port: sync from Mike/ralpi@c870efa1

AGENTS.md

@@ -124,6 +124,9 @@ Key config fields in `execution`:
 - `autoCommit` / `autoReview` / `saveReviews` — loop options (selectable at
   loop startup via `selectLoopOptions`; review is asked FIRST, commit is
   mandated when review is on)
+- `inactivityTimeoutMs` — hang detection: if no agent session event arrives
+  within this window (e.g. a bash subprocess that never returns), the task is
+  aborted (agent abort + bash subprocess kill). `0` = disabled (default)
 - `models` — slot-aware round-robin model list for parallel mode, with
   failover to the next model per task (only after exhausting same-model
   retries, see `maxSameModelAttempts`)

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mikefreno/omp-ralpi",
-  "version": "0.5.0",
+  "version": "0.6.0",
   "description": "Execute tasks from task files/PRD's using DAG-based dependency resolution with persistent progress tracking",
   "keywords": [
     "omp",

src/executor.ts

@@ -244,6 +244,13 @@ export async function runTask(
 	/** Review feedback from a rejected review — injected when re-executing
 	 *  a task in review-gated mode so the agent knows what to fix. */
 	reviewFeedback?: ReviewResult,
+	/** Session JSONL from a prior interrupted run of this task. When set and
+	 *  readable, the agent session reopens it and continues from the prior
+	 *  conversation instead of starting fresh. */
+	resumeSessionFile?: string,
+	/** Called with the session file path as soon as the agent session is
+	 *  created, so the caller can persist it for a later resume. */
+	onSessionFile?: (sessionFile: string) => void,
 ): Promise<{
 	success: boolean;
 	reflection?: Reflection;
@@ -253,6 +260,12 @@ export async function runTask(
 	outputPreview?: string;
 	commitMessages?: string[];
 	commitSummary?: string;
+	/** Path to the JSONL session file backing this run (for resume). */
+	sessionFile?: string;
+	/** True when a resume was requested but the session could not be opened
+	 *  from the file — the caller should clear the stored session file so
+	 *  retries start fresh. */
+	resumeFailed?: boolean;
 }> {
 	const startMs = Date.now();
 
@@ -384,6 +397,9 @@ export async function runTask(
 		config.thinkingLevel,
 		false, // noSkills — task sessions need skills
 		ctx.modelRegistry,
+		config.execution.inactivityTimeoutMs,
+		resumeSessionFile,
+		onSessionFile,
 	);
 
 	const durationMs = Date.now() - startMs;
@@ -408,6 +424,8 @@ export async function runTask(
 			success: false,
 			error: output.error,
 			durationMs,
+			sessionFile: output.sessionFile,
+			resumeFailed: output.resumeFailed,
 		};
 	}
 
@@ -438,6 +456,7 @@ export async function runTask(
 		outputPreview,
 		commitMessages,
 		commitSummary,
+		sessionFile: output.sessionFile,
 	};
 }
 
@@ -824,6 +843,12 @@ async function executeTask(
 		: null;
 	const worktreeDir = wt?.dir ?? projectDir;
 
+	// Session file from a prior interrupted run of this task. The first
+	// attempt reopens it so the agent continues with its prior conversation
+	// (tool calls, findings) instead of restarting from scratch; failover
+	// retries start fresh.
+	const resumeSessionFile = progress.getSessionFile(task.id);
+
 	while (modelAttempt < maxModelAttempts) {
 		// Model advancement happens in the cycling branch below (not here) so a
 		// same-model retry `continue` doesn't re-advance and accidentally swap
@@ -884,6 +909,11 @@ async function executeTask(
 				currentModel,
 				batchRender,
 				priorReview,
+				modelAttempt === 0 && sameModelAttempt === 0
+					? resumeSessionFile
+					: undefined,
+				(sessionFile) =>
+					progress.setSessionFile(task.id, sessionFile),
 			);
 
 			if (result.success) {
@@ -935,6 +965,10 @@ async function executeTask(
 						// A FAILED range computation (broken/stale base ref, git error) is
 						// logged as a distinct warning and is never treated as a clean,
 						// verified task — only a GENUINE "no changes" skips review.
+						// Accumulated rejected reviews from earlier passes — injected
+						// into the next review prompt so the reviewer sees prior
+						// findings and can verify they were addressed.
+						const priorReviews: ReviewResult[] = [];
 						while (true) {
 							if (!baseRef) {
 								sendChatMessage?.(
@@ -972,16 +1006,17 @@ async function executeTask(
 								reviewInfo.hash,
 								reviewInfo.subject,
 								reviewInfo.diff,
-								{
+							{
-									projectContext: config.prompts.projectContext,
+								projectContext: config.prompts.projectContext,
-									focus: config.prompts.reviewFocus,
+								focus: config.prompts.reviewFocus,
-									diffOptions: {
+								priorReviews,
-										extraPatterns: compileIgnorePatterns(
+								diffOptions: {
-											config.review.extraIgnorePatterns,
+									extraPatterns: compileIgnorePatterns(
-										),
+										config.review.extraIgnorePatterns,
-										ignorePaths: config.review.ignorePaths,
+									),
-									},
+									ignorePaths: config.review.ignorePaths,
 								},
+							},
 							);
 
 							const reviewModel = resolveFollowUpModel(
@@ -1003,6 +1038,7 @@ async function executeTask(
 									`review-${task.id}`,
 									config.execution.reviewTimeoutMs,
 									reviewModels,
+									config.execution.inactivityTimeoutMs,
 								);
 
 							if (!reviewResult.success) {
@@ -1096,6 +1132,9 @@ async function executeTask(
 								break; // changes already committed — merge proceeds
 							}
 
+							// Accumulate the rejected review so the next review pass
+							// sees prior findings and can verify they were addressed.
+							if (review) priorReviews.push(review);
 							attempt++;
 							reviewRetries++;
 							sendChatMessage?.(
@@ -1292,6 +1331,7 @@ async function executeTask(
 					finalCommitSummary,
 					finalReview,
 					reviewRetries,
+					result.sessionFile,
 				);
 				// Auto-update the PRD source file checkbox
 				try {
@@ -1305,6 +1345,12 @@ async function executeTask(
 
 			// Agent session failed (provider error).
 			// Pi's built-in in-call retry already exhausted for this attempt.
+			// A resumed session that couldn't be opened (corrupt/missing
+			// JSONL) must not be retried — forget it so later attempts (and
+			// future resumes) start fresh.
+			if (result.resumeFailed) {
+				progress.setSessionFile(task.id, undefined);
+			}
 			// Reattempt on the SAME model a few more times before cycling — a
 			// transient outage can outlast pi's per-prompt backoff window.
 			sameModelAttempt++;
@@ -1437,6 +1483,8 @@ async function runFollowUpSession(
 	widgetKeySuffix: string,
 	timeoutMs: number,
 	models: unknown[],
+	/** Inactivity timeout for the session (see runAgentSession). */
+	inactivityTimeoutMs = 0,
 ): Promise<{
 	result: Awaited<ReturnType<typeof runAgentSession>>;
 	toolCalls: ToolCallEntry[];
@@ -1535,6 +1583,7 @@ async function runFollowUpSession(
 					config.thinkingLevel,
 					false, // noSkills=false — follow-up sessions load skills too
 					ctx.modelRegistry,
+					inactivityTimeoutMs,
 				);
 
 				if (result.success) break;
@@ -1681,6 +1730,7 @@ async function runCommitSession(
 			`commit-${task.id}`,
 			config.execution.commitTimeoutMs,
 			commitModels,
+			config.execution.inactivityTimeoutMs,
 		);
 
 	if (commitResult.success) {
@@ -1778,6 +1828,52 @@ async function resolveConflictsSession(
 		return;
 	}
 
+	// Merge failed but produced no unmerged paths — no real conflicts to
+	// resolve. This can happen when the branch tip was already merged by the
+	// first attempt (mergeWorktree) before it aborted, or when the merge
+	// fails for a non-conflict reason (dirty index, stale ref). Don't spawn
+	// an agent session for zero conflicts — abort or complete and finish.
+	if (attempt.conflicts.length === 0) {
+		// Try to complete whatever merge state exists; if there's nothing to
+		// commit, abort to leave the working tree clean.
+		if (completeMerge(projectDir)) {
+			sendChatMessage?.(
+				`✓ conflicts auto-resolved for ${task.id} · ${task.title}`,
+			);
+			removeWorktree(projectDir, worktree);
+			progress.markCompleted(
+				task.id,
+				0,
+				undefined,
+				undefined,
+				undefined,
+				[],
+				"",
+				undefined,
+				0,
+			);
+			try {
+				updateTaskInFile(project.sourcePath, task.id, "completed");
+			} catch {
+				// Best-effort
+			}
+			return;
+		}
+		abortMerge(projectDir);
+		sendChatMessage?.(
+			`~ conflict resolution for ${task.id} · ${task.title} — merge produced no conflicts but could not be completed`,
+		);
+		progress.markFailed(
+			task.id,
+			`Merge of ${branch} produced no conflicts but could not be completed`,
+		);
+		try {
+			updateTaskInFile(project.sourcePath, task.id, "failed");
+		} catch {
+			// Best-effort
+		}
+		return;
+	}
 	// Conflicts exist — spawn a resolution agent session.
 	const prompt = buildConflictResolutionPrompt(
 		task,
@@ -1807,6 +1903,7 @@ async function resolveConflictsSession(
 		`resolve-${task.id}`,
 		config.execution.commitTimeoutMs,
 		models,
+		config.execution.inactivityTimeoutMs,
 	);
 
 	if (!result.success) {

src/progress.ts

@@ -185,11 +185,13 @@ export class ProgressTracker {
 	}
 
 	/** Mark a task as in progress */
-	markInProgress(taskId: string): void {
+	markInProgress(taskId: string, sessionFile?: string): void {
 		const prd = this.getPRD();
 		this.ensureTask(prd, taskId);
 		prd.tasks[taskId].status = "in_progress";
 		prd.tasks[taskId].startedAt = new Date().toISOString();
+		if (sessionFile !== undefined)
+			prd.tasks[taskId].sessionFile = sessionFile;
 		this.save();
 	}
 
@@ -204,6 +206,7 @@ export class ProgressTracker {
 		commitSummary?: string,
 		review?: ReviewResult,
 		reviewRetries?: number,
+		sessionFile?: string,
 	): void {
 		const prd = this.getPRD();
 		this.ensureTask(prd, taskId);
@@ -218,6 +221,7 @@ export class ProgressTracker {
 		if (review) prd.tasks[taskId].review = review;
 		if (reviewRetries !== undefined)
 			prd.tasks[taskId].reviewRetries = reviewRetries;
+		if (sessionFile !== undefined) prd.tasks[taskId].sessionFile = sessionFile;
 		this.save();
 	}
 
@@ -236,6 +240,22 @@ export class ProgressTracker {
 		return prd.tasks[taskId]?.status ?? "pending";
 	}
 
+	/** Get the persisted session file path for a task (for resume), if any. */
+	getSessionFile(taskId: string): string | undefined {
+		const prd = this.getPRD();
+		return prd.tasks[taskId]?.sessionFile;
+	}
+
+	/** Persist the session file path for a task without changing its status.
+	 *  Called as soon as an agent session is created so an interrupted run
+	 *  can be resumed from the JSONL history. Pass undefined to clear. */
+	setSessionFile(taskId: string, sessionFile: string | undefined): void {
+		const prd = this.getPRD();
+		this.ensureTask(prd, taskId);
+		prd.tasks[taskId].sessionFile = sessionFile;
+		this.save();
+	}
+
 	/** Get IDs of all completed tasks */
 	getCompletedTaskIds(): string[] {
 		const prd = this.getPRD();

src/prompts.ts

@@ -29,6 +29,12 @@ export interface ReviewPromptOptions {
 	focus?: string;
 	/** Noise-filter overrides (config.review.*). */
 	diffOptions?: DiffOptions;
+	/** Prior review results from earlier passes in a review-gated loop.
+	 *  Each rejected review is injected into the next review prompt so the
+	 *  reviewer can verify prior findings were addressed and catch new
+	 *  regressions introduced by the fix attempt — instead of re-reviewing
+	 *  from scratch. */
+	priorReviews?: ReviewResult[];
 }
 
 // ─── Task Prompt ─────────────────────────────────────────────────────────────
@@ -258,8 +264,12 @@ export function buildReviewPrompt(
 	parts.push(renderDiffSection(summary, filtered, "### Diff"));
 	parts.push("");
 
-	// ── Custom Review Focus ──
+	// ── Prior Review History (review-gated re-review) ──
 
+	parts.push(renderPriorReviews(opts.priorReviews ?? []));
+	if (opts.priorReviews && opts.priorReviews.length > 0) parts.push("");
+
+	// ── Custom Review Focus ──
 	if (opts.focus) {
 		parts.push("## Custom Review Focus");
 		parts.push(opts.focus);
@@ -360,7 +370,11 @@ export function buildReviewPromptUncommitted(
 	parts.push(
 		renderDiffSection(summary, filtered, "### Current Tracked Diff (git diff)"),
 	);
-	parts.push("");
+
+	// ── Prior Review History (review-gated re-review) ──
+
+	parts.push(renderPriorReviews(opts.priorReviews ?? []));
+	if (opts.priorReviews && opts.priorReviews.length > 0) parts.push("");
 
 	// ── Custom Review Focus ──
 
@@ -469,6 +483,45 @@ function renderDiffSection(
 	return lines.join("\n");
 }
 
+/**
+ * Render a "Prior Review History" section from earlier rejected reviews.
+ * Returns an empty string when there are no prior reviews so callers omit
+ * the section entirely.
+ *
+ * Each prior review's verdict, summary, and findings are listed so the
+ * reviewer can verify the developer addressed them and watch for new
+ * regressions — instead of re-reviewing from scratch on each pass.
+ */
+function renderPriorReviews(priorReviews: ReviewResult[]): string {
+	if (priorReviews.length === 0) return "";
+	const lines: string[] = [];
+	lines.push("## Prior Review History");
+	lines.push(
+		"Previous review pass(es) rejected this task. Verify each finding was",
+		"addressed in the current diff and watch for new regressions:",
+	);
+	lines.push("");
+	for (let i = 0; i < priorReviews.length; i++) {
+		const r = priorReviews[i];
+		if (!r) continue;
+		lines.push(`### Review ${i + 1} — ${r.verdict.toUpperCase()}`);
+		lines.push(`Summary: ${r.summary}`);
+		if (r.findings.length > 0) {
+			lines.push("Findings:");
+			for (const f of r.findings) {
+				const loc = f.file
+					? f.line
+						? ` (${f.file}:${f.line})`
+						: ` (${f.file})`
+					: "";
+				lines.push(`- [${f.severity}]${loc} ${f.message}`);
+			}
+		}
+		lines.push("");
+	}
+	return lines.join("\n");
+}
+
 function reviewInstructions(): string[] {
 	return [
 		"- **Correctness**: Does the implementation fulfill the task requirements?",

src/types.ts

@@ -161,6 +161,10 @@ export interface TaskProgressInfo {
 	commitSummary?: string;
 	/** Number of review-fix re-execution attempts made (review-gated mode) */
 	reviewRetries?: number;
+	/** Path to the JSONL session file backing this task's agent session,
+	 *  persisted so a resume can reopen it and continue where the
+	 *  interrupted session left off. */
+	sessionFile?: string;
 }
 
 export interface ProgressState {
@@ -205,6 +209,11 @@ export interface RalpiConfig {
 	execution: {
 		/** Task execution timeout in milliseconds */
 		timeoutMs: number;
+		/** Inactivity timeout in milliseconds — if no agent session event
+		 *  arrives within this window, the task is considered hung (e.g. a
+		 *  bash subprocess that never returns) and the session is aborted.
+		 *  0 = disabled. */
+		inactivityTimeoutMs: number;
 		/** Maximum parallel tasks (0 = unlimited) */
 		maxParallel: number;
 		/** Round-robin model list for parallel tasks (empty = inherit parent model) */
@@ -301,6 +310,7 @@ export const DEFAULT_CONFIG: RalpiConfig = {
 	},
 	execution: {
 		timeoutMs: 0, // 0 = inherit Pi's own defaults (no ralpi-level timeout)
+		inactivityTimeoutMs: 0, // 0 = disabled (no inactivity hang detection)
 		maxParallel: 3,
 		models: [],
 		autoCommit: true,

src/utils.ts

@@ -611,6 +611,18 @@ export async function runAgentSession(
    *  rate-limit normalization) are available. When omitted, the SDK creates
    *  a fresh registry from models.json only — extension providers are lost. */
   modelRegistry?: ModelRegistry,
+  /** Inactivity timeout in milliseconds — if no agent session event arrives
+   *  within this window, the task is considered hung (e.g. a bash subprocess
+   *  that never returns) and the session is aborted. 0 = disabled. */
+  inactivityTimeoutMs = 0,
+  /** Existing session JSONL file to resume. The session reopens the file and
+   *  appends to it, so the agent sees the full prior conversation and can
+   *  continue rather than redo prior tool calls. When the file is missing,
+   *  a fresh session is created instead (with a warning). */
+  resumeSessionFile?: string,
+  /** Called with the session file path as soon as the session is created, so
+   *  callers can persist it for resume before the session completes. */
+  onSessionFile?: (sessionFile: string) => void,
 ): Promise<{
   success: boolean;
   text: string;
@@ -618,6 +630,13 @@ export async function runAgentSession(
   toolUsage: ToolUsage;
   stopReason?: string;
   events: AgentSessionEvent[];
+  /** Path to the JSONL session file backing this session (set once the
+   *  session is created; enables resume). */
+  sessionFile?: string;
+  /** True when a resume was requested but the session could not be created
+   *  from the file (corrupt/unreadable JSONL). Callers should clear the
+   *  stored session file so retries start fresh. */
+  resumeFailed?: boolean;
 }> {
   const toolUsage: ToolUsage = {
     read: 0,
@@ -638,12 +657,45 @@ export async function runAgentSession(
     session?: Awaited<ReturnType<typeof createAgentSession>>["session"];
   } = {};
 
+  let sessionFile: string | undefined;
+  let sessionCreated = false;
+  // Inactivity watchdog: aborts the session when no events arrive within
+  // inactivityTimeoutMs. The SDK emits an event for every tool start/end/
+  // update and message start/end, so silence means the agent is stuck
+  // (typically a hung bash subprocess producing no output).
+  let inactivityInterval: NodeJS.Timeout | null = null;
+  let inactivityAborted = false;
+  let lastEventTime = 0;
+
   try {
     // Loop sessions load the full normal omp context: extensions (so all
     // extension-provided tools register) and project context (AGENTS.md).
+    // Persist sessions under the ralpi project's `.ralpi/sessions/` so they
+    // survive worktree removal and are findable from the main repo on resume.
+    // Worktrees live inside `<project>/.ralpi/worktrees/...`, so walking up
+    // from the agent's cwd always finds the main project's `.ralpi` first.
+    const ralpiDir = findRalpiDir(cwd);
+    const sessionDir = ralpiDir
+      ? path.join(ralpiDir, ".ralpi", "sessions")
+      : path.join(cwd, ".ralpi", "sessions");
+
+    let sessionManager: SessionManager;
+    if (resumeSessionFile && fs.existsSync(resumeSessionFile)) {
+      sessionManager = await SessionManager.open(resumeSessionFile, sessionDir, undefined, {
+        initialCwd: cwd,
+      });
+    } else {
+      if (resumeSessionFile) {
+        console.warn(
+          `[ralpi] resume session file not found (${resumeSessionFile}) — starting a fresh session`,
+        );
+      }
+      sessionManager = SessionManager.create(cwd, sessionDir);
+    }
+
     const result = await createAgentSession({
       cwd,
-      sessionManager: SessionManager.inMemory(cwd),
+      sessionManager,
       settingsManager: Settings.init({ cwd, agentDir: getAgentDir() }),
       // Loop sessions intentionally load extensions (no disableExtensionDiscovery),
       // plus skills and project context via default discovery.
@@ -655,8 +707,12 @@ export async function runAgentSession(
       modelRegistry,
       agentRegistry: new AgentRegistry(),
     });
+    sessionCreated = true;
     sessionRef.session = result.session;
 
+    sessionFile = result.session.sessionFile;
+    if (sessionFile) onSessionFile?.(sessionFile);
+
     // Wire external abort signal
     const abortHandler = () => result.session.agent.abort();
     signal?.addEventListener("abort", abortHandler, { once: true });
@@ -664,8 +720,26 @@ export async function runAgentSession(
     let finalText = "";
     let errorMessage: string | undefined;
     let stopReason: string | undefined;
+    lastEventTime = Date.now();
+
+    // Inactivity watchdog: check the silence window on an interval and abort
+    // (plus kill any hung bash subprocess) when it is exceeded.
+    if (inactivityTimeoutMs > 0) {
+      const intervalMs = Math.min(inactivityTimeoutMs, 5000);
+      inactivityInterval = setInterval(() => {
+        if (!sessionRef.session) return;
+        if (Date.now() - lastEventTime <= inactivityTimeoutMs) return;
+        inactivityAborted = true;
+        sessionRef.session.agent.abort();
+        sessionRef.session.abortBash();
+        errorMessage = `Task aborted: inactivity timeout (no events for ${Math.round(inactivityTimeoutMs / 1000)}s)`;
+        if (inactivityInterval) clearInterval(inactivityInterval);
+        inactivityInterval = null;
+      }, intervalMs);
+    }
 
     const unsubscribe = result.session.subscribe((event) => {
+      lastEventTime = Date.now();
       onEvent?.(event);
 
       if (event.type === "message_end") {
@@ -677,7 +751,10 @@ export async function runAgentSession(
         };
         if (message.role !== "assistant") return;
         if (message.stopReason) stopReason = message.stopReason;
-        if (message.errorMessage) errorMessage = message.errorMessage;
+        // Keep the inactivity-timeout message: the abort's own errorMessage
+        // would otherwise clobber the (more useful) hang explanation.
+        if (message.errorMessage && !inactivityAborted)
+          errorMessage = message.errorMessage;
         const text = extractAssistantText(message.content);
         if (text) finalText = text;
       }
@@ -710,6 +787,7 @@ export async function runAgentSession(
         toolUsage,
         stopReason,
         events: [], // streamed to file
+        sessionFile,
       };
     }
 
@@ -719,6 +797,7 @@ export async function runAgentSession(
       toolUsage,
       stopReason,
       events: [],
+      sessionFile,
     };
   } catch (error) {
     if (timeoutHandle) clearTimeout(timeoutHandle);
@@ -728,9 +807,14 @@ export async function runAgentSession(
       error: error instanceof Error ? error.message : String(error),
       toolUsage,
       events: [],
+      sessionFile,
+      // A requested resume that failed to open (corrupt/unreadable file)
+      // should not be retried — callers clear the stored file and go fresh.
+      resumeFailed: resumeSessionFile !== undefined && !sessionCreated,
     };
   } finally {
     sessionRef.session?.dispose();
+    if (inactivityInterval) clearInterval(inactivityInterval);
   }
 }