@blockrun/runcode 2.3.0 → 2.3.1

package/dist/agent/loop.js

@@ -7,6 +7,7 @@ import { ModelClient } from './llm.js';
 import { autoCompactIfNeeded, microCompact } from './compact.js';
 import { estimateHistoryTokens, updateActualTokens, resetTokenAnchor } from './tokens.js';
 import { handleSlashCommand } from './commands.js';
+import { reduceTokens } from './reduce.js';
 import { PermissionManager } from './permissions.js';
 import { StreamingExecutor } from './streaming-executor.js';
 import { optimizeHistory, CAPPED_MAX_TOKENS, ESCALATED_MAX_TOKENS, getMaxOutputTokens } from './optimize.js';
@@ -253,7 +254,13 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                 history.length = 0;
                 history.push(...optimized);
             }
-            // 2. Microcompact: only when history has >15 messages (skip for short conversations)
+            // 2. Token reduction: age old results, normalize whitespace, trim verbose messages
+            const reduced = reduceTokens(history, config.debug);
+            if (reduced !== history) {
+                history.length = 0;
+                history.push(...reduced);
+            }
+            // 3. Microcompact: only when history has >15 messages (skip for short conversations)
             if (history.length > 15) {
                 const microCompacted = microCompact(history, 8);
                 if (microCompacted !== history) {

package/dist/agent/reduce.d.ts

@@ -0,0 +1,38 @@
+/**
+ * Token Reduction for runcode.
+ * Original implementation — reduces context size through intelligent pruning.
+ *
+ * Strategy: instead of compression/encoding, we PRUNE redundant content.
+ * The model doesn't need verbose tool outputs from 20 turns ago.
+ *
+ * Three reduction passes:
+ * 1. Tool result aging — progressively shorten old tool results
+ * 2. Whitespace normalization — remove excessive blank lines and indentation
+ * 3. Stale context removal — drop system info that's been superseded
+ */
+import type { Dialogue } from './types.js';
+/**
+ * Progressively shorten tool results based on age.
+ * Recent results: keep full. Older results: keep summary. Very old: keep one line.
+ *
+ * This is the biggest token saver — a 10KB bash output from 20 turns ago
+ * can be reduced to "✓ Bash: ran npm test (exit 0)" saving ~2500 tokens.
+ */
+export declare function ageToolResults(history: Dialogue[]): Dialogue[];
+/**
+ * Normalize whitespace in text messages.
+ * - Collapse 3+ blank lines to 2
+ * - Remove trailing spaces
+ * - Reduce indentation beyond 8 spaces to 8
+ */
+export declare function normalizeWhitespace(history: Dialogue[]): Dialogue[];
+/**
+ * Trim very long assistant text messages from old turns.
+ * Recent messages: keep full. Old long messages: keep first 1000 chars.
+ */
+export declare function trimOldAssistantMessages(history: Dialogue[]): Dialogue[];
+/**
+ * Run all token reduction passes on conversation history.
+ * Returns same reference if nothing changed (cheap identity check).
+ */
+export declare function reduceTokens(history: Dialogue[], debug?: boolean): Dialogue[];

package/dist/agent/reduce.js

@@ -0,0 +1,231 @@
+/**
+ * Token Reduction for runcode.
+ * Original implementation — reduces context size through intelligent pruning.
+ *
+ * Strategy: instead of compression/encoding, we PRUNE redundant content.
+ * The model doesn't need verbose tool outputs from 20 turns ago.
+ *
+ * Three reduction passes:
+ * 1. Tool result aging — progressively shorten old tool results
+ * 2. Whitespace normalization — remove excessive blank lines and indentation
+ * 3. Stale context removal — drop system info that's been superseded
+ */
+// ─── 1. Tool Result Aging ─────────────────────────────────────────────────
+/**
+ * Progressively shorten tool results based on age.
+ * Recent results: keep full. Older results: keep summary. Very old: keep one line.
+ *
+ * This is the biggest token saver — a 10KB bash output from 20 turns ago
+ * can be reduced to "✓ Bash: ran npm test (exit 0)" saving ~2500 tokens.
+ */
+export function ageToolResults(history) {
+    // Find all tool_result positions
+    const toolPositions = [];
+    for (let i = 0; i < history.length; i++) {
+        const msg = history[i];
+        if (msg.role === 'user' &&
+            Array.isArray(msg.content) &&
+            msg.content.some(p => p.type === 'tool_result')) {
+            toolPositions.push(i);
+        }
+    }
+    if (toolPositions.length <= 3)
+        return history; // Nothing to age
+    const result = [...history];
+    const totalResults = toolPositions.length;
+    for (let idx = 0; idx < toolPositions.length; idx++) {
+        const pos = toolPositions[idx];
+        const age = totalResults - idx; // Higher = older
+        const msg = result[pos];
+        if (!Array.isArray(msg.content))
+            continue;
+        const parts = msg.content;
+        let modified = false;
+        const aged = parts.map(part => {
+            if (part.type !== 'tool_result')
+                return part;
+            const content = typeof part.content === 'string'
+                ? part.content
+                : JSON.stringify(part.content);
+            const charLen = content.length;
+            // Recent 3 results: keep full
+            if (age <= 3)
+                return part;
+            // Age 4-8: keep first 500 chars
+            if (age <= 8 && charLen > 500) {
+                modified = true;
+                const truncated = content.slice(0, 500);
+                const lastNl = truncated.lastIndexOf('\n');
+                const clean = lastNl > 250 ? truncated.slice(0, lastNl) : truncated;
+                return {
+                    ...part,
+                    content: `${clean}\n... (${charLen - clean.length} chars omitted, ${age} turns ago)`,
+                };
+            }
+            // Age 9-15: keep first 200 chars
+            if (age <= 15 && charLen > 200) {
+                modified = true;
+                const firstLine = content.split('\n')[0].slice(0, 150);
+                return {
+                    ...part,
+                    content: `${firstLine}\n... (${charLen} chars, ${age} turns ago)`,
+                };
+            }
+            // Age 16+: one line summary
+            if (age > 15 && charLen > 80) {
+                modified = true;
+                const summary = content.split('\n')[0].slice(0, 60);
+                return {
+                    ...part,
+                    content: part.is_error
+                        ? `[Error: ${summary}...]`
+                        : `[Result: ${summary}...]`,
+                };
+            }
+            return part;
+        });
+        if (modified) {
+            result[pos] = { role: 'user', content: aged };
+        }
+    }
+    return result;
+}
+// ─── 2. Whitespace Normalization ──────────────────────────────────────────
+/**
+ * Normalize whitespace in text messages.
+ * - Collapse 3+ blank lines to 2
+ * - Remove trailing spaces
+ * - Reduce indentation beyond 8 spaces to 8
+ */
+export function normalizeWhitespace(history) {
+    let modified = false;
+    const result = history.map(msg => {
+        if (typeof msg.content !== 'string')
+            return msg;
+        const original = msg.content;
+        const cleaned = original
+            .replace(/[ \t]+$/gm, '') // Trailing spaces
+            .replace(/\n{4,}/g, '\n\n\n') // Max 3 consecutive newlines
+            .replace(/^( {9,})/gm, '        '); // Cap indentation at 8 spaces
+        if (cleaned !== original) {
+            modified = true;
+            return { ...msg, content: cleaned };
+        }
+        return msg;
+    });
+    return modified ? result : history;
+}
+// ─── 3. Verbose Assistant Message Trimming ────────────────────────────────
+/**
+ * Trim very long assistant text messages from old turns.
+ * Recent messages: keep full. Old long messages: keep first 1000 chars.
+ */
+export function trimOldAssistantMessages(history) {
+    const MAX_OLD_ASSISTANT_CHARS = 1500;
+    const KEEP_RECENT = 4; // Keep last 4 assistant messages full
+    let assistantCount = 0;
+    for (const msg of history) {
+        if (msg.role === 'assistant')
+            assistantCount++;
+    }
+    if (assistantCount <= KEEP_RECENT)
+        return history;
+    let seenAssistant = 0;
+    let modified = false;
+    const result = history.map(msg => {
+        if (msg.role !== 'assistant')
+            return msg;
+        seenAssistant++;
+        // Keep recent messages full
+        if (assistantCount - seenAssistant < KEEP_RECENT)
+            return msg;
+        if (typeof msg.content === 'string' && msg.content.length > MAX_OLD_ASSISTANT_CHARS) {
+            modified = true;
+            const truncated = msg.content.slice(0, MAX_OLD_ASSISTANT_CHARS);
+            const lastNl = truncated.lastIndexOf('\n');
+            const clean = lastNl > MAX_OLD_ASSISTANT_CHARS / 2 ? truncated.slice(0, lastNl) : truncated;
+            return { ...msg, content: clean + '\n... (response truncated)' };
+        }
+        // Also handle content array with text parts
+        if (Array.isArray(msg.content)) {
+            const parts = msg.content;
+            let totalChars = 0;
+            for (const p of parts) {
+                if (p.type === 'text')
+                    totalChars += p.text.length;
+            }
+            if (totalChars > MAX_OLD_ASSISTANT_CHARS) {
+                modified = true;
+                const trimmedParts = parts.map(p => {
+                    if (p.type !== 'text' || p.text.length <= 500)
+                        return p;
+                    return { ...p, text: p.text.slice(0, 500) + '\n... (trimmed)' };
+                });
+                return { ...msg, content: trimmedParts };
+            }
+        }
+        return msg;
+    });
+    return modified ? result : history;
+}
+// ─── Pipeline ─────────────────────────────────────────────────────────────
+/**
+ * Run all token reduction passes on conversation history.
+ * Returns same reference if nothing changed (cheap identity check).
+ */
+export function reduceTokens(history, debug) {
+    if (history.length < 8)
+        return history; // Skip for short conversations
+    let current = history;
+    let totalSaved = 0;
+    // Pass 1: Age old tool results
+    const aged = ageToolResults(current);
+    if (aged !== current) {
+        const before = estimateChars(current);
+        current = aged;
+        const saved = before - estimateChars(current);
+        totalSaved += saved;
+    }
+    // Pass 2: Normalize whitespace
+    const normalized = normalizeWhitespace(current);
+    if (normalized !== current) {
+        const before = estimateChars(current);
+        current = normalized;
+        totalSaved += before - estimateChars(current);
+    }
+    // Pass 3: Trim old verbose assistant messages
+    const trimmed = trimOldAssistantMessages(current);
+    if (trimmed !== current) {
+        const before = estimateChars(current);
+        current = trimmed;
+        totalSaved += before - estimateChars(current);
+    }
+    if (debug && totalSaved > 500) {
+        const tokensSaved = Math.round(totalSaved / 4);
+        console.error(`[runcode] Token reduction: ~${tokensSaved} tokens saved`);
+    }
+    return current;
+}
+function estimateChars(history) {
+    let total = 0;
+    for (const msg of history) {
+        if (typeof msg.content === 'string') {
+            total += msg.content.length;
+        }
+        else if (Array.isArray(msg.content)) {
+            for (const p of msg.content) {
+                if ('type' in p) {
+                    if (p.type === 'text')
+                        total += p.text.length;
+                    else if (p.type === 'tool_result') {
+                        total += typeof p.content === 'string' ? p.content.length : JSON.stringify(p.content).length;
+                    }
+                    else if (p.type === 'tool_use') {
+                        total += JSON.stringify(p.input).length;
+                    }
+                }
+            }
+        }
+    }
+    return total;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/runcode",
-  "version": "2.3.0",
+  "version": "2.3.1",
   "description": "RunCode — AI coding agent powered by 41+ models. Pay per use with USDC.",
   "type": "module",
   "bin": {