kimi-vercel-ai-sdk-provider 0.4.0 → 0.5.0

package/src/multi-agent/workflows.ts

@@ -0,0 +1,539 @@
+/**
+ * Multi-agent workflow implementations.
+ * @module
+ */
+
+import type {
+  AgentStep,
+  GenerateResult,
+  LanguageModelUsage,
+  MultiAgentConfig,
+  MultiAgentMetadata,
+  MultiAgentResult,
+  WorkflowContext
+} from './types';
+import { extractPrimaryCode } from '../code-validation/detector';
+import { DEFAULT_SYSTEM_PROMPTS } from './types';
+
+// ============================================================================
+// Types
+// ============================================================================
+
+/**
+ * Function type for generating text.
+ */
+export type GenerateTextFunction = (modelId: string, prompt: string, systemPrompt?: string) => Promise<GenerateResult>;
+
+/**
+ * Function type for validating code.
+ */
+export type ValidateCodeFunction = (code: string) => Promise<{
+  valid: boolean;
+  errors: string[];
+  fixedCode?: string;
+}>;
+
+/**
+ * Options for running a workflow.
+ */
+export interface WorkflowRunnerOptions {
+  /**
+   * The user's prompt.
+   */
+  prompt: string;
+
+  /**
+   * Workflow configuration.
+   */
+  config: MultiAgentConfig;
+
+  /**
+   * Function to generate text.
+   */
+  generateText: GenerateTextFunction;
+
+  /**
+   * Optional function to validate code.
+   */
+  validateCode?: ValidateCodeFunction;
+}
+
+// ============================================================================
+// WorkflowRunner Class
+// ============================================================================
+
+/**
+ * Runner for multi-agent workflows.
+ */
+export class WorkflowRunner {
+  private generateText: GenerateTextFunction;
+  private validateCode?: ValidateCodeFunction;
+
+  constructor(generateText: GenerateTextFunction, validateCode?: ValidateCodeFunction) {
+    this.generateText = generateText;
+    this.validateCode = validateCode;
+  }
+
+  /**
+   * Run a multi-agent workflow.
+   */
+  async run(prompt: string, config: MultiAgentConfig): Promise<MultiAgentResult> {
+    const startTime = Date.now();
+
+    switch (config.workflow) {
+      case 'planner-executor':
+        return this.runPlannerExecutor(prompt, config, startTime);
+
+      case 'proposer-critic':
+        return this.runProposerCritic(prompt, config, startTime);
+
+      case 'debate':
+        return this.runDebate(prompt, config, startTime);
+
+      case 'custom':
+        if (!config.customWorkflow) {
+          throw new Error('Custom workflow requires customWorkflow function');
+        }
+        return this.runCustom(prompt, config, startTime);
+
+      default:
+        throw new Error(`Unknown workflow type: ${config.workflow}`);
+    }
+  }
+
+  /**
+   * Planner-Executor workflow.
+   */
+  private async runPlannerExecutor(
+    prompt: string,
+    config: MultiAgentConfig,
+    startTime: number
+  ): Promise<MultiAgentResult> {
+    const steps: AgentStep[] = [];
+    let totalUsage: LanguageModelUsage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 };
+
+    const modelA = config.modelA || 'kimi-k2.5-thinking';
+    const modelB = config.modelB || 'kimi-k2.5';
+
+    // Step 1: Planner creates plan
+    const planStart = Date.now();
+    const plannerPrompt = `${config.systemPrompts?.planner || DEFAULT_SYSTEM_PROMPTS.planner}
+
+Problem to solve:
+${prompt}
+
+Provide a detailed step-by-step plan.`;
+
+    const planResult = await this.generateText(modelA, plannerPrompt);
+
+    steps.push({
+      agent: 'A (Planner)',
+      role: 'planner',
+      action: 'Create implementation plan',
+      input: prompt,
+      output: planResult.text,
+      timestamp: Date.now(),
+      durationMs: Date.now() - planStart,
+      usage: planResult.usage,
+      model: modelA
+    });
+
+    totalUsage = this.addUsage(totalUsage, planResult.usage);
+
+    // Step 2: Executor implements
+    const execStart = Date.now();
+    const executorPrompt = `${config.systemPrompts?.executor || DEFAULT_SYSTEM_PROMPTS.executor}
+
+Plan to implement:
+${planResult.text}
+
+Original problem: ${prompt}
+
+Implement the solution following the plan.`;
+
+    const execResult = await this.generateText(modelB, executorPrompt);
+
+    steps.push({
+      agent: 'B (Executor)',
+      role: 'executor',
+      action: 'Implement solution',
+      input: planResult.text,
+      output: execResult.text,
+      timestamp: Date.now(),
+      durationMs: Date.now() - execStart,
+      usage: execResult.usage,
+      model: modelB
+    });
+
+    totalUsage = this.addUsage(totalUsage, execResult.usage);
+
+    // Step 3: Optional validation
+    let validation: MultiAgentResult['validation'];
+    if (config.validateCode && this.validateCode) {
+      const code = extractPrimaryCode(execResult.text);
+      if (code) {
+        const validStart = Date.now();
+        const validResult = await this.validateCode(code);
+
+        steps.push({
+          agent: 'Validator',
+          role: 'validator',
+          action: 'Validate code',
+          input: code,
+          output: validResult.valid ? 'Code is valid' : `Errors: ${validResult.errors.join(', ')}`,
+          timestamp: Date.now(),
+          durationMs: Date.now() - validStart
+        });
+
+        validation = {
+          valid: validResult.valid,
+          errors: validResult.errors,
+          finalCode: validResult.fixedCode || code
+        };
+      }
+    }
+
+    return {
+      text: execResult.text,
+      reasoning: planResult.text,
+      intermediateSteps: steps,
+      usage: totalUsage,
+      metadata: this.buildMetadata(config, steps, startTime, [modelA, modelB]),
+      validation
+    };
+  }
+
+  /**
+   * Proposer-Critic workflow.
+   */
+  private async runProposerCritic(
+    prompt: string,
+    config: MultiAgentConfig,
+    startTime: number
+  ): Promise<MultiAgentResult> {
+    const steps: AgentStep[] = [];
+    let totalUsage: LanguageModelUsage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 };
+
+    const modelA = config.modelA || 'kimi-k2.5';
+    const modelB = config.modelB || 'kimi-k2.5-thinking';
+    const iterations = config.iterations || 2;
+
+    let currentSolution = '';
+    let critique = '';
+
+    for (let i = 0; i < iterations; i++) {
+      // Proposer generates/improves solution
+      const proposeStart = Date.now();
+      const proposerPrompt =
+        i === 0
+          ? `${config.systemPrompts?.proposer || DEFAULT_SYSTEM_PROMPTS.proposer}
+
+Problem: ${prompt}
+
+Provide your solution.`
+          : `${config.systemPrompts?.proposer || DEFAULT_SYSTEM_PROMPTS.proposer}
+
+Problem: ${prompt}
+
+Previous solution:
+${currentSolution}
+
+Critique/feedback to address:
+${critique}
+
+Provide an improved solution addressing the feedback.`;
+
+      const proposeResult = await this.generateText(modelA, proposerPrompt);
+      currentSolution = proposeResult.text;
+
+      steps.push({
+        agent: 'A (Proposer)',
+        role: 'proposer',
+        action: `Generate solution (iteration ${i + 1})`,
+        input: i === 0 ? prompt : critique,
+        output: currentSolution,
+        timestamp: Date.now(),
+        durationMs: Date.now() - proposeStart,
+        usage: proposeResult.usage,
+        model: modelA
+      });
+
+      totalUsage = this.addUsage(totalUsage, proposeResult.usage);
+
+      // Critic reviews (except on last iteration)
+      if (i < iterations - 1) {
+        const critiqueStart = Date.now();
+        const criticPrompt = `${config.systemPrompts?.critic || DEFAULT_SYSTEM_PROMPTS.critic}
+
+Original problem: ${prompt}
+
+Solution to review:
+${currentSolution}
+
+Provide detailed feedback and suggestions for improvement.`;
+
+        const critiqueResult = await this.generateText(modelB, criticPrompt);
+        critique = critiqueResult.text;
+
+        steps.push({
+          agent: 'B (Critic)',
+          role: 'critic',
+          action: `Critique solution (iteration ${i + 1})`,
+          input: currentSolution,
+          output: critique,
+          timestamp: Date.now(),
+          durationMs: Date.now() - critiqueStart,
+          usage: critiqueResult.usage,
+          model: modelB
+        });
+
+        totalUsage = this.addUsage(totalUsage, critiqueResult.usage);
+      }
+    }
+
+    // Optional validation
+    let validation: MultiAgentResult['validation'];
+    if (config.validateCode && this.validateCode) {
+      const code = extractPrimaryCode(currentSolution);
+      if (code) {
+        const validResult = await this.validateCode(code);
+        validation = {
+          valid: validResult.valid,
+          errors: validResult.errors,
+          finalCode: validResult.fixedCode || code
+        };
+      }
+    }
+
+    return {
+      text: currentSolution,
+      reasoning: critique,
+      intermediateSteps: steps,
+      usage: totalUsage,
+      metadata: this.buildMetadata(config, steps, startTime, [modelA, modelB]),
+      validation
+    };
+  }
+
+  /**
+   * Debate workflow.
+   */
+  private async runDebate(prompt: string, config: MultiAgentConfig, startTime: number): Promise<MultiAgentResult> {
+    const steps: AgentStep[] = [];
+    let totalUsage: LanguageModelUsage = { promptTokens: 0, completionTokens: 0, totalTokens: 0 };
+
+    const modelA = config.modelA || 'kimi-k2.5';
+    const modelB = config.modelB || 'kimi-k2.5';
+    const iterations = config.iterations || 2;
+
+    let positionA = '';
+    let positionB = '';
+
+    // Initial positions
+    const initialPrompt = `Problem: ${prompt}
+
+Provide your solution and reasoning.`;
+
+    // Agent A's initial position
+    const aStart = Date.now();
+    const aResult = await this.generateText(modelA, initialPrompt);
+    positionA = aResult.text;
+
+    steps.push({
+      agent: 'A',
+      role: 'proposer',
+      action: 'Initial position',
+      input: prompt,
+      output: positionA,
+      timestamp: Date.now(),
+      durationMs: Date.now() - aStart,
+      usage: aResult.usage,
+      model: modelA
+    });
+
+    totalUsage = this.addUsage(totalUsage, aResult.usage);
+
+    // Agent B's initial position
+    const bStart = Date.now();
+    const bResult = await this.generateText(modelB, initialPrompt);
+    positionB = bResult.text;
+
+    steps.push({
+      agent: 'B',
+      role: 'proposer',
+      action: 'Initial position',
+      input: prompt,
+      output: positionB,
+      timestamp: Date.now(),
+      durationMs: Date.now() - bStart,
+      usage: bResult.usage,
+      model: modelB
+    });
+
+    totalUsage = this.addUsage(totalUsage, bResult.usage);
+
+    // Debate rounds
+    for (let i = 0; i < iterations - 1; i++) {
+      // A responds to B
+      const aDebateStart = Date.now();
+      const aDebatePrompt = `Problem: ${prompt}
+
+Your previous position:
+${positionA}
+
+Other perspective:
+${positionB}
+
+Consider the other perspective. Either defend your position with additional reasoning, or revise it based on valid points. Provide your updated solution.`;
+
+      const aDebateResult = await this.generateText(modelA, aDebatePrompt);
+      positionA = aDebateResult.text;
+
+      steps.push({
+        agent: 'A',
+        role: 'proposer',
+        action: `Debate round ${i + 1}`,
+        input: positionB,
+        output: positionA,
+        timestamp: Date.now(),
+        durationMs: Date.now() - aDebateStart,
+        usage: aDebateResult.usage,
+        model: modelA
+      });
+
+      totalUsage = this.addUsage(totalUsage, aDebateResult.usage);
+
+      // B responds to A
+      const bDebateStart = Date.now();
+      const bDebatePrompt = `Problem: ${prompt}
+
+Your previous position:
+${positionB}
+
+Other perspective:
+${positionA}
+
+Consider the other perspective. Either defend your position with additional reasoning, or revise it based on valid points. Provide your updated solution.`;
+
+      const bDebateResult = await this.generateText(modelB, bDebatePrompt);
+      positionB = bDebateResult.text;
+
+      steps.push({
+        agent: 'B',
+        role: 'proposer',
+        action: `Debate round ${i + 1}`,
+        input: positionA,
+        output: positionB,
+        timestamp: Date.now(),
+        durationMs: Date.now() - bDebateStart,
+        usage: bDebateResult.usage,
+        model: modelB
+      });
+
+      totalUsage = this.addUsage(totalUsage, bDebateResult.usage);
+    }
+
+    // Synthesize final answer (use last response as primary)
+    const finalText = positionA.length > positionB.length ? positionA : positionB;
+
+    return {
+      text: finalText,
+      reasoning: `Debate synthesis from ${iterations} rounds`,
+      intermediateSteps: steps,
+      usage: totalUsage,
+      metadata: this.buildMetadata(config, steps, startTime, [modelA, modelB])
+    };
+  }
+
+  /**
+   * Custom workflow.
+   */
+  private async runCustom(prompt: string, config: MultiAgentConfig, startTime: number): Promise<MultiAgentResult> {
+    const steps: AgentStep[] = [];
+    const modelA = config.modelA || 'kimi-k2.5';
+    const modelB = config.modelB || 'kimi-k2.5';
+
+    const context: WorkflowContext = {
+      generateWithModelA: (p: string) => this.generateText(modelA, p),
+      generateWithModelB: (p: string) => this.generateText(modelB, p),
+      validateCode: this.validateCode,
+      config,
+      addStep: (step) => {
+        steps.push({
+          ...step,
+          timestamp: Date.now(),
+          durationMs: 0
+        });
+      }
+    };
+
+    const result = await config.customWorkflow!(prompt, context);
+
+    // Merge steps if result doesn't include them
+    if (result.intermediateSteps.length === 0 && steps.length > 0) {
+      result.intermediateSteps = steps;
+    }
+
+    // Update metadata with timing
+    result.metadata.durationMs = Date.now() - startTime;
+
+    return result;
+  }
+
+  /**
+   * Build metadata for result.
+   */
+  private buildMetadata(
+    config: MultiAgentConfig,
+    steps: AgentStep[],
+    startTime: number,
+    models: string[]
+  ): MultiAgentMetadata {
+    return {
+      workflow: config.workflow,
+      iterations: steps.length,
+      durationMs: Date.now() - startTime,
+      models: [...new Set(models)],
+      validationEnabled: config.validateCode ?? false,
+      success: true
+    };
+  }
+
+  /**
+   * Add usage stats.
+   */
+  private addUsage(a: LanguageModelUsage, b?: LanguageModelUsage): LanguageModelUsage {
+    if (!b) {
+      return a;
+    }
+    return {
+      promptTokens: a.promptTokens + b.promptTokens,
+      completionTokens: a.completionTokens + b.completionTokens,
+      totalTokens: a.totalTokens + b.totalTokens
+    };
+  }
+}
+
+// ============================================================================
+// Utility Functions
+// ============================================================================
+
+/**
+ * Create an empty multi-agent result.
+ */
+export function createEmptyMultiAgentResult(workflow: MultiAgentConfig['workflow'], error: string): MultiAgentResult {
+  return {
+    text: '',
+    intermediateSteps: [],
+    usage: { promptTokens: 0, completionTokens: 0, totalTokens: 0 },
+    metadata: {
+      workflow,
+      iterations: 0,
+      durationMs: 0,
+      models: [],
+      validationEnabled: false,
+      success: false,
+      error
+    }
+  };
+}

package/src/project-tools/index.ts

@@ -0,0 +1,16 @@
+/**
+ * Project tools module exports.
+ * @module
+ */
+
+export type { GenerateTextFunction, ScaffolderOptions } from './scaffolder';
+export type {
+  OutputFormat,
+  ProjectFile,
+  ProjectMetadata,
+  ProjectTemplate,
+  ProjectType,
+  ScaffoldConfig,
+  ScaffoldResult
+} from './types';
+export { ProjectScaffolder, createEmptyScaffoldResult } from './scaffolder';