@prometheus-ai/agent 0.5.4 → 0.5.8

package/src/tools/index.ts

@@ -1,12 +1,14 @@
 import type { AgentTelemetryConfig, AgentTool } from "@prometheus-ai/agent-core";
-import type { ToolChoice } from "@prometheus-ai/ai";
+import type { FetchImpl, Model, ToolChoice } from "@prometheus-ai/ai";
 import type { InMemorySnapshotStore } from "@prometheus-ai/hashline";
 import { logger } from "@prometheus-ai/utils";
 import type { AsyncJobManager } from "../async/job-manager";
+import type { Rule } from "../capability/rule";
 import type { PromptTemplate } from "../config/prompt-templates";
 import type { Settings } from "../config/settings";
 import { EditTool } from "../edit";
 import { checkPythonKernelAvailability } from "../eval/py/kernel";
+import type { ToolPathWithSource } from "../extensibility/custom-tools";
 import type { Skill } from "../extensibility/skills";
 import type { GoalModeState, GoalRuntime } from "../goals";
 import { GoalTool } from "../goals/tools/goal-tool";
@@ -16,13 +18,15 @@ import { LspTool } from "../lsp";
 import type { MCPManager } from "../mcp";
 import type { MnemopiSessionState } from "../mnemopi/state";
 import type { PlanModeState } from "../plan-mode/state";
-import { type AgentRegistry, MAIN_AGENT_ID } from "../registry/agent-registry";
+import type { AgentRegistry } from "../registry/agent-registry";
 import type { ArtifactManager } from "../session/artifacts";
 import type { ClientBridge } from "../session/client-bridge";
 import type { CustomMessage } from "../session/messages";
+import type { UsageStatistics } from "../session/session-entries";
 import type { ToolChoiceQueue } from "../session/tool-choice-queue";
 import { TaskTool } from "../task";
 import type { AgentOutputManager } from "../task/output-manager";
+import { canSpawnAtDepth } from "../task/types";
 import { countToolsForAutoDiscovery, resolveEffectiveToolDiscoveryMode } from "../tool-discovery/mode";
 import type { DiscoverableTool, DiscoverableToolSearchIndex } from "../tool-discovery/tool-index";
 import type { EventBus } from "../utils/event-bus";
@@ -40,8 +44,10 @@ import { resolveEvalBackends } from "./eval-backends";
 import { FindTool } from "./find";
 import { GithubTool } from "./gh";
 import { InspectImageTool } from "./inspect-image";
-import { IrcTool } from "./irc";
+import { IrcTool, isIrcEnabled } from "./irc";
 import { JobTool } from "./job";
+import { LearnTool } from "./learn";
+import { ManageSkillTool } from "./manage-skill";
 import { MemoryEditTool } from "./memory-edit";
 import { MemoryRecallTool } from "./memory-recall";
 import { MemoryReflectTool } from "./memory-reflect";
@@ -59,11 +65,7 @@ import { type TodoPhase, TodoTool } from "./todo";
 import { WriteTool } from "./write";
 import { YieldTool } from "./yield";
 
-// Exa MCP tools (22 tools)
-
 export * from "../edit";
-export * from "../exa";
-export type * from "../exa/types";
 export * from "../goals";
 export * from "../lsp";
 export * from "../session/streaming-output";
@@ -84,6 +86,8 @@ export * from "./image-gen";
 export * from "./inspect-image";
 export * from "./irc";
 export * from "./job";
+export * from "./learn";
+export * from "./manage-skill";
 export * from "./memory-edit";
 export * from "./memory-recall";
 export * from "./memory-reflect";
@@ -117,12 +121,44 @@ export type {
 	DiscoverableToolSource,
 } from "../tool-discovery/tool-index";
 
+/**
+ * A late LSP diagnostics result that arrived after the edit/write tool already
+ * returned. Surfaced to the model and the transcript via
+ * {@link ToolSession.queueDeferredDiagnostics}, batched through the session
+ * yield queue like background-job results.
+ */
+export interface DeferredDiagnosticsEntry {
+	/** Absolute path the diagnostics belong to (the renderer shortens it). */
+	path: string;
+	/** One-line severity summary, e.g. "2 errors". */
+	summary: string;
+	/** Formatted, ready-to-display diagnostic lines. */
+	messages: string[];
+	/** True when any message is error severity. */
+	errored: boolean;
+	/**
+	 * Evaluated at injection time (in the dispatcher's stale check): drop the entry
+	 * when a newer mutation to the same file has superseded it, so the model never
+	 * sees diagnostics for stale content.
+	 */
+	isStale(): boolean;
+}
+
 /** Session context for tool factories */
 export interface ToolSession {
 	/** Current working directory */
 	cwd: string;
 	/** Whether UI is available */
 	hasUI: boolean;
+	/**
+	 * Suppress the spawn specialization/coordination advisory appended to `task`
+	 * results. Set by internal/programmatic callers (e.g. the commit agent's
+	 * file-analysis fan-out) whose results are consumed by code — not by a model
+	 * orchestrating further spawns — so the nudge would only be noise.
+	 */
+	suppressSpawnAdvisory?: boolean;
+	/** Optional fetch implementation injected into the URL read pipeline (tests, proxies). Defaults to global fetch. */
+	fetch?: FetchImpl;
 	/** Skip Python kernel availability check and warmup */
 	skipPythonPreflight?: boolean;
 	/** Pre-loaded context files (AGENTS.md, etc) */
@@ -133,6 +169,21 @@ export interface ToolSession {
 	skills?: Skill[];
 	/** Pre-loaded prompt templates */
 	promptTemplates?: PromptTemplate[];
+	/** Pre-loaded rules (forwarded to subagents to skip re-discovery). */
+	rules?: Rule[];
+	/**
+	 * Pre-discovered extension source paths. Forwarded to subagents so they
+	 * skip the FS scan but still re-bind extensions to their own session-scoped
+	 * `ExtensionAPI` (cwd, eventBus, runtime). Inline extension factories
+	 * (`<inline-N>`) are NOT included — those are session-local.
+	 */
+	extensionPaths?: string[];
+	/**
+	 * Pre-discovered custom-tool source paths from `.prometheus/tools/`, `.claude/tools/`,
+	 * plugins, etc. Forwarded to subagents so they skip the FS scan but still
+	 * re-bind tools to their own session-scoped `CustomToolAPI`.
+	 */
+	customToolPaths?: ToolPathWithSource[];
 	/** Whether LSP integrations are enabled */
 	enableLsp?: boolean;
 	/** Whether an edit-capable tool is available in this session (controls hashline output) */
@@ -159,7 +210,7 @@ export interface ToolSession {
 	getSessionId?: () => string | null;
 	/** Get Hindsight runtime state for this agent session. */
 	getHindsightSessionState?: () => HindsightSessionState | undefined;
-	/** Get Prometheus Memory runtime state for this agent session. */
+	/** Get Mnemopi runtime state for this agent session. */
 	getMnemopiSessionState?: () => MnemopiSessionState | undefined;
 	/** Agent identity used for IRC routing. Returns the registry id (e.g. "Main", "AuthLoader"). */
 	getAgentId?: () => string | null;
@@ -179,6 +230,8 @@ export interface ToolSession {
 	getModelString?: () => string | undefined;
 	/** Get the current session model string, regardless of how it was chosen */
 	getActiveModelString?: () => string | undefined;
+	/** Get the current session model object (provider/api capabilities), regardless of how it was chosen. */
+	getActiveModel?: () => Model | undefined;
 	/** Auth storage for passing to subagents (avoids re-discovery) */
 	authStorage?: import("../session/auth-storage").AuthStorage;
 	/** Model registry for passing to subagents (avoids re-discovery) */
@@ -215,15 +268,13 @@ export interface ToolSession {
 	/** Goal runtime for the active agent session. */
 	getGoalRuntime?: () => GoalRuntime | undefined;
 	/** Get cumulative session usage statistics (input/output tokens, cost). */
-	getUsageStatistics?: () => import("../session/session-manager").UsageStatistics;
+	getUsageStatistics?: () => UsageStatistics;
 	/** Current per-turn token budget {total, spent, hard} for the eval `budget` helper. */
 	getTurnBudget?: () => { total: number | null; spent: number; hard: boolean };
 	/** Record output tokens consumed by an eval-spawned subagent toward the current turn budget. */
 	recordEvalSubagentUsage?: (output: number) => void;
 	/** Bridge to the connected client (e.g. ACP editor host). Tools should route fs/terminal/permission requests through this when available. */
 	getClientBridge?: () => ClientBridge | undefined;
-	/** Get compact conversation context for subagents (excludes tool results, system prompts) */
-	getCompactContext?: () => string;
 	/** Get cached todo phases for this session. */
 	getTodoPhases?: () => TodoPhase[];
 	/** Replace cached todo phases for this session. */
@@ -282,8 +333,24 @@ export interface ToolSession {
 	 *  model for each file. Lazily initialized by `getDiagnosticsLedger`. */
 	diagnosticsLedger?: import("../lsp/diagnostics-ledger").DiagnosticsLedger;
 
+	/** Per-session ledger of consecutive byte-identical no-op edits, keyed by
+	 *  canonical file path. The hashline executor escalates a soft no-op hint
+	 *  to a thrown error once the same payload no-ops `NOOP_HARD_LIMIT` times,
+	 *  breaking subagent loops that ignore the textual hint (issue #2081).
+	 *  Lazily initialized by `getNoopLoopGuard`. */
+	noopLoopGuard?: import("../edit/hashline/noop-loop-guard").NoopLoopGuard;
+
 	/** Queue a hidden message to be injected at the next agent turn. */
 	queueDeferredMessage?(message: CustomMessage): void;
+	/** Queue late LSP diagnostics (arrived after an edit/write returned) to be shown
+	 *  in the transcript and delivered to the model at the next yield, like background
+	 *  job results. */
+	queueDeferredDiagnostics?(entry: DeferredDiagnosticsEntry): void;
+	/** Bump and return the session-global mutation counter for `path`. Edit/write
+	 *  tools call this on every file mutation so stale late-diagnostics can be dropped. */
+	bumpFileMutationVersion?(path: string): number;
+	/** Read the current session-global mutation counter for `path` (0 if never mutated). */
+	getFileMutationVersion?(path: string): number;
 	/** Get the active OpenTelemetry config so subagent dispatch can forward
 	 *  the parent's tracer/hooks with the subagent's own identity stamped. */
 	getTelemetry?: () => AgentTelemetryConfig | undefined;
@@ -310,6 +377,38 @@ export function computeEssentialBuiltinNames(settings: Settings): string[] {
 	return [...DEFAULT_ESSENTIAL_TOOL_NAMES];
 }
 
+/**
+ * Filter the initial active tool set when `tools.discoveryMode === "all"`.
+ *
+ * Non-essential discoverable built-ins are hidden — the model rediscovers them
+ * via `search_tool_bm25` and activates them on demand. A tool survives hiding
+ * when it is essential, explicitly requested, restored from a prior selection,
+ * or required by a forced tool_choice feature (`forceActive`). The last case is
+ * load-bearing: a named tool_choice (e.g. the eager `todo` prelude) must
+ * reference a tool present in the request, or the provider rejects it with 400.
+ */
+export function filterInitialToolsForDiscoveryAll(
+	initialToolNames: string[],
+	opts: {
+		loadModeOf: (name: string) => BuiltinToolLoadMode | undefined;
+		essentialNames: ReadonlySet<string>;
+		explicitlyRequested: ReadonlySet<string>;
+		restored: ReadonlySet<string>;
+		forceActive: ReadonlySet<string>;
+	},
+): string[] {
+	return initialToolNames.filter(name => {
+		const loadMode = opts.loadModeOf(name);
+		if (!loadMode) return true; // not a built-in — leave MCP/custom/extension to existing logic
+		if (loadMode === "essential") return true;
+		if (opts.essentialNames.has(name)) return true;
+		if (opts.explicitlyRequested.has(name)) return true;
+		if (opts.restored.has(name)) return true;
+		if (opts.forceActive.has(name)) return true;
+		return false;
+	});
+}
+
 /**
  * Public callable factory map. External callers may invoke `BUILTIN_TOOLS.read(session)` or
  * `BUILTIN_TOOLS[name](session)` to construct a tool directly.
@@ -334,7 +433,7 @@ export const BUILTIN_TOOLS: Record<string, ToolFactory> = {
 	checkpoint: CheckpointTool.createIf,
 	rewind: RewindTool.createIf,
 	task: s => TaskTool.create(s),
-	job: JobTool.createIf,
+	job: s => new JobTool(s),
 	irc: IrcTool.createIf,
 	todo: s => new TodoTool(s),
 	web_search: s => new WebSearchTool(s),
@@ -344,6 +443,8 @@ export const BUILTIN_TOOLS: Record<string, ToolFactory> = {
 	retain: MemoryRetainTool.createIf,
 	recall: MemoryRecallTool.createIf,
 	reflect: MemoryReflectTool.createIf,
+	learn: LearnTool.createIf,
+	manage_skill: ManageSkillTool.createIf,
 };
 
 export const HIDDEN_TOOLS: Record<string, ToolFactory> = {
@@ -384,7 +485,12 @@ export async function createTools(session: ToolSession, toolNames?: string[]): P
 		!allowJs &&
 		(requestedTools === undefined || requestedTools.includes("eval"))
 	) {
-		const availability = await logger.time("createTools:pythonCheck", checkPythonKernelAvailability, session.cwd);
+		const availability = await logger.time(
+			"createTools:pythonCheck",
+			checkPythonKernelAvailability,
+			session.cwd,
+			session.settings.get("python.interpreter")?.trim() || undefined,
+		);
 		pythonAvailable = availability.ok;
 		if (!availability.ok) {
 			logger.warn("Python kernel unavailable and JS backend disabled; eval will be unavailable", {
@@ -414,11 +520,26 @@ export async function createTools(session: ToolSession, toolNames?: string[]): P
 		) {
 			requestedTools.push("ast_edit");
 		}
-		if (["hindsight", "prometheus-memory"].includes(session.settings.get("memory.backend") ?? "")) {
+		if (["hindsight", "mnemopi"].includes(session.settings.get("memory.backend") ?? "")) {
 			for (const name of ["recall", "retain", "reflect"]) {
 				if (!requestedTools.includes(name)) requestedTools.push(name);
 			}
 		}
+		// Auto-learn tools are gated by `autolearn.enabled` but, like the memory
+		// tools above, must also be force-included into an explicit requestedTools
+		// list so a restricted top-level session whose controller/guidance is
+		// active still exposes the tools the nudge points at. Gated to top-level
+		// (taskDepth 0): the controller only runs there, so a subagent's explicit
+		// tool whitelist must never be silently widened with write-capable tools.
+		if (session.settings.get("autolearn.enabled") && (session.taskDepth ?? 0) === 0) {
+			if (!requestedTools.includes("manage_skill")) requestedTools.push("manage_skill");
+			if (
+				["hindsight", "mnemopi", "local"].includes(session.settings.get("memory.backend") ?? "") &&
+				!requestedTools.includes("learn")
+			) {
+				requestedTools.push("learn");
+			}
+		}
 	}
 	// Resolve effective tool discovery mode.
 	// tools.discoveryMode controls the new modes; mcp.discoveryMode remains a back-compat alias for "mcp-only".
@@ -432,7 +553,7 @@ export async function createTools(session: ToolSession, toolNames?: string[]): P
 	const isToolAllowed = (name: string) => {
 		if (name === "goal") return goalEnabled && goalModeActive;
 		if (name === "lsp") return enableLsp && session.settings.get("lsp.enabled");
-		if (name === "bash") return true;
+		if (name === "bash") return session.settings.get("bash.enabled");
 		if (name === "eval") return allowEval;
 		if (name === "debug") return session.settings.get("debug.enabled");
 		if (name === "todo") return !includeYield && session.settings.get("todo.enabled");
@@ -448,20 +569,20 @@ export async function createTools(session: ToolSession, toolNames?: string[]): P
 		if (name === "search_tool_bm25") return discoveryActive;
 		if (name === "browser") return session.settings.get("browser.enabled");
 		if (name === "checkpoint" || name === "rewind") return session.settings.get("checkpoint.enabled");
-		if (name === "irc") {
-			if (!session.settings.get("irc.enabled")) return false;
-			// Main agent only needs `irc` when subagents may run concurrently (async).
-			// In sync mode main blocks on `task`, so peer messaging from main is dead weight.
-			if (!session.settings.get("async.enabled") && session.getAgentId?.() === MAIN_AGENT_ID) return false;
-			return true;
-		}
+		if (name === "irc") return isIrcEnabled(session.settings, session.taskDepth ?? 0);
 		if (name === "retain" || name === "recall" || name === "reflect") {
-			return ["hindsight", "prometheus-memory"].includes(session.settings.get("memory.backend") ?? "");
+			return ["hindsight", "mnemopi"].includes(session.settings.get("memory.backend") ?? "");
+		}
+		if (name === "manage_skill") return session.settings.get("autolearn.enabled") && (session.taskDepth ?? 0) === 0;
+		if (name === "learn") {
+			return (
+				session.settings.get("autolearn.enabled") &&
+				(session.taskDepth ?? 0) === 0 &&
+				["hindsight", "mnemopi", "local"].includes(session.settings.get("memory.backend") ?? "")
+			);
 		}
 		if (name === "task") {
-			const maxDepth = session.settings.get("task.maxRecursionDepth") ?? 2;
-			const currentDepth = session.taskDepth ?? 0;
-			return maxDepth < 0 || currentDepth < maxDepth;
+			return canSpawnAtDepth(session.settings.get("task.maxRecursionDepth") ?? 2, session.taskDepth ?? 0);
 		}
 		return true;
 	};

package/src/tools/inspect-image-renderer.ts

@@ -2,8 +2,8 @@ import type { Component } from "@prometheus-ai/tui";
 import { Text } from "@prometheus-ai/tui";
 import type { RenderResultOptions } from "../extensibility/custom-tools/types";
 import type { Theme } from "../modes/theme/theme";
-import { renderStatusLine } from "../tui";
-import { formatExpandHint, replaceTabs, shortenPath, truncateToWidth } from "./render-utils";
+import { framedBlock, renderStatusLine } from "../tui";
+import { formatErrorDetail, formatExpandHint, replaceTabs, shortenPath, truncateToWidth } from "./render-utils";
 
 interface InspectImageRenderArgs {
 	path?: string;
@@ -27,17 +27,21 @@ const INSPECT_OUTPUT_COLLAPSED_LINES = 4;
 const INSPECT_OUTPUT_EXPANDED_LINES = 16;
 const INSPECT_OUTPUT_LINE_WIDTH = 120;
 
+function questionLine(question: string, uiTheme: Theme): string {
+	return `${uiTheme.fg("dim", "Question:")} ${uiTheme.fg("accent", truncateToWidth(replaceTabs(question), INSPECT_QUESTION_PREVIEW_WIDTH))}`;
+}
+
 export const inspectImageToolRenderer = {
 	renderCall(args: InspectImageRenderArgs, _options: RenderResultOptions, uiTheme: Theme): Component {
 		const rawPath = args.path ?? "";
 		const pathDisplay = rawPath ? shortenPath(rawPath) : "…";
-		const header = renderStatusLine({ icon: "pending", title: "Inspect Image", description: pathDisplay }, uiTheme);
+		const header = renderStatusLine({ icon: "pending", title: "Inspect", description: pathDisplay }, uiTheme);
 		const question = args.question?.trim();
-		if (!question) {
-			return new Text(header, 0, 0);
-		}
-		const questionLine = ` ${uiTheme.fg("dim", uiTheme.tree.last)} ${uiTheme.fg("dim", "Question:")} ${uiTheme.fg("accent", truncateToWidth(replaceTabs(question), INSPECT_QUESTION_PREVIEW_WIDTH))}`;
-		return new Text(`${header}\n${questionLine}`, 0, 0);
+		// Call is at most a status line plus a one-line question — too small to box.
+		// The container renders a lone Text cleanly with no chrome.
+		if (!question) return new Text(header, 0, 0);
+		const tree = ` ${uiTheme.fg("dim", uiTheme.tree.last)} ${questionLine(question, uiTheme)}`;
+		return new Text(`${header}\n${tree}`, 0, 0);
 	},
 
 	renderResult(
@@ -49,55 +53,80 @@ export const inspectImageToolRenderer = {
 		const details = result.details;
 		const rawPath = details?.imagePath ?? args?.path ?? "";
 		const pathDisplay = rawPath ? shortenPath(rawPath) : "image";
-		const metaParts: string[] = [];
-		if (details?.model) metaParts.push(details.model);
-		if (details?.mimeType) metaParts.push(details.mimeType);
+		const success = !result.isError;
 		const header = renderStatusLine(
-			{
-				icon: result.isError ? "error" : "success",
-				title: "Inspect Image",
-				description: pathDisplay,
-			},
+			success
+				? {
+						iconOverride: uiTheme.styledSymbol("tool.inspectImage", "accent"),
+						title: "Inspect",
+						description: pathDisplay,
+					}
+				: {
+						icon: "error",
+						title: "Inspect",
+						description: pathDisplay,
+					},
 			uiTheme,
 		);
 
-		const lines: string[] = [header];
 		const question = args?.question?.trim();
-		if (question) {
-			lines.push(
-				` ${uiTheme.fg("dim", uiTheme.tree.branch)} ${uiTheme.fg("dim", "Question:")} ${uiTheme.fg("accent", truncateToWidth(replaceTabs(question), INSPECT_QUESTION_PREVIEW_WIDTH))}`,
-			);
-		}
-
 		const outputText = result.content.find(content => content.type === "text")?.text?.trimEnd() ?? "";
-		if (!outputText) {
-			lines.push(uiTheme.fg("dim", "(no output)"));
-			if (metaParts.length > 0) {
-				lines.push("");
-				lines.push(uiTheme.fg("dim", metaParts.join(" · ")));
-			}
-			return new Text(lines.join("\n"), 0, 0);
-		}
 
-		lines.push("");
-		const outputLines = replaceTabs(outputText).split("\n");
-		const maxLines = options.expanded ? INSPECT_OUTPUT_EXPANDED_LINES : INSPECT_OUTPUT_COLLAPSED_LINES;
-		for (const line of outputLines.slice(0, maxLines)) {
-			lines.push(uiTheme.fg("toolOutput", truncateToWidth(line, INSPECT_OUTPUT_LINE_WIDTH)));
+		if (result.isError) {
+			return framedBlock(uiTheme, width => {
+				const bodyLines: string[] = [];
+				if (question) bodyLines.push(questionLine(question, uiTheme));
+				bodyLines.push(formatErrorDetail(outputText || "inspection failed", uiTheme));
+				return {
+					header,
+					sections: [{ lines: bodyLines }],
+					state: "error",
+					borderColor: "error",
+					applyBg: false,
+					width,
+				};
+			});
 		}
 
-		if (outputLines.length > maxLines) {
-			const remaining = outputLines.length - maxLines;
-			const hint = formatExpandHint(uiTheme, options.expanded, true);
-			lines.push(`${uiTheme.fg("dim", `… ${remaining} more lines`)}${hint ? ` ${hint}` : ""}`);
-		}
+		const metaParts: string[] = [];
+		if (details?.model) metaParts.push(details.model);
+		if (details?.mimeType) metaParts.push(details.mimeType);
+		const metaLine = metaParts.length > 0 ? uiTheme.fg("dim", metaParts.join(" · ")) : "";
 
-		if (metaParts.length > 0) {
-			lines.push("");
-			lines.push(uiTheme.fg("dim", metaParts.join(" · ")));
+		// No answer text: nothing worth boxing — keep it to a clean status line
+		// (plus a trailing meta line, when present).
+		if (!outputText) {
+			return new Text(metaLine ? `${header}\n${metaLine}` : header, 0, 0);
 		}
 
-		return new Text(lines.join("\n"), 0, 0);
+		return framedBlock(uiTheme, width => {
+			const bodyLines: string[] = [];
+			if (question) {
+				bodyLines.push(questionLine(question, uiTheme));
+				bodyLines.push("");
+			}
+
+			const outputLines = replaceTabs(outputText).split("\n");
+			const maxLines = options.expanded ? INSPECT_OUTPUT_EXPANDED_LINES : INSPECT_OUTPUT_COLLAPSED_LINES;
+			for (const line of outputLines.slice(0, maxLines)) {
+				bodyLines.push(uiTheme.fg("toolOutput", truncateToWidth(line, INSPECT_OUTPUT_LINE_WIDTH)));
+			}
+			if (outputLines.length > maxLines) {
+				const remaining = outputLines.length - maxLines;
+				const hint = formatExpandHint(uiTheme, options.expanded, true);
+				bodyLines.push(`${uiTheme.fg("dim", `… ${remaining} more lines`)}${hint ? ` ${hint}` : ""}`);
+			}
+
+			return {
+				header,
+				headerMeta: metaLine || undefined,
+				sections: [{ lines: bodyLines }],
+				state: "success",
+				borderColor: "borderMuted",
+				applyBg: false,
+				width,
+			};
+		});
 	},
 	mergeCallAndResult: true,
 };

package/src/tools/inspect-image.ts

@@ -2,9 +2,10 @@ import type { AgentTool, AgentToolContext, AgentToolResult, AgentToolUpdateCallb
 import { instrumentedCompleteSimple, resolveTelemetry } from "@prometheus-ai/agent-core";
 import { type Api, completeSimple, type Model } from "@prometheus-ai/ai";
 import { prompt } from "@prometheus-ai/utils";
-import * as z from "zod/v4";
+import { z } from "zod/v4";
 import { extractTextContent } from "../commit/utils";
-import { expandRoleAlias, resolveModelFromString } from "../config/model-resolver";
+
+import { expandRoleAlias, getModelMatchPreferences, resolveModelFromString } from "../config/model-resolver";
 import inspectImageDescription from "../prompts/tools/inspect-image.md" with { type: "text" };
 import inspectImageSystemPromptTemplate from "../prompts/tools/inspect-image-system.md" with { type: "text" };
 import {
@@ -12,6 +13,7 @@ import {
 	type LoadedImageInput,
 	loadImageInput,
 	MAX_IMAGE_INPUT_BYTES,
+	webpExclusionForModel,
 } from "../utils/image-loading";
 import type { ToolSession } from "./index";
 import { ToolError } from "./tool-errors";
@@ -71,7 +73,7 @@ export class InspectImageTool implements AgentTool<typeof inspectImageSchema, In
 			throw new ToolError("No models available for inspect_image.");
 		}
 
-		const matchPreferences = { usageOrder: this.session.settings.getStorage()?.getModelUsageOrder() };
+		const matchPreferences = getModelMatchPreferences(this.session.settings);
 		const resolvePattern = (pattern: string | undefined): Model<Api> | undefined => {
 			if (!pattern) return undefined;
 			const expanded = expandRoleAlias(pattern, this.session.settings);
@@ -80,8 +82,8 @@ export class InspectImageTool implements AgentTool<typeof inspectImageSchema, In
 
 		const activeModelPattern = this.session.getActiveModelString?.() ?? this.session.getModelString?.();
 		const model =
-			resolvePattern("prometheus/vision") ??
-			resolvePattern("prometheus/default") ??
+			resolvePattern("pi/vision") ??
+			resolvePattern("pi/default") ??
 			resolvePattern(activeModelPattern) ??
 			availableModels[0];
 		if (!model) {
@@ -108,6 +110,7 @@ export class InspectImageTool implements AgentTool<typeof inspectImageSchema, In
 				cwd: this.session.cwd,
 				autoResize: this.session.settings.get("images.autoResize"),
 				maxBytes: MAX_IMAGE_INPUT_BYTES,
+				excludeWebP: webpExclusionForModel(model),
 			});
 		} catch (error) {
 			if (error instanceof ImageInputTooLargeError) {
@@ -136,7 +139,10 @@ export class InspectImageTool implements AgentTool<typeof inspectImageSchema, In
 					},
 				],
 			},
-			{ apiKey, signal },
+			{
+				apiKey: modelRegistry.resolver(model, this.session.getSessionId?.() ?? undefined),
+				signal,
+			},
 			{ telemetry, oneshotKind: "inspect_image", completeImpl: this.completeImageRequest },
 		);