diff --git a/@omniroute/opencode-provider/src/index.ts b/@omniroute/opencode-provider/src/index.ts
index 4e2f070a19..454f45012c 100644
--- a/@omniroute/opencode-provider/src/index.ts
+++ b/@omniroute/opencode-provider/src/index.ts
@@ -78,6 +78,20 @@ export interface ModelCapabilities {
   tool_call?: boolean;
 }
 
+/**
+ * Default per-model context window sizes (tokens) for the curated default catalog.
+ * Matches the context lengths used by OmniRoute's provider registry.
+ */
+export const OMNIROUTE_DEFAULT_MODEL_CONTEXT_LENGTHS: Record<string, number> = {
+  "cc/claude-opus-4-7": 200_000,
+  "cc/claude-sonnet-4-6": 200_000,
+  "cc/claude-haiku-4-5-20251001": 200_000,
+  "claude-opus-4-5-thinking": 200_000,
+  "claude-sonnet-4-5-thinking": 200_000,
+  "gemini-3.1-pro-high": 1_000_000,
+  "gemini-3-flash": 1_000_000,
+};
+
 /**
  * Default per-model capability hints for the curated default catalog.
  *
@@ -141,6 +155,19 @@ export interface OpenCodeModelEntry {
   reasoning?: boolean;
   temperature?: boolean;
   tool_call?: boolean;
+  /**
+   * Context window limit. OpenCode reads this to determine usable context
+   * length for compaction, overflow detection, and router decisions.
+   * Maps to `limit.context` in OpenCode's provider config schema.
+   */
+  limit?: {
+    /** Maximum context length in tokens (e.g. 200000 for Claude, 1000000 for Gemini). */
+    context: number;
+    /** Optional per-request max input tokens. */
+    input?: number;
+    /** Optional max output tokens. */
+    output?: number;
+  };
 }
 
 export interface OpenCodeProviderEntry {
@@ -235,6 +262,14 @@ export function createOmniRouteProvider(options: OmniRouteProviderOptions): Open
     if (typeof merged.reasoning === "boolean") entry.reasoning = merged.reasoning;
     if (typeof merged.temperature === "boolean") entry.temperature = merged.temperature;
     if (typeof merged.tool_call === "boolean") entry.tool_call = merged.tool_call;
+
+    // Include context window limit when known — OpenCode reads this to
+    // determine usable context length for compaction & overflow detection.
+    const contextLength = OMNIROUTE_DEFAULT_MODEL_CONTEXT_LENGTHS[id];
+    if (typeof contextLength === "number" && contextLength > 0) {
+      entry.limit = { context: contextLength };
+    }
+
     models[id] = entry;
   }
 
@@ -445,6 +480,8 @@ async function fetchJSON<T>(url: string, apiKey: string, timeoutMs: number): Pro
 export interface OmniRouteLiveModel {
   id: string;
   name: string;
+  /** Context window length in tokens (e.g. 200000 for Claude, 1000000 for Gemini). */
+  contextLength?: number;
 }
 
 /**
@@ -514,7 +551,18 @@ export async function fetchLiveModels(
             ? r.display_name.trim()
             : id;
 
-    models.push({ id, name: name || id });
+    // Extract context_length from OmniRoute's /v1/models response.
+    // OmniRoute returns context_length in snake_case for both synced
+    // models (with inputTokenLimit) and custom models; the catalog's
+    // getDefaultContextFallback also injects it from registry defaults.
+    const contextLength =
+      typeof r.context_length === "number" && r.context_length > 0
+        ? r.context_length
+        : typeof r.max_context_window_tokens === "number" && r.max_context_window_tokens > 0
+          ? r.max_context_window_tokens
+          : undefined;
+
+    models.push({ id, name: name || id, ...(contextLength ? { contextLength } : {}) });
   }
 
   return models;
diff --git a/@omniroute/opencode-provider/tests/index.test.ts b/@omniroute/opencode-provider/tests/index.test.ts
index 324ae68bfa..9e14ba3bff 100644
--- a/@omniroute/opencode-provider/tests/index.test.ts
+++ b/@omniroute/opencode-provider/tests/index.test.ts
@@ -15,6 +15,7 @@ import {
   mergeIntoExistingConfig,
   normalizeBaseURL,
   OMNIROUTE_DEFAULT_MODEL_CAPABILITIES,
+  OMNIROUTE_DEFAULT_MODEL_CONTEXT_LENGTHS,
   OMNIROUTE_DEFAULT_OPENCODE_MODELS,
   OMNIROUTE_MCP_DEFAULT_SCOPES,
   OMNIROUTE_PROVIDER_NPM,
@@ -394,6 +395,77 @@ test("OMNIROUTE_DEFAULT_OPENCODE_MODELS includes cc/ prefixed models", () => {
   assert.ok(defaults.length >= 7, "should have at least 7 models");
 });
 
+test("OMNIROUTE_DEFAULT_MODEL_CONTEXT_LENGTHS covers every default model id", () => {
+  for (const id of OMNIROUTE_DEFAULT_OPENCODE_MODELS) {
+    const ctx = OMNIROUTE_DEFAULT_MODEL_CONTEXT_LENGTHS[id];
+    assert.ok(
+      typeof ctx === "number" && ctx > 0,
+      `default context_length for ${id} missing — should be a positive number`
+    );
+    // Sanity: context should be at least 8K, at most 2M tokens
+    assert.ok(ctx >= 8_000, `${id} context_length ${ctx} seems too low`);
+    assert.ok(ctx <= 2_000_000, `${id} context_length ${ctx} seems too high`);
+  }
+});
+
+test("createOmniRouteProvider emits limit.context on default model entries", () => {
+  const provider = createOmniRouteProvider({
+    baseURL: "http://localhost:20128",
+    apiKey: "sk_omniroute",
+  });
+  const entry = provider.models["cc/claude-opus-4-7"];
+  assert.ok(entry.limit, "model entry should have a limit field");
+  assert.equal(entry.limit!.context, 200_000);
+});
+
+test("createOmniRouteProvider omits limit.context for unknown model ids", () => {
+  const provider = createOmniRouteProvider({
+    baseURL: "http://localhost:20128",
+    apiKey: "sk_omniroute",
+    models: ["completely-unknown-model"],
+  });
+  const entry = provider.models["completely-unknown-model"];
+  assert.equal(entry.limit, undefined);
+});
+
+test("createOmniRouteProvider serialises limit.context to JSON", () => {
+  const provider = createOmniRouteProvider({
+    baseURL: "http://localhost:20128",
+    apiKey: "sk_omniroute",
+  });
+  const round = JSON.parse(JSON.stringify(provider));
+  for (const id of OMNIROUTE_DEFAULT_OPENCODE_MODELS) {
+    const expectedContext = OMNIROUTE_DEFAULT_MODEL_CONTEXT_LENGTHS[id];
+    assert.equal(
+      round.models[id].limit?.context,
+      expectedContext,
+      `${id} should serialise limit.context=${expectedContext}`
+    );
+  }
+});
+
+test("fetchLiveModels extracts context_length from snake_case field", async () => {
+  const { url, close } = await startMockServer(() => ({
+    data: [
+      { id: "cc/claude-opus-4-7", name: "Claude Opus 4.7", context_length: 200_000 },
+      { id: "gemini-3.1-pro-high", name: "Gemini 3.1 Pro", context_length: 1_000_000 },
+      { id: "no-context", name: "No Context" },
+    ],
+  }));
+  try {
+    const models = await fetchLiveModels(url, "sk_test");
+    const claude = models.find((m) => m.id === "cc/claude-opus-4-7");
+    assert.ok(claude, "claude model should be present");
+    assert.equal(claude!.contextLength, 200_000);
+    const gemini = models.find((m) => m.id === "gemini-3.1-pro-high");
+    assert.equal(gemini!.contextLength, 1_000_000);
+    const noCtx = models.find((m) => m.id === "no-context");
+    assert.equal(noCtx!.contextLength, undefined);
+  } finally {
+    close();
+  }
+});
+
 test("OMNIROUTE_DEFAULT_MODEL_CAPABILITIES covers every default model id", () => {
   for (const id of OMNIROUTE_DEFAULT_OPENCODE_MODELS) {
     const caps = OMNIROUTE_DEFAULT_MODEL_CAPABILITIES[id];