feat(dsh): add Responses-API route for TokenPlan Qwen models and enhance event text handling

packages/dsh/cordis.patch.yml

@@ -69,6 +69,54 @@
             compat:
               thinkingFormat: deepseek
 
+      # Responses-API route for the TokenPlan Qwen thinking models: same
+      # gateway and key as bailian-tokenplan, but OpenAI Responses protocol
+      # (`/responses`) instead of chat/completions. Reasoning arrives as
+      # native `reasoning` output items; the gateway's
+      # `response.reasoning_text.delta` SSE event is one pi-ai's
+      # openai-responses adapter explicitly consumes. The three models stay
+      # listed on the completions provider above, so both routes remain
+      # selectable; delete them there to make Responses the only route.
+      #
+      # No `thinkingFormat: qwen` here — that compat maps the completions
+      # `reasoning_content` field, pi-ai types it only for
+      # openai-completions, and resolution rejects such switches on this
+      # protocol. Reasoning effort instead goes through each model's
+      # `reasoningEfforts`, whose wire spellings land in the gateway's
+      # `reasoning.effort` request parameter. The four levels below were the
+      # only ones probed; xhigh/max were not offered to the gateway.
+      #
+      # Verified on 2026-08-15 per model against /responses: non-stream and
+      # stream bodies, one function call, and a colour question about a test
+      # PNG. qwen3.7-max rejects image content with HTTP 400, so it carries
+      # no `input: [text, image]`.
+      bailian-tokenplan-responses:
+        displayName: Aliyun Bailian TokenPlan (Responses)
+        api: openai-responses
+        baseURL: https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1
+        apiKeyEnv: BAILIAN_TOKENPLAN_API_KEY
+        models:
+          - id: qwen3.8-max
+            input: [text, image]
+            reasoningEfforts:
+              minimal: minimal
+              low: low
+              medium: medium
+              high: high
+          - id: qwen3.7-plus
+            input: [text, image]
+            reasoningEfforts:
+              minimal: minimal
+              low: low
+              medium: medium
+              high: high
+          - id: qwen3.7-max
+            reasoningEfforts:
+              minimal: minimal
+              low: low
+              medium: medium
+              high: high
+
 - insert:
     - id: bailian-tool-vision
       name: bailian-cli-dsh/tool-vision

packages/dsh/src/tool-managed-agent/index.ts

@@ -128,11 +128,36 @@ interface SessionRunResponse {
   events?: readonly { type?: string; content?: unknown; role?: string }[];
 }
 
-/** Assistant-visible text of a finished remote session. */
+/**
+ * Text of one sanitized envelope event. `bl --output json` emits the SDK's
+ * sanitized `SessionEvent` shape, where `content` is an ARRAY of content
+ * blocks (`[{ type: "text", text }]`) — never a plain string — and `type` is
+ * the provider's raw event type. Tolerate a legacy string `content` too.
+ */
+function eventText(content: unknown): string {
+  if (typeof content === "string") return content;
+  if (!Array.isArray(content)) return "";
+  return content
+    .map((block) =>
+      block !== null &&
+      typeof block === "object" &&
+      typeof (block as { text?: unknown }).text === "string"
+        ? (block as { text: string }).text
+        : "",
+    )
+    .join("");
+}
+
+/**
+ * Assistant-visible text of a finished remote session. The envelope echoes
+ * the user prompt as `type: "message", role: "user"`, so keep only non-user
+ * message events and join their text blocks.
+ */
 function assistantText(response: SessionRunResponse): string {
   return (response.events ?? [])
-    .filter((event) => event.type === "message" && typeof event.content === "string")
-    .map((event) => event.content as string)
+    .filter((event) => event.type === "message" && event.role !== "user")
+    .map((event) => eventText(event.content))
+    .filter((text) => text.length > 0)
     .join("\n")
     .trim();
 }