fix: stop forcing native ollama num_ctx

steipete · steipete · commit a8c745a6230e · 2026-05-10T09:34:36.000+01:00
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -20,6 +20,7 @@ Docs: https://docs.openclaw.ai
 
 ### Fixes
 
+- Ollama: stop native `/api/chat` requests from copying catalog `contextWindow` or `maxTokens` into `options.num_ctx` unless `params.num_ctx` is explicitly configured, avoiding pathological prompt-ingestion latency on local large-context models. Fixes #62267. Thanks @BenSHPD.
 - Ollama: keep the model idle watchdog enabled for `*:cloud` models routed through a local Ollama host, so cloud-backed tool-loop stalls fail over visibly instead of inheriting local-model no-idle behavior. Fixes #79350. Thanks @geek111.
 - Voice/Ollama: honor routed voice agent `tools.allow` for classic embedded voice responses, including empty allowlists, so no-tool Ollama agents do not receive tool schemas. Fixes #79506. Thanks @donkeykong91.
 - Gateway: reread config from disk after the first in-process restart loop startup, preventing SIGUSR1 restarts from reusing a stale startup snapshot and dropping config written after boot. Fixes #79947. Thanks @TheLevti.
diff --git a/extensions/ollama/src/stream-runtime.test.ts b/extensions/ollama/src/stream-runtime.test.ts
@@ -208,7 +208,7 @@ describe("createConfiguredOllamaCompatStreamWrapper", () => {
         };
         expect(requestBody.think).toBe(false);
         expect(requestBody.options?.think).toBeUndefined();
-        expect(requestBody.options?.num_ctx).toBe(131072);
+        expect(requestBody.options?.num_ctx).toBeUndefined();
       },
     );
   });
@@ -310,7 +310,7 @@ describe("createConfiguredOllamaCompatStreamWrapper", () => {
         };
         expect(requestBody.think).toBe("low");
         expect(requestBody.options?.think).toBeUndefined();
-        expect(requestBody.options?.num_ctx).toBe(131072);
+        expect(requestBody.options?.num_ctx).toBeUndefined();
       },
     );
   });
@@ -405,7 +405,7 @@ describe("createConfiguredOllamaCompatStreamWrapper", () => {
         };
         expect(requestBody.think).toBe("high");
         expect(requestBody.options?.think).toBeUndefined();
-        expect(requestBody.options?.num_ctx).toBe(131072);
+        expect(requestBody.options?.num_ctx).toBeUndefined();
       },
     );
   });
@@ -1609,9 +1609,7 @@ describe("createOllamaStreamFn", () => {
         if (!requestBody.options) {
           throw new Error("Expected Ollama request options");
         }
-        // Catalog `contextWindow` flows through as `num_ctx` so the request
-        // does not silently truncate to Ollama's small Modelfile default.
-        expect(requestBody.options?.num_ctx).toBe(131072);
+        expect(requestBody.options?.num_ctx).toBeUndefined();
         expect(requestBody.options.num_predict).toBe(123);
       },
     );
@@ -1694,7 +1692,7 @@ describe("createOllamaStreamFn", () => {
     );
   });
 
-  it("falls back to catalog contextWindow as num_ctx when params.num_ctx is unset", async () => {
+  it("does not fall back to catalog contextWindow as native Ollama num_ctx", async () => {
     await withMockNdjsonFetch(
       [
         '{"model":"m","created_at":"t","message":{"role":"assistant","content":"ok"},"done":false}',
@@ -1715,12 +1713,12 @@ describe("createOllamaStreamFn", () => {
         const requestBody = JSON.parse(requestInit.body) as {
           options?: { num_ctx?: number };
         };
-        expect(requestBody.options?.num_ctx).toBe(32768);
+        expect(requestBody.options?.num_ctx).toBeUndefined();
       },
     );
   });
 
-  it("falls back to catalog maxTokens as num_ctx when contextWindow is absent", async () => {
+  it("does not fall back to catalog maxTokens as native Ollama num_ctx", async () => {
     await withMockNdjsonFetch(
       [
         '{"model":"m","created_at":"t","message":{"role":"assistant","content":"ok"},"done":false}',
@@ -1744,7 +1742,7 @@ describe("createOllamaStreamFn", () => {
         const requestBody = JSON.parse(requestInit.body) as {
           options?: { num_ctx?: number };
         };
-        expect(requestBody.options?.num_ctx).toBe(65536);
+        expect(requestBody.options?.num_ctx).toBeUndefined();
       },
     );
   });
diff --git a/extensions/ollama/src/stream.ts b/extensions/ollama/src/stream.ts
@@ -293,29 +293,16 @@ function resolveOllamaNumCtx(model: ProviderRuntimeModel): number {
 /**
  * Resolves num_ctx for native /api/chat requests:
  *  1. explicit `params.num_ctx` set on the model wins,
- *  2. otherwise the catalog `contextWindow` / `maxTokens` is forwarded so
- *     OpenClaw's known model windows survive the trip and `/api/chat` does
- *     not silently truncate to Ollama's small Modelfile default (typically
- *     2048 tokens) — which is too small for a system prompt plus tool
- *     definitions and produces "model picks wrong tools / says nonsense"
- *     symptoms on agent turns,
- *  3. when neither is known, return undefined so the Modelfile decides.
+ *  2. otherwise return undefined so Ollama's model, OLLAMA_CONTEXT_LENGTH,
+ *     VRAM, or Modelfile policy decides.
  *
  * This intentionally differs from `resolveOllamaNumCtx` by not falling back
  * to `DEFAULT_CONTEXT_TOKENS`: that constant is a sane wrapper-side guess for
- * the OpenAI-compat path, but on the native path we prefer to leave num_ctx
- * absent rather than guess a window for an unknown model.
+ * the OpenAI-compat path, but native `/api/chat` should not force the full
+ * advertised catalog context for local models unless the operator opted in.
  */
 function resolveOllamaNativeNumCtx(model: ProviderRuntimeModel): number | undefined {
-  const configured = resolveOllamaConfiguredNumCtx(model);
-  if (configured !== undefined) {
-    return configured;
-  }
-  const catalog = model.contextWindow ?? model.maxTokens;
-  if (typeof catalog === "number" && Number.isFinite(catalog) && catalog > 0) {
-    return Math.floor(catalog);
-  }
-  return undefined;
+  return resolveOllamaConfiguredNumCtx(model);
 }
 
 function resolveOllamaModelOptions(model: ProviderRuntimeModel): Record<string, unknown> {
