fix(core): honor the -1 sentinels for max_turns and max_tokens (#56)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

packages/core/test/agent.test.ts

@@ -78,6 +78,7 @@ describe("metaMaxTokens (meta-request budget tightened by the per-model cap)", (
     expect(metaMaxTokens(300, 8000)).toBe(300); // ample cap: the small budget stays
     expect(metaMaxTokens(300, 128)).toBe(128); // pinned below the budget: the cap binds
     expect(metaMaxTokens(2048, 1024)).toBe(1024); // vision-describer budget, same rule
+    expect(metaMaxTokens(300, -1)).toBe(300); // -1 = uncapped: never tightens to -1 (issue #55 sibling)
   });
 });
 

packages/core/test/engine.test.ts

@@ -331,6 +331,47 @@ describe("ContextEngine ReAct loop (mock LLM, approve callback)", () => {
     ).toBe(true);
   });
 
+  it("maxTurns -1 removes the cap instead of stopping before the first turn (issue #55)", async () => {
+    // Two tool-call turns followed by a final text turn: with the old `0 >= -1` guard the
+    // engine emitted the stop note without ever calling the LLM.
+    let calls = 0;
+    const llm: LLMInterface = {
+      async *streamGenerate() {
+        calls += 1;
+        if (calls <= 2) {
+          yield toolCall({
+            name: "exec_command",
+            arguments: JSON.stringify({ cmd: "true" }),
+            toolCallId: `c${calls}`,
+            stopReason: "completed",
+          });
+        } else {
+          yield assistantText("Done");
+        }
+        yield tokenUsage(emptyTokenCounts(), {
+          cache_read: 0,
+          cache_write: 0,
+          output: 1,
+          total: 1,
+        });
+        return { status: "completed" };
+      },
+    };
+    const environment = new Environment({
+      workspaceDir: workspace,
+      toolConfig: execCommandToolConfig(),
+    });
+    const engine = new ContextEngine({ llm, environment, maxTurns: -1 });
+
+    const all = await collectRun(engine, [userText("go")], allowAll);
+    expect(calls).toBe(3);
+    const texts = all
+      .filter((m) => isCompleteModelMessage(m) && m.payload.type === "text")
+      .map((m) => (m.payload as TextPayload).text);
+    expect(texts.some((t) => t.includes("reached max turns"))).toBe(false);
+    expect(texts.some((t) => t === "Done")).toBe(true);
+  });
+
   it("max turns with pending tool outputs carries them over so the next run pairs the tool_call (issue #33)", async () => {
     const received: OmniMessage[][] = [];
     const llm: LLMInterface = {

packages/core/test/llm.test.ts

@@ -1052,6 +1052,11 @@ describe("config helpers", () => {
     expect("system_prompt" in minimal).toBe(false);
     expect("max_tokens" in minimal).toBe(false);
     expect("thinking_level" in minimal).toBe(false);
+
+    // max_tokens -1 (the config's "no cap" sentinel) stays OFF the wire — sent literally,
+    // providers reject a negative max_tokens with a 400.
+    const uncapped = buildUniConfig({ modelId: "m", tools: [], maxTokens: -1 });
+    expect("max_tokens" in uncapped).toBe(false);
   });
 
   it("omits tools when empty and never sets tool_choice (strict endpoints reject both)", () => {