diff --git a/src/resources/extensions/gsd/auto-model-selection.ts b/src/resources/extensions/gsd/auto-model-selection.ts
index 75f2b434f..57a2b8904 100644
--- a/src/resources/extensions/gsd/auto-model-selection.ts
+++ b/src/resources/extensions/gsd/auto-model-selection.ts
@@ -10,7 +10,7 @@ import type { ExtensionAPI, ExtensionContext } from "@gsd/pi-coding-agent";
 import type { GSDPreferences } from "./preferences.js";
 import { resolveModelWithFallbacksForUnit, resolveDynamicRoutingConfig } from "./preferences.js";
 import type { ComplexityTier } from "./complexity-classifier.js";
-import { classifyUnitComplexity, tierLabel } from "./complexity-classifier.js";
+import { classifyUnitComplexity, extractTaskMetadata, tierLabel } from "./complexity-classifier.js";
 import { resolveModelForComplexity, escalateTier, getEligibleModels, loadCapabilityOverrides, adjustToolSet, filterToolsForProvider } from "./model-router.js";
 import { getLedger, getProjectTotals } from "./metrics.js";
 import { unitPhaseLabel } from "./auto-dashboard.js";
@@ -120,6 +120,10 @@ export async function selectAndApplyModel(
     let routingTierLabel = "";
     let routingEligibleModels = availableModels;
 
+    const taskMetadataForPolicy = unitType === "execute-task"
+      ? extractTaskMetadata(unitId, basePath)
+      : undefined;
+
     if (uokFlags.modelPolicy) {
       const policy = applyModelPolicyFilter(
         availableModels,
@@ -128,6 +132,7 @@ export async function selectAndApplyModel(
           traceId: modelPolicyTraceId,
           turnId: modelPolicyTurnId,
           unitType,
+          taskMetadata: taskMetadataForPolicy,
           currentProvider: ctx.model?.provider,
           allowCrossProvider: routingConfig.cross_provider !== false,
           requiredTools: pi.getActiveTools(),
@@ -182,7 +187,13 @@ export async function selectAndApplyModel(
       const shouldClassify = !isHook || routingConfig.hooks !== false;
 
       if (shouldClassify) {
-        let classification = classifyUnitComplexity(unitType, unitId, basePath, budgetPct);
+        let classification = classifyUnitComplexity(
+          unitType,
+          unitId,
+          basePath,
+          budgetPct,
+          taskMetadataForPolicy,
+        );
         const availableModelIds = routingEligibleModels.map(m => m.id);
 
         // Escalate tier on retry when escalate_on_failure is enabled (default: true)
@@ -293,7 +304,8 @@ export async function selectAndApplyModel(
     let attemptedPolicyEligible = false;
 
     for (const modelId of modelsToTry) {
-      const model = resolveModelId(modelId, availableModels, ctx.model?.provider);
+      const resolutionPool = uokFlags.modelPolicy ? routingEligibleModels : availableModels;
+      const model = resolveModelId(modelId, resolutionPool, ctx.model?.provider);
 
       if (!model) {
         if (verbose) ctx.ui.notify(`Model ${modelId} not found, trying fallback.`, "info");
diff --git a/src/resources/extensions/gsd/tests/auto-model-selection.test.ts b/src/resources/extensions/gsd/tests/auto-model-selection.test.ts
index 1551888d4..7bb1cb7ba 100644
--- a/src/resources/extensions/gsd/tests/auto-model-selection.test.ts
+++ b/src/resources/extensions/gsd/tests/auto-model-selection.test.ts
@@ -227,6 +227,26 @@ test("model change notify in selectAndApplyModel is gated behind verbose flag",
   );
 });
 
+test("model policy resolves candidates from the policy-eligible pool", () => {
+  const src = readFileSync(join(__dirname, "..", "auto-model-selection.ts"), "utf-8");
+  assert.ok(
+    src.includes("const resolutionPool = uokFlags.modelPolicy ? routingEligibleModels : availableModels"),
+    "selectAndApplyModel should resolve model IDs against policy-eligible models when model policy is enabled",
+  );
+});
+
+test("model policy receives task metadata for requirement-vector decisions", () => {
+  const src = readFileSync(join(__dirname, "..", "auto-model-selection.ts"), "utf-8");
+  assert.ok(
+    src.includes("taskMetadata: taskMetadataForPolicy"),
+    "applyModelPolicyFilter should receive task metadata so requirement vectors are unit-aware",
+  );
+  assert.ok(
+    src.includes("extractTaskMetadata(unitId, basePath)"),
+    "execute-task dispatch should derive metadata before policy filtering",
+  );
+});
+
 test("resolveModelId: anthropic wins over claude-code when session provider is not claude-code", () => {
   const availableModels = [
     { id: "claude-sonnet-4-6", provider: "claude-code" },
diff --git a/src/resources/extensions/gsd/tests/uok-model-policy.test.ts b/src/resources/extensions/gsd/tests/uok-model-policy.test.ts
new file mode 100644
index 000000000..dd2b2b93a
--- /dev/null
+++ b/src/resources/extensions/gsd/tests/uok-model-policy.test.ts
@@ -0,0 +1,89 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { mkdtempSync, mkdirSync, readFileSync, rmSync } from "node:fs";
+import { join } from "node:path";
+import { tmpdir } from "node:os";
+
+import {
+  applyModelPolicyFilter,
+  buildRequirementVector,
+} from "../uok/model-policy.ts";
+import {
+  registerToolCompatibility,
+  resetToolCompatibilityRegistry,
+} from "@gsd/pi-coding-agent";
+
+test.afterEach(() => {
+  resetToolCompatibilityRegistry();
+});
+
+test("uok model policy builds requirement vectors from unit metadata", () => {
+  const requirements = buildRequirementVector("execute-task", {
+    tags: ["docs"],
+    fileCount: 8,
+    estimatedLines: 600,
+  });
+
+  assert.equal(requirements.instruction, 0.9);
+  assert.equal(requirements.coding, 0.3);
+  assert.equal(requirements.speed, 0.7);
+});
+
+test("uok model policy enforces provider/api/tool constraints and emits decision audit events", () => {
+  const basePath = mkdtempSync(join(tmpdir(), "gsd-uok-model-policy-"));
+  try {
+    mkdirSync(join(basePath, ".gsd"), { recursive: true });
+    registerToolCompatibility("screenshot", { producesImages: true });
+
+    const result = applyModelPolicyFilter(
+      [
+        { id: "openai-image", provider: "openai", api: "openai-responses" },
+        { id: "anthropic-ok", provider: "anthropic", api: "anthropic-messages" },
+        { id: "gemini-api-deny", provider: "google", api: "google-generative-ai" },
+        { id: "blocked-provider", provider: "blocked", api: "anthropic-messages" },
+      ],
+      {
+        basePath,
+        traceId: "trace-model-policy-1",
+        turnId: "turn-model-policy-1",
+        unitType: "execute-task",
+        taskMetadata: { tags: ["docs"] },
+        allowCrossProvider: true,
+        requiredTools: ["screenshot"],
+        allowedApis: ["anthropic-messages", "openai-responses"],
+        deniedProviders: ["blocked"],
+      },
+    );
+
+    assert.deepEqual(
+      result.eligible.map((m) => m.id),
+      ["anthropic-ok"],
+      "only the policy-compliant anthropic model should remain eligible",
+    );
+    assert.equal(result.decisions.length, 4);
+    assert.equal(result.decisions[0]?.allowed, false);
+    assert.match(result.decisions[0]?.reason ?? "", /tool policy denied/);
+    assert.equal(result.decisions[1]?.allowed, true);
+    assert.equal(result.decisions[2]?.allowed, false);
+    assert.match(result.decisions[2]?.reason ?? "", /transport\/api denied by policy/);
+    assert.equal(result.decisions[3]?.allowed, false);
+    assert.match(result.decisions[3]?.reason ?? "", /provider denied by policy/);
+
+    const auditLogPath = join(basePath, ".gsd", "audit", "events.jsonl");
+    const auditLines = readFileSync(auditLogPath, "utf-8")
+      .trim()
+      .split("\n")
+      .map((line) => JSON.parse(line) as { type: string; payload?: { reason?: string } });
+    const decisionTypes = auditLines.map((event) => event.type);
+
+    assert.equal(auditLines.length, 4);
+    assert.ok(decisionTypes.includes("model-policy-allow"));
+    assert.ok(decisionTypes.includes("model-policy-deny"));
+    assert.ok(
+      auditLines.some((event) => (event.payload?.reason ?? "").includes("tool policy denied")),
+      "audit stream should include explicit deny reasons",
+    );
+  } finally {
+    rmSync(basePath, { recursive: true, force: true });
+  }
+});