scream-code 0.12.6 → 0.12.7

package/dist/{app-NZLMIJND.mjs → app-DoxADJlo.mjs}

@@ -7,7 +7,7 @@ import { i as __require, o as __toESM, r as __exportAll, t as __commonJSMin } fr
 import "./suppress-sqlite-warning-C2VB0doZ.mjs";
 import { C as join$1, D as resolve$1, E as relative$1, S as isAbsolute$1, T as parse$7, a as isSupportedFile, b as basename$1, i as ingestFile, r as ingestDirectory, t as multiSearch, w as normalize, x as dirname$2, y as KnowledgeStore } from "./src-BH9W5k24.mjs";
 import { t as require_base64_js } from "./base64-js-DzVmk6Nb.mjs";
-import { a as setLocale, i as getLocale, n as assertScreamHostIdentity, o as t, r as createScreamDefaultHeaders, t as TextInputDialogComponent } from "./text-input-dialog-BI6Q02oM.mjs";
+import { a as setLocale, i as getLocale, n as assertScreamHostIdentity, o as t, r as createScreamDefaultHeaders, t as TextInputDialogComponent } from "./text-input-dialog-9ywV88Lh.mjs";
 import { createRequire } from "node:module";
 import { createHash, randomBytes, randomInt, randomUUID } from "node:crypto";
 import * as fs$1 from "node:fs/promises";
@@ -5600,7 +5600,7 @@ var _BetaToolRunner_instances, _BetaToolRunner_consumed, _BetaToolRunner_mutated
 /**
 * Just Promise.withResolvers(), which is not available in all environments.
 */
-function promiseWithResolvers() {
+function promiseWithResolvers$1() {
 	let resolve;
 	let reject;
 	return {
@@ -5646,7 +5646,7 @@ var BetaToolRunner = class {
 			...options,
 			headers: buildHeaders$2([{ "x-stainless-helper": helperValue }, options?.headers])
 		}, "f");
-		__classPrivateFieldSet$1(this, _BetaToolRunner_completion, promiseWithResolvers(), "f");
+		__classPrivateFieldSet$1(this, _BetaToolRunner_completion, promiseWithResolvers$1(), "f");
 		if (params.compactionControl?.enabled) console.warn("Anthropic: The `compactionControl` parameter is deprecated and will be removed in a future version. Use server-side compaction instead by passing `edits: [{ type: \"compact_20260112\" }]` in the params passed to `toolRunner()`. See https://platform.claude.com/docs/en/build-with-claude/compaction");
 	}
 	async *[(_BetaToolRunner_consumed = /* @__PURE__ */ new WeakMap(), _BetaToolRunner_mutated = /* @__PURE__ */ new WeakMap(), _BetaToolRunner_state = /* @__PURE__ */ new WeakMap(), _BetaToolRunner_options = /* @__PURE__ */ new WeakMap(), _BetaToolRunner_message = /* @__PURE__ */ new WeakMap(), _BetaToolRunner_toolResponse = /* @__PURE__ */ new WeakMap(), _BetaToolRunner_completion = /* @__PURE__ */ new WeakMap(), _BetaToolRunner_iterationCount = /* @__PURE__ */ new WeakMap(), _BetaToolRunner_instances = /* @__PURE__ */ new WeakSet(), _BetaToolRunner_checkAndCompact = async function _BetaToolRunner_checkAndCompact() {
@@ -5742,7 +5742,7 @@ var BetaToolRunner = class {
 			__classPrivateFieldSet$1(this, _BetaToolRunner_consumed, false, "f");
 			__classPrivateFieldGet$1(this, _BetaToolRunner_completion, "f").promise.catch(() => {});
 			__classPrivateFieldGet$1(this, _BetaToolRunner_completion, "f").reject(error);
-			__classPrivateFieldSet$1(this, _BetaToolRunner_completion, promiseWithResolvers(), "f");
+			__classPrivateFieldSet$1(this, _BetaToolRunner_completion, promiseWithResolvers$1(), "f");
 			throw error;
 		}
 	}
@@ -76344,14 +76344,84 @@ function maybeStatusCode(error) {
 }
 //#endregion
 //#region ../../packages/agent-core/src/agent/context/projector.ts
-function project(history) {
-	const usable = history.filter((message) => {
+/** Synthetic error text used when a tool result is missing and must be
+* filled in so the provider accepts the message sequence. */
+const SYNTHETIC_TOOL_RESULT_TEXT = "<system>ERROR: The tool call did not complete because of an interruption. Do not assume the tool executed successfully, and do not invent its result.</system>";
+function project(history, options) {
+	return repairToolExchangeAdjacency(mergeAdjacentUserMessages(history.filter((message) => {
 		return message.partial !== true && !(message.role === "assistant" && message.content.length === 0 && message.toolCalls.length === 0);
-	});
-	const last = usable.at(-1);
-	return mergeAdjacentUserMessages(last?.role === "assistant" && last.toolCalls.length > 0 ? usable.slice(0, -1) : usable);
+	}), options?.onAnomaly), options);
 }
-function mergeAdjacentUserMessages(history) {
+/**
+* Closes every tool exchange whose assistant `tool_use` is not fully answered
+* by a matching `tool_result`. A mid-history orphan (a later user/assistant
+* message follows) can never be in-flight, so it is always closed by
+* synthesizing an error result. A trailing orphan is closed only when
+* `synthesizeMissing` is set — otherwise it is left for the trim / replay
+* synthesis. This prevents the provider error "must be followed by tool
+* messages responding to each tool_call_id" after an interruption (e.g. a
+* network drop mid-batch).
+*/
+function repairToolExchangeAdjacency(messages, options) {
+	let lastNonToolIndex = messages.length - 1;
+	while (lastNonToolIndex >= 0 && messages[lastNonToolIndex]?.role === "tool") lastNonToolIndex -= 1;
+	const out = [];
+	const consumed = /* @__PURE__ */ new Set();
+	for (let i = 0; i < messages.length; i++) {
+		if (consumed.has(i)) continue;
+		const message = messages[i];
+		if (message.role !== "assistant" || message.toolCalls.length === 0) {
+			out.push(message);
+			continue;
+		}
+		out.push(message);
+		const pending = new Set(message.toolCalls.map((toolCall) => toolCall.id));
+		let foreignBetween = false;
+		for (let j = i + 1; j < messages.length && pending.size > 0; j++) {
+			if (consumed.has(j)) continue;
+			const next = messages[j];
+			const toolCallId = next.toolCallId;
+			if (next.role === "tool" && toolCallId !== void 0 && pending.has(toolCallId)) {
+				out.push(next);
+				consumed.add(j);
+				pending.delete(toolCallId);
+				if (foreignBetween) options?.onAnomaly?.({
+					kind: "tool_result_reordered",
+					toolCallId
+				});
+			} else foreignBetween = true;
+		}
+		const isMidHistory = i < lastNonToolIndex;
+		if (options?.synthesizeMissing === true || isMidHistory) for (const missingId of pending) {
+			out.push(makeSyntheticToolResult(missingId));
+			options?.onAnomaly?.({
+				kind: "tool_result_synthesized",
+				toolCallId: missingId,
+				trailing: !isMidHistory
+			});
+		}
+	}
+	return out;
+}
+function makeSyntheticToolResult(toolCallId) {
+	return {
+		role: "tool",
+		content: [{
+			type: "text",
+			text: SYNTHETIC_TOOL_RESULT_TEXT
+		}],
+		toolCalls: [],
+		toolCallId
+	};
+}
+/**
+* Drops a trailing open tool exchange from a projected message list: when the
+* last assistant message carries tool calls whose results never arrived and
+* nothing follows, truncate from that batch so the compacted history does not
+* carry an unterminated exchange (which would be rejected or, if synthesized,
+* pollute the compaction prompt).
+*/
+function mergeAdjacentUserMessages(history, _onAnomaly) {
 	const out = [];
 	for (const message of history) {
 		const previous = out.at(-1);
@@ -76621,12 +76691,23 @@ function extractFileOpsFromMessage(message, ops) {
 		}
 	}
 }
+/** Projects a `FileOperations` accumulator into the persistent, sorted
+*  file lists stored on `CompactionResult`. Unlike `formatFileOperations`
+*  this keeps the full list (no `FILE_LIMIT` elision) so later compactions
+*  can merge the previous round's file context without losing entries. */
+function computeFileLists(ops) {
+	const modified = new Set([...ops.edited, ...ops.written]);
+	return {
+		readFiles: [...ops.read].toSorted(),
+		modifiedFiles: [...modified].toSorted()
+	};
+}
 const FILE_LIMIT = 20;
 function formatFileOperations(ops) {
 	const modified = new Set([...ops.edited, ...ops.written]);
-	const readOnly = [...ops.read].filter((f) => !modified.has(f)).sort();
-	const modifiedFiles = [...modified].sort();
-	const all = [...new Set([...readOnly, ...modifiedFiles])].sort();
+	const readOnly = [...ops.read].filter((f) => !modified.has(f)).toSorted();
+	const modifiedFiles = [...modified].toSorted();
+	const all = [...new Set([...readOnly, ...modifiedFiles])].toSorted();
 	if (all.length === 0) return "";
 	const mode = /* @__PURE__ */ new Map();
 	for (const f of readOnly) mode.set(f, "Read");
@@ -76769,6 +76850,10 @@ var FullCompaction = class {
 	*  limit → overflow → compact cycle from consuming the entire
 	*  maxCompactionPerTurn budget with marginal savings. */
 	reactiveAttempted = false;
+	/** File lists from the most recent successful compaction. Merged into the
+	*  next compaction's file operations so file context accumulates across
+	*  repeated compactions instead of being reset each round. */
+	lastCompactionFiles;
 	compacting = null;
 	_compactedHistory = [];
 	strategy;
@@ -77000,15 +77085,25 @@ var FullCompaction = class {
 			const recent = originalHistory.slice(compactedCount);
 			const messagesToCompactForOps = originalHistory.slice(0, compactedCount);
 			const fileOps = createFileOps();
+			if (this.lastCompactionFiles !== void 0 && extractPreviousSummary(originalHistory) !== null) {
+				for (const f of this.lastCompactionFiles.readFiles) fileOps.read.add(f);
+				for (const f of this.lastCompactionFiles.modifiedFiles) fileOps.edited.add(f);
+			} else this.lastCompactionFiles = void 0;
 			for (const msg of messagesToCompactForOps) extractFileOpsFromMessage(msg, fileOps);
 			const toolCallHistory = formatToolCallHistory(messagesToCompactForOps);
 			const processedSummary = this.postProcessSummary(summary, fileOps, toolCallHistory);
 			const tokensAfter = estimateTokens$1(processedSummary) + estimateTokensForMessages(recent);
+			const fileLists = computeFileLists(fileOps);
+			const MAX_PERSISTED_FILES = 100;
+			const readFiles = fileLists.readFiles.slice(0, MAX_PERSISTED_FILES);
+			const modifiedFiles = fileLists.modifiedFiles.slice(0, MAX_PERSISTED_FILES);
 			const result = {
 				summary: processedSummary,
 				compactedCount,
 				tokensBefore,
 				tokensAfter,
+				...readFiles.length > 0 ? { readFiles } : {},
+				...modifiedFiles.length > 0 ? { modifiedFiles } : {},
 				...isUpdate ? { isUpdate: true } : {}
 			};
 			this.markCompleted();
@@ -77017,6 +77112,7 @@ var FullCompaction = class {
 				result
 			});
 			this.agent.context.applyCompaction(result);
+			this.lastCompactionFiles = fileLists;
 			this.lowWaterMark = Math.floor(this.effectiveTokenCount * 1.1);
 			await this.extractAndStoreMemos(processedSummary);
 			this.triggerPostCompactHook(data, result);
@@ -77145,7 +77241,9 @@ var FullCompaction = class {
 	*/
 	postProcessSummary(summary, fileOps, toolCallHistory) {
 		const todos = this.agent.tools.storeData()["todo"] ?? [];
-		const sections = [summary.trim()];
+		const base = summary.trim().replaceAll(/<files>[\s\S]*?<\/files>\s*/g, "").trimEnd();
+		const sections = [];
+		if (base.length > 0) sections.push(base);
 		if (todos.length > 0) {
 			const lines = todos.map((t) => {
 				return `- [${t.status === "done" ? "x" : t.status === "in_progress" ? "-" : " "}] ${t.title}`;
@@ -78804,12 +78902,15 @@ var ContextMemory = class {
 	* message mutated (compaction summary, micro-compaction truncation, or a
 	* projection repair) and the cache broke from that index.
 	*
-	* Behavior is otherwise identical to the getter - this is observation
-	* only, it does not alter the messages returned.
+	* Unlike the read-only `messages` getter, this path closes any trailing
+	* in-flight tool call by synthesizing an error result (synthesizeMissing):
+	* these messages go straight to the provider, which rejects an assistant
+	* tool_calls message with no matching tool result (e.g. after a network
+	* drop mid-batch).
 	*/
 	messagesForLLM() {
 		this.agent.microCompaction.detect();
-		const messages = project(this.agent.microCompaction.compact(this.history));
+		const messages = project(this.agent.microCompaction.compact(this.history), { synthesizeMissing: true });
 		this.observePrefixStability(messages);
 		return messages;
 	}
@@ -78911,6 +79012,31 @@ var ContextMemory = class {
 	hasOpenToolExchange() {
 		return this.pendingToolResultIds.size > 0;
 	}
+	/**
+	* Defensive teardown for a live turn that ended — normally, cancelled, or
+	* failed — while recorded tool calls were still awaiting results (e.g. the
+	* batch's result dispatch died after a `tool.call` was already recorded,
+	* like a network drop mid-execution). Synthesizes an error result for each
+	* dangling call so the exchange closes: left open, the assistant tool_calls
+	* message would have no matching tool message and the next request would be
+	* rejected by the provider ("must be followed by tool messages responding
+	* to each tool_call_id"). No-op when the exchange is already closed.
+	*/
+	closeAbandonedToolExchange(output) {
+		if (this.pendingToolResultIds.size === 0) return 0;
+		const interruptedToolCallIds = [...this.pendingToolResultIds];
+		for (const toolCallId of interruptedToolCallIds) this.appendLoopEvent({
+			type: "tool.result",
+			parentUuid: toolCallId,
+			toolCallId,
+			result: {
+				output,
+				isError: true
+			}
+		});
+		this.flushDeferredMessagesIfToolExchangeClosed();
+		return interruptedToolCallIds.length;
+	}
 	pushHistory(...messages) {
 		this._history.push(...messages);
 		for (const message of messages) {
@@ -96333,6 +96459,11 @@ var ToolCallDeduplicator = class {
 };
 //#endregion
 //#region ../../packages/agent-core/src/agent/turn/index.ts
+/** Builds the error text synthesized for tool calls abandoned when a live
+* turn ends (cancelled, failed, or completed) before their results arrived. */
+function abandonedToolResultOutput(ended) {
+	return `Tool call did not complete: ${ended.reason === "cancelled" ? "the turn was cancelled" : ended.reason === "failed" ? `the turn failed${ended.error !== void 0 ? ` (${ended.error.message})` : ""}` : "the turn ended"} before its result was recorded. Do not assume the tool completed successfully.`;
+}
 const GOAL_CONTINUATION_PROMPT = [
 	"Continue working toward the active goal.",
 	"Keep the self-audit brief. Do not explore unrelated interpretations once the goal can be",
@@ -96567,6 +96698,11 @@ var TurnFlow = class {
 			turnId,
 			reason: "completed"
 		};
+		try {
+			this.agent.context.closeAbandonedToolExchange(abandonedToolResultOutput(ended));
+		} catch (error) {
+			console.error("closeAbandonedToolExchange failed", error);
+		}
 		this.agent.usage.endTurn();
 		this.agent.emitEvent(ended);
 		return ended;
@@ -96646,6 +96782,11 @@ var TurnFlow = class {
 				};
 			}
 		}
+		try {
+			this.agent.context.closeAbandonedToolExchange(abandonedToolResultOutput(ended));
+		} catch (error) {
+			console.error("closeAbandonedToolExchange failed", error);
+		}
 		if (this.currentId === turnId) this.agent.usage.endTurn();
 		this.agent.emitEvent(ended);
 		if (standalone && this.currentId === turnId) this.activeTurn = null;
@@ -102568,7 +102709,7 @@ var SessionSubagentHost = class {
 				text: childPrompt
 			}], origin);
 			await runChildTurnToCompletion(child, options.signal);
-			let result = lastAssistantText(child);
+			let result = lastAssistantText$1(child);
 			let remainingContinuations = SUMMARY_CONTINUATION_ATTEMPTS;
 			while (remainingContinuations > 0 && result.length < SUMMARY_MIN_LENGTH) {
 				remainingContinuations -= 1;
@@ -102578,7 +102719,7 @@ var SessionSubagentHost = class {
 					text: summary_continuation_default
 				}], origin);
 				await runChildTurnToCompletion(child, options.signal);
-				result = lastAssistantText(child);
+				result = lastAssistantText$1(child);
 			}
 			const usage = child.usage.data().total;
 			let findingsBlock = "";
@@ -102694,7 +102835,7 @@ async function runChildTurnToCompletion(child, signal) {
 function throwIfSubagentStoppedAtMaxTokens(stopReason) {
 	if (stopReason === "max_tokens") throw new Error(`${SUBAGENT_MAX_TOKENS_ERROR}.`);
 }
-function lastAssistantText(agent) {
+function lastAssistantText$1(agent) {
 	for (const message of [...agent.context.history].toReversed()) {
 		if (message.role !== "assistant") continue;
 		const text = message.content.filter((part) => part.type === "text").map((part) => part.text).join("");
@@ -120191,10 +120332,10 @@ var ScreamAuthFacade = class {
 };
 //#endregion
 //#region ../../packages/node-sdk/src/rpc.ts
-const MAIN_AGENT_ID$2 = "main";
+const MAIN_AGENT_ID$3 = "main";
 var SDKRpcClient = class {
 	core;
-	interactiveAgentId = MAIN_AGENT_ID$2;
+	interactiveAgentId = MAIN_AGENT_ID$3;
 	ready;
 	rpc;
 	eventListeners = /* @__PURE__ */ new Set();
@@ -120772,7 +120913,7 @@ function errorMessage$2(error) {
 }
 //#endregion
 //#region ../../packages/node-sdk/src/session.ts
-const MAIN_AGENT_ID$1 = "main";
+const MAIN_AGENT_ID$2 = "main";
 var Session = class {
 	id;
 	workDir;
@@ -121212,7 +121353,7 @@ var Session = class {
 		this.emit({
 			type: "session.meta.updated",
 			sessionId: this.id,
-			agentId: MAIN_AGENT_ID$1,
+			agentId: MAIN_AGENT_ID$2,
 			title: patch.title,
 			patch
 		});
@@ -121634,7 +121775,7 @@ function optionalBuildString(value) {
 	return typeof value === "string" && value.length > 0 ? value : void 0;
 }
 const SCREAM_BUILD_INFO = {
-	version: optionalBuildString("0.12.6"),
+	version: optionalBuildString("0.12.7"),
 	channel: optionalBuildString(""),
 	commit: optionalBuildString(""),
 	buildTarget: optionalBuildString("darwin-arm64")
@@ -122906,6 +123047,12 @@ const BUILTIN_SLASH_COMMANDS = [
 		description: "registry.logout_desc",
 		priority: 93
 	},
+	{
+		name: "eval",
+		aliases: [],
+		description: "registry.eval_desc",
+		priority: 92
+	},
 	{
 		name: "exit",
 		aliases: ["quit", "q"],
@@ -123022,7 +123169,7 @@ function getCtrlDHint() {
 function getCtrlCHint() {
 	return t("constant.ctrl_c_hint");
 }
-const MAIN_AGENT_ID = "main";
+const MAIN_AGENT_ID$1 = "main";
 const EXIT_CONFIRM_WINDOW_MS = 1500;
 function isManagedUsageProvider(providerKey) {
 	return providerKey === DEFAULT_OAUTH_PROVIDER_NAME;
@@ -127448,7 +127595,7 @@ async function guidedGoalSetup(host) {
 		host.showNotice(t("goal.storm_breaker"), t("goal.conflict_loop"));
 		return;
 	}
-	const { TextInputDialogComponent } = await import("./text-input-dialog-BViNsv5f.mjs");
+	const { TextInputDialogComponent } = await import("./text-input-dialog-BIXonNNb.mjs");
 	const initialDesc = await promptText(host, TextInputDialogComponent, {
 		title: t("goal.setup_title_initial"),
 		subtitle: t("goal.setup_desc_hint"),
@@ -127469,7 +127616,7 @@ async function guidedGoalSetup(host) {
 	await showGoalConfigWizard(host, session, confirmed.trim() || objective, false);
 }
 async function showGoalConfigWizard(host, session, objective, replace) {
-	const { TextInputDialogComponent } = await import("./text-input-dialog-BViNsv5f.mjs");
+	const { TextInputDialogComponent } = await import("./text-input-dialog-BIXonNNb.mjs");
 	const turnInput = await promptNumber(host, TextInputDialogComponent, {
 		title: t("goal.wizard_title", { objective }),
 		subtitle: t("goal.budget_turns_hint"),
@@ -136032,6 +136179,429 @@ async function handleKnowledgeCommand(host, _args) {
 	showMenu();
 }
 //#endregion
+//#region ../../packages/evals/src/judge.ts
+/** Runs every rule against the input; a rule passes when its check is true. */
+function judgeOutput(input, rules) {
+	const passed = [];
+	const failed = [];
+	for (const rule of rules) if (rule.check(input)) passed.push(rule.name);
+	else failed.push(rule.name);
+	return {
+		passed,
+		failed
+	};
+}
+//#endregion
+//#region ../../packages/evals/src/harness.ts
+/**
+* Minimal end-to-end eval harness built on the public node-sdk
+* `ScreamHarness`. Each eval run gets an isolated temp **workspace** so runs
+* never touch real project files; the scream home defaults to the real one so
+* the provider/API-key configuration in `~/.scream-code/config.toml` is
+* honored (pass `screamHome` to override for fully isolated runs). Model
+* selection follows `SCREAM_EVAL_MODEL` (e.g. `provider/model`);
+* when unset the harness throws with a clear message.
+*/
+/** The main agent id; only its events settle the eval turn wait. */
+const MAIN_AGENT_ID = "main";
+/** The current scream default model, when resolvable from config. */
+const DEFAULT_EVAL_MODEL = process.env["SCREAM_EVAL_MODEL"];
+/**
+* Runs a single end-to-end prompt against a fresh isolated session and
+* returns the final assistant text plus token usage. The session, workspace
+* and scream home are torn down afterwards.
+*/
+async function runEvalPrompt(input, options = {}) {
+	const workDir = options.workDir ?? await mkdtemp(join(tmpdir(), "scream-eval-ws-"));
+	const screamHome = options.screamHome;
+	const createdWorkDir = options.workDir === void 0;
+	let harness;
+	let session;
+	try {
+		for (const fixture of (typeof input === "string" ? [] : input.fixtures) ?? []) await writeFile(join(workDir, fixture.name), fixture.content, "utf-8");
+		const prompt = typeof input === "string" ? input : input.prompt;
+		const model = options.model ?? DEFAULT_EVAL_MODEL;
+		if (model === void 0 || model.length === 0) throw new Error("No eval model configured. Set SCREAM_EVAL_MODEL, e.g. SCREAM_EVAL_MODEL=provider/model pnpm eval");
+		const [provider, modelId, ...rest] = model.split("/");
+		const usageProvider = rest.length > 0 ? model : provider ?? "unknown";
+		const usageModel = modelId === void 0 || rest.length > 0 ? model : modelId;
+		harness = new ScreamHarness({ homeDir: screamHome });
+		session = await harness.createSession({
+			workDir,
+			model,
+			thinking: options.thinking ?? "off",
+			permission: "yolo"
+		});
+		const { timedOut } = await promptAndWaitForTurnEnd(session, prompt);
+		let output = "";
+		try {
+			output = lastAssistantText((await session.getContext()).history);
+		} catch {
+			output = "";
+		}
+		const total = (await session.getUsage().catch(() => void 0))?.total;
+		let verifiedFile = false;
+		if (options.verifyFileAfterTurn !== void 0) try {
+			verifiedFile = await readFile(join(workDir, options.verifyFileAfterTurn.path), "utf-8") === options.verifyFileAfterTurn.content;
+		} catch {
+			verifiedFile = false;
+		}
+		return {
+			output,
+			timedOut,
+			verifiedFile,
+			usage: {
+				provider: usageProvider,
+				model: usageModel,
+				inputTokens: total?.inputOther ?? 0,
+				outputTokens: total?.output ?? 0,
+				totalTokens: (total?.inputOther ?? 0) + (total?.output ?? 0) + (total?.inputCacheRead ?? 0) + (total?.inputCacheCreation ?? 0)
+			}
+		};
+	} finally {
+		if (session !== void 0) await session.close({ extractMemories: false }).catch(() => {});
+		if (createdWorkDir) await rm(workDir, {
+			recursive: true,
+			force: true
+		}).catch(() => {});
+	}
+}
+function lastAssistantText(history) {
+	for (let i = history.length - 1; i >= 0; i -= 1) {
+		const message = history[i];
+		if (message.role !== "assistant") continue;
+		const text = (message.content ?? []).filter((part) => part.type === "text" && part.text !== void 0).map((part) => part.text).join("");
+		if (text.trim().length > 0) return text.trim();
+	}
+	return "";
+}
+/**
+* Sends a prompt and resolves when the main agent's turn ends. `Session.prompt`
+* only enqueues the RPC request; the assistant reply arrives asynchronously as
+* events, so we must wait for the `turn.ended` event to know the run finished.
+*
+* A hard timeout guards against a turn that never ends (e.g. a model looping
+* through tools instead of converging, or a network interruption). On timeout
+* we cancel the session so the caller's cleanup path still runs instead of
+* leaking the session and temp workspace. 90s is enough to complete any
+* legitimate single-turn tool round-trip; longer waits are almost always a
+* stuck model, so we fail fast instead of burning tokens.
+*/
+const TURN_WAIT_TIMEOUT_MS = 9e4;
+/**
+* Resolves with `timedOut: true` when the turn hit the timeout (model looping
+* or stuck) instead of throwing, so callers can still inspect side effects
+* (e.g. a file the Write tool already created). Real errors (turn failed /
+* error event) still reject.
+*/
+async function promptAndWaitForTurnEnd(session, prompt) {
+	const { promise, resolve, reject } = promiseWithResolvers();
+	let activeTurnId;
+	let activeAgentId;
+	let settled = false;
+	let timeout;
+	let graceTimer;
+	/** Set once the timeout fires; guards against the cancel-triggered
+	*  `turn.ended('cancelled')` being misread as a real failure. */
+	let timeoutFired = false;
+	const finish = (result) => {
+		if (settled) return;
+		settled = true;
+		if (timeout !== void 0) clearTimeout(timeout);
+		if (graceTimer !== void 0) clearTimeout(graceTimer);
+		if (result instanceof Error) reject(result);
+		else resolve(result);
+	};
+	const unsubscribe = session.onEvent((event) => {
+		if (event.agentId !== MAIN_AGENT_ID) return;
+		if (event.type === "error") {
+			finish(/* @__PURE__ */ new Error(`${event.code}: ${event.message}`));
+			return;
+		}
+		if (event.type === "turn.started" && activeTurnId === void 0) {
+			activeTurnId = event.turnId;
+			activeAgentId = event.agentId;
+			return;
+		}
+		if (activeTurnId === void 0 || activeAgentId === void 0 || !("turnId" in event) || event.turnId !== activeTurnId || event.agentId !== activeAgentId) return;
+		if (event.type === "turn.ended") if (event.reason === "completed") finish({ timedOut: false });
+		else if (timeoutFired && event.reason === "cancelled") return;
+		else finish(/* @__PURE__ */ new Error(`Turn ended with reason: ${event.reason}`));
+	});
+	try {
+		await session.prompt(prompt);
+		timeout = setTimeout(() => {
+			timeoutFired = true;
+			session.cancel().catch(() => {});
+			graceTimer = setTimeout(() => {
+				finish({ timedOut: true });
+			}, 500);
+		}, TURN_WAIT_TIMEOUT_MS);
+		return await promise;
+	} finally {
+		unsubscribe();
+		if (timeout !== void 0) clearTimeout(timeout);
+		if (graceTimer !== void 0) clearTimeout(graceTimer);
+	}
+}
+function promiseWithResolvers() {
+	let resolve;
+	let reject;
+	return {
+		promise: new Promise((res, rej) => {
+			resolve = res;
+			reject = rej;
+		}),
+		resolve,
+		reject
+	};
+}
+const EVAL_CASES = [
+	{
+		id: "smoke",
+		name: "Smoke — answers a trivial question",
+		input: "What is the capital of France? Answer in one word.",
+		rules: [{
+			name: "answers the capital",
+			check: (output) => {
+				const lower = output.toLowerCase();
+				return lower.includes("paris") || lower.includes("巴黎");
+			}
+		}],
+		extra: (result) => {
+			const failures = [];
+			if (result.usage.totalTokens <= 0) failures.push("usage.totalTokens must be > 0");
+			if (result.usage.model.length === 0) failures.push("usage.model must be non-empty");
+			return failures;
+		}
+	},
+	{
+		id: "read-file",
+		name: "Regression — reads a file via the Read tool",
+		input: {
+			prompt: "Read the file data.txt in the current directory using the Read tool, then tell me what it contains.",
+			fixtures: [{
+				name: "data.txt",
+				content: "The quick brown fox jumps over the lazy dog."
+			}]
+		},
+		rules: [{
+			name: "reports file content",
+			check: (output) => output.includes("quick brown fox")
+		}]
+	},
+	{
+		id: "write-file",
+		name: "Regression — writes a file via the Write tool",
+		input: { prompt: "Create a file named output.txt in the current directory using the Write tool. Write exactly the text: hello eval world. Make exactly one Write tool call and do not call any other tool afterwards. After the tool completes, immediately answer with the exact text you wrote and nothing else." },
+		rules: [{
+			name: "Write tool landed the file",
+			check: (output) => output.includes("hello eval world")
+		}],
+		verifyFile: {
+			path: "output.txt",
+			content: "hello eval world"
+		},
+		extra: (result) => {
+			const failures = [];
+			if (!result.verifiedFile) failures.push("output.txt was not created with the exact content");
+			return failures;
+		}
+	},
+	{
+		id: "isolated-workspace",
+		name: "Regression — runs in an isolated temp workspace",
+		input: {
+			prompt: "Run \"pwd\" in a shell and answer with only the absolute path that the command printed.",
+			fixtures: [{
+				name: "marker.txt",
+				content: "isolated"
+			}]
+		},
+		rules: [{
+			name: "prints an absolute path",
+			check: (output) => output.match(/(?:\/[\w.\-/]+|[A-Za-z]:[\\/][\w.\\/-]*)/)?.[0] !== void 0
+		}, {
+			name: "path is not the repo",
+			check: (output) => !output.match(/(?:\/[\w.\-/]+|[A-Za-z]:[\\/][\w.\\/-]*)/)?.[0]?.includes("scream-code")
+		}]
+	}
+];
+/** Runs every eval case sequentially and returns a structured report. */
+async function runAllEvals(options = {}) {
+	const results = [];
+	let passedCount = 0;
+	let failedCount = 0;
+	for (const caseDef of EVAL_CASES) {
+		let result;
+		try {
+			result = await runEvalPrompt(caseDef.input, {
+				model: options.model,
+				verifyFileAfterTurn: caseDef.verifyFile
+			});
+		} catch (error) {
+			results.push({
+				id: caseDef.id,
+				name: caseDef.name,
+				passed: false,
+				failedRules: [],
+				passedRules: [],
+				extraFailures: [],
+				output: "",
+				error: error instanceof Error ? error.message : String(error),
+				timedOut: false,
+				toolCheckFailed: false
+			});
+			failedCount += 1;
+			options.onProgress?.({
+				completed: results.length,
+				total: EVAL_CASES.length,
+				currentId: caseDef.id,
+				currentName: caseDef.name,
+				passed: passedCount,
+				failed: failedCount
+			});
+			continue;
+		}
+		const judged = judgeOutput(result.output, caseDef.rules);
+		const extraFailures = caseDef.extra?.(result) ?? [];
+		const passed = judged.failed.length === 0 && extraFailures.length === 0;
+		if (passed) passedCount += 1;
+		else failedCount += 1;
+		const toolCheckFailed = caseDef.verifyFile !== void 0 && !result.verifiedFile;
+		results.push({
+			id: caseDef.id,
+			name: caseDef.name,
+			passed,
+			failedRules: judged.failed,
+			passedRules: judged.passed,
+			extraFailures,
+			output: result.output,
+			timedOut: result.timedOut,
+			toolCheckFailed
+		});
+		options.onProgress?.({
+			completed: results.length,
+			total: EVAL_CASES.length,
+			currentId: caseDef.id,
+			currentName: caseDef.name,
+			passed: passedCount,
+			failed: failedCount
+		});
+	}
+	return {
+		results,
+		passed: passedCount,
+		failed: failedCount
+	};
+}
+//#endregion
+//#region src/tui/commands/eval.ts
+const EVAL_PANEL_DISMISS_MS = 6e4;
+let activeEvalPanel;
+let activeEvalTimer;
+/** Guards against overlapping runs (each run calls a real model and costs
+*  tokens; a second /eval while one is running is almost certainly a mistake). */
+let evalRunning = false;
+function dismissEvalPanel(state) {
+	if (activeEvalTimer !== void 0) {
+		clearTimeout(activeEvalTimer);
+		activeEvalTimer = void 0;
+	}
+	if (activeEvalPanel !== void 0) {
+		state.transcriptContainer.removeChild(activeEvalPanel);
+		activeEvalPanel = void 0;
+		state.ui.requestRender();
+	}
+}
+function clearEvalPanelState(state) {
+	dismissEvalPanel(state);
+}
+/** Model selector for eval runs; falls back to the active session model. */
+function resolveEvalModel(host) {
+	return host.state.appState.model;
+}
+/**
+* Runs the end-to-end evals in the background (they call a real model and can
+* take minutes) and renders the per-case report when done. Returns immediately
+* so the TUI keeps responding.
+*/
+function runEvalCommand(host) {
+	if (evalRunning) {
+		host.showStatus(t("dispatch.eval_running"));
+		return;
+	}
+	const model = resolveEvalModel(host);
+	if (model === void 0 || model.length === 0) {
+		host.showError(t("dispatch.eval_no_model"));
+		return;
+	}
+	const sessionId = host.state.appState.sessionId;
+	evalRunning = true;
+	host.showStatus(t("dispatch.eval_started", { model }));
+	(async () => {
+		let summary;
+		try {
+			summary = await runAllEvals({
+				model,
+				onProgress: ({ completed, total, passed, failed, currentName }) => {
+					host.showStatus(t("dispatch.eval_progress", {
+						completed: String(completed),
+						total: String(total),
+						passed: String(passed),
+						failed: String(failed),
+						name: currentName
+					}));
+				}
+			});
+		} catch (error) {
+			evalRunning = false;
+			host.showError(t("dispatch.eval_failed", { error: error instanceof Error ? error.message : String(error) }));
+			return;
+		}
+		evalRunning = false;
+		if (sessionId !== void 0 && sessionId !== host.state.appState.sessionId) return;
+		try {
+			renderEvalSummary(host, summary);
+		} catch (error) {
+			host.showError(t("dispatch.eval_failed", { error: error instanceof Error ? error.message : String(error) }));
+		}
+	})();
+}
+function renderEvalSummary(host, summary) {
+	const colors = host.state.theme.colors;
+	const lines = [chalk.hex(colors.primary)(`${summary.passed} ${t("dispatch.eval_passed")} · ${summary.failed} ${t("dispatch.eval_failed_count")}`), ""];
+	for (const result of summary.results) {
+		const mark = result.passed ? "✓" : "×";
+		const color = result.passed ? colors.success : colors.error;
+		lines.push(`${chalk.hex(color)(mark)} ${result.name}`);
+		if (result.passed) continue;
+		lines.push(`  ${chalk.hex(colors.error)(suggestionFor(result))}`);
+		const details = [];
+		for (const rule of result.failedRules) details.push(rule);
+		for (const failure of result.extraFailures) details.push(failure);
+		if (result.error !== void 0) details.push(result.error);
+		if (result.output.length > 0) details.push(`output: ${result.output.slice(0, 160)}`);
+		for (const detail of details) lines.push(`  ${chalk.hex(colors.textDim)(`· ${detail}`)}`);
+	}
+	lines.push("");
+	dismissEvalPanel(host.state);
+	const panel = new UsagePanelComponent(lines, colors.primary, " Eval ");
+	host.state.transcriptContainer.addChild(panel);
+	activeEvalPanel = panel;
+	activeEvalTimer = setTimeout(() => {
+		dismissEvalPanel(host.state);
+	}, EVAL_PANEL_DISMISS_MS);
+	host.state.ui.requestRender();
+}
+/** Picks a user-facing self-check hint based on how the case failed. */
+function suggestionFor(result) {
+	if (result.error !== void 0) return t("dispatch.eval_suggest_error");
+	if (result.timedOut) return t("dispatch.eval_suggest_timeout");
+	if (result.toolCheckFailed) return t("dispatch.eval_suggest_tool");
+	return t("dispatch.eval_suggest_answer");
+}
+//#endregion
 //#region src/tui/commands/dispatch.ts
 function dispatchInput(host, text) {
 	if (parseSlashInput(text) !== null) {
@@ -136205,6 +136775,9 @@ async function handleBuiltInSlashCommand(host, name, args) {
 		case "logout":
 			await handleLogoutCommand(host);
 			return;
+		case "eval":
+			runEvalCommand(host);
+			return;
 		case "memory":
 			await handleMemoryCommand(host, args);
 			return;
@@ -144537,7 +145110,7 @@ var SessionManager$1 = class {
 	}
 	resetSessionRuntime() {
 		this.host.state.queuedMessages = [];
-		this.host.harness.interactiveAgentId = MAIN_AGENT_ID;
+		this.host.harness.interactiveAgentId = MAIN_AGENT_ID$1;
 		this.host.streamingUI.discardPending();
 		this.host.streamingUI.resetToolCallState();
 		this.host.streamingUI.resetToolUi();
@@ -147194,6 +147767,7 @@ var ScreamTUI = class {
 		this.sessionEventHandler.stopAllMcpServerStatusSpinners();
 		clearGoalState();
 		clearInfoPanelState(this.state);
+		clearEvalPanelState(this.state);
 		this.state.terminal.write("\x1B[3J");
 		this.transcriptController.clearAndRedraw();
 		this.state.ui.requestRender(true);

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-NZLMIJND.mjs")).main();
+	(await import("./app-DoxADJlo.mjs")).main();
 } catch (error) {
 	process.stderr.write(`${error instanceof Error ? error.stack ?? error.message : String(error)}\n`);
 	process.exit(1);

package/dist/{text-input-dialog-BI6Q02oM.mjs → text-input-dialog-9ywV88Lh.mjs}

@@ -1058,6 +1058,17 @@ const dictionaries = {
 		"dispatch.tasks_browser_failed": "打开任务浏览器失败：{error}",
 		"dispatch.usage_failed": "显示使用情况失败：{error}",
 		"dispatch.status_failed": "显示状态报告失败：{error}",
+		"dispatch.eval_no_model": "没有可用的模型。请先通过 /connect 或 /config 配置模型。",
+		"dispatch.eval_started": "正在运行端到端评测（模型：{model}）…完成后将显示结果",
+		"dispatch.eval_running": "评测已在运行中，请等待完成",
+		"dispatch.eval_progress": "评测 {completed}/{total}：{name}（已通过 {passed}，未通过 {failed}）…",
+		"dispatch.eval_passed": "通过",
+		"dispatch.eval_failed_count": "未通过",
+		"dispatch.eval_failed": "评测失败：{error}",
+		"dispatch.eval_suggest_error": "可能是网络连接、API 配置或账户余额的问题，请检查后重试",
+		"dispatch.eval_suggest_timeout": "任务执行超时，可能是网络较慢或模型响应时间较长，可更换模型后重试",
+		"dispatch.eval_suggest_tool": "文件读写工具执行未成功，请检查工作目录权限；若重试后仍失败，请反馈",
+		"dispatch.eval_suggest_answer": "模型回答可能不完整或不正确，可尝试更换更强的模型后重试",
 		"constant.llm_not_set": "LLM 未设置，运行 /config 自定义模型配置",
 		"constant.no_active_session": "没有活动会话。运行 /config 自定义模型配置。",
 		"constant.ctrl_d_hint": "再次按 Ctrl+D 退出",
@@ -1121,6 +1132,7 @@ const dictionaries = {
 		"registry.update_desc": "手动更新 Scream Code 到最新版本",
 		"registry.version_desc": "显示版本信息",
 		"registry.logout_desc": "删除已配置的模型",
+		"registry.eval_desc": "运行端到端测试（Agent健康度检查）",
 		"registry.exit_desc": "退出应用",
 		"kw.hint_drag": "拖拽平移 · 滚轮缩放 · 单击展开/收起 · 双击详情",
 		"kw.detail_name": "名称",
@@ -2121,6 +2133,17 @@ const dictionaries = {
 		"dispatch.tasks_browser_failed": "Failed to open tasks browser: {error}",
 		"dispatch.usage_failed": "Failed to show usage: {error}",
 		"dispatch.status_failed": "Failed to show status report: {error}",
+		"dispatch.eval_no_model": "No model available. Configure one via /connect or /config first.",
+		"dispatch.eval_started": "Running end-to-end evals (model: {model})… results will appear when done",
+		"dispatch.eval_running": "An eval run is already in progress, please wait",
+		"dispatch.eval_progress": "Eval {completed}/{total}: {name} ({passed} passed, {failed} failed)…",
+		"dispatch.eval_passed": "passed",
+		"dispatch.eval_failed_count": "failed",
+		"dispatch.eval_failed": "Eval run failed: {error}",
+		"dispatch.eval_suggest_error": "This may be a network, API config, or account balance issue — please check and retry",
+		"dispatch.eval_suggest_timeout": "The task timed out; this is often a slow network or slow model response. Try a different model and retry",
+		"dispatch.eval_suggest_tool": "A file read/write tool did not complete. Please check directory permissions; if it still fails after retrying, please report it",
+		"dispatch.eval_suggest_answer": "The model answer may be incomplete or incorrect. Try a stronger model and retry",
 		"constant.llm_not_set": "LLM not set, run /config to configure model",
 		"constant.no_active_session": "No active session. Run /config to configure model.",
 		"constant.ctrl_d_hint": "Press Ctrl+D again to exit",
@@ -2184,6 +2207,7 @@ const dictionaries = {
 		"registry.update_desc": "Manually update Scream Code to latest version",
 		"registry.version_desc": "Show version info",
 		"registry.logout_desc": "Remove configured models",
+		"registry.eval_desc": "Run end-to-end tests (Agent health check)",
 		"registry.exit_desc": "Exit application",
 		"kw.hint_drag": "Drag to pan · Scroll to zoom · Click to expand/collapse · Double-click for details",
 		"kw.detail_name": "Name",

package/dist/{text-input-dialog-BViNsv5f.mjs → text-input-dialog-BIXonNNb.mjs}

@@ -3,5 +3,5 @@ import { fileURLToPath as __cjsShimFileURLToPath } from 'node:url';
 import { dirname as __cjsShimDirname } from 'node:path';
 const __filename = __cjsShimFileURLToPath(import.meta.url);
 const __dirname = __cjsShimDirname(__filename);
-import { t as TextInputDialogComponent } from "./text-input-dialog-BI6Q02oM.mjs";
+import { t as TextInputDialogComponent } from "./text-input-dialog-9ywV88Lh.mjs";
 export { TextInputDialogComponent };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "scream-code",
-  "version": "0.12.6",
+  "version": "0.12.7",
   "description": "A terminal-native AI agent for builders",
   "license": "MIT",
   "author": "ScreamCli",
@@ -74,6 +74,7 @@
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@scream-code/agent-core": "workspace:^",
     "@scream-code/config": "workspace:^",
+    "@scream-code/evals": "workspace:^",
     "@scream-code/knowledge": "workspace:*",
     "@scream-code/memory": "workspace:*",
     "@scream-code/scream-code-sdk": "workspace:^",