Use aieo's xAI provider instead of hand-rolled workarounds

package.json

@@ -103,7 +103,7 @@
     "@vercel/functions": "^3.1.4",
     "@xyflow/react": "^12.8.6",
     "ai": "^6.0.72",
-    "aieo": "^0.1.34",
+    "aieo": "^0.1.35",
     "autoprefixer": "^10.4.21",
     "axios": "1.10.0",
     "bcryptjs": "^3.0.2",

src/__tests__/unit/lib/ai/models-aieo-parity.test.ts

@@ -0,0 +1,37 @@
+import { describe, test, expect, afterEach } from "vitest";
+import { PROVIDERS, hasApiKeyForProvider } from "aieo";
+import { PROVIDER_API_KEY_ENV_VARS, PROVIDER_DISPLAY_LABELS } from "@/lib/ai/models";
+
+/**
+ * `src/lib/ai/models.ts` can't import aieo (it's pulled into client
+ * bundles), so it carries its own provider -> env-var table. This test
+ * pins that table to aieo's key lookup so a provider added to one side
+ * (e.g. xai) can't silently be missing or mis-keyed on the other.
+ */
+describe("models.ts provider table matches aieo", () => {
+  const saved: Record<string, string | undefined> = {};
+
+  afterEach(() => {
+    for (const [k, v] of Object.entries(saved)) {
+      if (v === undefined) delete process.env[k];
+      else process.env[k] = v;
+    }
+  });
+
+  test.each(PROVIDERS)("aieo provider %s has a hive env var and display label", (provider) => {
+    const enumKey = provider.toUpperCase();
+    expect(PROVIDER_API_KEY_ENV_VARS[enumKey]).toBeTruthy();
+    expect(PROVIDER_DISPLAY_LABELS[enumKey]).toBeTruthy();
+  });
+
+  test.each(PROVIDERS)("setting hive's env var for %s satisfies aieo's key lookup", (provider) => {
+    const envVar = PROVIDER_API_KEY_ENV_VARS[provider.toUpperCase()]!;
+    saved[envVar] = process.env[envVar];
+
+    delete process.env[envVar];
+    expect(hasApiKeyForProvider(provider)).toBe(false);
+
+    process.env[envVar] = "parity-test-key";
+    expect(hasApiKeyForProvider(provider)).toBe(true);
+  });
+});

src/app/api/agent/route.ts

@@ -362,14 +362,14 @@ async function createAgentSession(
     }
   }
 
-  // xAI bypass: the Bifrost VK reconciler falls back to anthropic's
-  // provider suffix for any model prefix it doesn't recognize, and its
-  // provider allow-list doesn't include xai — routing an xai/* session
-  // through Bifrost today would mint a VK pointed at the wrong
-  // provider. Skip the call for xai/* and use the direct XAI_API_KEY
-  // resolved above instead. Remove once an `aieo` gateway path + a
-  // Bifrost-side xai provider config both exist (see
-  // src/services/task-workflow.ts for the matching bypass).
+  // xAI bypass: the Bifrost VK provider allow-list (DEFAULT_PROVIDERS in
+  // src/services/bifrost/constants.ts) doesn't include xai, and the
+  // swarm gateways have no xai provider key configured — routing an
+  // xai/* session through Bifrost today would fail. Skip the call for
+  // xai/* and use the direct XAI_API_KEY resolved above instead. Remove
+  // once the gateways carry an xai key and DEFAULT_PROVIDERS lists it
+  // (aieo already maps xai onto the /openai/v1 gateway path). See
+  // src/services/task-workflow.ts for the matching bypass.
   const isXaiModel = effectiveModel?.startsWith("xai/") ?? false;
 
   // Bifrost routing for the goose-side LLM calls. When the rollout

src/app/org/[githubLogin]/_components/CanvasAgentSettingsPopover.tsx

@@ -70,22 +70,14 @@ export function CanvasAgentSettingsPopover({
     };
   }, []);
 
-  // Load the available models for the picker. xAI/Grok rows are excluded
-  // here specifically (not just left to the server-side key filter) —
-  // `aieo` (the canvas agent's LLM SDK, ^0.1.34) has no xai provider
-  // entry yet, so `runCanvasAgent` throws a visible error rather than
-  // silently answering as Anthropic if one is selected. Hiding it from
-  // this picker means users never hit that error in the first place.
-  // Remove this filter once aieo supports xai — see the "Add xAI"
-  // feature notes.
+  // Load the available models for the picker. /api/llm-models already
+  // filters out providers whose API key isn't configured server-side.
   useEffect(() => {
     let cancelled = false;
     fetch("/api/llm-models")
       .then((res) => (res.ok ? res.json() : null))
       .then((data) => {
-        if (!cancelled && data?.models) {
-          setModels((data.models as LlmModelOption[]).filter((m) => m.provider !== "XAI"));
-        }
+        if (!cancelled && data?.models) setModels(data.models);
       })
       .catch(() => {
         /* leave empty; picker stays hidden until a retry */

src/config/env.ts

@@ -166,7 +166,7 @@ export const optionalEnvVars = {
   // raw `process.env` read — so it's a typed, discoverable config value like
   // its sibling provider keys. Absence is a normal, expected state in any
   // environment that hasn't onboarded xAI yet: `/api/llm-models` filters out
-  // XAI rows when this is unset (see step 8 of the xAI feature), so pickers
+  // XAI rows when this is unset, so pickers
   // stay empty rather than erroring.
   XAI_API_KEY: process.env.XAI_API_KEY || "",
 } as const;

src/lib/ai/models.ts

@@ -61,16 +61,20 @@ export function isValidModel(model: unknown): model is string {
   return false;
 }
 
-// Map LlmProvider enum values to their API key environment variables
+// Map LlmProvider enum values to their API key environment variables.
+//
+// For the providers aieo knows about (anthropic/google/openai/openrouter/xai)
+// these MUST match the env vars aieo's own key lookup reads — aieo is the
+// source of truth for provider routing, and `runCanvasAgent` resolves keys
+// through it directly. This file is imported by client components, so it
+// can't import aieo itself; `models-aieo-parity.test.ts` guards the two
+// tables against drifting apart. AWS_BEDROCK/OTHER are hive-only.
 export const PROVIDER_API_KEY_ENV_VARS: Record<string, string | null> = {
   ANTHROPIC: "ANTHROPIC_API_KEY",
   OPENAI: "OPENAI_API_KEY",
   GOOGLE: "GOOGLE_API_KEY",
   AWS_BEDROCK: "AWS_BEDROCK_API_KEY",
   OPENROUTER: "OPENROUTER_API_KEY",
-  // Direct xAI credential — Grok models resolve to `xai/<name>` and use
-  // this key instead of routing through OpenRouter. See the "Add xAI"
-  // feature notes for the Bifrost/canvas-agent gaps this doesn't cover yet.
   XAI: "XAI_API_KEY",
   OTHER: null,
 };

src/lib/ai/runCanvasAgent.ts

@@ -697,22 +697,7 @@ export async function runCanvasAgent(
   // flag is on for the primary workspace — see `getBifrostForLLM` below.
   let provider: Provider = "anthropic";
   if (modelName?.includes("/")) {
-    const rawPrefix = modelName.split("/")[0];
-    // aieo (^0.1.34, the pinned version) has no "xai" entry in `PROVIDERS`
-    // — it's not just an unconfigured key, the provider itself doesn't
-    // exist on this path yet. Falling through to the generic branch
-    // below would silently answer a Grok selection as Anthropic, which
-    // is exactly the failure mode this feature must not have. Fail
-    // loudly instead. `CanvasAgentSettingsPopover` excludes xai/* rows
-    // from the picker, so this should only fire on a stale/tampered
-    // `chatAgentModel` preference. Remove this guard once `aieo` (or its
-    // replacement) supports xai — see the "Add xAI" feature notes.
-    if (rawPrefix === "xai") {
-      throw new Error(
-        `runCanvasAgent: model "${modelName}" is not available on this path — xAI/Grok is not yet supported by the canvas agent's LLM SDK (aieo). Choose a different model.`,
-      );
-    }
-    const prefix = rawPrefix as Provider;
+    const prefix = modelName.split("/")[0] as Provider;
     if (!PROVIDERS.includes(prefix)) {
       console.warn(
         `[runCanvasAgent] model "${modelName}" has unsupported provider prefix "${prefix}"; falling back to anthropic default`,

src/services/task-workflow.ts

@@ -853,15 +853,14 @@ export async function callStakworkAPI(params: {
   // tuning of ttlSeconds / maxCostUsd / maxSteps is intentionally
   // deferred to a follow-up so this initial wiring stays small.
   //
-  // xAI bypass: `reconcileBifrostVK` derives the `baseUrl` provider
-  // suffix from the model prefix and falls back to anthropic for any
-  // prefix it doesn't recognize, and `DEFAULT_PROVIDERS` doesn't list
-  // "xai" — so routing an `xai/*` selection through Bifrost today would
-  // mint a VK pointed at the wrong provider (or error). Skip the
-  // Bifrost call entirely for xai/* and fall through to the direct
-  // `vars.apiKey` (XAI_API_KEY) resolved above. Remove this bypass once
-  // an `aieo` version with an xAI gateway path AND a Bifrost-side xAI
-  // provider config both exist — until then this trades away per-agent
+  // xAI bypass: the Bifrost VK provider allow-list (`DEFAULT_PROVIDERS`)
+  // doesn't list "xai" and the swarm gateways have no xai provider key
+  // configured, so routing an `xai/*` selection through Bifrost today
+  // would fail. Skip the Bifrost call entirely for xai/* and fall
+  // through to the direct `vars.apiKey` (XAI_API_KEY) resolved above.
+  // Remove this bypass once the gateways carry an xai key and
+  // `DEFAULT_PROVIDERS` lists it (aieo already maps xai onto the
+  // /openai/v1 gateway path) — until then this trades away per-agent
   // cost attribution / macaroon observability for Grok runs only.
   const isXaiModel = effectiveModel?.startsWith("xai/") ?? false;
   const bifrost = isXaiModel