gsd-pi 2.50.0-dev.d210a87 → 2.51.0-dev.7d435fe

package/packages/pi-coding-agent/src/modes/rpc/rpc-types.ts

@@ -11,6 +11,13 @@ import type { SessionStats } from "../../core/agent-session.js";
 import type { BashResult } from "../../core/bash-executor.js";
 import type { CompactionResult } from "../../core/compaction/index.js";
 
+// ============================================================================
+// RPC Protocol Versioning
+// ============================================================================
+
+/** Supported protocol versions. v1 is the implicit default; v2 requires an init handshake. */
+export type RpcProtocolVersion = 1 | 2;
+
 // ============================================================================
 // RPC Commands (stdin)
 // ============================================================================
@@ -69,7 +76,12 @@ export type RpcCommand =
 	// Bridge-hosted native terminal
 	| { id?: string; type: "terminal_input"; data: string }
 	| { id?: string; type: "terminal_resize"; cols: number; rows: number }
-	| { id?: string; type: "terminal_redraw" };
+	| { id?: string; type: "terminal_redraw" }
+
+	// v2 Protocol
+	| { id?: string; type: "init"; protocolVersion: 2; clientId?: string }
+	| { id?: string; type: "shutdown"; graceful?: boolean }
+	| { id?: string; type: "subscribe"; events: string[] };
 
 // ============================================================================
 // RPC Slash Command (for get_commands response)
@@ -120,9 +132,9 @@ export interface RpcSessionState {
 // Success responses with data
 export type RpcResponse =
 	// Prompting (async - events follow)
-	| { id?: string; type: "response"; command: "prompt"; success: true }
-	| { id?: string; type: "response"; command: "steer"; success: true }
-	| { id?: string; type: "response"; command: "follow_up"; success: true }
+	| { id?: string; type: "response"; command: "prompt"; success: true; runId?: string }
+	| { id?: string; type: "response"; command: "steer"; success: true; runId?: string }
+	| { id?: string; type: "response"; command: "follow_up"; success: true; runId?: string }
 	| { id?: string; type: "response"; command: "abort"; success: true }
 	| { id?: string; type: "response"; command: "new_session"; success: true; data: { cancelled: boolean } }
 
@@ -216,9 +228,54 @@ export type RpcResponse =
 	| { id?: string; type: "response"; command: "terminal_resize"; success: true }
 	| { id?: string; type: "response"; command: "terminal_redraw"; success: true }
 
+	// v2 Protocol
+	| { id?: string; type: "response"; command: "init"; success: true; data: RpcInitResult }
+	| { id?: string; type: "response"; command: "shutdown"; success: true }
+	| { id?: string; type: "response"; command: "subscribe"; success: true }
+
 	// Error response (any command can fail)
 	| { id?: string; type: "response"; command: string; success: false; error: string };
 
+// ============================================================================
+// v2 Protocol Types
+// ============================================================================
+
+/** Result of the init handshake (v2 only) */
+export interface RpcInitResult {
+	protocolVersion: 2;
+	sessionId: string;
+	capabilities: {
+		events: string[];
+		commands: string[];
+	};
+}
+
+/** v2 execution_complete event — emitted when a prompt/steer/follow_up finishes */
+export interface RpcExecutionCompleteEvent {
+	type: "execution_complete";
+	runId: string;
+	status: "completed" | "error" | "cancelled";
+	reason?: string;
+	stats: SessionStats;
+}
+
+/** v2 cost_update event — emitted per-turn with running cost data */
+export interface RpcCostUpdateEvent {
+	type: "cost_update";
+	runId: string;
+	turnCost: number;
+	cumulativeCost: number;
+	tokens: {
+		input: number;
+		output: number;
+		cacheRead: number;
+		cacheWrite: number;
+	};
+}
+
+/** Discriminated union of all v2-only event types */
+export type RpcV2Event = RpcExecutionCompleteEvent | RpcCostUpdateEvent;
+
 // ============================================================================
 // Extension UI Events (stdout)
 // ============================================================================

package/packages/pi-coding-agent/src/utils/shell.ts

@@ -192,7 +192,6 @@ export function killProcessTree(pid: number): void {
 		try {
 			spawn("taskkill", ["/F", "/T", "/PID", String(pid)], {
 				stdio: "ignore",
-				detached: true,
 			});
 		} catch {
 			// Ignore errors if taskkill fails

package/pkg/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@glittercowboy/gsd",
-  "version": "2.50.0",
+  "version": "2.51.0",
   "piConfig": {
     "name": "gsd",
     "configDir": ".gsd"

package/src/resources/extensions/async-jobs/async-bash-tool.ts

@@ -14,7 +14,7 @@ import {
 	DEFAULT_MAX_LINES,
 } from "@gsd/pi-coding-agent";
 import { Type } from "@sinclair/typebox";
-import { spawn } from "node:child_process";
+import { spawn, spawnSync } from "node:child_process";
 import { createWriteStream } from "node:fs";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
@@ -38,17 +38,24 @@ function getTempFilePath(): string {
 }
 
 /**
- * Kill a process and its children. Uses process group kill on Unix.
+ * Kill a process and its children (cross-platform).
+ * Uses process group kill on Unix; taskkill /F /T on Windows.
  */
 function killTree(pid: number): void {
-	try {
-		// Kill the process group (negative PID)
-		process.kill(-pid, "SIGTERM");
-	} catch {
+	if (process.platform === "win32") {
 		try {
-			process.kill(pid, "SIGTERM");
+			spawnSync("taskkill", ["/F", "/T", "/PID", String(pid)], {
+				timeout: 5_000,
+				stdio: "ignore",
+			});
+		} catch {
+			try { process.kill(pid, "SIGTERM"); } catch { /* already exited */ }
+		}
+	} else {
+		try {
+			process.kill(-pid, "SIGTERM");
 		} catch {
-			// Already exited
+			try { process.kill(pid, "SIGTERM"); } catch { /* already exited */ }
 		}
 	}
 }
@@ -118,9 +125,13 @@ function executeBashInBackground(
 		const rewrittenCommand = rewriteCommandWithRtk(command);
 		const resolvedCommand = sanitizeCommand(rewrittenCommand);
 
+		// On Windows, detached: true sets CREATE_NEW_PROCESS_GROUP which can
+		// cause EINVAL in VSCode/ConPTY terminal contexts.  The bg-shell
+		// extension already guards this (process-manager.ts); align here.
+		// Process-tree cleanup uses taskkill /F /T on Windows regardless.
 		const child = spawn(shell, [...args, resolvedCommand], {
 			cwd,
-			detached: true,
+			detached: process.platform !== "win32",
 			env: { ...process.env },
 			stdio: ["ignore", "pipe", "pipe"],
 		});
@@ -143,8 +154,8 @@ function executeBashInBackground(
 				// If the process ignores SIGTERM, escalate to SIGKILL
 				sigkillHandle = setTimeout(() => {
 					if (child.pid) {
-						try { process.kill(-child.pid, "SIGKILL"); } catch { /* ignore */ }
-						try { process.kill(child.pid, "SIGKILL"); } catch { /* ignore */ }
+						// killTree already uses taskkill /F /T on Windows
+						killTree(child.pid);
 					}
 
 					// Hard deadline: if even SIGKILL doesn't trigger 'close',

package/src/resources/extensions/claude-code-cli/stream-adapter.ts

@@ -113,6 +113,20 @@ function makeErrorMessage(model: string, errorMsg: string): AssistantMessage {
 	};
 }
 
+/**
+ * Generator exhaustion without a terminal result means the SDK stream was
+ * interrupted mid-turn. Surface it as an error so downstream recovery logic
+ * can classify and retry it instead of treating it as a clean completion.
+ */
+export function makeStreamExhaustedErrorMessage(model: string, lastTextContent: string): AssistantMessage {
+	const errorMsg = "stream_exhausted_without_result";
+	const message = makeErrorMessage(model, errorMsg);
+	if (lastTextContent) {
+		message.content = [{ type: "text", text: lastTextContent }];
+	}
+	return message;
+}
+
 // ---------------------------------------------------------------------------
 // streamSimple implementation
 // ---------------------------------------------------------------------------
@@ -339,26 +353,11 @@ async function pumpSdkMessages(
 			}
 		}
 
-		// Generator exhausted without a result message (unexpected)
-		const fallbackContent: AssistantMessage["content"] = [];
-		if (lastTextContent) {
-			fallbackContent.push({ type: "text", text: lastTextContent });
-		}
-		if (fallbackContent.length === 0) {
-			fallbackContent.push({ type: "text", text: "(Claude Code session ended without a response)" });
-		}
-
-		const fallback: AssistantMessage = {
-			role: "assistant",
-			content: fallbackContent,
-			api: "anthropic-messages",
-			provider: "claude-code",
-			model: modelId,
-			usage: { ...ZERO_USAGE },
-			stopReason: "stop",
-			timestamp: Date.now(),
-		};
-		stream.push({ type: "done", reason: "stop", message: fallback });
+		// Generator exhaustion without a terminal result is a stream interruption,
+		// not a successful completion. Emitting an error lets GSD classify it as a
+		// transient provider failure instead of advancing auto-mode state.
+		const fallback = makeStreamExhaustedErrorMessage(modelId, lastTextContent);
+		stream.push({ type: "error", reason: "error", error: fallback });
 	} catch (err) {
 		const errorMsg = err instanceof Error ? err.message : String(err);
 		stream.push({

package/src/resources/extensions/claude-code-cli/tests/stream-adapter.test.ts

@@ -0,0 +1,21 @@
+import { describe, test } from "node:test";
+import assert from "node:assert/strict";
+import { makeStreamExhaustedErrorMessage } from "../stream-adapter.ts";
+
+describe("stream-adapter — exhausted stream fallback (#2575)", () => {
+	test("generator exhaustion becomes an error message instead of clean completion", () => {
+		const message = makeStreamExhaustedErrorMessage("claude-sonnet-4-20250514", "partial answer");
+
+		assert.equal(message.stopReason, "error");
+		assert.equal(message.errorMessage, "stream_exhausted_without_result");
+		assert.deepEqual(message.content, [{ type: "text", text: "partial answer" }]);
+	});
+
+	test("generator exhaustion without prior text still exposes a classifiable error", () => {
+		const message = makeStreamExhaustedErrorMessage("claude-sonnet-4-20250514", "");
+
+		assert.equal(message.stopReason, "error");
+		assert.equal(message.errorMessage, "stream_exhausted_without_result");
+		assert.match(String((message.content[0] as any)?.text ?? ""), /Claude Code error: stream_exhausted_without_result/);
+	});
+});

package/src/resources/extensions/gsd/auto-dispatch.ts

@@ -626,6 +626,25 @@ export const DISPATCH_RULES: DispatchRule[] = [
     match: async ({ state, mid, midTitle, basePath }) => {
       if (state.phase !== "completing-milestone") return null;
 
+      // Safety guard (#2675): block completion when VALIDATION verdict is
+      // needs-remediation. The state machine treats needs-remediation as
+      // terminal (to prevent validate-milestone loops per #832), but
+      // completing-milestone should NOT proceed — remediation work is needed.
+      const validationFile = resolveMilestoneFile(basePath, mid, "VALIDATION");
+      if (validationFile) {
+        const validationContent = await loadFile(validationFile);
+        if (validationContent) {
+          const verdict = extractVerdict(validationContent);
+          if (verdict === "needs-remediation") {
+            return {
+              action: "stop",
+              reason: `Cannot complete milestone ${mid}: VALIDATION verdict is "needs-remediation". Address the remediation findings and re-run validation, or update the verdict manually.`,
+              level: "warning",
+            };
+          }
+        }
+      }
+
       // Safety guard (#1368): verify all roadmap slices have SUMMARY files.
       const missingSlices = findMissingSummaries(basePath, mid);
       if (missingSlices.length > 0) {

package/src/resources/extensions/gsd/auto-start.ts

@@ -67,6 +67,7 @@ import {
   getDebugLogPath,
 } from "./debug-logger.js";
 import { parseUnitId } from "./unit-id.js";
+import { setLogBasePath } from "./workflow-logger.js";
 import type { AutoSession } from "./auto/session.js";
 import {
   existsSync,
@@ -461,6 +462,7 @@ export async function bootstrapAutoSession(
     s.verbose = verboseMode;
     s.cmdCtx = ctx;
     s.basePath = base;
+    setLogBasePath(base);
     s.unitDispatchCount.clear();
     s.unitRecoveryCount.clear();
     s.lastBudgetAlertLevel = 0;

package/src/resources/extensions/gsd/auto-timers.ts

@@ -15,6 +15,8 @@ import { computeBudgets, resolveExecutorContextWindow } from "./context-budget.j
 import {
   getInFlightToolCount,
   getOldestInFlightToolStart,
+  clearInFlightTools,
+  hasInteractiveToolInFlight,
 } from "./auto-tool-tracking.js";
 import { detectWorkingTreeActivity } from "./auto-supervisor.js";
 import { closeoutUnit, type CloseoutOptions } from "./auto-unit-closeout.js";
@@ -146,7 +148,17 @@ export function startUnitSupervision(sctx: SupervisionContext): void {
 
       // Agent has tool calls currently executing — not idle, just waiting.
       // But only suppress recovery if the tool started recently.
+      let stalledToolDetected = false;
       if (getInFlightToolCount() > 0) {
+        // User-interactive tools (ask_user_questions, secure_env_collect) block
+        // waiting for human input by design — never treat them as stalled (#2676).
+        if (hasInteractiveToolInFlight()) {
+          writeUnitRuntimeRecord(s.basePath, unitType, unitId, s.currentUnit.startedAt, {
+            lastProgressAt: Date.now(),
+            lastProgressKind: "interactive-tool-waiting",
+          });
+          return;
+        }
         const oldestStart = getOldestInFlightToolStart()!;
         const toolAgeMs = Date.now() - oldestStart;
         if (toolAgeMs < idleTimeoutMs) {
@@ -156,6 +168,12 @@ export function startUnitSupervision(sctx: SupervisionContext): void {
           });
           return;
         }
+        // Tool has been in-flight longer than idle timeout — treat as hung.
+        // Clear the stale entries so subsequent ticks don't re-detect them,
+        // and set the flag so the filesystem-activity check below does not
+        // override the stall verdict (#2527).
+        stalledToolDetected = true;
+        clearInFlightTools();
         ctx.ui.notify(
           `Stalled tool detected: a tool has been in-flight for ${Math.round(toolAgeMs / 60000)}min. Treating as hung — attempting idle recovery.`,
           "warning",
@@ -163,7 +181,9 @@ export function startUnitSupervision(sctx: SupervisionContext): void {
       }
 
       // Check if the agent is producing work on disk.
-      if (detectWorkingTreeActivity(s.basePath)) {
+      // Skip this when a stalled tool was just detected — filesystem changes
+      // from earlier in the task should not override the stall verdict (#2527).
+      if (!stalledToolDetected && detectWorkingTreeActivity(s.basePath)) {
         writeUnitRuntimeRecord(s.basePath, unitType, unitId, s.currentUnit.startedAt, {
           lastProgressAt: Date.now(),
           lastProgressKind: "filesystem-activity",
@@ -180,6 +200,10 @@ export function startUnitSupervision(sctx: SupervisionContext): void {
       const recovery = await recoverTimedOutUnit(ctx, pi, unitType, unitId, "idle", buildRecoveryContext());
       if (recovery === "recovered") return;
 
+      // Guard: recoverTimedOutUnit is async — pauseAuto/stopAuto may have
+      // set s.currentUnit = null during the await (#2527).
+      if (!s.currentUnit) return;
+
       writeUnitRuntimeRecord(s.basePath, unitType, unitId, s.currentUnit.startedAt, {
         phase: "paused",
       });

package/src/resources/extensions/gsd/auto-tool-tracking.ts

@@ -4,15 +4,27 @@
  * can distinguish "waiting for tool completion" from "truly idle".
  */
 
-const inFlightTools = new Map<string, number>();
+interface InFlightTool {
+  startedAt: number;
+  toolName: string;
+}
+
+const inFlightTools = new Map<string, InFlightTool>();
+
+/**
+ * Tools that block waiting for human input by design.
+ * The idle watchdog must not treat these as stalled.
+ */
+const INTERACTIVE_TOOLS = new Set(["ask_user_questions", "secure_env_collect"]);
 
 /**
  * Mark a tool execution as in-flight.
- * Records start time so the idle watchdog can detect tools hung longer than the idle timeout.
+ * Records start time and tool name so the idle watchdog can detect tools
+ * hung longer than the idle timeout while exempting interactive tools.
  */
-export function markToolStart(toolCallId: string, isActive: boolean): void {
+export function markToolStart(toolCallId: string, isActive: boolean, toolName?: string): void {
   if (!isActive) return;
-  inFlightTools.set(toolCallId, Date.now());
+  inFlightTools.set(toolCallId, { startedAt: Date.now(), toolName: toolName ?? "unknown" });
 }
 
 /**
@@ -29,7 +41,7 @@ export function getOldestInFlightToolAgeMs(): number {
   if (inFlightTools.size === 0) return 0;
   let oldestStart = Infinity;
   for (const t of inFlightTools.values()) {
-    if (t < oldestStart) oldestStart = t;
+    if (t.startedAt < oldestStart) oldestStart = t.startedAt;
   }
   return Date.now() - oldestStart;
 }
@@ -48,11 +60,23 @@ export function getOldestInFlightToolStart(): number | undefined {
   if (inFlightTools.size === 0) return undefined;
   let oldest = Infinity;
   for (const t of inFlightTools.values()) {
-    if (t < oldest) oldest = t;
+    if (t.startedAt < oldest) oldest = t.startedAt;
   }
   return oldest;
 }
 
+/**
+ * Returns true if any currently in-flight tool is a user-interactive tool
+ * (e.g. ask_user_questions, secure_env_collect) that blocks waiting for
+ * human input. These must be exempt from idle stall detection.
+ */
+export function hasInteractiveToolInFlight(): boolean {
+  for (const { toolName } of inFlightTools.values()) {
+    if (INTERACTIVE_TOOLS.has(toolName)) return true;
+  }
+  return false;
+}
+
 /**
  * Clear all in-flight tool tracking state.
  */

package/src/resources/extensions/gsd/auto-worktree.ts

@@ -82,6 +82,10 @@ const ROOT_STATE_FILES = [
   "QUEUE.md",
   "completed-units.json",
   "metrics.json",
+  // NOTE: preferences.md is intentionally NOT in ROOT_STATE_FILES.
+  // Forward-sync (main → worktree) is handled explicitly in syncGsdStateToWorktree().
+  // Back-sync (worktree → main) must NEVER overwrite the project root's copy
+  // because the project root is authoritative for preferences (#2684).
 ] as const;
 
 /**
@@ -416,6 +420,22 @@ export function syncGsdStateToWorktree(
     }
   }
 
+  // Forward-sync preferences.md from project root to worktree (additive only).
+  // NOT in ROOT_STATE_FILES because syncWorktreeStateBack() must never overwrite
+  // the project root's preferences — the project root is authoritative (#2684).
+  {
+    const src = join(mainGsd, "preferences.md");
+    const dst = join(wtGsd, "preferences.md");
+    if (existsSync(src) && !existsSync(dst)) {
+      try {
+        cpSync(src, dst);
+        synced.push("preferences.md");
+      } catch {
+        /* non-fatal */
+      }
+    }
+  }
+
   // Sync milestones: copy entire milestone directories that are missing
   const mainMilestonesDir = join(mainGsd, "milestones");
   const wtMilestonesDir = join(wtGsd, "milestones");
@@ -946,6 +966,7 @@ function copyPlanningArtifacts(srcBase: string, wtPath: string): void {
     "STATE.md",
     "KNOWLEDGE.md",
     "OVERRIDES.md",
+    "preferences.md",
   ]) {
     safeCopy(join(srcGsd, file), join(dstGsd, file), { force: true });
   }

package/src/resources/extensions/gsd/auto.ts

@@ -73,6 +73,7 @@ import {
   getOldestInFlightToolAgeMs as _getOldestInFlightToolAgeMs,
   getInFlightToolCount,
   getOldestInFlightToolStart,
+  hasInteractiveToolInFlight,
   clearInFlightTools,
 } from "./auto-tool-tracking.js";
 import { closeoutUnit } from "./auto-unit-closeout.js";
@@ -114,6 +115,7 @@ import {
   formatCost,
   formatTokenCount,
 } from "./metrics.js";
+import { setLogBasePath } from "./workflow-logger.js";
 import { join } from "node:path";
 import { readFileSync, existsSync, mkdirSync, writeFileSync, unlinkSync } from "node:fs";
 import { atomicWriteSync } from "./atomic-write.js";
@@ -374,8 +376,8 @@ export function getAutoModeStartModel(): {
 }
 
 // Tool tracking — delegates to auto-tool-tracking.ts
-export function markToolStart(toolCallId: string): void {
-  _markToolStart(toolCallId, s.active);
+export function markToolStart(toolCallId: string, toolName?: string): void {
+  _markToolStart(toolCallId, s.active, toolName);
 }
 
 export function markToolEnd(toolCallId: string): void {
@@ -1102,6 +1104,7 @@ export async function startAuto(
     s.stepMode = requestedStepMode;
     s.cmdCtx = ctx;
     s.basePath = base;
+    setLogBasePath(base);
     s.unitDispatchCount.clear();
     s.unitLifetimeDispatches.clear();
     if (!getLedger()) initMetrics(base);