@danypops/papyrus 0.27.10 → 0.27.12

package/extension/src/discuss-ask-view.ts

@@ -63,19 +63,10 @@ export interface AskQuestionParams {
 	displayMode?: AskDisplayMode;
 	timeout?: number;
 	/**
-	 * Joins a second concurrent call for the same key to the first's in-flight promise instead of
-	 * opening a second picker. Pass a stable id (the target Discussion's id) whenever the caller
-	 * cannot otherwise guarantee only one live ask is ever issued for that same question.
-	 */
-	key?: string;
-	/**
-	 * Streamed once before blocking on the human. A live ask can legitimately sit pending far
-	 * longer than a typical tool call (real human response time, not milliseconds) -- without any
-	 * progress signal, a tool call sitting silent that long looks indistinguishable from a dead
-	 * one to anything upstream watching for stalled calls. pi-ask-user's own original code (the
-	 * prior art this view is adapted from) sent exactly this same heartbeat before presenting its
-	 * UI; dropping it during the port was the regression that let two independent executions of
-	 * the same live ask run concurrently, each opening its own picker for the same question.
+	 * Streamed once before blocking on the human, matching pi-ask-user's own original code (the
+	 * prior art this view is adapted from) -- gives the tool call's progress UI something to show
+	 * during a wait that legitimately runs far longer than a typical tool call (real human response
+	 * time, not milliseconds).
 	 */
 	onUpdate?: AgentToolUpdateCallback;
 	/**
@@ -1099,9 +1090,8 @@ async function askViaDialogs(
  * driver (extension/src/index.ts's driveActiveTasks, on agent_settled) queue a "continue the
  * active task" nudge as a `deliverAs: "nextTurn"` message while this exact live ask is still
  * awaiting an answer -- starting a second, concurrent turn that reasons about the very Discussion
- * this call is already resolving, independently of it. A live-observed bug (two pickers for the
- * same question, one orphaned and later auto-resolving with fabricated "defer" text) traced back
- * to exactly this race. driveActiveTasks checks isLiveAskPending() and skips queuing while true.
+ * this call is already resolving, independently of it. driveActiveTasks checks isLiveAskPending()
+ * and skips queuing while true.
  */
 let livePendingCount = 0;
 
@@ -1109,16 +1099,6 @@ export function isLiveAskPending(): boolean {
 	return livePendingCount > 0;
 }
 
-/**
- * Keyed reentrancy join: whatever the exact external cause (an upstream retry, a duplicate turn,
- * anything outside code this package controls -- verified Pi's own ctx.ui.custom() is a clean,
- * single-shot, well-guarded call with no retry/timeout logic of its own), a second concurrent
- * askQuestion() call for the SAME key must never open a second picker for the same question. It
- * joins the already-in-flight promise instead. Keyed by the target Discussion's id (stable across
- * a genuine duplicate call, unlike a fresh toolCallId each retry might mint).
- */
-const pendingByKey = new Map<string, Promise<AskAnswer | undefined>>();
-
 /**
  * Discuss's live:true synchronous ask -- interactive AskComponent when a real TUI is available,
  * dialog fallback (ctx.ui.select/input) in RPC/headless mode, no-op undefined without any
@@ -1127,36 +1107,7 @@ const pendingByKey = new Map<string, Promise<AskAnswer | undefined>>();
  */
 export async function askQuestion(ctx: ExtensionContext, params: AskQuestionParams): Promise<AskAnswer | undefined> {
 	if (!ctx.hasUI || !ctx.ui) return undefined;
-	if (params.key !== undefined) {
-		const existing = pendingByKey.get(params.key);
-		if (existing) return existing;
-	}
-	const promise = askQuestionUnguarded(ctx, params);
-	if (params.key !== undefined) {
-		const key = params.key;
-		pendingByKey.set(key, promise);
-		void promise.finally(() => {
-			if (pendingByKey.get(key) === promise) pendingByKey.delete(key);
-		});
-	}
-	return promise;
-}
-
-/**
- * Sent at least this often while blocked on the human -- well under any 8-second window a
- * watchdog upstream might require to consider a tool call still alive. A single upfront ping
- * (this view's original heartbeat) only covers the first ~8s; for genuinely long human response
- * times (the whole point of a live ask) it goes silent again after that, indistinguishable from
- * dead. A live-observed bug traced back to exactly this: cancellation always cited "idle timeout,
- * no interaction within 8000ms" regardless of whether the real wait was 5 seconds or 18 minutes
- * -- a fixed re-check window, not a measure of total elapsed time -- meaning periodic liveness,
- * not a one-time ping, is what's required here.
- */
-let liveAskHeartbeatIntervalMs = 4_000;
-
-/** Test seam: real tiny interval, not a faked global timer -- exercises the periodic heartbeat deterministically and fast. */
-export function setLiveAskHeartbeatIntervalMsForTests(ms: number): void {
-	liveAskHeartbeatIntervalMs = ms;
+	return askQuestionUnguarded(ctx, params);
 }
 
 async function askQuestionUnguarded(ctx: ExtensionContext, params: AskQuestionParams): Promise<AskAnswer | undefined> {
@@ -1169,15 +1120,12 @@ async function askQuestionUnguarded(ctx: ExtensionContext, params: AskQuestionPa
 	const displayMode: AskDisplayMode = params.displayMode ?? envDisplayMode ?? "overlay";
 	const normalizedContext = params.context?.trim() || undefined;
 
-	const sendHeartbeat = () => params.onUpdate?.({ content: [{ type: "text", text: "Waiting for human input..." }], details: undefined });
-	sendHeartbeat();
-	const heartbeatTimer = params.onUpdate ? setInterval(sendHeartbeat, liveAskHeartbeatIntervalMs) : undefined;
+	params.onUpdate?.({ content: [{ type: "text", text: "Waiting for human input..." }], details: undefined });
 	livePendingCount += 1;
 	try {
 		return await askQuestionBlocking(ctx, params, options, allowMultiple, allowFreeform, allowComment, displayMode, normalizedContext);
 	} finally {
 		livePendingCount -= 1;
-		if (heartbeatTimer !== undefined) clearInterval(heartbeatTimer);
 	}
 }
 

package/extension/src/discuss.ts

@@ -84,8 +84,8 @@ export async function showDiscussions(ctx: ExtensionCommandContext): Promise<voi
 				const pending = (() => { try { return readDiscussionExtra(discussion.extra); } catch { return undefined; } })();
 				const question = `Reply to "${discussion.title}":`;
 				const answer = pending?.pendingOptions && pending.pendingOptions.length > 0 && pending.pendingOptionsMode
-					? await askQuestion(commandCtx, { question, options: pending.pendingOptions.map((title) => ({ title })), allowMultiple: pending.pendingOptionsMode === "multi", key: discussion.id })
-					: await askQuestion(commandCtx, { question, key: discussion.id });
+					? await askQuestion(commandCtx, { question, options: pending.pendingOptions.map((title) => ({ title })), allowMultiple: pending.pendingOptionsMode === "multi" })
+					: await askQuestion(commandCtx, { question });
 				if (!answer) return; // canceled
 				await callService("discuss.reply", { id: discussion.id, actor: ACTOR, content: answer.content, ...(answer.selected ? { selected: answer.selected } : {}), source: SOURCE });
 				commandCtx.ui.notify(answer.selected ? `Selected: ${answer.selected.join(", ")}` : "Reply added.", "info");

package/extension/src/domain-tools.ts

@@ -50,10 +50,9 @@ async function liveAnswer(ctx: ExtensionContext, discussion: Artifact, onUpdate:
 			allowMultiple: pending.pendingOptionsMode === "multi",
 			onUpdate,
 			signal,
-			key: discussion.id,
 		});
 	}
-	return askQuestion(ctx, { question, onUpdate, signal, key: discussion.id });
+	return askQuestion(ctx, { question, onUpdate, signal });
 }
 
 /**

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@danypops/papyrus",
-  "version": "0.27.10",
+  "version": "0.27.12",
   "description": "Daemon-backed graph artifacts, evidence-bearing tasks, rules, skills, and native TUI workflows for Pi",
   "type": "module",
   "keywords": ["pi-package"],

package/src/ops.ts

@@ -529,14 +529,17 @@ export function runGates(db: Db, artifactId: string, options: GateRunOptions = {
 				}
 			}
 			case "command": {
-				const { execSync } = require_("node:child_process");
-				try {
-					const output = execSync(gate.target, { encoding: "utf-8", timeout: GATE_COMMAND_TIMEOUT_MS, stdio: ["pipe", "pipe", "pipe"], ...(cwd ? { cwd } : {}) }).trim();
-					const passed = gate.expect ? output.includes(gate.expect) : true;
-					return { gate, passed, output: output.slice(0, GATE_OUTPUT_LIMIT) };
-				} catch (e) {
-					return { gate, passed: false, output: e instanceof Error ? e.message.slice(0, GATE_OUTPUT_LIMIT) : "command failed" };
-				}
+				// spawnSync + manual stdout/stderr concatenation, not execSync: execSync's return value is
+				// stdout only. Many real commands (bun test's own per-test lines and its pass/fail summary
+				// among them) write their actual output to stderr, so an execSync-based match against
+				// gate.expect saw only the first line of a banner and never the result -- every such gate
+				// failed regardless of whether the command actually passed.
+				const { spawnSync } = require_("node:child_process");
+				const result = spawnSync(gate.target, { shell: true, encoding: "utf-8", timeout: GATE_COMMAND_TIMEOUT_MS, ...(cwd ? { cwd } : {}) });
+				if (result.error) return { gate, passed: false, output: result.error.message.slice(0, GATE_OUTPUT_LIMIT) };
+				const combined = `${result.stdout ?? ""}${result.stderr ?? ""}`.trim();
+				const passed = result.status === 0 && (gate.expect ? combined.includes(gate.expect) : true);
+				return { gate, passed, output: combined.slice(0, GATE_OUTPUT_LIMIT) || (result.status === 0 ? "ok" : `command exited with code ${result.status}`) };
 			}
 			case "test": {
 				const { execSync } = require_("node:child_process");
@@ -564,7 +567,7 @@ export function runGates(db: Db, artifactId: string, options: GateRunOptions = {
  *      indefinitely after Papyrus considers the gate "timed out". Spawning detached (its own
  *      process group) and killing the negated pid on our own timer reaches the whole tree.
  */
-function executeGateCommand(command: string, timeout: number, cwd?: string): Promise<{ passed: boolean; output: string }> {
+function executeGateCommand(command: string, timeout: number, cwd?: string): Promise<{ passed: boolean; output: string; matchable: string }> {
 	// `spawn(..., { shell: true, detached: true })` instead of the `exec()` convenience wrapper:
 	// `detached` (needed to make the shell the leader of its own process group, so the negated pid
 	// below reaches every descendant, not just the shell) is not part of Node's `exec()`/
@@ -587,17 +590,21 @@ function executeGateCommand(command: string, timeout: number, cwd?: string): Pro
 		child.stdout?.on("data", append);
 		child.stderr?.on("data", append);
 
-		const finish = (result: { passed: boolean; output: string }): void => {
+		const finish = (result: { passed: boolean; output: string; matchable: string }): void => {
 			if (settled) return;
 			settled = true;
 			clearTimeout(timer);
 			resolve(result);
 		};
 
-		child.on("error", (error) => finish({ passed: false, output: error.message.slice(0, GATE_OUTPUT_LIMIT) }));
+		child.on("error", (error) => finish({ passed: false, output: error.message.slice(0, GATE_OUTPUT_LIMIT), matchable: error.message }));
 		child.on("close", (code) => {
-			const output = buffered.trim().slice(0, GATE_OUTPUT_LIMIT);
-			finish({ passed: code === 0, output: output || (code === 0 ? "ok" : `command exited with code ${code}`) });
+			// `matchable` carries the full (GATE_MAX_BUFFER_BYTES-bounded) buffer so the caller's
+			// gate.expect substring check sees the whole run, not just the first GATE_OUTPUT_LIMIT
+			// characters -- a real bun test run's pass/fail summary is its last line, not its first.
+			const full = buffered.trim();
+			const output = full.slice(0, GATE_OUTPUT_LIMIT);
+			finish({ passed: code === 0, output: output || (code === 0 ? "ok" : `command exited with code ${code}`), matchable: full });
 		});
 
 		const timer = setTimeout(() => {
@@ -609,7 +616,7 @@ function executeGateCommand(command: string, timeout: number, cwd?: string): Pro
 					child.kill("SIGKILL");
 				}
 			}
-			finish({ passed: false, output: `gate command timed out after ${timeout}ms` });
+			finish({ passed: false, output: `gate command timed out after ${timeout}ms`, matchable: "" });
 		}, timeout);
 	});
 }
@@ -651,7 +658,7 @@ export async function runGatesAsync(db: Db, artifactId: string, options: GateRun
 			const executed = await executeGateCommand(command, timeout, options.cwd);
 			results.push({
 				gate,
-				passed: executed.passed && (gate.expect ? executed.output.includes(gate.expect) : true),
+				passed: executed.passed && (gate.expect ? executed.matchable.includes(gate.expect) : true),
 				output: executed.output,
 			});
 		} else {