fix(autocomplete): cap reserved output tokens in pruneLength

core/autocomplete/templating/__tests__/renderPrompt.vitest.ts

@@ -26,12 +26,14 @@ vi.mock("../../../llm/countTokens", () => {
   const pruneLinesFromBottom = (str: string, allowed: number) =>
     str.slice(0, allowed);
   const getTokenCountingBufferSafety = () => 0;
+  const MIN_RESPONSE_TOKENS = 1000;
 
   return {
     countTokens,
     pruneLinesFromTop,
     pruneLinesFromBottom,
     getTokenCountingBufferSafety,
+    MIN_RESPONSE_TOKENS,
   };
 });
 
@@ -254,6 +256,31 @@ describe("renderPromptWithTokenLimit parity & pruning", () => {
 
     expect(compiledPrefix.length).toBeLessThan(120);
   });
+
+  it("does not wipe a small prompt when default maxTokens exceeds a small context length", () => {
+    // Regression test: a model whose contextLength (e.g. Ollama's num_ctx) is smaller
+    // than the default maxTokens reservation (4096) used to compute a negative prompt
+    // budget, which caused every prefix/suffix to be pruned to nothing regardless of
+    // how small the actual prompt was.
+    const shortPrefix = "A".repeat(50);
+
+    const helper = makeHelper({ prunedPrefix: shortPrefix });
+
+    const llmStub = {
+      contextLength: 2048,
+      completionOptions: {},
+      model: "test-model",
+    } as any;
+
+    const { prefix: compiledPrefix } = renderPromptWithTokenLimit({
+      snippetPayload: emptySnippetPayload,
+      workspaceDirs: ["file:///workspace"],
+      helper,
+      llm: llmStub,
+    });
+
+    expect(compiledPrefix.includes(shortPrefix)).toBe(true);
+  });
 });
 
 describe("stop-token merging", () => {

core/autocomplete/templating/index.ts

@@ -9,6 +9,7 @@ import { DEFAULT_MAX_TOKENS } from "../../llm/constants.js";
 import {
   countTokens,
   getTokenCountingBufferSafety,
+  MIN_RESPONSE_TOKENS,
   pruneLinesFromBottom,
   pruneLinesFromTop,
 } from "../../llm/countTokens";
@@ -204,7 +205,12 @@ function pruneLength(llm: ILLM, prompt: string): number {
   const contextLength = llm.contextLength;
   const reservedTokens = llm.completionOptions.maxTokens ?? DEFAULT_MAX_TOKENS;
   const safetyBuffer = getTokenCountingBufferSafety(contextLength);
-  const maxAllowedPromptTokens = contextLength - reservedTokens - safetyBuffer;
+  // Reserve at most MIN_RESPONSE_TOKENS for the completion, same as compileChatMessages.
+  // Otherwise a large maxTokens (the default is 4096) can consume the entire context
+  // window on small-context models, leaving no room for the prompt and silently
+  // pruning it to nothing on every request.
+  const minOutputTokens = Math.min(MIN_RESPONSE_TOKENS, reservedTokens);
+  const maxAllowedPromptTokens = contextLength - minOutputTokens - safetyBuffer;
   const promptTokenCount = countTokens(prompt, llm.model);
   return promptTokenCount - maxAllowedPromptTokens;
 }

core/llm/countTokens.ts

@@ -562,6 +562,7 @@ export {
   countTokens,
   countTokensAsync,
   extractToolSequence,
+  MIN_RESPONSE_TOKENS,
   pruneLinesFromBottom,
   pruneLinesFromTop,
   pruneRawPromptFromTop,