@gajae-code/coding-agent 0.11.9 → 0.11.11

package/src/task/render.ts

@@ -28,6 +28,12 @@ import {
 	type SubmitReviewDetails,
 } from "../tools/review";
 import { Ellipsis, Hasher, type RenderCache, renderStatusLine } from "../tui";
+import {
+	collectProviderDegradationGroups,
+	hasActiveProviderRetryInTaskDetails,
+	providerProgressAgeLabel,
+	providerRetryPhaseLabel,
+} from "./provider-retry-status";
 import type { TaskResultReceipt } from "./receipt";
 import { subprocessToolRegistry } from "./subprocess-tool-registry";
 import type { AgentProgress, TaskParams, TaskToolDetails } from "./types";
@@ -123,6 +129,15 @@ function formatModelSubstitutionWarning(warning: { requested: string; effective:
 	return `Requested model substituted: ${warning.requested} -> ${warning.effective}`;
 }
 
+function renderProviderDegradationNotices(progress: readonly AgentProgress[], theme: Theme): string[] {
+	return collectProviderDegradationGroups(progress).map(group =>
+		theme.fg(
+			"warning",
+			truncateToWidth(replaceTabs(`provider degraded: ${group.count} subagents retrying on ${group.provider}`), 100),
+		),
+	);
+}
+
 function formatJsonScalar(value: unknown, _theme: Theme): string {
 	if (value === null) return "null";
 	if (typeof value === "string") {
@@ -548,14 +563,11 @@ function renderAgentProgress(
 	const titlePart = description ? `${theme.bold(displayId)}: ${description}` : displayId;
 	let statusLine = `${prefix} ${theme.fg(iconColor, icon)} ${theme.fg("accent", titlePart)}`;
 
-	// Show retry-blocked badge so the parent immediately sees that a child
-	// is sleeping on a provider 429, not silently progressing. Wins over the
-	// generic running spinner because "we're waiting on a quota window" is
-	// the operationally meaningful state.
+	// A provider recovery loop is operationally distinct from normal agent work.
 	if (progress.retryState && progress.status === "running") {
-		statusLine += ` ${formatBadge("retrying", "warning", theme)}`;
+		statusLine += ` ${formatBadge("provider degraded", "warning", theme)}`;
 	} else if (progress.retryFailure && (progress.status === "failed" || progress.status === "aborted")) {
-		statusLine += ` ${formatBadge("rate-limited", "error", theme)}`;
+		statusLine += ` ${formatBadge("provider failed", "error", theme)}`;
 	} else if (progress.status === "failed" || progress.status === "aborted") {
 		const statusLabel = progress.status === "failed" ? "failed" : "aborted";
 		statusLine += ` ${formatBadge(statusLabel, iconColor, theme)}`;
@@ -615,8 +627,10 @@ function renderAgentProgress(
 	// keep spinning while a child sleeps on a 3-hour provider rate-limit.
 	if (progress.retryState && progress.status === "running") {
 		const attemptLabel = progress.retryState.unbounded
-			? `attempt ${progress.retryState.attempt}`
-			: `${progress.retryState.attempt}/${progress.retryState.maxAttempts}`;
+			? `attempt ${progress.retryState.attempt}, unbounded`
+			: `attempt ${progress.retryState.attempt} of ${progress.retryState.maxAttempts}, bounded`;
+		const phaseLabel = providerRetryPhaseLabel(progress.retryState.kind);
+		const progressAge = staticTime ? "" : ` · ${providerProgressAgeLabel(progress.retryState, Date.now())}`;
 		// `staticTime` omits the wall-clock countdown so a cached await body stays a
 		// pure function of its key (the producer already drops time-only churn).
 		let waitLabel = "";
@@ -624,7 +638,7 @@ function renderAgentProgress(
 			const remainingMs = Math.max(0, progress.retryState.startedAtMs + progress.retryState.delayMs - Date.now());
 			waitLabel = remainingMs > 0 ? ` in ${formatDuration(remainingMs)}` : " now";
 		}
-		const summary = `retrying ${attemptLabel}${waitLabel}: ${truncateToWidth(replaceTabs(progress.retryState.errorMessage), 60)}`;
+		const summary = `${phaseLabel} · retrying ${attemptLabel}${waitLabel}${progressAge}: ${truncateToWidth(replaceTabs(progress.retryState.errorMessage), 60)}`;
 		lines.push(`${continuePrefix}${theme.tree.hook} ${theme.fg("warning", summary)}`);
 	} else if (progress.retryFailure && progress.status !== "running") {
 		const summary = `auto-retry gave up after ${progress.retryFailure.attempt} attempt${
@@ -953,23 +967,28 @@ export function renderResult(
 				.u32(spinnerFrame ?? 0)
 				.u32(width)
 				.digest();
-			if (cached?.key === key) return cached.lines;
+			const hasDynamicRetry = hasActiveProviderRetryInTaskDetails(details);
+			if (!hasDynamicRetry && cached?.key === key) return cached.lines;
 
 			const lines: string[] = [];
 
-			const shouldRenderProgress =
-				Boolean(details.progress && details.progress.length > 0) && (isPartial || details.results.length === 0);
+			const hasProgress = Boolean(details.progress && details.progress.length > 0);
+			const shouldRenderProgress = hasProgress && (isPartial || details.results.length === 0);
+			const hasResults = details.results.length > 0;
+			if (hasResults) {
+				details.results.forEach((res, i) => {
+					const isLast = !shouldRenderProgress && i === details.results.length - 1;
+					lines.push(...renderAgentResult(res, isLast, expanded, theme));
+				});
+			}
 			if (shouldRenderProgress && details.progress) {
+				lines.push(...renderProviderDegradationNotices(details.progress, theme));
 				details.progress.forEach((progress, i) => {
 					const isLast = i === details.progress!.length - 1;
 					lines.push(...renderAgentProgress(progress, isLast, expanded, theme, spinnerFrame));
 				});
-			} else if (details.results && details.results.length > 0) {
-				details.results.forEach((res, i) => {
-					const isLast = i === details.results.length - 1;
-					lines.push(...renderAgentResult(res, isLast, expanded, theme));
-				});
-
+			}
+			if (hasResults && !shouldRenderProgress) {
 				const abortedCount = details.results.filter(r => r.status === "aborted").length;
 				const mergeFailedCount = details.results.filter(r => r.status === "merge_failed").length;
 				const successCount = details.results.filter(r => r.status === "completed").length;
@@ -997,7 +1016,7 @@ export function renderResult(
 			if (lines.length === 0) {
 				const text = fallbackText.trim() ? fallbackText : "No results";
 				const result = [theme.fg("dim", truncateToWidth(text, width))];
-				cached = { key, lines: result };
+				if (!hasDynamicRetry) cached = { key, lines: result };
 				return result;
 			}
 
@@ -1019,9 +1038,9 @@ export function renderResult(
 			}
 
 			const indented = lines.map(line =>
-				line.length > 0 ? truncateToWidth(`   ${line}`, width, Ellipsis.Omit) : "",
+				line.length > 0 ? truncateToWidth(`   ${replaceTabs(line)}`, width, Ellipsis.Omit) : "",
 			);
-			cached = { key, lines: indented };
+			if (!hasDynamicRetry) cached = { key, lines: indented };
 			return indented;
 		},
 		invalidate() {
@@ -1065,16 +1084,16 @@ function renderNestedTaskTree(
 ): string[] {
 	const lines: string[] = [];
 	for (const details of detailsList) {
-		const hasResults = Boolean(details.results && details.results.length > 0);
-		if (hasResults) {
+		const inflight = details.progress;
+		const hasInflight = Boolean(inflight && inflight.length > 0);
+		if (details.results && details.results.length > 0) {
 			details.results.forEach((result, index) => {
-				const isLast = index === details.results.length - 1;
+				const isLast = !hasInflight && index === details.results.length - 1;
 				lines.push(...renderAgentResult(result, isLast, expanded, theme));
 			});
-			continue;
 		}
-		const inflight = details.progress;
 		if (inflight && inflight.length > 0) {
+			lines.push(...renderProviderDegradationNotices(inflight, theme));
 			inflight.forEach((prog, index) => {
 				const isLast = index === inflight.length - 1;
 				lines.push(...renderAgentProgress(prog, isLast, expanded, theme, spinnerFrame, staticTime));

package/src/task/types.ts

@@ -1,6 +1,6 @@
 import type { ThinkingLevel } from "@gajae-code/agent-core";
 import type { Usage } from "@gajae-code/ai";
-import { $pickenv } from "@gajae-code/utils";
+import { $env } from "@gajae-code/utils";
 import * as z from "zod/v4";
 import { isValidTaskId, TASK_ID_DESCRIPTION } from "./id";
 import type { TaskResultReceipt } from "./receipt";
@@ -14,23 +14,28 @@ export type AgentSource = "bundled" | "user" | "project";
 export type ForkContextPolicy = "forbidden" | "allowed";
 export type ForkContextMode = "none" | "receipt" | "last-turn" | "bounded" | "full";
 
-const parseNumber = (value: string | undefined, defaultValue: number): number => {
-	if (value) {
-		try {
-			const number = Number.parseInt(value, 10);
-			if (!Number.isNaN(number) && number > 0) {
-				return number;
-			}
-		} catch {}
+const parsePositiveIntegerEnvironment = (keys: string[], defaultValue: number): number => {
+	for (const key of keys) {
+		const value = $env[key];
+		if (!value || value.trim().length === 0) continue;
+		if (!/^\d+$/.test(value)) return defaultValue;
+		const number = Number(value);
+		return Number.isSafeInteger(number) && number > 0 ? number : defaultValue;
 	}
 	return defaultValue;
 };
 
 /** Maximum output bytes per agent */
-export const MAX_OUTPUT_BYTES = parseNumber($pickenv("GJC_TASK_MAX_OUTPUT_BYTES", "PI_TASK_MAX_OUTPUT_BYTES"), 500_000);
+export const MAX_OUTPUT_BYTES = parsePositiveIntegerEnvironment(
+	["GJC_TASK_MAX_OUTPUT_BYTES", "PI_TASK_MAX_OUTPUT_BYTES"],
+	500_000,
+);
 
 /** Maximum output lines per agent */
-export const MAX_OUTPUT_LINES = parseNumber($pickenv("GJC_TASK_MAX_OUTPUT_LINES", "PI_TASK_MAX_OUTPUT_LINES"), 5000);
+export const MAX_OUTPUT_LINES = parsePositiveIntegerEnvironment(
+	["GJC_TASK_MAX_OUTPUT_LINES", "PI_TASK_MAX_OUTPUT_LINES"],
+	5000,
+);
 
 /** EventBus channel for raw subagent events */
 export const TASK_SUBAGENT_EVENT_CHANNEL = "task:subagent:event";
@@ -92,6 +97,12 @@ const createTaskItemSchema = (_contextEnabled: boolean) =>
 		id: z.string().max(48).refine(isValidTaskId, TASK_ID_DESCRIPTION).describe("filesystem-safe task identifier"),
 		description: z.string().describe("ui label, not seen by subagent"),
 		assignment: z.string().describe(assignmentDescription),
+		executionMode: z
+			.enum(["default", "ultragoal-red-team"])
+			.optional()
+			.describe(
+				"typed executor mode: default keeps ordinary executor behavior; ultragoal-red-team injects the Ultragoal QA/red-team prompt fragment. Prefer this over free-form assignment text (#2698).",
+			),
 		inheritContext: z
 			.enum(["none", "receipt", "last-turn", "bounded", "full"])
 			.optional()
@@ -285,6 +296,9 @@ export interface AgentProgress {
 		attempt: number;
 		maxAttempts: number;
 		unbounded?: boolean;
+		kind: "first_event_timeout" | "idle_stream_stall" | "provider_error";
+		provider?: string;
+		lastProviderProgressAtMs?: number;
 		delayMs: number;
 		errorMessage: string;
 		startedAtMs: number;

package/src/task/ultragoal-redteam-activation.ts

@@ -0,0 +1,75 @@
+/**
+ * Decide whether an executor assignment should inject the ultragoal red-team
+ * prompt fragment.
+ *
+ * Preference order (parse, don't re-validate later):
+ * 1. Explicit typed `executionMode` on the task / ExecutorOptions — authoritative.
+ * 2. Assignment text heuristics — only explicit Ultragoal QA/red-team phrasing.
+ *
+ * Bare mentions of `executorQa` (docs, quality-gate field names, negative
+ * instructions) must not flip the mode (#2698 / #2456).
+ *
+ * Mirrors `executor.md`:
+ * "activates only when the assignment explicitly labels Executor as Ultragoal
+ * completion QA/red-team or asks for `executorQa` red-team evidence."
+ */
+
+export const EXECUTOR_EXECUTION_MODES = ["default", "ultragoal-red-team"] as const;
+export type ExecutorExecutionMode = (typeof EXECUTOR_EXECUTION_MODES)[number];
+
+const ULTRAGOAL_COMPLETION_QA = /\bultragoal\s+completion\s+(?:qa|red[-\s]?team)\b/i;
+
+/** `executorQa` within a short window of red-team / matrix / evidence framing. */
+const EXECUTOR_QA_EVIDENCE =
+	/\bexecutorQa\b[\s\S]{0,120}\b(?:red[-\s]?team|matrix|evidence)\b|\b(?:red[-\s]?team|matrix|evidence)\b[\s\S]{0,120}\bexecutorQa\b/i;
+
+/** Common Ultragoal skill spawn phrasing: "executor QA/red-team lane". */
+const EXECUTOR_QA_LANE = /\bexecutor\b[\s\S]{0,48}\b(?:qa|red[-\s]?team)\s+lane\b/i;
+
+export function isExecutorExecutionMode(value: unknown): value is ExecutorExecutionMode {
+	return value === "default" || value === "ultragoal-red-team";
+}
+
+/**
+ * Parse a caller-supplied mode. Unknown values fail closed to `undefined`
+ * so the assignment heuristic (or default off) still applies — never invent on.
+ */
+export function parseExecutorExecutionMode(value: unknown): ExecutorExecutionMode | undefined {
+	if (typeof value !== "string") return undefined;
+	const normalized = value.trim().toLowerCase().replaceAll("_", "-");
+	if (normalized === "default" || normalized === "ordinary" || normalized === "implement") {
+		return "default";
+	}
+	if (
+		normalized === "ultragoal-red-team" ||
+		normalized === "ultragoal-redteam" ||
+		normalized === "red-team" ||
+		normalized === "redteam" ||
+		normalized === "executor-qa"
+	) {
+		return "ultragoal-red-team";
+	}
+	return undefined;
+}
+
+export function assignmentRequestsUltragoalRedTeam(assignment: string | undefined): boolean {
+	const text = assignment?.trim() ?? "";
+	if (text.length === 0) return false;
+	return ULTRAGOAL_COMPLETION_QA.test(text) || EXECUTOR_QA_EVIDENCE.test(text) || EXECUTOR_QA_LANE.test(text);
+}
+
+export interface ResolveUltragoalRedTeamArgs {
+	readonly executionMode?: unknown;
+	readonly assignment?: string;
+}
+
+/**
+ * Final activation decision for the executor prompt fragment.
+ * Typed mode wins; assignment text is a compatibility fallback only.
+ */
+export function resolveUltragoalRedTeamActivation(args: ResolveUltragoalRedTeamArgs): boolean {
+	const mode = parseExecutorExecutionMode(args.executionMode);
+	if (mode === "ultragoal-red-team") return true;
+	if (mode === "default") return false;
+	return assignmentRequestsUltragoalRedTeam(args.assignment);
+}

package/src/tools/bash-allowed-prefixes.ts

@@ -18,7 +18,14 @@ export interface BashRestrictionOptions {
 }
 
 const SHELL_CONTROL_CHARS = new Set([";", "|", "&", "<", ">", "(", ")"]);
-const UNSAFE_UNQUOTED_EXPANSION_CHARS = new Set(["$", "*", "?", "[", "]", "{", "}", "~"]);
+// `~` is handled separately: bash performs tilde expansion only at positions where a
+// tilde can open an expansion - the start of a word, and inside an assignment word
+// directly after the first `=` or after a `:` in the assigned value. A tilde anywhere
+// else (e.g. the git revision `HEAD~1`) is a literal character.
+const UNSAFE_UNQUOTED_EXPANSION_CHARS = new Set(["$", "*", "?", "[", "]", "{", "}"]);
+// A bash assignment word has a plain assignment name before its first unquoted `=`,
+// in either the `name=value` or the compound `name+=value` form.
+const ASSIGNMENT_NAME_PATTERN = /^[A-Za-z_][A-Za-z0-9_]*\+?$/u;
 const ALLOWED_STATE_ACTIONS = new Set(["read", "write", "contract"]);
 const CANONICAL_STATE_TARGETS = new Set<string>(CANONICAL_GJC_WORKFLOW_SKILLS);
 const READ_ONLY_COMMANDS = new Set(["grep", "rg", "tree", "ls", "pwd", "wc", "du", "file", "stat"]);
@@ -29,6 +36,14 @@ function parseShellWords(command: string): { words: string[]; reason?: string }
 	let quote: "single" | "double" | null = null;
 
 	let wordStarted = false;
+	// Per-token tilde-expansion state. `assignmentNameCandidate` stays true while the
+	// characters before the first unquoted `=` could still form an assignment name (any
+	// quoting in that region disqualifies the token), `sawAssignmentEquals` records that
+	// the token is an assignment word, and `tildeExpandable` means an unquoted `~` at the
+	// current index would actually be expanded by bash.
+	let assignmentNameCandidate = true;
+	let sawAssignmentEquals = false;
+	let tildeExpandable = false;
 	for (let index = 0; index < command.length; index += 1) {
 		const char = command[index]!;
 		const next = command[index + 1];
@@ -63,11 +78,15 @@ function parseShellWords(command: string): { words: string[]; reason?: string }
 		if (char === "'") {
 			quote = "single";
 			wordStarted = true;
+			if (!sawAssignmentEquals) assignmentNameCandidate = false;
+			tildeExpandable = false;
 			continue;
 		}
 		if (char === '"') {
 			quote = "double";
 			wordStarted = true;
+			if (!sawAssignmentEquals) assignmentNameCandidate = false;
+			tildeExpandable = false;
 			continue;
 		}
 		if (char === "`" || (char === "$" && next === "(")) {
@@ -82,17 +101,42 @@ function parseShellWords(command: string): { words: string[]; reason?: string }
 		if (UNSAFE_UNQUOTED_EXPANSION_CHARS.has(char)) {
 			return { words, reason: `shell expansion character '${char}' is not allowed in restricted bash commands` };
 		}
+		// Tilde expansion is positional: bash expands a tilde only where one can open an
+		// expansion, i.e. at the start of a word or, inside an assignment word, directly
+		// after the first `=` or after a `:` in the assigned value. A tilde anywhere else
+		// is literal, which is what git revision syntax such as `HEAD~1` relies on.
+		if (char === "~" && (!wordStarted || tildeExpandable)) {
+			return { words, reason: "shell expansion character '~' is not allowed in restricted bash commands" };
+		}
 		if (/\s/u.test(char)) {
 			if (wordStarted) {
 				words.push(current);
 				current = "";
 				wordStarted = false;
 			}
+			assignmentNameCandidate = true;
+			sawAssignmentEquals = false;
+			tildeExpandable = false;
 			continue;
 		}
 		if (char === "\\") {
 			return { words, reason: "backslash escapes are not allowed in restricted bash commands" };
 		}
+		if (char === "=" && !sawAssignmentEquals) {
+			// Only a plain assignment name before the first unquoted `=` makes this an
+			// assignment word whose value bash scans for tilde expansions.
+			if (assignmentNameCandidate && ASSIGNMENT_NAME_PATTERN.test(current)) {
+				sawAssignmentEquals = true;
+				tildeExpandable = true;
+			} else {
+				assignmentNameCandidate = false;
+				tildeExpandable = false;
+			}
+		} else if (char === ":" && sawAssignmentEquals) {
+			tildeExpandable = true;
+		} else {
+			tildeExpandable = false;
+		}
 		current += char;
 		wordStarted = true;
 	}

package/src/tools/index.ts

@@ -275,9 +275,11 @@ export interface ToolSession {
 	/** Agent identity used for IRC routing. Returns the registry id (e.g. "0-Main", "0-AuthLoader"). */
 	getAgentId?: () => string | null;
 	/** Look up a registered tool by name (used by the eval js backend's tool bridge). */
+	getToolByName?: (name: string) => AgentTool | undefined;
+	/** Look up a registered tool with the session's execution guards applied. */
+	getToolForExecution?: (name: string) => AgentTool | undefined;
 	/** Purge undelivered queued custom messages matching the predicate. Returns counts. */
 	purgeQueuedCustomMessages?: (predicate: (message: CustomMessage) => boolean) => PurgeQueuedCustomMessagesResult;
-	getToolByName?: (name: string) => AgentTool | undefined;
 	/** Agent registry for IRC routing across live sessions. */
 	agentRegistry?: AgentRegistry;
 	/** Optional restricted bash command prefixes for read-only role agents and constrained modes. */