@avasis-ai/synthcode 1.0.7 → 1.0.8

package/dist/index.cjs

@@ -99,6 +99,7 @@ __export(src_exports, {
   AnthropicProvider: () => AnthropicProvider,
   BaseProvider: () => BaseProvider,
   BashTool: () => BashTool,
+  ClusterProvider: () => ClusterProvider,
   ContextManager: () => ContextManager,
   CostTracker: () => CostTracker,
   DEFAULT_COMPACT_THRESHOLD: () => DEFAULT_COMPACT_THRESHOLD,
@@ -124,6 +125,7 @@ __export(src_exports, {
   WebFetchTool: () => WebFetchTool,
   agentLoop: () => agentLoop,
   anthropic: () => anthropic,
+  cluster: () => cluster,
   createProvider: () => createProvider,
   createStreamAggregator: () => createStreamAggregator,
   defineTool: () => defineTool,
@@ -2400,6 +2402,338 @@ var OllamaProvider = class {
   }
 };
 
+// src/llm/cluster.ts
+var DEFAULT_BASE = "http://localhost:11434/v1";
+function defaultCluster(baseURL) {
+  const url = baseURL || DEFAULT_BASE;
+  return [
+    { model: "gemma4:31b", role: "planner", baseURL: url, temperature: 0.3 },
+    { model: "gemma4:26b", role: "worker", baseURL: url, temperature: 0.3 },
+    { model: "gemma4:26b", role: "worker", baseURL: url, temperature: 0.35 },
+    { model: "gemma4:26b", role: "reviewer", baseURL: url, temperature: 0.3 },
+    { model: "gemma4:e4b", role: "drafter", baseURL: url, temperature: 0.3 }
+  ];
+}
+function classifyComplexity(messages, hasTools) {
+  let totalTokens = 0;
+  let turns = messages.length;
+  let hasToolResults = false;
+  let hasCode = false;
+  let hasErrors = false;
+  for (const m of messages) {
+    const text = typeof m.content === "string" ? m.content : JSON.stringify(m.content);
+    totalTokens += text.length / 4;
+    if (m.role === "tool") hasToolResults = true;
+    if (text.includes("```") || text.includes("function ") || text.includes("class ")) hasCode = true;
+    if (text.includes("error") || text.includes("Error") || text.includes("FAIL")) hasErrors = true;
+  }
+  if (totalTokens > 1e4 || hasToolResults && hasErrors) return "complex";
+  if (totalTokens > 2e3 || hasTools || hasCode || hasErrors || turns > 4) return "medium";
+  return "simple";
+}
+function selectSlots(config, complexity) {
+  const planners = config.slots.filter((s) => s.role === "planner");
+  const workers = config.slots.filter((s) => s.role === "worker");
+  const reviewers = config.slots.filter((s) => s.role === "reviewer");
+  const drafters = config.slots.filter((s) => s.role === "drafter");
+  switch (complexity) {
+    case "complex":
+      return { planner: planners[0], workers, reviewers, drafters: drafters[0] };
+    case "medium":
+      return { planner: planners[0], workers: workers.slice(0, 1), reviewers: reviewers.slice(0, 1), drafters: drafters[0] };
+    case "simple":
+    default:
+      return { planner: planners[0], workers: workers.slice(0, 1), reviewers: reviewers.slice(0, 1), drafters: drafters[0] };
+  }
+}
+async function callModel(slot, request, baseURL, timeoutMs, extraOpts) {
+  const url = (slot.baseURL || baseURL) + "/chat/completions";
+  const messages = [];
+  if (request.systemPrompt) {
+    messages.push({ role: "system", content: request.systemPrompt });
+  }
+  for (const m of request.messages) {
+    if (m.role === "tool") {
+      messages.push({ role: "tool", tool_call_id: m.tool_use_id, content: m.content });
+      continue;
+    }
+    if (m.role === "assistant" && Array.isArray(m.content)) {
+      const textParts = m.content.filter((b) => b.type === "text");
+      const toolParts = m.content.filter((b) => b.type === "tool_use");
+      const msg = {};
+      if (textParts.length) msg.content = textParts.map((p) => p.text).join("");
+      if (toolParts.length) {
+        msg.tool_calls = toolParts.map((b) => {
+          const tb = b;
+          return { id: tb.id, type: "function", function: { name: tb.name, arguments: JSON.stringify(tb.input) } };
+        });
+      }
+      msg.role = "assistant";
+      messages.push(msg);
+      continue;
+    }
+    messages.push({ role: m.role, content: m.content });
+  }
+  const body = {
+    model: slot.model,
+    messages,
+    stream: false,
+    options: {
+      num_predict: request.maxOutputTokens || slot.maxTokens || 4096,
+      temperature: request.temperature ?? slot.temperature ?? 0.3,
+      top_p: 0.95,
+      top_k: 64
+    },
+    ...extraOpts
+  };
+  if (request.tools?.length) {
+    body.tools = request.tools.map((t) => ({
+      type: "function",
+      function: { name: t.name, description: t.description, parameters: t.input_schema }
+    }));
+  }
+  const start = Date.now();
+  let response;
+  try {
+    response = await fetch(url, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+      signal: AbortSignal.timeout(timeoutMs)
+    });
+  } catch (err) {
+    if (err instanceof RetryableError) throw err;
+    throw new RetryableError(`Cluster slot ${slot.model} connection failed: ${err instanceof Error ? err.message : String(err)}`);
+  }
+  if (!response.ok) {
+    const text2 = await response.text();
+    if (response.status === 429 || response.status === 503 || response.status === 529) {
+      throw new RetryableError(`Cluster slot ${slot.model} rate limited: ${response.status}`);
+    }
+    throw new Error(`Cluster slot ${slot.model} error ${response.status}: ${text2.slice(0, 200)}`);
+  }
+  const data = await response.json();
+  const choice = data.choices?.[0];
+  if (!choice) throw new Error(`Cluster slot ${slot.model} returned no choices`);
+  let text = choice.message?.content || "";
+  text = text.replace(/<think[^>]*>[\s\S]*?<\/think>/gi, "").trim();
+  text = text.replace(/<thinking>[\s\S]*?<\/thinking>/gi, "").trim();
+  text = text.replace(/<channel>thought[\s\S]*?<channel|>/gi, "").trim();
+  text = text.replace(/\[Thinking[^\]]*\]/gi, "").trim();
+  return {
+    content: text,
+    usage: {
+      inputTokens: data.usage?.prompt_tokens ?? 0,
+      outputTokens: data.usage?.completion_tokens ?? 0,
+      cacheReadTokens: 0,
+      cacheWriteTokens: 0
+    },
+    stopReason: choice.finish_reason === "tool_calls" ? "tool_use" : "end_turn",
+    model: slot.model,
+    durationMs: Date.now() - start
+  };
+}
+function mergeUsage(...usages) {
+  return {
+    inputTokens: usages.reduce((s, u) => s + u.inputTokens, 0),
+    outputTokens: usages.reduce((s, u) => s + u.outputTokens, 0),
+    cacheReadTokens: usages.reduce((s, u) => s + (u.cacheReadTokens || 0), 0),
+    cacheWriteTokens: usages.reduce((s, u) => s + (u.cacheWriteTokens || 0), 0)
+  };
+}
+var ClusterProvider = class {
+  model;
+  config;
+  stats;
+  constructor(config) {
+    this.config = {
+      baseURL: DEFAULT_BASE,
+      timeoutMs: 18e4,
+      strategy: "auto",
+      draftCount: 1,
+      debateRounds: 1,
+      slots: defaultCluster(config?.baseURL),
+      ...config
+    };
+    const models = [...new Set(this.config.slots.map((s) => s.model))];
+    this.model = `cluster[${models.join(",")}]`;
+    this.stats = { calls: 0, tokensIn: 0, tokensOut: 0, byModel: {} };
+    for (const m of models) this.stats.byModel[m] = 0;
+  }
+  getStats() {
+    return { ...this.stats };
+  }
+  async chat(request) {
+    const strategy = this.config.strategy || "auto";
+    const complexity = strategy === "auto" ? classifyComplexity(request.messages, !!request.tools?.length) : "medium";
+    let result;
+    switch (strategy === "auto" ? complexity : "simple") {
+      case "simple":
+        result = await this.speculative(request);
+        break;
+      case "complex":
+        result = await this.debate(request);
+        break;
+      default:
+        result = await this.draftVerify(request);
+        break;
+    }
+    const content = [];
+    if (result.content) {
+      content.push({ type: "text", text: result.content });
+    }
+    return {
+      content,
+      stopReason: result.stopReason,
+      usage: result.usage
+    };
+  }
+  async speculative(request) {
+    const slots = selectSlots(this.config, "simple");
+    const drafter = slots.drafters;
+    const worker = slots.workers[0];
+    if (!drafter || !worker) {
+      return this.fallback(request);
+    }
+    const draft = await this.safeCall(drafter, request, { num_predict: request.maxOutputTokens || 4096 });
+    if (!draft || draft.content.length < 20) {
+      const fb = await this.safeCall(worker, request);
+      return fb || this.fallback(request);
+    }
+    const verifyMessages = [
+      ...request.messages.slice(0, -1),
+      { role: "user", content: `${typeof request.messages[request.messages.length - 1].content === "string" ? request.messages[request.messages.length - 1].content : ""}
+
+Here is a draft response. Review it. If it is correct and complete, return exactly the same text. If it has errors, fix them. Return ONLY the final corrected text, nothing else.
+
+DRAFT:
+${draft.content.slice(0, 8e3)}` }
+    ];
+    const verifyRequest = { ...request, messages: verifyMessages };
+    const verified = await this.safeCall(worker, verifyRequest);
+    if (!verified) return { ...draft, model: this.model };
+    const similarity2 = jaccardSimilarity(draft.content, verified.content);
+    if (similarity2 > 0.7) {
+      return { ...draft, usage: mergeUsage(draft.usage) };
+    }
+    return { ...verified, usage: mergeUsage(draft.usage, verified.usage) };
+  }
+  async draftVerify(request) {
+    const slots = selectSlots(this.config, "medium");
+    const worker = slots.workers[0];
+    const reviewer = slots.reviewers[0];
+    if (!worker) return this.fallback(request);
+    const primary = await this.safeCall(worker, request);
+    if (!primary) return this.fallback(request);
+    if (!reviewer || primary.content.length < 50) {
+      return primary;
+    }
+    const reviewMessages = [
+      { role: "user", content: `Review this response for correctness. Fix any errors. Return ONLY the corrected text.
+
+ORIGINAL REQUEST:
+${typeof request.messages[request.messages.length - 1].content === "string" ? request.messages[request.messages.length - 1].content : ""}
+
+RESPONSE TO REVIEW:
+${primary.content.slice(0, 6e3)}` }
+    ];
+    const reviewed = await this.safeCall(reviewer, { ...request, messages: reviewMessages, maxOutputTokens: request.maxOutputTokens });
+    if (!reviewed) return primary;
+    return { ...reviewed, usage: mergeUsage(primary.usage, reviewed.usage) };
+  }
+  async debate(request) {
+    const slots = selectSlots(this.config, "complex");
+    const workers = slots.workers;
+    const planner = slots.planner;
+    if (!workers.length) return this.fallback(request);
+    const roundRounds = this.config.debateRounds || 1;
+    let currentContent = "";
+    for (let round = 0; round < roundRounds; round++) {
+      const workerSlot = workers[round % workers.length];
+      const implMessages = round === 0 ? request.messages : [
+        ...request.messages.slice(0, -1),
+        {
+          role: "user",
+          content: `${typeof request.messages[request.messages.length - 1].content === "string" ? request.messages[request.messages.length - 1].content : ""}
+
+PREVIOUS ATTEMPT (Round ${round}):
+${currentContent.slice(0, 4e3)}
+
+Improve this. Fix any issues.`
+        }
+      ];
+      const impl = await this.safeCall(workerSlot, { ...request, messages: implMessages });
+      if (!impl) continue;
+      currentContent = impl.content;
+    }
+    if (!currentContent) return this.fallback(request);
+    if (planner && workers.length > 1) {
+      const candidates = [];
+      for (const w of workers) {
+        const c = await this.safeCall(w, request);
+        if (c && c.content.length > 30) candidates.push(c);
+      }
+      if (candidates.length > 1) {
+        const best = candidates.sort((a, b) => b.content.length - a.content.length)[0];
+        const arbMessages = [
+          {
+            role: "user",
+            content: `Select the best response or synthesize a better one. Return ONLY the final text.
+
+TASK:
+${typeof request.messages[request.messages.length - 1].content === "string" ? request.messages[request.messages.length - 1].content : ""}
+
+CANDIDATES:
+${candidates.map((c, i) => `--- Candidate ${i + 1} (${c.model}) ---
+${c.content.slice(0, 3e3)}`).join("\n\n")}`
+          }
+        ];
+        const arbitrated = await this.safeCall(planner, { ...request, messages: arbMessages, maxOutputTokens: request.maxOutputTokens });
+        if (arbitrated && arbitrated.content.length > 20) {
+          const allUsages = candidates.map((c) => c.usage).concat(arbitrated.usage);
+          return { ...arbitrated, usage: mergeUsage(...allUsages) };
+        }
+      }
+    }
+    return {
+      content: currentContent,
+      usage: { inputTokens: 0, outputTokens: 0 },
+      stopReason: "end_turn",
+      model: this.model,
+      durationMs: 0
+    };
+  }
+  async safeCall(slot, request, extraOpts) {
+    try {
+      const result = await callModel(slot, request, this.config.baseURL || DEFAULT_BASE, this.config.timeoutMs || 18e4, extraOpts);
+      this.stats.calls++;
+      this.stats.tokensIn += result.usage.inputTokens;
+      this.stats.tokensOut += result.usage.outputTokens;
+      this.stats.byModel[slot.model] = (this.stats.byModel[slot.model] || 0) + 1;
+      return result;
+    } catch (e) {
+      return null;
+    }
+  }
+  async fallback(request) {
+    const anySlot = this.config.slots[0];
+    const result = await callModel(anySlot, request, this.config.baseURL || DEFAULT_BASE, this.config.timeoutMs || 18e4);
+    this.stats.calls++;
+    this.stats.tokensIn += result.usage.inputTokens;
+    this.stats.tokensOut += result.usage.outputTokens;
+    this.stats.byModel[anySlot.model] = (this.stats.byModel[anySlot.model] || 0) + 1;
+    return result;
+  }
+};
+function jaccardSimilarity(a, b) {
+  const aWords = new Set(a.toLowerCase().split(/\s+/).filter((w) => w.length > 2));
+  const bWords = new Set(b.toLowerCase().split(/\s+/).filter((w) => w.length > 2));
+  if (aWords.size === 0 && bWords.size === 0) return 1;
+  const intersection = [...aWords].filter((w) => bWords.has(w)).length;
+  const union = (/* @__PURE__ */ new Set([...aWords, ...bWords])).size;
+  return union > 0 ? intersection / union : 0;
+}
+
 // src/llm/index.ts
 function anthropic(config) {
   return new AnthropicProvider(config);
@@ -2410,6 +2744,9 @@ function openai(config) {
 function ollama(config) {
   return new OllamaProvider(config);
 }
+function cluster(config) {
+  return new ClusterProvider(config);
+}
 var CustomProvider = class {
   model;
   chatFn;
@@ -2868,6 +3205,7 @@ describe("Agent", () => { it("should have tools registered", () => { expect(true
   AnthropicProvider,
   BaseProvider,
   BashTool,
+  ClusterProvider,
   ContextManager,
   CostTracker,
   DEFAULT_COMPACT_THRESHOLD,
@@ -2893,6 +3231,7 @@ describe("Agent", () => { it("should have tools registered", () => { expect(true
   WebFetchTool,
   agentLoop,
   anthropic,
+  cluster,
   createProvider,
   createStreamAggregator,
   defineTool,