@blockrun/runcode 2.3.0 → 2.4.0

package/dist/compression/layers/deduplication.js

@@ -0,0 +1,97 @@
+/**
+ * Layer 1: Message Deduplication
+ *
+ * Removes exact duplicate messages from conversation history.
+ * Common in heartbeat patterns and repeated tool calls.
+ *
+ * Safe for LLM: Identical messages add no new information.
+ * Expected savings: 2-5%
+ */
+import { createHash } from 'node:crypto';
+/**
+ * Generate a hash for a message based on its semantic content.
+ * Uses role + content + tool_call_id to identify duplicates.
+ */
+function hashMessage(message) {
+    const parts = [
+        message.role,
+        message.content || "",
+        message.tool_call_id || "",
+        message.name || "",
+    ];
+    // Include tool_calls if present
+    if (message.tool_calls) {
+        parts.push(JSON.stringify(message.tool_calls.map((tc) => ({
+            name: tc.function.name,
+            args: tc.function.arguments,
+        }))));
+    }
+    const content = parts.join("|");
+    return createHash("md5").update(content).digest("hex");
+}
+/**
+ * Remove exact duplicate messages from the conversation.
+ *
+ * Strategy:
+ * - Keep first occurrence of each unique message
+ * - Preserve order for semantic coherence
+ * - Never dedupe system messages (they set context)
+ * - Allow duplicate user messages (user might repeat intentionally)
+ * - CRITICAL: Never dedupe assistant messages with tool_calls that are
+ *   referenced by subsequent tool messages (breaks Anthropic tool_use/tool_result pairing)
+ */
+export function deduplicateMessages(messages) {
+    const seen = new Set();
+    const result = [];
+    let duplicatesRemoved = 0;
+    // First pass: collect all tool_call_ids that are referenced by tool messages
+    // These tool_calls MUST be preserved to maintain tool_use/tool_result pairing
+    const referencedToolCallIds = new Set();
+    for (const message of messages) {
+        if (message.role === "tool" && message.tool_call_id) {
+            referencedToolCallIds.add(message.tool_call_id);
+        }
+    }
+    for (const message of messages) {
+        // Always keep system messages (they set important context)
+        if (message.role === "system") {
+            result.push(message);
+            continue;
+        }
+        // Always keep user messages (user might repeat intentionally)
+        if (message.role === "user") {
+            result.push(message);
+            continue;
+        }
+        // Always keep tool messages (they are results of tool calls)
+        // Removing them would break the tool_use/tool_result pairing
+        if (message.role === "tool") {
+            result.push(message);
+            continue;
+        }
+        // For assistant messages with tool_calls, check if any are referenced
+        // by subsequent tool messages - if so, we MUST keep this message
+        if (message.role === "assistant" && message.tool_calls) {
+            const hasReferencedToolCall = message.tool_calls.some((tc) => referencedToolCallIds.has(tc.id));
+            if (hasReferencedToolCall) {
+                // This assistant message has tool_calls that are referenced - keep it
+                result.push(message);
+                continue;
+            }
+        }
+        // For other assistant messages, check for duplicates
+        const hash = hashMessage(message);
+        if (!seen.has(hash)) {
+            seen.add(hash);
+            result.push(message);
+        }
+        else {
+            duplicatesRemoved++;
+        }
+    }
+    return {
+        messages: result,
+        duplicatesRemoved,
+        originalCount: messages.length,
+    };
+}

package/dist/compression/layers/dictionary.d.ts

@@ -0,0 +1,20 @@
+/**
+ * Layer 3: Dictionary Encoding
+ *
+ * Replaces frequently repeated long phrases with short codes.
+ * Uses a static codebook of common patterns from production logs.
+ *
+ * Safe for LLM: Reversible substitution with codebook header.
+ * Expected savings: 4-8%
+ */
+import { NormalizedMessage } from "../types.js";
+export interface DictionaryResult {
+    messages: NormalizedMessage[];
+    substitutionCount: number;
+    usedCodes: Set<string>;
+    charsSaved: number;
+}
+/**
+ * Apply dictionary encoding to all messages.
+ */
+export declare function encodeMessages(messages: NormalizedMessage[]): DictionaryResult;

package/dist/compression/layers/dictionary.js

@@ -0,0 +1,67 @@
+/**
+ * Layer 3: Dictionary Encoding
+ *
+ * Replaces frequently repeated long phrases with short codes.
+ * Uses a static codebook of common patterns from production logs.
+ *
+ * Safe for LLM: Reversible substitution with codebook header.
+ * Expected savings: 4-8%
+ */
+import { getInverseCodebook } from "../codebook.js";
+/**
+ * Apply dictionary encoding to a string.
+ * Returns the encoded string and stats.
+ */
+function encodeContent(content, inverseCodebook) {
+    let encoded = content;
+    let substitutions = 0;
+    let charsSaved = 0;
+    const codes = new Set();
+    // Sort phrases by length (longest first) to avoid partial matches
+    const phrases = Object.keys(inverseCodebook).sort((a, b) => b.length - a.length);
+    for (const phrase of phrases) {
+        const code = inverseCodebook[phrase];
+        const regex = new RegExp(escapeRegex(phrase), "g");
+        const matches = encoded.match(regex);
+        if (matches && matches.length > 0) {
+            encoded = encoded.replace(regex, code);
+            substitutions += matches.length;
+            charsSaved += matches.length * (phrase.length - code.length);
+            codes.add(code);
+        }
+    }
+    return { encoded, substitutions, codes, charsSaved };
+}
+/**
+ * Escape special regex characters in a string.
+ */
+function escapeRegex(str) {
+    return str.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+/**
+ * Apply dictionary encoding to all messages.
+ */
+export function encodeMessages(messages) {
+    const inverseCodebook = getInverseCodebook();
+    let totalSubstitutions = 0;
+    let totalCharsSaved = 0;
+    const allUsedCodes = new Set();
+    const result = messages.map((message) => {
+        if (!message.content || typeof message.content !== "string")
+            return message;
+        const { encoded, substitutions, codes, charsSaved } = encodeContent(message.content, inverseCodebook);
+        totalSubstitutions += substitutions;
+        totalCharsSaved += charsSaved;
+        codes.forEach((code) => allUsedCodes.add(code));
+        return {
+            ...message,
+            content: encoded,
+        };
+    });
+    return {
+        messages: result,
+        substitutionCount: totalSubstitutions,
+        usedCodes: allUsedCodes,
+        charsSaved: totalCharsSaved,
+    };
+}

package/dist/compression/layers/dynamic-codebook.d.ts

@@ -0,0 +1,25 @@
+/**
+ * L7: Dynamic Codebook Builder
+ *
+ * Inspired by claw-compactor's frequency-based codebook.
+ * Builds codebook from actual content being compressed,
+ * rather than relying on static patterns.
+ *
+ * Finds phrases that appear 3+ times and replaces with short codes.
+ */
+import { NormalizedMessage } from "../types.js";
+interface DynamicCodebookResult {
+    messages: NormalizedMessage[];
+    charsSaved: number;
+    dynamicCodes: Record<string, string>;
+    substitutions: number;
+}
+/**
+ * Apply dynamic codebook to messages.
+ */
+export declare function applyDynamicCodebook(messages: NormalizedMessage[]): DynamicCodebookResult;
+/**
+ * Generate header for dynamic codes (to include in system message).
+ */
+export declare function generateDynamicCodebookHeader(codebook: Record<string, string>): string;
+export {};

package/dist/compression/layers/dynamic-codebook.js

@@ -0,0 +1,145 @@
+/**
+ * L7: Dynamic Codebook Builder
+ *
+ * Inspired by claw-compactor's frequency-based codebook.
+ * Builds codebook from actual content being compressed,
+ * rather than relying on static patterns.
+ *
+ * Finds phrases that appear 3+ times and replaces with short codes.
+ */
+// Config
+const MIN_PHRASE_LENGTH = 20;
+const MAX_PHRASE_LENGTH = 200;
+const MIN_FREQUENCY = 3;
+const MAX_ENTRIES = 100;
+const CODE_PREFIX = "$D"; // Dynamic codes: $D01, $D02, etc.
+/**
+ * Find repeated phrases in content.
+ */
+function findRepeatedPhrases(allContent) {
+    const phrases = new Map();
+    // Split by sentence-like boundaries
+    const segments = allContent.split(/(?<=[.!?\n])\s+/);
+    for (const segment of segments) {
+        const trimmed = segment.trim();
+        if (trimmed.length >= MIN_PHRASE_LENGTH &&
+            trimmed.length <= MAX_PHRASE_LENGTH) {
+            phrases.set(trimmed, (phrases.get(trimmed) || 0) + 1);
+        }
+    }
+    // Also find repeated lines
+    const lines = allContent.split("\n");
+    for (const line of lines) {
+        const trimmed = line.trim();
+        if (trimmed.length >= MIN_PHRASE_LENGTH &&
+            trimmed.length <= MAX_PHRASE_LENGTH) {
+            phrases.set(trimmed, (phrases.get(trimmed) || 0) + 1);
+        }
+    }
+    return phrases;
+}
+/**
+ * Build dynamic codebook from message content.
+ */
+function buildDynamicCodebook(messages) {
+    // Combine all content
+    let allContent = "";
+    for (const msg of messages) {
+        if (msg.content) {
+            allContent += msg.content + "\n";
+        }
+    }
+    // Find repeated phrases
+    const phrases = findRepeatedPhrases(allContent);
+    // Filter by frequency and sort by savings potential
+    const candidates = [];
+    for (const [phrase, count] of phrases.entries()) {
+        if (count >= MIN_FREQUENCY) {
+            // Savings = (phrase length - code length) * occurrences
+            const codeLength = 4; // e.g., "$D01"
+            const savings = (phrase.length - codeLength) * count;
+            if (savings > 50) {
+                candidates.push({ phrase, count, savings });
+            }
+        }
+    }
+    // Sort by savings (descending) and take top entries
+    candidates.sort((a, b) => b.savings - a.savings);
+    const topCandidates = candidates.slice(0, MAX_ENTRIES);
+    // Build codebook
+    const codebook = {};
+    topCandidates.forEach((c, i) => {
+        const code = `${CODE_PREFIX}${String(i + 1).padStart(2, "0")}`;
+        codebook[code] = c.phrase;
+    });
+    return codebook;
+}
+/**
+ * Escape special regex characters.
+ */
+function escapeRegex(str) {
+    return str.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+}
+/**
+ * Apply dynamic codebook to messages.
+ */
+export function applyDynamicCodebook(messages) {
+    // Build codebook from content
+    const codebook = buildDynamicCodebook(messages);
+    if (Object.keys(codebook).length === 0) {
+        return {
+            messages,
+            charsSaved: 0,
+            dynamicCodes: {},
+            substitutions: 0,
+        };
+    }
+    // Create inverse map for replacement
+    const phraseToCode = {};
+    for (const [code, phrase] of Object.entries(codebook)) {
+        phraseToCode[phrase] = code;
+    }
+    // Sort phrases by length (longest first) to avoid partial replacements
+    const sortedPhrases = Object.keys(phraseToCode).sort((a, b) => b.length - a.length);
+    let charsSaved = 0;
+    let substitutions = 0;
+    // Apply replacements
+    const result = messages.map((msg) => {
+        if (!msg.content || typeof msg.content !== "string")
+            return msg;
+        let content = msg.content;
+        for (const phrase of sortedPhrases) {
+            const code = phraseToCode[phrase];
+            const regex = new RegExp(escapeRegex(phrase), "g");
+            const matches = content.match(regex);
+            if (matches) {
+                content = content.replace(regex, code);
+                charsSaved += (phrase.length - code.length) * matches.length;
+                substitutions += matches.length;
+            }
+        }
+        return { ...msg, content };
+    });
+    return {
+        messages: result,
+        charsSaved,
+        dynamicCodes: codebook,
+        substitutions,
+    };
+}
+/**
+ * Generate header for dynamic codes (to include in system message).
+ */
+export function generateDynamicCodebookHeader(codebook) {
+    if (Object.keys(codebook).length === 0)
+        return "";
+    const entries = Object.entries(codebook)
+        .slice(0, 20) // Limit header size
+        .map(([code, phrase]) => {
+        // Truncate long phrases in header
+        const displayPhrase = phrase.length > 40 ? phrase.slice(0, 37) + "..." : phrase;
+        return `${code}=${displayPhrase}`;
+    })
+        .join(", ");
+    return `[DynDict: ${entries}]`;
+}

package/dist/compression/layers/json-compact.d.ts

@@ -0,0 +1,22 @@
+/**
+ * Layer 5: JSON Compaction
+ *
+ * Minifies JSON in tool_call arguments and tool results.
+ * Removes pretty-print whitespace from JSON strings.
+ *
+ * Safe for LLM: JSON semantics unchanged.
+ * Expected savings: 2-4%
+ */
+import { NormalizedMessage } from "../types.js";
+export interface JsonCompactResult {
+    messages: NormalizedMessage[];
+    charsSaved: number;
+}
+/**
+ * Apply JSON compaction to all messages.
+ *
+ * Targets:
+ * - tool_call arguments (in assistant messages)
+ * - tool message content (often JSON)
+ */
+export declare function compactMessagesJson(messages: NormalizedMessage[]): JsonCompactResult;

package/dist/compression/layers/json-compact.js

@@ -0,0 +1,74 @@
+/**
+ * Layer 5: JSON Compaction
+ *
+ * Minifies JSON in tool_call arguments and tool results.
+ * Removes pretty-print whitespace from JSON strings.
+ *
+ * Safe for LLM: JSON semantics unchanged.
+ * Expected savings: 2-4%
+ */
+/**
+ * Compact a JSON string by parsing and re-stringifying without formatting.
+ */
+function compactJson(jsonString) {
+    try {
+        const parsed = JSON.parse(jsonString);
+        return JSON.stringify(parsed);
+    }
+    catch {
+        // Not valid JSON, return as-is
+        return jsonString;
+    }
+}
+/**
+ * Check if a string looks like JSON (starts with { or [).
+ */
+function looksLikeJson(str) {
+    const trimmed = str.trim();
+    return ((trimmed.startsWith("{") && trimmed.endsWith("}")) ||
+        (trimmed.startsWith("[") && trimmed.endsWith("]")));
+}
+/**
+ * Compact tool_call arguments in a message.
+ */
+function compactToolCalls(toolCalls) {
+    return toolCalls.map((tc) => ({
+        ...tc,
+        function: {
+            ...tc.function,
+            arguments: compactJson(tc.function.arguments),
+        },
+    }));
+}
+/**
+ * Apply JSON compaction to all messages.
+ *
+ * Targets:
+ * - tool_call arguments (in assistant messages)
+ * - tool message content (often JSON)
+ */
+export function compactMessagesJson(messages) {
+    let charsSaved = 0;
+    const result = messages.map((message) => {
+        const newMessage = { ...message };
+        // Compact tool_calls arguments
+        if (message.tool_calls && message.tool_calls.length > 0) {
+            const originalLength = JSON.stringify(message.tool_calls).length;
+            newMessage.tool_calls = compactToolCalls(message.tool_calls);
+            const newLength = JSON.stringify(newMessage.tool_calls).length;
+            charsSaved += originalLength - newLength;
+        }
+        // Compact tool message content if it looks like JSON
+        if (message.role === "tool" && message.content && typeof message.content === "string" && looksLikeJson(message.content)) {
+            const originalLength = message.content.length;
+            const compacted = compactJson(message.content);
+            charsSaved += originalLength - compacted.length;
+            newMessage.content = compacted;
+        }
+        return newMessage;
+    });
+    return {
+        messages: result,
+        charsSaved,
+    };
+}

package/dist/compression/layers/observation.d.ts

@@ -0,0 +1,20 @@
+/**
+ * L6: Observation Compression (AGGRESSIVE)
+ *
+ * Inspired by claw-compactor's 97% compression on tool results.
+ * Tool call results (especially large ones) are summarized to key info only.
+ *
+ * This is the biggest compression win - tool outputs can be 10KB+ but
+ * only ~200 chars of actual useful information.
+ */
+import { NormalizedMessage } from "../types.js";
+interface ObservationResult {
+    messages: NormalizedMessage[];
+    charsSaved: number;
+    observationsCompressed: number;
+}
+/**
+ * Compress tool results in messages.
+ */
+export declare function compressObservations(messages: NormalizedMessage[]): ObservationResult;
+export {};

package/dist/compression/layers/observation.js

@@ -0,0 +1,126 @@
+/**
+ * L6: Observation Compression (AGGRESSIVE)
+ *
+ * Inspired by claw-compactor's 97% compression on tool results.
+ * Tool call results (especially large ones) are summarized to key info only.
+ *
+ * This is the biggest compression win - tool outputs can be 10KB+ but
+ * only ~200 chars of actual useful information.
+ */
+// Max length for tool results before compression kicks in
+const TOOL_RESULT_THRESHOLD = 500;
+// Max length to compress tool results down to
+const COMPRESSED_RESULT_MAX = 300;
+/**
+ * Extract key information from tool result.
+ * Keeps: errors, key values, status, first/last important lines.
+ */
+function compressToolResult(content) {
+    if (!content || content.length <= TOOL_RESULT_THRESHOLD) {
+        return content;
+    }
+    const lines = content.split("\n").map((l) => l.trim()).filter(Boolean);
+    // Priority 1: Error messages (always keep)
+    const errorLines = lines.filter((l) => /error|exception|failed|denied|refused|timeout|invalid/i.test(l) &&
+        l.length < 200);
+    // Priority 2: Status/result lines
+    const statusLines = lines.filter((l) => /success|complete|created|updated|found|result|status|total|count/i.test(l) &&
+        l.length < 150);
+    // Priority 3: Key JSON fields (extract important values)
+    const jsonMatches = [];
+    const jsonPattern = /"(id|name|status|error|message|count|total|url|path)":\s*"?([^",}\n]+)"?/gi;
+    let match;
+    while ((match = jsonPattern.exec(content)) !== null) {
+        jsonMatches.push(`${match[1]}: ${match[2].slice(0, 50)}`);
+    }
+    // Priority 4: First and last meaningful lines
+    const firstLine = lines[0]?.slice(0, 100);
+    const lastLine = lines.length > 1 ? lines[lines.length - 1]?.slice(0, 100) : "";
+    // Build compressed observation
+    const parts = [];
+    if (errorLines.length > 0) {
+        parts.push("[ERR] " + errorLines.slice(0, 3).join(" | "));
+    }
+    if (statusLines.length > 0) {
+        parts.push(statusLines.slice(0, 3).join(" | "));
+    }
+    if (jsonMatches.length > 0) {
+        parts.push(jsonMatches.slice(0, 5).join(", "));
+    }
+    if (parts.length === 0) {
+        // Fallback: keep first/last lines with truncation marker
+        parts.push(firstLine || "");
+        if (lines.length > 2) {
+            parts.push(`[...${lines.length - 2} lines...]`);
+        }
+        if (lastLine && lastLine !== firstLine) {
+            parts.push(lastLine);
+        }
+    }
+    let result = parts.join("\n");
+    // Final length cap
+    if (result.length > COMPRESSED_RESULT_MAX) {
+        result = result.slice(0, COMPRESSED_RESULT_MAX - 20) + "\n[...truncated]";
+    }
+    return result;
+}
+/**
+ * Compress large repeated content blocks.
+ * Detects when same large block appears multiple times.
+ */
+function deduplicateLargeBlocks(messages) {
+    const blockHashes = new Map(); // hash -> first occurrence index
+    let charsSaved = 0;
+    const result = messages.map((msg, idx) => {
+        if (!msg.content || typeof msg.content !== "string" || msg.content.length < 500) {
+            return msg;
+        }
+        // Hash first 200 chars as block identifier
+        const blockKey = msg.content.slice(0, 200);
+        if (blockHashes.has(blockKey)) {
+            const firstIdx = blockHashes.get(blockKey);
+            const original = msg.content;
+            const compressed = `[See message #${firstIdx + 1} - same content]`;
+            charsSaved += original.length - compressed.length;
+            return { ...msg, content: compressed };
+        }
+        blockHashes.set(blockKey, idx);
+        return msg;
+    });
+    return { messages: result, charsSaved };
+}
+/**
+ * Compress tool results in messages.
+ */
+export function compressObservations(messages) {
+    let charsSaved = 0;
+    let observationsCompressed = 0;
+    // First pass: compress individual tool results
+    let result = messages.map((msg) => {
+        // Only compress tool role messages (these are tool call results)
+        if (msg.role !== "tool" || !msg.content || typeof msg.content !== "string") {
+            return msg;
+        }
+        const original = msg.content;
+        if (original.length <= TOOL_RESULT_THRESHOLD) {
+            return msg;
+        }
+        const compressed = compressToolResult(original);
+        const saved = original.length - compressed.length;
+        if (saved > 50) {
+            charsSaved += saved;
+            observationsCompressed++;
+            return { ...msg, content: compressed };
+        }
+        return msg;
+    });
+    // Second pass: deduplicate large repeated blocks
+    const dedupResult = deduplicateLargeBlocks(result);
+    result = dedupResult.messages;
+    charsSaved += dedupResult.charsSaved;
+    return {
+        messages: result,
+        charsSaved,
+        observationsCompressed,
+    };
+}

package/dist/compression/layers/paths.d.ts

@@ -0,0 +1,23 @@
+/**
+ * Layer 4: Path Shortening
+ *
+ * Detects common filesystem path prefixes and replaces them with short codes.
+ * Common in coding assistant contexts with repeated file paths.
+ *
+ * Safe for LLM: Lossless abbreviation with path map header.
+ * Expected savings: 1-3%
+ */
+import { NormalizedMessage } from "../types.js";
+export interface PathShorteningResult {
+    messages: NormalizedMessage[];
+    pathMap: Record<string, string>;
+    charsSaved: number;
+}
+/**
+ * Apply path shortening to all messages.
+ */
+export declare function shortenPaths(messages: NormalizedMessage[]): PathShorteningResult;
+/**
+ * Generate the path map header for the codebook.
+ */
+export declare function generatePathMapHeader(pathMap: Record<string, string>): string;