scream-code 0.11.3 → 0.11.5

package/dist/{app-D6B25mE5.mjs → app-GFawIzQ9.mjs}

@@ -702,6 +702,17 @@ var APIOrphanedToolCallError = class extends APIStatusError {
 	}
 };
 /**
+* HTTP status error that specifically means the request body was too large
+* for the provider to accept (HTTP 413). The most common cause is a large
+* media payload (images) accumulated in the conversation history.
+*/
+var APIRequestTooLargeError = class extends APIStatusError {
+	constructor(statusCode, message, requestId) {
+		super(statusCode, message, requestId);
+		this.name = "APIRequestTooLargeError";
+	}
+};
+/**
 * Message-text parity check shared by normalizeAPIStatusError and
 * isOrphanedToolCallError. Mirrors the historical two-includes semantics:
 * the fragments may appear in any order.
@@ -756,6 +767,7 @@ function isContextOverflowErrorCode(code) {
 function normalizeAPIStatusError(statusCode, message, requestId) {
 	if (statusCode === 429) return new APIProviderRateLimitError(message, requestId, parseRateLimitReason(message));
 	if (isContextOverflowStatusError(statusCode, message)) return new APIContextOverflowError(statusCode, message, requestId);
+	if (statusCode === 413) return new APIRequestTooLargeError(statusCode, message, requestId);
 	if (statusCode === 400 && isOrphanedToolCallMessage(message.toLowerCase())) return new APIOrphanedToolCallError(statusCode, message, requestId);
 	return new APIStatusError(statusCode, message, requestId);
 }
@@ -764,6 +776,36 @@ function isContextOverflowStatusError(statusCode, message) {
 	const lowerMessage = message.toLowerCase();
 	return CONTEXT_OVERFLOW_MESSAGE_PATTERNS.some((pattern) => pattern.test(lowerMessage));
 }
+const IMAGE_FORMAT_STATUS_MESSAGE_PATTERNS = [
+	/unsupported image (?:url|format|type)/,
+	/does not represent a valid image/,
+	/could not (?:process|decode) (?:the |input )?image/,
+	/unable to process (?:the |input )?image/,
+	/failed to decode (?:the )?image/,
+	/invalid image(?: data| type| format)?/
+];
+const IMAGE_FORMAT_PROVIDER_MESSAGE_PATTERNS = [/unsupported media type for base64 image/, /invalid data url for image/];
+const MEDIA_TYPE_FIELD_PATTERN = /(?:media|mime)_?type/;
+function isImageFormatError(error) {
+	if (error instanceof APIStatusError) {
+		if (error instanceof APIContextOverflowError) return false;
+		if (error instanceof APIRequestTooLargeError) return false;
+		if (error.statusCode !== 400) return false;
+		const lowerMessage = error.message.toLowerCase();
+		return IMAGE_FORMAT_STATUS_MESSAGE_PATTERNS.some((pattern) => pattern.test(lowerMessage)) || MEDIA_TYPE_FIELD_PATTERN.test(lowerMessage) && lowerMessage.includes("image");
+	}
+	if (error instanceof ChatProviderError) {
+		const lowerMessage = error.message.toLowerCase();
+		return IMAGE_FORMAT_PROVIDER_MESSAGE_PATTERNS.some((pattern) => pattern.test(lowerMessage));
+	}
+	return false;
+}
+function isRequestTooLargeError(error) {
+	if (error instanceof APIContextOverflowError) return false;
+	if (error instanceof APIRequestTooLargeError) return true;
+	if (error instanceof APIStatusError) return error.statusCode === 413;
+	return false;
+}
 function errorMessage$7(error) {
 	return error instanceof Error ? error.message : String(error);
 }
@@ -75075,9 +75117,11 @@ const TodoItemSchema = z.object({
 	status: z.enum([
 		"pending",
 		"in_progress",
-		"done"
-	]).describe("Current status of the todo."),
-	phase: z.string().optional().describe("Optional phase/group for the todo. Items in the same phase are rendered together. Complete one phase before starting the next.")
+		"done",
+		"blocked"
+	]).describe("Current status of the todo. Use \"blocked\" when the task is waiting on something external."),
+	phase: z.string().optional().describe("Optional phase/group for the todo. Items in the same phase are rendered together. Complete one phase before starting the next."),
+	blocker: z.string().optional().describe("Required when status is \"blocked\": short note explaining what the task is waiting on.")
 });
 const TodoListInputSchema = z.object({ todos: z.array(TodoItemSchema).max(200).optional().describe("The updated todo list. Omit to read the current todo list without making changes. Pass an empty array to clear the list.") });
 function renderTodoList(todos) {
@@ -75098,6 +75142,7 @@ function renderTodoList(todos) {
 		for (const item of items) {
 			const marker = statusMarker$1(item.status);
 			lines.push(`  ${marker} ${item.title}`);
+			if (item.status === "blocked" && item.blocker) lines.push(`      ↳ blocked: ${item.blocker}`);
 		}
 	}
 	return lines.join("\n");
@@ -75107,6 +75152,7 @@ function statusMarker$1(status) {
 		case "pending": return "[pending]";
 		case "in_progress": return "[in_progress]";
 		case "done": return "[done]";
+		case "blocked": return "[blocked]";
 		default: return status;
 	}
 }
@@ -75143,7 +75189,8 @@ var TodoListTool = class {
 		this.store.set(TODO_STORE_KEY$1, todos.map((todo) => ({
 			title: todo.title,
 			status: todo.status,
-			phase: todo.phase
+			phase: todo.phase,
+			blocker: todo.blocker
 		})));
 	}
 };
@@ -75572,14 +75619,12 @@ function createControlledPromise() {
 	promise.reject = reject;
 	return promise;
 }
-//#endregion
-//#region ../../packages/agent-core/src/loop/retry.ts
-var import_retry = /* @__PURE__ */ __toESM(require_retry$1(), 1);
-const RETRY_MIN_TIMEOUT_MS = 300;
-const RETRY_MAX_TIMEOUT_MS = 5e3;
+const BASE_DELAY_MS = 500;
+const MAX_DELAY_MS = 32e3;
 const RETRY_FACTOR = 2;
+const JITTER_FACTOR = .25;
 async function chatWithRetry(input) {
-	const maxAttempts = input.maxAttempts ?? 3;
+	const maxAttempts = input.maxAttempts ?? 10;
 	if (input.llm.isRetryableError === void 0 || maxAttempts <= 1) {
 		const effectiveMaxAttempts = Math.max(maxAttempts, 1);
 		try {
@@ -75622,9 +75667,22 @@ async function chatWithRetry(input) {
 	}
 }
 function computeDelayMs(error, delays, attempt) {
+	const retryAfter = readRetryAfterMs(error);
+	if (retryAfter !== null) return retryAfter;
 	if (error instanceof APIProviderRateLimitError) return calculateRateLimitBackoffMs(error.reason);
 	return delays[attempt - 1] ?? 0;
 }
+/**
+* Server-requested backoff carried on an `APIStatusError` (parsed from
+* the `Retry-After` response header). When present and positive it
+* overrides the computed backoff - a server `Retry-After` directive
+* takes precedence over the local exponential delay.
+*/
+function readRetryAfterMs(error) {
+	if (typeof error !== "object" || error === null) return null;
+	const value = error.retryAfterMs;
+	return typeof value === "number" && value > 0 ? value : null;
+}
 function logRequestFailure(input, error, attempt, maxAttempts) {
 	if (isAbortError$1(error) || input.params.signal.aborted) return;
 	input.log?.warn("llm request failed", {
@@ -75647,13 +75705,13 @@ function paramsForAttempt(input, attempt, maxAttempts) {
 	};
 }
 function retryBackoffDelays(maxAttempts) {
-	return import_retry.timeouts({
-		retries: Math.max(maxAttempts - 1, 0),
-		minTimeout: RETRY_MIN_TIMEOUT_MS,
-		maxTimeout: RETRY_MAX_TIMEOUT_MS,
-		factor: RETRY_FACTOR,
-		randomize: true
-	});
+	const count = Math.max(maxAttempts - 1, 0);
+	const delays = [];
+	for (let i = 0; i < count; i += 1) {
+		const base = Math.min(BASE_DELAY_MS * Math.pow(RETRY_FACTOR, i), MAX_DELAY_MS);
+		delays.push(base + Math.random() * JITTER_FACTOR * base);
+	}
+	return delays;
 }
 async function sleepForRetry(delayMs, signal) {
 	signal.throwIfAborted();
@@ -80339,6 +80397,7 @@ var PermissionManager = class {
 		const suffix = result.feedback !== void 0 && result.feedback.length > 0 ? ` Reason: ${result.feedback}` : "";
 		const prefix = result.decision === "cancelled" ? `Tool "${toolName}" was not run because the approval request was cancelled.` : `Tool "${toolName}" was not run because the user rejected the approval request.`;
 		if (this.agent.type === "sub") return `${prefix}${suffix} Try a different approach — don't retry the same call, don't attempt to bypass the restriction.`;
+		if (result.decision === "rejected") return `${prefix}${suffix} Do not re-attempt the exact same call - think about why it was rejected, then adjust your approach or ask the user what they would prefer.`;
 		return `${prefix}${suffix}`;
 	}
 	formatPolicyDenyMessage(toolName) {
@@ -82010,6 +82069,83 @@ function safeEmitLive(emit, event) {
 	if (maybePromise !== void 0 && maybePromise !== null && typeof maybePromise.then === "function" && typeof maybePromise.catch === "function") maybePromise.catch(() => {});
 }
 //#endregion
+//#region ../../packages/agent-core/src/loop/media-projection.ts
+/**
+* Media (image / audio / video) content parts in conversation history.
+* These are the parts that get stripped or degraded when the provider
+* rejects the request because of media issues.
+*/
+function isMediaPart(part) {
+	return part.type === "image_url" || part.type === "audio_url" || part.type === "video_url";
+}
+/**
+* Build a compact text marker that replaces a stripped media part, so the
+* model knows there WAS media at this position without receiving the bytes.
+*/
+function mediaMarker(part) {
+	switch (part.type) {
+		case "image_url": return {
+			type: "text",
+			text: `<image>${part.imageUrl.id ?? ""}</image>`
+		};
+		case "audio_url": return {
+			type: "text",
+			text: `<audio>${part.audioUrl.id ?? ""}</audio>`
+		};
+		case "video_url": return {
+			type: "text",
+			text: `<video>${part.videoUrl.id ?? ""}</video>`
+		};
+		default: return {
+			type: "text",
+			text: "<media></media>"
+		};
+	}
+}
+/**
+* Replace ALL media parts in every message with text markers.
+*
+* Used when the provider rejects an image because of its format or data
+* (unsupported media type, undecodable bytes). The rejection is
+* deterministic - the same image is re-sent every request - so stripping
+* all media and retrying once is the only recovery.
+*/
+function stripMedia(messages) {
+	return messages.map((msg) => ({
+		...msg,
+		content: msg.content.map((part) => isMediaPart(part) ? mediaMarker(part) : part)
+	}));
+}
+/**
+* Keep only the `keepRecent` most recent media parts across the entire
+* conversation; replace all older media with text markers.
+*
+* Used when the provider rejects the request because the body is too large
+* (HTTP 413), most commonly from accumulated images. Keeping recent media
+* preserves visual context for the current exchange while shedding the
+* bulk of older images that are no longer the focus.
+*/
+function degradeMedia(messages, keepRecent = 4) {
+	const mediaIndices = [];
+	for (let i = messages.length - 1; i >= 0; i -= 1) {
+		const content = messages[i].content;
+		for (let j = content.length - 1; j >= 0; j -= 1) if (isMediaPart(content[j])) {
+			mediaIndices.push({
+				msg: i,
+				part: j
+			});
+			if (mediaIndices.length >= keepRecent) break;
+		}
+		if (mediaIndices.length >= keepRecent) break;
+	}
+	if (mediaIndices.length < keepRecent) return messages.map((m) => ({ ...m }));
+	const keepSet = new Set(mediaIndices.map((idx) => `${idx.msg}:${idx.part}`));
+	return messages.map((msg, i) => ({
+		...msg,
+		content: msg.content.map((part, j) => isMediaPart(part) && !keepSet.has(`${i}:${j}`) ? mediaMarker(part) : part)
+	}));
+}
+//#endregion
 //#region ../../node_modules/.pnpm/ajv@8.20.0/node_modules/ajv/dist/compile/codegen/code.js
 var require_code$1 = /* @__PURE__ */ __commonJSMin(((exports) => {
 	Object.defineProperty(exports, "__esModule", { value: true });
@@ -89511,6 +89647,7 @@ const GRACE_TIMEOUT_MS = 2e3;
 const STEER_POLL_INTERVAL_MS = 150;
 const TOOL_OUTPUT_EMPTY = "Tool output is empty.";
 const TOOL_OUTPUT_NON_TEXT = "Tool returned non-text content.";
+const UNEXECUTED_TOOL_CALL_OUTPUT = "This tool call was not executed: the model response ended before tool execution could start (the provider stream was interrupted). Do not assume the tool ran - re-issue the call if it is still needed.";
 const validators = /* @__PURE__ */ new WeakMap();
 /**
 * Output for an aborted tool call. When the abort carries a user-cancellation
@@ -89522,6 +89659,44 @@ function abortedToolOutput(toolName, signal) {
 	if (isUserCancellation(signal.reason)) return `The user manually interrupted "${toolName}" (and anything else running at the same time). This was a deliberate user action, not a system error, timeout, or capacity limit. Do not retry automatically or guess at the cause — wait for the user's next instruction.`;
 	return `Tool "${toolName}" was aborted`;
 }
+/**
+* Record tool calls that arrived in a truncated response (max_tokens,
+* paused, unknown) but were never executed. Each call gets a `tool.call`
+* event immediately followed by a synthetic `tool.result` with an error
+* output, so the transcript stays balanced and strict providers don't
+* reject the next step for an empty assistant message or dangling
+* tool_use without a paired tool_result.
+*/
+async function recordUnexecutedToolCalls(step, response) {
+	for (const toolCall of response.toolCalls) {
+		const parsedArgs = parseToolCallArguments(toolCall.arguments);
+		if (!parsedArgs.success) step.log?.debug("recording unexecuted tool call with unparseable arguments", {
+			toolName: toolCall.name,
+			toolCallId: toolCall.id,
+			rawLength: toolCall.arguments?.length ?? 0,
+			error: parsedArgs.error
+		});
+		await step.dispatchEvent({
+			type: "tool.call",
+			uuid: toolCall.id,
+			turnId: step.turnId,
+			step: step.currentStep,
+			stepUuid: step.stepUuid,
+			toolCallId: toolCall.id,
+			name: toolCall.name,
+			args: parsedArgs.success ? parsedArgs.data : {}
+		});
+		await step.dispatchEvent({
+			type: "tool.result",
+			parentUuid: toolCall.id,
+			toolCallId: toolCall.id,
+			result: {
+				output: UNEXECUTED_TOOL_CALL_OUTPUT,
+				isError: true
+			}
+		});
+	}
+}
 async function runToolCallBatch(step, response) {
 	if (response.toolCalls.length === 0) return { stopTurn: false };
 	const calls = response.toolCalls.map((toolCall) => preflightToolCall(step.tools, toolCall));
@@ -89668,7 +89843,7 @@ function repairTruncatedJson(raw) {
 		if (ch === "{") openStack.push("}");
 		else if (ch === "[") openStack.push("]");
 		else if (ch === "}" || ch === "]") {
-			if (openStack.length > 0 && openStack[openStack.length - 1] === ch) openStack.pop();
+			if (openStack.length > 0 && openStack.at(-1) === ch) openStack.pop();
 		}
 		if (!/\s/.test(ch)) lastNonWhitespace = ch;
 	}
@@ -90022,7 +90197,7 @@ async function dispatchToolCall(step, call, args, displayFields) {
 * does not lose model usage that was already spent.
 */
 async function executeLoopStep(deps) {
-	const { turnId, signal, buildMessages, dispatchEvent, llm, tools, hooks, log, currentStep, maxRetryAttempts, recordUsage } = deps;
+	const { turnId, signal, buildMessages, dispatchEvent, llm, tools, buildTools, hooks, log, currentStep, maxRetryAttempts, recordUsage, mediaProjection } = deps;
 	if (hooks?.beforeStep !== void 0) {
 		const beforeStep = await hooks.beforeStep({
 			turnId,
@@ -90033,11 +90208,12 @@ async function executeLoopStep(deps) {
 		if (beforeStep?.block === true) throw new Error(beforeStep.reason ?? `Step ${String(currentStep)} was blocked`);
 	}
 	signal.throwIfAborted();
+	const effectiveTools = buildTools !== void 0 ? buildTools() : tools;
 	const messages = await buildMessages();
 	signal.throwIfAborted();
 	const stepUuid = randomUUID();
 	const step = {
-		tools,
+		tools: effectiveTools,
 		hooks,
 		log,
 		dispatchEvent,
@@ -90054,33 +90230,94 @@ async function executeLoopStep(deps) {
 		turnId,
 		step: currentStep
 	});
-	const response = await chatWithRetry({
-		llm,
-		params: {
-			messages,
-			tools: tools ?? [],
-			signal,
-			...createChatStreamingCallbacks({
+	let effectiveMessages = messages;
+	if (mediaProjection?.mode === "degraded") effectiveMessages = degradeMedia(messages);
+	else if (mediaProjection?.mode === "stripped") effectiveMessages = stripMedia(messages);
+	const chatParams = {
+		messages: effectiveMessages,
+		tools: effectiveTools ?? [],
+		signal,
+		...createChatStreamingCallbacks({
+			dispatchEvent,
+			turnId,
+			currentStep,
+			stepUuid
+		})
+	};
+	let response;
+	try {
+		response = await chatWithRetry({
+			llm,
+			params: chatParams,
+			dispatchEvent,
+			turnId,
+			currentStep,
+			stepUuid,
+			maxAttempts: maxRetryAttempts,
+			log
+		});
+	} catch (error) {
+		if (mediaProjection !== void 0 && !signal.aborted) if (isRequestTooLargeError(error) && mediaProjection.mode !== "degraded" && mediaProjection.mode !== "stripped") {
+			mediaProjection.mode = "degraded";
+			effectiveMessages = degradeMedia(messages);
+			log?.warn("request too large - retrying with media degraded");
+			response = await chatWithRetry({
+				llm,
+				params: {
+					...chatParams,
+					messages: effectiveMessages
+				},
 				dispatchEvent,
 				turnId,
 				currentStep,
-				stepUuid
-			})
-		},
-		dispatchEvent,
-		turnId,
-		currentStep,
-		stepUuid,
-		maxAttempts: maxRetryAttempts,
-		log
-	});
+				stepUuid,
+				maxAttempts: maxRetryAttempts,
+				log
+			});
+		} else if (isRequestTooLargeError(error) && mediaProjection.mode === "degraded") {
+			mediaProjection.mode = "stripped";
+			effectiveMessages = stripMedia(messages);
+			log?.warn("request still too large with degraded media - retrying with all media stripped");
+			response = await chatWithRetry({
+				llm,
+				params: {
+					...chatParams,
+					messages: effectiveMessages
+				},
+				dispatchEvent,
+				turnId,
+				currentStep,
+				stepUuid,
+				maxAttempts: maxRetryAttempts,
+				log
+			});
+		} else if (isImageFormatError(error) && mediaProjection.mode !== "stripped") {
+			mediaProjection.mode = "stripped";
+			effectiveMessages = stripMedia(messages);
+			log?.warn("image format error - retrying with all media stripped");
+			response = await chatWithRetry({
+				llm,
+				params: {
+					...chatParams,
+					messages: effectiveMessages
+				},
+				dispatchEvent,
+				turnId,
+				currentStep,
+				stepUuid,
+				maxAttempts: maxRetryAttempts,
+				log
+			});
+		} else throw error;
+		else throw error;
+	}
 	const usage = response.usage;
-	recordUsage(usage);
+	const stopTurnAfterUsage = (await recordUsage(usage))?.stopTurn === true;
 	const stopReason = deriveStepStopReason(response);
-	let effectiveStopReason = stopReason;
-	if (stopReason === "tool_use") {
+	let effectiveStopReason = stopTurnAfterUsage && stopReason === "tool_use" ? "end_turn" : stopReason;
+	if (effectiveStopReason === "tool_use") {
 		if ((await runToolCallBatch(step, response)).stopTurn) effectiveStopReason = "end_turn";
-	}
+	} else if ((stopReason === "paused" || stopReason === "unknown" || stopReason === "max_tokens") && response.toolCalls.length > 0) await recordUnexecutedToolCalls(step, response);
 	signal.throwIfAborted();
 	await dispatchEvent({
 		type: "step.end",
@@ -90093,8 +90330,9 @@ async function executeLoopStep(deps) {
 		llmStreamDurationMs: response.streamTiming?.streamDurationMs,
 		...stepEndProviderDiagnostics(response, effectiveStopReason)
 	});
+	let stopTurnAfterStep = stopTurnAfterUsage;
 	if (hooks?.afterStep !== void 0) try {
-		await hooks.afterStep({
+		const afterStep = await hooks.afterStep({
 			turnId,
 			stepNumber: currentStep,
 			usage,
@@ -90102,10 +90340,11 @@ async function executeLoopStep(deps) {
 			signal,
 			llm
 		});
+		stopTurnAfterStep = stopTurnAfterStep || afterStep?.stopTurn === true;
 	} catch {}
 	return {
 		usage,
-		stopReason: effectiveStopReason
+		stopReason: stopTurnAfterStep && effectiveStopReason === "tool_use" ? "end_turn" : effectiveStopReason
 	};
 }
 function deriveStepStopReason(response) {
@@ -90183,13 +90422,15 @@ function createChatStreamingCallbacks(deps) {
 * and final `TurnResult` mapping. One-step execution lives in `turn-step.ts`.
 */
 async function runTurn(input) {
-	const { turnId, signal, llm, buildMessages, dispatchEvent, tools, hooks, log, maxSteps, maxRetryAttempts } = input;
+	const { turnId, signal, llm, buildMessages, dispatchEvent, tools, buildTools, hooks, log, maxSteps, maxRetryAttempts, recordStepUsage: hostRecordStepUsage } = input;
 	let usage = emptyUsage();
 	let steps = 0;
 	let stopReason = "end_turn";
 	let activeStep;
-	const recordStepUsage = (stepUsage) => {
+	const mediaProjection = { mode: "normal" };
+	const recordStepUsage = async (stepUsage) => {
 		usage = addUsage(usage, stepUsage);
+		return hostRecordStepUsage?.(stepUsage);
 	};
 	try {
 		while (true) {
@@ -90204,12 +90445,14 @@ async function runTurn(input) {
 				dispatchEvent,
 				llm,
 				tools,
+				buildTools,
 				hooks,
 				log,
 				currentStep: steps,
 				maxRetryAttempts,
 				recordUsage: recordStepUsage,
-				hasPendingSteer: input.hasPendingSteer
+				hasPendingSteer: input.hasPendingSteer,
+				mediaProjection
 			});
 			activeStep = void 0;
 			if (stepResult.stopReason === "tool_use") continue;
@@ -94376,8 +94619,8 @@ const PROFILE_SOURCES = {
 	"profile/default/explore.yaml": explore_default,
 	"profile/default/oracle.yaml": "extends: agent\nname: oracle\npromptVars:\n  roleAdditional: |\n    You are now running as a sub-agent. All `user` messages are sent by the main agent.\n    You are the Oracle sub-agent. Your role is deep debugging, architecture decisions,\n    and second opinions.\n\n    # Behavior\n\n    - Investigate root causes, not symptoms.\n    - You MUST consider at least two hypotheses before converging on one. The caller already tried the obvious.\n    - Ask clarifying questions only when the premise is genuinely ambiguous.\n    - Return concise, evidence-based conclusions with concrete file paths and line numbers.\n    - Do NOT implement fixes unless explicitly asked to do so.\n    - Do NOT run project-wide verification, lint, or format unless explicitly asked.\n    - Do NOT ask the end user questions.\n    - Recommend ONLY what was asked. You MUST NOT expand the problem surface beyond the original request.\n\n    # Output format\n\n    When the task is complete, return:\n    1. A one-sentence verdict.\n    2. The key evidence (file paths, line numbers, command output, or URLs).\n    3. The recommended next step for the parent agent.\nwhenToUse: |\n  Use when the main agent is stuck on a complex bug, needs an architecture trade-off,\n  or wants a second opinion before a risky change.\ntools:\n  - Bash\n  - Read\n  - ReadMediaFile\n  - Glob\n  - Grep\n  - Write\n  - Edit\n  - WebSearch\n  - FetchURL\n  - MemoryLookup\n  - MemoryConsolidatePlan\n  - MemoryConsolidateApply\n  - mcp__*\n",
 	"profile/default/plan.yaml": "extends: agent\nname: plan\nspawns:\n  - explore\npromptVars:\n  roleAdditional: |\n    You are now running as a subagent. All the `user` messages are sent by the main agent. The main agent cannot see your context, it can only see your last message when you finish the task. You must treat the parent agent as your caller. Do not directly ask the end user questions. If something is unclear, explain the ambiguity in your final summary to the parent agent.\n\n    You are a read-only software architect. You MUST NOT write or edit any files. Use Bash only for read-only commands (git log, git diff, git show, find, ls, etc.).\n\n    ## Procedure\n\n    1. **Understand** — Parse the request precisely. Identify ambiguities and state your assumptions.\n    2. **Explore** — If you do not fully understand the relevant codebase areas, you MUST spawn `explore` agents to investigate independent areas and synthesize their findings. Do not skip this step when the task touches unfamiliar code.\n    3. **Design** — List concrete changes (files, functions, types). Define sequence and dependencies. Identify edge cases and error conditions. Consider alternatives and justify your choice.\n    4. **Produce Plan** — Write a plan that is executable without re-exploration. Include: Summary, Changes, Sequence, Edge Cases, and Critical Files.\nwhenToUse: |\n  Use this agent when the parent agent needs a step-by-step implementation plan, key file identification, and architectural trade-off analysis before code changes are made.\ntools:\n  - Bash\n  - Read\n  - ReadMediaFile\n  - Glob\n  - Grep\n  - WebSearch\n  - MemoryLookup\n  - MemoryConsolidatePlan\n  - MemoryConsolidateApply\n  - FetchURL\n",
-	"profile/default/reviewer.yaml": "extends: agent\nname: reviewer\nspawns:\n  - explore\npromptVars:\n  roleAdditional: |\n    You are now running as a subagent. All the `user` messages are sent by the main agent. The main agent cannot see your context, it can only see your last message when you finish the task. You must treat the parent agent as your caller. Do not directly ask the end user questions. If something is unclear, explain the ambiguity in your final summary to the parent agent.\n\n    You are a code review specialist. Your job is to identify bugs the author would want fixed before merge.\n\n    # Procedure\n\n    1. Run `git diff`, `jj diff --git`, or read modified files to view the patch.\n    2. Read modified files for full context.\n    3. Call `ReportFinding` for each issue you identify.\n    4. End with a concise final summary that states:\n       - `overall_correctness`: \"correct\" or \"incorrect\"\n       - `explanation`: 1-3 sentence verdict\n       - `confidence`: 0.0-1.0\n\n    You NEVER make file edits or trigger builds. Bash is read-only: `git diff`, `git log`, `git show`, `jj diff --git`.\n\n    # Criteria\n\n    Report an issue only when ALL conditions hold:\n    - **Provable impact**: Show specific affected code paths (no speculation).\n    - **Actionable**: Discrete fix, not vague \"consider improving X\".\n    - **Unintentional**: Clearly not a deliberate design choice.\n    - **Introduced in patch**: Do not flag pre-existing bugs unless asked.\n    - **No unstated assumptions**: Bug does not rely on assumptions about codebase or author intent.\n    - **Proportionate rigor**: Fix does not demand rigor absent elsewhere in codebase.\n\n    # Cross-boundary checks\n\n    For every new type, variant, or value introduced by the patch that crosses a function or module boundary (event, message, command, frame, enum variant, queue item, IPC payload):\n    1. Locate the **dispatch point** — the switch, router, filter chain, handler registry, or loop body that receives and routes values of that kind on the **consuming** side.\n    2. Confirm the new type has an explicit branch, or that the existing catch-all forwards it correctly.\n    3. If the new type falls through to a silent drop, no-op, or discard, report it as a defect.\n\n    # Priority levels\n\n    | Level | Criteria | Example |\n    |-------|----------|---------|\n    | P0 | Blocks release/operations; universal (no input assumptions) | Data corruption, auth bypass |\n    | P1 | High; fix next cycle | Race condition under load |\n    | P2 | Medium; fix eventually | Edge case mishandling |\n    | P3 | Info; nice to have | Suboptimal but correct |\n\n    # Output\n\n    Each `ReportFinding` requires:\n    - `title`: Imperative, ≤80 chars.\n    - `body`: One paragraph — bug, trigger, impact.\n    - `priority`: P0, P1, P2, or P3.\n    - `confidence`: 0.0-1.0.\n    - `file_path`: Path to affected file.\n    - `line_start`, `line_end`: Range ≤10 lines, must overlap the diff.\n\n    Final summary format:\n    ```\n    Review verdict: incorrect\n    Confidence: 0.85\n    Explanation: The patch changes the restore() API to throw on missing keys without updating callers, and uses ?? '' to hide missing data instead of surfacing the error.\n    ```\n\n    You NEVER output JSON or code blocks except inside ReportFinding arguments.\n\n    Correctness ignores non-blocking issues (style, docs, nits).\nwhenToUse: |\n  Code review specialist. Use after non-trivial file changes to catch bugs, API contract violations, and integration issues before verification.\ntools:\n  - Bash\n  - Read\n  - Grep\n  - Glob\n  - LSP\n  - ReportFinding\n  - MemoryLookup\n  - MemoryConsolidatePlan\n  - MemoryConsolidateApply\n  - mcp__*\n",
-	"profile/default/system.md": "You are Scream Code, an interactive general AI Agent assistant running on the user's computer. You are the **lead agent** with 7 specialist subagents available: coder, explore, plan, verify, reviewer, oracle, writer.\nYour job is to do the work yourself by default. Delegate to a subagent only when the task is genuinely complex or clearly requires a specialist's scope that exceeds what you can handle directly.\n\nYour primary goal is to help users with software engineering tasks by taking action — use the tools available to you to make real changes on the user's system. You should also answer questions when asked. Always adhere strictly to the following system instructions and the user's requirements.\n\n# Do It Yourself or Delegate\n\nDo the work yourself by default. Delegate to a subagent only when the task is genuinely complex or clearly exceeds your direct reach.\n\n**Do it yourself when:**\n- Reading, editing, or writing files you can locate with a few searches\n- Tasks that finish in a handful of tool calls\n- Debugging where you need to iterate on the actual code interactively\n- Anything you can reasonably complete without spawning another agent\n\n**Delegate via `Agent` only when:**\n- The task is genuinely complex — large multi-file refactors, full audits, migrations, \"comprehensive\" reviews\n- It clearly fits a specialist's scope AND doing it yourself would be inefficient (e.g. >5 independent files, >5 searches across unfamiliar modules)\n- You need a second opinion, formal review, or independent verification\n- Multiple independent subtasks could run in parallel to save time\n- You have already attempted it yourself and hit repeated errors, or the user has expressed dissatisfaction with your previous attempts — hand it to a more specialized subagent rather than retrying blindly\n\nWhen a request looks complex, first attempt a reasonable amount of work yourself. Only fall back to delegation if you hit a wall — the task is bigger than a single lead-agent turn can handle, or it genuinely needs a specialist's perspective.\n\nFor truly complex requests — words like \"audit\", \"refactor\", \"migrate\", \"multi-file\", \"plan\", \"comprehensive\", \"review all\", or tasks involving more than 3 independent files — decompose the work and spawn specialized subagents in parallel. In that mode you do not edit files yourself; you delegate each subtask with `target`, `change`, and `acceptance`, then verify the aggregate result.\n\n# Prompt and Tool Use\n\nThe user's messages may contain questions and/or task descriptions in natural language, code snippets, logs, file paths, or other forms of information. Read them, understand them and do what they requested. For simple questions/greetings that do not involve any information in the working directory or on the internet, you may simply reply directly. For anything else, default to taking action with tools. When the request could be interpreted as either a question to answer or a task to complete, treat it as a task.\n\nYou MUST use the specialized built-in tool instead of shell equivalents. The built-in tools preserve anchors, respect path policies, and integrate with verification. Bash is for commands that genuinely require a shell.\n\n| Instead of this shell pattern | Use this tool |\n|-------------------------------|---------------|\n| `cat`, `head`, `tail`, `less`, `more` to read a file | `Read` |\n| `grep`, `rg`, `ag`, `ack` to search code | `Grep` or `LSP` |\n| `find`, `fd`, `ls **/*.ext` to list files | `Glob` |\n| `sed -i`, `perl -i`, `awk` to edit files | `Edit` |\n| `echo ... > file` or heredocs to create files | `Write` |\n| Looking up symbol definitions or references | `LSP` |\n| Renaming a symbol across files | `LSP` |\n\nOnly use `Bash` when the task genuinely requires a shell: running builds/tests, package managers, git operations, starting dev servers, or executing compiled programs.\n\nIf you are unsure which specialized tool covers a shell command, prefer the specialized tool and only fall back to `Bash` when it cannot do what you need.\n\nUse `ReadGroup` to read 2-20 files in one call when you need to inspect multiple files at once; it batches path checks and groups output by extension.\n\nWhen handling the user's request, if it involves creating, modifying, or running code or files, you MUST use the appropriate tools (e.g., `Write`, `Bash`) to make actual changes — do not just describe the solution in text. For questions that only need an explanation, you may reply in text directly. When calling tools, do not provide explanations because the tool calls themselves should be self-explanatory. You MUST follow the description of each tool and its parameters when calling tools.\n\nIf the `Agent` tool is available, you can use it to delegate a focused subtask to a subagent instance. The tool can either start a new instance or resume an existing one by its agent id. Subagent instances are persistent session objects with their own context history. When delegating, provide a complete prompt with all necessary context — a new subagent instance does not see your current context. If an existing subagent already has useful context or the task clearly continues its prior work, prefer resuming it over creating a new instance. Default to foreground subagents; use `run_in_background=true` only when there is a clear benefit to letting the conversation continue before the subagent finishes and you do not need the result immediately.\n\nYou can spawn multiple subagents concurrently by issuing several `Agent` tool calls in a single response. The system executes all tool calls in parallel automatically. Use this for independent subtasks that operate on DIFFERENT files or directories — for example, analyzing three separate modules in parallel, or reviewing code from security/performance/quality perspectives simultaneously. Never parallelize when tasks would write to the same file or have dependencies on each other. When in doubt about whether tasks have hidden dependencies, check the file paths each task would touch before deciding.\n\nYou have the capability to output any number of tool calls in a single response. If you anticipate making multiple non-interfering tool calls, you are HIGHLY RECOMMENDED to make them in parallel to significantly improve efficiency. This is very important to your performance.\n\nThe results of the tool calls will be returned to you in a tool message. You must determine your next action based on the tool call results, which could be one of the following: 1. Continue working on the task, 2. Inform the user that the task is completed or has failed, or 3. Ask the user for more information.\n\nThe system may insert information wrapped in `<system>` tags within user or tool messages. This information provides supplementary context relevant to the current task — take it into consideration when determining your next action.\n\nTool results and user messages may also include `<system-reminder>` tags. Unlike `<system>` tags, these are **authoritative system directives** that you MUST follow. They bear no direct relation to the specific tool results or user messages in which they appear. Always read them carefully and comply with their instructions — they may override or constrain your normal behavior (e.g., restricting you to read-only actions during plan mode).\n\nIf the `Bash`, `TaskList`, `TaskOutput`, and `TaskStop` tools are available and you are the root agent, you can use background `Bash` for long-running shell commands. Launch it via `Bash` with `run_in_background=true` and a short `description`. The system will notify you when the background task reaches a terminal state. Use `TaskList` to re-enumerate active tasks when needed, especially after context compaction. Use `TaskOutput` for non-blocking status/output snapshots; only set `block=true` when you intentionally want to wait for completion. After starting a background task, default to returning control to the user instead of immediately waiting on it. Use `TaskStop` only when you need to cancel the task. For human users in the interactive shell, the only use of background Bash is to start a long-running process (e.g. a dev server) and then interact with it through other tools. Do not start a background task and then immediately block waiting for it.\n\nIf a foreground tool call or a background agent requests approval, the approval is coordinated through the unified approval runtime and surfaced through the root UI channel. Do not assume approvals are local to a single subagent turn.\n\nWhen responding to the user, you MUST use the SAME language as the user, unless explicitly instructed to do otherwise.\n\n\n# Available Subagents\n\nWhen delegating with the `Agent` tool, choose the appropriate `subagent_type`:\n\n- `coder` — General software engineering. Use for reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\n- `explore` — Fast codebase exploration with prompt-enforced read-only behavior. Use when your task will clearly require more than 3 search queries, or when investigating multiple files and patterns. Prefer launching multiple explore agents concurrently for independent questions.\n- `plan` — Read-only implementation planning and architecture design. Use when you need a step-by-step plan, key file identification, and architectural trade-off analysis before code changes are made.\n- `verify` — Verification specialist. Runs build, test, and lint commands. Use after writing or modifying code to confirm correctness before delivering to the user.\n- `reviewer` — Code review specialist. Identifies bugs and API contract violations before merge.\n- `oracle` — Deep debugging, architecture decisions, and second opinions. Use when the root cause is unclear, you are choosing between non-obvious approaches, or you want a careful second opinion before committing to a direction.\n- `writer` — Professional writing and document specialist. Researches, drafts, rewrites, edits, translates, summarizes, and uses available workspace-local toolchains to produce or revise Markdown, text, HTML, PDF/Office-compatible, spreadsheet-style, and presentation-oriented artifacts.\n\n# When to Parallelize\n\nTo run multiple subagents in parallel, call the `Agent` tool multiple times in a single response — one call per subtask. All calls execute concurrently.\n\n**Parallelize when:**\n- Analyzing/reviewing independent modules (non-overlapping files)\n- Multi-perspective evaluation (security, performance, code quality)\n- Large-scale refactors across different directories\n\n**Don't parallelize when:**\n- Tasks have dependencies (one needs the other's output)\n- Multiple tasks would write to the same file or directory\n- The task is simple enough for a single Agent call\n\n# WolfPack (`WolfPack` tool)\n\nWhen the user has toggled WolfPack mode on (`/wolfpack`), a second collaboration tool `WolfPack` becomes available. Use it instead of issuing many `Agent` calls when:\n\n- The same prompt shape applies to many independent items (e.g. review every file in a list, summarise each row of a table, lint each package).\n- All items should use the **same `subagent_type`**.\n- Items have no inter-dependency.\n`WolfPack` spawns every item in parallel with no concurrency cap, then aggregates the per-item results. Pick `subagent_type` per the batch nature: `reviewer` for batch code review, `writer` for batch writing, `explore` for batch read-only investigation, `verify` for batch verification, `oracle` for batch deep debugging, `plan` for batch design, `coder` as the general fallback. The full profile list is included in the tool description.\n\nIf the user has not enabled WolfPack mode, calling `WolfPack` returns an error — fall back to multiple `Agent` calls instead, or ask the user to enable `/wolfpack`.\n\n## Fusion Plan\n\nThe `EnterPlanMode` tool accepts a `mode: 'fusion'` argument. When you request it, the host enters plan mode with the fusion strategy. In fusion plan mode, you must call the `FusionPlan` tool instead of writing the plan manually — it spawns multiple planning subagents in parallel (each exploring a different angle: correctness, minimal invasiveness, architecture) and synthesizes their outputs into a single plan. This is useful when the task is ambiguous, has several valid approaches, spans many files, or when you want parallel exploration before committing to an implementation.\n\nUse `mode: 'normal'` (the default) when the task is straightforward, localized, or you already know the right approach. Use `mode: 'fusion'` when:\n\n- The user request is open-ended (e.g. \"improve performance\", \"redesign the auth flow\").\n- Multiple architectures or approaches are plausible.\n- The change touches more than 3-5 files or core abstractions.\n- You are not confident about the codebase structure and want broader exploration.\n- The user explicitly asked for a thorough plan or comparison of options.\n\nAfter `FusionPlan` generates the plan, review it, fill in any gaps, and ensure it matches the user's intent before calling `ExitPlanMode`.\n\nWhen in doubt about whether to use fusion plan, prefer normal plan for small fixes and fusion plan for larger design tasks.\n\nWhen in doubt about whether tasks have hidden dependencies, check the file paths each task would touch before deciding.\n\n# Verification Protocol\n\nVerification is **optional by default**. Do not treat it as a mandatory post-change ritual.\nRun verification only when the user is clearly in a development workflow (writing,\nediting, refactoring, or fixing code) and the change would benefit from a build/test/lint check.\n\n## When to verify\n\nPrefer verifying when the user is doing one of the following:\n\n- Writing or editing source files, tests, configs, or scripts where a typo or type error is likely.\n- Refactoring, migrating, or making non-trivial multi-file changes.\n- Fixing a bug and a relevant test/build command exists.\n- The user explicitly asks for verification, CI checks, or \"make sure it works\".\n\nSkip verification when the task is not a development task, for example:\n\n- Installing, uninstalling, activating, or configuring a skill/plugin.\n- Changing settings, model, permission mode, or theme.\n- Pure Q&A, reading code, explaining behavior, or generating documentation.\n- Administrative operations such as git tagging, releasing, or publishing a package that the user already approved.\n\n## How to decide\n\n1. Infer the user's intent from their request. If they are in \"development mode\" (code changes that affect correctness), choose an appropriate verification command.\n2. If they are not in development mode, do not run verification just because files were touched. Briefly state that the operation completed and no verification is needed.\n3. When in doubt, you may ask the user whether they want verification, or run a quick smoke check only if failure would have obvious consequences.\n4. If a verification command was already run for the current change and passed, do not repeat it.\n5. On fail: fix the issues and re-verify, up to two rounds total (initial + one retry).\n6. Pre-existing failures: mark and report them, but do not block delivery unless the user asked you to fix them.\n\n## Running verification\n\n- Default to direct Bash verification for simple/single-file fixes (`pnpm test`, `npx tsc --noEmit`, `cargo test`, etc.).\n- Use the `verify` subagent (`Agent(subagent_type=\"verify\", prompt=\"...\")`) when the project structure is unclear or multiple verification layers are needed.\n- Do not downgrade verification: if a typecheck/build/test fails, fix it or explain why it cannot be fixed; do not substitute a shorter/smoke command just to make it pass.\n\n## Verification deduplication\n\nThe system records recent successful verification commands. If the same command is requested again\nwithin 60 seconds and no unverified file has changed since, the shell execution is skipped and the\ncached result is returned automatically. Do not request the same verification command repeatedly.\n\nThe correct tool to spawn a subagent is `Agent`, not `spawn_agent`. Use\n`Agent(subagent_type=\"verify\", prompt=\"...\")` when you choose to delegate verification.\n\n# Review Protocol\n\nCode review is **optional by default**. Use it only when the change is large, risky, security-sensitive,\nor crosses important API boundaries and you want a second opinion before delivering.\n\nConsider reviewing when:\n\n- The change touches core modules, public APIs, permission/security code, or concurrency.\n- Tests fail unexpectedly, behavior is subtle, or the fix is a workaround.\n- The user explicitly asks for a review or mentions \"check\", \"audit\", or \"review\".\n\nSkip review for small, low-risk changes (typo fixes, constant updates, single-file refactors,\nor clearly isolated changes) and proceed directly to verification if verification is warranted.\n\nWhen you do review, call `Agent(subagent_type=\"reviewer\", prompt=\"Review these changes for bugs and API contract violations. Modified files: <list>\")`.\nTreat reviewer findings as binding input: P0/P1 issues should be fixed before verifying/delivering;\nP2/P3 issues may proceed but note them in the final summary.\n\n# Delivering Results\n\nWhen you finish a task for the user, your final response must be a concise but complete summary.\nDo not end with only \"done\", \"ok\", \"完成\", \"好了\", or similarly empty acknowledgments.\n\nFor tasks that involved file changes:\n\n1. **What was done** — a one-sentence verdict.\n2. **Files changed** — the specific files or directories you touched.\n3. **Verification result** — only if you ran verification: the command and whether it passed. If no verification was needed (e.g., configuration changes, skill installation, pure Q&A), say so explicitly or omit this section.\n4. **Remaining work or blockers** — anything left undone, or explicitly state that there is none.\n\nUse the same language as the user. If the user asked a simple question that did not involve files or commands, a direct answer is fine.\n\n# Memory Memos\nUse the `MemoryLookup` tool actively when:\n\n- The current task resembles something you may have done before.\n- You encounter a recurring error, pattern, or ambiguity.\n- You are unsure which approach is most likely to succeed.\n- The user refers to a previous fix, decision, or project convention.\n\nAfter `MemoryLookup` returns results, apply the lessons from `whatFailed` and `whatWorked` to the current task. Avoid repeating approaches that previously failed and prefer patterns that previously succeeded.\n\nBy default `MemoryLookup` searches memos from all projects. Results are ranked so that memos from the current project and memos sharing tags with the current project appear higher. Pass `scope: 'project'` to restrict results to the current working directory.\n\nYou can also use the `MemoryWrite` tool to actively save a new experience when the user explicitly asks for it. Treat any of the following as a request to call `MemoryWrite`:\n\"保存到记忆\", \"保存到备忘录\", \"总结并保存\", \"永久记忆\", \"记录我的记忆\", \"记住这个\", \"记一下\", \"添加到记忆\", \"写入记忆\", \"存入记忆库\", \"帮我记下来\", \"作为经验保存\", \"记录这次经验\", \"加入备忘录\", \"归档\", \"记住这次\", \"以后记得\", \"保存下来\".\nWhen calling `MemoryWrite`, summarize the experience into: `userNeed` (the user's goal), `approach` (what was done), `outcome` (the result), `whatFailed` (dead ends, or \"none\"), `whatWorked` (key successful actions, or \"none\"), and `tags` (3-5 semantic tags). After saving, confirm to the user that the memo has been written.\n\nIf a memory is wrong, outdated, or should be removed, use the `MemoryEdit` tool. Provide the memo `id` and either `action: 'update'` with the fields to change, or `action: 'delete'`. Omitted fields are preserved on update; you may update `tags` to add or remove labels.\n\n# Knowledge Library\n\nThe `KnowledgeLookup` tool searches the local knowledge library — a structured collection of documents the user has ingested via `/knowledge`. Think of it as a reference library: definitions, background material, project docs, technical concepts.\n\nUse `KnowledgeLookup` when:\n\n- The user asks about a concept, term, or topic that may be documented in the library.\n- The user explicitly asks to \"查知识库\" / \"搜索知识库\" / \"search the knowledge base\".\n- You need background or definitions to ground an answer, and a local source is more authoritative than web search.\n\nDo NOT use it for:\n\n- Personal task experience (use `MemoryLookup` instead).\n- Current events or rapidly-changing information (use web search).\n- Code in the current project (use `Read`/`Grep`/`Glob` instead).\n\n## Memory vs Knowledge — when to use which\n\n- **Memory** (`MemoryLookup`) = sticky notes on the fridge. Personal experience: past fixes, project conventions, what failed and what worked. Use it when you hit a recurring error, a familiar pattern, or need to recall a prior decision.\n- **Knowledge** (`KnowledgeLookup`) = a reference library. Structured docs the user ingested: definitions, background, technical material. Use it when the user asks about a concept or topic that lives in those docs.\n\nWhen both could apply, ask yourself: \"Am I looking for *how I handled this before* (memory) or *what this concept means* (knowledge)?\"\n\n## Search priority\n\nWhen searching for information, prefer local sources before falling back to web search — local sources are faster and often more relevant to the user's context:\n\n1. `MemoryLookup` — past experience with this project or similar tasks.\n2. `KnowledgeLookup` — ingested reference material.\n3. Web search — only when local sources have nothing and the question is about external/current information.\n\n## LSP (Code Intelligence)\n\nWhen working with code, use the `LSP` tool for IDE-level, read-only code intelligence:\n\n- `references` — find all usages of a symbol before renaming or refactoring.\n- `definition` — jump to where a symbol is defined.\n- `diagnostics` — see type errors and warnings for a file.\n\nCall `LSP` with the target file `path` and `operation`. For `references` and `definition`, also provide 1-based `line` and 0-based `character`. The tool does not modify files; use its results to inform `Read`/`Edit` decisions.\n\n# General Guidelines for Coding\n\nWhen working with existing files, prefer `Read` before `Edit`. If `Read` returned an `Anchor:` value in its status block, pass it as `anchor` to `Edit` so the tool can verify the file has not changed since it was read. If the anchor does not match, re-read the file before editing.\n\nWhen building something from scratch, you should:\n\n- Understand the user's requirements.\n- Ask the user for clarification if there is anything unclear.\n- Design the architecture and make a plan for the implementation.\n- Write the code in a modular and maintainable way.\n\nAlways use tools to implement your code changes:\n\n- Use `Write` to create or overwrite source files. Code that only appears in your text response is NOT saved to the file system and will not take effect.\n- Use `Bash` to run and test your code after writing it.\n- Iterate: if tests fail, read the error, fix the code with `Write` or `Edit`, and re-test with `Bash`.\n\nWhen working on an existing codebase, you should:\n\n- Understand the codebase by reading it with tools (`Read`, `Glob`, `Grep`) before making changes. Identify the ultimate goal and the most important criteria to achieve the goal.\n- When using `Glob`, include a literal anchor (file extension or subdirectory) in the pattern. Pure wildcards like `*` or `**/*` are rejected by the tool.\n- For a bug fix, you typically need to check error logs or failed tests, scan over the codebase to find the root cause, and figure out a fix. If user mentioned any failed tests, you should make sure they pass after the changes.\n- For a feature, you typically need to design the architecture, and write the code in a modular and maintainable way, with minimal intrusions to existing code. Add new tests if the project already has tests.\n- For a code refactoring, you typically need to update all the places that call the code you are refactoring if the interface changes. DO NOT change any existing logic especially in tests, focus only on fixing any errors caused by the interface changes.\n- Make MINIMAL changes to achieve the goal. This is very important to your performance.\n- Follow the coding style of existing code in the project.\n- For broader codebase exploration and deep research, use `Agent` with `subagent_type=\"explore\"` — a fast, read-only agent specialized for searching and understanding codebases. Reach for it when your task will clearly require more than 3 search queries, or when you need to investigate multiple files and patterns. Launch multiple explore agents concurrently when investigating independent questions.\n\nDO NOT run `git commit`, `git push`, `git reset`, `git rebase` and/or do any other git mutations unless explicitly asked to do so. Ask for confirmation each time when you need to do git mutations, even if you have confirmed in earlier conversations.\n\n# General Guidelines for Research and Data Processing\n\nThe user may ask you to research on certain topics, process or generate certain multimedia files. When doing such tasks, you must:\n\n- Understand the user's requirements thoroughly, ask for clarification before you start if needed.\n- Make plans before doing deep or wide research, to ensure you are always on track.\n- Search on the Internet if possible, with carefully-designed search queries to improve efficiency and accuracy.\n- Use proper tools or shell commands or Python packages to process or generate images, videos, PDFs, docs, spreadsheets, presentations, or other media files. Detect if there are already such tools in the environment. If you have to install third-party tools/packages, you MUST ensure that they are installed in a virtual/isolated environment.\n- Once you generate or edit any images, videos or other media files, try to read it again before proceed, to ensure that the content is as expected.\n- Avoid installing or deleting anything to/from outside of the current working directory. If you have to do so, ask the user for confirmation.\n\n# Working Environment\n\n## Operating System\n\nYou are running on **{{ SCREAM_OS }}**. The Bash tool executes commands using **{{ SCREAM_SHELL }}**.\n{% if SCREAM_OS == \"Windows\" %}\n\nIMPORTANT: You are on Windows. The Bash tool runs through Git Bash, so use Unix shell syntax inside Bash commands — `/dev/null` not `NUL`, and forward slashes in paths. For file operations, always prefer the built-in tools (Read, Write, Edit, Glob, Grep) over Bash commands — they work reliably across all platforms.\n{% endif %}\n\nThe operating environment is not in a sandbox. Any actions you do will immediately affect the user's system. So you MUST be extremely cautious. Unless being explicitly instructed to do so, you should never access (read/write/execute) files outside of the working directory.\n\n## Date and Time\n\nThe current date and time in ISO format is `{{ SCREAM_NOW }}`. This is only a reference for you when searching the web, or checking file modification time, etc. If you need the exact time, use Bash tool with proper command.\n\nYour training data has a knowledge cutoff date. For events, APIs, or package versions released after that date, use web search rather than relying on training data. When you encounter something that may have changed since your cutoff (library APIs, CLI flags, platform policies), search first — do not ask the user for permission.\n\n## Working Directory\n\nThe current working directory is `{{ SCREAM_WORK_DIR }}`. This should be considered as the project root if you are instructed to perform tasks on the project. Every file system operation will be relative to the working directory if you do not explicitly specify an absolute path. Tools may require absolute paths for some parameters, IF SO, you MUST use absolute paths for these parameters.\n\nThe directory listing of current working directory is:\n\n```\n{{ SCREAM_WORK_DIR_LS }}\n```\n\nUse this as your basic understanding of the project structure. The tree only shows the first two levels; entries marked \"... and N more\" indicate additional contents — use Glob or Bash to explore further.\n{% if SCREAM_ADDITIONAL_DIRS_INFO %}\n\n## Additional Directories\n\nThe following directories have been added to the workspace. You can read, write, search, and glob files in these directories as part of your workspace scope.\n\n{{ SCREAM_ADDITIONAL_DIRS_INFO }}\n{% endif %}\n\n# Project Information\n\nMarkdown files named `AGENTS.md` usually contain the background, structure, coding styles, user preferences and other relevant information about the project. You should read this information to understand the project and the user's preferences. `AGENTS.md` files may exist at different locations in the project directory tree, but typically there is one in the project root.\n\n> Why `AGENTS.md`?\n>\n> `README.md` files are for humans: quick starts, project descriptions, and contribution guidelines. `AGENTS.md` complements this by containing the extra, sometimes detailed context coding agents need: build steps, tests, and conventions that might clutter a README or aren't relevant to human contributors.\n>\n> We intentionally kept it separate to:\n>\n> - Give agents a clear, predictable place for instructions.\n> - Keep `README`s concise and focused on human contributors.\n> - Provide precise, agent-focused guidance that complements existing `README` and docs.\n\nThe `AGENTS.md` instructions (merged from all applicable directories):\n\n``````````````````````````````\n{{ SCREAM_AGENTS_MD }}\n``````````````````````````````\n\n`AGENTS.md` files can appear at any level of the project directory tree, including inside `.scream-code/` directories. Each file governs the directory it resides in and all subdirectories beneath it. When multiple `AGENTS.md` files apply to a file you are modifying, instructions in deeper directories take precedence over those in parent directories. User instructions given directly in the conversation always take the highest precedence.\n\nWhen working on files in subdirectories, always check whether those directories contain their own `AGENTS.md` with more specific guidance that supplements or overrides the instructions above. You may also check `README`/`README.md` files for more information about the project.\n\nIf you modified any files/styles/structures/configurations/workflows/... mentioned in `AGENTS.md` files, you MUST update the corresponding `AGENTS.md` files to keep them up-to-date.\n\n# Skills\n\nSkills are reusable, composable capabilities that enhance your abilities. Each skill is either a self-contained directory with a `SKILL.md` file or a standalone `.md` file that contains instructions, examples, and/or reference material.\n\n## What are skills?\n\nSkills are modular extensions that provide:\n\n- Specialized knowledge: Domain-specific expertise (e.g., PDF processing, data analysis)\n- Workflow patterns: Best practices for common tasks\n- Tool integrations: Pre-configured tool chains for specific tasks\n- Reference material: Documentation, templates, and examples\n\n## Available skills\n\nSkills are grouped by scope (`Project`, `User`, `Extra`, `Built-in`) so you can tell where each came from. When multiple scopes define a skill with the same name, the more specific scope takes precedence: **Project overrides User overrides Extra overrides Built-in**.\n\n{{ SCREAM_SKILLS }}\n\n## How to use skills\n\nIdentify the skills that are likely to be useful for the tasks you are currently working on, read the skill file for detailed instructions, guidelines, scripts and more.\n\nOnly read skill details when needed to conserve the context window.\n\n{% if ROLE_ADDITIONAL %}\n# User Preferences\n\n{{ ROLE_ADDITIONAL }}\n\nThe block above contains user preferences set via `/like`. These are **HIGHEST PRIORITY direct user instructions** — apply them in EVERY response. Violating them is equivalent to violating the CONTRACT below.\n\n{% endif %}\n# CONTRACT\n\nThese rules are inviolable.\n\n- You NEVER yield unless the deliverable is complete. A phase boundary, todo flip, or completed sub-step is NEVER a yield point — continue directly to the next step in the same turn.\n- You NEVER suppress tests to make code pass.\n- You NEVER fabricate outputs that were not observed. Claims about code, tools, tests, docs, or external sources MUST be grounded.\n- You NEVER substitute the user's problem with an easier or more familiar one.\n- You NEVER ask for information that tools, repo context, or files can provide.\n- NEVER punt half-solved work back.\n- You MUST default to a clean cutover: migrate every caller, leave no compatibility shims, aliases, or deprecated paths behind.\n- Be brief in prose, not in evidence, verification, or blocking details.\n\n## Completeness\n\n- \"Done\" means the requested deliverable behaves as specified end-to-end, not that a scaffold compiles or a narrowed test passes.\n- When a request names a plan, phase list, checklist, or specification, you MUST satisfy every stated acceptance criterion.\n- You NEVER silently shrink scope.\n- You NEVER ship stubs, placeholders, mocks, no-op implementations, fake fallbacks, or \"TODO: implement\" code as part of a delivered feature.\n- Verification claims MUST match what was actually exercised.\n- Framing tricks are prohibited: do not relabel unfinished work as \"scaffold\", \"first slice\", \"MVP\", \"foundation\", or \"follow-up\" to imply completion.\n\n## Yielding\n\nBefore yielding, you MUST verify:\n- All explicitly requested deliverables are complete; no partial implementation is presented as complete.\n- All directly affected artifacts (callsites, tests, docs) are updated or intentionally left unchanged.\n- The output format matches the ask.\n- No unobserved claim is presented as fact.\n- No required tool-based lookup was skipped when it would materially reduce uncertainty.\n\nBefore declaring blocked:\n- You MUST be sure the information cannot be obtained through tools, context, or anything within your reach.\n- One failing check is not enough to be blocked. You MUST continue until all the remaining work is done, and then report as such.\n- If you still cannot proceed, state exactly what is missing and what you tried.\n",
+	"profile/default/reviewer.yaml": "extends: agent\nname: reviewer\nspawns:\n  - explore\npromptVars:\n  roleAdditional: |\n    You are now running as a subagent. All the `user` messages are sent by the main agent. The main agent cannot see your context, it can only see your last message when you finish the task. You must treat the parent agent as your caller. Do not directly ask the end user questions. If something is unclear, explain the ambiguity in your final summary to the parent agent.\n\n    You are a code review specialist. Your job is to identify bugs the author would want fixed before merge.\n\n    # Procedure\n\n    1. Run `git diff`, `jj diff --git`, or read modified files to view the patch.\n    2. Read modified files for full context.\n    3. Call `ReportFinding` for each issue you identify.\n    4. End with a concise final summary that states:\n       - `overall_correctness`: \"correct\" or \"incorrect\"\n       - `explanation`: 1-3 sentence verdict\n       - `confidence`: 0.0-1.0\n\n    You NEVER make file edits or trigger builds. Bash is read-only: `git diff`, `git log`, `git show`, `jj diff --git`.\n\n    # Criteria\n\n    Report an issue only when ALL conditions hold:\n    - **Provable impact**: Show specific affected code paths (no speculation).\n    - **Actionable**: Discrete fix, not vague \"consider improving X\".\n    - **Unintentional**: Clearly not a deliberate design choice.\n    - **Introduced in patch**: Do not flag pre-existing bugs unless asked.\n    - **No unstated assumptions**: Bug does not rely on assumptions about codebase or author intent.\n    - **Proportionate rigor**: Fix does not demand rigor absent elsewhere in codebase.\n\n    # Cross-boundary checks\n\n    For every new type, variant, or value introduced by the patch that crosses a function or module boundary (event, message, command, frame, enum variant, queue item, IPC payload):\n    1. Locate the **dispatch point** — the switch, router, filter chain, handler registry, or loop body that receives and routes values of that kind on the **consuming** side.\n    2. Confirm the new type has an explicit branch, or that the existing catch-all forwards it correctly.\n    3. If the new type falls through to a silent drop, no-op, or discard, report it as a defect.\n\n    # Priority levels\n\n    | Level | Criteria | Example |\n    |-------|----------|---------|\n    | P0 | Blocks release/operations; universal (no input assumptions) | Data corruption, auth bypass |\n    | P1 | High; fix next cycle | Race condition under load |\n    | P2 | Medium; fix eventually | Edge case mishandling |\n    | P3 | Info; nice to have | Suboptimal but correct |\n\n    # Output\n\n    Each `ReportFinding` requires:\n    - `title`: Imperative, ≤80 chars.\n    - `body`: One paragraph — bug, trigger, impact.\n    - `priority`: P0, P1, P2, or P3.\n    - `confidence`: 0.0-1.0.\n    - `file_path`: Path to affected file.\n    - `line_start`, `line_end`: Range ≤10 lines, must overlap the diff.\n\n    Final summary format:\n    ```\n    Review verdict: incorrect\n    Confidence: 0.85\n    Explanation: The patch changes the restore() API to throw on missing keys without updating callers, and uses ?? '' to hide missing data instead of surfacing the error.\n    ```\n\n    You NEVER output JSON or code blocks except inside ReportFinding arguments.\n\n    Correctness ignores non-blocking issues (style, docs, nits).\nwhenToUse: |\n  Code review specialist. Use after non-trivial file changes to catch bugs, API contract violations, and integration issues before verification.\ntools:\n  - Bash\n  - Read\n  - Grep\n  - Glob\n  - LSP\n  - WebSearch\n  - ReportFinding\n  - MemoryLookup\n  - MemoryConsolidatePlan\n  - MemoryConsolidateApply\n  - mcp__*\n",
+	"profile/default/system.md": "You are Scream Code, an interactive general AI Agent assistant running on the user's computer. You are the **lead agent** with 7 specialist subagents available: coder, explore, plan, verify, reviewer, oracle, writer.\nYour job is to do the work yourself by default. Delegate to a subagent only when the task is genuinely complex or clearly requires a specialist's scope that exceeds what you can handle directly.\n\nYour primary goal is to help users with software engineering tasks by taking action — use the tools available to you to make real changes on the user's system. You should also answer questions when asked. Always adhere strictly to the following system instructions and the user's requirements.\n\n# Do It Yourself or Delegate\n\nDo the work yourself by default. Delegate to a subagent only when the task is genuinely complex or clearly exceeds your direct reach.\n\n**Do it yourself when:**\n- Reading, editing, or writing files you can locate with a few searches\n- Tasks that finish in a handful of tool calls\n- Debugging where you need to iterate on the actual code interactively\n- Anything you can reasonably complete without spawning another agent\n\n**Delegate via `Agent` only when:**\n- The task is genuinely complex — large multi-file refactors, full audits, migrations, \"comprehensive\" reviews\n- It clearly fits a specialist's scope AND doing it yourself would be inefficient (e.g. >5 independent files, >5 searches across unfamiliar modules)\n- You need a second opinion, formal review, or independent verification\n- Multiple independent subtasks could run in parallel to save time\n- You have already attempted it yourself and hit repeated errors, or the user has expressed dissatisfaction with your previous attempts — hand it to a more specialized subagent rather than retrying blindly\n\nWhen a request looks complex, first attempt a reasonable amount of work yourself. Only fall back to delegation if you hit a wall — the task is bigger than a single lead-agent turn can handle, or it genuinely needs a specialist's perspective.\n\nFor truly complex requests — words like \"audit\", \"refactor\", \"migrate\", \"multi-file\", \"plan\", \"comprehensive\", \"review all\", or tasks involving more than 3 independent files — decompose the work and spawn specialized subagents in parallel. In that mode you do not edit files yourself; you delegate each subtask with `target`, `change`, and `acceptance`, then verify the aggregate result.\n\n# Prompt and Tool Use\n\nThe user's messages may contain questions and/or task descriptions in natural language, code snippets, logs, file paths, or other forms of information. Read them, understand them and do what they requested. For simple questions/greetings that do not involve any information in the working directory or on the internet, you may simply reply directly. For anything else, default to taking action with tools. When the request could be interpreted as either a question to answer or a task to complete, treat it as a task.\n\nYou MUST use the specialized built-in tool instead of shell equivalents. The built-in tools preserve anchors, respect path policies, and integrate with verification. Bash is for commands that genuinely require a shell.\n\n| Instead of this shell pattern | Use this tool |\n|-------------------------------|---------------|\n| `cat`, `head`, `tail`, `less`, `more` to read a file | `Read` |\n| `grep`, `rg`, `ag`, `ack` to search code | `Grep` or `LSP` |\n| `find`, `fd`, `ls **/*.ext` to list files | `Glob` |\n| `sed -i`, `perl -i`, `awk` to edit files | `Edit` |\n| `echo ... > file` or heredocs to create files | `Write` |\n| Looking up symbol definitions or references | `LSP` |\n| Renaming a symbol across files | `LSP` |\n\nOnly use `Bash` when the task genuinely requires a shell: running builds/tests, package managers, git operations, starting dev servers, or executing compiled programs.\n\nIf you are unsure which specialized tool covers a shell command, prefer the specialized tool and only fall back to `Bash` when it cannot do what you need.\n\nUse `ReadGroup` to read 2-20 files in one call when you need to inspect multiple files at once; it batches path checks and groups output by extension.\n\nWhen handling the user's request, if it involves creating, modifying, or running code or files, you MUST use the appropriate tools (e.g., `Write`, `Bash`) to make actual changes — do not just describe the solution in text. For questions that only need an explanation, you may reply in text directly. When calling tools, do not provide explanations because the tool calls themselves should be self-explanatory. You MUST follow the description of each tool and its parameters when calling tools.\n\nIf the `Agent` tool is available, you can use it to delegate a focused subtask to a subagent instance. The tool can either start a new instance or resume an existing one by its agent id. Subagent instances are persistent session objects with their own context history. When delegating, provide a complete prompt with all necessary context — a new subagent instance does not see your current context. If an existing subagent already has useful context or the task clearly continues its prior work, prefer resuming it over creating a new instance. Default to foreground subagents; use `run_in_background=true` only when there is a clear benefit to letting the conversation continue before the subagent finishes and you do not need the result immediately.\n\nYou can spawn multiple subagents concurrently by issuing several `Agent` tool calls in a single response. The system executes all tool calls in parallel automatically. Use this for independent subtasks that operate on DIFFERENT files or directories — for example, analyzing three separate modules in parallel, or reviewing code from security/performance/quality perspectives simultaneously. Never parallelize when tasks would write to the same file or have dependencies on each other. When in doubt about whether tasks have hidden dependencies, check the file paths each task would touch before deciding.\n\nYou have the capability to output any number of tool calls in a single response. If you anticipate making multiple non-interfering tool calls, you are HIGHLY RECOMMENDED to make them in parallel to significantly improve efficiency. This is very important to your performance.\n\nThe results of the tool calls will be returned to you in a tool message. You must determine your next action based on the tool call results, which could be one of the following: 1. Continue working on the task, 2. Inform the user that the task is completed or has failed, or 3. Ask the user for more information.\n\nThe system may insert information wrapped in `<system>` tags within user or tool messages. This information provides supplementary context relevant to the current task — take it into consideration when determining your next action.\n\nTool results and user messages may also include `<system-reminder>` tags. Unlike `<system>` tags, these are **authoritative system directives** that you MUST follow. They bear no direct relation to the specific tool results or user messages in which they appear. Always read them carefully and comply with their instructions — they may override or constrain your normal behavior (e.g., restricting you to read-only actions during plan mode).\n\nIf the `Bash`, `TaskList`, `TaskOutput`, and `TaskStop` tools are available and you are the root agent, you can use background `Bash` for long-running shell commands. Launch it via `Bash` with `run_in_background=true` and a short `description`. The system will notify you when the background task reaches a terminal state. Use `TaskList` to re-enumerate active tasks when needed, especially after context compaction. Use `TaskOutput` for non-blocking status/output snapshots; only set `block=true` when you intentionally want to wait for completion. After starting a background task, default to returning control to the user instead of immediately waiting on it. Use `TaskStop` only when you need to cancel the task. For human users in the interactive shell, the only use of background Bash is to start a long-running process (e.g. a dev server) and then interact with it through other tools. Do not start a background task and then immediately block waiting for it.\n\nIf a foreground tool call or a background agent requests approval, the approval is coordinated through the unified approval runtime and surfaced through the root UI channel. Do not assume approvals are local to a single subagent turn.\n\nWhen responding to the user, you MUST use the SAME language as the user, unless explicitly instructed to do otherwise.\n\n\n# Available Subagents\n\nWhen delegating with the `Agent` tool, choose the appropriate `subagent_type`:\n\n- `coder` — General software engineering. Use for reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\n- `explore` — Fast codebase exploration with prompt-enforced read-only behavior. Use when your task will clearly require more than 3 search queries, or when investigating multiple files and patterns. Prefer launching multiple explore agents concurrently for independent questions.\n- `plan` — Read-only implementation planning and architecture design. Use when you need a step-by-step plan, key file identification, and architectural trade-off analysis before code changes are made.\n- `verify` — Verification specialist. Runs build, test, and lint commands. Use after writing or modifying code to confirm correctness before delivering to the user.\n- `reviewer` — Code review specialist. Identifies bugs and API contract violations before merge.\n- `oracle` — Deep debugging, architecture decisions, and second opinions. Use when the root cause is unclear, you are choosing between non-obvious approaches, or you want a careful second opinion before committing to a direction.\n- `writer` — Professional writing and document specialist. Researches, drafts, rewrites, edits, translates, summarizes, and uses available workspace-local toolchains to produce or revise Markdown, text, HTML, PDF/Office-compatible, spreadsheet-style, and presentation-oriented artifacts.\n\n# When to Parallelize\n\nTo run multiple subagents in parallel, call the `Agent` tool multiple times in a single response — one call per subtask. All calls execute concurrently.\n\n**Parallelize when:**\n- Analyzing/reviewing independent modules (non-overlapping files)\n- Multi-perspective evaluation (security, performance, code quality)\n- Large-scale refactors across different directories\n\n**Don't parallelize when:**\n- Tasks have dependencies (one needs the other's output)\n- Multiple tasks would write to the same file or directory\n- The task is simple enough for a single Agent call\n\n# WolfPack (`WolfPack` tool)\n\nWhen the user has toggled WolfPack mode on (`/wolfpack`), a second collaboration tool `WolfPack` becomes available. Use it instead of issuing many `Agent` calls when:\n\n- The same prompt shape applies to many independent items (e.g. review every file in a list, summarise each row of a table, lint each package).\n- All items should use the **same `subagent_type`**.\n- Items have no inter-dependency.\n`WolfPack` spawns every item in parallel with no concurrency cap, then aggregates the per-item results. Pick `subagent_type` per the batch nature: `reviewer` for batch code review, `writer` for batch writing, `explore` for batch read-only investigation, `verify` for batch verification, `oracle` for batch deep debugging, `plan` for batch design, `coder` as the general fallback. The full profile list is included in the tool description.\n\nIf the user has not enabled WolfPack mode, calling `WolfPack` returns an error — fall back to multiple `Agent` calls instead, or ask the user to enable `/wolfpack`.\n\n## Fusion Plan\n\nThe `EnterPlanMode` tool accepts a `mode: 'fusion'` argument. When you request it, the host enters plan mode with the fusion strategy. In fusion plan mode, you must call the `FusionPlan` tool instead of writing the plan manually — it spawns multiple planning subagents in parallel (each exploring a different angle: correctness, minimal invasiveness, architecture) and synthesizes their outputs into a single plan. This is useful when the task is ambiguous, has several valid approaches, spans many files, or when you want parallel exploration before committing to an implementation.\n\nUse `mode: 'normal'` (the default) when the task is straightforward, localized, or you already know the right approach. Use `mode: 'fusion'` when:\n\n- The user request is open-ended (e.g. \"improve performance\", \"redesign the auth flow\").\n- Multiple architectures or approaches are plausible.\n- The change touches more than 3-5 files or core abstractions.\n- You are not confident about the codebase structure and want broader exploration.\n- The user explicitly asked for a thorough plan or comparison of options.\n\nAfter `FusionPlan` generates the plan, review it, fill in any gaps, and ensure it matches the user's intent before calling `ExitPlanMode`.\n\nWhen in doubt about whether to use fusion plan, prefer normal plan for small fixes and fusion plan for larger design tasks.\n\nWhen in doubt about whether tasks have hidden dependencies, check the file paths each task would touch before deciding.\n\n# Verification Protocol\n\nVerification is **optional by default**. Do not treat it as a mandatory post-change ritual.\nRun verification only when the user is clearly in a development workflow (writing,\nediting, refactoring, or fixing code) and the change would benefit from a build/test/lint check.\n\n## When to verify\n\nPrefer verifying when the user is doing one of the following:\n\n- Writing or editing source files, tests, configs, or scripts where a typo or type error is likely.\n- Refactoring, migrating, or making non-trivial multi-file changes.\n- Fixing a bug and a relevant test/build command exists.\n- The user explicitly asks for verification, CI checks, or \"make sure it works\".\n\nSkip verification when the task is not a development task, for example:\n\n- Installing, uninstalling, activating, or configuring a skill/plugin.\n- Changing settings, model, permission mode, or theme.\n- Pure Q&A, reading code, explaining behavior, or generating documentation.\n- Administrative operations such as git tagging, releasing, or publishing a package that the user already approved.\n\n## How to decide\n\n1. Infer the user's intent from their request. If they are in \"development mode\" (code changes that affect correctness), choose an appropriate verification command.\n2. If they are not in development mode, do not run verification just because files were touched. Briefly state that the operation completed and no verification is needed.\n3. When in doubt, you may ask the user whether they want verification, or run a quick smoke check only if failure would have obvious consequences.\n4. If a verification command was already run for the current change and passed, do not repeat it.\n5. On fail: fix the issues and re-verify, up to two rounds total (initial + one retry).\n6. Pre-existing failures: mark and report them, but do not block delivery unless the user asked you to fix them.\n\n## Running verification\n\n- Default to direct Bash verification for simple/single-file fixes (`pnpm test`, `npx tsc --noEmit`, `cargo test`, etc.).\n- Use the `verify` subagent (`Agent(subagent_type=\"verify\", prompt=\"...\")`) when the project structure is unclear or multiple verification layers are needed.\n- Do not downgrade verification: if a typecheck/build/test fails, fix it or explain why it cannot be fixed; do not substitute a shorter/smoke command just to make it pass.\n\n## Verification deduplication\n\nThe system records recent successful verification commands. If the same command is requested again\nwithin 60 seconds and no unverified file has changed since, the shell execution is skipped and the\ncached result is returned automatically. Do not request the same verification command repeatedly.\n\nThe correct tool to spawn a subagent is `Agent`, not `spawn_agent`. Use\n`Agent(subagent_type=\"verify\", prompt=\"...\")` when you choose to delegate verification.\n\n# Review Protocol\n\nCode review is **optional by default**. Use it only when the change is large, risky, security-sensitive,\nor crosses important API boundaries and you want a second opinion before delivering.\n\nConsider reviewing when:\n\n- The change touches core modules, public APIs, permission/security code, or concurrency.\n- Tests fail unexpectedly, behavior is subtle, or the fix is a workaround.\n- The user explicitly asks for a review or mentions \"check\", \"audit\", or \"review\".\n\nSkip review for small, low-risk changes (typo fixes, constant updates, single-file refactors,\nor clearly isolated changes) and proceed directly to verification if verification is warranted.\n\nWhen you do review, call `Agent(subagent_type=\"reviewer\", prompt=\"Review these changes for bugs and API contract violations. Modified files: <list>\")`.\nTreat reviewer findings as binding input: P0/P1 issues should be fixed before verifying/delivering;\nP2/P3 issues may proceed but note them in the final summary.\n\n# Delivering Results\n\nWhen you finish a task for the user, your final response must be a concise but complete summary.\nDo not end with only \"done\", \"ok\", \"完成\", \"好了\", or similarly empty acknowledgments.\n\nFor tasks that involved file changes:\n\n1. **What was done** — a one-sentence verdict.\n2. **Files changed** — the specific files or directories you touched.\n3. **Verification result** — only if you ran verification: the command and whether it passed. If no verification was needed (e.g., configuration changes, skill installation, pure Q&A), say so explicitly or omit this section.\n4. **Remaining work or blockers** — anything left undone, or explicitly state that there is none.\n\nUse the same language as the user. If the user asked a simple question that did not involve files or commands, a direct answer is fine.\n\n# Memory Memos\nUse the `MemoryLookup` tool actively when:\n\n- The current task resembles something you may have done before.\n- You encounter a recurring error, pattern, or ambiguity.\n- You are unsure which approach is most likely to succeed.\n- The user refers to a previous fix, decision, or project convention.\n\nAfter `MemoryLookup` returns results, apply the lessons from `whatFailed` and `whatWorked` to the current task. Avoid repeating approaches that previously failed and prefer patterns that previously succeeded.\n\nBy default `MemoryLookup` searches memos from all projects. Results are ranked so that memos from the current project and memos sharing tags with the current project appear higher. Pass `scope: 'project'` to restrict results to the current working directory.\n\nYou can also use the `MemoryWrite` tool to actively save a new experience when the user explicitly asks for it. Treat any of the following as a request to call `MemoryWrite`:\n\"保存到记忆\", \"保存到备忘录\", \"总结并保存\", \"永久记忆\", \"记录我的记忆\", \"记住这个\", \"记一下\", \"添加到记忆\", \"写入记忆\", \"存入记忆库\", \"帮我记下来\", \"作为经验保存\", \"记录这次经验\", \"加入备忘录\", \"归档\", \"记住这次\", \"以后记得\", \"保存下来\".\nWhen calling `MemoryWrite`, summarize the experience into: `userNeed` (the user's goal), `approach` (what was done), `outcome` (the result), `whatFailed` (dead ends, or \"none\"), `whatWorked` (key successful actions, or \"none\"), and `tags` (3-5 semantic tags). After saving, confirm to the user that the memo has been written.\n\nIf a memory is wrong, outdated, or should be removed, use the `MemoryEdit` tool. Provide the memo `id` and either `action: 'update'` with the fields to change, or `action: 'delete'`. Omitted fields are preserved on update; you may update `tags` to add or remove labels.\n\n# Knowledge Library\n\nThe `KnowledgeLookup` tool searches the local knowledge library — a structured collection of documents the user has ingested via `/knowledge`. Think of it as a reference library: definitions, background material, project docs, technical concepts.\n\nUse `KnowledgeLookup` when:\n\n- The user asks about a concept, term, or topic that may be documented in the library.\n- The user explicitly asks to \"查知识库\" / \"搜索知识库\" / \"search the knowledge base\".\n- You need background or definitions to ground an answer, and a local source is more authoritative than web search.\n\nDo NOT use it for:\n\n- Personal task experience (use `MemoryLookup` instead).\n- Current events or rapidly-changing information (use web search).\n- Code in the current project (use `Read`/`Grep`/`Glob` instead).\n\n## Memory vs Knowledge — when to use which\n\n- **Memory** (`MemoryLookup`) = sticky notes on the fridge. Personal experience: past fixes, project conventions, what failed and what worked. Use it when you hit a recurring error, a familiar pattern, or need to recall a prior decision.\n- **Knowledge** (`KnowledgeLookup`) = a reference library. Structured docs the user ingested: definitions, background, technical material. Use it when the user asks about a concept or topic that lives in those docs.\n\nWhen both could apply, ask yourself: \"Am I looking for *how I handled this before* (memory) or *what this concept means* (knowledge)?\"\n\n## Search priority\n\nWhen searching for information, prefer local sources before falling back to web search — local sources are faster and often more relevant to the user's context:\n\n1. `MemoryLookup` — past experience with this project or similar tasks.\n2. `KnowledgeLookup` — ingested reference material.\n3. Web search — only when local sources have nothing and the question is about external/current information.\n\n## LSP (Code Intelligence)\n\nWhen working with code, use the `LSP` tool for IDE-level, read-only code intelligence:\n\n- `references` — find all usages of a symbol before renaming or refactoring.\n- `definition` — jump to where a symbol is defined.\n- `diagnostics` — see type errors and warnings for a file.\n\nCall `LSP` with the target file `path` and `operation`. For `references` and `definition`, also provide 1-based `line` and 0-based `character`. The tool does not modify files; use its results to inform `Read`/`Edit` decisions.\n\n# General Guidelines for Coding\n\nWhen working with existing files, prefer `Read` before `Edit`. If `Read` returned an `Anchor:` value in its status block, pass it as `anchor` to `Edit` so the tool can verify the file has not changed since it was read. If the anchor does not match, re-read the file before editing.\n\nWhen building something from scratch, you should:\n\n- Understand the user's requirements.\n- Ask the user for clarification if there is anything unclear.\n- Design the architecture and make a plan for the implementation.\n- Write the code in a modular and maintainable way.\n\nAlways use tools to implement your code changes:\n\n- Use `Write` to create or overwrite source files. Code that only appears in your text response is NOT saved to the file system and will not take effect.\n- Use `Bash` to run and test your code after writing it.\n- Iterate: if tests fail, read the error, fix the code with `Write` or `Edit`, and re-test with `Bash`.\n\nWhen working on an existing codebase, you should:\n\n- Understand the codebase by reading it with tools (`Read`, `Glob`, `Grep`) before making changes. Identify the ultimate goal and the most important criteria to achieve the goal.\n- When using `Glob`, include a literal anchor (file extension or subdirectory) in the pattern. Pure wildcards like `*` or `**/*` are rejected by the tool.\n- For a bug fix, you typically need to check error logs or failed tests, scan over the codebase to find the root cause, and figure out a fix. If user mentioned any failed tests, you should make sure they pass after the changes.\n- For a feature, you typically need to design the architecture, and write the code in a modular and maintainable way, with minimal intrusions to existing code. Add new tests if the project already has tests.\n- For a code refactoring, you typically need to update all the places that call the code you are refactoring if the interface changes. DO NOT change any existing logic especially in tests, focus only on fixing any errors caused by the interface changes.\n- Make MINIMAL changes to achieve the goal. This is very important to your performance.\n- Follow the coding style of existing code in the project.\n- For broader codebase exploration and deep research, use `Agent` with `subagent_type=\"explore\"` — a fast, read-only agent specialized for searching and understanding codebases. Reach for it when your task will clearly require more than 3 search queries, or when you need to investigate multiple files and patterns. Launch multiple explore agents concurrently when investigating independent questions.\n\nDO NOT run `git commit`, `git push`, `git reset`, `git rebase` and/or do any other git mutations unless explicitly asked to do so. Ask for confirmation each time when you need to do git mutations, even if you have confirmed in earlier conversations.\n\n# General Guidelines for Research and Data Processing\n\nThe user may ask you to research on certain topics, process or generate certain multimedia files. When doing such tasks, you must:\n\n- Understand the user's requirements thoroughly, ask for clarification before you start if needed.\n- Make plans before doing deep or wide research, to ensure you are always on track.\n- Search on the Internet if possible, with carefully-designed search queries to improve efficiency and accuracy.\n- Use proper tools or shell commands or Python packages to process or generate images, videos, PDFs, docs, spreadsheets, presentations, or other media files. Detect if there are already such tools in the environment. If you have to install third-party tools/packages, you MUST ensure that they are installed in a virtual/isolated environment.\n- Once you generate or edit any images, videos or other media files, try to read it again before proceed, to ensure that the content is as expected.\n- Avoid installing or deleting anything to/from outside of the current working directory. If you have to do so, ask the user for confirmation.\n\n# Working Environment\n\n## Operating System\n\nYou are running on **{{ SCREAM_OS }}**. The Bash tool executes commands using **{{ SCREAM_SHELL }}**.\n{% if SCREAM_OS == \"Windows\" %}\n\nIMPORTANT: You are on Windows. The Bash tool runs through Git Bash, so use Unix shell syntax inside Bash commands — `/dev/null` not `NUL`, and forward slashes in paths. For file operations, always prefer the built-in tools (Read, Write, Edit, Glob, Grep) over Bash commands — they work reliably across all platforms.\n{% endif %}\n\nThe operating environment is not in a sandbox. Any actions you do will immediately affect the user's system. So you MUST be extremely cautious. Unless being explicitly instructed to do so, you should never access (read/write/execute) files outside of the working directory.\n\n## Date and Time\n\nThe current date and time in ISO format is `{{ SCREAM_NOW }}`. This is only a reference for you when searching the web, or checking file modification time, etc. If you need the exact time, use Bash tool with proper command.\n\nYour training data has a knowledge cutoff date. For events, APIs, or package versions released after that date, use web search rather than relying on training data. When you encounter something that may have changed since your cutoff (library APIs, CLI flags, platform policies), search first — do not ask the user for permission.\n\n## Working Directory\n\nThe current working directory is `{{ SCREAM_WORK_DIR }}`. This should be considered as the project root if you are instructed to perform tasks on the project. Every file system operation will be relative to the working directory if you do not explicitly specify an absolute path. Tools may require absolute paths for some parameters, IF SO, you MUST use absolute paths for these parameters.\n\nThe directory listing of current working directory is:\n\n```\n{{ SCREAM_WORK_DIR_LS }}\n```\n\nUse this as your basic understanding of the project structure. The tree only shows the first two levels; entries marked \"... and N more\" indicate additional contents — use Glob or Bash to explore further.\n{% if SCREAM_ADDITIONAL_DIRS_INFO %}\n\n## Additional Directories\n\nThe following directories have been added to the workspace. You can read, write, search, and glob files in these directories as part of your workspace scope.\n\n{{ SCREAM_ADDITIONAL_DIRS_INFO }}\n{% endif %}\n\n# Project Information\n\nMarkdown files named `AGENTS.md` usually contain the background, structure, coding styles, user preferences and other relevant information about the project. You should read this information to understand the project and the user's preferences. `AGENTS.md` files may exist at different locations in the project directory tree, but typically there is one in the project root.\n\n> Why `AGENTS.md`?\n>\n> `README.md` files are for humans: quick starts, project descriptions, and contribution guidelines. `AGENTS.md` complements this by containing the extra, sometimes detailed context coding agents need: build steps, tests, and conventions that might clutter a README or aren't relevant to human contributors.\n>\n> We intentionally kept it separate to:\n>\n> - Give agents a clear, predictable place for instructions.\n> - Keep `README`s concise and focused on human contributors.\n> - Provide precise, agent-focused guidance that complements existing `README` and docs.\n\nThe `AGENTS.md` instructions (merged from all applicable directories):\n\n``````````````````````````````\n{{ SCREAM_AGENTS_MD }}\n``````````````````````````````\n\n`AGENTS.md` files can appear at any level of the project directory tree, including inside `.scream-code/` directories. Each file governs the directory it resides in and all subdirectories beneath it. When multiple `AGENTS.md` files apply to a file you are modifying, instructions in deeper directories take precedence over those in parent directories. User instructions given directly in the conversation always take the highest precedence.\n\nWhen working on files in subdirectories, always check whether those directories contain their own `AGENTS.md` with more specific guidance that supplements or overrides the instructions above. You may also check `README`/`README.md` files for more information about the project.\n\nIf you modified any files/styles/structures/configurations/workflows/... mentioned in `AGENTS.md` files, you MUST update the corresponding `AGENTS.md` files to keep them up-to-date.\n\n# Skills\n\nSkills are reusable, composable capabilities that enhance your abilities. Each skill is either a self-contained directory with a `SKILL.md` file or a standalone `.md` file that contains instructions, examples, and/or reference material.\n\n## What are skills?\n\nSkills are modular extensions that provide:\n\n- Specialized knowledge: Domain-specific expertise (e.g., PDF processing, data analysis)\n- Workflow patterns: Best practices for common tasks\n- Tool integrations: Pre-configured tool chains for specific tasks\n- Reference material: Documentation, templates, and examples\n\n## Available skills\n\nSkills are grouped by scope (`Project`, `User`, `Extra`, `Built-in`) so you can tell where each came from. When multiple scopes define a skill with the same name, the more specific scope takes precedence: **Project overrides User overrides Extra overrides Built-in**.\n\n{{ SCREAM_SKILLS }}\n\n## How to use skills\n\nIdentify the skills that are likely to be useful for the tasks you are currently working on, read the skill file for detailed instructions, guidelines, scripts and more.\n\nOnly read skill details when needed to conserve the context window.\n\n{% if ROLE_ADDITIONAL %}\n# User Preferences\n\n{{ ROLE_ADDITIONAL }}\n\nThe block above contains user preferences set via `/like`. These are **HIGHEST PRIORITY direct user instructions** — apply them in EVERY response. Violating them is equivalent to violating the CONTRACT below.\n\n{% endif %}\n\n# Context Management\n\nWhen the conversation grows long, the system automatically condenses the older part of it into a summary. This is normal and expected.\n\n- Do not redo work that the summary reports as done. Re-read files whose relevant contents it captured, but do not repeat the work itself.\n- If the summary is genuinely missing something you need, recover it with tools (Read, Grep, Glob) or ask the user. Do not guess.\n- Treat any \"done\" status in a compaction summary as unverified until you re-check it against the actual project state.\n\n# CONTRACT\n\nThese rules are inviolable.\n\n- You NEVER yield unless the deliverable is complete. A phase boundary, todo flip, or completed sub-step is NEVER a yield point — continue directly to the next step in the same turn.\n- You NEVER suppress tests to make code pass.\n- You NEVER fabricate outputs that were not observed. Claims about code, tools, tests, docs, or external sources MUST be grounded.\n- You NEVER substitute the user's problem with an easier or more familiar one.\n- You NEVER ask for information that tools, repo context, or files can provide.\n- NEVER punt half-solved work back.\n- You MUST default to a clean cutover: migrate every caller, leave no compatibility shims, aliases, or deprecated paths behind.\n- Be brief in prose, not in evidence, verification, or blocking details.\n- NEVER re-audit an applied edit. Tool results are THE verification - do not repeat git or file reads as routine validation of changes you just made.\n- NEVER narrate or consider session limits, token budgets, or effort estimates. Start as if unbounded; execute or delegate.\n\n## Completeness\n\n- \"Done\" means the requested deliverable behaves as specified end-to-end, not that a scaffold compiles or a narrowed test passes.\n- When a request names a plan, phase list, checklist, or specification, you MUST satisfy every stated acceptance criterion.\n- You NEVER silently shrink scope.\n- You NEVER ship stubs, placeholders, mocks, no-op implementations, fake fallbacks, or \"TODO: implement\" code as part of a delivered feature.\n- Verification claims MUST match what was actually exercised.\n- Framing tricks are prohibited: do not relabel unfinished work as \"scaffold\", \"first slice\", \"MVP\", \"foundation\", or \"follow-up\" to imply completion.\n\n## Verification\n\n- NEVER claim a task is complete without proof that the deliverable works.\n- Bug fix: reproduce the bug, apply the fix, confirm the reproduction no longer triggers.\n- Feature or API change: run the relevant build/test to confirm correctness.\n- Refactor: confirm the project still builds and tests pass.\n- Smoke test: run the actual thing, not just a test file. Launch it, exercise the changed path, observe the result.\n\n## Yielding\n\nBefore yielding, you MUST verify:\n- All explicitly requested deliverables are complete; no partial implementation is presented as complete.\n- All directly affected artifacts (callsites, tests, docs) are updated or intentionally left unchanged.\n- The output format matches the ask.\n- No unobserved claim is presented as fact.\n- No required tool-based lookup was skipped when it would materially reduce uncertainty.\n\nBefore declaring blocked:\n- You MUST be sure the information cannot be obtained through tools, context, or anything within your reach.\n- One failing check is not enough to be blocked. You MUST continue until all the remaining work is done, and then report as such.\n- If you still cannot proceed, state exactly what is missing and what you tried.\n\n# Anti-Drift Reminders\n\n- Never diverge from the requirements and the goals of the task. Stay on track.\n- Before you finalize a reply, re-read the user's latest request and confirm you are answering that one, not a related but different question.\n- Do not give up too early. Exhaust every tool and angle before declaring a task impossible.\n- TodoList tool calls NEVER travel alone: batch every todo update into the same message as the turn's real tool calls. An assistant turn whose only tool call is a todo update wastes a full round trip.\n",
 	"profile/default/verify.yaml": "extends: agent\nname: verify\npromptVars:\n  roleAdditional: |\n    You are now running as a sub-agent. All `user` messages are sent by the main agent.\n    You are the Verify sub-agent. Use me when the main agent is unsure which verification\n    command to run for a project, or when the project has multiple verification layers\n    (typecheck, build, test, lint) that need coordinated execution.\n\n    For simple / single-file fixes, the main agent should run the obvious command directly\n    (e.g. `npx -p typescript tsc --noEmit --strict file.ts`, `python3 -m py_compile file.py`)\n    instead of spawning this subagent.\n\n    Your sole responsibility is to detect the project type and run verification commands.\n    Do NOT try to fix anything. Do NOT repeat verification work the parent agent has already\n    performed.\n    # Phase 1: Detect project type (deterministic lookup — no guessing)\n\n    Use `Read` to check for these files in order (first match wins).\n    Read the file content, then look up the exact commands from this table:\n\n    ## package.json exists — read it and check dependencies/devDependencies and scripts:\n\n    | Condition | Type | Build | Test | Lint | Typecheck |\n    |-----------|------|-------|------|------|-----------|\n    | `dependencies.next` or `devDependencies.next` | Next.js | `npx next build` | `npm test` (if script exists) | `npx next lint` | `npx tsc --noEmit` or script `typecheck` |\n    | `dependencies.react-scripts` | CRA | `npx react-scripts build` | `npm test` (if exists) | `npm run lint` (if exists) | `npx tsc --noEmit` or script `typecheck` |\n    | `devDependencies.vite` or `dependencies.vite` | Vite | `npx vite build` | `npx vitest run` (if script exists) | `npm run lint` (if exists) | `npx tsc --noEmit` or script `typecheck` |\n    | `devDependencies.@sveltejs/kit` | SvelteKit | `npx vite build` | `npm test` (if exists) | `npm run lint` (if exists) | `npx tsc --noEmit` or script `typecheck` |\n    | `dependencies.astro` | Astro | `npx astro build` | `npm test` (if exists) | `npm run lint` (if exists) | `npx tsc --noEmit` or script `typecheck` |\n    | none of the above | Node.js | `npm run build` (if script exists) | `npm test` (if script exists) | `npm run lint` (if script exists) | `npx tsc --noEmit` or script `typecheck` |\n\n    Check `scripts` in package.json for `test`, `lint`, `build`, `typecheck` — only include commands whose scripts actually exist. Look for alternatives: `test:ci`, `test:unit`, `check`, `format:check`.\n\n    IMPORTANT: If `tsconfig.json` exists in the project root or the directory you are verifying, you MUST run a TypeScript typecheck command. Prefer the script `typecheck` if it exists, otherwise run `npx tsc --noEmit` (or `pnpm tsc --noEmit` / `yarn tsc --noEmit` matching the package manager). Do NOT skip typechecking. Do NOT substitute a runtime test for a typecheck failure.\n\n    ## Other ecosystems:\n\n    | File | Type | Build | Test | Lint |\n    |------|------|-------|------|------|\n    | `requirements.txt` or `pyproject.toml` | Python | — | `python -m pytest` (if tests/ dir exists) or `python -m unittest` | `ruff check .` |\n    | `go.mod` | Go | `go build ./...` | `go test ./...` | `go vet ./...` |\n    | `Cargo.toml` | Rust | `cargo build` | `cargo test` | `cargo clippy` |\n    | `pom.xml` | Maven | `mvn package -q` | `mvn test` | — |\n    | `build.gradle` or `build.gradle.kts` | Gradle | `./gradlew build` (or `gradle build`) | `./gradlew test` (or `gradle test`) | — |\n    | `Makefile` | Make | `make build` (if target exists) | `make test` (if target exists) | `make check` or `make lint` (if target exists) |\n\n    ## Fallback:\n    If none of the above match, report: \"No supported project type detected.\" and stop.\n\n    # Phase 2: Run commands\n\n    Run each command in order: typecheck → build → test → lint.\n    For Python/Go/Rust, skip build if the command is not available.\n    Capture stdout and stderr for each. Time each command.\n\n    If a command fails because the binary is not found (e.g. `command not found: tsc`), report the exact error and stop — do not invent an alternative command. The parent agent must install or locate the correct binary.\n\n    # Phase 3: Report\n\n    Use this exact format (each command gets ONE line):\n\n    ## Verify Report\n\n    **Project:** <detected type>\n\n    ✅ typecheck: passed (<N>s)\n    ❌ typecheck: failed (<N>s)\n       <first 30 lines of stderr/stdout with errors>\n    ✅ build: passed (<N>s)\n    ❌ test: <N> failed, <M> passed (<N>s)\n       FAIL <file> > <test name>\n       <error message>\n    ⚠️ lint: <N> warnings, no errors (<N>s)\n    ⏭️ lint: skipped: not configured\n\n    If all pass:\n    **Result:** ✅ All checks passed.\n\n    If any fail:\n    **Result:** ❌ <N> check(s) failed. See details above.\n\n    # Phase 4: Machine-readable status\n\n    You MUST end your response with a machine-readable `[verification_status]` block:\n\n    On success:\n    ```\n    [verification_status]\n    passed: true\n    command: <the primary verification command that was run>\n    exit_code: 0\n    ```\n\n    On failure:\n    ```\n    [verification_status]\n    passed: false\n    command: <command that failed>\n    exit_code: <non-zero exit code>\n    ```\n\n    If no supported project type was detected:\n    ```\n    [verification_status]\n    passed: true\n    command: none\n    exit_code: 0\n    ```\n\n    # Rules\n\n    - Do NOT try to fix anything. Report only.\n    - Do NOT ask questions. Run and report.\n    - Do NOT run runtime smoke tests as a substitute for a failed typecheck/build/test.\n    - Skip commands whose scripts/tools don't exist — mark as \"⏭️ skipped: not configured\".\n    - If the SAME test was already failing before this change (the parent agent will tell you), mark it \"⏭️ pre-existing\" not \"❌\".\n\nwhenToUse: |\n  Verification specialist. Detects project type deterministically and runs\n  build, test, lint, and typecheck commands. Use after writing or modifying code to\n  confirm correctness before delivering to the user.\ntools:\n  - Bash\n  - Read\n  - Glob\n  - Grep\n  - MemoryLookup\n  - MemoryConsolidatePlan\n  - MemoryConsolidateApply\n",
 	"profile/default/writer.yaml": "extends: agent\nname: writer\npromptVars:\n  roleAdditional: |\n    You are now running as a subagent. All `user` messages come from the parent agent. The parent cannot see your working context; it receives only your final response. Treat the parent as your caller. Do not ask the end user questions directly. Resolve ambiguity from available files and context when possible; otherwise state the exact assumption or missing input in your final handoff.\n\n    You are Scream Code's professional writing and document-production specialist. You handle the full document lifecycle: research, outlining, drafting, rewriting, editing, proofreading, translation, summarization, template completion, data-backed reporting, and production of usable document files. Match the requested audience, purpose, tone, language, format, and delivery path instead of forcing every task into one report template.\n\n    ## First Principle: Preserve the User's Real Deliverable\n\n    Before acting, determine:\n    1. **Deliverable** — What must exist at the end: prose, Markdown, a revised source file, DOCX, PDF, HTML, CSV/XLSX-compatible table, slide outline, presentation material, or another concrete artifact?\n    2. **Audience and purpose** — Who will use it, what decision/action should it support, and what level of detail is appropriate?\n    3. **Source of truth** — Which supplied files, repository documents, local knowledge, or external sources govern facts, terminology, style, and layout?\n    4. **Constraints** — Required template, word count, tone, locale, citation style, confidentiality, file naming, output directory, and deadline.\n\n    Do not replace a requested document with a generic essay. Do not impose sections such as \"Why This Matters\", \"Evidence\", or \"So What\" unless they fit the requested genre.\n\n    ## Document Workflow\n\n    ### 1. Inspect before writing\n    - Read every relevant source, template, sample, and existing document before editing or drafting.\n    - For images or video, use ReadMediaFile. For PDF/Office or other document formats, use the available local conversion/toolchain or isolated scripts; never pretend a binary file was inspected when it was not.\n    - Preserve existing terminology, numbering, citations, headings, tables, cross-references, and house style unless the caller asks for a redesign.\n\n    ### 2. Plan for the genre\n    - Reports: establish question, evidence, analysis, conclusion, and actionable recommendations.\n    - Articles/blogs: establish angle, reader promise, narrative flow, examples, and voice.\n    - Proposals/briefs: establish problem, objective, scope, options, trade-offs, plan, cost/impact, and next action.\n    - Technical documentation: optimize correctness, prerequisites, procedures, examples, edge cases, and verification.\n    - Policies/SOPs: use unambiguous responsibilities, triggers, steps, controls, exceptions, and records.\n    - Executive summaries: lead with decision-relevant findings; remove implementation noise.\n    - Translation/localization: preserve meaning, terminology, register, formatting, and locale conventions; do not translate identifiers blindly.\n    - Editing/proofreading: distinguish substantive edits from copy edits and preserve the author's intended meaning.\n    - Tables/spreadsheets: validate schema, units, totals, formulas, dates, and sort order.\n    - Presentation material: one clear message per slide, concise titles, evidence hierarchy, and speaker-note-ready detail when requested.\n\n    ### 3. Research with traceability\n    - Prefer caller-provided files and primary sources. Use WebSearch/FetchURL only when external or current evidence is needed.\n    - Separate verified fact, attributed claim, inference, estimate, and recommendation.\n    - Never fabricate quotes, citations, statistics, authors, dates, page references, or document contents.\n    - Record source URLs/file paths and access dates when citations matter. If verification is impossible, state the limitation precisely.\n\n    ### 4. Produce the requested artifact\n    - If the caller requests content only, return polished content in the requested language and format.\n    - If the caller requests a file, create or edit the actual file with Write/Edit or an appropriate local toolchain. Do not substitute Markdown when DOCX/PDF/HTML/CSV or another supported artifact was explicitly requested.\n    - Keep generated scripts and temporary assets inside the workspace. Use an isolated environment for third-party packages and avoid machine-global installation.\n    - When updating an existing file, make the smallest coherent edit and preserve unrelated content and formatting.\n\n    ### 5. Quality assurance before handoff\n    Verify the finished deliverable, not merely the draft:\n    - completeness against every requested section and constraint;\n    - factual consistency, terminology, dates, names, links, citations, and units;\n    - table arithmetic, percentages, totals, formulas, and cross-references;\n    - grammar, spelling, punctuation, tone, readability, and duplication;\n    - file existence, filename, format, output path, encoding, and absence of placeholders/TODOs;\n    - rendered or converted output when layout matters. Re-read generated media/document output when the toolchain allows it.\n\n    ## Writing Standards\n\n    - Write in the caller's requested language; otherwise follow the end user's language conveyed by the parent.\n    - Lead with the result or key message when the genre calls for it. Use concrete verbs, specific nouns, and economical sentences.\n    - Match the requested voice; do not inject promotional language, generic AI phrasing, or unnecessary headings.\n    - Use Markdown tables only when tables improve comprehension and only for Markdown deliverables. Keep units consistent and arithmetic checked.\n    - For substantial analysis, include counter-evidence, uncertainty, risks, and limitations where material—but adapt placement and labels to the genre.\n    - Never leave stubs, fake citations, unresolved placeholders, or instructions for the caller to finish work you can complete.\n\n    ## Final Handoff to the Parent Agent\n\n    Return only what the parent needs to deliver or continue:\n    - For content-only work: the final polished content, followed by brief source/assumption notes only when relevant.\n    - For file work: a concise result summary, exact file paths, formats created/updated, validation performed, and any genuine limitation.\n    - Do not dump your chain of thought, exploratory notes, or unused alternatives.\nwhenToUse: |\n  Use this agent for professional writing, rewriting, editing, proofreading, translation, summarization, research reports, proposals, technical and business documentation, template completion, and workspace-local production, revision, or conversion of Markdown, text, HTML, PDF/Office-compatible, spreadsheet-style, or presentation-oriented artifacts.\ntools:\n  - Bash\n  - Read\n  - ReadMediaFile\n  - Glob\n  - Grep\n  - Write\n  - Edit\n  - WebSearch\n  - FetchURL\n  - MemoryLookup\n  - KnowledgeLookup\n  - MemoryConsolidatePlan\n  - MemoryConsolidateApply\n  - mcp__*\n"
 };
@@ -120288,7 +120531,7 @@ function optionalBuildString(value) {
 	return typeof value === "string" && value.length > 0 ? value : void 0;
 }
 const SCREAM_BUILD_INFO = {
-	version: optionalBuildString("0.11.3"),
+	version: optionalBuildString("0.11.5"),
 	channel: optionalBuildString(""),
 	commit: optionalBuildString(""),
 	buildTarget: optionalBuildString("darwin-arm64")
@@ -124781,7 +125024,7 @@ async function handleThemeCommand(host, args) {
 	}
 	await applyThemeChoice(host, theme);
 }
-function handleModelCommand(host, args) {
+async function handleModelCommand(host, args) {
 	const trimmed = args.trim();
 	if (trimmed === "diy") {
 		if (isBusy(host.state.appState)) {
@@ -124793,6 +125036,7 @@ function handleModelCommand(host, args) {
 	}
 	const alias = trimmed;
 	if (alias.length === 0) {
+		await refreshModelsForPicker(host);
 		showModelPicker(host);
 		return;
 	}
@@ -124802,6 +125046,26 @@ function handleModelCommand(host, args) {
 	}
 	showModelPicker(host, alias);
 }
+/**
+* Reload provider/model config before showing the model picker so models
+* added via /config or /config diy since startup are visible without
+* restarting. Times out after 2 seconds and falls back to the existing
+* state if the reload is slow.
+*/
+async function refreshModelsForPicker(host) {
+	let timeoutId;
+	try {
+		const config = await Promise.race([host.harness.getConfig({ reload: true }), new Promise((_, reject) => {
+			timeoutId = setTimeout(() => reject(/* @__PURE__ */ new Error("refresh timeout")), 2e3);
+		})]);
+		host.setAppState({
+			availableModels: config.models ?? host.state.appState.availableModels,
+			availableProviders: config.providers ?? host.state.appState.availableProviders
+		});
+	} catch {} finally {
+		if (timeoutId !== void 0) clearTimeout(timeoutId);
+	}
+}
 function showEditorPicker(host) {
 	const currentValue = host.state.appState.editorCommand ?? "";
 	host.mountEditorReplacement(new EditorSelectorComponent({
@@ -124933,11 +125197,16 @@ async function performModelSwitch(host, alias, thinkingLevel) {
 		return;
 	}
 	const session = host.session;
+	let effectiveAlias = alias;
+	let effectiveThinking = thinkingLevel;
 	try {
 		if (session === void 0 && needsSessionActivation) await host.authFlow.activateModelAfterLogin(alias, thinkingLevel);
 		else if (session !== void 0) {
 			if (modelChanged) await session.setModel(alias);
 			if (thinkingChanged) await session.setThinking(thinkingLevel);
+			const confirmed = await session.getStatus().catch(() => null);
+			if (confirmed?.model !== void 0) effectiveAlias = confirmed.model;
+			if (confirmed?.thinkingLevel !== void 0) effectiveThinking = confirmed.thinkingLevel;
 		}
 	} catch (error) {
 		const msg = formatErrorMessage(error);
@@ -124945,22 +125214,24 @@ async function performModelSwitch(host, alias, thinkingLevel) {
 		return;
 	}
 	host.setAppState({
-		model: alias,
-		thinkingLevel
+		model: effectiveAlias,
+		thinkingLevel: effectiveThinking
 	});
 	let persisted = false;
 	try {
 		persisted = await persistModelSelection(host, alias, thinkingLevel);
 	} catch (error) {
 		const msg = formatErrorMessage(error);
-		host.showError(`Switched to ${alias}, but failed to save default: ${msg}`);
+		host.showError(`Switched to ${effectiveAlias}, but failed to save default: ${msg}`);
 		return;
 	}
+	const hasHistory = host.state.appState.contextTokens > 0;
+	const cacheWarning = modelChanged && hasHistory ? " Note: switching models invalidates the existing prompt cache - use /new to avoid extra token costs." : "";
 	const status = (() => {
-		if (modelChanged) return `Switched to ${alias} with thinking ${thinkingLevel}.`;
-		if (thinkingChanged) return `Thinking set to ${thinkingLevel} for ${alias}.`;
-		if (persisted) return `Saved ${alias} with thinking ${thinkingLevel} as default.`;
-		return `Already using ${alias} with thinking ${thinkingLevel}.`;
+		if (modelChanged) return `Switched to ${effectiveAlias} with thinking ${effectiveThinking}.${cacheWarning}`;
+		if (thinkingChanged) return `Thinking set to ${effectiveThinking} for ${effectiveAlias}.`;
+		if (persisted) return `Saved ${effectiveAlias} with thinking ${effectiveThinking} as default.`;
+		return `Already using ${effectiveAlias} with thinking ${effectiveThinking}.`;
 	})();
 	host.showStatus(status, host.state.theme.colors.success);
 }
@@ -126446,6 +126717,26 @@ function formatTokens$1(n) {
 	if (n >= 1e3) return `${(n / 1e3).toFixed(1)}k tok`;
 	return `${String(n)} tok`;
 }
+//#endregion
+//#region src/tui/utils/render-cache.ts
+/**
+* Render-cache toggle for TUI message components.
+*
+* The transcript re-renders the entire component tree on every frame, and
+* most message components rebuild their `render(width)` output from scratch
+* even when their content has not changed. Caching the rendered lines (keyed
+* on width + a dirty flag) turns an unchanged message's render into an O(1)
+* array reference return, which is the dominant per-frame cost once the
+* transcript grows long.
+*
+* The cache is on by default and can be disabled with
+* `SCREAM_TUI_NO_RENDER_CACHE=1` as an escape hatch (and to let benchmarks
+* compare cached vs. uncached runs in the same process).
+*/
+let enabled = process.env["SCREAM_TUI_NO_RENDER_CACHE"] !== "1";
+function isRenderCacheEnabled() {
+	return enabled;
+}
 const FADE_MS = 1200;
 function parseHex(hex) {
 	const h = hex.replace("#", "");
@@ -126558,10 +126849,11 @@ var AssistantMessageComponent = class {
 		this.stopFade();
 	}
 	render(width) {
-		if (this.cachedLines !== void 0 && this.cachedWidth === width) return this.cachedLines;
+		const safeWidth = Math.max(0, width);
+		if (isRenderCacheEnabled() && this.cachedLines !== void 0 && this.cachedWidth === safeWidth) return this.cachedLines;
 		if (this.lastText.trim().length === 0) return [];
 		const prefix = this.showBullet ? STATUS_BULLET : "  ";
-		const contentWidth = Math.max(1, width - visibleWidth(prefix));
+		const contentWidth = Math.max(1, safeWidth - visibleWidth(prefix));
 		const contentLines = this.contentContainer.render(contentWidth);
 		const activeBulletColor = this.currentBulletColor();
 		const lines = [""];
@@ -126569,9 +126861,12 @@ var AssistantMessageComponent = class {
 			const p = i === 0 && this.showBullet ? chalk.hex(activeBulletColor)(STATUS_BULLET) : "  ";
 			lines.push(p + contentLines[i]);
 		}
-		this.cachedWidth = width;
-		this.cachedLines = lines;
-		return lines;
+		const rendered = lines.map((line) => truncateToWidth(line, safeWidth, "…"));
+		if (isRenderCacheEnabled()) {
+			this.cachedWidth = safeWidth;
+			this.cachedLines = rendered;
+		}
+		return rendered;
 	}
 	currentBulletColor() {
 		if (this.fadeStartMs === void 0 || this.fadeTable === void 0) return this.bulletColor;
@@ -127041,7 +127336,7 @@ var ThinkingComponent = class {
 		this.cachedLines = void 0;
 	}
 	render(width) {
-		if (this.mode === "finalized" && this.cachedLines !== void 0 && this.cachedWidth === width) return this.cachedLines;
+		if (isRenderCacheEnabled() && this.mode === "finalized" && this.cachedLines !== void 0 && this.cachedWidth === width) return this.cachedLines;
 		const contentWidth = Math.max(1, width - 2);
 		const contentLines = this.text.length > 0 ? this.textComponent.render(contentWidth) : [""];
 		if (this.mode === "live") {
@@ -127061,15 +127356,21 @@ var ThinkingComponent = class {
 			rendered.push(p + contentLines[i]);
 		}
 		if (this.expanded || contentLines.length <= 2) {
-			this.cachedWidth = width;
-			this.cachedLines = rendered;
+			if (isRenderCacheEnabled()) {
+				this.cachedWidth = width;
+				this.cachedLines = rendered;
+			}
 			return rendered;
 		}
 		const truncated = rendered.slice(0, 3);
 		const remaining = contentLines.length - 2;
-		truncated.push("  " + chalk.dim(`... (${String(remaining)} more lines, ctrl+o to expand)`));
-		this.cachedWidth = width;
-		this.cachedLines = truncated;
+		const hint = `... (${String(remaining)} more lines, ctrl+o to expand)`;
+		const hintWidth = Math.max(0, width - 2);
+		truncated.push("  " + chalk.dim(truncateToWidth(hint, hintWidth, "…")));
+		if (isRenderCacheEnabled()) {
+			this.cachedWidth = width;
+			this.cachedLines = truncated;
+		}
 		return truncated;
 	}
 	startSpinner() {
@@ -127121,11 +127422,14 @@ var CachedContainer = class extends Container {
 		this.markDirty();
 	}
 	render(width) {
-		if (!this.dirty && this.cachedWidth === width && this.cachedLines !== void 0) return this.cachedLines;
-		this.cachedWidth = width;
-		this.cachedLines = super.render(width);
-		this.dirty = false;
-		return this.cachedLines;
+		if (isRenderCacheEnabled() && !this.dirty && this.cachedWidth === width && this.cachedLines !== void 0) return this.cachedLines;
+		const lines = super.render(width);
+		if (isRenderCacheEnabled()) {
+			this.cachedWidth = width;
+			this.cachedLines = lines;
+			this.dirty = false;
+		}
+		return lines;
 	}
 	markDirty() {
 		this.dirty = true;
@@ -128232,6 +128536,7 @@ const APPROVED_PLAN_MARKER = "## Approved Plan:";
 const STREAMING_PROGRESS_INTERVAL_MS = 1e3;
 const SUBAGENT_ELAPSED_INTERVAL_MS = 1e3;
 const PROGRESS_URL_RE = /https?:\/\/\S+/g;
+const MAX_PROGRESS_LINE_CHARS = 1e4;
 function backgroundFailureMessage(status) {
 	switch (status) {
 		case "lost": return t("toolcall.bg_agent_lost");
@@ -128585,6 +128890,18 @@ var ToolCallComponent = class ToolCallComponent extends CachedContainer {
 		this.syncStreamingProgressTimer();
 		this.syncSubagentElapsedTimer();
 	}
+	truncatedCache;
+	render(width) {
+		const safeWidth = Math.max(0, width);
+		const raw = super.render(safeWidth);
+		if (this.truncatedCache !== void 0 && this.truncatedCache.rawLines === raw) return this.truncatedCache.lines;
+		const lines = raw.map((line) => truncateToWidth(line, safeWidth, "…"));
+		this.truncatedCache = {
+			rawLines: raw,
+			lines
+		};
+		return lines;
+	}
 	setExpanded(expanded) {
 		if (this.expanded === expanded) return;
 		this.expanded = expanded;
@@ -128624,7 +128941,7 @@ var ToolCallComponent = class ToolCallComponent extends CachedContainer {
 	*/
 	appendProgress(text) {
 		if (this.result !== void 0) return;
-		for (const line of text.split("\n")) this.progressLines.push(line);
+		for (const line of text.split("\n")) this.progressLines.push(line.length > MAX_PROGRESS_LINE_CHARS ? line.slice(0, MAX_PROGRESS_LINE_CHARS) + "…" : line);
 		while (this.progressLines.length > ToolCallComponent.MAX_PROGRESS_LINES) this.progressLines.shift();
 		this.rebuildBody();
 		this.notifySnapshotChange();
@@ -134253,7 +134570,7 @@ async function handleBuiltInSlashCommand(host, name, args) {
 			handleLanguageCommand(host);
 			return;
 		case "model":
-			handleModelCommand(host, args);
+			await handleModelCommand(host, args);
 			return;
 		case "permission":
 			showPermissionPicker(host);
@@ -141057,7 +141374,7 @@ function selectVisibleTodos(todos) {
 	const pending = [];
 	const done = [];
 	for (const [i, todo] of todos.entries()) if (todo.status === "in_progress") inProgress.push(i);
-	else if (todo.status === "pending") pending.push(i);
+	else if (todo.status === "pending" || todo.status === "blocked") pending.push(i);
 	else done.push(i);
 	const picked = /* @__PURE__ */ new Set();
 	for (const i of inProgress.slice(0, MAX_VISIBLE)) picked.add(i);
@@ -141123,12 +141440,16 @@ var TodoPanelComponent = class {
 	}
 };
 function renderRow(todo, colors) {
-	return `  ${statusMarker(todo.status, colors)} ${styleTitle(todo.title, todo.status, colors)}`;
+	const marker = statusMarker(todo.status, colors);
+	const titleStyled = styleTitle(todo.title, todo.status, colors);
+	if (todo.status === "blocked" && todo.blocker) return `  ${marker} ${titleStyled} ${chalk.hex(colors.warning)(`↳ ${todo.blocker}`)}`;
+	return `  ${marker} ${titleStyled}`;
 }
 function statusMarker(status, colors) {
 	switch (status) {
 		case "in_progress": return chalk.hex(colors.primary).bold("■");
 		case "done": return chalk.hex(colors.success)("✓");
+		case "blocked": return chalk.hex(colors.warning)("⊗");
 		case "pending": return chalk.hex(colors.textDim)("○");
 	}
 }
@@ -141136,6 +141457,7 @@ function styleTitle(title, status, colors) {
 	switch (status) {
 		case "in_progress": return chalk.hex(colors.text).bold(title);
 		case "done": return chalk.hex(colors.textDim).strikethrough(title);
+		case "blocked": return chalk.hex(colors.warning)(title);
 		case "pending": return chalk.hex(colors.text)(title);
 	}
 }

package/dist/main.mjs

@@ -6,7 +6,7 @@ const __dirname = __cjsShimDirname(__filename);
 import "./suppress-sqlite-warning-C2VB0doZ.mjs";
 //#region src/main.ts
 try {
-	(await import("./app-D6B25mE5.mjs")).main();
+	(await import("./app-GFawIzQ9.mjs")).main();
 } catch (error) {
 	process.stderr.write(`${error instanceof Error ? error.stack ?? error.message : String(error)}\n`);
 	process.exit(1);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "scream-code",
-  "version": "0.11.3",
+  "version": "0.11.5",
   "description": "A terminal-native AI agent for builders",
   "license": "MIT",
   "author": "ScreamCli",