@diousk/pi-subagents-fast 0.20.0

package/src/output-file.ts

@@ -0,0 +1,155 @@
+/**
+ * output-file.ts — Streaming JSONL output file for agent transcripts.
+ *
+ * Creates a per-agent output file that streams conversation turns as JSONL,
+ * matching Claude Code's task output file format.
+ */
+
+import { appendFileSync, chmodSync, mkdirSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import type { AgentSession, AgentSessionEvent } from "@earendil-works/pi-coding-agent";
+
+/**
+ * Project/global default for writing a subagent's `.output` transcript; a custom
+ * agent's `output_transcript` overrides it per agent.
+ *
+ * State lives here rather than in an index.ts closure because both spawn paths
+ * need it — the top-level Agent tool and the nested delegation tools. Same
+ * reason `scopeModels` lives in model-scope.ts: a setting only one path can read
+ * is a setting the other path silently ignores.
+ */
+let outputTranscriptDefault = true;
+
+export function getOutputTranscriptDefault(): boolean { return outputTranscriptDefault; }
+export function setOutputTranscriptDefault(b: boolean): void { outputTranscriptDefault = b; }
+
+/**
+ * Encode a cwd path as a filesystem-safe directory name. Handles:
+ *   - POSIX:   "/home/user/project"        → "home-user-project"
+ *   - Windows: "C:\Users\foo\project"      → "Users-foo-project"
+ *   - UNC:     "\\\\server\\share\\project"  → "server-share-project"
+ */
+export function encodeCwd(cwd: string): string {
+  return cwd
+    .replace(/[/\\]/g, "-")        // both separators → dash
+    .replace(/^[A-Za-z]:-/, "")    // strip Windows drive prefix ("C:-")
+    .replace(/^-+/, "");           // strip leading dashes (POSIX root, UNC)
+}
+
+/**
+ * The per-session scratch directory, created if missing.
+ * Mirrors Claude Code's layout: /tmp/{prefix}-{uid}/{encoded-cwd}/{sessionId}/tasks
+ *
+ * Shared with the workflow tool, which persists each invocation's script here so
+ * iterating on one is edit-file-then-rerun — the same convention, one directory.
+ */
+export function sessionTaskDir(cwd: string, sessionId: string): string {
+  const encoded = encodeCwd(cwd);
+  const root = join(tmpdir(), `pi-subagents-${process.getuid?.() ?? 0}`);
+  mkdirSync(root, { recursive: true, mode: 0o700 });
+  // chmod is a no-op on Windows and throws on some Windows filesystems.
+  // On Unix we still want to enforce 0o700 past umask, so only swallow on Windows.
+  try {
+    chmodSync(root, 0o700);
+  } catch (err) {
+    if (process.platform !== "win32") throw err;
+  }
+  const dir = join(root, encoded, sessionId, "tasks");
+  mkdirSync(dir, { recursive: true });
+  return dir;
+}
+
+/** Create the output file path, ensuring the directory exists. */
+export function createOutputFilePath(cwd: string, agentId: string, sessionId: string): string {
+  return join(sessionTaskDir(cwd, sessionId), `${agentId}.output`);
+}
+
+/**
+ * Ensure a transcript file exists without disturbing what is already in it.
+ *
+ * A resume reuses the agent's existing transcript (same deterministic path), so
+ * it must never call `writeInitialEntry` — that truncates, discarding turns the
+ * completion notification still points the user at, and any history the session
+ * has since compacted away is gone for good. Appending nothing creates the file
+ * when this is the agent's first transcript and is a no-op when it is not.
+ */
+export function ensureOutputFile(path: string): void {
+  try {
+    appendFileSync(path, "", "utf-8");
+  } catch { /* ignore — streaming writes are best-effort too */ }
+}
+
+/** Write the initial user prompt entry. */
+export function writeInitialEntry(path: string, agentId: string, prompt: string, cwd: string): void {
+  const entry = {
+    isSidechain: true,
+    agentId,
+    type: "user",
+    message: { role: "user", content: prompt },
+    timestamp: new Date().toISOString(),
+    cwd,
+  };
+  writeFileSync(path, JSON.stringify(entry) + "\n", "utf-8");
+}
+
+/**
+ * Subscribe to session events and flush new messages to the output file on each turn_end.
+ * Returns a cleanup function that does a final flush and unsubscribes.
+ */
+export function streamToOutputFile(
+  session: AgentSession,
+  path: string,
+  agentId: string,
+  cwd: string,
+  startIndex?: number,
+): () => void {
+  // Index of the first message this stream is responsible for. A spawn writes
+  // messages[0] as the initial prompt entry, so it starts at 1. A resume hands
+  // in the session's length as of just before the run: the session already
+  // holds every prior turn, and re-emitting those would duplicate history that
+  // is already in the file.
+  let writtenCount = startIndex ?? 1;
+
+  const flush = () => {
+    const messages = session.messages;
+    while (writtenCount < messages.length) {
+      const msg = messages[writtenCount];
+      const entry = {
+        isSidechain: true,
+        agentId,
+        type: msg.role === "assistant" ? "assistant" : msg.role === "user" ? "user" : "toolResult",
+        message: msg,
+        timestamp: new Date().toISOString(),
+        cwd,
+      };
+      try {
+        appendFileSync(path, JSON.stringify(entry) + "\n", "utf-8");
+      } catch { /* ignore write errors */ }
+      writtenCount++;
+    }
+  };
+
+  const unsubscribe = session.subscribe((event: AgentSessionEvent) => {
+    if (event.type === "turn_end") flush();
+    // Compaction replaces session.messages with a shorter, summarized array,
+    // leaving writtenCount past the new end — without re-anchoring, the flush
+    // loop would never match again and streaming would halt for good (#145).
+    // Flush before it runs so any not-yet-flushed tail still reaches the file,
+    // then re-anchor to the rebuilt array once it lands. The re-anchor is
+    // deferred a microtask because on the overflow-retry path pi trims the
+    // trailing error assistant message AFTER emitting compaction_end —
+    // anchoring synchronously would sit one past the trimmed array and skip
+    // the first post-compaction message. Aborted/failed compactions leave
+    // session.messages untouched, so only successful ones re-anchor.
+    if (event.type === "compaction_start") flush();
+    if (event.type === "compaction_end" && !event.aborted && event.result) {
+      queueMicrotask(() => { writtenCount = session.messages.length; });
+    }
+  });
+
+  return () => {
+    flush();
+    unsubscribe();
+  };
+}

package/src/prompts.ts

@@ -0,0 +1,142 @@
+/**
+ * prompts.ts — System prompt builder for agents.
+ */
+
+import type { AgentConfig, EnvInfo } from "./types.js";
+
+/** Extra sections to inject into the system prompt (memory, skills, etc.). */
+export interface PromptExtras {
+  /** Persistent memory content to inject (first 200 lines of MEMORY.md + instructions). */
+  memoryBlock?: string;
+  /** Preloaded skill contents to inject. */
+  skillBlocks?: { name: string; content: string }[];
+  /**
+   * Parent directory the worktree copy was created from. Set only for
+   * `isolation: "worktree"` spawns — triggers the block that tells the agent
+   * to stay in the copy.
+   */
+  worktreeBase?: string;
+  /**
+   * Set only for a workflow's own children, and only when they have no
+   * `StructuredOutput` tool to answer through.
+   *
+   * A workflow child's final text is not read by a human — it is the value
+   * `agent()` resolves to, and the script interpolates it straight into the
+   * next stage's prompt. Without this, children answer the way every other
+   * subagent does (a report addressed to a reader), and the padding becomes
+   * input tokens for the stage downstream. Claude Code's `Workflow` tool
+   * documents this contract to the script-writing model; this is the end of it
+   * that makes the documentation true.
+   *
+   * Deliberately NOT applied to every subagent. In pi an ordinary agent's
+   * output IS read by a human — through FleetView, the conversation viewer and
+   * `get_subagent_result` — so terse raw data would be the wrong answer there.
+   */
+  workflowChild?: boolean;
+}
+
+/**
+ * Build the system prompt for an agent from its config.
+ *
+ * - "replace" mode: env header + config.systemPrompt (full control, no parent identity)
+ * - "append" mode: parent system prompt + sub-agent context + env header + config.systemPrompt
+ * - "append" with empty systemPrompt: pure parent clone
+ *
+ * Both modes include an `<active_agent name="${config.name}"/>` tag so downstream
+ * extensions (e.g. permission/policy systems) can resolve per-agent policy
+ * inside the child session by parsing the system prompt. In replace mode the tag
+ * is prepended; in append mode it follows the shared inherited content so the
+ * parent prompt forms an identical, cacheable byte prefix with the parent
+ * session (the LLM's KV cache can then reuse those tokens across every spawn).
+ *
+ * @param parentSystemPrompt  The parent agent's effective system prompt (for append mode).
+ * @param extras  Optional extra sections to inject (memory, preloaded skills).
+ */
+export function buildAgentPrompt(
+  config: AgentConfig,
+  cwd: string,
+  env: EnvInfo,
+  parentSystemPrompt?: string,
+  extras?: PromptExtras,
+): string {
+  const activeAgentTag = `<active_agent name="${config.name}"/>\n\n`;
+
+  const envBlock = `# Environment
+Working directory: ${cwd}
+${env.isGitRepo ? `Git repository: yes\nBranch: ${env.branch}` : "Not a git repository"}
+Platform: ${env.platform}`;
+
+  // A worktree agent is told its cwd twice: by the env block above (the copy)
+  // and by whatever names the main checkout — the inherited parent prompt in
+  // append mode, or the task prompt in either mode. It follows the latter and
+  // works in the shared tree (#187), so resolve the contradiction explicitly.
+  const worktreeBlock = extras?.worktreeBase
+    ? `\n\n<worktree_isolation>
+Your working directory is an isolated git worktree copy of ${extras.worktreeBase}.
+Work only inside it — never in ${extras.worktreeBase}, even if other instructions name that path as your working directory.
+</worktree_isolation>`
+    : "";
+
+  // The script, not a person, reads what this child returns — see
+  // `PromptExtras.workflowChild` for why only workflow children get this.
+  const workflowBlock = extras?.workflowChild
+    ? `\n\n<workflow_child>
+Your final message IS the return value of this task. A workflow script captures it and passes it to the next stage; no person reads it.
+Return only the answer, in exactly the shape the prompt asks for — no preamble, no summary of what you did, no offer to continue.
+</workflow_child>`
+    : "";
+
+  // Build optional extras suffix
+  const extraSections: string[] = [];
+  if (extras?.memoryBlock) {
+    extraSections.push(extras.memoryBlock);
+  }
+  if (extras?.skillBlocks?.length) {
+    for (const skill of extras.skillBlocks) {
+      extraSections.push(`\n# Preloaded Skill: ${skill.name}\n${skill.content}`);
+    }
+  }
+  const extrasSuffix = extraSections.length > 0 ? "\n\n" + extraSections.join("\n") : "";
+
+  if (config.promptMode === "append") {
+    const identity = parentSystemPrompt || genericBase;
+
+    const bridge = `<sub_agent_context>
+You are operating as a sub-agent invoked to handle a specific task.
+- Use the read tool instead of cat/head/tail
+- Use the edit tool instead of sed/awk
+- Use the write tool instead of echo/heredoc
+- Use the find tool instead of bash find/ls for file search
+- Use the grep tool instead of bash grep/rg for content search
+- Make independent tool calls in parallel
+- Use absolute file paths
+- Do not use emojis
+- Be concise but complete
+</sub_agent_context>`;
+
+    const customSection = config.systemPrompt?.trim()
+      ? `\n\n<agent_instructions>\n${config.systemPrompt}\n</agent_instructions>`
+      : "";
+
+    // Place shared/stable content first so the LLM's KV cache can reuse the
+    // inherited prefix across all subagent invocations. The parent prompt is
+    // placed verbatim (no wrapper tag) so it forms an identical byte prefix
+    // with the parent session, maximising KV cache hits. The <active_agent>
+    // tag and env block vary per call and are placed after the cached prefix.
+    return identity + "\n\n" + bridge + "\n\n" + activeAgentTag + envBlock + worktreeBlock + workflowBlock + customSection + extrasSuffix;
+  }
+
+  // "replace" mode — env header + the config's full system prompt
+  const replaceHeader = `You are a pi coding agent sub-agent.
+You have been invoked to handle a specific task autonomously.
+
+${envBlock}`;
+
+  return activeAgentTag + replaceHeader + worktreeBlock + workflowBlock + "\n\n" + config.systemPrompt + extrasSuffix;
+}
+
+/** Fallback base prompt when parent system prompt is unavailable in append mode. */
+const genericBase = `# Role
+You are a general-purpose coding agent for complex, multi-step tasks.
+You have full access to read, write, edit files, and execute commands.
+Do what has been asked; nothing more, nothing less.`;

package/src/schedule-store.ts

@@ -0,0 +1,153 @@
+/**
+ * schedule-store.ts — File-backed store for scheduled subagents.
+ *
+ * Session-scoped: each pi session owns its own schedules at
+ * `<cwd>/.pi/subagent-schedules/<sessionId>.json`. `/new` starts a fresh
+ * empty store; `/resume` reloads.
+ *
+ * Concurrency model lifted from pi-chonky-tasks/src/task-store.ts: every
+ * mutation acquires a PID-based exclusion lock, re-reads the latest state
+ * from disk, applies the change, atomic-writes via temp+rename, releases.
+ */
+
+import { existsSync, mkdirSync, readFileSync, renameSync, unlinkSync, writeFileSync } from "node:fs";
+import { dirname, join } from "node:path";
+import type { ScheduledSubagent, ScheduleStoreData } from "./types.js";
+
+const LOCK_RETRY_MS = 50;
+const LOCK_MAX_RETRIES = 100;
+
+function isProcessRunning(pid: number): boolean {
+  try { process.kill(pid, 0); return true; } catch { return false; }
+}
+
+function acquireLock(lockPath: string): void {
+  for (let i = 0; i < LOCK_MAX_RETRIES; i++) {
+    try {
+      writeFileSync(lockPath, `${process.pid}`, { flag: "wx" });
+      return;
+    } catch (e: any) {
+      if (e.code === "EEXIST") {
+        try {
+          const pid = parseInt(readFileSync(lockPath, "utf-8"), 10);
+          if (pid && !isProcessRunning(pid)) {
+            unlinkSync(lockPath);
+            continue;
+          }
+        } catch { /* ignore — try again */ }
+        const start = Date.now();
+        while (Date.now() - start < LOCK_RETRY_MS) { /* busy wait */ }
+        continue;
+      }
+      throw e;
+    }
+  }
+  throw new Error(`Failed to acquire schedule lock: ${lockPath}`);
+}
+
+function releaseLock(lockPath: string): void {
+  try { unlinkSync(lockPath); } catch { /* ignore */ }
+}
+
+/** Resolve the storage path for a session-scoped store. */
+export function resolveStorePath(cwd: string, sessionId: string): string {
+  return join(cwd, ".pi", "subagent-schedules", `${sessionId}.json`);
+}
+
+export class ScheduleStore {
+  private filePath: string;
+  private lockPath: string;
+  private jobs = new Map<string, ScheduledSubagent>();
+
+  constructor(filePath: string) {
+    this.filePath = filePath;
+    this.lockPath = filePath + ".lock";
+    this.load();
+  }
+
+  /** Create the backing directory lazily — only when we're about to persist. */
+  private ensureDir(): void {
+    mkdirSync(dirname(this.filePath), { recursive: true });
+  }
+
+  /** Load from disk into the in-memory cache. Silent on parse errors. */
+  private load(): void {
+    if (!existsSync(this.filePath)) return;
+    try {
+      const data: ScheduleStoreData = JSON.parse(readFileSync(this.filePath, "utf-8"));
+      this.jobs.clear();
+      for (const j of data.jobs ?? []) this.jobs.set(j.id, j);
+    } catch { /* corrupt — start fresh, next save rewrites */ }
+  }
+
+  /** Atomic write via temp file + rename (POSIX-atomic). */
+  private save(): void {
+    const data: ScheduleStoreData = { version: 1, jobs: [...this.jobs.values()] };
+    const tmp = this.filePath + ".tmp";
+    writeFileSync(tmp, JSON.stringify(data, null, 2));
+    renameSync(tmp, this.filePath);
+  }
+
+  /** Acquire lock → reload → mutate → save → release. */
+  private withLock<T>(fn: () => T): T {
+    this.ensureDir();
+    acquireLock(this.lockPath);
+    try {
+      this.load();
+      const result = fn();
+      this.save();
+      return result;
+    } finally {
+      releaseLock(this.lockPath);
+    }
+  }
+
+  /** Read-only — returns a snapshot of the in-memory cache. */
+  list(): ScheduledSubagent[] {
+    return [...this.jobs.values()];
+  }
+
+  /** Read-only check — uses the cache. */
+  hasName(name: string, exceptId?: string): boolean {
+    for (const j of this.jobs.values()) {
+      if (j.id !== exceptId && j.name === name) return true;
+    }
+    return false;
+  }
+
+  get(id: string): ScheduledSubagent | undefined {
+    return this.jobs.get(id);
+  }
+
+  add(job: ScheduledSubagent): void {
+    this.withLock(() => {
+      this.jobs.set(job.id, job);
+    });
+  }
+
+  update(id: string, patch: Partial<ScheduledSubagent>): ScheduledSubagent | undefined {
+    // No-op fast path — an unknown id changes nothing, so don't lock or touch
+    // disk (which would otherwise lazily create the backing directory).
+    if (!this.jobs.has(id)) return undefined;
+    return this.withLock(() => {
+      const existing = this.jobs.get(id);
+      if (!existing) return undefined;
+      const updated = { ...existing, ...patch };
+      this.jobs.set(id, updated);
+      return updated;
+    });
+  }
+
+  remove(id: string): boolean {
+    // No-op fast path — see update().
+    if (!this.jobs.has(id)) return false;
+    return this.withLock(() => this.jobs.delete(id));
+  }
+
+  /** Delete the backing file (used when no jobs remain, optional cleanup). */
+  deleteFileIfEmpty(): void {
+    if (this.jobs.size === 0 && existsSync(this.filePath)) {
+      try { unlinkSync(this.filePath); } catch { /* ignore */ }
+    }
+  }
+}