@quilltap/plugin-utils 2.2.20 → 2.4.0

package/dist/index.mjs

@@ -352,6 +352,36 @@ function buildRequestAbortSignal(params, defaultMs = DEFAULT_REQUEST_TIMEOUT_MS)
   return AbortSignal.timeout(resolveRequestTimeoutMs(params, defaultMs));
 }
 
+// src/providers/profile-parameters.ts
+function applyProfileParameters(body, params, allowlist, normalize) {
+  const profile = params.profileParameters;
+  if (!profile || typeof profile !== "object") return;
+  const bag = profile;
+  for (const key of allowlist) {
+    const raw = bag[key];
+    if (raw === void 0 || raw === null) continue;
+    if (typeof raw === "string" && raw === "") continue;
+    const value = normalize ? normalize(key, raw, params, body) : raw;
+    if (value === void 0) continue;
+    body[key] = value;
+  }
+}
+
+// src/providers/system-messages.ts
+function collapseLeadingSystemMessages(messages) {
+  let runLength = 0;
+  while (runLength < messages.length && messages[runLength].role === "system") {
+    runLength++;
+  }
+  if (runLength < 2) return messages;
+  const run = messages.slice(0, runLength);
+  const merged = {
+    ...run[0],
+    content: run.map((m) => m.content ?? "").filter((c) => c.length > 0).join("\n\n")
+  };
+  return [merged, ...messages.slice(runLength)];
+}
+
 // src/providers/openai-compatible.ts
 var DEFAULT_MAX_RETRIES = 2;
 var OpenAICompatibleProvider = class {
@@ -367,6 +397,28 @@ var OpenAICompatibleProvider = class {
     this.supportedMimeTypes = [];
     /** Web search is not supported */
     this.supportsWebSearch = false;
+    /**
+     * Whether the endpoint tolerates more than one leading `system` message.
+     *
+     * True for every hosted service — and so the default, which keeps the
+     * outbound bytes of OpenRouter, Z.AI and every other subclass exactly as they
+     * were. A subclass that points at a local runtime applying the *model's* chat
+     * template sets this false: several template families reject a system message
+     * anywhere but index 0, and Quilltap's context builder deliberately emits up
+     * to three of them so its cache breakpoints survive. See Bug 82 and
+     * {@link collapseLeadingSystemMessages}.
+     */
+    this.acceptsRepeatedSystemMessages = true;
+    /**
+     * Profile-parameter keys this provider forwards from
+     * `LLMParams.profileParameters` onto the request body.
+     *
+     * **Empty by default**, so a subclass that declares nothing sends exactly the
+     * bytes it sent before this mechanism existed. Subclasses opt in by
+     * overriding with their own allow-list; `model`, `messages`, `stream`,
+     * `stream_options` and `tools` must never appear in one.
+     */
+    this.profileParamAllowlist = [];
     if (typeof config === "string") {
       this.baseUrl = config;
       this.providerName = "OpenAICompatible";
@@ -384,6 +436,14 @@ var OpenAICompatibleProvider = class {
     }
     this.logger = createPluginLogger(`${this.providerName}Provider`);
   }
+  /**
+   * Apply {@link acceptsRepeatedSystemMessages} to a mapped wire-message array.
+   * Called by both body builds so the streaming and non-streaming shapes cannot
+   * drift apart.
+   */
+  applySystemMessagePolicy(messages) {
+    return this.acceptsRepeatedSystemMessages ? messages : collapseLeadingSystemMessages(messages);
+  }
   /**
    * Collects attachment failures for messages with attachments.
    * Since attachments are not supported, all attachments are marked as failed.
@@ -454,6 +514,53 @@ var OpenAICompatibleProvider = class {
   buildRequestOptions(params) {
     return buildSdkRequestOptions(params);
   }
+  /**
+   * Per-key hook for allow-listed profile parameters. The default forwards the
+   * stored value verbatim.
+   *
+   * Override to reshape a value (the editor stores flat strings where some APIs
+   * want objects), to gate a key on the model, or — by writing into `body` and
+   * returning `undefined` — to send it under a different key entirely.
+   */
+  normalizeProfileParam(_key, value, _params, _body) {
+    return value;
+  }
+  /**
+   * Apply {@link profileParamAllowlist} to a request body under construction.
+   * Called by both body builds so the streaming and non-streaming shapes cannot
+   * drift apart.
+   */
+  applyProfileParams(body, params) {
+    applyProfileParameters(
+      body,
+      params,
+      this.profileParamAllowlist,
+      (key, value, callParams, target) => this.normalizeProfileParam(key, value, callParams, target)
+    );
+  }
+  /**
+   * Normalize an OpenAI-shaped `tool_calls` array into Quilltap's `ToolCall`
+   * shape, tolerating servers that hand back already-parsed argument objects
+   * where the OpenAI wire format specifies a JSON string.
+   */
+  normalizeToolCalls(toolCalls) {
+    if (!toolCalls || toolCalls.length === 0) return [];
+    const out = [];
+    for (const raw of toolCalls) {
+      const tc = raw;
+      if (!tc?.function?.name) continue;
+      const args = tc.function.arguments;
+      out.push({
+        id: tc.id ?? "",
+        type: "function",
+        function: {
+          name: tc.function.name,
+          arguments: typeof args === "string" ? args : JSON.stringify(args ?? {})
+        }
+      });
+    }
+    return out;
+  }
   /**
    * Sends a message and returns the complete response.
    *
@@ -465,7 +572,7 @@ var OpenAICompatibleProvider = class {
     this.validateApiKeyRequirement(apiKey);
     const attachmentResults = this.collectAttachmentFailures(params);
     const client = this.createClient(apiKey);
-    const messages = params.messages.filter((m) => {
+    const mappedMessages = params.messages.filter((m) => {
       if (m.role === "tool" && !m.toolCallId) return false;
       return true;
     }).map((m) => {
@@ -492,17 +599,27 @@ var OpenAICompatibleProvider = class {
         content: m.content
       };
     });
+    const messages = this.applySystemMessagePolicy(mappedMessages);
+    const body = {
+      model: params.model,
+      messages,
+      temperature: params.temperature ?? 0.7,
+      max_tokens: params.maxTokens ?? 4096,
+      top_p: params.topP ?? 1,
+      stop: params.stop,
+      ...params.cacheKey ? { user: params.cacheKey } : {},
+      // Tools arrive only once the runtime gate (the profile's "Allow tool
+      // use") has passed, so no separate guard is needed here.
+      ...params.tools && params.tools.length > 0 ? { tools: params.tools, tool_choice: params.toolChoice ?? "auto" } : {}
+    };
+    this.applyProfileParams(body, params);
     try {
-      const response = await client.chat.completions.create({
-        model: params.model,
-        messages,
-        temperature: params.temperature ?? 0.7,
-        max_tokens: params.maxTokens ?? 4096,
-        top_p: params.topP ?? 1,
-        stop: params.stop,
-        ...params.cacheKey ? { user: params.cacheKey } : {}
-      }, this.buildRequestOptions(params));
+      const response = await client.chat.completions.create(
+        body,
+        this.buildRequestOptions(params)
+      );
       const choice = response.choices[0];
+      const toolCalls = this.normalizeToolCalls(choice.message.tool_calls);
       return {
         content: choice.message.content ?? "",
         finishReason: choice.finish_reason,
@@ -512,7 +629,8 @@ var OpenAICompatibleProvider = class {
           totalTokens: response.usage?.total_tokens ?? 0
         },
         raw: response,
-        attachmentResults
+        attachmentResults,
+        ...toolCalls.length > 0 ? { toolCalls } : {}
       };
     } catch (error) {
       this.logger.error(
@@ -534,7 +652,7 @@ var OpenAICompatibleProvider = class {
     this.validateApiKeyRequirement(apiKey);
     const attachmentResults = this.collectAttachmentFailures(params);
     const client = this.createClient(apiKey);
-    const messages = params.messages.filter((m) => {
+    const mappedMessages = params.messages.filter((m) => {
       if (m.role === "tool" && !m.toolCallId) return false;
       return true;
     }).map((m) => {
@@ -561,23 +679,33 @@ var OpenAICompatibleProvider = class {
         content: m.content
       };
     });
+    const messages = this.applySystemMessagePolicy(mappedMessages);
+    const body = {
+      model: params.model,
+      messages,
+      temperature: params.temperature ?? 0.7,
+      max_tokens: params.maxTokens ?? 4096,
+      top_p: params.topP ?? 1,
+      stop: params.stop,
+      stream: true,
+      stream_options: { include_usage: true },
+      ...params.cacheKey ? { user: params.cacheKey } : {},
+      ...params.tools && params.tools.length > 0 ? { tools: params.tools, tool_choice: params.toolChoice ?? "auto" } : {}
+    };
+    this.applyProfileParams(body, params);
     try {
-      const stream = await client.chat.completions.create({
-        model: params.model,
-        messages,
-        temperature: params.temperature ?? 0.7,
-        max_tokens: params.maxTokens ?? 4096,
-        top_p: params.topP ?? 1,
-        stop: params.stop,
-        stream: true,
-        stream_options: { include_usage: true },
-        ...params.cacheKey ? { user: params.cacheKey } : {}
-      }, this.buildRequestOptions(params));
+      const stream = await client.chat.completions.create(
+        body,
+        this.buildRequestOptions(params)
+      );
       let chunkCount = 0;
       let accumulatedUsage = null;
+      const toolCallAccumulator = /* @__PURE__ */ new Map();
+      let finishReason = null;
       for await (const chunk of stream) {
         chunkCount++;
-        const content = chunk.choices[0]?.delta?.content;
+        const choice = chunk.choices[0];
+        const content = choice?.delta?.content;
         const hasUsage = chunk.usage;
         if (hasUsage) {
           accumulatedUsage = {
@@ -586,6 +714,17 @@ var OpenAICompatibleProvider = class {
             total_tokens: chunk.usage?.total_tokens
           };
         }
+        if (choice?.delta?.tool_calls) {
+          for (const tcDelta of choice.delta.tool_calls) {
+            const idx = tcDelta.index ?? 0;
+            const existing = toolCallAccumulator.get(idx) ?? { id: "", name: "", arguments: "" };
+            if (tcDelta.id) existing.id = tcDelta.id;
+            if (tcDelta.function?.name) existing.name = tcDelta.function.name;
+            if (tcDelta.function?.arguments) existing.arguments += tcDelta.function.arguments;
+            toolCallAccumulator.set(idx, existing);
+          }
+        }
+        if (choice?.finish_reason) finishReason = choice.finish_reason;
         if (content) {
           yield {
             content,
@@ -593,6 +732,11 @@ var OpenAICompatibleProvider = class {
           };
         }
       }
+      const toolCalls = Array.from(toolCallAccumulator.values()).filter((tc) => tc.name).map((tc) => ({
+        id: tc.id,
+        type: "function",
+        function: { name: tc.name, arguments: tc.arguments }
+      }));
       yield {
         content: "",
         done: true,
@@ -601,7 +745,22 @@ var OpenAICompatibleProvider = class {
           completionTokens: accumulatedUsage.completion_tokens ?? 0,
           totalTokens: accumulatedUsage.total_tokens ?? 0
         } : void 0,
-        attachmentResults
+        attachmentResults,
+        // Only synthesized when the stream actually carried tool calls: the
+        // host's tool detection reads `rawResponse`, and an empty one on every
+        // ordinary turn would be noise in the logs for no gain.
+        ...toolCalls.length > 0 ? {
+          toolCalls,
+          rawResponse: {
+            choices: [
+              {
+                index: 0,
+                message: { role: "assistant", content: "", tool_calls: toolCalls },
+                finish_reason: finishReason
+              }
+            ]
+          }
+        } : {}
       };
     } catch (error) {
       this.logger.error(
@@ -902,7 +1061,7 @@ function rewriteLocalhostUrl(url) {
 }
 
 // src/version.generated.ts
-var PLUGIN_UTILS_VERSION = "2.2.20";
+var PLUGIN_UTILS_VERSION = "2.4.0";
 export {
   BUILTIN_TOOL_NAMES,
   DEFAULT_REQUEST_TIMEOUT_MS,
@@ -913,9 +1072,11 @@ export {
   __injectCoreLoggerFactory,
   __injectQuilltapVersion,
   applyDescriptionLimit,
+  applyProfileParameters,
   buildRequestAbortSignal,
   buildSdkClientOptions,
   buildSdkRequestOptions,
+  collapseLeadingSystemMessages,
   convertFromAnthropicFormat,
   convertFromGoogleFormat,
   convertOpenAIToAnthropicFormat,