fix: classify current prompt intent, not full history

[evansantos]

Problem

The routing classifier was passing the entire conversation history (235+ messages, 300k+ chars) to classifyWithRouter. This caused every single request — even a simple "oi" — to be classified as reasoning tier with lengthy-content signal.

Fix

Changed the classification input to only include:

• System prompt (for persona/context awareness)
• Last 3 messages (conversational flow)
• Current prompt (the actual request to classify)

This ensures the classifier evaluates the intent of the current request, not the volume of prior conversation.

Before

User: "oi"
🔀 Routing: reasoning tier (0.89) → opus | signals: [lengthy-content]

After

User: "oi"  
🔀 Routing: simple tier → haiku

User: "design a distributed cache"
🔀 Routing: complex tier → opus

Files Changed

• src/index.ts — llm_input hook routing classification

Fixes #9

---

🐕 GitSniff Summary

What this PR does

This pull request refines the routing classification logic within the slimclaw plugin by reducing the input context provided to the classifyWithRouter function. It now intelligently filters the conversation history, including only the system prompt, the last three messages, and the current prompt, to accurately assess the intent of the current request rather than the cumulative volume of the entire conversation. This change significantly improves the precision of AI model routing, preventing simple requests from being misclassified as complex, leading to more efficient resource utilization.

Key Changes

• Modified the llm_input hook to pass a curated set of messages (system prompt, last 3 messages, current prompt) to classifyWithRouter.
• Ensured that the classifier evaluates the intent of the current request rather than the entire conversation history.
• Corrected the routing behavior, allowing simple prompts to be classified as 'simple tier' and complex prompts as 'complex tier'.
• Addressed issue Routing classifier should evaluate current prompt, not full history #9 by preventing over-classification of prompts due to lengthy input.

Review Score: Excellent 🟢

Tip

No major issues found. Safe to merge.

_{🐕 Reviewed by GitSniff}

[gitsniff]

Score: Excellent ✅

The PR addresses a critical issue where the routing classifier was over-classifying prompts due to excessive context. The fix involves adjusting the classifyWithRouter input to be more focused, incorporating only the system prompt, the last three messages, and the current prompt. This ensures accurate and contextually relevant routing decisions, improving the efficiency and correctness of the slimclaw plugin.

1 finding posted as inline comments below.

_{Review completed in 27s | Basic Plan | Gemini 2.5 Flash}

[gitsniff]

ℹ️ Info

The history.slice(-3) operation will always return an array, even if empty. If history is undefined or null, recentHistory will be an empty array, and the loop will not execute. Therefore, the if (!msg) continue; check at src/index.ts:352 is redundant and can be removed.

Fix: Remove the if (!msg) continue; statement as msg will always be defined within the loop if recentHistory contains elements. The recentHistory array will not contain null or undefined elements itself.

🤖 Prompt for AI Agents

In src/index.ts around line 352:

Issue: The `history.slice(-3)` operation will always return an array, even if empty. If `history` is `undefined` or `null`, `recentHistory` will be an empty array, and the loop will not execute. Therefore, the `if (!msg) continue;` check at `src/index.ts:352` is redundant and can be removed.

Remove the `if (!msg) continue;` statement as `msg` will always be defined within the loop if `recentHistory` contains elements. The `recentHistory` array will not contain `null` or `undefined` elements itself.