Add multi-turn tool message probe

Author: xiaoshuo1988130

bin/deepseek-compat-kit.js

@@ -452,6 +452,16 @@ async function probeEndpoint(args) {
     expectStream: true,
   }));
 
+  report.checks.push(await runProbeCheck({
+    name: "multi_turn_tool_messages",
+    capability: "multi_turn_tool_messages",
+    description: "Endpoint accepts a follow-up request containing assistant tool_calls, reasoning_content, and tool results.",
+    impact: "Multi-turn tool-calling agents can pass DeepSeek reasoning_content back alongside tool results.",
+    recommendation: "If this warns or fails, confirm that the framework preserves reasoning_content and that the provider accepts DeepSeek tool-call message history.",
+    request: buildMultiTurnToolProbeRequest(model),
+    baseUrl,
+  }));
+
   report.checks.push(await runProbeCheck({
     name: "strict_schema_request",
     capability: "strict_schema",
@@ -609,6 +619,53 @@ function buildProbeRequest({ model, stream }) {
   };
 }
 
+function buildMultiTurnToolProbeRequest(model) {
+  return {
+    model,
+    messages: [
+      { role: "user", content: "Use the weather tool for Paris." },
+      {
+        role: "assistant",
+        content: null,
+        reasoning_content: "I need to call the weather tool before answering.",
+        tool_calls: [{
+          id: "call_probe_weather",
+          type: "function",
+          function: {
+            name: "get_weather",
+            arguments: "{\"city\":\"Paris\"}",
+          },
+        }],
+      },
+      {
+        role: "tool",
+        tool_call_id: "call_probe_weather",
+        content: "{\"city\":\"Paris\",\"weather\":\"sunny\"}",
+      },
+      { role: "user", content: "Reply with exactly: ok" },
+    ],
+    tools: [{
+      type: "function",
+      function: {
+        name: "get_weather",
+        description: "Return simple weather data for a city.",
+        parameters: {
+          type: "object",
+          properties: {
+            city: {
+              type: "string",
+              description: "City name.",
+            },
+          },
+          required: ["city"],
+          additionalProperties: false,
+        },
+      },
+    }],
+    max_tokens: 8,
+  };
+}
+
 function buildStrictSchemaProbeRequest(model) {
   return {
     model,

docs/capability-probe.md

@@ -63,6 +63,7 @@ Use `--profile` to tune report guidance:
 | --- | --- | --- |
 | `chat_completions` | Minimal non-streaming `POST /chat/completions` | Verifies the basic OpenAI-compatible request path. |
 | `streaming` | `stream: true` with event-stream-like response | Verifies whether streaming clients can parse incremental responses. |
+| `multi_turn_tool_messages` | Follow-up request containing assistant `tool_calls`, `reasoning_content`, and a matching `tool` result | Verifies whether Agent loops can pass DeepSeek reasoning content back through multi-turn tool-call history. |
 | `strict_schema` | Minimal strict tool schema request | Verifies whether tool-calling agents can send DeepSeek strict-mode compatible schemas. |
 
 ## Reading the Report
@@ -92,6 +93,12 @@ If `streaming` warns or fails:
 - Check whether a relay buffers responses instead of returning `text/event-stream`.
 - Re-run the probe after changing provider settings.
 
+If `multi_turn_tool_messages` warns or fails:
+
+- Confirm that the framework preserves `reasoning_content` from the previous assistant tool-call turn.
+- Confirm that the provider accepts assistant messages containing both `tool_calls` and `reasoning_content`.
+- Re-run the same probe directly against the official DeepSeek endpoint to separate framework, relay, and self-hosted endpoint behavior.
+
 If `strict_schema` warns or fails:
 
 - Run `compile-schema --dry-run` on generated tool schemas.

docs/getting-started.md

@@ -24,7 +24,7 @@ You can also run a small functional endpoint probe against a mock or self-hosted
 npx deepseek-compat-kit probe --endpoint http://127.0.0.1:9000 --model mock-model --out ./capability-report.json --markdown ./Capability_Report.md
 ```
 
-`probe` checks basic chat completions, streaming response shape, and a minimal strict tool schema request. It is not a benchmark or load test. The Markdown report is designed for team handoff or upstream issue triage.
+`probe` checks basic chat completions, streaming response shape, multi-turn tool-call message history with `reasoning_content`, and a minimal strict tool schema request. It is not a benchmark or load test. The Markdown report is designed for team handoff or upstream issue triage.
 
 Then use the real proxy against DeepSeek:
 

test/cli.test.js

@@ -177,8 +177,10 @@ test("compile-schema dry-run prints repair plan without writing files", () => {
 });
 
 test("probe writes endpoint capability report against mock upstream", async (t) => {
+  const requestBodies = [];
   const mock = http.createServer((request, response) => {
     collectRequestJson(request).then((body) => {
+      requestBodies.push(body);
       const pathname = new URL(request.url, "http://127.0.0.1").pathname;
       if (request.method !== "POST" || pathname !== "/chat/completions") {
         response.writeHead(404, { "content-type": "application/json" });
@@ -232,30 +234,37 @@ test("probe writes endpoint capability report against mock upstream", async (t)
   assert.equal(report.profile_guidance.name, "Third-party relay or API gateway");
   assert.match(report.profile_guidance.strict_schema_hint, /relay preserves DeepSeek strict schema semantics/);
   assert.equal(report.summary.status, "PASS");
-  assert.equal(report.summary.passed, 3);
+  assert.equal(report.summary.passed, 4);
   assert.deepEqual(report.summary.capabilities, {
     chat_completions: "PASS",
     streaming: "PASS",
+    multi_turn_tool_messages: "PASS",
     strict_schema: "PASS",
   });
   assert.deepEqual(report.checks.map((check) => check.name), [
     "chat_completions",
     "streaming",
+    "multi_turn_tool_messages",
     "strict_schema_request",
   ]);
   assert.deepEqual(report.checks.map((check) => check.capability), [
     "chat_completions",
     "streaming",
+    "multi_turn_tool_messages",
     "strict_schema",
   ]);
-  assert.match(report.checks[2].recommendation, /compile-schema/);
+  assert.match(report.checks[2].recommendation, /reasoning_content/);
+  assert.match(report.checks[3].recommendation, /compile-schema/);
+  assert.ok(requestBodies.some((body) => body.messages?.some((message) => message.reasoning_content)));
+  assert.ok(requestBodies.some((body) => body.messages?.some((message) => message.role === "tool" && message.tool_call_id === "call_probe_weather")));
 
   const markdown = fs.readFileSync(markdownPath, "utf8");
   assert.match(markdown, /# DeepSeek CompatKit Capability Report/);
   assert.match(markdown, /## Profile Guidance/);
   assert.match(markdown, /Third-party relay or API gateway/);
   assert.match(markdown, /Status: \*\*PASS\*\*/);
   assert.match(markdown, /\| `chat_completions` \| `chat_completions` \| PASS \| 200 \|/);
+  assert.match(markdown, /\| `multi_turn_tool_messages` \| `multi_turn_tool_messages` \| PASS \| 200 \|/);
   assert.match(markdown, /## Recommendations/);
   assert.match(markdown, /No immediate compatibility issues/);
 });