if (!task) continue;
+      const candidates = rankModels(
+        providers,
+        (providerId) => providerId === pid && hasKey(providerId),
+        cat,
+        modelsPerCat
+      );
+      if (candidates.length === 0) continue;
+      if (!collected.has(cat)) collected.set(cat, /* @__PURE__ */ new Map());
+      for (const c of candidates) {
+        total++;
+        const modelKey = `${pid}/${c.model}`;
+        try {
+          const start = Date.now();
+          const resp = await prov.complete(
+            {
+              model: c.model,
+              system: [{ type: "text", text: "Be thorough and correct." }],
+              messages: [
+                { role: "user", content: [{ type: "text", text: task.prompt }] }
+              ],
+              maxTokens: 1024
+            },
+            { signal: AbortSignal.timeout(45e3) }
+          );
+          const respText = resp.content[0] && "text" in resp.content[0] ? resp.content[0].text : "";
+          const respUsage = resp.usage;
+          collected.get(cat)?.set(modelKey, {
+            model: modelKey,
+            latency: Date.now() - start,
+            tokens: (respUsage?.input ?? 0) + (respUsage?.output ?? 0),
+            text: respText
+          });
+          ok++;
+        } catch {
+          collected.get(cat)?.set(modelKey, {
+            model: modelKey,
+            latency: -1,
+            tokens: 0,
+            text: ""
+          });
+        }
+      }
+    }
+  }
+  writeLine(`${color2.dim(`Phase 1: ${ok}/${total} calls succeeded`)}`);
+  writeLine();
+  if (collected.size === 0) {
+    writeLine(color2.amber("No responses collected. Check provider configuration."));
+    return 0;
+  }
+  const leaderProvider = createProviderForId(config.provider, config);
+  if (leaderProvider) {
+    writeLine(`${color2.bold("Phase 2")} \u2014 ${color2.dim("leader ranks responses")}`);
+    writeLine();
+  }
+  const rankings = /* @__PURE__ */ new Map();
+  for (const [cat, responses] of collected) {
+    const valid = Array.from(responses.values()).filter((r) => r.latency >= 0);
+    if (valid.length < 2) {
+      if (valid.length === 1) {
+        const m = valid[0]?.model;
+        if (m !== void 0) {
+          if (!rankings.has(m)) rankings.set(m, /* @__PURE__ */ new Map());
+          rankings.get(m)?.set(cat, { rank: 1, total: 1 });
+        }
+      }
+      continue;
+    }
+    const task = EVAL_TASKS[cat];
+    if (leaderProvider) {
+      const ranked = await rankResponses(leaderProvider, leaderModel, task.prompt, valid);
+      for (let i = 0; i < valid.length; i++) {
+        const m = valid[ranked[i] ?? i]?.model;
+        if (m === void 0) continue;
+        if (!rankings.has(m)) rankings.set(m, /* @__PURE__ */ new Map());
+        rankings.get(m)?.set(cat, { rank: i + 1, total: valid.length });
+      }
+    } else {
+      for (const r of valid) {
+        if (!rankings.has(r.model)) rankings.set(r.model, /* @__PURE__ */ new Map());
+        rankings.get(r.model)?.set(cat, { rank: 1, total: valid.length });
+      }
+    }
+  }
+  writeLine(`${color2.bold("Competency Report")}`);
+  writeLine();
+  const allModels = [.../* @__PURE__ */ new Set([...rankings.keys()])].sort();
+  const catList = [...collected.keys()];
+  const modelColWidth = Math.max(24, ...allModels.map((m) => m.length)) + 2;
+  const cw = 12;
+  writeLine(
+    `  ${color2.dim("model".padEnd(modelColWidth))}` + catList.map((c) => color2.dim((EVAL_TASKS[c]?.label ?? c).slice(0, cw).padEnd(cw + 2))).join("")
+  );
+  writeLine(`  ${color2.dim("\u2500".repeat(modelColWidth + catList.length * (cw + 2)))}`);
+  for (const model of allModels) {
+    const mr = rankings.get(model);
+    const provFromModel = model.split("/")[0] ?? "";
+    const modelColor = provFromModel === config.provider ? color2.cyan : color2.green;
+    let row = `  ${modelColor(model.padEnd(modelColWidth))}`;
+    for (const cat of catList) {
+      const e = mr.get(cat);
+      if (e) {
+        const pct = Math.round((1 - (e.rank - 1) / Math.max(1, e.total - 1)) * 100);
+        const pc = pct >= 80 ? color2.green : pct >= 50 ? color2.amber : color2.red;
+        row += `${pc(`#${e.rank} ${pct}%`.padEnd(cw + 2))}`;
+      } else {
+        row += color2.dim("\u2014".padEnd(cw + 2));
+      }
+    }
+    writeLine(row);
+  }
+  writeLine();
+  writeLine(color2.dim("#1 100% = best in category. \u2014 = not tested."));
+  writeLine();
+  writeLine(color2.dim("Pin: wstack setmodel set <role> <provider>/<model>"));
+  writeLine(
+    color2.dim(
+      "Full: wstack modeldiag eval          Providers: wstack modeldiag eval --providers=id1,id2"
+    )
+  );
+  writeLine(
+    color2.dim("Max:  wstack modeldiag eval --max=3   Quick: wstack modeldiag eval --quick")
+  );
+  return 0;
+}
+
+// src/subcommands/handlers/modeldiag-bench.ts
+async function runModeldiagBench(args, deps, providers, config, hasKey) {
+  const writeLine = (line = "") => {
+    deps.renderer.write(`${line}
+`);
+  };
+  const benchArgs = args.slice(1);
+  if (benchArgs.length < 2) {
+    writeLine(
+      `${color3.amber("Usage:")} wstack modeldiag bench <role> "<test prompt>" [--providers=p1,p2]`
+    );
+    writeLine();
+    writeLine(
+      color3.dim(
+        'Example: wstack modeldiag bench verify "Write a function that checks if a string is a palindrome"'
+      )
+    );
+    writeLine(
+      color3.dim(
+        "Tests the top 5 candidate models for the role with your prompt and reports results."
+      )
+    );
+    writeLine(color3.dim("Add --providers=anthropic,google to test across multiple providers."));
+    return 0;
+  }
+  const providersEqIdx = benchArgs.findIndex((a) => a.startsWith("--providers="));
+  let providerFilter;
+  if (providersEqIdx >= 0) {
+    const rawFilter = benchArgs[providersEqIdx]?.replace("--providers=", "").split(",");
+    if (rawFilter) {
+      providerFilter = rawFilter.map((s) => s.trim()).filter(Boolean);
+      benchArgs.splice(providersEqIdx, 1);
+    }
+  } else {
+    const rawFilter = deps.flags?.["providers"] ?? deps.flags?.["provider"];
+    if (typeof rawFilter === "string") {
+      providerFilter = rawFilter.split(",").map((s) => s.trim()).filter(Boolean);
+    }
+  }
+  const benchRole = benchArgs[0];
+  if (!benchRole) {
+    writeLine(
+      `${color3.amber("No benchmark role specified")}. Usage: wstack diag bench <role> [prompt]`
+    );
+    return 1;
+  }
+  const benchPrompt = benchArgs.slice(1).join(" ");
+  const cat = roleCat(benchRole);
+  const candidates = rankModels(providers, hasKey, cat, 5);
+  if (candidates.length === 0) {
+    writeLine(
+      `${color3.amber("No candidate models found")} for role "${benchRole}" (category: ${cat}).`
+    );
+    return 0;
+  }
+  let targetCandidates;
+  if (providerFilter && providerFilter.length > 0) {
+    targetCandidates = candidates.filter((c) => providerFilter.includes(c.provider));
+    if (targetCandidates.length === 0) {
+      writeLine(
+        `${color3.amber("No candidates match the specified providers")}: ${providerFilter.join(", ")}`
+      );
+      writeLine(
+        `Candidate providers: ${[...new Set(candidates.map((c) => c.provider))].join(", ")}`
+      );
+      return 0;
+    }
+  } else {
+    targetCandidates = candidates;
+  }
+  writeLine(
+    `${color3.bold("Model Benchmark")} \u2014 ${color3.amber(benchRole)} ${color3.dim(`(category: ${cat})`)}`
+  );
+  writeLine(
+    `${color3.dim("Prompt:")} "${benchPrompt.slice(0, 120)}${benchPrompt.length > 120 ? "\u2026" : ""}"`
+  );
+  writeLine();
+  writeLine(
+    `  ${color3.dim("#  model".padEnd(52))} ${color3.dim("score".padEnd(12))} ${color3.dim("latency".padEnd(10))} ${color3.dim("tokens".padEnd(14))} ${color3.dim("first line")}`
+  );
+  writeLine(`  ${color3.dim("\u2500".repeat(108))}`);
+  const providerInstances = /* @__PURE__ */ new Map();
+  for (const pid of [...new Set(targetCandidates.map((c) => c.provider))]) {
+    const prov = createProviderForId(pid, config);
+    if (prov) providerInstances.set(pid, prov);
+  }
+  let idx = 0;
+  for (const c of targetCandidates.slice(0, 20)) {
+    idx++;
+    const label = `${idx}`.padStart(2);
+    const modelKey = `${c.provider}/${c.model}`;
+    const prov = providerInstances.get(c.provider);
+    if (!prov) {
+      writeLine(
+        `  ${label} ${color3.red(modelKey.padEnd(50))} ${scoreBar(c.score, 110).slice(0, 11)}  ${color3.red("NO PROVIDER")}`
+      );
+      continue;
+    }
+    try {
+      const start = Date.now();
+      const resp = await prov.complete(
+        {
+          model: c.model,
+          messages: [{ role: "user", content: [{ type: "text", text: benchPrompt }] }],
+          maxTokens: 256
+        },
+        { signal: AbortSignal.timeout(3e4) }
+      );
+      const latency = Date.now() - start;
+      const firstText = resp.content[0] && "text" in resp.content[0] ? resp.content[0].text : "";
+      const firstLineClean = firstText.replace(/\n/g, " ").slice(0, 80) || color3.dim("(empty)");
+      const provColor = c.provider === config.provider ? color3.green : color3.cyan;
+      const usage = resp.usage;
+      writeLine(
+        `  ${label} ${provColor(modelKey.padEnd(50))} ${scoreBar(c.score, 110).slice(0, 11)}  ${color3.amber(fmtMs(latency).padEnd(8))} ${color3.dim(`in${usage?.input ?? "?"}/out${usage?.output ?? "?"}`.padEnd(12))} ${firstLineClean}`
+      );
+    } catch (err) {
+      const errMsg = toErrorMessage(err);
+      writeLine(
+        `  ${label} ${color3.red(modelKey.padEnd(50))} ${scoreBar(c.score, 110).slice(0, 11)}  ${color3.red("FAILED")}    ${color3.dim(errMsg.slice(0, 40))}`
+      );
+    }
+  }
+  const testedProviders = [...new Set(targetCandidates.map((c) => c.provider))];
+  writeLine();
+  writeLine(
+    color3.dim(
+      `Tested ${idx} model(s) across ${testedProviders.length} provider(s): ${testedProviders.join(", ")}.`
+    )
+  );
+  writeLine(color3.dim("Pin the best: wstack setmodel set <role> <provider>/<model>"));
+  return 0;
+}
+
+// src/subcommands/handlers/modeldiag-test.ts
+import { color as color4, toErrorMessage as toErrorMessage3 } from "@wrongstack/core/utils";
+import { makeProviderFromConfig as makeProviderFromConfig2, setOAuthTokenPersister } from "@wrongstack/providers";
+
+// src/subcommands/handlers/model-smoke-test.ts
+import { ProviderError } from "@wrongstack/core/types";
+import { toErrorMessage as toErrorMessage2 } from "@wrongstack/core/utils";
+var DEFAULT_TIMEOUT_MS = 45e3;
+var DEFAULT_MAX_TOKENS = 32;
+function csv(value) {
+  if (typeof value !== "string") return void 0;
+  const values = value.split(",").map((part) => part.trim()).filter(Boolean);
+  return values.length > 0 ? values : void 0;
+}
+function positiveInteger(value, fallback, label) {
+  if (value === void 0) return fallback;
+  const parsed = typeof value === "string" ? Number(value) : Number.NaN;
+  if (!Number.isSafeInteger(parsed) || parsed <= 0) {
+    throw new Error(`${label} must be a positive integer.`);
+  }
+  return parsed;
+}
+function parseModelSmokeOptions(args, flags = {}) {
+  const local = /* @__PURE__ */ new Map();
+  for (let i = 0; i < args.length; i++) {
+    const arg = args[i];
+    if (!arg?.startsWith("--")) continue;
+    const eq = arg.indexOf("=");
+    if (eq >= 0) {
+      local.set(arg.slice(2, eq), arg.slice(eq + 1));
+      continue;
+    }
+    const name = arg.slice(2);
+    const next = args[i + 1];
+    if (next && !next.startsWith("--")) {
+      local.set(name, next);
+      i++;
+    } else {
+      local.set(name, true);
+    }
+  }
+  const get = (name) => local.get(name) ?? flags[name];
+  const providerFilter = csv(get("providers") ?? get("provider"));
+  const modelFilter = csv(get("models") ?? get("model"));
+  return {
+    allModels: get("all-models") === true || get("all-models") === "true",
+    planOnly: get("plan") === true || get("plan") === "true",
+    json: get("json") === true || get("json") === "true",
+    timeoutMs: positiveInteger(get("timeout"), DEFAULT_TIMEOUT_MS, "--timeout"),
+    maxTokens: positiveInteger(get("max-tokens"), DEFAULT_MAX_TOKENS, "--max-tokens"),
+    ...providerFilter ? { providerFilter } : {},
+    ...modelFilter ? { modelFilter } : {}
+  };
+}
+function pushUnique(target, values) {
+  const seen = new Set(target);
+  for (const value of values) {
+    const normalized = value?.trim();
+    if (!normalized || seen.has(normalized)) continue;
+    seen.add(normalized);
+    target.push(normalized);
+  }
+}
+async function catalogModelsFor(providerId, providerConfig, registry) {
+  const direct = await registry.getProvider(providerId).catch(() => void 0);
+  const inherited = direct ?? (providerConfig?.type && providerConfig.type !== providerId ? await registry.getProvider(providerConfig.type).catch(() => void 0) : void 0);
+  return inherited?.models ?? [];
+}
+function supportsTextCompletion(model) {
+  const outputs = model?.modalities?.output;
+  return !outputs?.length || outputs.includes("text");
+}
+async function buildModelSmokeTargets(config, registry, options) {
+  const providerIds = Object.keys(config.providers ?? {});
+  if (config.provider && !providerIds.includes(config.provider)) providerIds.push(config.provider);
+  const providerFilter = options.providerFilter ? new Set(options.providerFilter.map((id) => id.toLowerCase())) : void 0;
+  const modelFilter = options.modelFilter ? new Set(options.modelFilter) : void 0;
+  const targets = [];
+  for (const providerId of providerIds) {
+    if (providerFilter && !providerFilter.has(providerId.toLowerCase())) continue;
+    const saved = config.providers?.[providerId];
+    const catalogModels = await catalogModelsFor(providerId, saved, registry);
+    const catalogById = new Map(catalogModels.map((model) => [model.id, model]));
+    const textCatalogModels = catalogModels.filter(supportsTextCompletion);
+    const candidates = [];
+    if (options.allModels) {
+      pushUnique(candidates, saved?.models ?? []);
+      pushUnique(candidates, Object.keys(saved?.customModels ?? {}));
+      pushUnique(
+        candidates,
+        textCatalogModels.map((model) => model.id)
+      );
+      if (providerId === config.provider) pushUnique(candidates, [config.model]);
+      pushUnique(candidates, [saved?.model]);
+    } else {
+      const representative = (providerId === config.provider ? config.model : void 0) ?? saved?.model ?? saved?.models?.find(Boolean) ?? Object.keys(saved?.customModels ?? {})[0] ?? await registry.suggestModel(providerId).catch(() => void 0) ?? textCatalogModels[0]?.id;
+      pushUnique(candidates, [representative]);
+    }
+    for (const modelId of candidates) {
+      if (modelFilter && !modelFilter.has(modelId)) continue;
+      if (!supportsTextCompletion(catalogById.get(modelId))) continue;
+      targets.push({ providerId, modelId });
+    }
+  }
+  return targets;
+}
+function classifyFailure(err) {
+  if (err instanceof ProviderError || ProviderError.isProviderError(err)) {
+    const providerError = err;
+    return {
+      error: providerError.describe(),
+      errorKind: providerError.kind,
+      httpStatus: providerError.status
+    };
+  }
+  const isTimeout = err instanceof Error && (err.name === "TimeoutError" || err.name === "AbortError") || /timed?\s*out/i.test(toErrorMessage2(err));
+  return {
+    error: toErrorMessage2(err),
+    errorKind: isTimeout ? "timeout" : "unknown"
+  };
+}
+async function runModelSmokeTests(params) {
+  const results = [];
+  let providerId;
+  let provider;
+  let providerError;
+  for (let index = 0; index < params.targets.length; index++) {
+    const target = params.targets[index];
+    params.onTargetStart?.(target, index, params.targets.length);
+    const started = Date.now();
+    if (target.providerId !== providerId) {
+      providerId = target.providerId;
+      provider = void 0;
+      providerError = void 0;
+      try {
+        provider = await params.createProvider(target.providerId);
+      } catch (err) {
+        providerError = err;
+      }
+    }
+    if (!provider) {
+      const failure = classifyFailure(providerError ?? new Error("Provider could not be created."));
+      const result = {
+        providerId: target.providerId,
+        modelId: target.modelId,
+        status: "failed",
+        latencyMs: Date.now() - started,
+        ...failure
+      };
+      results.push(result);
+      params.onTargetComplete?.(result, index, params.targets.length);
+      continue;
+    }
+    try {
+      const response = await provider.complete(
+        {
+          model: target.modelId,
+          messages: [
+            {
+              role: "user",
+              content: [{ type: "text", text: "Reply with exactly: OK" }]
+            }
+          ],
+          maxTokens: params.options.maxTokens
+        },
+        { signal: AbortSignal.timeout(params.options.timeoutMs) }
+      );
+      const result = {
+        providerId: target.providerId,
+        modelId: target.modelId,
+        status: "passed",
+        latencyMs: Date.now() - started,
+        stopReason: response.stopReason,
+        usage: response.usage
+      };
+      results.push(result);
+      params.onTargetComplete?.(result, index, params.targets.length);
+    } catch (err) {
+      const result = {
+        providerId: target.providerId,
+        modelId: target.modelId,
+        status: "failed",
+        latencyMs: Date.now() - started,
+        ...classifyFailure(err)
+      };
+      results.push(result);
+      params.onTargetComplete?.(result, index, params.targets.length);
+    }
+  }
+  return results;
+}
+
+// src/subcommands/handlers/modeldiag-test.ts
+async function runModeldiagTest(args, deps, config) {
+  const writeLine = (line = "") => {
+    deps.renderer.write(`${line}
+`);
+  };
+  if (args.slice(1).includes("--help") || deps.flags?.["help"] === true) {
+    writeLine(`${color4.bold("Usage:")} wstack modeldiag test [options]`);
+    writeLine();
+    writeLine("  (no flags)              Probe one representative model per configured provider");
+    writeLine("  --all-models             Probe every catalog-visible model sequentially");
+    writeLine("  --plan                   Show targets without making model calls");
+    writeLine("  --provider=<id,...>      Restrict provider ids");
+    writeLine("  --model=<id,...>         Restrict model ids");
+    writeLine("  --timeout=<ms>           Per-model timeout (default 45000)");
+    writeLine("  --max-tokens=<n>         Output cap for the tiny probe (default 32)");
+    writeLine("  --json                   Emit machine-readable output");
+    writeLine("  --yes                    Confirm a run larger than 50 requests");
+    return 0;
+  }
+  let smokeOptions;
+  try {
+    smokeOptions = parseModelSmokeOptions(args.slice(1), deps.flags);
+  } catch (err) {
+    writeLine(`${color4.red("Invalid smoke-test options:")} ${toErrorMessage3(err)}`);
+    return 2;
+  }
+  if ("mergeOverlay" in deps.modelsRegistry) {
+    await discoverAndMergeProviders({
+      config,
+      registry: deps.modelsRegistry,
+      cacheDir: deps.paths.cacheDir
+    });
+  }
+  const targets = await buildModelSmokeTargets(
+    config,
+    deps.modelsRegistry,
+    smokeOptions
+  );
+  const byProvider = /* @__PURE__ */ new Map();
+  for (const target of targets) {
+    byProvider.set(target.providerId, (byProvider.get(target.providerId) ?? 0) + 1);
+  }
+  if (smokeOptions.planOnly) {
+    const plan = {
+      mode: smokeOptions.allModels ? "all-models" : "representative",
+      targetCount: targets.length,
+      providers: Object.fromEntries(byProvider),
+      targets
+    };
+    if (smokeOptions.json) {
+      writeLine(JSON.stringify(plan, null, 2));
+    } else {
+      writeLine(`${color4.bold("Provider/Model Smoke Test Plan")}`);
+      writeLine();
+      for (const [providerId, count] of byProvider) {
+        writeLine(`  ${color4.cyan(providerId.padEnd(24))} ${count} model(s)`);
+      }
+      writeLine();
+      writeLine(
+        color4.dim(
+          `${targets.length} request(s) would run sequentially. Remove --plan to execute.`
+        )
+      );
+    }
+    return targets.length > 0 ? 0 : 2;
+  }
+  if (targets.length === 0) {
+    const message = "No provider/model targets matched. Check the active profile or remove filters.";
+    if (smokeOptions.json) {
+      writeLine(
+        JSON.stringify({ results: [], summary: { passed: 0, failed: 0 }, error: message })
+      );
+    } else {
+      writeLine(color4.amber(message));
+    }
+    return 2;
+  }
+  const confirmed = deps.flags?.["yes"] === true || deps.flags?.["yes"] === "true" || args.slice(1).includes("--yes");
+  if (targets.length > 50 && !confirmed) {
+    const message = `${targets.length} live requests are planned. Review with --plan, then add --yes to run them.`;
+    if (smokeOptions.json) {
+      writeLine(JSON.stringify({ error: message, targetCount: targets.length }));
+    } else {
+      writeLine(color4.amber(message));
+    }
+    return 2;
+  }
+  if (!smokeOptions.json) {
+    writeLine(`${color4.bold("Live Provider/Model Smoke Test")}`);
+    writeLine(
+      color4.dim(
+        `${targets.length} request(s), sequential, ${smokeOptions.timeoutMs}ms timeout, max ${smokeOptions.maxTokens} output tokens.`
+      )
+    );
+    writeLine();
+  }
+  const resultLines = (result, index, total) => {
+    if (smokeOptions.json) return;
+    const label = `[${String(index + 1).padStart(String(total).length)}/${total}]`;
+    const target = `${result.providerId}/${result.modelId}`;
+    if (result.status === "passed") {
+      writeLine(
+        `  ${color4.green("\u2713")} ${color4.dim(label)} ${color4.cyan(target)} ${color4.dim(
+          `${fmtMs(result.latencyMs)} \xB7 in${result.usage.input}/out${result.usage.output} \xB7 ${result.stopReason}`
+        )}`
+      );
+    } else {
+      const status = result.httpStatus ? `HTTP ${result.httpStatus} \xB7 ` : "";
+      writeLine(
+        `  ${color4.red("\u2717")} ${color4.dim(label)} ${color4.red(target)} ${color4.dim(
+          `${fmtMs(result.latencyMs)} \xB7 ${status}${result.errorKind}`
+        )}`
+      );
+      writeLine(`      ${color4.red(result.error.slice(0, 240))}`);
+    }
+  };
+  let oauthWriteChain = Promise.resolve();
+  let oauthWriteScheduled = false;
+  const profilePath = activeProfileConfigPath(deps.paths, config);
+  setOAuthTokenPersister((providerId, credentials) => {
+    oauthWriteScheduled = true;
+    oauthWriteChain = oauthWriteChain.then(
+      () => mutateConfigProviders(profilePath, deps.vault, (all) => {
+        const providerConfig = all[providerId];
+        if (!providerConfig) return;
+        const keys = normalizeKeys(providerConfig);
+        const active = providerConfig.activeKey ? keys.find((key) => key.label === providerConfig.activeKey) : keys[0];
+        if (!active) return;
+        active.apiKey = credentials.accessToken;
+        active.refreshToken = credentials.refreshToken;
+        active.expiresAt = new Date(credentials.expiresAt).toISOString();
+        if (credentials.accountId) active.accountId = credentials.accountId;
+        writeKeysBack(providerConfig, keys);
+      })
+    );
+  });
+  let results;
+  try {
+    results = await runModelSmokeTests({
+      targets,
+      options: smokeOptions,
+      createProvider: async (providerId) => {
+        const saved = config.providers?.[providerId];
+        const resolved = await deps.modelsRegistry.getProvider(providerId).catch(() => void 0) ?? (saved?.type && saved.type !== providerId ? await deps.modelsRegistry.getProvider(saved.type).catch(() => void 0) : void 0);
+        const providerConfig = {
+          ...providerId === config.provider ? {
+            ...config.apiKey ? { apiKey: config.apiKey } : {},
+            ...config.baseUrl ? { baseUrl: config.baseUrl } : {}
+          } : {},
+          ...saved,
+          type: providerId,
+          ...saved?.family ?? resolved?.family ? { family: saved?.family ?? resolved?.family } : {},
+          ...saved?.baseUrl ?? resolved?.apiBase ? { baseUrl: saved?.baseUrl ?? resolved?.apiBase } : {},
+          ...saved?.envVars ?? resolved?.envVars ? { envVars: saved?.envVars ?? resolved?.envVars } : {}
+        };
+        return makeProviderFromConfig2(providerId, providerConfig);
+      },
+      onTargetComplete: resultLines
+    });
+  } finally {
+    setOAuthTokenPersister(void 0);
+  }
+  let oauthPersistenceError;
+  if (oauthWriteScheduled) {
+    try {
+      await oauthWriteChain;
+    } catch (err) {
+      oauthPersistenceError = toErrorMessage3(err);
+    }
+  }
+  const passed = results.filter((result) => result.status === "passed").length;
+  const failed = results.length - passed;
+  const summary = { passed, failed, total: results.length };
+  if (smokeOptions.json) {
+    writeLine(
+      JSON.stringify(
+        {
+          results,
+          summary,
+          ...oauthPersistenceError ? {
+            warning: `OAuth token refresh worked in-memory but could not be saved: ${oauthPersistenceError}`
+          } : {}
+        },
+        null,
+        2
+      )
+    );
+  } else {
+    writeLine();
+    writeLine(
+      `${color4.bold("Summary:")} ${color4.green(`${passed} passed`)} \xB7 ${failed > 0 ? color4.red(`${failed} failed`) : color4.dim("0 failed")}`
+    );
+    if (!smokeOptions.allModels) {
+      writeLine(color4.dim("Use --all-models to test every catalog-visible model sequentially."));
+    }
+    if (oauthPersistenceError) {
+      writeLine(
+        color4.amber(
+          `OAuth token refresh worked in-memory but could not be saved: ${oauthPersistenceError}`
+        )
+      );
+    }
+  }
+  return failed === 0 ? 0 : 1;
+}
+
+// src/subcommands/handlers/modeldiag.ts
+async function readProviders(cachePath) {
+  if (!cachePath) {
+    return `${color5.red("Models cache not available")}.`;
+  }
+  try {
+    const raw = await fs.readFile(cachePath, "utf8");
+    const parsed = JSON.parse(raw);
+    const payload = parsed.payload ?? parsed;
+    return Object.entries(payload).map(([id, p]) => ({
+      id: p.id ?? id,
+      name: p.name ?? id,
+      family: p.npm ?? id,
+      models: Object.values(p.models ?? {}).map(
+        (m) => ({
+          id: m.id,
+          name: m.name,
+          capabilities: {
+            contextWindow: m.limit?.context,
+            maxOutputTokens: m.limit?.output
+          },
+          pricing: m.cost
+        })
+      )
+    }));
+  } catch {
+    return `${color5.amber("Models cache not available")}. Run wstack sync-models.`;
+  }
+}
+function checkHasKey(pid, config) {
+  if (pid === config.provider && config.provider) return true;
+  const pc = config.providers?.[pid];
+  if (!pc) return false;
+  if (typeof pc.apiKey === "string" && pc.apiKey.length > 0) return true;
+  if (Array.isArray(pc.apiKeys) && pc.apiKeys.some((k) => k?.apiKey)) return true;
+  return false;
+}
+var modeldiagCmd = async (args, deps) => {
+  const sub = args[0]?.toLowerCase() || "full";
+  const cacheResult = await readProviders(deps.paths.modelsCache);
+  if (typeof cacheResult === "string") {
+    deps.renderer.write(`${cacheResult}
+`);
+    return cacheResult.includes(color5.red("")) ? 1 : 0;
+  }
+  const providers = cacheResult;
+  const config = deps.config;
+  const modelMatrix = config.modelMatrix ?? {};
+  function hasKey(pid) {
+    return checkHasKey(pid, config);
+  }
+  function writeLine(line = "") {
+    deps.renderer.write(`${line}
+`);
+  }
+  if (sub === "keys") {
+    writeLine(`${color5.bold("API Key Status")}`);
+    writeLine();
+    for (const prov of providers) {
+      const k = hasKey(prov.id);
+      writeLine(`  ${checkMark(k)} ${color5.bold(prov.id.padEnd(18))} ${color5.dim(prov.name)}`);
+    }
+    writeLine();
+    writeLine(`${color5.dim(`Leader: ${config.provider}/${config.model}`)}`);
+    return 0;
+  }
+  if (sub === "caps") {
+    writeLine(`${color5.bold("Model Capabilities")} ${color5.dim("\u2014 matched to known profiles")}`);
+    writeLine();
+    for (const prov of providers) {
+      if (!hasKey(prov.id)) continue;
+      writeLine(`  ${color5.bold(prov.id)} ${color5.dim(`(${prov.name})`)}`);
+      const tiers = {
+        premium: [],
         standard: [],
         budget: [],
         unknown: []
@@ -595,24 +1170,24 @@ var modeldiagCmd = async (args, deps) => {
       for (const tier of ["premium", "standard", "budget", "unknown"]) {
         const tierModels = tiers[tier];
         if (tierModels.length === 0) continue;
-        const label = tier === "unknown" ? color2.dim("unmatched") : `${costLabel(tier)} ${tier}`;
+        const label = tier === "unknown" ? color5.dim("unmatched") : `${costLabel(tier)} ${tier}`;
         writeLine(`    ${label}`);
         for (const m of tierModels) {
           const cap = m.capabilities;
           const ctx = cap?.contextWindow ?? 0;
           const maxOut = cap?.maxOutputTokens ?? 0;
           const profile = findProfile(prov.id, m.id);
-          const family = profile ? `${speedLabel(profile.speedTier)} ${color2.green(profile.family)}` : color2.dim("no profile match");
-          const pricing = m.pricing ? `${color2.dim("in")}${fmtPrice(m.pricing.input)} ${color2.dim("out")}${fmtPrice(m.pricing.output)}` : color2.dim("pricing ?");
+          const family = profile ? `${speedLabel(profile.speedTier)} ${color5.green(profile.family)}` : color5.dim("no profile match");
+          const pricing = m.pricing ? `${color5.dim("in")}${fmtPrice(m.pricing.input)} ${color5.dim("out")}${fmtPrice(m.pricing.output)}` : color5.dim("pricing ?");
           writeLine(
-            `      ${color2.cyan(m.id.padEnd(34))}${ctx > 0 ? `ctx ${fmtTokens(ctx).padEnd(6)}` : color2.dim("ctx ?  ")}${maxOut > 0 ? `out ${fmtTokens(maxOut).padEnd(6)}` : "        "}${family}   ${pricing}`
+            `      ${color5.cyan(m.id.padEnd(34))}${ctx > 0 ? `ctx ${fmtTokens(ctx).padEnd(6)}` : color5.dim("ctx ?  ")}${maxOut > 0 ? `out ${fmtTokens(maxOut).padEnd(6)}` : "        "}${family}   ${pricing}`
           );
         }
       }
       writeLine();
     }
     writeLine(
-      color2.dim(
+      color5.dim(
         "Prices in USD per 1M tokens (input/output). ctx = context window, out = max output."
       )
     );
@@ -621,10 +1196,10 @@ var modeldiagCmd = async (args, deps) => {
   async function renderSuggest() {
     writeLine();
     writeLine(
-      `${color2.bold("Agent \u2192 Model Suggestions")} ${color2.amber("(heuristic \u2014 untested)")}`
+      `${color5.bold("Agent \u2192 Model Suggestions")} ${color5.amber("(heuristic \u2014 untested)")}`
     );
     writeLine(
-      color2.dim(
+      color5.dim(
         'These are profile-based best guesses. Test them with wstack modeldiag bench <role> "<prompt>".'
       )
     );
@@ -632,7 +1207,7 @@ var modeldiagCmd = async (args, deps) => {
     const keyedProviders = providers.filter((p) => hasKey(p.id));
     if (keyedProviders.length === 0) {
       writeLine(
-        `  ${color2.amber("No providers have API keys configured. Add keys with wstack auth.")}`
+        `  ${color5.amber("No providers have API keys configured. Add keys with wstack auth.")}`
       );
     } else {
       const roles = [
@@ -654,606 +1229,68 @@ var modeldiagCmd = async (args, deps) => {
           const entry = modelMatrix[role];
           const p = entry.provider ?? config.provider;
           writeLine(
-            `  ${color2.dim(role.padEnd(20))} \u2192 ${color2.cyan(`${p}/${entry.model}`)}  ${color2.dim("(user-configured)")}`
+            `  ${color5.dim(role.padEnd(20))} \u2192 ${color5.cyan(`${p}/${entry.model}`)}  ${color5.dim("(user-configured)")}`
           );
           continue;
         }
         const cat = roleCat(role);
         const ranked = rankModels(providers, hasKey, cat, 3);
         if (ranked.length === 0) {
-          writeLine(`  ${color2.dim(role.padEnd(20))} \u2192 ${color2.dim("no candidates")}`);
+          writeLine(`  ${color5.dim(role.padEnd(20))} \u2192 ${color5.dim("no candidates")}`);
           continue;
         }
         const best = ranked[0];
-        const family = best.profile ? ` ${color2.dim(`(${best.profile.family})`)}` : "";
+        const family = best.profile ? ` ${color5.dim(`(${best.profile.family})`)}` : "";
         const bar = scoreBar(best.score, 110);
         writeLine(
-          `  ${color2.amber(role.padEnd(20))} \u2192 ${color2.cyan(`${best.provider}/${best.model}`)}${family}`
+          `  ${color5.amber(role.padEnd(20))} \u2192 ${color5.cyan(`${best.provider}/${best.model}`)}${family}`
         );
-        writeLine(`  ${" ".repeat(22)}  ${bar}  ${color2.dim(cat)}`);
+        writeLine(`  ${" ".repeat(22)}  ${bar}  ${color5.dim(cat)}`);
         if (ranked.length > 1 && (ranked[1]?.score ?? Number.NEGATIVE_INFINITY) >= best.score - 15) {
           for (const alt of ranked.slice(1)) {
             const af = alt.profile ? ` (${alt.profile.family})` : "";
             writeLine(
-              `  ${" ".repeat(22)}  ${color2.dim(`${alt.provider}/${alt.model}${af}  score ${alt.score}`)}`
+              `  ${" ".repeat(22)}  ${color5.dim(`${alt.provider}/${alt.model}${af}  score ${alt.score}`)}`
             );
           }
         }
       }
       writeLine();
       writeLine(
-        `  ${color2.bold("leader".padEnd(20))} \u2192 ${color2.cyan(`${config.provider}/${config.model}`)}`
+        `  ${color5.bold("leader".padEnd(20))} \u2192 ${color5.cyan(`${config.provider}/${config.model}`)}`
       );
     }
   }
   if (sub === "suggest") {
     await renderSuggest();
     writeLine();
-    writeLine(color2.dim("Pin a suggestion:  wstack setmodel set <role> <provider>/<model>"));
-    writeLine(color2.dim('Test candidates:   wstack modeldiag bench <role> "<test prompt>"'));
+    writeLine(color5.dim("Pin a suggestion:  wstack setmodel set <role> <provider>/<model>"));
+    writeLine(color5.dim('Test candidates:   wstack modeldiag bench <role> "<test prompt>"'));
     return 0;
   }
   if (sub === "test") {
-    if (args.slice(1).includes("--help") || deps.flags?.["help"] === true) {
-      writeLine(`${color2.bold("Usage:")} wstack modeldiag test [options]`);
-      writeLine();
-      writeLine("  (no flags)              Probe one representative model per configured provider");
-      writeLine("  --all-models             Probe every catalog-visible model sequentially");
-      writeLine("  --plan                   Show targets without making model calls");
-      writeLine("  --provider=<id,...>      Restrict provider ids");
-      writeLine("  --model=<id,...>         Restrict model ids");
-      writeLine("  --timeout=<ms>           Per-model timeout (default 45000)");
-      writeLine("  --max-tokens=<n>         Output cap for the tiny probe (default 32)");
-      writeLine("  --json                   Emit machine-readable output");
-      writeLine("  --yes                    Confirm a run larger than 50 requests");
-      return 0;
-    }
-    let smokeOptions;
-    try {
-      smokeOptions = parseModelSmokeOptions(args.slice(1), deps.flags);
-    } catch (err) {
-      writeLine(`${color2.red("Invalid smoke-test options:")} ${toErrorMessage2(err)}`);
-      return 2;
-    }
-    if ("mergeOverlay" in deps.modelsRegistry) {
-      await discoverAndMergeProviders({
-        config,
-        registry: deps.modelsRegistry,
-        cacheDir: deps.paths.cacheDir
-      });
-    }
-    const targets = await buildModelSmokeTargets(
-      config,
-      deps.modelsRegistry,
-      smokeOptions
-    );
-    const byProvider = /* @__PURE__ */ new Map();
-    for (const target of targets) {
-      byProvider.set(target.providerId, (byProvider.get(target.providerId) ?? 0) + 1);
-    }
-    if (smokeOptions.planOnly) {
-      const plan = {
-        mode: smokeOptions.allModels ? "all-models" : "representative",
-        targetCount: targets.length,
-        providers: Object.fromEntries(byProvider),
-        targets
-      };
-      if (smokeOptions.json) {
-        writeLine(JSON.stringify(plan, null, 2));
-      } else {
-        writeLine(`${color2.bold("Provider/Model Smoke Test Plan")}`);
-        writeLine();
-        for (const [providerId, count] of byProvider) {
-          writeLine(`  ${color2.cyan(providerId.padEnd(24))} ${count} model(s)`);
-        }
-        writeLine();
-        writeLine(
-          color2.dim(
-            `${targets.length} request(s) would run sequentially. Remove --plan to execute.`
-          )
-        );
-      }
-      return targets.length > 0 ? 0 : 2;
-    }
-    if (targets.length === 0) {
-      const message = "No provider/model targets matched. Check the active profile or remove filters.";
-      if (smokeOptions.json) {
-        writeLine(
-          JSON.stringify({ results: [], summary: { passed: 0, failed: 0 }, error: message })
-        );
-      } else {
-        writeLine(color2.amber(message));
-      }
-      return 2;
-    }
-    const confirmed = deps.flags?.["yes"] === true || deps.flags?.["yes"] === "true" || args.slice(1).includes("--yes");
-    if (targets.length > 50 && !confirmed) {
-      const message = `${targets.length} live requests are planned. Review with --plan, then add --yes to run them.`;
-      if (smokeOptions.json) {
-        writeLine(JSON.stringify({ error: message, targetCount: targets.length }));
-      } else {
-        writeLine(color2.amber(message));
-      }
-      return 2;
-    }
-    if (!smokeOptions.json) {
-      writeLine(`${color2.bold("Live Provider/Model Smoke Test")}`);
-      writeLine(
-        color2.dim(
-          `${targets.length} request(s), sequential, ${smokeOptions.timeoutMs}ms timeout, max ${smokeOptions.maxTokens} output tokens.`
-        )
-      );
-      writeLine();
-    }
-    const resultLines = (result, index, total) => {
-      if (smokeOptions.json) return;
-      const label = `[${String(index + 1).padStart(String(total).length)}/${total}]`;
-      const target = `${result.providerId}/${result.modelId}`;
-      if (result.status === "passed") {
-        writeLine(
-          `  ${color2.green("\u2713")} ${color2.dim(label)} ${color2.cyan(target)} ${color2.dim(
-            `${fmtMs(result.latencyMs)} \xB7 in${result.usage.input}/out${result.usage.output} \xB7 ${result.stopReason}`
-          )}`
-        );
-      } else {
-        const status = result.httpStatus ? `HTTP ${result.httpStatus} \xB7 ` : "";
-        writeLine(
-          `  ${color2.red("\u2717")} ${color2.dim(label)} ${color2.red(target)} ${color2.dim(
-            `${fmtMs(result.latencyMs)} \xB7 ${status}${result.errorKind}`
-          )}`
-        );
-        writeLine(`      ${color2.red(result.error.slice(0, 240))}`);
-      }
-    };
-    let oauthWriteChain = Promise.resolve();
-    let oauthWriteScheduled = false;
-    const profilePath = activeProfileConfigPath(deps.paths, config);
-    setOAuthTokenPersister((providerId, credentials) => {
-      oauthWriteScheduled = true;
-      oauthWriteChain = oauthWriteChain.then(
-        () => mutateConfigProviders(profilePath, deps.vault, (all) => {
-          const providerConfig = all[providerId];
-          if (!providerConfig) return;
-          const keys = normalizeKeys(providerConfig);
-          const active = providerConfig.activeKey ? keys.find((key) => key.label === providerConfig.activeKey) : keys[0];
-          if (!active) return;
-          active.apiKey = credentials.accessToken;
-          active.refreshToken = credentials.refreshToken;
-          active.expiresAt = new Date(credentials.expiresAt).toISOString();
-          if (credentials.accountId) active.accountId = credentials.accountId;
-          writeKeysBack(providerConfig, keys);
-        })
-      );
-    });
-    let results;
-    try {
-      results = await runModelSmokeTests({
-        targets,
-        options: smokeOptions,
-        createProvider: async (providerId) => {
-          const saved = config.providers?.[providerId];
-          const resolved = await deps.modelsRegistry.getProvider(providerId).catch(() => void 0) ?? (saved?.type && saved.type !== providerId ? await deps.modelsRegistry.getProvider(saved.type).catch(() => void 0) : void 0);
-          const providerConfig = {
-            ...providerId === config.provider ? {
-              ...config.apiKey ? { apiKey: config.apiKey } : {},
-              ...config.baseUrl ? { baseUrl: config.baseUrl } : {}
-            } : {},
-            ...saved,
-            type: providerId,
-            ...saved?.family ?? resolved?.family ? { family: saved?.family ?? resolved?.family } : {},
-            ...saved?.baseUrl ?? resolved?.apiBase ? { baseUrl: saved?.baseUrl ?? resolved?.apiBase } : {},
-            ...saved?.envVars ?? resolved?.envVars ? { envVars: saved?.envVars ?? resolved?.envVars } : {}
-          };
-          return makeProviderFromConfig(providerId, providerConfig);
-        },
-        onTargetComplete: resultLines
-      });
-    } finally {
-      setOAuthTokenPersister(void 0);
-    }
-    let oauthPersistenceError;
-    if (oauthWriteScheduled) {
-      try {
-        await oauthWriteChain;
-      } catch (err) {
-        oauthPersistenceError = toErrorMessage2(err);
-      }
-    }
-    const passed = results.filter((result) => result.status === "passed").length;
-    const failed = results.length - passed;
-    const summary = { passed, failed, total: results.length };
-    if (smokeOptions.json) {
-      writeLine(
-        JSON.stringify(
-          {
-            results,
-            summary,
-            ...oauthPersistenceError ? {
-              warning: `OAuth token refresh worked in-memory but could not be saved: ${oauthPersistenceError}`
-            } : {}
-          },
-          null,
-          2
-        )
-      );
-    } else {
-      writeLine();
-      writeLine(
-        `${color2.bold("Summary:")} ${color2.green(`${passed} passed`)} \xB7 ${failed > 0 ? color2.red(`${failed} failed`) : color2.dim("0 failed")}`
-      );
-      if (!smokeOptions.allModels) {
-        writeLine(color2.dim("Use --all-models to test every catalog-visible model sequentially."));
-      }
-      if (oauthPersistenceError) {
-        writeLine(
-          color2.amber(
-            `OAuth token refresh worked in-memory but could not be saved: ${oauthPersistenceError}`
-          )
-        );
-      }
-    }
-    return failed === 0 ? 0 : 1;
+    return runModeldiagTest(args, deps, config);
   }
   if (sub === "bench") {
-    const benchArgs = args.slice(1);
-    if (benchArgs.length < 2) {
-      writeLine(
-        `${color2.amber("Usage:")} wstack modeldiag bench <role> "<test prompt>" [--providers=p1,p2]`
-      );
-      writeLine();
-      writeLine(
-        color2.dim(
-          'Example: wstack modeldiag bench verify "Write a function that checks if a string is a palindrome"'
-        )
-      );
-      writeLine(
-        color2.dim(
-          "Tests the top 5 candidate models for the role with your prompt and reports results."
-        )
-      );
-      writeLine(color2.dim("Add --providers=anthropic,google to test across multiple providers."));
-      return 0;
-    }
-    const providersEqIdx = benchArgs.findIndex((a) => a.startsWith("--providers="));
-    let providerFilter;
-    if (providersEqIdx >= 0) {
-      const rawFilter = benchArgs[providersEqIdx]?.replace("--providers=", "").split(",");
-      if (rawFilter) {
-        providerFilter = rawFilter.map((s) => s.trim()).filter(Boolean);
-        benchArgs.splice(providersEqIdx, 1);
-      }
-    } else {
-      const rawFilter = deps.flags?.["providers"] ?? deps.flags?.["provider"];
-      if (typeof rawFilter === "string") {
-        providerFilter = rawFilter.split(",").map((s) => s.trim()).filter(Boolean);
-      }
-    }
-    const benchRole = benchArgs[0];
-    if (!benchRole) {
-      writeLine(
-        `${color2.amber("No benchmark role specified")}. Usage: wstack diag bench <role> [prompt]`
-      );
-      return 1;
-    }
-    const benchPrompt = benchArgs.slice(1).join(" ");
-    const cat = roleCat(benchRole);
-    const candidates = rankModels(providers, hasKey, cat, 5);
-    if (candidates.length === 0) {
-      writeLine(
-        `${color2.amber("No candidate models found")} for role "${benchRole}" (category: ${cat}).`
-      );
-      return 0;
-    }
-    let targetCandidates;
-    if (providerFilter && providerFilter.length > 0) {
-      targetCandidates = candidates.filter((c) => providerFilter.includes(c.provider));
-      if (targetCandidates.length === 0) {
-        writeLine(
-          `${color2.amber("No candidates match the specified providers")}: ${providerFilter.join(", ")}`
-        );
-        writeLine(
-          `Candidate providers: ${[...new Set(candidates.map((c) => c.provider))].join(", ")}`
-        );
-        return 0;
-      }
-    } else {
-      targetCandidates = candidates;
-    }
-    writeLine(
-      `${color2.bold("Model Benchmark")} \u2014 ${color2.amber(benchRole)} ${color2.dim(`(category: ${cat})`)}`
-    );
-    writeLine(
-      `${color2.dim("Prompt:")} "${benchPrompt.slice(0, 120)}${benchPrompt.length > 120 ? "\u2026" : ""}"`
-    );
-    writeLine();
-    writeLine(
-      `  ${color2.dim("#  model".padEnd(52))} ${color2.dim("score".padEnd(12))} ${color2.dim("latency".padEnd(10))} ${color2.dim("tokens".padEnd(14))} ${color2.dim("first line")}`
-    );
-    writeLine(`  ${color2.dim("\u2500".repeat(108))}`);
-    const providerInstances = /* @__PURE__ */ new Map();
-    for (const pid of [...new Set(targetCandidates.map((c) => c.provider))]) {
-      const prov = createProviderForId(pid, config);
-      if (prov) providerInstances.set(pid, prov);
-    }
-    let idx = 0;
-    for (const c of targetCandidates.slice(0, 20)) {
-      idx++;
-      const label = `${idx}`.padStart(2);
-      const modelKey = `${c.provider}/${c.model}`;
-      const prov = providerInstances.get(c.provider);
-      if (!prov) {
-        writeLine(
-          `  ${label} ${color2.red(modelKey.padEnd(50))} ${scoreBar(c.score, 110).slice(0, 11)}  ${color2.red("NO PROVIDER")}`
-        );
-        continue;
-      }
-      try {
-        const start = Date.now();
-        const resp = await prov.complete(
-          {
-            model: c.model,
-            messages: [{ role: "user", content: [{ type: "text", text: benchPrompt }] }],
-            maxTokens: 256
-          },
-          { signal: AbortSignal.timeout(3e4) }
-        );
-        const latency = Date.now() - start;
-        const firstText = resp.content[0] && "text" in resp.content[0] ? resp.content[0].text : "";
-        const firstLineClean = firstText.replace(/\n/g, " ").slice(0, 80) || color2.dim("(empty)");
-        const provColor = c.provider === config.provider ? color2.green : color2.cyan;
-        const usage = resp.usage;
-        writeLine(
-          `  ${label} ${provColor(modelKey.padEnd(50))} ${scoreBar(c.score, 110).slice(0, 11)}  ${color2.amber(fmtMs(latency).padEnd(8))} ${color2.dim(`in${usage?.input ?? "?"}/out${usage?.output ?? "?"}`.padEnd(12))} ${firstLineClean}`
-        );
-      } catch (err) {
-        const errMsg = toErrorMessage2(err);
-        writeLine(
-          `  ${label} ${color2.red(modelKey.padEnd(50))} ${scoreBar(c.score, 110).slice(0, 11)}  ${color2.red("FAILED")}    ${color2.dim(errMsg.slice(0, 40))}`
-        );
-      }
-    }
-    const testedProviders = [...new Set(targetCandidates.map((c) => c.provider))];
-    writeLine();
-    writeLine(
-      color2.dim(
-        `Tested ${idx} model(s) across ${testedProviders.length} provider(s): ${testedProviders.join(", ")}.`
-      )
-    );
-    writeLine(color2.dim("Pin the best: wstack setmodel set <role> <provider>/<model>"));
-    return 0;
+    return runModeldiagBench(args, deps, providers, config, hasKey);
   }
   if (sub === "eval" || sub === "evall") {
-    const evalArgs = args.slice(1);
-    const providersEq = evalArgs.find((a) => a.startsWith("--providers="));
-    const providersValue = providersEq?.replace("--providers=", "") ?? (typeof deps.flags?.["providers"] === "string" ? deps.flags["providers"] : typeof deps.flags?.["provider"] === "string" ? deps.flags["provider"] : void 0);
-    const providerFilter = providersValue ? providersValue.split(",").map((s) => s.trim()).filter(Boolean) : void 0;
-    const maxEq = evalArgs.find((a) => a.startsWith("--max="));
-    const maxValue = maxEq?.replace("--max=", "") ?? (typeof deps.flags?.["max"] === "string" ? deps.flags["max"] : void 0);
-    const maxModels = maxValue ? Math.max(1, parseInt(maxValue, 10) || 2) : 2;
-    const quick = evalArgs.includes("--quick") || deps.flags?.["quick"] === true || deps.flags?.["quick"] === "true";
-    const modelsPerCat = quick ? 1 : maxModels;
-    const roleFilter = evalArgs.find((a) => !a.startsWith("--"));
-    const targetCategories = roleFilter ? EVAL_CATEGORIES.includes(roleCat(roleFilter)) ? [roleCat(roleFilter)] : [] : EVAL_CATEGORIES;
-    if (targetCategories.length === 0 && roleFilter) {
-      writeLine(
-        `${color2.amber("Unknown role/category")}: "${roleFilter}". Try: ${EVAL_CATEGORIES.join(", ")}`
-      );
-      return 1;
-    }
-    const keyedProviderIds = providers.filter((p) => hasKey(p.id)).map((p) => p.id);
-    let targetProviderIds;
-    if (providerFilter && providerFilter.length > 0) {
-      const unknown = providerFilter.filter((pid) => !keyedProviderIds.includes(pid));
-      targetProviderIds = providerFilter.filter((pid) => keyedProviderIds.includes(pid));
-      if (targetProviderIds.length === 0) {
-        const noKeyMsg = unknown.length > 0 ? `None of the specified providers (${unknown.join(", ")}) have API keys. Add keys with wstack auth.` : "None of the specified providers have API keys configured.";
-        writeLine(`${color2.amber(noKeyMsg)}`);
-        return 0;
-      }
-    } else if (keyedProviderIds.length === 0) {
-      writeLine(`${color2.amber("No providers have API keys. Add keys with wstack auth.")}`);
-      return 0;
-    } else if (keyedProviderIds.length === 1) {
-      targetProviderIds = keyedProviderIds;
-    } else {
-      const providerList = keyedProviderIds.map((pid, i) => {
-        const info = providers.find((p) => p.id === pid);
-        return `  ${color2.cyan(String(i + 1))}) ${color2.bold(pid.padEnd(16))} ${color2.dim(info?.name ?? "")}`;
-      }).join("\n");
-      deps.renderer.write(
-        `
-${color2.bold("Select providers to evaluate")}
-
-${providerList}
-
-${color2.dim('Enter numbers or provider IDs (comma-separated, or "all"):')}
-`
-      );
-      const input = await deps.reader.readLine("  > ");
-      const selected = input.trim().toLowerCase();
-      if (selected === "" || selected === "all") {
-        targetProviderIds = keyedProviderIds;
-      } else {
-        targetProviderIds = [];
-        for (const part of selected.split(",").map((s) => s.trim())) {
-          const idx = parseInt(part, 10);
-          if (idx >= 1 && idx <= keyedProviderIds.length) {
-            const pid = keyedProviderIds[idx - 1];
-            if (!targetProviderIds.includes(pid)) targetProviderIds.push(pid);
-          } else if (keyedProviderIds.includes(part)) {
-            if (!targetProviderIds.includes(part)) targetProviderIds.push(part);
-          }
-        }
-      }
-      if (targetProviderIds.length === 0) {
-        writeLine(color2.dim("No providers selected."));
-        return 0;
-      }
-    }
-    const leaderModel = config.model ?? "unknown";
-    const unknownProviders = providerFilter ? providerFilter.filter((pid) => !keyedProviderIds.includes(pid)) : [];
-    const warningLine = unknownProviders.length > 0 ? `  ${color2.amber("\u26A0 skipped (no key):")} ${unknownProviders.join(", ")}
-` : "";
-    writeLine(`${color2.bold("Model Competency Evaluation")}`);
-    writeLine(
-      color2.dim(
-        `Providers: ${targetProviderIds.join(", ")}  |  ${targetCategories.length} cats  |  ${modelsPerCat} model(s)/cat/provider`
-      )
-    );
-    writeLine(warningLine);
-    writeLine(color2.dim(`Leader (ranker): ${config.provider}/${leaderModel}`));
-    writeLine();
-    const collected = /* @__PURE__ */ new Map();
-    let total = 0;
-    let ok = 0;
-    for (const pid of targetProviderIds) {
-      const prov = createProviderForId(pid, config);
-      if (!prov) {
-        writeLine(color2.dim(`  \u2298 ${pid}: provider unavailable, skipping`));
-        continue;
-      }
-      for (const cat of targetCategories) {
-        const task = EVAL_TASKS[cat];
-        if (!task) continue;
-        const candidates = rankModels(
-          providers,
-          (providerId) => providerId === pid && hasKey(providerId),
-          cat,
-          modelsPerCat
-        );
-        if (candidates.length === 0) continue;
-        if (!collected.has(cat)) collected.set(cat, /* @__PURE__ */ new Map());
-        for (const c of candidates) {
-          total++;
-          const modelKey = `${pid}/${c.model}`;
-          try {
-            const start = Date.now();
-            const resp = await prov.complete(
-              {
-                model: c.model,
-                system: [{ type: "text", text: "Be thorough and correct." }],
-                messages: [
-                  { role: "user", content: [{ type: "text", text: task.prompt }] }
-                ],
-                maxTokens: 1024
-              },
-              { signal: AbortSignal.timeout(45e3) }
-            );
-            const respText = resp.content[0] && "text" in resp.content[0] ? resp.content[0].text : "";
-            const respUsage = resp.usage;
-            collected.get(cat)?.set(modelKey, {
-              model: modelKey,
-              latency: Date.now() - start,
-              tokens: (respUsage?.input ?? 0) + (respUsage?.output ?? 0),
-              text: respText
-            });
-            ok++;
-          } catch {
-            collected.get(cat)?.set(modelKey, {
-              model: modelKey,
-              latency: -1,
-              tokens: 0,
-              text: ""
-            });
-          }
-        }
-      }
-    }
-    writeLine(`${color2.dim(`Phase 1: ${ok}/${total} calls succeeded`)}`);
-    writeLine();
-    if (collected.size === 0) {
-      writeLine(color2.amber("No responses collected. Check provider configuration."));
-      return 0;
-    }
-    const leaderProvider = createProviderForId(config.provider, config);
-    if (leaderProvider) {
-      writeLine(`${color2.bold("Phase 2")} \u2014 ${color2.dim("leader ranks responses")}`);
-      writeLine();
-    }
-    const rankings = /* @__PURE__ */ new Map();
-    for (const [cat, responses] of collected) {
-      const valid = Array.from(responses.values()).filter((r) => r.latency >= 0);
-      if (valid.length < 2) {
-        if (valid.length === 1) {
-          const m = valid[0]?.model;
-          if (m !== void 0) {
-            if (!rankings.has(m)) rankings.set(m, /* @__PURE__ */ new Map());
-            rankings.get(m)?.set(cat, { rank: 1, total: 1 });
-          }
-        }
-        continue;
-      }
-      const task = EVAL_TASKS[cat];
-      if (leaderProvider) {
-        const ranked = await rankResponses(leaderProvider, leaderModel, task.prompt, valid);
-        for (let i = 0; i < valid.length; i++) {
-          const m = valid[ranked[i] ?? i]?.model;
-          if (m === void 0) continue;
-          if (!rankings.has(m)) rankings.set(m, /* @__PURE__ */ new Map());
-          rankings.get(m)?.set(cat, { rank: i + 1, total: valid.length });
-        }
-      } else {
-        for (const r of valid) {
-          if (!rankings.has(r.model)) rankings.set(r.model, /* @__PURE__ */ new Map());
-          rankings.get(r.model)?.set(cat, { rank: 1, total: valid.length });
-        }
-      }
-    }
-    writeLine(`${color2.bold("Competency Report")}`);
-    writeLine();
-    const allModels = [.../* @__PURE__ */ new Set([...rankings.keys()])].sort();
-    const catList = [...collected.keys()];
-    const modelColWidth = Math.max(24, ...allModels.map((m) => m.length)) + 2;
-    const cw = 12;
-    writeLine(
-      `  ${color2.dim("model".padEnd(modelColWidth))}` + catList.map((c) => color2.dim((EVAL_TASKS[c]?.label ?? c).slice(0, cw).padEnd(cw + 2))).join("")
-    );
-    writeLine(`  ${color2.dim("\u2500".repeat(modelColWidth + catList.length * (cw + 2)))}`);
-    for (const model of allModels) {
-      const mr = rankings.get(model);
-      const provFromModel = model.split("/")[0] ?? "";
-      const modelColor = provFromModel === config.provider ? color2.cyan : color2.green;
-      let row = `  ${modelColor(model.padEnd(modelColWidth))}`;
-      for (const cat of catList) {
-        const e = mr.get(cat);
-        if (e) {
-          const pct = Math.round((1 - (e.rank - 1) / Math.max(1, e.total - 1)) * 100);
-          const pc = pct >= 80 ? color2.green : pct >= 50 ? color2.amber : color2.red;
-          row += `${pc(`#${e.rank} ${pct}%`.padEnd(cw + 2))}`;
-        } else {
-          row += color2.dim("\u2014".padEnd(cw + 2));
-        }
-      }
-      writeLine(row);
-    }
-    writeLine();
-    writeLine(color2.dim("#1 100% = best in category. \u2014 = not tested."));
-    writeLine();
-    writeLine(color2.dim("Pin: wstack setmodel set <role> <provider>/<model>"));
-    writeLine(
-      color2.dim(
-        "Full: wstack modeldiag eval          Providers: wstack modeldiag eval --providers=id1,id2"
-      )
-    );
-    writeLine(
-      color2.dim("Max:  wstack modeldiag eval --max=3   Quick: wstack modeldiag eval --quick")
-    );
-    return 0;
+    return runModeldiagEval(args, deps, providers, config, hasKey);
   }
-  writeLine(`${color2.bold("API Key Status")}`);
+  writeLine(`${color5.bold("API Key Status")}`);
   writeLine();
   for (const prov of providers) {
     const k = hasKey(prov.id);
-    writeLine(`  ${checkMark(k)} ${color2.bold(prov.id.padEnd(18))} ${color2.dim(prov.name)}`);
+    writeLine(`  ${checkMark(k)} ${color5.bold(prov.id.padEnd(18))} ${color5.dim(prov.name)}`);
   }
   writeLine();
-  writeLine(`${color2.dim(`Leader: ${config.provider}/${config.model}`)}`);
+  writeLine(`${color5.dim(`Leader: ${config.provider}/${config.model}`)}`);
   writeLine();
-  writeLine(`${color2.bold("Model Capabilities")} ${color2.dim("\u2014 matched to known profiles")}`);
+  writeLine(`${color5.bold("Model Capabilities")} ${color5.dim("\u2014 matched to known profiles")}`);
   writeLine();
   for (const prov of providers) {
     if (!hasKey(prov.id)) continue;
-    writeLine(`  ${color2.bold(prov.id)} ${color2.dim(`(${prov.name})`)}`);
+    writeLine(`  ${color5.bold(prov.id)} ${color5.dim(`(${prov.name})`)}`);
     const tiers = {
       premium: [],
       standard: [],
@@ -1267,17 +1304,17 @@ ${color2.dim('Enter numbers or provider IDs (comma-separated, or "all"):')}
     for (const tier of ["premium", "standard", "budget", "unknown"]) {
       const tierModels = tiers[tier];
       if (tierModels.length === 0) continue;
-      const label = tier === "unknown" ? color2.dim("unmatched") : `${costLabel(tier)} ${tier}`;
+      const label = tier === "unknown" ? color5.dim("unmatched") : `${costLabel(tier)} ${tier}`;
       writeLine(`    ${label}`);
       for (const m of tierModels) {
         const cap = m.capabilities;
         const ctx = cap?.contextWindow ?? 0;
         const maxOut = cap?.maxOutputTokens ?? 0;
         const profile = findProfile(prov.id, m.id);
-        const family = profile ? `${speedLabel(profile.speedTier)} ${color2.green(profile.family)}` : color2.dim("no profile match");
-        const pricing = m.pricing ? `${color2.dim("in")}${fmtPrice(m.pricing.input)} ${color2.dim("out")}${fmtPrice(m.pricing.output)}` : color2.dim("pricing ?");
+        const family = profile ? `${speedLabel(profile.speedTier)} ${color5.green(profile.family)}` : color5.dim("no profile match");
+        const pricing = m.pricing ? `${color5.dim("in")}${fmtPrice(m.pricing.input)} ${color5.dim("out")}${fmtPrice(m.pricing.output)}` : color5.dim("pricing ?");
         writeLine(
-          `      ${color2.cyan(m.id.padEnd(34))}${ctx > 0 ? `ctx ${fmtTokens(ctx).padEnd(6)}` : color2.dim("ctx ?  ")}${maxOut > 0 ? `out ${fmtTokens(maxOut).padEnd(6)}` : "        "}${family}   ${pricing}`
+          `      ${color5.cyan(m.id.padEnd(34))}${ctx > 0 ? `ctx ${fmtTokens(ctx).padEnd(6)}` : color5.dim("ctx ?  ")}${maxOut > 0 ? `out ${fmtTokens(maxOut).padEnd(6)}` : "        "}${family}   ${pricing}`
         );
       }
     }
@@ -1285,11 +1322,11 @@ ${color2.dim('Enter numbers or provider IDs (comma-separated, or "all"):')}
   }
   await renderSuggest();
   writeLine();
-  writeLine(color2.dim("Pin a suggestion:  wstack setmodel set <role> <provider>/<model>"));
-  writeLine(color2.dim('Test candidates:   wstack modeldiag bench <role> "<test prompt>"'));
+  writeLine(color5.dim("Pin a suggestion:  wstack setmodel set <role> <provider>/<model>"));
+  writeLine(color5.dim('Test candidates:   wstack modeldiag bench <role> "<test prompt>"'));
   return 0;
 };
 export {
   modeldiagCmd
 };
-//# sourceMappingURL=modeldiag-6VS5KDP5.js.map
+//# sourceMappingURL=modeldiag-ICMPHT5L.js.map