@prometheus-ai/agent 0.5.3 → 0.5.8

package/src/tools/debug.ts

@@ -9,7 +9,7 @@ import type {
 } from "@prometheus-ai/agent-core";
 import { type Component, Text } from "@prometheus-ai/tui";
 import { isEnoent, prompt } from "@prometheus-ai/utils";
-import * as z from "zod/v4";
+import { z } from "zod/v4";
 import {
 	type DapBreakpointRecord,
 	type DapCapabilities,
@@ -22,6 +22,7 @@ import {
 	type DapFunctionBreakpointRecord,
 	type DapInstructionBreakpointRecord,
 	type DapModule,
+	type DapResolvedAdapter,
 	type DapScope,
 	type DapSessionSummary,
 	type DapSource,
@@ -30,6 +31,8 @@ import {
 	type DapVariable,
 	dapSessionManager,
 	getAvailableAdapters,
+	type LaunchProgramKind,
+	resolveLaunchOverrides,
 	selectAttachAdapter,
 	selectLaunchAdapter,
 } from "../dap";
@@ -489,16 +492,23 @@ function getConfiguredAdapters(cwd: string): string {
 	const adapters = getAvailableAdapters(cwd).map(adapter => adapter.name);
 	return adapters.length > 0 ? adapters.join(", ") : "none";
 }
-async function validateLaunchProgram(program: string, cwd: string): Promise<void> {
-	let isDirectory: boolean;
+
+async function classifyLaunchProgram(program: string): Promise<LaunchProgramKind> {
 	try {
-		isDirectory = (await fs.stat(program)).isDirectory();
+		return (await fs.stat(program)).isDirectory() ? "directory" : "file";
 	} catch (error) {
-		if (isEnoent(error)) return;
+		if (isEnoent(error)) return "missing";
 		throw error;
 	}
-	if (!isDirectory) return;
+}
 
+function validateLaunchProgram(
+	program: string,
+	cwd: string,
+	programKind: LaunchProgramKind,
+	adapter: DapResolvedAdapter,
+): void {
+	if (programKind !== "directory" || adapter.acceptsDirectoryProgram) return;
 	const displayPath = formatPathRelativeToCwd(program, cwd, { trailingSlash: true });
 	throw new ToolError(
 		`launch program resolves to a directory: ${displayPath}. Pass an executable file path, or for Python use adapter "debugpy" with program set to the .py file.`,
@@ -582,10 +592,13 @@ export const debugToolRenderer = {
 	): Component {
 		const outputBlock = new CachedOutputBlock();
 		return markFramedBlockComponent({
-			render(width: number): string[] {
+			render(width: number): readonly string[] {
 				const action = (args?.action ?? result.details?.action ?? "debug").replaceAll("_", " ");
-				const status = options.isPartial ? "running" : result.isError ? "error" : "success";
-				const header = `${formatStatusIcon(status, theme, options.spinnerFrame)} Debug ${action}`;
+				const success = !options.isPartial && !result.isError;
+				const statusIcon = success
+					? theme.styledSymbol("tool.debug", "accent")
+					: formatStatusIcon(options.isPartial ? "running" : "error", theme, options.spinnerFrame);
+				const header = `${statusIcon} Debug ${action}`;
 				const summaryLines = result.details?.snapshot
 					? formatSessionSnapshot(result.details.snapshot).map(line => replaceTabs(line))
 					: [];
@@ -676,8 +689,8 @@ export class DebugTool implements AgentTool<typeof debugSchema, DebugToolDetails
 				}
 				const commandCwd = params.cwd ? resolveToCwd(params.cwd, this.session.cwd) : this.session.cwd;
 				const program = resolveToCwd(params.program, commandCwd);
-				await validateLaunchProgram(program, commandCwd);
-				const adapter = selectLaunchAdapter(program, commandCwd, params.adapter);
+				const programKind = await classifyLaunchProgram(program);
+				const adapter = selectLaunchAdapter(program, commandCwd, params.adapter, programKind);
 				if (!adapter) {
 					if (params.adapter === "debugpy") {
 						throw new ToolError("adapter 'debugpy' is not available: python not found in PATH");
@@ -686,8 +699,10 @@ export class DebugTool implements AgentTool<typeof debugSchema, DebugToolDetails
 						`No debugger adapter available. Installed adapters: ${getConfiguredAdapters(commandCwd)}`,
 					);
 				}
+				validateLaunchProgram(program, commandCwd, programKind, adapter);
+				const extraLaunchArguments = resolveLaunchOverrides(adapter, program, programKind);
 				const snapshot = await dapSessionManager.launch(
-					{ adapter, program, args: params.args, cwd: commandCwd },
+					{ adapter, program, args: params.args, cwd: commandCwd, extraLaunchArguments },
 					combinedSignal,
 					timeoutSec * 1000,
 				);

package/src/tools/eval-render.ts

@@ -16,9 +16,8 @@ import type { EvalCellResult, EvalLanguage, EvalStatusEvent, EvalToolDetails } f
 import type { RenderResultOptions } from "../extensibility/custom-tools/types";
 import { formatContextUsage } from "../modes/components/status-line/context-thresholds";
 import { truncateToVisualLines } from "../modes/components/visual-truncate";
-import { shimmerEnabled } from "../modes/theme/shimmer";
 import { getMarkdownTheme, type Theme } from "../modes/theme/theme";
-import { borderShimmerTick, markFramedBlockComponent, renderCodeCell } from "../tui";
+import { markFramedBlockComponent, renderCodeCell } from "../tui";
 import {
 	JSON_TREE_MAX_DEPTH_COLLAPSED,
 	JSON_TREE_MAX_DEPTH_EXPANDED,
@@ -34,20 +33,13 @@ import {
 	formatDuration,
 	formatStatusIcon,
 	formatTitle,
+	previewWindowRows,
 	replaceTabs,
 	shortenPath,
 	truncateToWidth,
 	wrapBrackets,
 } from "./render-utils";
 export const EVAL_DEFAULT_PREVIEW_LINES = 10;
-/**
- * Rows of source kept in the *pending* eval preview. The window follows the
- * streaming edge (newest lines pinned to the bottom) so you can watch the code
- * being written, while staying bounded — a volatile tool block taller than the
- * viewport would otherwise strand its scrolled-off head out of native scrollback
- * on ED3-risk terminals. Matches the streaming windows used by edit/write.
- */
-export const EVAL_STREAMING_PREVIEW_LINES = 12;
 
 function languageForHighlighter(language: EvalLanguage | undefined): "python" | "javascript" {
 	return language === "js" ? "javascript" : "python";
@@ -181,7 +173,7 @@ function renderAgentProgressEvents(events: EvalStatusEvent[], theme: Theme, spin
 		const status = agentEventStatus(event.status);
 		const iconStatus =
 			status === "completed"
-				? "success"
+				? "done"
 				: status === "failed"
 					? "error"
 					: status === "aborted"
@@ -191,10 +183,13 @@ function renderAgentProgressEvents(events: EvalStatusEvent[], theme: Theme, spin
 							: "running";
 		const iconColor =
 			status === "completed" ? "success" : status === "failed" || status === "aborted" ? "error" : "accent";
-		const icon = formatStatusIcon(iconStatus, theme, status === "running" ? spinnerFrame : undefined);
+		const icon =
+			status === "completed"
+				? theme.styledSymbol("tool.eval", "accent")
+				: theme.fg(iconColor, formatStatusIcon(iconStatus, theme, status === "running" ? spinnerFrame : undefined));
 
 		const id = eventString(event.id) ?? "agent";
-		let line = `${prefix} ${theme.fg(iconColor, icon)} ${theme.fg("accent", theme.bold(id))}`;
+		let line = `${prefix} ${icon} ${theme.fg("accent", theme.bold(id))}`;
 
 		if (status === "failed" || status === "aborted") {
 			line += ` ${formatBadge(status, iconColor, theme)}`;
@@ -255,7 +250,7 @@ function formatStatusEvent(event: EvalStatusEvent, theme: Theme): string {
 		sh: "icon.package",
 		env: "icon.package",
 		batch: "icon.package",
-		llm: "icon.package",
+		completion: "icon.package",
 		log: "icon.package",
 		phase: "icon.package",
 	};
@@ -324,7 +319,7 @@ function formatStatusEvent(event: EvalStatusEvent, theme: Theme): string {
 		case "batch":
 			parts.push(`${data.files} file${(data.files as number) !== 1 ? "s" : ""} processed`);
 			break;
-		case "llm":
+		case "completion":
 			if (data.model) parts.push(String(data.model));
 			if (data.tier && data.tier !== data.model) parts.push(`(${data.tier})`);
 			parts.push(`${data.chars ?? 0} chars`);
@@ -461,7 +456,7 @@ function formatCellOutputLines(
 	previewLines: number,
 	theme: Theme,
 	width: number,
-): { lines: string[]; hiddenCount: number } {
+): { lines: readonly string[]; hiddenCount: number } {
 	if (!cell.output) {
 		return { lines: [], hiddenCount: 0 };
 	}
@@ -498,9 +493,8 @@ export const evalToolRenderer = {
 		let cached: { key: string; width: number; result: string[] } | undefined;
 
 		return markFramedBlockComponent({
-			render: (width: number): string[] => {
-				const animate = options.isPartial && shimmerEnabled();
-				const key = `${animate ? borderShimmerTick() : 0}|${options.expanded ? 1 : 0}|${cells.map(c => `${c.language}:${c.title ?? ""}:${c.code.length}`).join("|")}`;
+			render: (width: number): readonly string[] => {
+				const key = `${options.expanded ? 1 : 0}|${previewWindowRows()}|${cells.map(c => `${c.language}:${c.title ?? ""}:${c.code.length}`).join("|")}`;
 				if (cached && cached.key === key && cached.width === width) {
 					return cached.result;
 				}
@@ -517,17 +511,12 @@ export const evalToolRenderer = {
 							title: cell.title,
 							status: "pending",
 							width,
-							codeMaxLines: EVAL_STREAMING_PREVIEW_LINES,
-							// Follow the streaming edge with a bounded tail window so the
-							// newest source stays visible as it is written, instead of
-							// rendering every line of a >100-line `code` — which would
-							// overflow the viewport and, because a tool block is volatile
-							// (it collapses to a capped result), strand its scrolled-off head
-							// out of native scrollback, cutting the box top. `Ctrl+O` lifts
-							// the window via `expanded` for a deliberate full view.
+							// Viewport-sized tail window following the newest streamed code
+							// line; renderResult keeps the same cap so the cell never snaps
+							// open on completion. Only ctrl+o uncaps.
 							codeTail: true,
+							codeMaxLines: previewWindowRows(),
 							expanded: options.expanded,
-							animate,
 						},
 						uiTheme,
 					);
@@ -587,11 +576,10 @@ export const evalToolRenderer = {
 			let cached: { key: string; width: number; result: string[] } | undefined;
 
 			return markFramedBlockComponent({
-				render: (width: number): string[] => {
+				render: (width: number): readonly string[] => {
 					const expanded = options.renderContext?.expanded ?? options.expanded;
 					const previewLines = options.renderContext?.previewLines ?? EVAL_DEFAULT_PREVIEW_LINES;
-					const animate = options.isPartial && shimmerEnabled();
-					const key = `${expanded}|${previewLines}|${options.spinnerFrame}|${animate ? borderShimmerTick() : 0}`;
+					const key = `${expanded}|${previewLines}|${options.spinnerFrame}|${previewWindowRows()}`;
 					if (cached && cached.key === key && cached.width === width) {
 						return cached.result;
 					}
@@ -628,10 +616,13 @@ export const evalToolRenderer = {
 								duration: cell.durationMs,
 								output: outputLines.length > 0 ? outputLines.join("\n") : undefined,
 								outputMaxLines: outputLines.length,
-								codeMaxLines: expanded ? Number.POSITIVE_INFINITY : EVAL_DEFAULT_PREVIEW_LINES,
+								// Same viewport-sized tail window as the pending preview so the
+								// cell never snaps open on completion; only ctrl+o uncaps.
+								// `output` keeps its own preview cap from above.
+								codeTail: true,
+								codeMaxLines: previewWindowRows(),
 								expanded,
 								width,
-								animate,
 							},
 							uiTheme,
 						);
@@ -711,12 +702,12 @@ export const evalToolRenderer = {
 		const textContent = `\n${styledOutput}`;
 
 		let cachedWidth: number | undefined;
-		let cachedLines: string[] | undefined;
+		let cachedLines: readonly string[] | undefined;
 		let cachedSkipped: number | undefined;
 		let cachedPreviewLines: number | undefined;
 
 		return {
-			render: (width: number): string[] => {
+			render: (width: number): readonly string[] => {
 				const previewLines = options.renderContext?.previewLines ?? EVAL_DEFAULT_PREVIEW_LINES;
 				if (cachedLines === undefined || cachedWidth !== width || cachedPreviewLines !== previewLines) {
 					const result = truncateToVisualLines(textContent, previewLines, width);
@@ -760,6 +751,12 @@ export const evalToolRenderer = {
 			},
 		};
 	},
+
 	mergeCallAndResult: true,
 	inline: true,
+	// Collapsed pending preview shows tail-window code cells; the result render
+	// interleaves each cell's output under its code, re-laying-out every row
+	// below the first cell. Expanded output is top-anchored enough for the
+	// transcript to commit its settled prefix.
+	provisionalPendingPreview: "collapsed",
 };

package/src/tools/eval.ts

@@ -1,7 +1,7 @@
 import type { AgentTool, AgentToolContext, AgentToolResult, AgentToolUpdateCallback } from "@prometheus-ai/agent-core";
 import type { ImageContent } from "@prometheus-ai/ai";
 import { prompt } from "@prometheus-ai/utils";
-import * as z from "zod/v4";
+import { z } from "zod/v4";
 import { jsBackend, pythonBackend } from "../eval";
 import type { ExecutorBackend, ExecutorBackendResult } from "../eval/backend";
 import { EVAL_TIMEOUT_PAUSE_OP, EVAL_TIMEOUT_RESUME_OP } from "../eval/bridge-timeout";
@@ -10,6 +10,7 @@ import { defaultEvalSessionId } from "../eval/session-id";
 import type { EvalCellResult, EvalDisplayOutput, EvalLanguage, EvalStatusEvent, EvalToolDetails } from "../eval/types";
 import evalDescription from "../prompts/tools/eval.md" with { type: "text" };
 import { DEFAULT_MAX_BYTES, OutputSink, type OutputSummary, TailBuffer } from "../session/streaming-output";
+import { webpExclusionForModel } from "../utils/image-loading";
 import { formatDimensionNote, resizeImage } from "../utils/image-resize";
 import type { ToolSession } from ".";
 import { truncateForPrompt } from "./approval";
@@ -30,7 +31,7 @@ const evalCellSchema = z.object({
 	language: z.enum(["py", "js"]).describe('runtime: "py" for the IPython kernel, "js" for the persistent JS VM'),
 	code: z.string().describe("cell body, verbatim. Use top-level await freely."),
 	title: z.string().optional().describe('short label shown in transcript (e.g. "imports", "load config")'),
-	timeout: z.number().int().min(1).max(600).optional().describe("per-cell timeout in seconds (1-600, default 30)"),
+	timeout: z.number().int().min(1).max(3600).optional().describe("per-cell timeout in seconds (1-3600, default 30)"),
 	reset: z
 		.boolean()
 		.optional()
@@ -88,12 +89,21 @@ function formatDisplayOutputsForText(outputs: EvalDisplayOutput[]): string {
 export interface EvalToolDescriptionOptions {
 	py?: boolean;
 	js?: boolean;
+	/**
+	 * Whether `agent()` is allowed in this session. Driven by the parent's
+	 * spawn policy (`getSessionSpawns`). Defaults to `true` for backward
+	 * compatibility — when the session forbids spawning, the prelude doc
+	 * omits the `agent()` entry so the model does not promise itself a
+	 * helper that will only ever throw "spawns disabled".
+	 */
+	spawns?: boolean;
 }
 
 export function getEvalToolDescription(options: EvalToolDescriptionOptions = {}): string {
 	const py = options.py ?? true;
 	const js = options.js ?? true;
-	return prompt.render(evalDescription, { py, js });
+	const spawns = options.spawns ?? true;
+	return prompt.render(evalDescription, { py, js, spawns });
 }
 
 export interface EvalToolOptions {
@@ -169,7 +179,9 @@ export class EvalTool implements AgentTool<typeof evalSchema> {
 	get description(): string {
 		if (!this.session) return getEvalToolDescription();
 		const backends = resolveEvalBackends(this.session);
-		return getEvalToolDescription({ py: backends.python, js: backends.js });
+		const sessionSpawns = this.session.getSessionSpawns?.() ?? "*";
+		const spawnsAllowed = sessionSpawns !== "" && sessionSpawns !== null;
+		return getEvalToolDescription({ py: backends.python, js: backends.js, spawns: spawnsAllowed });
 	}
 	readonly parameters = evalSchema;
 	readonly concurrency = "exclusive";
@@ -208,6 +220,7 @@ export class EvalTool implements AgentTool<typeof evalSchema> {
 			throw new ToolError("Eval tool requires a session when not using proxy executor");
 		}
 		const session = this.session;
+		const excludeWebP = webpExclusionForModel(session.getActiveModel?.());
 
 		const cells: ResolvedEvalCell[] = [];
 		for (let i = 0; i < params.cells.length; i++) {
@@ -315,7 +328,7 @@ export class EvalTool implements AgentTool<typeof evalSchema> {
 					const cell = cells[i];
 					const backend = cell.resolved.backend;
 					// The per-cell `timeout` is a budget on the cell runtime's *own*
-					// work. Host-side `agent()`/`parallel()`/`llm()` bridge calls suspend
+					// work. Host-side `agent()`/`parallel()`/`completion()` bridge calls suspend
 					// that budget entirely and restart a fresh timeout window when control
 					// returns to Python/JS. Compute, stdout, `log()`/`phase()`, and
 					// ordinary tool calls all count against the budget. The watchdog drives
@@ -347,8 +360,6 @@ export class EvalTool implements AgentTool<typeof evalSchema> {
 							session,
 							idleTimeoutMs,
 							reset: cell.reset,
-							artifactPath,
-							artifactId,
 							onChunk: chunk => {
 								outputSink!.push(chunk);
 							},
@@ -381,11 +392,14 @@ export class EvalTool implements AgentTool<typeof evalSchema> {
 							cellDisplayOutputs.push(output);
 						}
 						if (output.type === "image") {
-							const resized = await resizeImage({
-								type: "image",
-								data: output.data,
-								mimeType: output.mimeType,
-							});
+							const resized = await resizeImage(
+								{
+									type: "image",
+									data: output.data,
+									mimeType: output.mimeType,
+								},
+								{ excludeWebP },
+							);
 							const image: ImageContent = {
 								type: "image",
 								data: resized.data,