fix(headroom): skip unsafe responses tool history (#2132)

Guard openai-responses compression: skip Headroom when body.input
contains non-message items (function_call, function_call_output,
reasoning) to preserve the Responses contract instead of collapsing
them into chat messages.

Co-authored-by: Cursor <cursoragent@cursor.com>

open-sse/rtk/headroom.js

@@ -73,6 +73,14 @@ function maskEndpoint(endpoint) {
   }
 }
 
+function hasUnsafeResponsesInputForCompression(body) {
+  if (!Array.isArray(body?.input)) return false;
+  return body.input.some((item) => {
+    if (!item || typeof item !== "object" || Array.isArray(item)) return false;
+    return typeof item.type === "string" && item.type !== "message";
+  });
+}
+
 // POST messages to Headroom /v1/compress; returns compressed messages + stats or null.
 async function callCompress(url, messages, model, timeoutMs, compressUserMessages, diagnostics) {
   const endpoint = buildCompressEndpoint(url);
@@ -143,6 +151,10 @@ export async function compressWithHeadroom(body, { enabled, url, model, format,
     // messages. Translate input -> OpenAI -> compress -> translate back to input so
     // body.input keeps the Responses contract (the proxy only understands OpenAI). (#1998)
     if (format === "openai-responses") {
+      if (hasUnsafeResponsesInputForCompression(body)) {
+        setDiagnostic(diagnostics, "skipped: openai-responses tool/reasoning input is not safe to compress");
+        return null;
+      }
       const oai = openaiResponsesToOpenAIRequest(model, body, false);
       if (!Array.isArray(oai?.messages)) return null;
       const data = await callCompress(url, oai.messages, model, timeoutMs, compressUserMessages, diagnostics || {});

tests/unit/headroom-responses-format.test.js

@@ -47,4 +47,61 @@ describe("compressWithHeadroom openai-responses format (#1998)", () => {
     expect(Array.isArray(body.input[0].content)).toBe(true);
     expect(typeof body.input[0].content).not.toBe("string");
   });
+
+  it("skips Responses tool/reasoning history instead of collapsing it into a message (#2132)", async () => {
+    global.fetch = vi.fn(async () => ({
+      ok: true,
+      json: async () => ({
+        messages: [{ role: "user", content: "compressed tool history" }],
+        tokens_saved: 10,
+      }),
+    }));
+
+    const input = [
+      {
+        type: "message",
+        role: "user",
+        content: [{ type: "input_text", text: "investigate bug" }],
+      },
+      {
+        type: "function_call",
+        call_id: "call_apply_patch_123",
+        name: "apply_patch",
+        arguments: "*** Begin Patch\n*** End Patch",
+      },
+      {
+        type: "function_call_output",
+        call_id: "call_apply_patch_123",
+        output: "ok",
+      },
+      {
+        type: "reasoning",
+        summary: [{ type: "summary_text", text: "Need a plan" }],
+      },
+    ];
+    const body = {
+      input: structuredClone(input),
+      tools: [
+        {
+          type: "custom",
+          name: "apply_patch",
+          format: { type: "grammar", syntax: "lark", definition: "start: /.+/" },
+        },
+      ],
+    };
+    const diagnostics = {};
+
+    const data = await compressWithHeadroom(body, {
+      enabled: true,
+      url: "http://headroom.test",
+      model: "gpt-5",
+      format: "openai-responses",
+      diagnostics,
+    });
+
+    expect(data).toBeNull();
+    expect(global.fetch).not.toHaveBeenCalled();
+    expect(body.input).toEqual(input);
+    expect(diagnostics.reason).toBe("skipped: openai-responses tool/reasoning input is not safe to compress");
+  });
 });