scream-code 0.12.6 → 0.12.8

package/dist/{app-NZLMIJND.mjs → app-CyJPMDaK.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);
@@ -76395,10 +76465,10 @@ function stripContextMetadata(message) {
 }
 //#endregion
 //#region ../../packages/agent-core/src/agent/compaction/compaction-instruction.md
-var compaction_instruction_default = "\n--- This message is a direct task, not part of the above conversation ---\n\nYou are now given a task to compact this conversation context according to specific priorities and output requirements.\n\nOutput text only. DO NOT CALL ANY TOOLS. Calling tools will be rejected and fails the task. You already have all the information you need in the conversation history. You have only one chance.\n\nThe goal of compaction is to keep essential code patterns, technical details, and architectural decisions for continuing development without losing context after the above messages are cleared work.\n\n{{ customInstruction }}\n\n<!-- Memory Memo Extraction (PRIORITY — do not skip) -->\n\n## 任务经验提取\n\nAFTER completing the compaction summary below, scan the messages being compacted for **completed task loops**. A task loop is \"completed\" when:\n- The user made a clear request or asked a specific question\n- You provided a solution or answer\n- The outcome is clear (success, partial success, or failure)\n\nFor each completed task loop found, output a structured experience record **at the very end of your response**:\n\n```memory-memo\n{\n  \"userNeed\": \"<the user's need or goal, one sentence>\",\n  \"approach\": \"<what was done — the approach taken, 2-4 sentences>\",\n  \"outcome\": \"<final result, e.g. '完成', '部分完成', '失败: reason'>\",\n  \"whatFailed\": \"<dead ends tried — things that didn't work, or 'none'>\",\n  \"whatWorked\": \"<key actions that ultimately worked, or 'none'>\",\n  \"tags\": [\"<tag1>\", \"<tag2>\", \"<tag3>\"],\n  \"note\": \"<optional: a soft, AI-readable note/suggestion that makes the record easier for future turns to understand — e.g. '这类任务先查接口可用性再写代码'. Omit if none>\"\n}\n```\n\nGuidelines:\n- Record important failed attempts in \"whatFailed\" to help avoid repeating mistakes.\n- Record key successful actions in \"whatWorked\" to help reuse effective approaches.\n- Include 3-5 semantic \"tags\" summarizing the task domain, tech stack, or action type (e.g. [\"react\", \"auth\", \"部署\"]).\n- \"note\" is optional: a one-sentence advisory note that helps future AI understand/reuse the record faster. It is a soft suggestion, not a user-enforced rule.\n- Skip in-progress work unless it contains a valuable error+fix experience.\n- Merge closely related sub-tasks into a single record.\n- Use the exact field names and JSON format shown above (no extra fields beyond \"note\").\n\nIf no completed task loops are found in the compacted messages, output:\n```memory-memo\n{\"none\": true}\n```\n\n<!-- Compression Priorities (in order) -->\n\n1. **Current Task State**: What is being worked on RIGHT NOW\n2. **Errors & Solutions**: All encountered errors and their resolutions\n3. **Code Evolution**: Final working versions only (remove intermediate attempts)\n4. **System Context**: Project structure, dependencies, environment setup\n5. **Design Decisions**: Architectural choices and their rationale\n6. **TODO Items**: Unfinished tasks and known issues\n\n<!-- Required Output Structure -->\n\n## Current Focus\n\n[What we're working on now]\n\n## Environment\n\n- [Key setup/config points]\n- ...\n\n## Completed Tasks\n\n- [Task]: [Brief outcome]\n- ...\n\n## Active Issues\n\n- [Issue]: [Status/Next steps]\n- ...\n\n## Code State\n\n### [Critical file name]\n\n[Brief description of the file's purpose and current state]\n\n```\n[The latest version of critical code snippets in this file, <20 lines]\n```\n\n### [Critical file name]\n\n- [Useful classes/methods/functions]: [Brief description/usage]\n- ...\n\n<!-- Omit non-critical code, intermediate attempts, and resolved errors -->\n\n## Important Context\n\n- [Any crucial information not covered above]\n- ...\n\n## All User Messages\n\n- [Detailed non tool use user message]\n- ...\n\n## Skill candidates\n\nScan the compressed messages for reusable processes (project-specific build\nsteps, recurring debugging patterns, or tool workflows). If any is genuinely\nworth capturing as a reusable skill, output at the very end, at most one:\n\n[[skill-candidate: <name>|<one-line purpose>|<evidence>]]\n\nOmit if none.\n";
+var compaction_instruction_default = "\n--- This message is a direct task, not part of the above conversation ---\n\nYou are now given a task to compact this conversation context according to specific priorities and output requirements.\n\nOutput text only. DO NOT CALL ANY TOOLS. Calling tools will be rejected and fails the task. You already have all the information you need in the conversation history. You have only one chance.\n\nThe goal of compaction is to keep essential code patterns, technical details, and architectural decisions for continuing development without losing context after the above messages are cleared work.\n\n{{ customInstruction }}\n\n<!-- Memory Memo Extraction (PRIORITY — do not skip) -->\n\n## 任务经验提取\n\nAFTER completing the compaction summary below, scan the messages being compacted for **task loops**. A task loop is \"completed\" when:\n- The user made a clear request or asked a specific question\n- You provided a solution or answer\n- The outcome is clear (success, partial success, or failure)\n\n**You MUST output at least one memory-memo block** (or the `{\"none\": true}` marker) at the very end of your response — omitting the section entirely is not allowed. Record completed task loops as full experience records; record **in-progress work** as a lower-priority record whose \"outcome\" is \"进行中\" (so the ongoing task survives compaction and can be resumed later).\n\nFor each task loop found, output a structured experience record **at the very end of your response**:\n\n```memory-memo\n{\n  \"userNeed\": \"<the user's need or goal, one sentence>\",\n  \"approach\": \"<what was done — the approach taken, 2-4 sentences>\",\n  \"outcome\": \"<final result, e.g. '完成', '部分完成', '失败: reason', or '进行中' for in-progress work>\",\n  \"whatFailed\": \"<dead ends tried — things that didn't work, or 'none'>\",\n  \"whatWorked\": \"<key actions that ultimately worked, or 'none'>\",\n  \"tags\": [\"<tag1>\", \"<tag2>\", \"<tag3>\"],\n  \"note\": \"<optional: a soft, AI-readable note/suggestion that makes the record easier for future turns to understand — e.g. '这类任务先查接口可用性再写代码'. Omit if none>\"\n}\n```\n\nGuidelines:\n- Record important failed attempts in \"whatFailed\" to help avoid repeating mistakes.\n- Record key successful actions in \"whatWorked\" to help reuse effective approaches.\n- Include 3-5 semantic \"tags\" summarizing the task domain, tech stack, or action type (e.g. [\"react\", \"auth\", \"部署\"]).\n- \"note\" is optional: a one-sentence advisory note that helps future AI understand/reuse the record faster. It is a soft suggestion, not a user-enforced rule.\n- For in-progress work: record it with \"outcome\": \"进行中\" and describe where the task stands and what remains — this is what lets the agent resume seamlessly after compaction.\n- Merge closely related sub-tasks into a single record.\n- Use the exact field names and JSON format shown above (no extra fields beyond \"note\").\n\nIf no task loops (completed or in-progress) are found in the compacted messages, output:\n```memory-memo\n{\"none\": true}\n```\n\n<!-- Compression Priorities (in order) -->\n\n1. **Current Task State**: What is being worked on RIGHT NOW\n2. **Errors & Solutions**: All encountered errors and their resolutions\n3. **Code Evolution**: Final working versions only (remove intermediate attempts)\n4. **System Context**: Project structure, dependencies, environment setup\n5. **Design Decisions**: Architectural choices and their rationale\n6. **Next Steps**: The ordered, actionable plan going forward\n7. **TODO Items**: Unfinished tasks and known issues\n\n<!-- Required Output Structure -->\n\n## Current Focus\n\n[What we're working on now]\n\n## Environment\n\n- [Key setup/config points]\n- ...\n\n## Completed Tasks\n\n- [Task]: [Brief outcome]\n- ...\n\n## Active Issues\n\n- [Issue]: [Status/Next steps]\n- ...\n\n## Key Decisions\n\n- [Decision]: [rationale] — preserve architectural choices and why they were made\n- ...\n\n## Next Steps\n\n1. [The very next actionable step]\n2. [Then this]\n3. ...\n\n## Code State\n\n### [Critical file name]\n\n[Brief description of the file's purpose and current state]\n\n```\n[The latest version of critical code snippets in this file, <20 lines]\n```\n\n### [Critical file name]\n\n- [Useful classes/methods/functions]: [Brief description/usage]\n- ...\n\n<!-- Omit non-critical code, intermediate attempts, and resolved errors -->\n\n## Important Context\n\n- [Any crucial information not covered above]\n- ...\n\n## All User Messages\n\n- [Detailed non tool use user message]\n- ...\n\n## Skill candidates\n\nScan the compressed messages for reusable processes (project-specific build\nsteps, recurring debugging patterns, or tool workflows). If any is genuinely\nworth capturing as a reusable skill, output at the very end, at most one:\n\n[[skill-candidate: <name>|<one-line purpose>|<evidence>]]\n\n- The evidence must be concrete and verifiable: exact file paths, commands, or\n  step sequences from the conversation (e.g. \"run `pnpm vitest run -t compaction`\n  in packages/agent-core to reproduce the compaction snapshots\"). Vague evidence\n  like \"used a script\" is not acceptable.\n- Only emit a candidate when the process is genuinely reusable beyond this one\n  task; do not emit for one-off actions.\n\nOmit if none.\n";
 //#endregion
 //#region ../../packages/agent-core/src/agent/compaction/compaction-update-instruction.md
-var compaction_update_instruction_default = "\n--- This message is a direct task, not part of the above conversation ---\n\nYou are now given a task to UPDATE an existing compaction summary with new messages that came in after the last compaction.\n\nA previous compaction summary already exists at the top of the conversation (the first assistant message). You must NOT discard it. Instead, merge the new information into the existing structure, keeping all previously captured state intact unless explicitly contradicted or resolved by the new messages.\n\nOutput text only. DO NOT CALL ANY TOOLS. Calling tools will be rejected and fails the task. You have only one chance.\n\n{{ customInstruction }}\n\n<!-- Previous Summary Reference -->\n\nThe first assistant message above is the PREVIOUS summary. Treat it as the source of truth for everything that happened before the new messages. Do not restate it verbatim — produce an updated, merged summary.\n\n<!-- Memory Memo Extraction (PRIORITY — do not skip) -->\n\n## 任务经验提取\n\nAFTER completing the updated summary below, scan ONLY the new messages being compacted (not those already covered by the previous summary) for **completed task loops**. A task loop is \"completed\" when:\n- The user made a clear request or asked a specific question\n- You provided a solution or answer\n- The outcome is clear (success, partial success, or failure)\n\nFor each completed task loop found, output a structured experience record **at the very end of your response**:\n\n```memory-memo\n{\n  \"userNeed\": \"<the user's need or goal, one sentence>\",\n  \"approach\": \"<what was done — the approach taken, 2-4 sentences>\",\n  \"outcome\": \"<final result, e.g. '完成', '部分完成', '失败: reason'>\",\n  \"whatFailed\": \"<dead ends tried — things that didn't work, or 'none'>\",\n  \"whatWorked\": \"<key actions that ultimately worked, or 'none'>\",\n  \"tags\": [\"<tag1>\", \"<tag2>\", \"<tag3>\"]\n}\n```\n\nGuidelines:\n- Record important failed attempts in \"whatFailed\" to help avoid repeating mistakes.\n- Record key successful actions in \"whatWorked\" to help reuse effective approaches.\n- Include 3-5 semantic \"tags\" summarizing the task domain, tech stack, or action type.\n- Skip in-progress work unless it contains a valuable error+fix experience.\n- Merge closely related sub-tasks into a single record.\n- Use the exact field names and JSON format shown above.\n\nIf no completed task loops are found in the new compacted messages, output:\n```memory-memo\n{\"none\": true}\n```\n\n<!-- Update Rules -->\n\n1. Preserve the section structure of the previous summary (Current Focus, Environment, Completed Tasks, Active Issues, Code State, Important Context, All User Messages).\n2. Move newly-completed tasks from \"Current Focus\" / \"Active Issues\" into \"Completed Tasks\".\n3. Update \"Current Focus\" to reflect what is being worked on RIGHT NOW.\n4. Append new user messages to \"All User Messages\" — do not repeat those already captured.\n5. Refresh \"Code State\" code snippets only if newer versions exist in the new messages.\n6. Drop resolved issues from \"Active Issues\"; add newly-discovered ones.\n7. Do not invent new information. If the new messages say nothing about a section, carry the previous content forward unchanged.\n\n<!-- Required Output Structure -->\n\n## Current Focus\n\n[What we're working on now]\n\n## Environment\n\n- [Key setup/config points]\n- ...\n\n## Completed Tasks\n\n- [Task]: [Brief outcome]\n- ...\n\n## Active Issues\n\n- [Issue]: [Status/Next steps]\n- ...\n\n## Code State\n\n### [Critical file name]\n\n[Brief description of the file's purpose and current state]\n\n```\n[The latest version of critical code snippets in this file, <20 lines]\n```\n\n## Important Context\n\n- [Any crucial information not covered above]\n- ...\n\n## All User Messages\n\n- [Detailed non tool use user message]\n- ...\n";
+var compaction_update_instruction_default = "\n--- This message is a direct task, not part of the above conversation ---\n\nYou are now given a task to UPDATE an existing compaction summary with new messages that came in after the last compaction.\n\nA previous compaction summary already exists at the top of the conversation (the first assistant message). You must NOT discard it. Instead, merge the new information into the existing structure, keeping all previously captured state intact unless explicitly contradicted or resolved by the new messages.\n\nOutput text only. DO NOT CALL ANY TOOLS. Calling tools will be rejected and fails the task. You have only one chance.\n\n{{ customInstruction }}\n\n<!-- Previous Summary Reference -->\n\nThe first assistant message above is the PREVIOUS summary. Treat it as the source of truth for everything that happened before the new messages. Do not restate it verbatim — produce an updated, merged summary.\n\n<!-- Memory Memo Extraction (PRIORITY — do not skip) -->\n\n## 任务经验提取\n\nAFTER completing the updated summary below, scan ONLY the new messages being compacted (not those already covered by the previous summary) for **task loops**. A task loop is \"completed\" when:\n- The user made a clear request or asked a specific question\n- You provided a solution or answer\n- The outcome is clear (success, partial success, or failure)\n\n**You MUST output at least one memory-memo block** (or the `{\"none\": true}` marker) at the very end of your response — omitting the section entirely is not allowed. Record completed task loops as full experience records; record **in-progress work** as a lower-priority record whose \"outcome\" is \"进行中\" (so the ongoing task survives compaction and can be resumed later).\n\nFor each task loop found, output a structured experience record **at the very end of your response**:\n\n```memory-memo\n{\n  \"userNeed\": \"<the user's need or goal, one sentence>\",\n  \"approach\": \"<what was done — the approach taken, 2-4 sentences>\",\n  \"outcome\": \"<final result, e.g. '完成', '部分完成', '失败: reason', or '进行中' for in-progress work>\",\n  \"whatFailed\": \"<dead ends tried — things that didn't work, or 'none'>\",\n  \"whatWorked\": \"<key actions that ultimately worked, or 'none'>\",\n  \"tags\": [\"<tag1>\", \"<tag2>\", \"<tag3>\"],\n  \"note\": \"<optional: a soft, AI-readable note/suggestion that makes the record easier for future turns to understand — e.g. '这类任务先查接口可用性再写代码'. Omit if none>\"\n}\n```\n\nGuidelines:\n- Record important failed attempts in \"whatFailed\" to help avoid repeating mistakes.\n- Record key successful actions in \"whatWorked\" to help reuse effective approaches.\n- Include 3-5 semantic \"tags\" summarizing the task domain, tech stack, or action type.\n- For in-progress work: record it with \"outcome\": \"进行中\" and describe where the task stands and what remains — this is what lets the agent resume seamlessly after compaction.\n- Merge closely related sub-tasks into a single record.\n- Use the exact field names and JSON format shown above.\n\nIf no task loops (completed or in-progress) are found in the new compacted messages, output:\n```memory-memo\n{\"none\": true}\n```\n\n<!-- Update Rules -->\n\n1. Preserve the section structure of the previous summary (Current Focus, Environment, Completed Tasks, Active Issues, Key Decisions, Next Steps, Code State, Important Context, All User Messages).\n2. Move newly-completed tasks from \"Current Focus\" / \"Active Issues\" into \"Completed Tasks\".\n3. Update \"Current Focus\" to reflect what is being worked on RIGHT NOW.\n4. Append new user messages to \"All User Messages\" — do not repeat those already captured.\n5. Refresh \"Code State\" code snippets only if newer versions exist in the new messages.\n6. Drop resolved issues from \"Active Issues\"; add newly-discovered ones.\n7. Merge new decisions into \"Key Decisions\" (keep prior decisions unless explicitly overturned); merge new steps into \"Next Steps\" (reorder if priorities changed).\n8. Do not invent new information. If the new messages say nothing about a section, carry the previous content forward unchanged.\n\n<!-- Required Output Structure -->\n\n## Current Focus\n\n[What we're working on now]\n\n## Environment\n\n- [Key setup/config points]\n- ...\n\n## Completed Tasks\n\n- [Task]: [Brief outcome]\n- ...\n\n## Active Issues\n\n- [Issue]: [Status/Next steps]\n- ...\n\n## Key Decisions\n\n- [Decision]: [rationale] — preserve architectural choices and why they were made\n- ...\n\n## Next Steps\n\n1. [The very next actionable step]\n2. [Then this]\n3. ...\n\n## Code State\n\n### [Critical file name]\n\n[Brief description of the file's purpose and current state]\n\n```\n[The latest version of critical code snippets in this file, <20 lines]\n```\n\n## Important Context\n\n- [Any crucial information not covered above]\n- ...\n\n## All User Messages\n\n- [Detailed non tool use user message]\n- ...\n";
 //#endregion
 //#region ../../packages/agent-core/src/agent/compaction/render-messages.ts
 const TOOL_RESULT_MAX_CHARS = 2e3;
@@ -76490,8 +76560,9 @@ const DEFAULT_COMPACTION_CONFIG = {
 	blockRatio: .85,
 	reservedContextSize: 5e4,
 	maxCompactionPerTurn: 3,
-	maxRecentMessages: 4,
-	maxRecentUserMessages: 2,
+	maxRecentMessages: 16,
+	maxRecentUserMessages: 8,
+	maxRecentTokens: 2e4,
 	maxRecentSizeRatio: .2,
 	minOverflowReductionRatio: .05,
 	turnGrowthMultiplier: 2.5
@@ -76533,7 +76604,7 @@ var DefaultCompactionStrategy = class {
 			if (m2.role === "user") recentUserMessages++;
 			recentSize += estimateTokensForMessage(m2);
 			if (canSplitAfter(messages, splitIndex)) bestN = splitIndex + 1;
-			if ((recentMessages >= this.config.maxRecentMessages || recentUserMessages >= this.config.maxRecentUserMessages || recentSize >= this.maxSize * this.config.maxRecentSizeRatio) && bestN !== void 0) break;
+			if ((recentMessages >= this.config.maxRecentMessages || recentUserMessages >= this.config.maxRecentUserMessages || recentSize >= this.config.maxRecentTokens || recentSize >= this.maxSize * this.config.maxRecentSizeRatio) && bestN !== void 0) break;
 		}
 		return bestN ?? 0;
 	}
@@ -76621,12 +76692,23 @@ 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 +76851,10 @@ 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 +77086,25 @@ 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 processedSummary = this.postProcessSummary(summary, fileOps, toolCallHistory, compactedCount);
 			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();
@@ -77017,8 +77113,9 @@ var FullCompaction = class {
 				result
 			});
 			this.agent.context.applyCompaction(result);
+			this.lastCompactionFiles = fileLists;
 			this.lowWaterMark = Math.floor(this.effectiveTokenCount * 1.1);
-			await this.extractAndStoreMemos(processedSummary);
+			await this.extractAndStoreMemos(processedSummary, messagesToCompactForOps);
 			this.triggerPostCompactHook(data, result);
 			this.detectSkillCandidates(processedSummary);
 			this.consecutiveCompactionFailures = 0;
@@ -77082,16 +77179,27 @@ var FullCompaction = class {
 			}
 		});
 	}
-	/** Extract memory memos from compaction summary and store them. */
-	async extractAndStoreMemos(summary) {
+	/** Extract memory memos from compaction summary and store them. When the
+	*  summary carries no memory-memo block, fall back to recording the most
+	*  recent real user message from the compacted history as a low-confidence
+	*  in-progress memo, so an ongoing task survives compaction even if the
+	*  model omitted the extraction section. */
+	async extractAndStoreMemos(summary, messagesToCompact) {
 		const memoStore = this.agent.memoStore;
 		if (!memoStore) {
 			this.agent.log.info("Memory memo store not available, skipping extraction");
 			return;
 		}
 		this.agent.log.info("Scanning compaction summary for memory memos", { summaryLen: summary.length });
-		const memos = parseMemoryMemos(summary);
+		let memos = parseMemoryMemos(summary);
 		this.agent.log.info("Memory memo parse result", { memoCount: memos.length });
+		if (memos.length === 0 && messagesToCompact !== void 0) {
+			const fallback = /```memory-memo[\s\S]*?"none"\s*:\s*true[\s\S]*?```/.test(summary) ? void 0 : this.buildFallbackMemo(messagesToCompact);
+			if (fallback !== void 0) {
+				memos = [fallback];
+				this.agent.log.info("Compaction summary carried no memory-memo; stored fallback memo", { userNeed: fallback.userNeed.slice(0, 120) });
+			}
+		}
 		if (memos.length === 0) return;
 		const sessionId = this.agent.homedir ? basename$1(dirname$2(dirname$2(this.agent.homedir))) : "unknown";
 		const sessionTitle = await this.agent.getSessionTitle();
@@ -77110,6 +77218,28 @@ var FullCompaction = class {
 			sessionId
 		});
 	}
+	/** Build a low-confidence fallback memo from the most recent real user
+	*  message in the compacted history (origin.kind === 'user'), so an ongoing
+	*  task is not lost when the summary omits the memory-memo section. */
+	buildFallbackMemo(messages) {
+		for (let i = messages.length - 1; i >= 0; i--) {
+			const msg = messages[i];
+			if (msg === void 0 || msg.role !== "user" || msg.origin?.kind !== "user") continue;
+			const text = msg.content.filter((p) => p.type === "text").map((p) => p.text).join("").trim();
+			if (text.length === 0) continue;
+			return createMemoryMemo({
+				sourceSessionId: "",
+				userNeed: text.slice(0, 300),
+				approach: "（压缩时自动兜底记录：压缩摘要未生成完整任务经验提取，此处记录最近一条用户请求，用于跨压缩恢复进行中任务。请结合会话上下文核实。）",
+				outcome: "进行中",
+				whatFailed: "none",
+				whatWorked: "none",
+				tags: ["compaction-fallback"],
+				note: "低置信度自动记录：建议后续在上下文恢复后核实并补充完整经验。",
+				extractionSource: "compaction"
+			});
+		}
+	}
 	/**
 	* Detects [[skill-candidate: <name>|<purpose>|<evidence>]] markers in the
 	* compaction summary and emits a `skill_candidate` event for each one (at
@@ -77142,10 +77272,18 @@ var FullCompaction = class {
 	* the compaction summary so active tasks and file context survive
 	* compression. Without this, both are lost after compaction because the
 	* original messages containing them are removed from the context window.
+	*
+	* When messages were actually compacted (`compactedCount > 0`), a leading
+	* elision note tells the model how much of the conversation is missing and
+	* that the summary is working notes, not verbatim history — so the tail
+	* messages after the summary are not mistaken for a continuous conversation.
 	*/
-	postProcessSummary(summary, fileOps, toolCallHistory) {
+	postProcessSummary(summary, fileOps, toolCallHistory, compactedCount) {
 		const todos = this.agent.tools.storeData()["todo"] ?? [];
-		const sections = [summary.trim()];
+		const base = summary.trim().replaceAll(/<files>[\s\S]*?<\/files>\s*/g, "").replace(/^> The conversation before this point was compacted:[\s\S]*?continue from here\.\s*\n+/m, "").trimEnd();
+		const sections = [];
+		if (compactedCount > 0) sections.push(`> The conversation before this point was compacted: ${String(compactedCount)} earlier message(s) were omitted and are covered by this summary. Treat the summary below as working notes, not verbatim history — the messages after it continue from here.`);
+		if (base.length > 0) sections.push(base);
 		if (todos.length > 0) {
 			const lines = todos.map((t) => {
 				return `- [${t.status === "done" ? "x" : t.status === "in_progress" ? "-" : " "}] ${t.title}`;
@@ -78804,12 +78942,15 @@ 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 +79052,31 @@ 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 +96499,11 @@ 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 +96738,11 @@ 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 +96822,11 @@ 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 +102749,7 @@ 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 +102759,7 @@ var SessionSubagentHost = class {
 					text: summary_continuation_default
 				}], origin);
 				await runChildTurnToCompletion(child, options.signal);
-				result = lastAssistantText(child);
+				result = lastAssistantText$1(child);
 			}
 			const usage = child.usage.data().total;
 			let findingsBlock = "";
@@ -102694,7 +102875,7 @@ async function runChildTurnToCompletion(child, signal) {
 function throwIfSubagentStoppedAtMaxTokens(stopReason) {
 	if (stopReason === "max_tokens") throw new Error(`${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 +120372,10 @@ 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 +120953,7 @@ 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 +121393,7 @@ 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 +121815,7 @@ function optionalBuildString(value) {
 	return typeof value === "string" && value.length > 0 ? value : void 0;
 }
 const SCREAM_BUILD_INFO = {
-	version: optionalBuildString("0.12.6"),
+	version: optionalBuildString("0.12.8"),
 	channel: optionalBuildString(""),
 	commit: optionalBuildString(""),
 	buildTarget: optionalBuildString("darwin-arm64")
@@ -122621,28 +122802,46 @@ const BUILTIN_SLASH_COMMANDS = [
 		name: "auto",
 		aliases: [],
 		description: "registry.auto_desc",
-		priority: 125,
+		priority: 220,
 		availability: "always"
 	},
 	{
 		name: "yes",
 		aliases: ["yolo"],
 		description: "registry.yolo_desc",
-		priority: 124,
+		priority: 219,
 		availability: "always"
 	},
 	{
 		name: "ask",
 		aliases: ["ask"],
 		description: "registry.ask_desc",
-		priority: 123,
+		priority: 218,
+		availability: "always"
+	},
+	{
+		name: "goal",
+		aliases: ["goaloff"],
+		description: "registry.goal_desc",
+		argumentHint: "[objective]",
+		priority: 217,
+		availability: (args) => {
+			const trimmed = args.trim();
+			return trimmed === "" || trimmed === "status" || trimmed === "pause" || trimmed === "off" ? "always" : "idle-only";
+		}
+	},
+	{
+		name: "wolfpack",
+		aliases: ["wp"],
+		description: "registry.wolfpack_desc",
+		priority: 216,
 		availability: "always"
 	},
 	{
 		name: "rlm",
 		aliases: ["rlm"],
 		description: "registry.rlm_desc",
-		priority: 121,
+		priority: 215,
 		availability: "always"
 	},
 	{
@@ -122650,39 +122849,21 @@ const BUILTIN_SLASH_COMMANDS = [
 		aliases: [],
 		description: "registry.rlm_max_depth_desc",
 		argumentHint: "[N]",
-		priority: 119,
-		availability: "always"
-	},
-	{
-		name: "wolfpack",
-		aliases: ["wp"],
-		description: "registry.wolfpack_desc",
-		priority: 122,
+		priority: 214,
 		availability: "always"
 	},
 	{
 		name: "sessions",
 		aliases: ["resume"],
 		description: "registry.sessions_desc",
-		priority: 119
-	},
-	{
-		name: "goal",
-		aliases: ["goaloff"],
-		description: "registry.goal_desc",
-		argumentHint: "[objective]",
-		priority: 120,
-		availability: (args) => {
-			const trimmed = args.trim();
-			return trimmed === "" || trimmed === "status" || trimmed === "pause" || trimmed === "off" ? "always" : "idle-only";
-		}
+		priority: 213
 	},
 	{
 		name: "memory",
 		aliases: ["memo", "mem"],
 		description: "registry.memory_desc",
 		argumentHint: "[query]",
-		priority: 120,
+		priority: 212,
 		availability: "always"
 	},
 	{
@@ -122690,96 +122871,61 @@ const BUILTIN_SLASH_COMMANDS = [
 		aliases: ["know"],
 		description: "registry.knowledge_desc",
 		argumentHint: "[query]",
-		priority: 119,
+		priority: 211,
 		availability: "always"
 	},
-	{
-		name: "new",
-		aliases: ["clear"],
-		description: "registry.new_desc",
-		priority: 120
-	},
 	{
 		name: "model",
 		aliases: [],
 		description: "registry.model_desc",
 		argumentHint: "[alias]",
-		priority: 120
+		priority: 210
+	},
+	{
+		name: "new",
+		aliases: ["clear"],
+		description: "registry.new_desc",
+		priority: 209
 	},
 	{
 		name: "compact",
 		aliases: [],
 		description: "registry.compact_desc",
-		priority: 119
+		priority: 208
 	},
 	{
-		name: "make-skill",
-		aliases: ["makeskill", "craftskill"],
-		description: "registry.make_skill_desc",
-		priority: 118,
-		availability: "idle-only"
+		name: "fusionplan",
+		aliases: ["fp"],
+		description: "registry.fusionplan_desc",
+		priority: 207,
+		availability: (args) => args.trim().toLowerCase() === "clear" ? "idle-only" : "always"
 	},
 	{
 		name: "plan",
 		aliases: [],
 		description: "registry.plan_desc",
-		priority: 118,
-		availability: (args) => args.trim().toLowerCase() === "clear" ? "idle-only" : "always"
-	},
-	{
-		name: "fusionplan",
-		aliases: ["fp"],
-		description: "registry.fusionplan_desc",
-		priority: 118,
+		priority: 206,
 		availability: (args) => args.trim().toLowerCase() === "clear" ? "idle-only" : "always"
 	},
 	{
 		name: "tasks",
 		aliases: ["task"],
 		description: "registry.tasks_desc",
-		priority: 117,
-		availability: "always"
-	},
-	{
-		name: "help",
-		aliases: ["h", "?"],
-		description: "registry.help_desc",
-		priority: 116,
-		availability: "always"
-	},
-	{
-		name: "status",
-		aliases: [],
-		description: "registry.status_desc",
-		priority: 115,
-		availability: "always"
-	},
-	{
-		name: "usage",
-		aliases: [],
-		description: "registry.usage_desc",
-		priority: 114,
+		priority: 205,
 		availability: "always"
 	},
 	{
 		name: "btw",
 		aliases: [],
 		description: "registry.btw_desc",
-		priority: 113,
+		priority: 204,
 		availability: "always"
 	},
 	{
 		name: "like",
 		aliases: [],
 		description: "registry.like_desc",
-		priority: 113,
-		availability: "always"
-	},
-	{
-		name: "mcp",
-		aliases: [],
-		description: "registry.mcp_desc",
-		priority: 112,
+		priority: 203,
 		availability: "always"
 	},
 	{
@@ -122790,121 +122936,162 @@ const BUILTIN_SLASH_COMMANDS = [
 			"plugins"
 		],
 		description: "registry.skill_desc",
-		priority: 110,
+		priority: 202,
 		availability: "always"
 	},
 	{
-		name: "cc",
+		name: "fork",
 		aliases: [],
-		description: "registry.cc_desc",
-		priority: 109,
+		description: "registry.fork_desc",
+		priority: 201
+	},
+	{
+		name: "title",
+		aliases: ["rename"],
+		description: "registry.title_desc",
+		priority: 200,
 		availability: "always"
 	},
 	{
-		name: "cc-connect",
+		name: "config",
 		aliases: [],
-		description: "registry.cc_connect_desc",
-		priority: 109,
+		description: "registry.config_desc",
+		priority: 199
+	},
+	{
+		name: "help",
+		aliases: ["h", "?"],
+		description: "registry.help_desc",
+		priority: 198,
 		availability: "always"
 	},
 	{
-		name: "revoke",
-		aliases: [],
-		description: "registry.revoke_desc",
-		priority: 108,
+		name: "make-skill",
+		aliases: ["makeskill", "craftskill"],
+		description: "registry.make_skill_desc",
+		priority: 197,
 		availability: "idle-only"
 	},
 	{
-		name: "fork",
+		name: "mcp",
 		aliases: [],
-		description: "registry.fork_desc",
-		priority: 105
+		description: "registry.mcp_desc",
+		priority: 196,
+		availability: "always"
 	},
 	{
-		name: "title",
-		aliases: ["rename"],
-		description: "registry.title_desc",
-		priority: 104,
+		name: "status",
+		aliases: [],
+		description: "registry.status_desc",
+		priority: 195,
 		availability: "always"
 	},
 	{
-		name: "config",
+		name: "usage",
 		aliases: [],
-		description: "registry.config_desc",
-		priority: 103
+		description: "registry.usage_desc",
+		priority: 194,
+		availability: "always"
 	},
 	{
-		name: "permission",
+		name: "revoke",
 		aliases: [],
-		description: "registry.permission_desc",
-		priority: 102,
+		description: "registry.revoke_desc",
+		priority: 193,
+		availability: "idle-only"
+	},
+	{
+		name: "cc",
+		aliases: [],
+		description: "registry.cc_desc",
+		priority: 192,
+		availability: "always"
+	},
+	{
+		name: "cc-connect",
+		aliases: [],
+		description: "registry.cc_connect_desc",
+		priority: 191,
 		availability: "always"
 	},
 	{
 		name: "theme",
 		aliases: [],
 		description: "registry.theme_desc",
-		priority: 101,
+		priority: 190,
 		availability: "always"
 	},
 	{
 		name: "language",
 		aliases: ["lang"],
 		description: "registry.language_desc",
-		priority: 102,
+		priority: 189,
+		availability: "always"
+	},
+	{
+		name: "permission",
+		aliases: [],
+		description: "registry.permission_desc",
+		priority: 188,
 		availability: "always"
 	},
 	{
 		name: "editor",
 		aliases: [],
 		description: "registry.editor_desc",
-		priority: 100,
+		priority: 187,
 		availability: "always"
 	},
 	{
 		name: "settings",
 		aliases: [],
 		description: "registry.settings_desc",
-		priority: 99,
+		priority: 186,
 		availability: "always"
 	},
 	{
 		name: "init",
 		aliases: [],
 		description: "registry.init_desc",
-		priority: 98
+		priority: 185
 	},
 	{
 		name: "export-md",
 		aliases: ["export"],
 		description: "registry.export_md_desc",
-		priority: 97
+		priority: 184
 	},
 	{
 		name: "export-debug-zip",
 		aliases: [],
 		description: "registry.export_debug_desc",
-		priority: 96
+		priority: 183
+	},
+	{
+		name: "eval",
+		aliases: [],
+		description: "registry.eval_desc",
+		priority: 182
 	},
 	{
 		name: "update",
 		aliases: [],
 		description: "registry.update_desc",
-		priority: 95,
+		priority: 181,
 		availability: "idle-only"
 	},
 	{
 		name: "version",
 		aliases: [],
 		description: "registry.version_desc",
-		priority: 94,
+		priority: 180,
 		availability: "always"
 	},
 	{
 		name: "logout",
 		aliases: ["disconnect"],
 		description: "registry.logout_desc",
-		priority: 93
+		priority: 179
 	},
 	{
 		name: "exit",
@@ -123022,7 +123209,7 @@ 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;
@@ -125546,7 +125733,15 @@ const HEADING_HASH_PREFIX = /^((?:\u001B\[[0-9;]*m)*)#{1,6}[ \t]+/;
 * takes one formatter function per token; tokens not listed here fall back to
 * its DEFAULT_THEME.
 */
-function createCodeHighlightTheme(colors) {
+/**
+* Markdown code-block highlight theme: green-dominant mapping (keyword,
+* function, built_in → primary; strings → success; numbers → warning;
+* comments → textDim). Kept distinct from the shared preview theme in
+* code-highlight-theme.ts on purpose — markdown code blocks use the green
+* primary hue, while file-preview panels use the classic blue/red/yellow
+* scheme mapped to the same palette. Both follow the active theme.
+*/
+function createMarkdownCodeHighlightTheme(colors) {
 	const keyword = chalk.hex(colors.primary);
 	const str = chalk.hex(colors.success);
 	const comment = chalk.hex(colors.textDim);
@@ -125596,7 +125791,7 @@ function createMarkdownTheme(colors) {
 	const stripHash = (text) => text.replace(HEADING_HASH_PREFIX, "$1");
 	const muted = chalk.hex(colors.textMuted);
 	const border = chalk.hex(colors.border);
-	const codeTheme = createCodeHighlightTheme(colors);
+	const codeTheme = createMarkdownCodeHighlightTheme(colors);
 	return {
 		heading: (text) => chalk.bold.hex(colors.text)(stripHash(text)),
 		link: (text) => chalk.hex(colors.mdLink)(text),
@@ -127448,7 +127643,7 @@ 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 +127664,7 @@ 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"),
@@ -128834,6 +129029,60 @@ var CachedContainer = class extends Container {
 	}
 };
 //#endregion
+//#region src/tui/theme/code-highlight-theme.ts
+function createCodeHighlightTheme(colors) {
+	const keyword = chalk.hex(colors.mdCodeBlock);
+	const str = chalk.hex(colors.error);
+	const num = chalk.hex(colors.primary);
+	const fn = chalk.hex(colors.warning);
+	const builtin = chalk.hex(colors.planMode);
+	const cls = chalk.hex(colors.accent);
+	const text = chalk.hex(colors.text);
+	return {
+		keyword,
+		built_in: builtin,
+		type: keyword,
+		literal: builtin,
+		number: num,
+		regexp: str,
+		string: str,
+		subst: str,
+		symbol: num,
+		class: cls,
+		function: fn,
+		title: fn,
+		params: text,
+		comment: num,
+		doctag: num,
+		meta: chalk.hex(colors.textMuted),
+		"meta-keyword": keyword,
+		"meta-string": str,
+		section: keyword,
+		tag: keyword,
+		name: fn,
+		"builtin-name": builtin,
+		attr: fn,
+		attribute: fn,
+		variable: text,
+		bullet: num,
+		code: str,
+		emphasis: (s) => chalk.italic(s),
+		strong: (s) => chalk.bold(s),
+		formula: text,
+		link: chalk.hex(colors.mdLink),
+		quote: chalk.hex(colors.mdQuote),
+		addition: chalk.hex(colors.diffAdded),
+		deletion: chalk.hex(colors.diffRemoved),
+		"selector-tag": keyword,
+		"selector-id": cls,
+		"selector-class": cls,
+		"selector-attr": fn,
+		"selector-pseudo": fn,
+		"template-tag": str,
+		"template-variable": fn
+	};
+}
+//#endregion
 //#region src/tui/components/media/code-highlight.ts
 /**
 * Shared syntax-highlighting helpers for code previews
@@ -128871,13 +129120,14 @@ function langFromPath(filePath) {
 	const lang = EXT_LANG_MAP[ext] ?? ext;
 	return supportsLanguage(lang) ? lang : void 0;
 }
-function highlightLines(code, lang) {
+function highlightLines(code, lang, colors) {
 	const normalizedLang = lang?.trim().toLowerCase();
 	if (!normalizedLang || !supportsLanguage(normalizedLang)) return code.split("\n");
 	try {
 		return highlight(code, {
 			language: normalizedLang,
-			ignoreIllegals: true
+			ignoreIllegals: true,
+			theme: createCodeHighlightTheme(colors)
 		}).split("\n");
 	} catch {
 		return code.split("\n");
@@ -128939,10 +129189,12 @@ function visualizeIndent(line) {
 }
 /**
 * Render a diff line's code with leading-whitespace visualization (tabs as
-* `->`, spaces as `·`, dimmed) and optional syntax highlighting. Highlighting
-* runs on the raw code first; indent visualization then runs on the
-* highlighted string - leading whitespace carries no token color, so it is
-* still plain and can be replaced and dimmed without disturbing the tokens.
+* `->`, spaces as `·`, dimmed) and optional syntax highlighting. Indent
+* visualization runs on the raw code FIRST, then highlighting runs on the
+* stripped remainder: once the text is ANSI-tinted (the theme's `default`
+* color wraps whitespace too), a leading tab is no longer the first raw
+* character and would be missed. The dimmed indent itself carries no token
+* color.
 *
 * When highlighting is on and produced token colors, the diff line color is
 * layered as the base foreground: syntax tokens override it, but every ANSI
@@ -128951,17 +129203,18 @@ function visualizeIndent(line) {
 * When highlighting is off (streaming) or produced no token colors, the code
 * part is colored with the diff line color instead.
 */
-function renderDiffCode(code, colorFn, highlight, lang) {
-	const { text, indentEnd } = visualizeIndent(highlight ? highlightLines(code, lang)[0] ?? code : code);
+function renderDiffCode(code, colorFn, highlight, lang, colors) {
+	const { text, indentEnd } = visualizeIndent(code);
 	const indent = text.slice(0, indentEnd);
 	const rest = text.slice(indentEnd);
 	const dimIndent = indent.length > 0 ? chalk.dim(indent) : indent;
-	if (highlight && ANSI_RE.test(rest)) {
+	const source = highlight ? highlightLines(rest, lang, colors)[0] ?? rest : rest;
+	if (highlight && ANSI_RE.test(source)) {
 		const prefix = extractAnsiPrefix(colorFn);
-		if (prefix.length > 0) return dimIndent + prefix + rest.replace(ANSI_RESET_RE, (m) => m + prefix) + "\x1B[39m";
-		return dimIndent + rest;
+		if (prefix.length > 0) return dimIndent + prefix + source.replace(ANSI_RESET_RE, (m) => m + prefix) + "\x1B[39m";
+		return dimIndent + source;
 	}
-	return dimIndent + colorFn(rest);
+	return dimIndent + colorFn(source);
 }
 /**
 * Compute word-level diff between two single lines and highlight the changed
@@ -129095,7 +129348,7 @@ function renderDiffLines(oldText, newText, path, colors, isIncomplete = false, o
 		const line = shown[i];
 		const marker = line.kind === "add" ? "+" : "-";
 		const color = line.kind === "add" ? s.add : s.del;
-		output.push(s.gutter(String(line.lineNum).padStart(4) + " ") + color(`${marker} `) + renderDiffCode(line.code, color, doHighlight, lang));
+		output.push(s.gutter(String(line.lineNum).padStart(4) + " ") + color(`${marker} `) + renderDiffCode(line.code, color, doHighlight, lang, colors));
 		i += 1;
 	}
 	const hidden = changedLines.length - shown.length;
@@ -129146,11 +129399,11 @@ function buildClusters(diffLines, contextLines) {
 		removedCount: removed
 	};
 }
-function formatDiffRow(line, s, doHighlight, lang) {
+function formatDiffRow(line, s, doHighlight, lang, colors) {
 	const gutter = s.gutter(String(line.lineNum).padStart(4) + " ");
-	if (line.kind === "add") return gutter + s.add("+ ") + renderDiffCode(line.code, s.add, doHighlight, lang);
-	if (line.kind === "delete") return gutter + s.del("- ") + renderDiffCode(line.code, s.del, doHighlight, lang);
-	return gutter + "  " + renderDiffCode(line.code, (x) => x, doHighlight, lang);
+	if (line.kind === "add") return gutter + s.add("+ ") + renderDiffCode(line.code, s.add, doHighlight, lang, colors);
+	if (line.kind === "delete") return gutter + s.del("- ") + renderDiffCode(line.code, s.del, doHighlight, lang, colors);
+	return gutter + "  " + renderDiffCode(line.code, (x) => x, doHighlight, lang, colors);
 }
 /**
 * Render a diff with surrounding context, eliding unchanged middle
@@ -129221,7 +129474,7 @@ function renderDiffLinesClustered(oldText, newText, path, colors, opts = {}) {
 				i += 2;
 				continue;
 			}
-			output.push(formatDiffRow(line, s, doHighlight, lang));
+			output.push(formatDiffRow(line, s, doHighlight, lang, colors));
 			body++;
 			if (line.kind !== "context") shownChanges++;
 			prevEnd = i;
@@ -131104,7 +131357,7 @@ var ToolCallComponent = class ToolCallComponent extends CachedContainer {
 		if (name === "Write") {
 			const content = str(this.toolCall.args["content"]);
 			if (content.length === 0) return;
-			const allLines = highlightLines(content, langFromPath(str(this.toolCall.args["file_path"] ?? this.toolCall.args["path"])));
+			const allLines = highlightLines(content, langFromPath(str(this.toolCall.args["file_path"] ?? this.toolCall.args["path"])), this.colors);
 			const writeShouldCap = !this.expanded;
 			const shown = writeShouldCap ? allLines.slice(0, 10) : allLines;
 			const remaining = allLines.length - shown.length;
@@ -131209,7 +131462,7 @@ var ToolCallComponent = class ToolCallComponent extends CachedContainer {
 		const fragmentEnd = Math.min(streamText.length, fragmentStart + ToolCallComponent.WRITE_STREAM_TAIL_RAW_CHARS);
 		const fragment = unescapeJsonStringValue(streamText, fragmentStart, fragmentEnd);
 		if (fragment.length === 0) return;
-		const lines = highlightLines(fragment, this.writeStreamLang);
+		const lines = highlightLines(fragment, this.writeStreamLang, this.colors);
 		const displayLines = lines.slice(-10);
 		const firstFragmentLineNo = this.writeStreamNlCount + (tailStart > contentStart ? 2 : 1);
 		const skipped = lines.length - displayLines.length;
@@ -136032,6 +136285,429 @@ 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 +136881,9 @@ async function handleBuiltInSlashCommand(host, name, args) {
 		case "logout":
 			await handleLogoutCommand(host);
 			return;
+		case "eval":
+			runEvalCommand(host);
+			return;
 		case "memory":
 			await handleMemoryCommand(host, args);
 			return;
@@ -137591,6 +138270,8 @@ var SessionEventHandler = class {
 		this.renderedSkillActivationIds.clear();
 		this.renderedMcpServerStatusKeys.clear();
 		this.stopAllMcpServerStatusSpinners();
+		this.pendingSkillCandidates = [];
+		this.promptedSkillCandidates.clear();
 		this.host.setAppState({ subagentUsage: {} });
 	}
 	startSubscription() {
@@ -137846,6 +138527,12 @@ var SessionEventHandler = class {
 		const todos = this.host.state.todoPanel.getTodos();
 		if (todos.length > 0 && todos.every((todo) => todo.status === "done")) this.host.streamingUI.setTodoList([]);
 		this.host.streamingUI.resetToolUi();
+		if (this.pendingSkillCandidates.length > 0 && this.host.state.queuedMessages.length === 0) {
+			const pending = this.pendingSkillCandidates;
+			this.pendingSkillCandidates = [];
+			const session = this.host.session;
+			if (session !== void 0) for (const candidate of pending) this.promptSkillCandidate(session, candidate.name, candidate.purpose);
+		}
 		this.host.streamingUI.finalizeTurn(sendQueued);
 	}
 	handleStepBegin(event) {
@@ -138220,6 +138907,10 @@ var SessionEventHandler = class {
 	}
 	/** Candidates already prompted this session, to avoid repeating the request. */
 	promptedSkillCandidates = /* @__PURE__ */ new Set();
+	/** Skill candidates detected while a turn was still active. Flushed at the
+	*  next turn end so the AskUserQuestion dialog never interrupts an ongoing
+	*  response. */
+	pendingSkillCandidates = [];
 	/**
 	* Handles a `skill_candidate` event emitted after compaction detected a
 	* reusable process in the summary. Best-effort and non-fatal: if the session
@@ -138227,6 +138918,10 @@ var SessionEventHandler = class {
 	* no confirmation. When it works, the model runs an AskUserQuestion dialog
 	* (确定/忽略) so the user can confirm or dismiss with a single keystroke,
 	* then generates the skill via MakeSkillPlanTool on confirmation.
+	*
+	* If a turn is still streaming (the compaction finished while the agent was
+	* mid-response), the candidate is queued and prompted at the next turn end —
+	* never interrupting the running turn.
 	*/
 	handleSkillCandidate(event) {
 		try {
@@ -138236,10 +138931,25 @@ var SessionEventHandler = class {
 			this.promptedSkillCandidates.add(name);
 			const session = this.host.session;
 			if (session === void 0) return;
-			const request = `检测到可复用过程「${name}」${purpose ? `（${purpose}）` : ""}。请用 AskUserQuestion 工具询问用户是否将其保存为可复用技能（选项：生成 / 忽略）。用户选择「生成」后，用 MakeSkillPlanTool 生成该技能；选择「忽略」则直接说明已跳过。`;
-			session.prompt(request).catch(() => {});
+			if (this.host.streamingUI.hasActiveTurn()) {
+				this.pendingSkillCandidates.push({
+					name,
+					purpose
+				});
+				return;
+			}
+			this.promptSkillCandidate(session, name, purpose);
 		} catch {}
 	}
+	/** Fire the AskUserQuestion prompt for a skill candidate. On failure, surface
+	*  a status hint instead of failing silently so the user still learns the
+	*  candidate was detected. */
+	promptSkillCandidate(session, name, purpose) {
+		const request = `检测到可复用过程「${name}」${purpose ? `（${purpose}）` : ""}。请用 AskUserQuestion 工具询问用户是否将其保存为可复用技能（选项：生成 / 忽略）。用户选择「生成」后，用 MakeSkillPlanTool 生成该技能；选择「忽略」则直接说明已跳过。`;
+		session.prompt(request).catch(() => {
+			this.host.showStatus(`检测到可复用过程「${name}」，但询问失败。可稍后手动处理。`, this.host.state.theme.colors.warning);
+		});
+	}
 	finishCompaction(sendQueued) {
 		if (!this.host.streamingUI.hasActiveTurn()) {
 			this.host.setAppState({
@@ -144537,7 +145247,7 @@ 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();
@@ -144590,7 +145300,7 @@ function renderDisplayBlock(block, s, colors) {
 		});
 		case "file_content": {
 			const lang = block.language ?? langFromPath(block.path);
-			const allLines = highlightLines(replaceTabs(block.content), lang);
+			const allLines = highlightLines(replaceTabs(block.content), lang, colors);
 			const shown = allLines.slice(0, CONTENT_SUMMARY_MAX_LINES);
 			const lines = [s.strong(block.path)];
 			for (const [i, line] of shown.entries()) lines.push(s.gutter(String(i + 1).padStart(4) + "  ") + line);
@@ -145001,7 +145711,7 @@ function buildDiffBody(block, colors) {
 }
 function buildFileContentBody(block, colors) {
 	const lang = block.language ?? langFromPath(block.path);
-	const highlighted = highlightLines(block.content, lang);
+	const highlighted = highlightLines(block.content, lang, colors);
 	const gutter = chalk.hex(colors.diffGutter);
 	return {
 		lines: highlighted.map((line, i) => gutter(String(i + 1).padStart(4) + "  ") + line),
@@ -147194,6 +147904,7 @@ 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);