@gajae-code/coding-agent 0.7.11 → 0.8.0

package/src/gjc-runtime/session-state-sidecar.ts

@@ -1,13 +1,18 @@
+import * as fsSync from "node:fs";
 import * as fs from "node:fs/promises";
 import * as path from "node:path";
 import type { AssistantMessage } from "@gajae-code/ai";
-import { logger } from "@gajae-code/utils";
+import { logger, postmortem } from "@gajae-code/utils";
 
 export const GJC_COORDINATOR_SESSION_STATE_FILE_ENV = "GJC_COORDINATOR_SESSION_STATE_FILE";
 export const GJC_COORDINATOR_SESSION_ID_ENV = "GJC_COORDINATOR_SESSION_ID";
+export const GJC_COORDINATOR_SESSION_BRANCH_ENV = "GJC_COORDINATOR_SESSION_BRANCH";
 
 export type RuntimeState = "ready_for_input" | "running" | "needs_user_input" | "completed" | "errored";
 
+type FinalResponseSource = "agent_end" | "launch_error";
+const MAX_PUBLIC_ERROR_MESSAGE_LENGTH = 2000;
+
 interface RuntimeStateEvent {
 	type: string;
 	messages?: unknown[];
@@ -17,6 +22,7 @@ interface RuntimeStateContext {
 	sessionId: string;
 	cwd: string;
 	sessionFile?: string | null;
+	branch?: string | null;
 }
 
 interface RuntimeStateSidecarPayload {
@@ -26,6 +32,7 @@ interface RuntimeStateSidecarPayload {
 	ready_for_input?: unknown;
 	cwd?: unknown;
 	session_file?: unknown;
+	final_response?: { source?: unknown };
 }
 
 export type TerminalRuntimeStateStatus =
@@ -103,7 +110,7 @@ function assistantText(assistant: AssistantMessage | undefined): string | null {
 function finalResponseForEvent(event: RuntimeStateEvent): {
 	text: string | null;
 	format: "markdown";
-	source: "agent_end";
+	source: FinalResponseSource;
 	artifact_path: null;
 	truncated: false;
 } | null {
@@ -127,6 +134,118 @@ function stateForEvent(event: RuntimeStateEvent): RuntimeState | null {
 	return null;
 }
 
+function readPreviousPayload(stateFile: string): Record<string, unknown> {
+	try {
+		return JSON.parse(fsSync.readFileSync(stateFile, "utf8")) as Record<string, unknown>;
+	} catch {
+		return {};
+	}
+}
+
+function shouldPreserveTerminalPayload(previous: RuntimeStateSidecarPayload): boolean {
+	if (previous.state !== "completed" && previous.state !== "errored") return false;
+	const source = previous.final_response?.source;
+	return source === "agent_end" || source === "launch_error";
+}
+
+function branchForContext(context: RuntimeStateContext): string | null {
+	return context.branch ?? (process.env[GJC_COORDINATOR_SESSION_BRANCH_ENV]?.trim() || null);
+}
+
+function basePayload(input: {
+	context: RuntimeStateContext;
+	previous: Record<string, unknown>;
+	state: RuntimeState;
+	now: string;
+	source: string;
+	event: string;
+	reason: string | null;
+}): Record<string, unknown> {
+	return {
+		schema_version: 1,
+		session_id: process.env[GJC_COORDINATOR_SESSION_ID_ENV]?.trim() || input.context.sessionId,
+		state: input.state,
+		ready_for_input: input.state === "completed" || input.state === "ready_for_input",
+		updated_at: input.now,
+		current_turn_id: typeof input.previous.current_turn_id === "string" ? input.previous.current_turn_id : null,
+		last_turn_id: typeof input.previous.last_turn_id === "string" ? input.previous.last_turn_id : null,
+		live: input.state === "running",
+		reason: input.reason,
+		source: input.source,
+		event: input.event,
+		cwd: input.context.cwd,
+		workdir: input.context.cwd,
+		branch: branchForContext(input.context),
+		session_file: input.context.sessionFile ?? null,
+	};
+}
+function publicSafeErrorMessage(message: string): string {
+	const normalized = message.replace(/[\u0000-\u0008\u000b\u000c\u000e-\u001f\u007f]/g, " ").trim();
+	if (normalized.length <= MAX_PUBLIC_ERROR_MESSAGE_LENGTH) return normalized;
+	return `${normalized.slice(0, MAX_PUBLIC_ERROR_MESSAGE_LENGTH)}…`;
+}
+
+function errorMessageForPostmortem(reason: postmortem.Reason): string {
+	return publicSafeErrorMessage(`GJC process cleanup ran for ${reason}`);
+}
+
+function numericProcessExitCode(defaultCode: number | null): number | null {
+	return typeof process.exitCode === "number" ? process.exitCode : defaultCode;
+}
+
+function postmortemExitDetails(reason: postmortem.Reason): {
+	state: RuntimeState;
+	reason: string;
+	exitKind: string;
+	exitCode: number | null;
+	signal: string | null;
+	error?: { code: string; message: string; recoverable: true };
+} {
+	if (reason === postmortem.Reason.EXIT || reason === postmortem.Reason.MANUAL) {
+		const exitCode = numericProcessExitCode(0) ?? 0;
+		const state: RuntimeState = exitCode === 0 ? "completed" : "errored";
+		return {
+			state,
+			reason: reason === postmortem.Reason.EXIT ? "process_exit" : "manual_cleanup",
+			exitKind: reason,
+			exitCode,
+			signal: null,
+			...(state === "errored"
+				? {
+						error: {
+							code: "process_exit",
+							message: publicSafeErrorMessage(`GJC process exited with code ${exitCode}`),
+							recoverable: true,
+						},
+					}
+				: {}),
+		};
+	}
+	const signalByReason: Partial<Record<postmortem.Reason, string>> = {
+		[postmortem.Reason.SIGINT]: "SIGINT",
+		[postmortem.Reason.SIGTERM]: "SIGTERM",
+		[postmortem.Reason.SIGHUP]: "SIGHUP",
+	};
+	return {
+		state: "errored",
+		reason,
+		exitKind: reason,
+		exitCode: numericProcessExitCode(null),
+		signal: signalByReason[reason] ?? null,
+		error: { code: reason, message: errorMessageForPostmortem(reason), recoverable: true },
+	};
+}
+
+function writeStateFileSync(stateFile: string, payload: Record<string, unknown>): void {
+	fsSync.mkdirSync(path.dirname(stateFile), { recursive: true });
+	fsSync.writeFileSync(stateFile, `${JSON.stringify(payload, null, 2)}\n`);
+}
+
+async function writeStateFile(stateFile: string, payload: Record<string, unknown>): Promise<void> {
+	await fs.mkdir(path.dirname(stateFile), { recursive: true });
+	await Bun.write(stateFile, `${JSON.stringify(payload, null, 2)}\n`);
+}
+
 export async function persistCoordinatorRuntimeStateFromEvent(
 	event: RuntimeStateEvent,
 	context: RuntimeStateContext,
@@ -136,42 +255,66 @@ export async function persistCoordinatorRuntimeStateFromEvent(
 	const state = stateForEvent(event);
 	if (!state) return;
 	const now = new Date().toISOString();
-	let previous: Record<string, unknown> = {};
-	try {
-		previous = JSON.parse(await Bun.file(stateFile).text()) as Record<string, unknown>;
-	} catch {
-		previous = {};
-	}
+	const previous = readPreviousPayload(stateFile);
 	const finalResponse = finalResponseForEvent(event);
 	const payload = {
-		schema_version: 1,
-		session_id: process.env[GJC_COORDINATOR_SESSION_ID_ENV]?.trim() || context.sessionId,
-		state,
-		ready_for_input: state === "completed" || state === "ready_for_input",
-		updated_at: now,
-		current_turn_id: typeof previous.current_turn_id === "string" ? previous.current_turn_id : null,
-		last_turn_id: typeof previous.last_turn_id === "string" ? previous.last_turn_id : null,
-		live: typeof previous.live === "boolean" ? previous.live : null,
-		reason: null,
-		source: "agent_session_event",
-		event: event.type,
-		cwd: context.cwd,
-		session_file: context.sessionFile ?? null,
+		...basePayload({ context, previous, state, now, source: "agent_session_event", event: event.type, reason: null }),
+		...(state === "completed" || state === "errored" ? { ended_at: now } : {}),
 		...(finalResponse ? { final_response: finalResponse } : {}),
 		...(state === "errored"
 			? {
 					error: {
 						code: "agent_error",
-						message: lastAssistant(event.messages)?.errorMessage ?? "agent_error",
+						message: publicSafeErrorMessage(lastAssistant(event.messages)?.errorMessage ?? "agent_error"),
 						recoverable: true,
 					},
 				}
 			: {}),
 	};
 	try {
-		await fs.mkdir(path.dirname(stateFile), { recursive: true });
-		await Bun.write(stateFile, `${JSON.stringify(payload, null, 2)}\n`);
+		await writeStateFile(stateFile, payload);
 	} catch (error) {
 		logger.warn("Failed to persist coordinator runtime state", { error: String(error), stateFile });
 	}
 }
+
+export function persistCoordinatorRuntimeStateFromPostmortem(
+	reason: postmortem.Reason,
+	context: RuntimeStateContext,
+): void {
+	const stateFile = process.env[GJC_COORDINATOR_SESSION_STATE_FILE_ENV]?.trim();
+	if (!stateFile) return;
+	const previous = readPreviousPayload(stateFile);
+	if (shouldPreserveTerminalPayload(previous as RuntimeStateSidecarPayload)) return;
+	const now = new Date().toISOString();
+	const details = postmortemExitDetails(reason);
+	const payload = {
+		...basePayload({
+			context,
+			previous,
+			state: details.state,
+			now,
+			source: "process_postmortem",
+			event: "process_exit",
+			reason: details.reason,
+		}),
+		ended_at: now,
+		detected_at: now,
+		exit_kind: details.exitKind,
+		exit_code: details.exitCode,
+		signal: details.signal,
+		...(details.error ? { error: details.error } : {}),
+	};
+	try {
+		writeStateFileSync(stateFile, payload);
+	} catch (error) {
+		logger.warn("Failed to persist coordinator runtime state during postmortem", { error: String(error), stateFile });
+	}
+}
+
+export function registerCoordinatorRuntimeStateFinalizer(context: RuntimeStateContext): () => void {
+	if (!process.env[GJC_COORDINATOR_SESSION_STATE_FILE_ENV]?.trim()) return () => {};
+	return postmortem.register("coordinator-runtime-state", reason => {
+		persistCoordinatorRuntimeStateFromPostmortem(reason, context);
+	});
+}

package/src/gjc-runtime/team-runtime.ts

@@ -1931,16 +1931,14 @@ export function buildWorkerCommand(
 	const workspace = worker.worktree_path
 		? `Worker worktree: ${worker.worktree_path}.`
 		: `Worker cwd: ${config.leader.cwd}.`;
-	// The worker prompt body is dispatched into the pane through
-	// `tmux send-keys`, which treats embedded LF characters as Enter
-	// keypresses. A multi-line prompt therefore lands in the pane as
-	// multiple prompt bodies followed by premature Enter presses, which on
-	// Windows + psmux/ConPTY causes the worker CLI to bail out before the
-	// startup ACK fires. Normalize the body to a single line by replacing
-	// any LF/CRLF with a space, and strip a defensive U+FEFF in case a
-	// caller managed to inject a UTF-8 BOM into the task text. Empty /
-	// whitespace-only bodies fall back to a one-line placeholder so the
-	// worker never sits idle at an empty prompt.
+	// The worker prompt body must stay single-line because fallback startup on
+	// Windows/psmux dispatches the command through `tmux send-keys`, where
+	// embedded LF characters are treated as Enter keypresses. POSIX tmux starts
+	// workers by passing the command directly to `split-window`, but keeping the
+	// body normalized there too makes fake-tmux logs and shell argv handling
+	// deterministic. Strip a defensive U+FEFF in case a caller managed to inject
+	// a UTF-8 BOM into the task text. Empty / whitespace-only bodies fall back to
+	// a one-line placeholder so the worker never sits idle at an empty prompt.
 	const normalizePrompt = (raw: string): string =>
 		raw
 			.replace(/[\uFEFF\u200B]/g, "")
@@ -1995,6 +1993,11 @@ export function buildWorkerCommand(
 	}
 	return `${joined} ${config.worker_command} ${quote(prompt)}`;
 }
+
+function shouldDispatchWorkerWithSendKeys(tmuxCommand: string, platform: NodeJS.Platform = process.platform): boolean {
+	return platform === "win32" || path.basename(tmuxCommand).toLowerCase() === "psmux";
+}
+
 interface GjcTeamInitialLane {
 	label: string;
 	title: string;
@@ -2098,22 +2101,24 @@ async function startTmuxSession(
 			const splitDirection: string = worker.index === 1 ? "-h" : "-v";
 			const splitTarget: string =
 				worker.index === 1 ? config.leader.pane_id : (rightStackRootPaneId ?? config.leader.pane_id);
-			const split: Bun.SyncSubprocess<"pipe", "pipe"> = Bun.spawnSync(
-				[
-					config.tmux_command,
-					"split-window",
-					splitDirection,
-					"-t",
-					splitTarget,
-					"-d",
-					"-P",
-					"-F",
-					"#{pane_id}",
-					"-c",
-					worker.worktree_path ?? config.leader.cwd,
-				],
-				{ stdout: "pipe", stderr: "pipe" },
-			);
+			const workerCommand = buildWorkerCommand(config, worker);
+			const workerCwd = worker.worktree_path ?? config.leader.cwd;
+			const useSendKeysFallback = shouldDispatchWorkerWithSendKeys(config.tmux_command);
+			const splitArgs = [
+				config.tmux_command,
+				"split-window",
+				splitDirection,
+				"-t",
+				splitTarget,
+				"-d",
+				"-P",
+				"-F",
+				"#{pane_id}",
+				"-c",
+				workerCwd,
+				...(useSendKeysFallback ? [] : [workerCommand]),
+			];
+			const split: Bun.SyncSubprocess<"pipe", "pipe"> = Bun.spawnSync(splitArgs, { stdout: "pipe", stderr: "pipe" });
 			if (split.exitCode !== 0)
 				throw new Error(split.stderr.toString().trim() || `tmux_split_failed:${config.tmux_target}:${worker.id}`);
 			const paneId: string = split.stdout.toString().trim().split(/\r?\n/)[0]?.trim() ?? "";
@@ -2121,31 +2126,31 @@ async function startTmuxSession(
 			rollbackPaneIds.push(paneId);
 			if (worker.index === 1) rightStackRootPaneId = paneId;
 			workers.push({ ...worker, pane_id: paneId });
-			// On psmux/ConPTY (Windows pwsh default-shell) panes, tmux writes the
-			// split-window command argv to the new pane's stdin but pwsh's
-			// readline never receives an Enter, so the worker never starts.
-			// Create the pane without a command, then dispatch the worker
-			// command through send-keys + Enter so it actually executes.
-			// Two-step dispatch because tmux's `-l` (literal mode) flag is
-			// global per send-keys invocation, not per arg: passing `-l
-			// <body> Enter` would treat "Enter" as literal text too. Sending
-			// the body in literal mode first and the Enter keypress second
-			// keeps the body verbatim while still submitting the prompt as a
-			// keystroke.
-			const body = buildWorkerCommand(config, worker);
-			Bun.spawnSync([config.tmux_command, "send-keys", "-l", "-t", paneId, body], {
-				stdout: "ignore",
-				stderr: "ignore",
-			});
-			const sendKeys = Bun.spawnSync([config.tmux_command, "send-keys", "-t", paneId, "Enter"], {
-				stdout: "ignore",
-				stderr: "ignore",
-			});
-			// void-cast the exit code so the linter does not flag an unused
-			// expression; the value is intentionally discarded here because
-			// the actual spawn outcome is recovered by the leader through
-			// the worker startup-ack watcher, not via the spawn exit code.
-			void sendKeys.exitCode;
+			if (useSendKeysFallback) {
+				// On psmux/ConPTY (Windows pwsh default-shell) panes, tmux writes the
+				// split-window command argv to the new pane's stdin but pwsh's readline
+				// never receives an Enter, so the worker never starts. Create the pane
+				// without a command, then dispatch the worker command through send-keys +
+				// Enter so it actually executes.
+				// Two-step dispatch because tmux's `-l` (literal mode) flag is global per
+				// send-keys invocation, not per arg: passing `-l <body> Enter` would treat
+				// "Enter" as literal text too. Sending the body in literal mode first and
+				// the Enter keypress second keeps the body verbatim while still submitting
+				// the prompt as a keystroke.
+				Bun.spawnSync([config.tmux_command, "send-keys", "-l", "-t", paneId, workerCommand], {
+					stdout: "ignore",
+					stderr: "ignore",
+				});
+				const sendKeys = Bun.spawnSync([config.tmux_command, "send-keys", "-t", paneId, "Enter"], {
+					stdout: "ignore",
+					stderr: "ignore",
+				});
+				// void-cast the exit code so the linter does not flag an unused expression;
+				// the value is intentionally discarded here because the actual spawn outcome
+				// is recovered by the leader through the worker startup-ack watcher, not via
+				// the spawn exit code.
+				void sendKeys.exitCode;
+			}
 		}
 		Bun.spawnSync([config.tmux_command, "select-layout", "-t", config.tmux_target, "main-vertical"], {
 			stdout: "ignore",