@ai-devkit/agent-manager 0.36.0 → 0.38.0

package/src/utils/jsonlTail.ts

@@ -0,0 +1,349 @@
+/**
+ * Tail reader for append-only JSONL transcripts.
+ *
+ * Reading the last N conversation messages from a large transcript should not
+ * require reading (and JSON-parsing) the whole file. This reader:
+ *
+ * 1. On the first call for a file, reads backward from the end in fixed-size
+ *    chunks until enough messages are collected (or the start is reached).
+ * 2. Remembers `{dev, ino, offset}` plus a small window of parsed items, and
+ *    on later calls parses only bytes appended since `offset`.
+ * 3. Resets (re-reads from the end) when the file is truncated or replaced by
+ *    a different inode.
+ *
+ * Each adapter supplies a {@link JsonlTailSource}: `parseLine` maps one raw
+ * line to a candidate item (or null) and the optional `finalize` turns a
+ * contiguous window of items into messages. `finalize` exists for formats
+ * where one line's visibility depends on other lines (e.g. Codex drops an
+ * `event_msg` mirrored by a `response_item`); line-independent formats can
+ * omit it and return a {@link ConversationMessage} directly from `parseLine`.
+ */
+
+import * as fs from "fs";
+import type { ConversationMessage } from "../adapters/AgentAdapter.js";
+
+/**
+ * Normalize a `tail` option: a positive integer, or undefined for "full
+ * conversation" (the pre-existing behaviour).
+ */
+export function normalizeTail(tail: number | undefined): number | undefined {
+  if (tail === undefined || !Number.isFinite(tail)) return undefined;
+  const n = Math.floor(tail);
+  return n > 0 ? n : undefined;
+}
+
+/** Keep the last `tail` messages of an already fully-parsed conversation. */
+export function sliceTail<T>(messages: T[], tail: number | undefined): T[] {
+  const n = normalizeTail(tail);
+  return n !== undefined && messages.length > n ? messages.slice(-n) : messages;
+}
+
+export const DEFAULT_TAIL_CHUNK_SIZE = 64 * 1024;
+const DEFAULT_MAX_FILES = 16;
+const NEWLINE = 0x0a;
+const ANCHOR_BYTES = 32;
+
+export interface JsonlTailSource<Item> {
+  /** Map one raw JSONL line (no trailing newline) to an item, or null to skip it. */
+  parseLine(line: string): Item | null;
+  /**
+   * Convert a contiguous window of items (oldest first) into messages. Must
+   * behave like the adapter's full parse restricted to that window. Defaults
+   * to treating every item as a {@link ConversationMessage}.
+   */
+  finalize?(items: Item[]): ConversationMessage[];
+}
+
+export interface JsonlTailReaderOptions {
+  /** Bytes read per backward step. Defaults to 64 KiB. */
+  chunkSize?: number;
+  /** Number of (file, variant, tail) windows kept in the LRU cache. */
+  maxFiles?: number;
+}
+
+interface TailState {
+  dev: number;
+  ino: number;
+  size: number;
+  mtimeMs: number;
+  /** Byte offset just past the last newline consumed; appends are read from here. */
+  offset: number;
+  /**
+   * Up to {@link ANCHOR_BYTES} bytes ending at `offset`. Re-checked before an
+   * incremental read so a same-inode truncate-and-regrow is detected.
+   */
+  anchor: Buffer;
+  /** Parsed items from complete (newline-terminated) lines, oldest first. */
+  items: unknown[];
+  /** Last computed result, returned as-is while the file is unchanged. */
+  result: ConversationMessage[];
+}
+
+export class JsonlTailReader {
+  private readonly chunkSize: number;
+  private readonly maxFiles: number;
+  private readonly cache = new Map<string, TailState>();
+
+  constructor(options: JsonlTailReaderOptions = {}) {
+    this.chunkSize = Math.max(1, options.chunkSize ?? DEFAULT_TAIL_CHUNK_SIZE);
+    this.maxFiles = Math.max(1, options.maxFiles ?? DEFAULT_MAX_FILES);
+  }
+
+  /**
+   * Return the last `tail` messages of `filePath`.
+   *
+   * @param variant  Cache discriminator for sources whose output differs by
+   *                 option (e.g. verbose vs. non-verbose).
+   * @param overscan Extra messages collected beyond `tail` before trimming,
+   *                 for sources whose `finalize` looks at neighbouring lines.
+   */
+  read<Item>(
+    filePath: string,
+    tail: number,
+    source: JsonlTailSource<Item>,
+    variant = "",
+    overscan = 0,
+  ): ConversationMessage[] {
+    const key = `${variant}\0${tail}\0${filePath}`;
+    const threshold = tail + Math.max(0, overscan);
+
+    let stat: fs.Stats;
+    try {
+      stat = fs.statSync(filePath);
+    } catch {
+      this.cache.delete(key);
+      return [];
+    }
+
+    const cached = this.cache.get(key);
+    if (
+      cached &&
+      cached.dev === stat.dev &&
+      cached.ino === stat.ino &&
+      cached.size === stat.size &&
+      cached.mtimeMs === stat.mtimeMs
+    ) {
+      this.touch(key, cached);
+      return cached.result.slice();
+    }
+
+    let fd: number;
+    try {
+      fd = fs.openSync(filePath, "r");
+    } catch {
+      this.cache.delete(key);
+      return [];
+    }
+
+    try {
+      const current = fs.fstatSync(fd);
+      const canAppend =
+        cached !== undefined &&
+        cached.dev === current.dev &&
+        cached.ino === current.ino &&
+        current.size >= cached.offset &&
+        // A very large append is cheaper to re-read from the end.
+        current.size - cached.offset <= Math.max(this.chunkSize * 16, 1024 * 1024);
+
+      const state =
+        (canAppend && this.readAppended(fd, current, cached, tail, threshold, source)) ||
+        this.readBackward(fd, current, tail, threshold, source);
+      this.touch(key, state);
+      return state.result.slice();
+    } catch {
+      this.cache.delete(key);
+      return [];
+    } finally {
+      fs.closeSync(fd);
+    }
+  }
+
+  /** Forget all cached windows. */
+  clear(): void {
+    this.cache.clear();
+  }
+
+  private readBackward<Item>(
+    fd: number,
+    stat: fs.Stats,
+    tail: number,
+    threshold: number,
+    source: JsonlTailSource<Item>,
+  ): TailState {
+    const reversed: Item[] = [];
+    let partial: Item | null = null;
+    let offset = 0;
+    let sawNewline = false;
+    // Pieces of the line currently being assembled, in file order.
+    let pending: Buffer[] = [];
+    let pos = stat.size;
+
+    const takeLine = (piece: Buffer): void => {
+      const lineBuf = pending.length === 0 ? piece : Buffer.concat([piece, ...pending]);
+      pending = [];
+      if (!sawNewline) {
+        // Bytes after the file's last newline: an unterminated (possibly
+        // still being written) line. Parsed tentatively, never cached.
+        sawNewline = true;
+        partial = this.parse(lineBuf, source);
+        return;
+      }
+      const item = this.parse(lineBuf, source);
+      if (item !== null) reversed.push(item);
+    };
+
+    while (pos > 0) {
+      const len = Math.min(this.chunkSize, pos);
+      pos -= len;
+      const buf = readExactly(fd, pos, len);
+
+      let end = len;
+      let nl = end > 0 ? buf.lastIndexOf(NEWLINE, end - 1) : -1;
+      while (nl !== -1) {
+        if (!sawNewline) offset = pos + nl + 1;
+        takeLine(buf.subarray(nl + 1, end));
+        end = nl;
+        nl = end > 0 ? buf.lastIndexOf(NEWLINE, end - 1) : -1;
+      }
+      if (end > 0) pending.unshift(buf.subarray(0, end));
+
+      if (this.countMessages(reversed, partial, source, threshold) >= threshold) break;
+    }
+
+    if (pos === 0 && pending.length > 0) {
+      // The first line of the file (no newline before it).
+      const first = pending.shift()!;
+      takeLine(first);
+    }
+
+    const items = reversed.reverse();
+    return {
+      dev: stat.dev,
+      ino: stat.ino,
+      size: stat.size,
+      mtimeMs: stat.mtimeMs,
+      offset,
+      anchor: readAnchor(fd, offset),
+      items,
+      result: this.finish(items, partial, source, tail),
+    };
+  }
+
+  /**
+   * Parse bytes appended since the cached offset. Returns null when the
+   * bytes before the offset no longer match (file rewritten in place).
+   */
+  private readAppended<Item>(
+    fd: number,
+    stat: fs.Stats,
+    cached: TailState,
+    tail: number,
+    threshold: number,
+    source: JsonlTailSource<Item>,
+  ): TailState | null {
+    let items = cached.items as Item[];
+    let offset = cached.offset;
+    let partial: Item | null = null;
+
+    const anchorLength = cached.anchor.length;
+    const raw = readExactly(fd, cached.offset - anchorLength, stat.size - offset + anchorLength);
+    if (!raw.subarray(0, anchorLength).equals(cached.anchor)) return null;
+    const buf = raw.subarray(anchorLength);
+    let anchor = cached.anchor;
+
+    if (buf.length > 0) {
+      const added: Item[] = [];
+      let start = 0;
+      let nl = buf.indexOf(NEWLINE, start);
+      while (nl !== -1) {
+        const item = this.parse(buf.subarray(start, nl), source);
+        if (item !== null) added.push(item);
+        start = nl + 1;
+        nl = buf.indexOf(NEWLINE, start);
+      }
+      offset = cached.offset + start;
+      if (start > 0) anchor = tailBytes(raw.subarray(0, anchorLength + start));
+      if (start < buf.length) partial = this.parse(buf.subarray(start), source);
+
+      if (added.length > 0) {
+        items = items.concat(added);
+        // Bound the retained window; keep a generous margin so sources whose
+        // finalize() drops items (e.g. de-duplication) still cover `tail`.
+        if (items.length > threshold * 4) items = items.slice(-threshold * 2);
+      }
+    }
+
+    return {
+      dev: stat.dev,
+      ino: stat.ino,
+      size: stat.size,
+      mtimeMs: stat.mtimeMs,
+      offset,
+      anchor,
+      items,
+      result: this.finish(items, partial, source, tail),
+    };
+  }
+
+  private parse<Item>(lineBuf: Buffer, source: JsonlTailSource<Item>): Item | null {
+    const line = lineBuf.toString("utf-8");
+    if (line.trim().length === 0) return null;
+    return source.parseLine(line);
+  }
+
+  private countMessages<Item>(
+    reversed: Item[],
+    partial: Item | null,
+    source: JsonlTailSource<Item>,
+    threshold: number,
+  ): number {
+    const available = reversed.length + (partial === null ? 0 : 1);
+    if (!source.finalize || available < threshold) return available;
+    const window = reversed.slice().reverse();
+    if (partial !== null) window.push(partial);
+    return source.finalize(window).length;
+  }
+
+  private finish<Item>(
+    items: Item[],
+    partial: Item | null,
+    source: JsonlTailSource<Item>,
+    tail: number,
+  ): ConversationMessage[] {
+    const window = partial === null ? items : [...items, partial];
+    const messages = source.finalize
+      ? source.finalize(window)
+      : (window as unknown as ConversationMessage[]).slice();
+    return messages.length > tail ? messages.slice(-tail) : messages;
+  }
+
+  private touch(key: string, state: TailState): void {
+    this.cache.delete(key);
+    this.cache.set(key, state);
+    while (this.cache.size > this.maxFiles) {
+      const oldest = this.cache.keys().next().value;
+      if (oldest === undefined) break;
+      this.cache.delete(oldest);
+    }
+  }
+}
+
+function readExactly(fd: number, position: number, length: number): Buffer {
+  const buf = Buffer.allocUnsafe(length);
+  let filled = 0;
+  while (filled < length) {
+    const n = fs.readSync(fd, buf, filled, length - filled, position + filled);
+    if (n === 0) break;
+    filled += n;
+  }
+  return filled === length ? buf : buf.subarray(0, filled);
+}
+
+function readAnchor(fd: number, offset: number): Buffer {
+  const start = Math.max(0, offset - ANCHOR_BYTES);
+  return Buffer.from(readExactly(fd, start, offset - start));
+}
+
+function tailBytes(buf: Buffer): Buffer {
+  return Buffer.from(buf.subarray(Math.max(0, buf.length - ANCHOR_BYTES)));
+}

package/src/utils/matching.ts

@@ -7,10 +7,10 @@
 
 import * as path from "path";
 import type { ProcessInfo } from "../adapters/AgentAdapter.js";
+import { MATCH_TOLERANCE_MS } from "./matchingConstants.js";
 import type { SessionFile } from "./session.js";
 
-/** Maximum allowed delta between process start time and session file birth time. */
-const TOLERANCE_MS = 3 * 60 * 1000; // 3 minutes
+export { MATCH_TOLERANCE_MS };
 
 /**
  * Result of matching a process to a session file.
@@ -55,7 +55,7 @@ export function matchProcessesToSessions(
       if (proc.cwd !== session.resolvedCwd) continue;
 
       const deltaMs = Math.abs(processStartMs - session.birthtimeMs);
-      if (deltaMs > TOLERANCE_MS) continue;
+      if (deltaMs > MATCH_TOLERANCE_MS) continue;
 
       candidates.push({ process: proc, session, deltaMs });
     }

package/src/utils/matchingConstants.ts

@@ -0,0 +1,15 @@
+/**
+ * Session Matching Constants
+ *
+ * Kept apart from matching.ts so session locators can import them even when
+ * adapter tests fully mock matching.ts.
+ */
+
+/**
+ * Maximum allowed delta between process start time and session file birth time.
+ *
+ * Single source of truth for matchProcessesToSessions and for locator
+ * discovery windows: a window narrower than this would skip files the matcher
+ * would accept.
+ */
+export const MATCH_TOLERANCE_MS = 3 * 60 * 1000; // 3 minutes