@unerr-ai/unerr 0.1.0 → 0.1.2

package/dist/intelligence/smart-truncate.js

@@ -0,0 +1,158 @@
+/**
+ * Structural-Priority Truncation — renders entity content within a token budget.
+ *
+ * Priority hierarchy (adapted from rtk's AggressiveFilter + graphify's _subgraph_to_text):
+ *   1. Metadata — always included (name, kind, file, community, blast radius, corrections)
+ *   2. Imports  — included if budget allows
+ *   3. Signatures — function/class/struct signatures without bodies
+ *   4. Bodies — full implementations, included line-by-line until budget exhausted
+ *
+ * Contracts:
+ *   SC-10: Token accounting overhead <1ms
+ *   SC-12: Never cuts mid-line — operates on complete lines only
+ *   Decision L.5: Default budget 2000 tokens
+ *   Decision L.6: Structural-priority over character-cut
+ *
+ * Leapfrog Sprint C, Task C.2
+ */
+const CHARS_PER_TOKEN = 4;
+const TRUNCATION_MARKER_RESERVE = 120;
+/**
+ * Estimate token count from character length.
+ * Uses 4 chars/token (conservative for code — actual is ~3.5 for English, ~4.5 for code).
+ */
+export function estimateTokens(text) {
+    return Math.ceil(text.length / CHARS_PER_TOKEN);
+}
+/**
+ * Truncate entity content within a token budget using structural priority.
+ *
+ * Small entities fit entirely within the budget and return `truncated: false`.
+ * Large entities are truncated at the appropriate structural level, with a
+ * marker indicating how many lines were omitted and the token_budget needed
+ * to retrieve the full content.
+ *
+ * Performance: <1ms for any input size (string concatenation + length checks only).
+ */
+export function smartTruncate(input) {
+    const budget = Math.max(input.budget, 100);
+    const charBudget = budget * CHARS_PER_TOKEN;
+    const sections = [
+        input.metadata,
+        input.imports,
+        input.signatures,
+        input.bodies,
+    ].filter(Boolean);
+    const fullContent = sections.join("\n\n");
+    const fullTokens = estimateTokens(fullContent);
+    if (fullContent.length <= charBudget) {
+        return {
+            content: fullContent,
+            tokens_used: fullTokens,
+            tokens_budget: budget,
+            truncated: false,
+            truncation_level: "full",
+            full_tokens_estimate: fullTokens,
+        };
+    }
+    let result = input.metadata;
+    let level = "metadata_only";
+    // Priority 2: imports
+    if (input.imports &&
+        result.length + input.imports.length + TRUNCATION_MARKER_RESERVE <
+            charBudget) {
+        result += `\n\n${input.imports}`;
+    }
+    // Priority 3: signatures
+    if (input.signatures &&
+        result.length + input.signatures.length + TRUNCATION_MARKER_RESERVE <
+            charBudget) {
+        result += `\n\n${input.signatures}`;
+        level = "signatures_only";
+    }
+    // Priority 4: bodies — line-by-line (SC-12: never cut mid-line)
+    if (input.bodies) {
+        const remainingChars = charBudget - result.length - TRUNCATION_MARKER_RESERVE;
+        if (remainingChars > 0) {
+            const bodyLines = input.bodies.split("\n");
+            const includedLines = [];
+            let bodyChars = 0;
+            for (const line of bodyLines) {
+                const lineLen = line.length + 1; // +1 for newline
+                if (bodyChars + lineLen > remainingChars)
+                    break;
+                includedLines.push(line);
+                bodyChars += lineLen;
+            }
+            if (includedLines.length > 0) {
+                result += `\n\n${includedLines.join("\n")}`;
+                level =
+                    includedLines.length === bodyLines.length
+                        ? "full"
+                        : "signatures_and_bodies";
+            }
+        }
+    }
+    if (level !== "full") {
+        const totalBodyLines = input.bodies ? input.bodies.split("\n").length : 0;
+        const includedBodyLines = level === "signatures_and_bodies" || level === "signatures_only"
+            ? countIncludedBodyLines(result, input.bodies)
+            : 0;
+        const omittedLines = totalBodyLines - includedBodyLines;
+        if (omittedLines > 0) {
+            result += `\n\n// ... ${omittedLines} lines omitted. Request with token_budget: ${fullTokens} for full content.`;
+        }
+        else {
+            result += `\n\n// ... truncated. Request with token_budget: ${fullTokens} for full content.`;
+        }
+    }
+    return {
+        content: result,
+        tokens_used: estimateTokens(result),
+        tokens_budget: budget,
+        truncated: level !== "full",
+        truncation_level: level,
+        full_tokens_estimate: fullTokens,
+    };
+}
+/**
+ * Count how many body lines were included in the result.
+ */
+function countIncludedBodyLines(result, bodies) {
+    if (!bodies)
+        return 0;
+    const bodyLines = bodies.split("\n");
+    let count = 0;
+    for (const line of bodyLines) {
+        if (result.includes(line)) {
+            count++;
+        }
+        else {
+            break;
+        }
+    }
+    return count;
+}
+/**
+ * Apply token budget to an array of result items (for list-returning tools).
+ * Keeps items until the budget is exhausted, then appends a summary.
+ */
+export function truncateResultList(items, budget, serialize) {
+    const charBudget = Math.max(budget, 100) * CHARS_PER_TOKEN;
+    let chars = 0;
+    const kept = [];
+    for (const item of items) {
+        const serialized = serialize(item);
+        if (chars + serialized.length + 100 > charBudget && kept.length > 0) {
+            break;
+        }
+        kept.push(item);
+        chars += serialized.length;
+    }
+    return {
+        items: kept,
+        truncated: kept.length < items.length,
+        total: items.length,
+        tokens_used: estimateTokens(kept.map(serialize).join("\n")),
+    };
+}

package/dist/intelligence/subgraph-cache.js

@@ -0,0 +1,88 @@
+/**
+ * Pre-emptive Subgraph Cache — preloads 1-hop neighbors on first directory access.
+ *
+ * S.3: When an agent first queries a file in a directory, pre-warm the cache
+ * with intelligence for all sibling files in that directory. This eliminates
+ * latency on subsequent queries within the same module.
+ */
+const TTL_MS = 5 * 60 * 1000;
+const MAX_ENTRIES = 500;
+/** RC-5: Maximum tracked directories before eviction. */
+const MAX_DIRECTORIES = 200;
+export function createSubgraphCache() {
+    const cache = new Map();
+    const directoriesWarmed = new Set();
+    function get(filePath) {
+        const entry = cache.get(filePath);
+        if (!entry)
+            return null;
+        if (Date.now() - entry.cachedAt > TTL_MS) {
+            cache.delete(filePath);
+            return null;
+        }
+        return entry;
+    }
+    function warmDirectory(dirPath, files, loader) {
+        if (directoriesWarmed.has(dirPath))
+            return;
+        directoriesWarmed.add(dirPath);
+        // RC-5: Evict oldest directories if set exceeds limit
+        if (directoriesWarmed.size > MAX_DIRECTORIES) {
+            const iter = directoriesWarmed.values();
+            for (let i = 0; i < 50; i++) {
+                const v = iter.next().value;
+                if (v !== undefined)
+                    directoriesWarmed.delete(v);
+            }
+        }
+        for (const file of files.slice(0, 20)) {
+            if (cache.has(file))
+                continue;
+            if (cache.size >= MAX_ENTRIES)
+                break;
+            try {
+                const data = loader(file);
+                cache.set(file, { filePath: file, cachedAt: Date.now(), data });
+            }
+            catch {
+                /* skip failed loads */
+            }
+        }
+    }
+    function has(filePath) {
+        const entry = cache.get(filePath);
+        if (!entry)
+            return false;
+        if (Date.now() - entry.cachedAt > TTL_MS) {
+            cache.delete(filePath);
+            return false;
+        }
+        return true;
+    }
+    function invalidate(filePath) {
+        cache.delete(filePath);
+    }
+    function invalidateDirectory(dirPath) {
+        directoriesWarmed.delete(dirPath);
+        for (const [key] of cache) {
+            if (key.startsWith(dirPath))
+                cache.delete(key);
+        }
+    }
+    function size() {
+        return cache.size;
+    }
+    function clear() {
+        cache.clear();
+        directoriesWarmed.clear();
+    }
+    return {
+        get,
+        warmDirectory,
+        has,
+        invalidate,
+        invalidateDirectory,
+        size,
+        clear,
+    };
+}