ccs/src/glmt/reasoning-enforcer.ts

/**
 * ReasoningEnforcer - Inject explicit reasoning instructions into prompts
 *
 * Purpose: Force GLM models to use structured reasoning output format (<reasoning_content>)
 * This complements API parameters (reasoning: true) with explicit prompt instructions.
 *
 * Strategy:
 *   1. If system prompt exists: Prepend reasoning instruction
 *   2. If no system prompt: Prepend to first user message
 *   3. Select prompt template based on effort level (low/medium/high/max)
 *   4. Preserve message structure (string vs array content)
 */

type EffortLevel = 'low' | 'medium' | 'high' | 'max';

interface ContentBlock {
  type: string;
  text?: string;
  [key: string]: unknown;
}

interface Message {
  role: string;
  content: string | ContentBlock[];
}

interface ThinkingConfig {
  thinking?: boolean;
  effort?: string;
}

interface ReasoningEnforcerOptions {
  enabled?: boolean;
  prompts?: Record<EffortLevel, string>;
}

export class ReasoningEnforcer {
  private enabled: boolean;
  private prompts: Record<EffortLevel, string>;

  constructor(options: ReasoningEnforcerOptions = {}) {
    this.enabled = options.enabled ?? false; // Opt-in by default
    this.prompts = options.prompts || this.getDefaultPrompts();
  }

  /**
   * Inject reasoning instruction into messages
   * @param messages - Messages array to modify
   * @param thinkingConfig - { thinking: boolean, effort: string }
   * @returns Modified messages array
   */
  injectInstruction(messages: Message[], thinkingConfig: ThinkingConfig = {}): Message[] {
    // Only inject if enabled or thinking explicitly requested
    if (!this.enabled && !thinkingConfig.thinking) {
      return messages;
    }

    // Clone messages to avoid mutation
    const modifiedMessages: Message[] = JSON.parse(JSON.stringify(messages));

    // Select prompt based on effort level
    const effort = (thinkingConfig.effort?.toLowerCase() || 'medium') as EffortLevel;
    const prompt = this.selectPrompt(effort);

    // Strategy 1: Inject into system prompt (preferred)
    const systemIndex = modifiedMessages.findIndex((m) => m.role === 'system');
    if (systemIndex >= 0) {
      const systemMsg = modifiedMessages[systemIndex];

      if (typeof systemMsg.content === 'string') {
        systemMsg.content = `${prompt}\n\n${systemMsg.content}`;
      } else if (Array.isArray(systemMsg.content)) {
        systemMsg.content.unshift({
          type: 'text',
          text: prompt,
        });
      }

      return modifiedMessages;
    }

    // Strategy 2: Prepend to first user message
    const userIndex = modifiedMessages.findIndex((m) => m.role === 'user');
    if (userIndex >= 0) {
      const userMsg = modifiedMessages[userIndex];

      if (typeof userMsg.content === 'string') {
        userMsg.content = `${prompt}\n\n${userMsg.content}`;
      } else if (Array.isArray(userMsg.content)) {
        userMsg.content.unshift({
          type: 'text',
          text: prompt,
        });
      }

      return modifiedMessages;
    }

    // No system or user messages found (edge case)
    return modifiedMessages;
  }

  /**
   * Select prompt template based on effort level
   */
  private selectPrompt(effort: EffortLevel): string {
    return this.prompts[effort] || this.prompts.medium;
  }

  /**
   * Get default prompt templates
   */
  private getDefaultPrompts(): Record<EffortLevel, string> {
    return {
      low: `You are an expert reasoning model using GLM-4.6 architecture.

CRITICAL: Before answering, write 2-3 sentences of reasoning in <reasoning_content> tags.

OUTPUT FORMAT:
<reasoning_content>
(Brief analysis: what is the problem? what's the approach?)
</reasoning_content>

(Write your final answer here)`,

      medium: `You are an expert reasoning model using GLM-4.6 architecture.

CRITICAL REQUIREMENTS:
1. Always think step-by-step before answering
2. Write your reasoning process explicitly in <reasoning_content> tags
3. Never skip your chain of thought, even for simple problems

OUTPUT FORMAT:
<reasoning_content>
(Write your detailed thinking here: analyze the problem, explore approaches,
evaluate trade-offs, and arrive at a conclusion)
</reasoning_content>

(Write your final answer here based on your reasoning above)`,

      high: `You are an expert reasoning model using GLM-4.6 architecture.

CRITICAL REQUIREMENTS:
1. Think deeply and systematically before answering
2. Write comprehensive reasoning in <reasoning_content> tags
3. Explore multiple approaches and evaluate trade-offs
4. Show all steps in your problem-solving process

OUTPUT FORMAT:
<reasoning_content>
(Write exhaustive analysis here:
 - Problem decomposition
 - Multiple approach exploration
 - Trade-off analysis for each approach
 - Edge case consideration
 - Final conclusion with justification)
</reasoning_content>

(Write your final answer here based on your systematic reasoning above)`,

      max: `You are an expert reasoning model using GLM-4.6 architecture.

CRITICAL REQUIREMENTS:
1. Think exhaustively from first principles
2. Write extremely detailed reasoning in <reasoning_content> tags
3. Analyze ALL possible angles, approaches, and edge cases
4. Challenge your own assumptions and explore alternatives
5. Provide rigorous justification for every claim

OUTPUT FORMAT:
<reasoning_content>
(Write comprehensive analysis here:
 - First principles breakdown
 - Exhaustive approach enumeration
 - Comparative analysis of all approaches
 - Edge case and failure mode analysis
 - Assumption validation
 - Counter-argument consideration
 - Final conclusion with rigorous justification)
</reasoning_content>

(Write your final answer here based on your exhaustive reasoning above)`,
    };
  }
}

export default ReasoningEnforcer;