scream-code 0.5.1 → 0.5.2

package/dist/main.mjs

@@ -56244,6 +56244,15 @@ function buildGoalCompletionSummaryPrompt(goal) {
 		"Write a concise final message for the user. State that the goal is complete, summarize the main work completed, and mention any validation you ran. Do not call more goal tools."
 	].join("\n");
 }
+function buildGradingFeedbackPrompt(reason) {
+	return [
+		"Goal verification failed. An independent reviewer found that the completion criteria were not genuinely met.",
+		"",
+		`Reviewer feedback:\n${reason}`,
+		"",
+		"Address every issue listed above before calling UpdateGoal with complete again. Do not re-submit until all issues are resolved."
+	].join("\n");
+}
 function buildGoalBlockedReasonPrompt(goal) {
 	return [
 		buildGoalBlockedMessage(goal),
@@ -56279,13 +56288,28 @@ const UpdateGoalToolInputSchema = z.object({ status: z.enum([
 	"paused",
 	"blocked"
 ]).describe("The lifecycle status to set for the current goal.") }).strict();
+const MAX_GRADER_OUTPUT_CHARS = 4e3;
+function extractRecentOutput(history) {
+	const parts = [];
+	for (let i = history.length - 1; i >= 0; i--) {
+		const msg = history[i];
+		if (msg === void 0 || msg.role !== "assistant") continue;
+		const text = msg.content.filter((p) => p.type === "text" && typeof p.text === "string").map((p) => p.text).join("");
+		if (text) parts.unshift(text);
+		if (parts.join("\n\n").length >= MAX_GRADER_OUTPUT_CHARS) break;
+	}
+	const joined = parts.join("\n\n");
+	return joined.length > MAX_GRADER_OUTPUT_CHARS ? `${joined.slice(0, MAX_GRADER_OUTPUT_CHARS)}…` : joined;
+}
 var UpdateGoalTool = class {
 	agent;
+	grader;
 	name = "UpdateGoal";
 	description = "Update the current goal's lifecycle status. Use `complete` when the goal is achieved, `blocked` when you cannot proceed, `paused` to park it, or `active` to resume.";
 	parameters = toInputJsonSchema(UpdateGoalToolInputSchema);
-	constructor(agent) {
+	constructor(agent, grader) {
 		this.agent = agent;
+		this.grader = grader;
 	}
 	resolveExecution(args) {
 		const goal = this.agent.goal;
@@ -56297,17 +56321,7 @@ var UpdateGoalTool = class {
 					await goal.resumeGoal({}, "model");
 					return { output: "Goal resumed." };
 				}
-				if (args.status === "complete") {
-					const completed = await goal.markComplete({}, "model");
-					if (completed !== null) this.agent.context.appendSystemReminder(buildGoalCompletionSummaryPrompt(completed), {
-						kind: "system_trigger",
-						name: GOAL_COMPLETION_REMINDER_NAME
-					});
-					return {
-						output: "Goal marked complete.",
-						stopTurn: true
-					};
-				}
+				if (args.status === "complete") return this.handleComplete(goal);
 				if (args.status === "blocked") {
 					const blocked = await goal.markBlocked({}, "model");
 					if (blocked !== null) this.agent.context.appendSystemReminder(buildGoalBlockedReasonPrompt(blocked), {
@@ -56327,6 +56341,39 @@ var UpdateGoalTool = class {
 			}
 		};
 	}
+	async handleComplete(goal) {
+		const goalState = goal.getGoal().goal;
+		if (!goalState) return { output: "No active goal." };
+		const output = extractRecentOutput(this.agent.context.history);
+		await goal.pauseGoal({ reason: "verifying" }, "system");
+		let pass;
+		let reason;
+		try {
+			const result = await this.grader(goalState.objective, goalState.completionCriterion, output);
+			pass = result.pass;
+			reason = result.reason;
+		} catch {
+			pass = true;
+			reason = "Grader unavailable";
+		}
+		await goal.resumeGoal({}, "system");
+		if (pass) {
+			const completed = await goal.markComplete({}, "model");
+			if (completed !== null) this.agent.context.appendSystemReminder(buildGoalCompletionSummaryPrompt(completed), {
+				kind: "system_trigger",
+				name: GOAL_COMPLETION_REMINDER_NAME
+			});
+			return {
+				output: `Goal verified and marked complete.\n${reason}`,
+				stopTurn: true
+			};
+		}
+		this.agent.context.appendSystemReminder(buildGradingFeedbackPrompt(reason), {
+			kind: "system_trigger",
+			name: "goal_grading_feedback"
+		});
+		return { output: `Verification failed: ${reason}. Continue working.` };
+	}
 };
 //#endregion
 //#region ../../packages/agent-core/src/tools/builtin/goal/write-goal-note.ts
@@ -75991,6 +76038,8 @@ function buildGoalReminder(goal) {
 	lines.push("When you discover important facts, verify a hypothesis, or hit a dead end, call WriteGoalNote to record it. Future turns will read these notes automatically. Keep notes concise and actionable.");
 	lines.push("");
 	lines.push("Goal mode is iterative. Keep the self-audit brief each turn. Do not explore unrelated interpretations once the goal can be decided. If the objective is simple, already answered, impossible, unsafe, or contradictory, do not run another goal turn. Explain briefly if useful, then call UpdateGoal with `complete` or `blocked` in the same turn. Otherwise, self-audit against the objective and any completion criteria above, then do one coherent slice of work toward the objective. Use multiple turns when the task naturally has multiple phases. Call UpdateGoal with `complete` only when all required work is done, any stated validation has passed, and there is no useful next action. Do not mark complete after only producing a plan, summary, first pass, or partial result. If an external condition or required user input prevents progress, or the objective cannot be completed as stated, call UpdateGoal with `blocked`. Otherwise keep working — after your turn ends you will be prompted to continue. Call UpdateGoal as soon as the goal is genuinely done or cannot proceed; don't keep going once there is nothing left to do.");
+	lines.push("");
+	lines.push("When you call UpdateGoal with `complete`, an independent reviewer will verify that the completion criteria are met. Provide a clear summary of what you accomplished in your final response so the reviewer can evaluate it.");
 	return lines.join("\n");
 }
 function maxBudgetFraction(goal) {
@@ -76488,7 +76537,12 @@ const DEFAULT_APPROVE_TOOLS = new Set([
 	"Agent",
 	"AskUserQuestion",
 	"Skill",
-	"WolfPack"
+	"WolfPack",
+	"CreateGoal",
+	"UpdateGoal",
+	"GetGoal",
+	"SetGoalBudget",
+	"WriteGoalNote"
 ]);
 var DefaultToolApprovePermissionPolicy = class {
 	name = "default-tool-approve";
@@ -90872,6 +90926,151 @@ const DEFAULT_AGENT_PROFILES = loadAgentProfilesFromSources([
 ].map((file) => `profile/default/${file}`), PROFILE_SOURCES);
 //#endregion
 //#region ../../packages/agent-core/src/agent/tool/index.ts
+const CRITERIA_SYSTEM_PROMPT = [
+	"You generate concrete, verifiable acceptance criteria for a given objective.",
+	"Criteria must be specific and testable — state what should work end-to-end, not just what should exist.",
+	"Avoid vague criteria like \"feature works\" or \"code is correct\".",
+	"Respond with JSON only: {\"criteria\": [\"criterion 1\", \"criterion 2\", ...]}"
+].join(" ");
+const GRADER_SYSTEM_PROMPT = [
+	"You are a strict goal completion evaluator. Your default judgment is FAIL. Only PASS when there is clear, specific evidence that every acceptance criterion is genuinely met end-to-end.",
+	"Evaluate across three dimensions:",
+	"- Completeness: every acceptance criterion is individually met with concrete evidence in the output. Partial completion is FAIL.",
+	"- Conformance: the work matches what was asked — no scope drift, no over-engineering, no cutting corners.",
+	"- Substance: the output is real, finished, working work — not just a plan, outline, scaffold, stub, mock, or partial implementation, unless the objective specifically asks for those. Surface-level appearance without end-to-end correctness is FAIL.",
+	"When FAIL, you MUST list specific issues with actionable fix directions. Do not accept plausible-sounding but unverified claims of completion.",
+	"Respond with JSON only."
+].join(" ");
+function buildCriteriaPrompt(objective) {
+	return [
+		"## Objective",
+		objective,
+		"",
+		"Generate 3-8 concrete, verifiable acceptance criteria for this objective.",
+		"Each criterion should describe a specific, testable behavior or outcome — focus on end-to-end correctness, not surface existence.",
+		"Respond with JSON: {\"criteria\": [\"criterion 1\", \"criterion 2\", ...]}"
+	].join("\n");
+}
+function buildGraderPrompt(objective, criteria, output) {
+	return [
+		"## Objective",
+		objective,
+		"",
+		"## Acceptance Criteria",
+		criteria,
+		"",
+		"## Agent Output",
+		output || "(no output captured)",
+		"",
+		"Evaluate each dimension independently against the acceptance criteria, then decide overall PASS/FAIL.",
+		"When FAIL, list every specific issue with an actionable fix direction so the agent knows exactly what to address next.",
+		"Respond with JSON:",
+		"{\"completeness\":{\"pass\":true/false,\"detail\":\"...\"},\"conformance\":{\"pass\":true/false,\"detail\":\"...\"},\"substance\":{\"pass\":true/false,\"detail\":\"...\"},\"issues\":[\"issue 1: what to fix\",\"issue 2: what to fix\"],\"pass\":true/false,\"reason\":\"overall summary\"}"
+	].join("\n");
+}
+function parseGraderResponse(text) {
+	try {
+		const match = text.match(/\{[\s\S]*\}/);
+		if (!match) return {
+			pass: true,
+			reason: "No JSON found in grader response",
+			summary: ""
+		};
+		const parsed = JSON.parse(match[0]);
+		const overallPass = parsed.pass === true;
+		const overallReason = typeof parsed.reason === "string" ? parsed.reason : "No reason provided";
+		if (![
+			parsed.completeness,
+			parsed.conformance,
+			parsed.substance
+		].some((d) => d !== void 0)) return {
+			pass: overallPass,
+			reason: overallReason,
+			summary: ""
+		};
+		const lines = [];
+		const failedDims = [];
+		for (const [name, dim] of Object.entries({
+			Completeness: parsed.completeness,
+			Conformance: parsed.conformance,
+			Substance: parsed.substance
+		})) {
+			if (dim === void 0) continue;
+			const ok = dim.pass === true;
+			const detail = typeof dim.detail === "string" ? dim.detail : "";
+			lines.push(`  ${ok ? "✓" : "✗"} ${name}: ${detail}`);
+			if (!ok) failedDims.push(`${name}: ${detail}`);
+		}
+		const issues = Array.isArray(parsed.issues) ? parsed.issues.filter((i) => typeof i === "string") : [];
+		if (issues.length > 0) {
+			lines.push("");
+			lines.push("  Issues to fix:");
+			for (const issue of issues) lines.push(`  - ${issue}`);
+		}
+		const summary = lines.join("\n");
+		const reasonParts = [overallReason];
+		if (failedDims.length > 0) reasonParts.push(failedDims.join("\n"));
+		if (issues.length > 0) reasonParts.push(`Issues to fix:\n${issues.map((i) => `- ${i}`).join("\n")}`);
+		return {
+			pass: overallPass,
+			reason: reasonParts.join("\n"),
+			summary
+		};
+	} catch {
+		return {
+			pass: true,
+			reason: "Failed to parse grader response",
+			summary: ""
+		};
+	}
+}
+function extractResponseText(response) {
+	return response.message.content.filter((p) => p.type === "text").map((p) => p.text).join("");
+}
+async function generateAcceptanceCriteria(agent, objective) {
+	const prompt = buildCriteriaPrompt(objective);
+	const text = extractResponseText(await agent.rawGenerate(agent.config.provider, CRITERIA_SYSTEM_PROMPT, [], [{
+		role: "user",
+		content: [{
+			type: "text",
+			text: prompt
+		}],
+		toolCalls: []
+	}]));
+	try {
+		const match = text.match(/\{[\s\S]*\}/);
+		if (!match) return "";
+		const parsed = JSON.parse(match[0]);
+		if (!Array.isArray(parsed.criteria)) return "";
+		return parsed.criteria.filter((c) => typeof c === "string").map((c, i) => `${i + 1}. ${c}`).join("\n");
+	} catch {
+		return "";
+	}
+}
+function createGoalGrader(agent) {
+	return async (objective, criterion, output) => {
+		let criteria;
+		if (criterion !== void 0) criteria = criterion;
+		else {
+			criteria = await generateAcceptanceCriteria(agent, objective);
+			if (!criteria) criteria = "No specific criteria defined. Evaluate based on whether the objective is clearly achieved.";
+		}
+		const user = buildGraderPrompt(objective, criteria, output);
+		const result = parseGraderResponse(extractResponseText(await agent.rawGenerate(agent.config.provider, GRADER_SYSTEM_PROMPT, [], [{
+			role: "user",
+			content: [{
+				type: "text",
+				text: user
+			}],
+			toolCalls: []
+		}])));
+		const reason = result.summary ? `${result.reason}\n${result.summary}` : result.reason;
+		return {
+			pass: result.pass,
+			reason
+		};
+	};
+}
 var ToolManager = class {
 	agent;
 	builtinTools = /* @__PURE__ */ new Map();
@@ -91150,7 +91349,7 @@ var ToolManager = class {
 			this.agent.cron && new CronListTool(this.agent.cron),
 			this.agent.cron && new CronDeleteTool(this.agent.cron),
 			this.agent.type === "main" && new CreateGoalTool(this.agent),
-			this.agent.type === "main" && new UpdateGoalTool(this.agent),
+			this.agent.type === "main" && new UpdateGoalTool(this.agent, createGoalGrader(this.agent)),
 			this.agent.type === "main" && new GetGoalTool(this.agent),
 			this.agent.type === "main" && new SetGoalBudgetTool(this.agent),
 			this.agent.type === "main" && new WriteGoalNoteTool(this.agent),
@@ -121976,6 +122175,9 @@ var GoalStatusMessageComponent = class {
 };
 //#endregion
 //#region src/tui/commands/goal.ts
+const GOAL_STATUS_DISMISS_MS = 1e4;
+let activeGoalPanel;
+let activeGoalTimer;
 const CONTROL_SUBCOMMANDS = new Set([
 	"pause",
 	"resume",
@@ -122125,13 +122327,28 @@ async function showGoalStatus(host) {
 	}
 	try {
 		const result = await session.getGoal();
-		host.state.transcriptContainer.addChild(new GoalStatusMessageComponent(result.goal, host.state.theme.colors));
+		dismissGoalPanel(host);
+		const panel = new GoalStatusMessageComponent(result.goal, host.state.theme.colors);
+		host.state.transcriptContainer.addChild(panel);
+		activeGoalPanel = panel;
+		activeGoalTimer = setTimeout(() => dismissGoalPanel(host), GOAL_STATUS_DISMISS_MS);
 		host.state.ui.requestRender();
 	} catch (error) {
 		const message = error instanceof Error ? error.message : String(error);
 		host.showError(`获取目标状态失败：${message}`);
 	}
 }
+function dismissGoalPanel(host) {
+	if (activeGoalTimer !== void 0) {
+		clearTimeout(activeGoalTimer);
+		activeGoalTimer = void 0;
+	}
+	if (activeGoalPanel !== void 0) {
+		host.state.transcriptContainer.removeChild(activeGoalPanel);
+		activeGoalPanel = void 0;
+		host.state.ui.requestRender();
+	}
+}
 //#endregion
 //#region src/utils/git/git-status.ts
 /**
@@ -122687,10 +122904,7 @@ var FooterComponent = class {
 		if (state.permissionMode === "yolo") left.push(chalk.hex(colors.warning).bold("YES"));
 		if (state.planMode) left.push(chalk.hex(colors.primary).bold("plan"));
 		if (state.wolfpackMode) left.push(chalk.hex(colors.primary).bold("wolfpack"));
-		if (state.goalActive && state.goal) {
-			const goalText = state.goal.length > 20 ? state.goal.slice(0, 20) + "…" : state.goal;
-			left.push(chalk.hex(colors.success).bold(`🎯 ${goalText}`));
-		}
+		if (state.goalActive) left.push(chalk.hex(colors.primary).bold("goal"));
 		const model = shortenModel(modelDisplayName(state));
 		if (model) {
 			const thinkingLabel = state.thinking ? " 思考中" : "";
@@ -135933,6 +136147,27 @@ var ScreamTUI = class ScreamTUI {
 		const footerWrap = new GutterContainer(1, 1);
 		footerWrap.addChild(this.state.footer);
 		this.state.ui.addChild(footerWrap);
+		this.installFixedBottomRegion();
+	}
+	/**
+	* Pin the editor + footer to the terminal bottom. The pi-tui patch adds a
+	* `fixedBottomLineCount` property: the last N rendered lines stay pinned
+	* while the transcript above scrolls independently.
+	*
+	* We override `doRender` to measure the editor + footer height each frame
+	* (they change with multi-line input or terminal resize) and set the count
+	* before the real render runs.
+	*/
+	installFixedBottomRegion() {
+		const ui = this.state.ui;
+		const originalDoRender = ui.doRender.bind(ui);
+		const editorContainer = this.state.editorContainer;
+		const footerWrap = ui.children.at(-1);
+		ui.doRender = () => {
+			const w = ui.terminal?.columns ?? 0;
+			if (w > 0) ui.fixedBottomLineCount = editorContainer.render(w).length + (footerWrap ? footerWrap.render(w).length : 0);
+			originalDoRender();
+		};
 	}
 	handlePlanToggle(next) {
 		handlePlanCommand(this, next ? "on" : "off");
@@ -136598,7 +136833,7 @@ const SHADOW_CHARS = new Set([
 ]);
 const SHEEN_STEP = 2;
 const SHEEN_INTERVAL_MS = 150;
-const LOADING_DURATION_MS = 2200;
+const LOADING_DURATION_MS = 1500;
 const THEME_ACCENT = {
 	dark: [
 		78,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "scream-code",
-  "version": "0.5.1",
+  "version": "0.5.2",
   "description": "The Starting Point for Next-Gen Agents",
   "license": "MIT",
   "author": "ScreamCli",
@@ -40,7 +40,7 @@
   },
   "publishConfig": {
     "access": "public",
-    "provenance": false
+    "provenance": true
   },
   "scripts": {
     "build": "tsdown",