@tekmidian/pai 0.55.0 → 0.57.0

package/dist/{run-DkTrZejF.mjs → run-9HbO1gTM.mjs}

@@ -466,6 +466,22 @@ function elapsedOf(ts, now = /* @__PURE__ */ new Date()) {
 	if (mins > 0) return `${mins}m ${String(sec).padStart(2, "0")}s`;
 	return `${sec}s`;
 }
+/** Shorter sleeps are legitimate polling — not worth a supervision event. */
+const SLEEP_FLOOR_SECS = 60;
+const SLEEP_RE = /^\s*sleep\s+([0-9]+(?:\.[0-9]+)?)([smhd]?)\s*$/;
+/**
+* Seconds of a standalone `sleep N` (N plain or with an s/m/h/d suffix), or
+* null when the command is not one long sleep: anchored to the whole command,
+* so a filename containing "sleep" or an `echo sleep 500` never trips it, and
+* floored at SLEEP_FLOOR_SECS so ordinary polling stays invisible.
+*/
+function parseSleepSecs(command) {
+	const m = command.match(SLEEP_RE);
+	if (!m) return null;
+	const mult = m[2] === "m" ? 60 : m[2] === "h" ? 3600 : m[2] === "d" ? 86400 : 1;
+	const secs = Number(m[1]) * mult;
+	return Number.isFinite(secs) && secs >= SLEEP_FLOOR_SECS ? secs : null;
+}
 /** One-line description of a tool call, e.g. "Bash: npm test". */
 function describeTool(name, inp) {
 	if (typeof inp !== "object" || inp === null) return name;
@@ -2410,6 +2426,7 @@ function fullContractPrompt(format) {
 		"will pick something up later, never leave a background command as your final action. Finish, then",
 		"report.",
 		"No narration, no timestamps, no greetings, no summaries of what you read; act, verify, then stop.",
+		"Never sleep longer than 60 seconds in one command; wait on a bounded timeout or event instead.",
 		...fileNotInlineBlock(),
 		"",
 		`Operator messages: a user turn starting with ${OPERATOR_MARK} was typed by the operator while you`,
@@ -2520,6 +2537,53 @@ function renderReport(c, prefix, r, cwd = "") {
 	if (r.notes) out.push(`${prefix}${c("bold", "notes")}  ${shortText(r.notes, 120)}`);
 	return out;
 }
+/** A summary that names its own deletion — the path is expected to be gone. */
+const DELETED_RE = /\b(delete|deleted|deleting|remove|removed|removing)\b/i;
+/**
+* Check every path a report's `changed` list claims against the filesystem:
+* it must exist (mtime at or after `startMs`, the run's start) unless its
+* summary says it was deleted, in which case it must be absent. Mechanical
+* only — counts, row totals and other free-text claims are not checked (a
+* worker's R report is trusted as written; a 2026-09-23 report claimed "181
+* rows saved" while it had written 31, and only its listed paths are
+* something the filesystem can confirm or refute).
+*/
+function verifyReportChanges(changed, cwd, startMs) {
+	const failures = [];
+	for (const ch of changed ?? []) {
+		const raw = (ch.path ?? "").trim();
+		if (!raw) continue;
+		const abs = isAbsolute(raw) ? raw : resolve(cwd, raw);
+		let mtimeMs;
+		try {
+			mtimeMs = statSync(abs).mtimeMs;
+		} catch {
+			mtimeMs = null;
+		}
+		if (DELETED_RE.test(ch.summary ?? "")) {
+			if (mtimeMs !== null) failures.push({
+				path: raw,
+				reason: "not-deleted"
+			});
+		} else if (mtimeMs === null) failures.push({
+			path: raw,
+			reason: "missing"
+		});
+		else if (mtimeMs < startMs) failures.push({
+			path: raw,
+			reason: "stale"
+		});
+	}
+	return {
+		ok: failures.length === 0,
+		failures
+	};
+}
+/** The one-line note appended to a report/ledger when verify fails. */
+function verifyFailNote(v) {
+	const paths = v.failures.map((f) => f.path).slice(0, 5).join(", ");
+	return `verify: ${v.failures.length} claimed change${v.failures.length === 1 ? "" : "s"} not confirmed: ${paths}`;
+}
 /** repo-relative display of a changed path when it lies under cwd. */
 function relShort(p, cwd) {
 	if (!cwd) return p;
@@ -2848,6 +2912,27 @@ function resetContextTokensOnCompact(status, e) {
 function isCompactBoundary(e) {
 	return e.type === "system" && (e.subtype === "compact_boundary" || e.subtype === "compact");
 }
+/** The command string of a Bash tool_use input, "" when there is none. */
+function bashCommand(input) {
+	if (typeof input === "object" && input !== null && typeof input.command === "string") return input.command;
+	return "";
+}
+function toolResultText(content) {
+	if (typeof content === "string") return content;
+	if (Array.isArray(content)) return content.map((c) => typeof c === "object" && c !== null && typeof c.text === "string" ? c.text : "").join("\n");
+	return "";
+}
+/**
+* Whether a tool_result block is a failed execution: the harness's own
+* is_error flag, or — the older Bash spelling — an "Exit code N" line with a
+* non-zero N in the result text. Feeds the consecFails counter supervision's
+* thrashing detector reads.
+*/
+function toolResultFailed(block) {
+	if (block.is_error === true) return true;
+	const m = toolResultText(block.content).match(/Exit code (\d+)/);
+	return m !== null && Number(m[1]) !== 0;
+}
 /**
 * The model the init event announces, adopted into the status when the spawn
 * could not name one (a caller that passed `--model` itself in the claude
@@ -2952,7 +3037,7 @@ async function runWorker(opts) {
 	const logDir = workersLogDir(config);
 	mkdirSync(logDir, { recursive: true });
 	if (opts.className === "plan" && !opts._planner) {
-		const { runPlanner } = await import("./planner-C2wWnZrZ.mjs");
+		const { runPlanner } = await import("./planner-s1iQQsjj.mjs");
 		return runPlanner(opts);
 	}
 	const parent = launchParent(opts.parent);
@@ -3495,8 +3580,21 @@ async function executeRun(a) {
 				for (const block of e.message?.content ?? []) if (block.type === "tool_use") {
 					status.tools += 1;
 					status.last = describeTool(block.name ?? "?", block.input);
+					if (block.name === "Bash") status.sleepSec = parseSleepSecs(bashCommand(block.input));
 				} else if (block.type === "text" && (block.text ?? "").trim()) status.last = "says: " + shortText(block.text, 70);
 				saveStatus(logDir, status);
+			} else if (e.type === "user") {
+				const blocks = e.message?.content;
+				if (Array.isArray(blocks)) {
+					let sawResult = false;
+					for (const block of blocks) {
+						if (block.type !== "tool_result") continue;
+						sawResult = true;
+						status.consecFails = toolResultFailed(block) ? (status.consecFails ?? 0) + 1 : 0;
+						status.sleepSec = null;
+					}
+					if (sawResult) saveStatus(logDir, status);
+				}
 			} else if (e.type === "result") {
 				const tokens = usageContextTokens(e.usage);
 				if (e.is_compact) status.contextTokens = tokens ?? status.contextTokens;
@@ -3561,6 +3659,20 @@ async function executeRun(a) {
 			} catch {}
 		}
 	}
+	if (headless && ctx.resultReport?.changed?.length) {
+		const verify = verifyReportChanges(ctx.resultReport.changed, worktree?.dir ?? cwd, t0);
+		if (!verify.ok) {
+			status.reportValid = false;
+			const note = verifyFailNote(verify);
+			status.last = shortText(note, 90);
+			ctx.resultReport.notes = ctx.resultReport.notes ? `${ctx.resultReport.notes} — ${note}` : note;
+			appendLedger(ledger, "REPORT-UNVERIFIED", {
+				id: wid,
+				failures: verify.failures.length,
+				paths: verify.failures.map((f) => f.path).slice(0, 5).join(",")
+			});
+		}
+	}
 	saveStatus(logDir, status);
 	appendLedger(ledger, "WORKER-END", {
 		id: wid,
@@ -3833,6 +3945,20 @@ async function executeCodexRun(a) {
 	status.rc = rc;
 	status.secs = secs;
 	status.last = shortText(report?.notes ?? finalText, 90) || (ok ? "done" : "failed");
+	if (report?.changed?.length) {
+		const verify = verifyReportChanges(report.changed, worktree?.dir ?? cwd, t0);
+		if (!verify.ok) {
+			status.reportValid = false;
+			const note = verifyFailNote(verify);
+			status.last = shortText(note, 90);
+			report.notes = report.notes ? `${report.notes} — ${note}` : note;
+			appendLedger(ledger, "REPORT-UNVERIFIED", {
+				id: wid,
+				failures: verify.failures.length,
+				paths: verify.failures.map((f) => f.path).slice(0, 5).join(",")
+			});
+		}
+	}
 	saveStatus(logDir, status);
 	appendLedger(ledger, "WORKER-END", {
 		id: wid,
@@ -4000,4 +4126,4 @@ function resultFromOutput(out) {
 
 //#endregion
 export { contextLabel as A, saveStatus as B, mergeWorker as C, UNLABELED as D, ledgerSummary as E, loadStatus as F, shortText as G, waitForTerminalStatus as H, loadStatuses as I, describeMcp as K, newWorkerId as L, elapsedOf as M, isChatPane as N, ageOf as O, isLive as P, nowStamp as R, discardWorker as S, appendLedger as T, longInlinePromptHint as U, setWorkerLabel as V, parseRunnerArgs as W, itermUuid as _, projectLaunchConfig as a, sessionTag as b, handoffFromInside as c, workerDepth as d, checkPaneForWorker as f, itermForClaudeSession as g, currentTabKey as h, usageContextTokens as i, contextPercent as j, alive as k, readInbox as l, openPaneForWorker as m, runWorker as n, parseWorkerReport as o, openFollowPane as p, expandMcpNames as q, testProvider as r, renderReport as s, printResult as t, sayToWorker as u, resolveSession as v, worktreesDir as w, workerInScope as x, sessionMapPath as y, ownsPid as z };
-//# sourceMappingURL=run-DkTrZejF.mjs.map
+//# sourceMappingURL=run-9HbO1gTM.mjs.map