@tagma/sdk 0.6.0 → 0.6.1

package/src/engine.ts

@@ -1,1220 +1,1220 @@
-import { resolve } from 'path';
-import { readdir, rm } from 'fs/promises';
-import type {
-  PipelineConfig,
-  TaskConfig,
-  TaskState,
-  TaskStatus,
-  TaskResult,
-  DriverPlugin,
-  TriggerPlugin,
-  CompletionPlugin,
-  MiddlewarePlugin,
-  MiddlewareContext,
-  DriverContext,
-  OnFailure,
-  PromptDocument,
-  Permissions,
-  AbortReason,
-  RunEventPayload,
-  RunTaskState,
-} from './types';
-import { buildDag, type Dag } from './dag';
-import { getHandler, hasHandler, loadPlugins } from './registry';
-import { runSpawn, runCommand } from './runner';
-import { parseDuration, nowISO, generateRunId } from './utils';
-import { promptDocumentFromString, serializePromptDocument } from './prompt-doc';
-import {
-  executeHook,
-  buildPipelineStartContext,
-  buildTaskContext,
-  buildPipelineCompleteContext,
-  buildPipelineErrorContext,
-  type PipelineInfo,
-  type TrackInfo,
-  type TaskInfo,
-} from './hooks';
-import { Logger, tailLines, clip } from './logger';
-import { InMemoryApprovalGateway, type ApprovalGateway } from './approval';
-
-// ═══ A7: Typed trigger errors ═══
-// Replace string-matching on error messages with structured error types so
-// coincidental substrings don't cause misclassification.
-
-export class TriggerBlockedError extends Error {
-  readonly code = 'TRIGGER_BLOCKED' as const;
-  constructor(message: string) {
-    super(message);
-    this.name = 'TriggerBlockedError';
-  }
-}
-
-export class TriggerTimeoutError extends Error {
-  readonly code = 'TRIGGER_TIMEOUT' as const;
-  constructor(message: string) {
-    super(message);
-    this.name = 'TriggerTimeoutError';
-  }
-}
-
-// ═══ Preflight Validation ═══
-
-function preflight(config: PipelineConfig, dag: Dag): void {
-  const errors: string[] = [];
-
-  for (const [, node] of dag.nodes) {
-    const task = node.task;
-    const track = node.track;
-    const driverName = task.driver ?? track.driver ?? config.driver ?? 'opencode';
-
-    // Pure command tasks don't use a driver — skip driver registration check.
-    const isCommandOnly = task.command && !task.prompt;
-
-    if (!isCommandOnly && !hasHandler('drivers', driverName)) {
-      errors.push(`Task "${node.taskId}": driver "${driverName}" not registered`);
-    }
-
-    if (task.trigger && !hasHandler('triggers', task.trigger.type)) {
-      errors.push(`Task "${node.taskId}": trigger type "${task.trigger.type}" not registered`);
-    }
-
-    if (task.completion && !hasHandler('completions', task.completion.type)) {
-      errors.push(
-        `Task "${node.taskId}": completion type "${task.completion.type}" not registered`,
-      );
-    }
-
-    const mws = task.middlewares ?? track.middlewares ?? [];
-    for (const mw of mws) {
-      if (!hasHandler('middlewares', mw.type)) {
-        errors.push(`Task "${node.taskId}": middleware type "${mw.type}" not registered`);
-      }
-    }
-
-    if (task.continue_from && hasHandler('drivers', driverName)) {
-      const driver = getHandler<DriverPlugin>('drivers', driverName);
-      if (!driver.capabilities.sessionResume) {
-        // buildDag has already qualified `continue_from` and stored the result
-        // on the node; preflight runs after buildDag, so the upstream id is
-        // always available here without re-resolving.
-        const upstreamId = node.resolvedContinueFrom;
-        if (upstreamId) {
-          const upstream = dag.nodes.get(upstreamId);
-          if (upstream) {
-            // A handoff is possible via session resume (already ruled out above),
-            // OR in-memory text injection through normalizedMap
-            // (when the upstream driver implements parseResult and returns normalizedOutput).
-            const upstreamDriverName =
-              upstream.task.driver ?? upstream.track.driver ?? config.driver ?? 'opencode';
-            const upstreamDriver = hasHandler('drivers', upstreamDriverName)
-              ? getHandler<DriverPlugin>('drivers', upstreamDriverName)
-              : null;
-            const canNormalize = typeof upstreamDriver?.parseResult === 'function';
-
-            if (!canNormalize) {
-              errors.push(
-                `Task "${node.taskId}" uses continue_from: "${task.continue_from}", ` +
-                  `but upstream task "${upstreamId}" its driver ` +
-                  `does not implement parseResult for text-injection handoff. ` +
-                  `Use a driver with parseResult, or remove continue_from.`,
-              );
-            }
-          }
-        }
-      }
-    }
-  }
-
-  if (errors.length > 0) {
-    throw new Error(`Preflight validation failed:\n  - ${errors.join('\n  - ')}`);
-  }
-}
-
-// ═══ Engine ═══
-
-export interface EngineResult {
-  readonly success: boolean;
-  readonly runId: string;
-  readonly logPath: string;
-  readonly summary: {
-    total: number;
-    success: number;
-    failed: number;
-    skipped: number;
-    timeout: number;
-    blocked: number;
-  };
-  readonly states: ReadonlyMap<string, TaskState>;
-}
-
-// ═══ Pipeline Events ═══
-//
-// The engine emits RunEventPayload values (defined in @tagma/types) via
-// `onEvent`. Every payload carries `runId`; the editor server stamps a
-// per-run `seq` before broadcasting. There is one event vocabulary
-// end-to-end — no server-side translation layer.
-
-// Re-export so SDK consumers can import the event type without reaching
-// into @tagma/types directly.
-export type { RunEventPayload } from './types';
-
-// ═══ Helpers ═══
-
-/**
- * Project the engine's internal TaskState onto the wire RunTaskState
- * shape. `logs` / `totalLogCount` default to empty — they are populated
- * on the server side from streamed `task_log` events, not from state.
- */
-function toRunTaskState(
-  taskId: string,
-  trackId: string,
-  taskName: string,
-  state: TaskState,
-): RunTaskState {
-  const result = state.result;
-  const cfg = state.config;
-  return {
-    taskId,
-    trackId,
-    taskName,
-    status: state.status,
-    startedAt: state.startedAt,
-    finishedAt: state.finishedAt,
-    durationMs: result?.durationMs ?? null,
-    exitCode: result?.exitCode ?? null,
-    stdout: result?.stdout ?? '',
-    stderr: result?.stderr ?? '',
-    stderrPath: result?.stderrPath ?? null,
-    sessionId: result?.sessionId ?? null,
-    normalizedOutput: result?.normalizedOutput ?? null,
-    resolvedDriver: cfg.driver ?? null,
-    resolvedModel: cfg.model ?? null,
-    resolvedPermissions: (cfg.permissions as Permissions | undefined) ?? null,
-    logs: [],
-    totalLogCount: 0,
-  };
-}
-
-export interface RunPipelineOptions {
-  readonly approvalGateway?: ApprovalGateway;
-  /**
-   * Maximum number of per-run log directories to retain under `<workDir>/.tagma/logs/`.
-   * Oldest directories are deleted after each run. Defaults to 20. Set to 0 to disable cleanup.
-   */
-  readonly maxLogRuns?: number;
-  /**
-   * Caller-supplied run ID. When provided the engine uses this instead of
-   * generating its own via `generateRunId()`, keeping the editor and SDK
-   * log directories aligned on the same ID.
-   */
-  readonly runId?: string;
-  /**
-   * External AbortSignal — aborting it cancels the pipeline immediately.
-   * Equivalent to the pipeline timeout firing, but caller-controlled.
-   */
-  readonly signal?: AbortSignal;
-  /**
-   * Called on every pipeline/task status transition.
-   * Use for real-time UI updates (e.g. updating a visual workflow graph).
-   */
-  readonly onEvent?: (event: RunEventPayload) => void;
-  /**
-   * Skip the engine's built-in `loadPlugins(config.plugins)` call.
-   * Use this when the host has already pre-loaded plugins from a custom
-   * resolution path (e.g. a user workspace's node_modules) so the engine
-   * doesn't re-resolve them via Node's default cwd-based import.
-   */
-  readonly skipPluginLoading?: boolean;
-}
-
-// Poll interval when no tasks are in-flight but non-terminal tasks remain
-// (e.g. tasks waiting on a file or manual trigger).
-const POLL_INTERVAL_MS = 50;
-
-// R15: cap on each normalized-output entry stored in normalizedMap so a
-// runaway parseResult can't accumulate hundreds of MB across tasks. 1 MB
-// is generous for any text-context handoff between AI tasks.
-const MAX_NORMALIZED_BYTES = 1_000_000;
-
-export async function runPipeline(
-  config: PipelineConfig,
-  workDir: string,
-  options: RunPipelineOptions = {},
-): Promise<EngineResult> {
-  const approvalGateway = options.approvalGateway ?? new InMemoryApprovalGateway();
-  const maxLogRuns = options.maxLogRuns ?? 20;
-
-  // Load any plugins declared in the pipeline config before preflight so that
-  // drivers, completions, and middlewares referenced in YAML are registered.
-  // Hosts that pre-load plugins from a custom path (e.g. the editor loading
-  // from the user's workspace node_modules) pass skipPluginLoading: true so
-  // we don't re-resolve via Node's cwd-based default import.
-  if (!options.skipPluginLoading && config.plugins?.length) {
-    await loadPlugins(config.plugins);
-  }
-
-  const dag = buildDag(config);
-  const runId = options.runId ?? generateRunId();
-  preflight(config, dag);
-
-  const startedAt = nowISO();
-  const pipelineInfo: PipelineInfo = { name: config.name, run_id: runId, started_at: startedAt };
-  // Forward every structured log line to subscribers as task_log events.
-  // Reading options.onEvent inside the callback (vs. capturing it once) keeps
-  // the SDK behavior correct if callers pass a fresh onEvent on each run.
-  const log = new Logger(workDir, runId, (record) => {
-    options.onEvent?.({
-      type: 'task_log',
-      runId,
-      taskId: record.taskId,
-      level: record.level,
-      timestamp: record.timestamp,
-      text: record.text,
-    });
-  });
-
-  try {
-    log.info('[pipeline]', `start "${config.name}" run_id=${runId}`);
-
-    // File-only: dump the resolved pipeline shape + DAG topology for post-mortem.
-    log.section('Pipeline configuration');
-    log.quiet(`name:          ${config.name}`);
-    log.quiet(`driver:        ${config.driver ?? '(default: opencode)'}`);
-    log.quiet(`timeout:       ${config.timeout ?? '(none)'}`);
-    log.quiet(`tracks:        ${config.tracks.length}`);
-    log.quiet(`tasks (total): ${dag.nodes.size}`);
-    log.quiet(`plugins:       ${(config.plugins ?? []).join(', ') || '(none)'}`);
-    log.quiet(
-      `hooks:         ${config.hooks ? Object.keys(config.hooks).join(', ') || '(none)' : '(none)'}`,
-    );
-
-    log.section('DAG topology');
-    for (const [id, node] of dag.nodes) {
-      const deps = node.dependsOn.length ? node.dependsOn.join(', ') : '(root)';
-      const kind = node.task.prompt ? 'ai' : 'cmd';
-      log.quiet(`  • ${id}  [${kind}]  track=${node.track.id}  deps=[${deps}]`);
-    }
-    log.quiet('');
-
-    // Initialize states (before hook, so we can return them even if blocked)
-    const states = new Map<string, TaskState>();
-    for (const [id, node] of dag.nodes) {
-      states.set(id, {
-        config: node.task,
-        trackConfig: node.track,
-        status: 'idle',
-        result: null,
-        startedAt: null,
-        finishedAt: null,
-      });
-    }
-
-    // Pipeline start hook (gate). Runs BEFORE the engine emits run_start so
-    // a blocked pipeline produces zero wire events (the server treats the
-    // thrown error as run_error). Hosts get a rich error message; nothing
-    // is ever half-broadcast.
-    const startHook = await executeHook(
-      config.hooks,
-      'pipeline_start',
-      buildPipelineStartContext(pipelineInfo),
-      workDir,
-    );
-    if (!startHook.allowed) {
-      console.error(`Pipeline blocked by pipeline_start hook (exit code ${startHook.exitCode})`);
-      await executeHook(
-        config.hooks,
-        'pipeline_error',
-        buildPipelineErrorContext(pipelineInfo, 'pipeline_blocked', 'pipeline_blocked'),
-        workDir,
-      );
-      return {
-        success: false,
-        runId,
-        logPath: log.path,
-        summary: {
-          total: dag.nodes.size,
-          success: 0,
-          failed: 0,
-          skipped: 0,
-          timeout: 0,
-          blocked: 0,
-        },
-        states: freezeStates(states),
-      };
-    }
-
-    // Pipeline approved — transition all tasks to waiting.
-    for (const [, state] of states) {
-      state.status = 'waiting';
-    }
-    // Emit run_start with a wire-shape snapshot so SSE subscribers can
-    // initialize their task maps on the same event stream that carries
-    // updates. No separate "server pre-broadcasts run_start" ceremony —
-    // the engine owns the lifecycle boundary.
-    const runStartTasks: RunTaskState[] = [];
-    for (const [id, node] of dag.nodes) {
-      const s = states.get(id)!;
-      runStartTasks.push(toRunTaskState(id, node.track.id, node.task.name ?? id, s));
-    }
-    emit({ type: 'run_start', runId, tasks: runStartTasks });
-
-    const sessionMap = new Map<string, string>();
-    const normalizedMap = new Map<string, string>();
-
-    // Pipeline timeout + abort reason tracking.
-    //
-    // `abortReason` replaces the previous `pipelineAborted: boolean`: it
-    // carries the concrete cause (timeout / stop_all / external) through
-    // to run_end and the pipeline_error hook so downstream consumers can
-    // distinguish them without scraping message strings.
-    const pipelineTimeoutMs = config.timeout ? parseDuration(config.timeout) : 0;
-    let abortReason: AbortReason | null = null;
-    const abortController = new AbortController();
-    let pipelineTimer: ReturnType<typeof setTimeout> | null = null;
-
-    if (pipelineTimeoutMs > 0) {
-      pipelineTimer = setTimeout(() => {
-        if (abortReason === null) abortReason = 'timeout';
-        abortController.abort();
-      }, pipelineTimeoutMs);
-    }
-
-    // When the pipeline is aborted (timeout, stop_all, external), drain
-    // all pending approvals so waiting triggers unblock immediately.
-    abortController.signal.addEventListener('abort', () => {
-      approvalGateway.abortAll('pipeline aborted');
-    });
-
-    // Wire external cancel signal into the internal abort controller.
-    const externalAbortHandler = () => {
-      if (abortReason === null) abortReason = 'external';
-      abortController.abort();
-    };
-    if (options.signal) {
-      if (options.signal.aborted) {
-        externalAbortHandler();
-      } else {
-        options.signal.addEventListener('abort', externalAbortHandler, { once: true });
-      }
-    }
-
-    // Bridge approval gateway events onto the wire stream so hosts (editor
-    // server, CLI adapters) see approvals on the same channel as task
-    // updates. The server no longer needs its own gateway subscription.
-    const unsubscribeApprovals = approvalGateway.subscribe((ev) => {
-      if (ev.type === 'requested') {
-        emit({
-          type: 'approval_request',
-          runId,
-          request: {
-            id: ev.request.id,
-            taskId: ev.request.taskId,
-            trackId: ev.request.trackId,
-            message: ev.request.message,
-            createdAt: ev.request.createdAt,
-            timeoutMs: ev.request.timeoutMs,
-            metadata: ev.request.metadata,
-          },
-        });
-        return;
-      }
-      if (ev.type === 'resolved' || ev.type === 'expired' || ev.type === 'aborted') {
-        const outcome =
-          ev.type === 'resolved'
-            ? ev.decision.outcome
-            : ev.type === 'expired'
-              ? 'timeout'
-              : 'aborted';
-        emit({
-          type: 'approval_resolved',
-          runId,
-          requestId: ev.request.id,
-          outcome,
-        });
-      }
-    });
-
-    // ── Helpers ──
-
-    function emit(event: RunEventPayload): void {
-      options.onEvent?.(event);
-    }
-
-    function setTaskStatus(taskId: string, newStatus: TaskStatus): void {
-      const state = states.get(taskId)!;
-      // Terminal lock: once a task reaches a terminal state it must not be
-      // re-transitioned. This prevents stop_all from marking running tasks as
-      // skipped and then having their in-flight processTask promise overwrite
-      // that with success/failed, producing an invalid double transition.
-      if (isTerminal(state.status)) return;
-      state.status = newStatus;
-      const result = state.result;
-      const cfg = state.config;
-      emit({
-        type: 'task_update',
-        runId,
-        taskId,
-        status: newStatus,
-        startedAt: state.startedAt ?? undefined,
-        finishedAt: state.finishedAt ?? undefined,
-        durationMs: result?.durationMs,
-        exitCode: result?.exitCode,
-        stdout: result?.stdout,
-        stderr: result?.stderr,
-        stderrPath: result?.stderrPath ?? null,
-        sessionId: result?.sessionId ?? null,
-        normalizedOutput: result?.normalizedOutput ?? null,
-        resolvedDriver: cfg.driver ?? null,
-        resolvedModel: cfg.model ?? null,
-        resolvedPermissions: (cfg.permissions as Permissions | undefined) ?? null,
-      });
-    }
-
-    function getOnFailure(taskId: string): OnFailure {
-      return dag.nodes.get(taskId)?.track.on_failure ?? 'skip_downstream';
-    }
-
-    function isDependencySatisfied(depId: string): 'satisfied' | 'unsatisfied' | 'skip' {
-      const depState = states.get(depId);
-      if (!depState) return 'skip';
-      switch (depState.status) {
-        case 'success':
-          return 'satisfied';
-        case 'skipped':
-          return 'skip';
-        case 'failed':
-        case 'timeout':
-        case 'blocked':
-          return getOnFailure(depId) === 'ignore' ? 'satisfied' : 'skip';
-        default:
-          return 'unsatisfied';
-      }
-    }
-
-    /**
-     * H3: "stop_all" historically only stopped tasks within the same track,
-     * which contradicted both its name and user expectations. It now stops
-     * the **entire pipeline**:
-     *   - In-flight tasks are signalled via the shared abort controller so
-     *     drivers / runner.ts can cancel cooperatively (returning
-     *     `failureKind: 'timeout'`).
-     *   - Still-waiting tasks across every track are immediately marked
-     *     skipped so the run completes promptly.
-     * The terminal lock in setTaskStatus prevents any later re-transition
-     * should a completed running task try to overwrite the skipped state.
-     */
-    function applyStopAll(_failedTrackId: string): void {
-      if (abortReason === null) abortReason = 'stop_all';
-      abortController.abort();
-      for (const [id, state] of states) {
-        if (state.status === 'waiting') {
-          state.finishedAt = nowISO();
-          setTaskStatus(id, 'skipped');
-        }
-      }
-    }
-
-    function buildTaskInfoObj(taskId: string): TaskInfo {
-      const state = states.get(taskId)!;
-      return {
-        id: taskId,
-        name: state.config.name,
-        type: state.config.prompt ? 'ai' : 'command',
-        status: state.status,
-        exit_code: state.result?.exitCode ?? null,
-        duration_ms: state.result?.durationMs ?? null,
-        stderr_path: state.result?.stderrPath ?? null,
-        session_id: state.result?.sessionId ?? null,
-        started_at: state.startedAt,
-        finished_at: state.finishedAt,
-      };
-    }
-
-    function trackInfoOf(taskId: string): TrackInfo {
-      const node = dag.nodes.get(taskId)!;
-      return { id: node.track.id, name: node.track.name };
-    }
-
-    async function fireHook(taskId: string, event: 'task_success' | 'task_failure'): Promise<void> {
-      await executeHook(
-        config.hooks,
-        event,
-        buildTaskContext(event, pipelineInfo, trackInfoOf(taskId), buildTaskInfoObj(taskId)),
-        workDir,
-        abortController.signal,
-      );
-    }
-
-    // ── Process a single task ──
-
-    async function processTask(taskId: string): Promise<void> {
-      const state = states.get(taskId)!;
-      const node = dag.nodes.get(taskId)!;
-      const task = node.task;
-      const track = node.track;
-
-      log.section(`Task ${taskId}`, taskId);
-      log.debug(
-        `[task:${taskId}]`,
-        `type=${task.prompt ? 'ai' : 'cmd'} track=${track.id} deps=[${node.dependsOn.join(', ') || '(root)'}]`,
-      );
-
-      // 1. Check dependencies
-      for (const depId of node.dependsOn) {
-        const result = isDependencySatisfied(depId);
-        if (result === 'skip') {
-          const depStatus = states.get(depId)?.status ?? 'unknown';
-          log.debug(`[task:${taskId}]`, `skipped (upstream "${depId}" status=${depStatus})`);
-          state.finishedAt = nowISO();
-          setTaskStatus(taskId, 'skipped');
-          return;
-        }
-        if (result === 'unsatisfied') return; // still waiting
-      }
-
-      // 2. Check trigger
-      if (task.trigger) {
-        log.debug(
-          `[task:${taskId}]`,
-          `trigger wait: type=${task.trigger.type} ${JSON.stringify(task.trigger)}`,
-        );
-        try {
-          const triggerPlugin = getHandler<TriggerPlugin>('triggers', task.trigger.type);
-          // R6: race the plugin's watch() against the pipeline's abort signal.
-          // Third-party triggers may forget to wire up ctx.signal — without
-          // this race, an aborted pipeline would hang forever waiting for the
-          // plugin's watch promise to resolve. The race resolves on whichever
-          // path settles first, and the cleanup paths in finally never run on
-          // the orphaned plugin promise (it's allowed to leak a watcher; the
-          // pipeline is being torn down anyway).
-          await new Promise<unknown>((resolve, reject) => {
-            let settled = false;
-            const onAbort = () => {
-              if (settled) return;
-              settled = true;
-              abortController.signal.removeEventListener('abort', onAbort);
-              reject(new Error('Pipeline aborted'));
-            };
-            if (abortController.signal.aborted) {
-              onAbort();
-              return;
-            }
-            abortController.signal.addEventListener('abort', onAbort, { once: true });
-            triggerPlugin
-              .watch(task.trigger as Record<string, unknown>, {
-                taskId: node.taskId,
-                trackId: track.id,
-                workDir: task.cwd ?? workDir,
-                signal: abortController.signal,
-                approvalGateway,
-              })
-              .then(
-                (v) => {
-                  if (settled) return;
-                  settled = true;
-                  abortController.signal.removeEventListener('abort', onAbort);
-                  resolve(v);
-                },
-                (e) => {
-                  if (settled) return;
-                  settled = true;
-                  abortController.signal.removeEventListener('abort', onAbort);
-                  reject(e);
-                },
-              );
-          });
-          log.debug(`[task:${taskId}]`, `trigger fired`);
-        } catch (err: unknown) {
-          // If pipeline was aborted while we were still waiting for the trigger,
-          // this task never entered running state → skipped, not timeout.
-          state.finishedAt = nowISO();
-          if (abortReason !== null) {
-            setTaskStatus(taskId, 'skipped');
-          } else if (err instanceof TriggerBlockedError) {
-            setTaskStatus(taskId, 'blocked'); // user/policy rejection
-          } else if (err instanceof TriggerTimeoutError) {
-            setTaskStatus(taskId, 'timeout'); // genuine trigger wait timeout
-          } else {
-            // A7 fallback: also check message strings for backward-compat with
-            // third-party trigger plugins that don't throw typed errors yet.
-            const msg = err instanceof Error ? err.message : String(err);
-            if (msg.includes('rejected') || msg.includes('denied')) {
-              setTaskStatus(taskId, 'blocked');
-            } else if (msg.includes('timeout')) {
-              setTaskStatus(taskId, 'timeout');
-            } else {
-              setTaskStatus(taskId, 'failed'); // plugin error, watcher crash, etc.
-            }
-          }
-          try {
-            await fireHook(taskId, 'task_failure');
-          } catch (hookErr) {
-            log.error(
-              `[task:${taskId}]`,
-              `hook execution failed: ${hookErr instanceof Error ? hookErr.message : String(hookErr)}`,
-            );
-          }
-          return;
-        }
-      }
-
-      // 3. task_start hook (gate)
-      const hookResult = await executeHook(
-        config.hooks,
-        'task_start',
-        buildTaskContext('task_start', pipelineInfo, trackInfoOf(taskId), buildTaskInfoObj(taskId)),
-        workDir,
-        abortController.signal,
-      );
-      if (hookResult.exitCode !== 0 || config.hooks?.task_start) {
-        log.debug(
-          `[task:${taskId}]`,
-          `task_start hook exit=${hookResult.exitCode} allowed=${hookResult.allowed}`,
-        );
-      }
-      if (!hookResult.allowed) {
-        state.finishedAt = nowISO();
-        setTaskStatus(taskId, 'blocked');
-        try {
-          await fireHook(taskId, 'task_failure');
-        } catch (hookErr) {
-          log.error(
-            `[task:${taskId}]`,
-            `hook execution failed: ${hookErr instanceof Error ? hookErr.message : String(hookErr)}`,
-          );
-        }
-        return;
-      }
-
-      // 4. Mark running — set startedAt before emitting so subscribers see a
-      // complete task_update (startedAt non-null) on the status transition.
-      state.startedAt = nowISO();
-      setTaskStatus(taskId, 'running');
-      log.info(
-        `[task:${taskId}]`,
-        task.command ? `running: ${task.command}` : `running (driver task)`,
-      );
-
-      // File-only: resolved config for this task
-      const resolvedDriver = task.driver ?? track.driver ?? config.driver ?? 'opencode';
-      const resolvedModel = task.model ?? track.model ?? config.model ?? '(default)';
-      const resolvedPerms = task.permissions ?? track.permissions ?? '(default)';
-      const resolvedCwd = task.cwd ?? track.cwd ?? workDir;
-      log.debug(
-        `[task:${taskId}]`,
-        `resolved: driver=${resolvedDriver} model=${resolvedModel} cwd=${resolvedCwd}`,
-      );
-      log.debug(`[task:${taskId}]`, `permissions: ${JSON.stringify(resolvedPerms)}`);
-      if (task.continue_from) {
-        log.debug(`[task:${taskId}]`, `continue_from: "${task.continue_from}"`);
-      }
-      if (task.timeout) {
-        log.debug(`[task:${taskId}]`, `timeout: ${task.timeout}`);
-      }
-
-      try {
-        let result: TaskResult;
-        const timeoutMs = task.timeout ? parseDuration(task.timeout) : undefined;
-
-        const runOpts = { timeoutMs, signal: abortController.signal };
-
-        if (task.command) {
-          log.debug(`[task:${taskId}]`, `command: ${task.command}`);
-          result = await runCommand(task.command, task.cwd ?? workDir, runOpts);
-        } else {
-          // AI task: apply middleware chain against a structured PromptDocument.
-          const driverName = task.driver ?? track.driver ?? config.driver ?? 'opencode';
-          const driver = getHandler<DriverPlugin>('drivers', driverName);
-
-          const originalLen = task.prompt!.length;
-          let doc: PromptDocument = promptDocumentFromString(task.prompt!);
-          const mws = task.middlewares !== undefined ? task.middlewares : track.middlewares;
-          if (mws && mws.length > 0) {
-            log.debug(
-              `[task:${taskId}]`,
-              `middleware chain: ${mws.map((m) => m.type).join(' → ')}`,
-            );
-            const mwCtx: MiddlewareContext = {
-              task,
-              track,
-              workDir: task.cwd ?? workDir,
-            };
-            for (const mwConfig of mws) {
-              const mwPlugin = getHandler<MiddlewarePlugin>('middlewares', mwConfig.type);
-              const beforeBlocks = doc.contexts.length;
-              const beforeLen = serializePromptDocument(doc).length;
-
-              // Prefer the structured API. Fall back to the legacy
-              // `enhance(string) → string` path so v0.x plugins keep
-              // working — that fallback loses context structure (the
-              // middleware's output becomes the new task body) but never
-              // silently drops content.
-              if (typeof mwPlugin.enhanceDoc === 'function') {
-                const next = await mwPlugin.enhanceDoc(
-                  doc,
-                  mwConfig as Record<string, unknown>,
-                  mwCtx,
-                );
-                if (
-                  !next ||
-                  typeof next !== 'object' ||
-                  !Array.isArray((next as PromptDocument).contexts) ||
-                  typeof (next as PromptDocument).task !== 'string'
-                ) {
-                  throw new Error(
-                    `middleware "${mwConfig.type}".enhanceDoc() returned a malformed PromptDocument`,
-                  );
-                }
-                doc = next as PromptDocument;
-              } else if (typeof mwPlugin.enhance === 'function') {
-                const asString = serializePromptDocument(doc);
-                const next = await mwPlugin.enhance(
-                  asString,
-                  mwConfig as Record<string, unknown>,
-                  mwCtx,
-                );
-                // R3: a middleware that returns undefined / null / a non-string
-                // would silently corrupt the prompt. Fail loud.
-                if (typeof next !== 'string') {
-                  throw new Error(
-                    `middleware "${mwConfig.type}".enhance() returned ${next === null ? 'null' : typeof next}, expected string`,
-                  );
-                }
-                // Legacy fallback: collapse the returned string into a
-                // fresh doc. Earlier structure is folded into the string
-                // (serializePromptDocument just ran), so bytes the driver
-                // sees match the old string pipeline.
-                doc = { contexts: [], task: next };
-              } else {
-                throw new Error(
-                  `middleware "${mwConfig.type}" provides neither enhanceDoc nor enhance`,
-                );
-              }
-              const afterLen = serializePromptDocument(doc).length;
-              const addedBlocks = doc.contexts.length - beforeBlocks;
-              log.debug(
-                `[task:${taskId}]`,
-                `  ${mwConfig.type}: ${beforeLen} → ${afterLen} chars` +
-                  (addedBlocks > 0
-                    ? ` (+${addedBlocks} context block${addedBlocks > 1 ? 's' : ''})`
-                    : ''),
-              );
-            }
-          }
-          const prompt = serializePromptDocument(doc);
-          log.debug(
-            `[task:${taskId}]`,
-            `prompt: ${originalLen} chars (final: ${prompt.length} chars, ${doc.contexts.length} block${doc.contexts.length === 1 ? '' : 's'})`,
-          );
-          log.quiet(`--- prompt (final) ---\n${clip(prompt)}\n--- end prompt ---`, taskId);
-
-          // H1: hand the driver a continue_from that has already been
-          // qualified by dag.ts. Without this, drivers like codex/opencode/
-          // claude-code look up maps directly with
-          // the user's raw (possibly bare) string, which races whenever two
-          // tracks share a task name. dag.ts has the only authoritative
-          // resolver, so we use its precomputed answer here.
-          // Drivers key sessionMap/normalizedMap by fully-qualified id. buildDag
-          // guarantees `resolvedContinueFrom` is set for every task that has a
-          // `continue_from`, so if we see the bare form here something upstream
-          // is broken — fail loud instead of silently miskeying the lookup.
-          if (task.continue_from && !node.resolvedContinueFrom) {
-            throw new Error(
-              `Internal: task "${taskId}" has continue_from "${task.continue_from}" ` +
-                `but no resolvedContinueFrom. buildDag should have qualified it.`,
-            );
-          }
-          const enrichedTask: TaskConfig = {
-            ...task,
-            prompt,
-            continue_from: node.resolvedContinueFrom,
-          };
-          const driverCtx: DriverContext = {
-            sessionMap,
-            normalizedMap,
-            workDir: task.cwd ?? workDir,
-            // Structured view for drivers that want fine-grained control
-            // over serialization (e.g. inserting [Previous Output] between
-            // contexts and task). Drivers that read task.prompt see the
-            // default serialization and need no changes.
-            promptDoc: doc,
-          };
-          const spec = await driver.buildCommand(enrichedTask, track, driverCtx);
-          log.debug(`[task:${taskId}]`, `driver=${driverName}`);
-          log.debug(`[task:${taskId}]`, `spawn args: ${JSON.stringify(spec.args)}`);
-          if (spec.cwd) log.debug(`[task:${taskId}]`, `spawn cwd: ${spec.cwd}`);
-          if (spec.env)
-            log.debug(
-              `[task:${taskId}]`,
-              `spawn env overrides: ${Object.keys(spec.env).join(', ')}`,
-            );
-          if (spec.stdin) log.debug(`[task:${taskId}]`, `spawn stdin: ${spec.stdin.length} chars`);
-          result = await runSpawn(spec, driver, runOpts);
-        }
-
-        // 6. Determine terminal status (without emitting yet — result must be complete first)
-        // H2: branch on failureKind so spawn errors no longer masquerade as
-        // timeouts. Old runners that don't set failureKind still work — we
-        // fall back to the historical `exitCode === -1 → timeout` heuristic so
-        // pre-existing third-party drivers don't regress.
-        let terminalStatus: TaskStatus;
-        const kind = result.failureKind;
-        if (kind === 'timeout') {
-          terminalStatus = 'timeout';
-        } else if (kind === 'spawn_error') {
-          terminalStatus = 'failed';
-        } else if (kind === undefined && result.exitCode === -1) {
-          // Legacy path: pre-H2 driver returned -1 with no kind. Treat as
-          // timeout for backward compatibility (the previous behaviour).
-          terminalStatus = 'timeout';
-        } else if (result.exitCode !== 0) {
-          terminalStatus = 'failed';
-        } else if (task.completion) {
-          const plugin = getHandler<CompletionPlugin>('completions', task.completion.type);
-          const completionCtx = { workDir: task.cwd ?? workDir, signal: abortController.signal };
-          const passed = await plugin.check(
-            task.completion as Record<string, unknown>,
-            result,
-            completionCtx,
-          );
-          // R4: strict boolean check. Truthy strings/numbers used to be coerced
-          // to success — a check returning "ok" would let a failing task pass.
-          if (typeof passed !== 'boolean') {
-            throw new Error(
-              `completion "${task.completion.type}".check() returned ${passed === null ? 'null' : typeof passed}, expected boolean`,
-            );
-          }
-          terminalStatus = passed ? 'success' : 'failed';
-        } else {
-          terminalStatus = 'success';
-        }
-
-        // Store normalized text separately (in-memory) for continue_from handoff.
-        // R15: clip oversized values so a runaway parseResult can't accumulate
-        // hundreds of MB across tasks.
-        if (result.normalizedOutput !== null) {
-          const clipped =
-            result.normalizedOutput.length > MAX_NORMALIZED_BYTES
-              ? result.normalizedOutput.slice(0, MAX_NORMALIZED_BYTES) +
-                `\n[…clipped at ${MAX_NORMALIZED_BYTES} bytes]`
-              : result.normalizedOutput;
-          normalizedMap.set(taskId, clipped);
-        }
-
-        if (result.stderr) {
-          const stderrPath = resolve(log.dir, `${taskId.replace(/\./g, '_')}.stderr`);
-          await Bun.write(stderrPath, result.stderr);
-          result = { ...result, stderrPath };
-        }
-
-        if (result.sessionId) {
-          // H1: qualified-only key.
-          sessionMap.set(taskId, result.sessionId);
-        }
-
-        // Set result and finishedAt before emitting terminal status so listeners see complete state
-        state.result = result;
-        state.finishedAt = nowISO();
-        setTaskStatus(taskId, terminalStatus);
-
-        // Log task outcome with relevant details
-        const durSec = (result.durationMs / 1000).toFixed(1);
-        if (terminalStatus === 'success') {
-          log.info(`[task:${taskId}]`, `success (${durSec}s)`);
-        } else {
-          log.error(
-            `[task:${taskId}]`,
-            `${terminalStatus} exit=${result.exitCode} duration=${durSec}s`,
-          );
-          if (result.stderr) {
-            const tail = tailLines(result.stderr, 10);
-            log.error(`[task:${taskId}]`, `stderr tail:\n${tail}`);
-          }
-        }
-
-        // File-only: full stdout/stderr dump (clipped) + extracted metadata
-        log.debug(
-          `[task:${taskId}]`,
-          `stdout: ${result.stdout.length} chars, stderr: ${result.stderr.length} chars`,
-        );
-        if (result.sessionId) {
-          log.debug(`[task:${taskId}]`, `sessionId: ${result.sessionId}`);
-        }
-        if (result.stderrPath) {
-          log.debug(`[task:${taskId}]`, `wrote stderr: ${result.stderrPath}`);
-        }
-        if (result.stdout) {
-          log.quiet(
-            `--- stdout (${taskId}) ---\n${clip(result.stdout)}\n--- end stdout ---`,
-            taskId,
-          );
-        }
-        if (result.stderr) {
-          log.quiet(
-            `--- stderr (${taskId}) ---\n${clip(result.stderr)}\n--- end stderr ---`,
-            taskId,
-          );
-        }
-        if (task.completion) {
-          log.debug(
-            `[task:${taskId}]`,
-            `completion check: type=${task.completion.type} result=${terminalStatus}`,
-          );
-        }
-      } catch (err: unknown) {
-        const errMsg = err instanceof Error ? (err.stack ?? err.message) : String(err);
-        log.error(`[task:${taskId}]`, `failed before execution: ${errMsg}`);
-        state.result = {
-          exitCode: -1,
-          stdout: '',
-          stderr: errMsg,
-          stderrPath: null,
-          durationMs: 0,
-          sessionId: null,
-          normalizedOutput: null,
-          // H2: Engine-level pre-execution errors (driver throw, middleware
-          // throw, getHandler 404) classify as spawn_error — the process never
-          // ran, so calling them "timeout" was actively misleading.
-          failureKind: 'spawn_error',
-        };
-        state.finishedAt = nowISO();
-        setTaskStatus(taskId, 'failed');
-      }
-
-      // 7. Fire hooks
-      const finalStatus: TaskStatus = state.status;
-      try {
-        await fireHook(taskId, finalStatus === 'success' ? 'task_success' : 'task_failure');
-      } catch (hookErr) {
-        log.error(
-          `[task:${taskId}]`,
-          `hook execution failed: ${hookErr instanceof Error ? hookErr.message : String(hookErr)}`,
-        );
-      }
-
-      // 8. Handle stop_all for failure states
-      if (finalStatus !== 'success' && getOnFailure(taskId) === 'stop_all') {
-        applyStopAll(node.track.id);
-      }
-    }
-
-    // ── Event loop ──
-    // Each task is launched as soon as ALL its deps reach a terminal state.
-    // We track in-flight tasks in `running` so a task completing mid-batch
-    // immediately unblocks its dependents without waiting for sibling tasks.
-    const running = new Map<string, Promise<void>>();
-
-    try {
-      while (abortReason === null) {
-        // Launch every task whose deps are all terminal and that isn't already in-flight
-        for (const [id, state] of states) {
-          if (state.status !== 'waiting' || running.has(id)) continue;
-          const node = dag.nodes.get(id)!;
-          const allDepsTerminal =
-            node.dependsOn.length === 0 ||
-            node.dependsOn.every((d) => isTerminal(states.get(d)!.status));
-          if (!allDepsTerminal) continue;
-          const p = processTask(id).finally(() => running.delete(id));
-          running.set(id, p);
-        }
-
-        // All tasks terminal — done
-        if ([...states.values()].every((s) => isTerminal(s.status))) break;
-
-        if (running.size === 0) {
-          // Nothing in-flight but non-terminal tasks exist (e.g. trigger-wait states
-          // that processTask hasn't been called for yet). Poll briefly.
-          await new Promise((r) => setTimeout(r, POLL_INTERVAL_MS));
-        } else {
-          // Wait for any one task to finish, then re-scan for new launchables.
-          await Promise.race(running.values());
-        }
-      }
-
-      if (abortReason !== null) {
-        // Wait for in-flight tasks to honour the abort signal before marking states.
-        if (running.size > 0) await Promise.allSettled(running.values());
-        for (const [id, state] of states) {
-          if (!isTerminal(state.status)) {
-            // By the time allSettled resolves, processTask's try/finally has already
-            // set running tasks to success/failed/timeout. The only non-terminal
-            // statuses remaining here are waiting/idle tasks that were never started.
-            state.finishedAt = nowISO();
-            setTaskStatus(id, 'skipped');
-          }
-        }
-      }
-    } finally {
-      if (pipelineTimer) clearTimeout(pipelineTimer);
-      // Clean up the external abort signal listener to prevent dead references
-      // accumulating on long-lived shared AbortControllers.
-      if (options.signal) {
-        options.signal.removeEventListener('abort', externalAbortHandler);
-      }
-      // Safety net: drain any approvals still pending at shutdown (e.g. crash path).
-      if (approvalGateway.pending().length > 0) {
-        approvalGateway.abortAll('pipeline finished');
-      }
-      // Detach gateway → onEvent bridge so a long-lived gateway (host-supplied)
-      // doesn't keep firing into a dead run.
-      unsubscribeApprovals();
-    }
-
-    // ── Summary ──
-    const summary = { total: 0, success: 0, failed: 0, skipped: 0, timeout: 0, blocked: 0 };
-    for (const [, state] of states) {
-      summary.total++;
-      switch (state.status) {
-        case 'success':
-          summary.success++;
-          break;
-        case 'failed':
-          summary.failed++;
-          break;
-        case 'skipped':
-          summary.skipped++;
-          break;
-        case 'timeout':
-          summary.timeout++;
-          break;
-        case 'blocked':
-          summary.blocked++;
-          break;
-      }
-    }
-
-    const finishedAt = nowISO();
-    const durationMs = new Date(finishedAt).getTime() - new Date(startedAt).getTime();
-
-    if (abortReason !== null) {
-      const reasonText =
-        abortReason === 'timeout'
-          ? 'Pipeline timeout exceeded'
-          : abortReason === 'stop_all'
-            ? 'Pipeline stopped (on_failure: stop_all)'
-            : 'Pipeline aborted by host';
-      await executeHook(
-        config.hooks,
-        'pipeline_error',
-        buildPipelineErrorContext(pipelineInfo, reasonText, undefined, abortReason),
-        workDir,
-      );
-    } else {
-      await executeHook(
-        config.hooks,
-        'pipeline_complete',
-        buildPipelineCompleteContext(
-          { ...pipelineInfo, finished_at: finishedAt, duration_ms: durationMs },
-          summary,
-        ),
-        workDir,
-      );
-    }
-
-    const allSuccess =
-      abortReason === null &&
-      summary.failed === 0 &&
-      summary.timeout === 0 &&
-      summary.blocked === 0;
-
-    log.section('Pipeline summary');
-    log.quiet(
-      `status:   ${abortReason !== null ? `aborted (${abortReason})` : 'completed'}`,
-    );
-    log.quiet(`duration: ${(durationMs / 1000).toFixed(1)}s`);
-    log.quiet(
-      `counts:   total=${summary.total} success=${summary.success} ` +
-        `failed=${summary.failed} skipped=${summary.skipped} ` +
-        `timeout=${summary.timeout} blocked=${summary.blocked}`,
-    );
-    log.quiet('');
-    log.quiet('per-task:');
-    for (const [id, state] of states) {
-      const dur =
-        state.result?.durationMs != null ? `${(state.result.durationMs / 1000).toFixed(1)}s` : '-';
-      const exit = state.result?.exitCode ?? '-';
-      log.quiet(`  ${state.status.padEnd(8)} ${id}  (exit=${exit}, ${dur})`);
-    }
-
-    log.info('[pipeline]', `completed "${config.name}"`);
-    log.info(
-      '[pipeline]',
-      `Total: ${summary.total} | Success: ${summary.success} | Failed: ${summary.failed} | Skipped: ${summary.skipped} | Timeout: ${summary.timeout} | Blocked: ${summary.blocked}`,
-    );
-    log.info('[pipeline]', `Duration: ${(durationMs / 1000).toFixed(1)}s`);
-    log.info('[pipeline]', `Log: ${log.path}`);
-
-    emit({ type: 'run_end', runId, success: allSuccess, abortReason });
-    return { success: allSuccess, runId, logPath: log.path, summary, states: freezeStates(states) };
-  } finally {
-    // Close the persistent log file handle before pruning.
-    log.close();
-    // Prune old per-run log directories on every exit path (normal, blocked, or thrown).
-    // Exclude the current runId so a concurrent run cannot delete its own live directory.
-    if (maxLogRuns > 0) {
-      await pruneLogDirs(resolve(workDir, '.tagma', 'logs'), maxLogRuns, runId);
-    }
-  }
-}
-
-/**
- * Delete the oldest subdirectories under `logsDir`, keeping only the most recent `keep`
- * total runs (including the currently-live run identified by `excludeRunId`).
- * Directories are sorted lexicographically; because runIds are prefixed with a base-36
- * timestamp, lexicographic order equals chronological order.
- *
- * `excludeRunId` is always skipped from deletion even if it would otherwise be pruned —
- * this prevents a concurrent run from removing a live log directory that is still in use.
- *
- * D10: The live run occupies one slot out of `keep`, so the maximum number of
- * *historical* dirs to retain is `keep - 1`. Without this adjustment the function
- * kept `keep` historical dirs plus 1 live dir = `keep + 1` total on disk.
- */
-async function pruneLogDirs(logsDir: string, keep: number, excludeRunId: string): Promise<void> {
-  let entries: string[];
-  try {
-    entries = await readdir(logsDir);
-  } catch {
-    return; // logsDir doesn't exist yet — nothing to prune
-  }
-
-  // Only consider directories that look like run IDs (run_<...>), excluding the live run.
-  const runDirs = entries.filter((e) => e.startsWith('run_') && e !== excludeRunId).sort();
-  // keep - 1 historical slots (1 slot is reserved for the live excludeRunId).
-  const historyKeep = Math.max(0, keep - 1);
-  const toDelete = runDirs.slice(0, Math.max(0, runDirs.length - historyKeep));
-
-  await Promise.all(
-    toDelete.map((dir) =>
-      rm(resolve(logsDir, dir), { recursive: true, force: true }).catch(() => {
-        // Ignore deletion errors — stale dirs are better than a crash
-      }),
-    ),
-  );
-}
-
-function isTerminal(status: TaskStatus): boolean {
-  return (
-    status === 'success' ||
-    status === 'failed' ||
-    status === 'timeout' ||
-    status === 'skipped' ||
-    status === 'blocked'
-  );
-}
-
-/** Return a deep-copied, caller-safe snapshot of the states map. */
-function freezeStates(states: Map<string, TaskState>): ReadonlyMap<string, TaskState> {
-  const copy = new Map<string, TaskState>();
-  for (const [id, s] of states) {
-    copy.set(id, {
-      config: { ...s.config },
-      trackConfig: { ...s.trackConfig },
-      status: s.status,
-      result: s.result ? { ...s.result } : null,
-      startedAt: s.startedAt,
-      finishedAt: s.finishedAt,
-    });
-  }
-  return copy;
-}
+import { resolve } from 'path';
+import { readdir, rm } from 'fs/promises';
+import type {
+  PipelineConfig,
+  TaskConfig,
+  TaskState,
+  TaskStatus,
+  TaskResult,
+  DriverPlugin,
+  TriggerPlugin,
+  CompletionPlugin,
+  MiddlewarePlugin,
+  MiddlewareContext,
+  DriverContext,
+  OnFailure,
+  PromptDocument,
+  Permissions,
+  AbortReason,
+  RunEventPayload,
+  RunTaskState,
+} from './types';
+import { buildDag, type Dag } from './dag';
+import { getHandler, hasHandler, loadPlugins } from './registry';
+import { runSpawn, runCommand } from './runner';
+import { parseDuration, nowISO, generateRunId } from './utils';
+import { promptDocumentFromString, serializePromptDocument } from './prompt-doc';
+import {
+  executeHook,
+  buildPipelineStartContext,
+  buildTaskContext,
+  buildPipelineCompleteContext,
+  buildPipelineErrorContext,
+  type PipelineInfo,
+  type TrackInfo,
+  type TaskInfo,
+} from './hooks';
+import { Logger, tailLines, clip } from './logger';
+import { InMemoryApprovalGateway, type ApprovalGateway } from './approval';
+
+// ═══ A7: Typed trigger errors ═══
+// Replace string-matching on error messages with structured error types so
+// coincidental substrings don't cause misclassification.
+
+export class TriggerBlockedError extends Error {
+  readonly code = 'TRIGGER_BLOCKED' as const;
+  constructor(message: string) {
+    super(message);
+    this.name = 'TriggerBlockedError';
+  }
+}
+
+export class TriggerTimeoutError extends Error {
+  readonly code = 'TRIGGER_TIMEOUT' as const;
+  constructor(message: string) {
+    super(message);
+    this.name = 'TriggerTimeoutError';
+  }
+}
+
+// ═══ Preflight Validation ═══
+
+function preflight(config: PipelineConfig, dag: Dag): void {
+  const errors: string[] = [];
+
+  for (const [, node] of dag.nodes) {
+    const task = node.task;
+    const track = node.track;
+    const driverName = task.driver ?? track.driver ?? config.driver ?? 'opencode';
+
+    // Pure command tasks don't use a driver — skip driver registration check.
+    const isCommandOnly = task.command && !task.prompt;
+
+    if (!isCommandOnly && !hasHandler('drivers', driverName)) {
+      errors.push(`Task "${node.taskId}": driver "${driverName}" not registered`);
+    }
+
+    if (task.trigger && !hasHandler('triggers', task.trigger.type)) {
+      errors.push(`Task "${node.taskId}": trigger type "${task.trigger.type}" not registered`);
+    }
+
+    if (task.completion && !hasHandler('completions', task.completion.type)) {
+      errors.push(
+        `Task "${node.taskId}": completion type "${task.completion.type}" not registered`,
+      );
+    }
+
+    const mws = task.middlewares ?? track.middlewares ?? [];
+    for (const mw of mws) {
+      if (!hasHandler('middlewares', mw.type)) {
+        errors.push(`Task "${node.taskId}": middleware type "${mw.type}" not registered`);
+      }
+    }
+
+    if (task.continue_from && hasHandler('drivers', driverName)) {
+      const driver = getHandler<DriverPlugin>('drivers', driverName);
+      if (!driver.capabilities.sessionResume) {
+        // buildDag has already qualified `continue_from` and stored the result
+        // on the node; preflight runs after buildDag, so the upstream id is
+        // always available here without re-resolving.
+        const upstreamId = node.resolvedContinueFrom;
+        if (upstreamId) {
+          const upstream = dag.nodes.get(upstreamId);
+          if (upstream) {
+            // A handoff is possible via session resume (already ruled out above),
+            // OR in-memory text injection through normalizedMap
+            // (when the upstream driver implements parseResult and returns normalizedOutput).
+            const upstreamDriverName =
+              upstream.task.driver ?? upstream.track.driver ?? config.driver ?? 'opencode';
+            const upstreamDriver = hasHandler('drivers', upstreamDriverName)
+              ? getHandler<DriverPlugin>('drivers', upstreamDriverName)
+              : null;
+            const canNormalize = typeof upstreamDriver?.parseResult === 'function';
+
+            if (!canNormalize) {
+              errors.push(
+                `Task "${node.taskId}" uses continue_from: "${task.continue_from}", ` +
+                  `but upstream task "${upstreamId}" its driver ` +
+                  `does not implement parseResult for text-injection handoff. ` +
+                  `Use a driver with parseResult, or remove continue_from.`,
+              );
+            }
+          }
+        }
+      }
+    }
+  }
+
+  if (errors.length > 0) {
+    throw new Error(`Preflight validation failed:\n  - ${errors.join('\n  - ')}`);
+  }
+}
+
+// ═══ Engine ═══
+
+export interface EngineResult {
+  readonly success: boolean;
+  readonly runId: string;
+  readonly logPath: string;
+  readonly summary: {
+    total: number;
+    success: number;
+    failed: number;
+    skipped: number;
+    timeout: number;
+    blocked: number;
+  };
+  readonly states: ReadonlyMap<string, TaskState>;
+}
+
+// ═══ Pipeline Events ═══
+//
+// The engine emits RunEventPayload values (defined in @tagma/types) via
+// `onEvent`. Every payload carries `runId`; the editor server stamps a
+// per-run `seq` before broadcasting. There is one event vocabulary
+// end-to-end — no server-side translation layer.
+
+// Re-export so SDK consumers can import the event type without reaching
+// into @tagma/types directly.
+export type { RunEventPayload } from './types';
+
+// ═══ Helpers ═══
+
+/**
+ * Project the engine's internal TaskState onto the wire RunTaskState
+ * shape. `logs` / `totalLogCount` default to empty — they are populated
+ * on the server side from streamed `task_log` events, not from state.
+ */
+function toRunTaskState(
+  taskId: string,
+  trackId: string,
+  taskName: string,
+  state: TaskState,
+): RunTaskState {
+  const result = state.result;
+  const cfg = state.config;
+  return {
+    taskId,
+    trackId,
+    taskName,
+    status: state.status,
+    startedAt: state.startedAt,
+    finishedAt: state.finishedAt,
+    durationMs: result?.durationMs ?? null,
+    exitCode: result?.exitCode ?? null,
+    stdout: result?.stdout ?? '',
+    stderr: result?.stderr ?? '',
+    stderrPath: result?.stderrPath ?? null,
+    sessionId: result?.sessionId ?? null,
+    normalizedOutput: result?.normalizedOutput ?? null,
+    resolvedDriver: cfg.driver ?? null,
+    resolvedModel: cfg.model ?? null,
+    resolvedPermissions: (cfg.permissions as Permissions | undefined) ?? null,
+    logs: [],
+    totalLogCount: 0,
+  };
+}
+
+export interface RunPipelineOptions {
+  readonly approvalGateway?: ApprovalGateway;
+  /**
+   * Maximum number of per-run log directories to retain under `<workDir>/.tagma/logs/`.
+   * Oldest directories are deleted after each run. Defaults to 20. Set to 0 to disable cleanup.
+   */
+  readonly maxLogRuns?: number;
+  /**
+   * Caller-supplied run ID. When provided the engine uses this instead of
+   * generating its own via `generateRunId()`, keeping the editor and SDK
+   * log directories aligned on the same ID.
+   */
+  readonly runId?: string;
+  /**
+   * External AbortSignal — aborting it cancels the pipeline immediately.
+   * Equivalent to the pipeline timeout firing, but caller-controlled.
+   */
+  readonly signal?: AbortSignal;
+  /**
+   * Called on every pipeline/task status transition.
+   * Use for real-time UI updates (e.g. updating a visual workflow graph).
+   */
+  readonly onEvent?: (event: RunEventPayload) => void;
+  /**
+   * Skip the engine's built-in `loadPlugins(config.plugins)` call.
+   * Use this when the host has already pre-loaded plugins from a custom
+   * resolution path (e.g. a user workspace's node_modules) so the engine
+   * doesn't re-resolve them via Node's default cwd-based import.
+   */
+  readonly skipPluginLoading?: boolean;
+}
+
+// Poll interval when no tasks are in-flight but non-terminal tasks remain
+// (e.g. tasks waiting on a file or manual trigger).
+const POLL_INTERVAL_MS = 50;
+
+// R15: cap on each normalized-output entry stored in normalizedMap so a
+// runaway parseResult can't accumulate hundreds of MB across tasks. 1 MB
+// is generous for any text-context handoff between AI tasks.
+const MAX_NORMALIZED_BYTES = 1_000_000;
+
+export async function runPipeline(
+  config: PipelineConfig,
+  workDir: string,
+  options: RunPipelineOptions = {},
+): Promise<EngineResult> {
+  const approvalGateway = options.approvalGateway ?? new InMemoryApprovalGateway();
+  const maxLogRuns = options.maxLogRuns ?? 20;
+
+  // Load any plugins declared in the pipeline config before preflight so that
+  // drivers, completions, and middlewares referenced in YAML are registered.
+  // Hosts that pre-load plugins from a custom path (e.g. the editor loading
+  // from the user's workspace node_modules) pass skipPluginLoading: true so
+  // we don't re-resolve via Node's cwd-based default import.
+  if (!options.skipPluginLoading && config.plugins?.length) {
+    await loadPlugins(config.plugins);
+  }
+
+  const dag = buildDag(config);
+  const runId = options.runId ?? generateRunId();
+  preflight(config, dag);
+
+  const startedAt = nowISO();
+  const pipelineInfo: PipelineInfo = { name: config.name, run_id: runId, started_at: startedAt };
+  // Forward every structured log line to subscribers as task_log events.
+  // Reading options.onEvent inside the callback (vs. capturing it once) keeps
+  // the SDK behavior correct if callers pass a fresh onEvent on each run.
+  const log = new Logger(workDir, runId, (record) => {
+    options.onEvent?.({
+      type: 'task_log',
+      runId,
+      taskId: record.taskId,
+      level: record.level,
+      timestamp: record.timestamp,
+      text: record.text,
+    });
+  });
+
+  try {
+    log.info('[pipeline]', `start "${config.name}" run_id=${runId}`);
+
+    // File-only: dump the resolved pipeline shape + DAG topology for post-mortem.
+    log.section('Pipeline configuration');
+    log.quiet(`name:          ${config.name}`);
+    log.quiet(`driver:        ${config.driver ?? '(default: opencode)'}`);
+    log.quiet(`timeout:       ${config.timeout ?? '(none)'}`);
+    log.quiet(`tracks:        ${config.tracks.length}`);
+    log.quiet(`tasks (total): ${dag.nodes.size}`);
+    log.quiet(`plugins:       ${(config.plugins ?? []).join(', ') || '(none)'}`);
+    log.quiet(
+      `hooks:         ${config.hooks ? Object.keys(config.hooks).join(', ') || '(none)' : '(none)'}`,
+    );
+
+    log.section('DAG topology');
+    for (const [id, node] of dag.nodes) {
+      const deps = node.dependsOn.length ? node.dependsOn.join(', ') : '(root)';
+      const kind = node.task.prompt ? 'ai' : 'cmd';
+      log.quiet(`  • ${id}  [${kind}]  track=${node.track.id}  deps=[${deps}]`);
+    }
+    log.quiet('');
+
+    // Initialize states (before hook, so we can return them even if blocked)
+    const states = new Map<string, TaskState>();
+    for (const [id, node] of dag.nodes) {
+      states.set(id, {
+        config: node.task,
+        trackConfig: node.track,
+        status: 'idle',
+        result: null,
+        startedAt: null,
+        finishedAt: null,
+      });
+    }
+
+    // Pipeline start hook (gate). Runs BEFORE the engine emits run_start so
+    // a blocked pipeline produces zero wire events (the server treats the
+    // thrown error as run_error). Hosts get a rich error message; nothing
+    // is ever half-broadcast.
+    const startHook = await executeHook(
+      config.hooks,
+      'pipeline_start',
+      buildPipelineStartContext(pipelineInfo),
+      workDir,
+    );
+    if (!startHook.allowed) {
+      console.error(`Pipeline blocked by pipeline_start hook (exit code ${startHook.exitCode})`);
+      await executeHook(
+        config.hooks,
+        'pipeline_error',
+        buildPipelineErrorContext(pipelineInfo, 'pipeline_blocked', 'pipeline_blocked'),
+        workDir,
+      );
+      return {
+        success: false,
+        runId,
+        logPath: log.path,
+        summary: {
+          total: dag.nodes.size,
+          success: 0,
+          failed: 0,
+          skipped: 0,
+          timeout: 0,
+          blocked: 0,
+        },
+        states: freezeStates(states),
+      };
+    }
+
+    // Pipeline approved — transition all tasks to waiting.
+    for (const [, state] of states) {
+      state.status = 'waiting';
+    }
+    // Emit run_start with a wire-shape snapshot so SSE subscribers can
+    // initialize their task maps on the same event stream that carries
+    // updates. No separate "server pre-broadcasts run_start" ceremony —
+    // the engine owns the lifecycle boundary.
+    const runStartTasks: RunTaskState[] = [];
+    for (const [id, node] of dag.nodes) {
+      const s = states.get(id)!;
+      runStartTasks.push(toRunTaskState(id, node.track.id, node.task.name ?? id, s));
+    }
+    emit({ type: 'run_start', runId, tasks: runStartTasks });
+
+    const sessionMap = new Map<string, string>();
+    const normalizedMap = new Map<string, string>();
+
+    // Pipeline timeout + abort reason tracking.
+    //
+    // `abortReason` replaces the previous `pipelineAborted: boolean`: it
+    // carries the concrete cause (timeout / stop_all / external) through
+    // to run_end and the pipeline_error hook so downstream consumers can
+    // distinguish them without scraping message strings.
+    const pipelineTimeoutMs = config.timeout ? parseDuration(config.timeout) : 0;
+    let abortReason: AbortReason | null = null;
+    const abortController = new AbortController();
+    let pipelineTimer: ReturnType<typeof setTimeout> | null = null;
+
+    if (pipelineTimeoutMs > 0) {
+      pipelineTimer = setTimeout(() => {
+        if (abortReason === null) abortReason = 'timeout';
+        abortController.abort();
+      }, pipelineTimeoutMs);
+    }
+
+    // When the pipeline is aborted (timeout, stop_all, external), drain
+    // all pending approvals so waiting triggers unblock immediately.
+    abortController.signal.addEventListener('abort', () => {
+      approvalGateway.abortAll('pipeline aborted');
+    });
+
+    // Wire external cancel signal into the internal abort controller.
+    const externalAbortHandler = () => {
+      if (abortReason === null) abortReason = 'external';
+      abortController.abort();
+    };
+    if (options.signal) {
+      if (options.signal.aborted) {
+        externalAbortHandler();
+      } else {
+        options.signal.addEventListener('abort', externalAbortHandler, { once: true });
+      }
+    }
+
+    // Bridge approval gateway events onto the wire stream so hosts (editor
+    // server, CLI adapters) see approvals on the same channel as task
+    // updates. The server no longer needs its own gateway subscription.
+    const unsubscribeApprovals = approvalGateway.subscribe((ev) => {
+      if (ev.type === 'requested') {
+        emit({
+          type: 'approval_request',
+          runId,
+          request: {
+            id: ev.request.id,
+            taskId: ev.request.taskId,
+            trackId: ev.request.trackId,
+            message: ev.request.message,
+            createdAt: ev.request.createdAt,
+            timeoutMs: ev.request.timeoutMs,
+            metadata: ev.request.metadata,
+          },
+        });
+        return;
+      }
+      if (ev.type === 'resolved' || ev.type === 'expired' || ev.type === 'aborted') {
+        const outcome =
+          ev.type === 'resolved'
+            ? ev.decision.outcome
+            : ev.type === 'expired'
+              ? 'timeout'
+              : 'aborted';
+        emit({
+          type: 'approval_resolved',
+          runId,
+          requestId: ev.request.id,
+          outcome,
+        });
+      }
+    });
+
+    // ── Helpers ──
+
+    function emit(event: RunEventPayload): void {
+      options.onEvent?.(event);
+    }
+
+    function setTaskStatus(taskId: string, newStatus: TaskStatus): void {
+      const state = states.get(taskId)!;
+      // Terminal lock: once a task reaches a terminal state it must not be
+      // re-transitioned. This prevents stop_all from marking running tasks as
+      // skipped and then having their in-flight processTask promise overwrite
+      // that with success/failed, producing an invalid double transition.
+      if (isTerminal(state.status)) return;
+      state.status = newStatus;
+      const result = state.result;
+      const cfg = state.config;
+      emit({
+        type: 'task_update',
+        runId,
+        taskId,
+        status: newStatus,
+        startedAt: state.startedAt ?? undefined,
+        finishedAt: state.finishedAt ?? undefined,
+        durationMs: result?.durationMs,
+        exitCode: result?.exitCode,
+        stdout: result?.stdout,
+        stderr: result?.stderr,
+        stderrPath: result?.stderrPath ?? null,
+        sessionId: result?.sessionId ?? null,
+        normalizedOutput: result?.normalizedOutput ?? null,
+        resolvedDriver: cfg.driver ?? null,
+        resolvedModel: cfg.model ?? null,
+        resolvedPermissions: (cfg.permissions as Permissions | undefined) ?? null,
+      });
+    }
+
+    function getOnFailure(taskId: string): OnFailure {
+      return dag.nodes.get(taskId)?.track.on_failure ?? 'skip_downstream';
+    }
+
+    function isDependencySatisfied(depId: string): 'satisfied' | 'unsatisfied' | 'skip' {
+      const depState = states.get(depId);
+      if (!depState) return 'skip';
+      switch (depState.status) {
+        case 'success':
+          return 'satisfied';
+        case 'skipped':
+          return 'skip';
+        case 'failed':
+        case 'timeout':
+        case 'blocked':
+          return getOnFailure(depId) === 'ignore' ? 'satisfied' : 'skip';
+        default:
+          return 'unsatisfied';
+      }
+    }
+
+    /**
+     * H3: "stop_all" historically only stopped tasks within the same track,
+     * which contradicted both its name and user expectations. It now stops
+     * the **entire pipeline**:
+     *   - In-flight tasks are signalled via the shared abort controller so
+     *     drivers / runner.ts can cancel cooperatively (returning
+     *     `failureKind: 'timeout'`).
+     *   - Still-waiting tasks across every track are immediately marked
+     *     skipped so the run completes promptly.
+     * The terminal lock in setTaskStatus prevents any later re-transition
+     * should a completed running task try to overwrite the skipped state.
+     */
+    function applyStopAll(_failedTrackId: string): void {
+      if (abortReason === null) abortReason = 'stop_all';
+      abortController.abort();
+      for (const [id, state] of states) {
+        if (state.status === 'waiting') {
+          state.finishedAt = nowISO();
+          setTaskStatus(id, 'skipped');
+        }
+      }
+    }
+
+    function buildTaskInfoObj(taskId: string): TaskInfo {
+      const state = states.get(taskId)!;
+      return {
+        id: taskId,
+        name: state.config.name,
+        type: state.config.prompt ? 'ai' : 'command',
+        status: state.status,
+        exit_code: state.result?.exitCode ?? null,
+        duration_ms: state.result?.durationMs ?? null,
+        stderr_path: state.result?.stderrPath ?? null,
+        session_id: state.result?.sessionId ?? null,
+        started_at: state.startedAt,
+        finished_at: state.finishedAt,
+      };
+    }
+
+    function trackInfoOf(taskId: string): TrackInfo {
+      const node = dag.nodes.get(taskId)!;
+      return { id: node.track.id, name: node.track.name };
+    }
+
+    async function fireHook(taskId: string, event: 'task_success' | 'task_failure'): Promise<void> {
+      await executeHook(
+        config.hooks,
+        event,
+        buildTaskContext(event, pipelineInfo, trackInfoOf(taskId), buildTaskInfoObj(taskId)),
+        workDir,
+        abortController.signal,
+      );
+    }
+
+    // ── Process a single task ──
+
+    async function processTask(taskId: string): Promise<void> {
+      const state = states.get(taskId)!;
+      const node = dag.nodes.get(taskId)!;
+      const task = node.task;
+      const track = node.track;
+
+      log.section(`Task ${taskId}`, taskId);
+      log.debug(
+        `[task:${taskId}]`,
+        `type=${task.prompt ? 'ai' : 'cmd'} track=${track.id} deps=[${node.dependsOn.join(', ') || '(root)'}]`,
+      );
+
+      // 1. Check dependencies
+      for (const depId of node.dependsOn) {
+        const result = isDependencySatisfied(depId);
+        if (result === 'skip') {
+          const depStatus = states.get(depId)?.status ?? 'unknown';
+          log.debug(`[task:${taskId}]`, `skipped (upstream "${depId}" status=${depStatus})`);
+          state.finishedAt = nowISO();
+          setTaskStatus(taskId, 'skipped');
+          return;
+        }
+        if (result === 'unsatisfied') return; // still waiting
+      }
+
+      // 2. Check trigger
+      if (task.trigger) {
+        log.debug(
+          `[task:${taskId}]`,
+          `trigger wait: type=${task.trigger.type} ${JSON.stringify(task.trigger)}`,
+        );
+        try {
+          const triggerPlugin = getHandler<TriggerPlugin>('triggers', task.trigger.type);
+          // R6: race the plugin's watch() against the pipeline's abort signal.
+          // Third-party triggers may forget to wire up ctx.signal — without
+          // this race, an aborted pipeline would hang forever waiting for the
+          // plugin's watch promise to resolve. The race resolves on whichever
+          // path settles first, and the cleanup paths in finally never run on
+          // the orphaned plugin promise (it's allowed to leak a watcher; the
+          // pipeline is being torn down anyway).
+          await new Promise<unknown>((resolve, reject) => {
+            let settled = false;
+            const onAbort = () => {
+              if (settled) return;
+              settled = true;
+              abortController.signal.removeEventListener('abort', onAbort);
+              reject(new Error('Pipeline aborted'));
+            };
+            if (abortController.signal.aborted) {
+              onAbort();
+              return;
+            }
+            abortController.signal.addEventListener('abort', onAbort, { once: true });
+            triggerPlugin
+              .watch(task.trigger as Record<string, unknown>, {
+                taskId: node.taskId,
+                trackId: track.id,
+                workDir: task.cwd ?? workDir,
+                signal: abortController.signal,
+                approvalGateway,
+              })
+              .then(
+                (v) => {
+                  if (settled) return;
+                  settled = true;
+                  abortController.signal.removeEventListener('abort', onAbort);
+                  resolve(v);
+                },
+                (e) => {
+                  if (settled) return;
+                  settled = true;
+                  abortController.signal.removeEventListener('abort', onAbort);
+                  reject(e);
+                },
+              );
+          });
+          log.debug(`[task:${taskId}]`, `trigger fired`);
+        } catch (err: unknown) {
+          // If pipeline was aborted while we were still waiting for the trigger,
+          // this task never entered running state → skipped, not timeout.
+          state.finishedAt = nowISO();
+          if (abortReason !== null) {
+            setTaskStatus(taskId, 'skipped');
+          } else if (err instanceof TriggerBlockedError) {
+            setTaskStatus(taskId, 'blocked'); // user/policy rejection
+          } else if (err instanceof TriggerTimeoutError) {
+            setTaskStatus(taskId, 'timeout'); // genuine trigger wait timeout
+          } else {
+            // A7 fallback: also check message strings for backward-compat with
+            // third-party trigger plugins that don't throw typed errors yet.
+            const msg = err instanceof Error ? err.message : String(err);
+            if (msg.includes('rejected') || msg.includes('denied')) {
+              setTaskStatus(taskId, 'blocked');
+            } else if (msg.includes('timeout')) {
+              setTaskStatus(taskId, 'timeout');
+            } else {
+              setTaskStatus(taskId, 'failed'); // plugin error, watcher crash, etc.
+            }
+          }
+          try {
+            await fireHook(taskId, 'task_failure');
+          } catch (hookErr) {
+            log.error(
+              `[task:${taskId}]`,
+              `hook execution failed: ${hookErr instanceof Error ? hookErr.message : String(hookErr)}`,
+            );
+          }
+          return;
+        }
+      }
+
+      // 3. task_start hook (gate)
+      const hookResult = await executeHook(
+        config.hooks,
+        'task_start',
+        buildTaskContext('task_start', pipelineInfo, trackInfoOf(taskId), buildTaskInfoObj(taskId)),
+        workDir,
+        abortController.signal,
+      );
+      if (hookResult.exitCode !== 0 || config.hooks?.task_start) {
+        log.debug(
+          `[task:${taskId}]`,
+          `task_start hook exit=${hookResult.exitCode} allowed=${hookResult.allowed}`,
+        );
+      }
+      if (!hookResult.allowed) {
+        state.finishedAt = nowISO();
+        setTaskStatus(taskId, 'blocked');
+        try {
+          await fireHook(taskId, 'task_failure');
+        } catch (hookErr) {
+          log.error(
+            `[task:${taskId}]`,
+            `hook execution failed: ${hookErr instanceof Error ? hookErr.message : String(hookErr)}`,
+          );
+        }
+        return;
+      }
+
+      // 4. Mark running — set startedAt before emitting so subscribers see a
+      // complete task_update (startedAt non-null) on the status transition.
+      state.startedAt = nowISO();
+      setTaskStatus(taskId, 'running');
+      log.info(
+        `[task:${taskId}]`,
+        task.command ? `running: ${task.command}` : `running (driver task)`,
+      );
+
+      // File-only: resolved config for this task
+      const resolvedDriver = task.driver ?? track.driver ?? config.driver ?? 'opencode';
+      const resolvedModel = task.model ?? track.model ?? config.model ?? '(default)';
+      const resolvedPerms = task.permissions ?? track.permissions ?? '(default)';
+      const resolvedCwd = task.cwd ?? track.cwd ?? workDir;
+      log.debug(
+        `[task:${taskId}]`,
+        `resolved: driver=${resolvedDriver} model=${resolvedModel} cwd=${resolvedCwd}`,
+      );
+      log.debug(`[task:${taskId}]`, `permissions: ${JSON.stringify(resolvedPerms)}`);
+      if (task.continue_from) {
+        log.debug(`[task:${taskId}]`, `continue_from: "${task.continue_from}"`);
+      }
+      if (task.timeout) {
+        log.debug(`[task:${taskId}]`, `timeout: ${task.timeout}`);
+      }
+
+      try {
+        let result: TaskResult;
+        const timeoutMs = task.timeout ? parseDuration(task.timeout) : undefined;
+
+        const runOpts = { timeoutMs, signal: abortController.signal };
+
+        if (task.command) {
+          log.debug(`[task:${taskId}]`, `command: ${task.command}`);
+          result = await runCommand(task.command, task.cwd ?? workDir, runOpts);
+        } else {
+          // AI task: apply middleware chain against a structured PromptDocument.
+          const driverName = task.driver ?? track.driver ?? config.driver ?? 'opencode';
+          const driver = getHandler<DriverPlugin>('drivers', driverName);
+
+          const originalLen = task.prompt!.length;
+          let doc: PromptDocument = promptDocumentFromString(task.prompt!);
+          const mws = task.middlewares !== undefined ? task.middlewares : track.middlewares;
+          if (mws && mws.length > 0) {
+            log.debug(
+              `[task:${taskId}]`,
+              `middleware chain: ${mws.map((m) => m.type).join(' → ')}`,
+            );
+            const mwCtx: MiddlewareContext = {
+              task,
+              track,
+              workDir: task.cwd ?? workDir,
+            };
+            for (const mwConfig of mws) {
+              const mwPlugin = getHandler<MiddlewarePlugin>('middlewares', mwConfig.type);
+              const beforeBlocks = doc.contexts.length;
+              const beforeLen = serializePromptDocument(doc).length;
+
+              // Prefer the structured API. Fall back to the legacy
+              // `enhance(string) → string` path so v0.x plugins keep
+              // working — that fallback loses context structure (the
+              // middleware's output becomes the new task body) but never
+              // silently drops content.
+              if (typeof mwPlugin.enhanceDoc === 'function') {
+                const next = await mwPlugin.enhanceDoc(
+                  doc,
+                  mwConfig as Record<string, unknown>,
+                  mwCtx,
+                );
+                if (
+                  !next ||
+                  typeof next !== 'object' ||
+                  !Array.isArray((next as PromptDocument).contexts) ||
+                  typeof (next as PromptDocument).task !== 'string'
+                ) {
+                  throw new Error(
+                    `middleware "${mwConfig.type}".enhanceDoc() returned a malformed PromptDocument`,
+                  );
+                }
+                doc = next as PromptDocument;
+              } else if (typeof mwPlugin.enhance === 'function') {
+                const asString = serializePromptDocument(doc);
+                const next = await mwPlugin.enhance(
+                  asString,
+                  mwConfig as Record<string, unknown>,
+                  mwCtx,
+                );
+                // R3: a middleware that returns undefined / null / a non-string
+                // would silently corrupt the prompt. Fail loud.
+                if (typeof next !== 'string') {
+                  throw new Error(
+                    `middleware "${mwConfig.type}".enhance() returned ${next === null ? 'null' : typeof next}, expected string`,
+                  );
+                }
+                // Legacy fallback: collapse the returned string into a
+                // fresh doc. Earlier structure is folded into the string
+                // (serializePromptDocument just ran), so bytes the driver
+                // sees match the old string pipeline.
+                doc = { contexts: [], task: next };
+              } else {
+                throw new Error(
+                  `middleware "${mwConfig.type}" provides neither enhanceDoc nor enhance`,
+                );
+              }
+              const afterLen = serializePromptDocument(doc).length;
+              const addedBlocks = doc.contexts.length - beforeBlocks;
+              log.debug(
+                `[task:${taskId}]`,
+                `  ${mwConfig.type}: ${beforeLen} → ${afterLen} chars` +
+                  (addedBlocks > 0
+                    ? ` (+${addedBlocks} context block${addedBlocks > 1 ? 's' : ''})`
+                    : ''),
+              );
+            }
+          }
+          const prompt = serializePromptDocument(doc);
+          log.debug(
+            `[task:${taskId}]`,
+            `prompt: ${originalLen} chars (final: ${prompt.length} chars, ${doc.contexts.length} block${doc.contexts.length === 1 ? '' : 's'})`,
+          );
+          log.quiet(`--- prompt (final) ---\n${clip(prompt)}\n--- end prompt ---`, taskId);
+
+          // H1: hand the driver a continue_from that has already been
+          // qualified by dag.ts. Without this, drivers like codex/opencode/
+          // claude-code look up maps directly with
+          // the user's raw (possibly bare) string, which races whenever two
+          // tracks share a task name. dag.ts has the only authoritative
+          // resolver, so we use its precomputed answer here.
+          // Drivers key sessionMap/normalizedMap by fully-qualified id. buildDag
+          // guarantees `resolvedContinueFrom` is set for every task that has a
+          // `continue_from`, so if we see the bare form here something upstream
+          // is broken — fail loud instead of silently miskeying the lookup.
+          if (task.continue_from && !node.resolvedContinueFrom) {
+            throw new Error(
+              `Internal: task "${taskId}" has continue_from "${task.continue_from}" ` +
+                `but no resolvedContinueFrom. buildDag should have qualified it.`,
+            );
+          }
+          const enrichedTask: TaskConfig = {
+            ...task,
+            prompt,
+            continue_from: node.resolvedContinueFrom,
+          };
+          const driverCtx: DriverContext = {
+            sessionMap,
+            normalizedMap,
+            workDir: task.cwd ?? workDir,
+            // Structured view for drivers that want fine-grained control
+            // over serialization (e.g. inserting [Previous Output] between
+            // contexts and task). Drivers that read task.prompt see the
+            // default serialization and need no changes.
+            promptDoc: doc,
+          };
+          const spec = await driver.buildCommand(enrichedTask, track, driverCtx);
+          log.debug(`[task:${taskId}]`, `driver=${driverName}`);
+          log.debug(`[task:${taskId}]`, `spawn args: ${JSON.stringify(spec.args)}`);
+          if (spec.cwd) log.debug(`[task:${taskId}]`, `spawn cwd: ${spec.cwd}`);
+          if (spec.env)
+            log.debug(
+              `[task:${taskId}]`,
+              `spawn env overrides: ${Object.keys(spec.env).join(', ')}`,
+            );
+          if (spec.stdin) log.debug(`[task:${taskId}]`, `spawn stdin: ${spec.stdin.length} chars`);
+          result = await runSpawn(spec, driver, runOpts);
+        }
+
+        // 6. Determine terminal status (without emitting yet — result must be complete first)
+        // H2: branch on failureKind so spawn errors no longer masquerade as
+        // timeouts. Old runners that don't set failureKind still work — we
+        // fall back to the historical `exitCode === -1 → timeout` heuristic so
+        // pre-existing third-party drivers don't regress.
+        let terminalStatus: TaskStatus;
+        const kind = result.failureKind;
+        if (kind === 'timeout') {
+          terminalStatus = 'timeout';
+        } else if (kind === 'spawn_error') {
+          terminalStatus = 'failed';
+        } else if (kind === undefined && result.exitCode === -1) {
+          // Legacy path: pre-H2 driver returned -1 with no kind. Treat as
+          // timeout for backward compatibility (the previous behaviour).
+          terminalStatus = 'timeout';
+        } else if (result.exitCode !== 0) {
+          terminalStatus = 'failed';
+        } else if (task.completion) {
+          const plugin = getHandler<CompletionPlugin>('completions', task.completion.type);
+          const completionCtx = { workDir: task.cwd ?? workDir, signal: abortController.signal };
+          const passed = await plugin.check(
+            task.completion as Record<string, unknown>,
+            result,
+            completionCtx,
+          );
+          // R4: strict boolean check. Truthy strings/numbers used to be coerced
+          // to success — a check returning "ok" would let a failing task pass.
+          if (typeof passed !== 'boolean') {
+            throw new Error(
+              `completion "${task.completion.type}".check() returned ${passed === null ? 'null' : typeof passed}, expected boolean`,
+            );
+          }
+          terminalStatus = passed ? 'success' : 'failed';
+        } else {
+          terminalStatus = 'success';
+        }
+
+        // Store normalized text separately (in-memory) for continue_from handoff.
+        // R15: clip oversized values so a runaway parseResult can't accumulate
+        // hundreds of MB across tasks.
+        if (result.normalizedOutput !== null) {
+          const clipped =
+            result.normalizedOutput.length > MAX_NORMALIZED_BYTES
+              ? result.normalizedOutput.slice(0, MAX_NORMALIZED_BYTES) +
+                `\n[…clipped at ${MAX_NORMALIZED_BYTES} bytes]`
+              : result.normalizedOutput;
+          normalizedMap.set(taskId, clipped);
+        }
+
+        if (result.stderr) {
+          const stderrPath = resolve(log.dir, `${taskId.replace(/\./g, '_')}.stderr`);
+          await Bun.write(stderrPath, result.stderr);
+          result = { ...result, stderrPath };
+        }
+
+        if (result.sessionId) {
+          // H1: qualified-only key.
+          sessionMap.set(taskId, result.sessionId);
+        }
+
+        // Set result and finishedAt before emitting terminal status so listeners see complete state
+        state.result = result;
+        state.finishedAt = nowISO();
+        setTaskStatus(taskId, terminalStatus);
+
+        // Log task outcome with relevant details
+        const durSec = (result.durationMs / 1000).toFixed(1);
+        if (terminalStatus === 'success') {
+          log.info(`[task:${taskId}]`, `success (${durSec}s)`);
+        } else {
+          log.error(
+            `[task:${taskId}]`,
+            `${terminalStatus} exit=${result.exitCode} duration=${durSec}s`,
+          );
+          if (result.stderr) {
+            const tail = tailLines(result.stderr, 10);
+            log.error(`[task:${taskId}]`, `stderr tail:\n${tail}`);
+          }
+        }
+
+        // File-only: full stdout/stderr dump (clipped) + extracted metadata
+        log.debug(
+          `[task:${taskId}]`,
+          `stdout: ${result.stdout.length} chars, stderr: ${result.stderr.length} chars`,
+        );
+        if (result.sessionId) {
+          log.debug(`[task:${taskId}]`, `sessionId: ${result.sessionId}`);
+        }
+        if (result.stderrPath) {
+          log.debug(`[task:${taskId}]`, `wrote stderr: ${result.stderrPath}`);
+        }
+        if (result.stdout) {
+          log.quiet(
+            `--- stdout (${taskId}) ---\n${clip(result.stdout)}\n--- end stdout ---`,
+            taskId,
+          );
+        }
+        if (result.stderr) {
+          log.quiet(
+            `--- stderr (${taskId}) ---\n${clip(result.stderr)}\n--- end stderr ---`,
+            taskId,
+          );
+        }
+        if (task.completion) {
+          log.debug(
+            `[task:${taskId}]`,
+            `completion check: type=${task.completion.type} result=${terminalStatus}`,
+          );
+        }
+      } catch (err: unknown) {
+        const errMsg = err instanceof Error ? (err.stack ?? err.message) : String(err);
+        log.error(`[task:${taskId}]`, `failed before execution: ${errMsg}`);
+        state.result = {
+          exitCode: -1,
+          stdout: '',
+          stderr: errMsg,
+          stderrPath: null,
+          durationMs: 0,
+          sessionId: null,
+          normalizedOutput: null,
+          // H2: Engine-level pre-execution errors (driver throw, middleware
+          // throw, getHandler 404) classify as spawn_error — the process never
+          // ran, so calling them "timeout" was actively misleading.
+          failureKind: 'spawn_error',
+        };
+        state.finishedAt = nowISO();
+        setTaskStatus(taskId, 'failed');
+      }
+
+      // 7. Fire hooks
+      const finalStatus: TaskStatus = state.status;
+      try {
+        await fireHook(taskId, finalStatus === 'success' ? 'task_success' : 'task_failure');
+      } catch (hookErr) {
+        log.error(
+          `[task:${taskId}]`,
+          `hook execution failed: ${hookErr instanceof Error ? hookErr.message : String(hookErr)}`,
+        );
+      }
+
+      // 8. Handle stop_all for failure states
+      if (finalStatus !== 'success' && getOnFailure(taskId) === 'stop_all') {
+        applyStopAll(node.track.id);
+      }
+    }
+
+    // ── Event loop ──
+    // Each task is launched as soon as ALL its deps reach a terminal state.
+    // We track in-flight tasks in `running` so a task completing mid-batch
+    // immediately unblocks its dependents without waiting for sibling tasks.
+    const running = new Map<string, Promise<void>>();
+
+    try {
+      while (abortReason === null) {
+        // Launch every task whose deps are all terminal and that isn't already in-flight
+        for (const [id, state] of states) {
+          if (state.status !== 'waiting' || running.has(id)) continue;
+          const node = dag.nodes.get(id)!;
+          const allDepsTerminal =
+            node.dependsOn.length === 0 ||
+            node.dependsOn.every((d) => isTerminal(states.get(d)!.status));
+          if (!allDepsTerminal) continue;
+          const p = processTask(id).finally(() => running.delete(id));
+          running.set(id, p);
+        }
+
+        // All tasks terminal — done
+        if ([...states.values()].every((s) => isTerminal(s.status))) break;
+
+        if (running.size === 0) {
+          // Nothing in-flight but non-terminal tasks exist (e.g. trigger-wait states
+          // that processTask hasn't been called for yet). Poll briefly.
+          await new Promise((r) => setTimeout(r, POLL_INTERVAL_MS));
+        } else {
+          // Wait for any one task to finish, then re-scan for new launchables.
+          await Promise.race(running.values());
+        }
+      }
+
+      if (abortReason !== null) {
+        // Wait for in-flight tasks to honour the abort signal before marking states.
+        if (running.size > 0) await Promise.allSettled(running.values());
+        for (const [id, state] of states) {
+          if (!isTerminal(state.status)) {
+            // By the time allSettled resolves, processTask's try/finally has already
+            // set running tasks to success/failed/timeout. The only non-terminal
+            // statuses remaining here are waiting/idle tasks that were never started.
+            state.finishedAt = nowISO();
+            setTaskStatus(id, 'skipped');
+          }
+        }
+      }
+    } finally {
+      if (pipelineTimer) clearTimeout(pipelineTimer);
+      // Clean up the external abort signal listener to prevent dead references
+      // accumulating on long-lived shared AbortControllers.
+      if (options.signal) {
+        options.signal.removeEventListener('abort', externalAbortHandler);
+      }
+      // Safety net: drain any approvals still pending at shutdown (e.g. crash path).
+      if (approvalGateway.pending().length > 0) {
+        approvalGateway.abortAll('pipeline finished');
+      }
+      // Detach gateway → onEvent bridge so a long-lived gateway (host-supplied)
+      // doesn't keep firing into a dead run.
+      unsubscribeApprovals();
+    }
+
+    // ── Summary ──
+    const summary = { total: 0, success: 0, failed: 0, skipped: 0, timeout: 0, blocked: 0 };
+    for (const [, state] of states) {
+      summary.total++;
+      switch (state.status) {
+        case 'success':
+          summary.success++;
+          break;
+        case 'failed':
+          summary.failed++;
+          break;
+        case 'skipped':
+          summary.skipped++;
+          break;
+        case 'timeout':
+          summary.timeout++;
+          break;
+        case 'blocked':
+          summary.blocked++;
+          break;
+      }
+    }
+
+    const finishedAt = nowISO();
+    const durationMs = new Date(finishedAt).getTime() - new Date(startedAt).getTime();
+
+    if (abortReason !== null) {
+      const reasonText =
+        abortReason === 'timeout'
+          ? 'Pipeline timeout exceeded'
+          : abortReason === 'stop_all'
+            ? 'Pipeline stopped (on_failure: stop_all)'
+            : 'Pipeline aborted by host';
+      await executeHook(
+        config.hooks,
+        'pipeline_error',
+        buildPipelineErrorContext(pipelineInfo, reasonText, undefined, abortReason),
+        workDir,
+      );
+    } else {
+      await executeHook(
+        config.hooks,
+        'pipeline_complete',
+        buildPipelineCompleteContext(
+          { ...pipelineInfo, finished_at: finishedAt, duration_ms: durationMs },
+          summary,
+        ),
+        workDir,
+      );
+    }
+
+    const allSuccess =
+      abortReason === null &&
+      summary.failed === 0 &&
+      summary.timeout === 0 &&
+      summary.blocked === 0;
+
+    log.section('Pipeline summary');
+    log.quiet(
+      `status:   ${abortReason !== null ? `aborted (${abortReason})` : 'completed'}`,
+    );
+    log.quiet(`duration: ${(durationMs / 1000).toFixed(1)}s`);
+    log.quiet(
+      `counts:   total=${summary.total} success=${summary.success} ` +
+        `failed=${summary.failed} skipped=${summary.skipped} ` +
+        `timeout=${summary.timeout} blocked=${summary.blocked}`,
+    );
+    log.quiet('');
+    log.quiet('per-task:');
+    for (const [id, state] of states) {
+      const dur =
+        state.result?.durationMs != null ? `${(state.result.durationMs / 1000).toFixed(1)}s` : '-';
+      const exit = state.result?.exitCode ?? '-';
+      log.quiet(`  ${state.status.padEnd(8)} ${id}  (exit=${exit}, ${dur})`);
+    }
+
+    log.info('[pipeline]', `completed "${config.name}"`);
+    log.info(
+      '[pipeline]',
+      `Total: ${summary.total} | Success: ${summary.success} | Failed: ${summary.failed} | Skipped: ${summary.skipped} | Timeout: ${summary.timeout} | Blocked: ${summary.blocked}`,
+    );
+    log.info('[pipeline]', `Duration: ${(durationMs / 1000).toFixed(1)}s`);
+    log.info('[pipeline]', `Log: ${log.path}`);
+
+    emit({ type: 'run_end', runId, success: allSuccess, abortReason });
+    return { success: allSuccess, runId, logPath: log.path, summary, states: freezeStates(states) };
+  } finally {
+    // Close the persistent log file handle before pruning.
+    log.close();
+    // Prune old per-run log directories on every exit path (normal, blocked, or thrown).
+    // Exclude the current runId so a concurrent run cannot delete its own live directory.
+    if (maxLogRuns > 0) {
+      await pruneLogDirs(resolve(workDir, '.tagma', 'logs'), maxLogRuns, runId);
+    }
+  }
+}
+
+/**
+ * Delete the oldest subdirectories under `logsDir`, keeping only the most recent `keep`
+ * total runs (including the currently-live run identified by `excludeRunId`).
+ * Directories are sorted lexicographically; because runIds are prefixed with a base-36
+ * timestamp, lexicographic order equals chronological order.
+ *
+ * `excludeRunId` is always skipped from deletion even if it would otherwise be pruned —
+ * this prevents a concurrent run from removing a live log directory that is still in use.
+ *
+ * D10: The live run occupies one slot out of `keep`, so the maximum number of
+ * *historical* dirs to retain is `keep - 1`. Without this adjustment the function
+ * kept `keep` historical dirs plus 1 live dir = `keep + 1` total on disk.
+ */
+async function pruneLogDirs(logsDir: string, keep: number, excludeRunId: string): Promise<void> {
+  let entries: string[];
+  try {
+    entries = await readdir(logsDir);
+  } catch {
+    return; // logsDir doesn't exist yet — nothing to prune
+  }
+
+  // Only consider directories that look like run IDs (run_<...>), excluding the live run.
+  const runDirs = entries.filter((e) => e.startsWith('run_') && e !== excludeRunId).sort();
+  // keep - 1 historical slots (1 slot is reserved for the live excludeRunId).
+  const historyKeep = Math.max(0, keep - 1);
+  const toDelete = runDirs.slice(0, Math.max(0, runDirs.length - historyKeep));
+
+  await Promise.all(
+    toDelete.map((dir) =>
+      rm(resolve(logsDir, dir), { recursive: true, force: true }).catch(() => {
+        // Ignore deletion errors — stale dirs are better than a crash
+      }),
+    ),
+  );
+}
+
+function isTerminal(status: TaskStatus): boolean {
+  return (
+    status === 'success' ||
+    status === 'failed' ||
+    status === 'timeout' ||
+    status === 'skipped' ||
+    status === 'blocked'
+  );
+}
+
+/** Return a deep-copied, caller-safe snapshot of the states map. */
+function freezeStates(states: Map<string, TaskState>): ReadonlyMap<string, TaskState> {
+  const copy = new Map<string, TaskState>();
+  for (const [id, s] of states) {
+    copy.set(id, {
+      config: { ...s.config },
+      trackConfig: { ...s.trackConfig },
+      status: s.status,
+      result: s.result ? { ...s.result } : null,
+      startedAt: s.startedAt,
+      finishedAt: s.finishedAt,
+    });
+  }
+  return copy;
+}