cascade-ai 0.70.0 → 0.72.0

package/dist/index.js

@@ -1862,7 +1862,7 @@ function reconcilePrice(model, live) {
 }
 
 // src/constants.ts
-var CASCADE_VERSION = "0.70.0";
+var CASCADE_VERSION = "0.72.0";
 var CASCADE_CONFIG_DIR = ".cascade";
 var CASCADE_MD_FILE = "CASCADE.md";
 var CASCADE_IGNORE_FILE = ".cascadeignore";
@@ -2269,9 +2269,6 @@ var DEFAULT_THEME = "cascade";
 var OLLAMA_BASE_URL = "http://localhost:11434";
 var LM_STUDIO_BASE_URL = "http://localhost:1234";
 var AZURE_BASE_URL_TEMPLATE = "https://{resource}.openai.azure.com";
-var GITHUB_MODELS_INFERENCE_URL = "https://models.github.ai/inference";
-var GITHUB_MODELS_CATALOG_URL = "https://models.github.ai/catalog/models";
-var GITHUB_MODELS_API_VERSION = "2026-03-10";
 var TOOL_NAMES = {
   SHELL: "shell",
   FILE_READ: "file_read",
@@ -2306,8 +2303,7 @@ var PROVIDER_DISPLAY_NAMES = {
   gemini: "Google Gemini",
   azure: "Azure OpenAI",
   "openai-compatible": "OpenAI-Compatible",
-  ollama: "Ollama (Local)",
-  "github-models": "GitHub Models"
+  ollama: "Ollama (Local)"
 };
 
 // src/utils/cost.ts
@@ -3350,206 +3346,6 @@ async function nodeHttpFetch(input, init = {}, redirectCount = 0, opts = {}) {
   });
 }
 
-// src/providers/github-models.ts
-var GITHUB_MODELS_MAX_OUTPUT_TOKENS = 4e3;
-var GITHUB_MODELS_MAX_INPUT_TOKENS = 8e3;
-var DEFAULT_CONTEXT_WINDOW = 128e3;
-var GITHUB_MODELS_MAX_CATALOG_PAGES = 20;
-var GITHUB_MODELS_CATALOG_ORIGIN = new URL(GITHUB_MODELS_CATALOG_URL).origin;
-function stripModelOwnerPrefix(id) {
-  const slash = (id || "").indexOf("/");
-  return slash === -1 ? id : id.slice(slash + 1);
-}
-function pickString(o, keys) {
-  for (const k of keys) {
-    const v = o[k];
-    if (typeof v === "string" && v.length > 0) return v;
-  }
-  return void 0;
-}
-function pickNumber(o, keys) {
-  for (const k of keys) {
-    const v = k.includes(".") ? o[k.split(".")[0]]?.[k.split(".")[1]] : o[k];
-    if (typeof v === "number" && Number.isFinite(v) && v > 0) return v;
-  }
-  return void 0;
-}
-function looksVisionCapable(o) {
-  for (const k of ["supported_input_modalities", "input_modalities", "modalities", "capabilities"]) {
-    const v = o[k];
-    if (Array.isArray(v) && v.some((x) => typeof x === "string" && /image|vision/i.test(x))) return true;
-  }
-  return false;
-}
-function looksToolCapable(o) {
-  const params = o["supported_parameters"];
-  if (Array.isArray(params) && params.some((x) => typeof x === "string" && /^tools?$|function.?call/i.test(x))) {
-    return true;
-  }
-  for (const k of ["capabilities", "supported_output_modalities"]) {
-    const v = o[k];
-    if (Array.isArray(v) && v.some((x) => typeof x === "string" && /tool|function.?call/i.test(x))) return true;
-  }
-  const caps = o["capabilities"];
-  if (caps && typeof caps === "object" && !Array.isArray(caps)) {
-    const c = caps;
-    if (c["function_calling"] === true || c["tool_calling"] === true || c["tools"] === true) return true;
-  }
-  return false;
-}
-function parseNextLink(linkHeader) {
-  if (!linkHeader) return null;
-  for (const part of linkHeader.split(",")) {
-    const match = part.match(/<([^>]+)>\s*;\s*rel="next"/);
-    if (match) return match[1];
-  }
-  return null;
-}
-var GitHubModelsProvider = class extends OpenAIProvider {
-  constructor(config, model) {
-    super({ ...config, apiKey: config.apiKey ?? "" }, model);
-    this.useMaxCompletionTokens = isReasoningModel(stripModelOwnerPrefix(model.id));
-    this.client = new OpenAI({
-      apiKey: config.apiKey ?? "",
-      baseURL: GITHUB_MODELS_INFERENCE_URL,
-      defaultHeaders: { "X-GitHub-Api-Version": GITHUB_MODELS_API_VERSION, "User-Agent": "Cascade-AI" }
-    });
-  }
-  /**
-   * `OpenAIProvider.generateStream()` builds its request budget as
-   * `options.maxTokens ?? this.model.maxOutputTokens` — an explicit per-call
-   * `maxTokens` (T1Administrator's final compilation asks for 8,000, well
-   * above GitHub's real ~4K hard cap) wins outright and is sent unclamped.
-   * GitHub then rejects the request instead of answering, so a run routed to
-   * this provider for that step fails outright rather than getting a
-   * (perfectly serviceable, just shorter) completion. Clamp here, once, right
-   * before delegating to the inherited implementation, rather than teaching
-   * every caller GitHub's specific cap.
-   */
-  async generateStream(options, onChunk) {
-    const clamped = options.maxTokens !== void 0 && options.maxTokens > GITHUB_MODELS_MAX_OUTPUT_TOKENS ? { ...options, maxTokens: GITHUB_MODELS_MAX_OUTPUT_TOKENS } : options;
-    return super.generateStream(clamped, onChunk);
-  }
-  /**
-   * The catalog is a GitHub REST resource, not an OpenAI `/models` list, and
-   * wants GitHub's own header trio — the plain `Accept: application/json` the
-   * openai-compatible provider sends is not the same request.
-   */
-  catalogHeaders() {
-    return {
-      Accept: "application/vnd.github+json",
-      Authorization: `Bearer ${this.config.apiKey ?? ""}`,
-      "X-GitHub-Api-Version": GITHUB_MODELS_API_VERSION,
-      // GitHub's REST API rejects any request with no User-Agent — a plain 403
-      // regardless of how valid the PAT is. nodeHttpFetch is a thin wrapper
-      // over node:http/https, neither of which sets one the way a browser's
-      // fetch or a full HTTP client library would, so this has to be explicit.
-      "User-Agent": "Cascade-AI"
-    };
-  }
-  async listModels() {
-    const raw = [];
-    let url = GITHUB_MODELS_CATALOG_URL;
-    let pages = 0;
-    const visited = /* @__PURE__ */ new Set();
-    while (url && pages < GITHUB_MODELS_MAX_CATALOG_PAGES) {
-      if (visited.has(url)) break;
-      visited.add(url);
-      const res = await nodeHttpFetch(url, { headers: this.catalogHeaders() }, 0, {
-        allowedRedirectOrigin: GITHUB_MODELS_CATALOG_ORIGIN
-      });
-      if (!res.ok) throw new Error(`GitHub Models catalog ${url} returned HTTP ${res.status}`);
-      const body = await res.json();
-      const pageItems = Array.isArray(body) ? body : Array.isArray(body?.models) ? body.models : Array.isArray(body?.data) ? body.data : [];
-      raw.push(...pageItems);
-      const next = parseNextLink(res.headers.get("link"));
-      if (next) {
-        const resolved = new URL(next, url);
-        if (resolved.origin !== GITHUB_MODELS_CATALOG_ORIGIN) {
-          throw new Error(`Refusing cross-origin GitHub Models pagination URL: ${resolved.origin}`);
-        }
-        url = resolved.href;
-      } else {
-        url = null;
-      }
-      pages++;
-    }
-    const entries = raw.map((m) => {
-      if (typeof m === "string") {
-        return { id: m, name: m, contextWindow: DEFAULT_CONTEXT_WINDOW, vision: false, tools: false };
-      }
-      if (!m || typeof m !== "object") return void 0;
-      const o = m;
-      const id = pickString(o, ["id", "name", "model"]);
-      if (!id) return void 0;
-      return {
-        id,
-        name: pickString(o, ["friendly_name", "display_name", "name"]) ?? id,
-        contextWindow: pickNumber(o, [
-          "context_window",
-          "context_length",
-          "max_input_tokens",
-          "limits.max_input_tokens",
-          "limits.max_context_tokens"
-        ]) ?? DEFAULT_CONTEXT_WINDOW,
-        vision: looksVisionCapable(o),
-        tools: looksToolCapable(o)
-      };
-    }).filter((e) => e !== void 0);
-    if (entries.length === 0) return [];
-    const chat = entries.filter((e) => isChatModel(e.id));
-    return (chat.length ? chat : entries).map((e) => ({
-      id: e.id,
-      // full `owner/model` — what the API is addressed by
-      name: e.name,
-      provider: "github-models",
-      // GitHub's real per-request input cap, not the base model's larger one
-      // (see GITHUB_MODELS_MAX_INPUT_TOKENS above) — the same reasoning
-      // already applied to maxOutputTokens below, for the input side.
-      contextWindow: Math.min(e.contextWindow, GITHUB_MODELS_MAX_INPUT_TOKENS),
-      isVisionCapable: e.vision,
-      // A REAL $0, not a missing number — usage is bundled into the quota of
-      // the plan the PAT belongs to and is never billed per token, so
-      // `pricingUnknown: false` is the honest value (see the flag's contract in
-      // core/router/pricing.ts: 0 alone is ambiguous). Set here rather than in
-      // pricing-data.json because resolvePricing() matches exactly on
-      // provider|model|region with no wildcard: every catalog model would need
-      // a hand-curated row, and the next one GitHub adds would read as "cost
-      // unknown" — an inconsistent picture for a billing reality that is
-      // identical across the entire catalog.
-      //
-      // `isLocal` is the OTHER way to reach a real $0, and it is the wrong one:
-      // it means "runs on this machine" and carries consequences that are
-      // simply false for a hosted API — the shared concurrency-1
-      // LocalRequestQueue (serialising these calls behind actual Ollama GPU
-      // inference), the 300s local rather than 120s cloud inference timeout,
-      // and a literal "[local]" tag in `cascade models`.
-      inputCostPer1kTokens: 0,
-      outputCostPer1kTokens: 0,
-      pricingUnknown: false,
-      maxOutputTokens: GITHUB_MODELS_MAX_OUTPUT_TOKENS,
-      supportsStreaming: true,
-      // Not every catalog model supports function/tool calling — a hardcoded
-      // `true` sent `tools` to models that reject the parameter outright and
-      // never engaged t3-worker.ts's existing text-tool fallback (only an
-      // explicit `false` does). See looksToolCapable's own comment for why
-      // "not positively advertised" defaults to false rather than true here.
-      supportsToolUse: e.tools,
-      isLocal: false
-    }));
-  }
-  async isAvailable() {
-    try {
-      const res = await nodeHttpFetch(GITHUB_MODELS_CATALOG_URL, { headers: this.catalogHeaders() }, 0, {
-        allowedRedirectOrigin: GITHUB_MODELS_CATALOG_ORIGIN
-      });
-      return res.ok;
-    } catch {
-      return false;
-    }
-  }
-};
-
 // src/providers/ollama.ts
 var TOOL_CAPABLE_FAMILIES = [
   "llama3.1",
@@ -4058,7 +3854,7 @@ var ModelSelector = class {
     if (overrideModelId.includes(":")) {
       const parts = overrideModelId.split(":");
       const prefix = parts[0].toLowerCase();
-      const validProviders = ["anthropic", "openai", "gemini", "azure", "openai-compatible", "ollama", "github-models"];
+      const validProviders = ["anthropic", "openai", "gemini", "azure", "openai-compatible", "ollama"];
       if (validProviders.includes(prefix)) {
         providerStr = prefix;
         actualId = parts.slice(1).join(":");
@@ -4078,31 +3874,6 @@ var ModelSelector = class {
       else if (this.availableProviders.has("openai-compatible")) providerStr = "openai-compatible";
       else if (this.availableProviders.size === 1) providerStr = Array.from(this.availableProviders)[0];
     }
-    if (providerStr === "github-models" && this.availableProviders.has("github-models")) {
-      const dynamicModel = {
-        id: actualId,
-        name: actualId,
-        provider: "github-models",
-        contextWindow: GITHUB_MODELS_MAX_INPUT_TOKENS,
-        isVisionCapable: false,
-        inputCostPer1kTokens: 0,
-        outputCostPer1kTokens: 0,
-        pricingUnknown: false,
-        maxOutputTokens: GITHUB_MODELS_MAX_OUTPUT_TOKENS,
-        supportsStreaming: true,
-        // Unconfirmed (no live catalog entry to read a real capability from)
-        // must default to false, not left undefined: t3-worker.ts's text-tool
-        // fallback only engages on a strict `=== false`, so an undefined
-        // value here sends native `tools` to an unverified multi-vendor
-        // model and any that don't support function calling reject the
-        // request outright. Same default-to-false-when-unconfirmed policy
-        // listModels() already applies for a real catalog entry.
-        supportsToolUse: false,
-        isLocal: false
-      };
-      this.addDynamicModel(dynamicModel);
-      return dynamicModel;
-    }
     if (providerStr && this.availableProviders.has(providerStr)) {
       const dynamicModel = withResolvedPricing({
         id: actualId,
@@ -4195,20 +3966,7 @@ var DEFAULT_PROVIDER_TPM = {
   gemini: 6e4,
   azure: 3e4,
   "openai-compatible": 3e4,
-  ollama: Number.POSITIVE_INFINITY,
-  // An order of magnitude below the others, on purpose. GitHub Models is
-  // rate-limited by REQUESTS (Free ≈10 RPM, Pro ≈20, plus low daily caps), and
-  // this bucket meters TOKENS — there is no request-count limiter in this
-  // codebase, and building one is not in scope. The approximation works because
-  // the router reserves `model.maxOutputTokens + 512` per call, and the
-  // provider reports GitHub's real ~4K per-request cap rather than the
-  // underlying model's larger one — so every call withdraws a near-constant
-  // ~4.5K regardless of how short the reply actually turns out to be. At 8K/min
-  // that lands around 1–2 calls per minute of headroom: comfortably inside even
-  // the Free tier, with room for the daily cap to matter more than this does.
-  // Users on a higher tier can raise it with no code change through the generic
-  // `config.rateLimits.providerTpm['github-models']` override.
-  "github-models": 8e3
+  ollama: Number.POSITIVE_INFINITY
 };
 var TpmLimiter = class {
   buckets = /* @__PURE__ */ new Map();
@@ -4564,7 +4322,6 @@ var ModelProfiler = class {
       (m) => !alreadyProfiled.has(m.id) && !SKIP_PATTERN.test(m.id) && !SKIP_PATTERN.test(m.name)
     );
     if (toProfile.length === 0) return;
-    const probesHitGithubModels = this.router?.getModelForTier("T3")?.provider === "github-models";
     const openRouterModels = await fetchOpenRouterModels();
     const orByNormalizedId = /* @__PURE__ */ new Map();
     for (const m of openRouterModels) {
@@ -4579,7 +4336,7 @@ var ModelProfiler = class {
         if (orMatch?.description) {
           specializations = extractSpecializations(orMatch.description);
         }
-        if (specializations.length === 0 && this.router && !probesHitGithubModels) {
+        if (specializations.length === 0 && this.router) {
           specializations = await queryModelDirectly(this.router);
         }
         this.store.saveModelProfile(model.id, model.provider, specializations);
@@ -4799,7 +4556,6 @@ var LiveDataProvider = class {
    */
   applyLivePricing(models) {
     return models.map((m) => {
-      if (m.provider === "github-models") return m;
       const p = this.getLivePrice(m.baseModelId ?? m.id);
       if (!p) return m;
       const reconciled = reconcilePrice(m, p);
@@ -4838,7 +4594,6 @@ var LiveDataProvider = class {
    */
   applyLiveCapabilities(models) {
     return models.map((m) => {
-      if (m.provider === "github-models") return m;
       const c = this.getCapability(m.baseModelId ?? m.id);
       if (!c) return m;
       const next = { ...m };
@@ -5142,16 +4897,13 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
       for (const model of azureDeployments) this.selector.addDynamicModel(model);
       availableProviders.add("azure");
     }
-    if (availableProviders.has("github-models")) {
-      await this.discoverGitHubModelsCatalog(config);
-    }
     await this.validateCloudProviderModels(config);
     for (const tier of ["T1", "T2", "T3"]) {
       const override = tier === "T1" ? config.models.t1 : tier === "T2" ? config.models.t2 : config.models.t3;
       if (!override || override === "auto") continue;
       const model = this.selector.selectForTier(tier, override);
       if (!model) {
-        const knownProviders = ["anthropic", "openai", "gemini", "azure", "openai-compatible", "ollama", "github-models"];
+        const knownProviders = ["anthropic", "openai", "gemini", "azure", "openai-compatible", "ollama"];
         const hasProviderPrefix = override.includes(":") && knownProviders.some((p) => override.startsWith(p + ":"));
         if (hasProviderPrefix) {
           const provider = override.split(":")[0];
@@ -5306,7 +5058,7 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
    * models without a package upgrade. Mirrors discoverOllamaModels.
    */
   async discoverProviderModels() {
-    const cloud = ["anthropic", "openai", "gemini", "azure", "openai-compatible", "github-models"];
+    const cloud = ["anthropic", "openai", "gemini", "azure", "openai-compatible"];
     const tasks = cloud.map(async (type) => {
       if (!this.selector.isProviderAvailable(type)) return;
       const seed = this.getAnyModelForProvider(type);
@@ -5821,26 +5573,6 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
       return false;
     }
   }
-  /**
-   * Registers GitHub Models' real catalog ids via addDynamicModel so init()'s
-   * tier-fill loop can pick one. Registration alone does not create a bound
-   * BaseProvider instance (see ensureProvider) — the tier-fill loop's own
-   * ensureProvider call does that for whichever model it actually selects,
-   * which is why this runs BEFORE that loop rather than only in the
-   * background discoverProviderModels() refresh.
-   */
-  async discoverGitHubModelsCatalog(config) {
-    const cfg = config.providers.find((p) => p.type === "github-models") ?? { type: "github-models" };
-    const seed = this.getAnyModelForProvider("github-models");
-    if (!seed) return;
-    try {
-      const provider = new GitHubModelsProvider(cfg, seed);
-      const models = await provider.listModels();
-      for (const m of models) this.selector.addDynamicModel(m);
-    } catch (err) {
-      console.warn("[router] GitHub Models catalog discovery failed:", err instanceof Error ? err.message : err);
-    }
-  }
   ensureProvider(model, configs) {
     const key = `${model.provider}:${model.id}`;
     if (this.providers.has(key)) return;
@@ -5865,8 +5597,6 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
         return new OllamaProvider(cfg, model);
       case "openai-compatible":
         return new OpenAICompatibleProvider(cfg, model);
-      case "github-models":
-        return new GitHubModelsProvider(cfg, model);
       default:
         throw new Error(`Unsupported provider type: ${String(cfg.type)}`);
     }
@@ -5874,7 +5604,7 @@ var CascadeRouter = class _CascadeRouter extends EventEmitter {
   getAnyModelForProvider(type) {
     const fromCatalog = Object.values(MODELS).find((m) => m.provider === type);
     if (fromCatalog) return fromCatalog;
-    if (type === "openai-compatible" || type === "azure" || type === "github-models") {
+    if (type === "openai-compatible" || type === "azure") {
       return {
         id: type,
         name: type,
@@ -7522,6 +7252,22 @@ ${assignment.expectedOutput}`;
   requiresArtifact() {
     return shouldRequireArtifact(this.assignment, this.tools.map((t) => t.name));
   }
+  /**
+   * Where a file a tool wrote actually lives.
+   *
+   * This used to be `process.cwd()`, which is only the workspace in a plain
+   * CLI run. The hosted server runs with cwd at the app directory while the
+   * run's workspace is the tenant's scratch dir, and Electron's cwd is the app
+   * bundle rather than the folder the user picked. So the tool wrote
+   * `<workspace>/report.docx`, verification stat'd `<cwd>/report.docx`, found
+   * nothing, and the worker looped and then threw WorkerStallError — surfacing
+   * as "no file was created" on a file that had in fact been created, followed
+   * by a failed node. Asking the registry is exact: it is the same root every
+   * registered tool was configured with (registry.registerDefaults).
+   */
+  artifactRoot() {
+    return this.toolRegistry.getWorkspaceRoot?.() ?? process.cwd();
+  }
   extractArtifactPaths(assignment) {
     const declared = (assignment.files ?? []).map((f) => f.trim()).filter((f) => f.includes("."));
     const haystack = `${assignment.description}
@@ -7533,11 +7279,11 @@ ${assignment.expectedOutput}`;
     const artifactPaths = this.extractArtifactPaths(assignment);
     if (!artifactPaths.length) return { ok: true, issues: [] };
     const issues = [];
-    const { exec: exec2 } = await import('child_process');
+    const { execFile: execFile4 } = await import('child_process');
     const { promisify: promisify4 } = await import('util');
-    const execAsync2 = promisify4(exec2);
+    const execFileAsync3 = promisify4(execFile4);
     for (const artifactPath of artifactPaths) {
-      const absolutePath = path26.resolve(process.cwd(), artifactPath);
+      const absolutePath = path26.resolve(this.artifactRoot(), artifactPath);
       try {
         const stat = await fs5.stat(absolutePath);
         if (!stat.isFile()) {
@@ -7573,20 +7319,18 @@ ${assignment.expectedOutput}`;
           continue;
         }
         const ext = path26.extname(absolutePath).toLowerCase();
-        try {
-          if (ext === ".ts" || ext === ".tsx") {
-            await execAsync2(`npx tsc --noEmit ${absolutePath}`, { timeout: 1e4 });
-          } else if (ext === ".js" || ext === ".jsx") {
-            await execAsync2(`node --check ${absolutePath}`, { timeout: 1e4 });
-          } else if (ext === ".py") {
-            await execAsync2(`python -m py_compile ${absolutePath}`, { timeout: 1e4 });
+        const check = ext === ".ts" || ext === ".tsx" ? { cmd: "npx", args: ["tsc", "--noEmit", absolutePath] } : ext === ".js" || ext === ".jsx" ? { cmd: process.execPath, args: ["--check", absolutePath] } : ext === ".py" ? { cmd: "python", args: ["-m", "py_compile", absolutePath] } : null;
+        if (check) {
+          try {
+            await execFileAsync3(check.cmd, check.args, { timeout: 1e4 });
+          } catch (err) {
+            const code = err.code;
+            if (code === "ENOENT" || code === "EINVAL") continue;
+            const e = err;
+            issues.push(`Semantic error in ${artifactPath}:
+${e.stderr || String(err)}
+${e.stdout || ""}`);
           }
-        } catch (err) {
-          const stderr = err?.stderr || String(err);
-          const stdout = err?.stdout || "";
-          issues.push(`Semantic error in ${artifactPath}:
-${stderr}
-${stdout}`);
         }
       } catch {
         issues.push(`Required artifact was not created: ${artifactPath}`);
@@ -7665,7 +7409,7 @@ HIERARCHY CONTEXT: ${this.hierarchyContext}` : ""),
   async checkAcceptance(assignment) {
     const criteria = assignment.acceptance ?? [];
     if (!criteria.length) return [];
-    const resolve2 = (target) => path26.resolve(process.cwd(), target);
+    const resolve2 = (target) => path26.resolve(this.artifactRoot(), target);
     return evaluateAcceptance(criteria, assignment.files ?? [], {
       stat: async (target) => {
         try {
@@ -13253,6 +12997,18 @@ var ToolRegistry = class extends EventEmitter {
     this.tools.set(tool.name, tool);
     this.emit("tool:added", tool.name);
   }
+  /**
+   * The directory every registered tool resolves relative paths against.
+   *
+   * Exposed so a caller that has to reason about a file a tool WROTE can look
+   * in the same place the tool wrote it. `process.cwd()` is not that place
+   * outside a plain CLI run: the hosted server's cwd is the app directory
+   * while the run's workspace is the tenant's scratch dir, and Electron's cwd
+   * is the app bundle rather than the user's chosen folder.
+   */
+  getWorkspaceRoot() {
+    return this.workspaceRoot;
+  }
   /**
    * Wait until a named tool is registered, resolving immediately if it already exists.
    * T3 workers can call this after encountering a missing-tool error to resume
@@ -14007,7 +13763,7 @@ var PermissionEscalator = class extends EventEmitter {
   }
 };
 var ProviderConfigSchema = z.object({
-  type: z.enum(["anthropic", "openai", "gemini", "azure", "openai-compatible", "ollama", "github-models"]),
+  type: z.enum(["anthropic", "openai", "gemini", "azure", "openai-compatible", "ollama"]),
   label: z.string().optional(),
   apiKey: z.string().optional(),
   baseUrl: z.string().url().optional(),
@@ -19283,6 +19039,19 @@ Original error: ${err.message}`
     const rows = provider ? this.db.prepare("SELECT metadata FROM model_cache WHERE provider = ?").all(provider) : this.db.prepare("SELECT metadata FROM model_cache").all();
     return rows.map((r) => JSON.parse(r.metadata));
   }
+  /**
+   * Deletes cached rows for a provider type that no longer exists in
+   * `ProviderType`, so it takes a raw string rather than the union.
+   *
+   * Needed because the cache outlives the code: the REPL's loadCache() only
+   * re-discovers when the cache is EMPTY or older than 24h, so rows left
+   * behind by a removed provider read as "cache populated" and suppress
+   * discovery for the providers that replaced it — `/model` then shows a
+   * configured provider with zero models until the rows age out.
+   */
+  purgeCachedModelsForRetiredProvider(providerType) {
+    return this.db.prepare("DELETE FROM model_cache WHERE provider = ?").run(providerType).changes;
+  }
   clearModelCache(provider) {
     if (provider) {
       this.db.prepare("DELETE FROM model_cache WHERE provider = ?").run(provider);
@@ -19630,7 +19399,73 @@ function mergeGlobalCredentials(workspaceProviders, globalProviders) {
   return merged;
 }
 
+// src/config/retired-providers.ts
+var RETIRED_PROVIDER_TYPES = {
+  "github-models": "GitHub retired GitHub Models on 30 July 2026; models.github.ai no longer responds."
+};
+function isRetiredProviderType(type) {
+  return typeof type === "string" && Object.hasOwn(RETIRED_PROVIDER_TYPES, type);
+}
+function didCleanupChangeAnything(c) {
+  return c.removed.length > 0 || c.clearedPins.length > 0;
+}
+function stripRetiredProviders(raw) {
+  const cleanup = { removed: [], clearedPins: [] };
+  if (typeof raw !== "object" || raw === null) return cleanup;
+  const cfg = raw;
+  if (Array.isArray(cfg["providers"])) {
+    const kept = cfg["providers"].filter((p) => {
+      const type = p?.type;
+      if (!isRetiredProviderType(type)) return true;
+      if (!cleanup.removed.includes(type)) cleanup.removed.push(type);
+      return false;
+    });
+    cfg["providers"] = kept;
+  }
+  cleanup.clearedPins = clearRetiredPins(cfg["models"]);
+  return cleanup;
+}
+function clearRetiredPins(models) {
+  const cleared = [];
+  if (typeof models !== "object" || models === null) return cleared;
+  const tiers = models;
+  for (const tier of ["t1", "t2", "t3"]) {
+    const pin = tiers[tier];
+    if (typeof pin !== "string") continue;
+    const prefix = pin.slice(0, pin.indexOf(":")).toLowerCase();
+    if (pin.includes(":") && isRetiredProviderType(prefix)) {
+      delete tiers[tier];
+      cleared.push(tier);
+    }
+  }
+  return cleared;
+}
+function filterRetiredCredentials(providers) {
+  const removed = [];
+  const kept = providers.filter((p) => {
+    if (!isRetiredProviderType(p.type)) return true;
+    if (!removed.includes(p.type)) removed.push(p.type);
+    return false;
+  });
+  return { kept, removed };
+}
+function describeCleanup(c) {
+  const parts = [];
+  for (const type of c.removed) {
+    parts.push(`removed the retired "${type}" provider (${RETIRED_PROVIDER_TYPES[type]})`);
+  }
+  if (c.clearedPins.length > 0) {
+    parts.push(`reset ${c.clearedPins.map((t) => t.toUpperCase()).join("/")} to Auto, since the pin named it`);
+  }
+  return `Cascade config migration: ${parts.join("; ")}.`;
+}
+
 // src/config/index.ts
+var KEY_OPTIONAL_PROVIDER_TYPES = /* @__PURE__ */ new Set(["ollama", "openai-compatible"]);
+function hasUsableProvider(providers) {
+  if (!providers?.length) return false;
+  return providers.some((p) => KEY_OPTIONAL_PROVIDER_TYPES.has(p.type) || !!p.apiKey || !!p.authToken);
+}
 var ConfigManager = class {
   config;
   keystore;
@@ -19639,6 +19474,29 @@ var ConfigManager = class {
   cascadeMd = null;
   workspacePath;
   globalDir;
+  /**
+   * What THIS load migrated out, if anything. Strictly per-load state: reset
+   * at the top of every `load()`, because two places read it as a statement
+   * about the load in progress — `injectEnvKeys()` asks whether an empty
+   * provider list was emptied by a retirement, and the end of `load()` warns
+   * and builds the user-facing notice from it. Left set, a second `load()` on
+   * the same instance (`startRepl()` does one after its setup wizard) would
+   * re-announce a migration that already happened and suppress the Ollama
+   * fallback for a list that is empty for some unrelated reason.
+   */
+  retiredCleanup;
+  /**
+   * Human-readable migration notice held for a UI to display. `console.warn`
+   * is not sufficient: the REPL clears the TTY immediately after load(), and
+   * the desktop emits from the main process where nothing is rendered.
+   */
+  pendingRetiredNotice;
+  /** Returns the pending migration notice once, then forgets it. */
+  takeRetiredNotice() {
+    const n = this.pendingRetiredNotice;
+    this.pendingRetiredNotice = void 0;
+    return n;
+  }
   /** `globalDirOverride` exists for tests — never point it at the real home dir there. */
   constructor(workspacePath = process.cwd(), globalDirOverride) {
     this.workspacePath = workspacePath;
@@ -19658,10 +19516,32 @@ var ConfigManager = class {
     this.cascadeMd = await loadCascadeMd(this.workspacePath);
     this.keystore = new Keystore(path26.join(this.globalDir, GLOBAL_KEYSTORE_FILE));
     this.store = new MemoryStore(path26.join(this.workspacePath, CASCADE_DB_FILE));
+    const globalCreds = filterRetiredCredentials(loadGlobalCredentials(this.globalDir));
+    if (globalCreds.removed.length > 0) {
+      try {
+        saveGlobalCredentials(this.globalDir, globalCreds.kept);
+      } catch (err) {
+        console.warn(`Could not rewrite the global credential store: ${err instanceof Error ? err.message : String(err)}`);
+      }
+      this.retiredCleanup = {
+        removed: [.../* @__PURE__ */ new Set([...this.retiredCleanup?.removed ?? [], ...globalCreds.removed])],
+        clearedPins: this.retiredCleanup?.clearedPins ?? []
+      };
+    }
     await this.injectEnvKeys();
-    this.config.providers = mergeGlobalCredentials(this.config.providers, loadGlobalCredentials(this.globalDir));
+    this.config.providers = mergeGlobalCredentials(this.config.providers, globalCreds.kept);
+    for (const type of Object.keys(RETIRED_PROVIDER_TYPES)) {
+      try {
+        this.store.purgeCachedModelsForRetiredProvider(type);
+      } catch {
+      }
+    }
     await this.ensureDefaultIdentity();
     if (mcpNamesChanged) await this.save();
+    if (this.retiredCleanup) {
+      console.warn(describeCleanup(this.retiredCleanup));
+      this.pendingRetiredNotice = describeCleanup(this.retiredCleanup);
+    }
   }
   getConfig() {
     return this.config;
@@ -19700,14 +19580,13 @@ var ConfigManager = class {
       anthropic: "ANTHROPIC_API_KEY",
       openai: "OPENAI_API_KEY",
       gemini: "GOOGLE_API_KEY",
-      azure: "AZURE_OPENAI_KEY",
+      azure: "AZURE_OPENAI_KEY"
       // Deliberately NOT `GITHUB_TOKEN`: that one is injected automatically by
       // GitHub Actions and is commonly exported in developer shells for `gh`
       // and git auth, and it will essentially never carry the fine-grained
       // `models: read` scope this provider needs. Adopting it would wire up a
       // provider that 403s on its first real call — a confusing failure for
       // something that is meant to be explicit opt-in BYOK.
-      "github-models": "GITHUB_MODELS_TOKEN"
     };
     const envKey = envMap[provider];
     if (envKey && process.env[envKey]) return process.env[envKey];
@@ -19719,10 +19598,21 @@ var ConfigManager = class {
     return configProvider?.apiKey;
   }
   async loadConfig() {
+    this.retiredCleanup = void 0;
     const configPath = path26.join(this.workspacePath, CASCADE_CONFIG_FILE);
     try {
       const raw = await fs5.readFile(configPath, "utf-8");
-      return validateConfig(JSON.parse(raw));
+      const parsed = JSON.parse(raw);
+      const cleanup = stripRetiredProviders(parsed);
+      if (didCleanupChangeAnything(cleanup)) {
+        this.retiredCleanup = cleanup;
+        try {
+          await fs5.writeFile(configPath, JSON.stringify(parsed, null, 2), "utf-8");
+        } catch (err) {
+          console.warn(`Could not persist the retired-provider migration: ${err instanceof Error ? err.message : String(err)}`);
+        }
+      }
+      return validateConfig(parsed);
     } catch (err) {
       if (err.code === "ENOENT") {
         return validateConfig({});
@@ -19731,27 +19621,25 @@ var ConfigManager = class {
     }
   }
   async injectEnvKeys() {
-    const isFirstRun = this.config.providers.length === 0;
+    const wasEmpty = this.config.providers.length === 0;
+    const emptiedByRetirement = wasEmpty && !!this.retiredCleanup;
     const envProviders = [
       { env: "ANTHROPIC_API_KEY", type: "anthropic" },
       { env: "OPENAI_API_KEY", type: "openai" },
       { env: "GOOGLE_API_KEY", type: "gemini" },
-      { env: "AZURE_OPENAI_KEY", type: "azure" },
-      // See getApiKey() above for why this is GITHUB_MODELS_TOKEN and never
-      // the ambient GITHUB_TOKEN.
-      { env: "GITHUB_MODELS_TOKEN", type: "github-models" }
+      { env: "AZURE_OPENAI_KEY", type: "azure" }
     ];
     for (const { env, type } of envProviders) {
       const key = process.env[env];
       if (!key) continue;
       const existing = this.config.providers.find((p) => p.type === type);
-      if (!existing && isFirstRun) {
+      if (!existing && wasEmpty) {
         this.config.providers.push({ type, apiKey: key });
       } else if (existing && !existing.apiKey) {
         existing.apiKey = key;
       }
     }
-    if (isFirstRun && !this.config.providers.find((p) => p.type === "ollama")) {
+    if (wasEmpty && !emptiedByRetirement && !this.config.providers.find((p) => p.type === "ollama")) {
       this.config.providers.push({ type: "ollama" });
     }
   }
@@ -22388,7 +22276,9 @@ function gatherSyncBundle(config) {
   if (Object.keys(prefs).length) bundle.prefs = prefs;
   return bundle;
 }
-function applySyncBundle(bundle, config) {
+function applySyncBundle(bundle, config, cleanup) {
+  const dropped = stripRetiredProviders(bundle);
+  if (cleanup) cleanup.removed = dropped.removed;
   const next = { ...config, tools: { ...config.tools } };
   if (bundle.providers) next.providers = mergeByKey(config.providers, bundle.providers, providerSig) ?? [];
   if (bundle.webSearch) next.tools.webSearch = { ...config.tools?.webSearch, ...bundle.webSearch };
@@ -22404,6 +22294,14 @@ function applySyncBundle(bundle, config) {
   }
   if (bundle.models) next.models = { ...config.models, ...bundle.models };
   if (bundle.budget) next.budget = { ...config.budget, ...bundle.budget };
+  const merged = { ...next.models };
+  const clearedFromMerge = clearRetiredPins(merged);
+  if (clearedFromMerge.length > 0) next.models = merged;
+  if (cleanup) {
+    cleanup.clearedPins = ["t1", "t2", "t3"].filter(
+      (t) => (dropped.clearedPins.includes(t) || clearedFromMerge.includes(t)) && next.models?.[t] === void 0
+    );
+  }
   if (bundle.prefs) {
     const p = bundle.prefs;
     if (p.autoBias !== void 0) next.autoBias = p.autoBias;
@@ -22414,6 +22312,6 @@ function applySyncBundle(bundle, config) {
   return next;
 }
 
-export { AZURE_BASE_MODELS, AZURE_BASE_URL_TEMPLATE, AuditLogger, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, CascadeIgnore, CascadeRouter, CascadeToolError, CloudClient, CodeSearchTool, ConfigManager, CurrentPageTool, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, DOCUMENT_MIME, DashboardServer, DependencyScheduler, FileMcpOAuthStore, GITHUB_MODELS_API_VERSION, GITHUB_MODELS_CATALOG_URL, GITHUB_MODELS_INFERENCE_URL, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, GenerateDocumentTool, GraphRetriever, GraphSearchTool, HooksRunner, Keystore, LLMReranker, LM_STUDIO_BASE_URL, MAX_PAGE_TEXT_CHARS, MCP_TOOL_PREFIX, MODELS, McpClient, McpOAuthProvider, MemoryStore, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, PROVIDER_DISPLAY_NAMES, RESUME_SCHEMA_VERSION, ResumeStore, Retriever, SqliteVectorStore, T1Administrator, T1_MODEL_PRIORITY, T2Manager, T2_MODEL_PRIORITY, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, TaskGraphValidationError, TaskScheduler, Telemetry, ToolRegistry, VISION_MODEL_PRIORITY, WorkspaceIndex, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildDocumentTools, buildManifest, buildSessionTranscript, bytesToBase64, cagCharBudget, chartKind, chartToTableRows, chatCompleterFromProviders, chunkCode, chunkText2 as chunkText, compileSectionGraph, compileSubtaskGraph, compileTaskGraph, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, disambiguateMcpServerNames, discoverMcpAuthServer, discoverMcpTools, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, evaluateAcceptance, extractCharts, failures, fileExt, fileOAuthProvider, gatherSyncBundle, hashContent, heuristicCodeChunker, inferAzureBaseModel, inlineRuns, isDocumentFormat, isMcpToolName, matchImageLine, mcpServerPrefix, mcpToolName, nodeHttpFetch, parseBlocks, parseChartSpec, parseDelimited, parseRankOrder, parseSlide, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, removeMcpServerDenials, renderDocument, renderDocx, renderPptx, renderXlsx, runCascade, sessionWorthRemembering, sniffImage, splitSlides, streamCascade, stripInline, summarizeCompleted, undecided, uniqueMcpServerName };
+export { AZURE_BASE_MODELS, AZURE_BASE_URL_TEMPLATE, AuditLogger, CASCADE_AUDIT_FILE, CASCADE_CONFIG_DIR, CASCADE_CONFIG_FILE, CASCADE_DASHBOARD_SECRET_FILE, CASCADE_DB_FILE, CASCADE_IGNORE_FILE, CASCADE_KEYSTORE_FILE, CASCADE_MD_FILE, CASCADE_VERSION, CHARS_PER_TOKEN, COMPLEXITY_T2_COUNT, Cascade, CascadeCancelledError, CascadeIgnore, CascadeRouter, CascadeToolError, CloudClient, CodeSearchTool, ConfigManager, CurrentPageTool, DEFAULT_API_PORT, DEFAULT_APPROVAL_REQUIRED, DEFAULT_AUTO_SUMMARIZE_AT, DEFAULT_CLOUD_URL, DEFAULT_CONTEXT_LIMIT, DEFAULT_DASHBOARD_PORT, DEFAULT_MAX_SESSION_MESSAGES, DEFAULT_RETENTION_DAYS, DEFAULT_THEME, DOCUMENT_MIME, DashboardServer, DependencyScheduler, FileMcpOAuthStore, GLOBAL_CONFIG_DIR, GLOBAL_CREDENTIALS_FILE, GLOBAL_DB_FILE, GLOBAL_KEYSTORE_FILE, GLOBAL_RUNTIME_DB_FILE, GenerateDocumentTool, GraphRetriever, GraphSearchTool, HooksRunner, Keystore, LLMReranker, LM_STUDIO_BASE_URL, MAX_PAGE_TEXT_CHARS, MCP_TOOL_PREFIX, MODELS, McpClient, McpOAuthProvider, MemoryStore, OLLAMA_BASE_URL, OpenAICompatibleEmbedder, PROVIDER_DISPLAY_NAMES, RESUME_SCHEMA_VERSION, ResumeStore, Retriever, SqliteVectorStore, T1Administrator, T1_MODEL_PRIORITY, T2Manager, T2_MODEL_PRIORITY, T3Worker, T3_MODEL_PRIORITY, THEME_NAMES, TOOL_NAMES, TaskGraphValidationError, TaskScheduler, Telemetry, ToolRegistry, VISION_MODEL_PRIORITY, WorkspaceIndex, applySyncBundle, azureModelForDeployment, beginMcpOAuth, buildDocumentTools, buildManifest, buildSessionTranscript, bytesToBase64, cagCharBudget, chartKind, chartToTableRows, chatCompleterFromProviders, chunkCode, chunkText2 as chunkText, compileSectionGraph, compileSubtaskGraph, compileTaskGraph, completeMcpOAuth, connectMcpWithLoopbackOAuth, createCascade, decryptJSON as decryptSyncBlob, diffManifest, disambiguateMcpServerNames, discoverMcpAuthServer, discoverMcpTools, distillSessionFacts, embedderFromProviders, encryptJSON as encryptSyncBlob, evaluateAcceptance, extractCharts, failures, fileExt, fileOAuthProvider, gatherSyncBundle, hasUsableProvider, hashContent, heuristicCodeChunker, inferAzureBaseModel, inlineRuns, isDocumentFormat, isMcpToolName, matchImageLine, mcpServerPrefix, mcpToolName, nodeHttpFetch, parseBlocks, parseChartSpec, parseDelimited, parseRankOrder, parseSlide, planRetrieval, preferIpv4Host, reciprocalRankFusion, refreshMcpToken, removeMcpServerDenials, renderDocument, renderDocx, renderPptx, renderXlsx, runCascade, sessionWorthRemembering, sniffImage, splitSlides, streamCascade, stripInline, summarizeCompleted, undecided, uniqueMcpServerName };
 //# sourceMappingURL=index.js.map
 //# sourceMappingURL=index.js.map