Source instrument specs from grid-defaults.json

grid-defaults.json

@@ -0,0 +1,127 @@
+{
+  "provider": "thegrid",
+  "bases": {
+    "inference": "https://api.thegrid.ai/v1",
+    "messages": "https://messages-beta.api.thegrid.ai"
+  },
+  "apiKeyEnv": "THEGRID_API_KEY",
+  "models": [
+    {
+      "id": "agent-max",
+      "maxTokens": 128000,
+      "contextWindow": 1000000,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 1.5885,
+      "costPer1mOut": 1.5885
+    },
+    {
+      "id": "agent-prime",
+      "maxTokens": 30000,
+      "contextWindow": 196608,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 0.1422,
+      "costPer1mOut": 0.1422
+    },
+    {
+      "id": "agent-standard",
+      "maxTokens": 16000,
+      "contextWindow": 128000,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 0.02,
+      "costPer1mOut": 0.02
+    },
+    {
+      "id": "code-max",
+      "maxTokens": 128000,
+      "contextWindow": 1000000,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 1.7553,
+      "costPer1mOut": 1.7553
+    },
+    {
+      "id": "code-prime",
+      "maxTokens": 30000,
+      "contextWindow": 196608,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 0.1493,
+      "costPer1mOut": 0.1493
+    },
+    {
+      "id": "code-standard",
+      "maxTokens": 16000,
+      "contextWindow": 128000,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 0.0288,
+      "costPer1mOut": 0.0288
+    },
+    {
+      "id": "text-max",
+      "maxTokens": 128000,
+      "contextWindow": 1000000,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 1.68,
+      "costPer1mOut": 1.68
+    },
+    {
+      "id": "text-prime",
+      "maxTokens": 30000,
+      "contextWindow": 196608,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 0.1017,
+      "costPer1mOut": 0.1017
+    },
+    {
+      "id": "text-standard",
+      "maxTokens": 16000,
+      "contextWindow": 128000,
+      "input": [
+        "text"
+      ],
+      "costPer1mIn": 0.03,
+      "costPer1mOut": 0.03
+    }
+  ],
+  "tools": [
+    {
+      "slug": "opencode",
+      "surface": "openai-completions",
+      "baseUrl": "https://api.thegrid.ai/v1"
+    },
+    {
+      "slug": "cline",
+      "surface": "openai-completions",
+      "baseUrl": "https://api.thegrid.ai/v1"
+    },
+    {
+      "slug": "kilocode",
+      "surface": "openai-completions",
+      "baseUrl": "https://api.thegrid.ai/v1"
+    },
+    {
+      "slug": "aider",
+      "surface": "openai-completions",
+      "baseUrl": "https://api.thegrid.ai/v1"
+    },
+    {
+      "slug": "hermes",
+      "surface": "anthropic-messages",
+      "baseUrl": "https://messages-beta.api.thegrid.ai"
+    }
+  ]
+}

packages/cli/src/__tests__/e2e/grid-doc-parity.contract.test.ts

@@ -195,7 +195,7 @@ describe("grid doc parity (configure uploads)", () => {
     expect(uploaded).toContain("key_env: THEGRID_API_KEY");
     expect(uploaded).toContain("provider: custom:thegrid");
     expect(uploaded).toContain("default: agent-prime");
-    expect(uploaded).toContain("context_length: 128000");
+    expect(uploaded).toContain("context_length: 196608");
     expect(uploaded).toContain("supports_vision: false");
     expect(uploaded).toContain("model_aliases:");
     expect(uploaded).toContain("agent_standard:");
@@ -214,7 +214,7 @@ describe("grid doc parity (configure uploads)", () => {
     expect(uploaded).toContain('"thegrid"');
     expect(uploaded).toContain("api.thegrid.ai");
     expect(uploaded).toContain("code-prime");
-    expect(uploaded).toContain('"context": 128000');
+    expect(uploaded).toContain('"context": 196608');
     expect(uploaded).toContain('"modalities"');
     expect(uploaded).toContain('"text"');
   });
@@ -224,7 +224,7 @@ describe("grid doc parity (configure uploads)", () => {
     const { agents: localAgents } = createCloudAgents(runner);
     await localAgents.pi.configure!("test-key");
     const uploaded = [...runner.uploads.values()].join("\n");
-    expect(uploaded).toContain('"contextWindow": 128000');
+    expect(uploaded).toContain('"contextWindow": 196608');
     expect(uploaded).toContain('"input"');
     expect(uploaded).toContain('"text"');
     expect(uploaded).toContain("code-prime");

packages/cli/src/__tests__/grid-defaults-parity.test.ts

@@ -0,0 +1,62 @@
+/**
+ * grid-defaults.json (repo root) is published by the Grid registry generator and is the source
+ * of truth for instrument specs. The CLI ships a fallback mirror for offline/bundled runs;
+ * these tests pin the mirror to the committed file so the two can never drift apart, and cover
+ * the pure spec builder the runtime refresh uses.
+ */
+
+import { readFileSync } from "node:fs";
+import { join } from "node:path";
+import { describe, expect, it } from "bun:test";
+import { resolveGridInstrumentModelSpec, specsFromGridDefaults } from "../shared/grid-instruments.js";
+import { OPENCLAW_GRID_MODEL_MAX_TOKENS } from "../shared/vendor-routing.js";
+
+const GRID_DEFAULTS_PATH = join(import.meta.dir, "../../../..", "grid-defaults.json");
+const gridDefaults = JSON.parse(readFileSync(GRID_DEFAULTS_PATH, "utf-8")) as {
+  models: Array<{ id: string; maxTokens: number; contextWindow: number; input: string[] }>;
+};
+
+describe("grid-defaults.json parity", () => {
+  it("has the registry-generator shape", () => {
+    expect(gridDefaults.models.length).toBeGreaterThan(0);
+    for (const model of gridDefaults.models) {
+      expect(model.contextWindow).toBeGreaterThan(0);
+      expect(model.maxTokens).toBeGreaterThan(0);
+      expect(model.maxTokens).toBeLessThanOrEqual(model.contextWindow);
+      expect(model.input.length).toBeGreaterThan(0);
+    }
+  });
+
+  it("fallback mirror matches the committed catalogue (update both together)", () => {
+    for (const model of gridDefaults.models) {
+      const spec = resolveGridInstrumentModelSpec(model.id);
+      expect(spec.contextWindow).toBe(model.contextWindow);
+      expect([...spec.input]).toEqual(model.input);
+      expect(spec.maxOutputTokens).toBe(Math.min(OPENCLAW_GRID_MODEL_MAX_TOKENS, model.maxTokens));
+    }
+  });
+
+  it("specsFromGridDefaults builds the same map the fallback mirrors", () => {
+    const specs = specsFromGridDefaults(gridDefaults);
+    expect(specs).not.toBeNull();
+    for (const model of gridDefaults.models) {
+      expect(specs?.[model.id]).toEqual({
+        contextWindow: model.contextWindow,
+        maxOutputTokens: Math.min(OPENCLAW_GRID_MODEL_MAX_TOKENS, model.maxTokens),
+        input: model.input as ("text" | "image")[],
+      });
+    }
+  });
+
+  it("specsFromGridDefaults rejects malformed bodies", () => {
+    expect(specsFromGridDefaults(null)).toBeNull();
+    expect(specsFromGridDefaults({})).toBeNull();
+    expect(specsFromGridDefaults({ models: [] })).toBeNull();
+    expect(
+      specsFromGridDefaults({ models: [{ id: "x", contextWindow: -1, maxTokens: 1, input: ["text"] }] }),
+    ).toBeNull();
+    expect(
+      specsFromGridDefaults({ models: [{ id: "x", contextWindow: 10, maxTokens: 1, input: ["video"] }] }),
+    ).toBeNull();
+  });
+});

packages/cli/src/__tests__/grid-instrument-model-spec.test.ts

@@ -6,10 +6,10 @@ import {
 import { OPENCLAW_GRID_MODEL_MAX_TOKENS } from "../shared/vendor-routing.js";
 
 describe("grid instrument model specs", () => {
-  it("uses Cortex catalogue context windows (not oversized defaults)", () => {
-    expect(resolveGridInstrumentModelSpec("agent-prime").contextWindow).toBe(128_000);
-    expect(resolveGridInstrumentModelSpec("code-prime").contextWindow).toBe(128_000);
-    expect(resolveGridInstrumentModelSpec("text-prime").contextWindow).toBe(128_000);
+  it("uses Grid catalogue context windows (grid-defaults.json, not oversized defaults)", () => {
+    expect(resolveGridInstrumentModelSpec("agent-prime").contextWindow).toBe(196_608);
+    expect(resolveGridInstrumentModelSpec("code-prime").contextWindow).toBe(196_608);
+    expect(resolveGridInstrumentModelSpec("text-prime").contextWindow).toBe(196_608);
     expect(resolveGridInstrumentModelSpec("agent-max").contextWindow).toBe(1_000_000);
     expect(resolveGridInstrumentModelSpec("code-max").contextWindow).toBe(1_000_000);
     expect(resolveGridInstrumentModelSpec("text-max").contextWindow).toBe(1_000_000);

packages/cli/src/index.ts

@@ -46,6 +46,7 @@ import {
 import { expandEqualsFlags, findUnknownFlag } from "./flags.js";
 import { agentKeys, cloudKeys, getCacheAge, loadManifest } from "./manifest.js";
 import { expandFastProvisionVariant, getFeatureFlag, initFeatureFlags } from "./shared/feature-flags.js";
+import { refreshGridInstrumentSpecs } from "./shared/grid-instruments.js";
 import { getInstallRefPath } from "./shared/paths.js";
 import { asyncTryCatch, asyncTryCatchIf, isFileError, isNetworkError, tryCatch, tryCatchIf } from "./shared/result.js";
 import { captureError, initTelemetry, setTelemetryContext } from "./shared/telemetry.js";
@@ -976,7 +977,11 @@ async function main(): Promise<void> {
   const preOutputIdx = args.indexOf("--output");
   const isJsonOutput = preOutputIdx !== -1 && args[preOutputIdx + 1] === "json";
 
+  // Refresh instrument specs from grid-defaults.json concurrently with the update check so
+  // provisioning uses live catalogue limits without adding startup latency. Never throws.
+  const specsRefresh = refreshGridInstrumentSpecs();
   await checkForUpdates(isJsonOutput);
+  await specsRefresh;
 
   const [prompt, filteredArgs] = await resolvePrompt(args);