@byfriends/cli 0.1.2 → 0.2.1

package/dist/main.mjs

@@ -8041,6 +8041,9 @@ Anthropic.Completions = Completions$2;
 Anthropic.Messages = Messages$2;
 Anthropic.Models = Models$2;
 Anthropic.Beta = Beta$1;
+//#endregion
+//#region ../../packages/kosong/src/capability.ts
+const UNKNOWN_CAPABILITY_MARKER = Symbol.for("byf.kosong.UNKNOWN_CAPABILITY");
 /**
 * Shared read-only default returned when a provider has not catalogued a
 * given model. Frozen so accidental mutation at one call site cannot leak
@@ -8056,9 +8059,32 @@ const UNKNOWN_CAPABILITY = Object.freeze(Object.defineProperty({
 	thinking_xhigh: false,
 	thinking_max: false,
 	max_context_tokens: 0
-}, Symbol.for("byf.kosong.UNKNOWN_CAPABILITY"), { value: true }));
+}, UNKNOWN_CAPABILITY_MARKER, { value: true }));
+function isUnknownCapability(capability) {
+	if (capability === UNKNOWN_CAPABILITY) return true;
+	if (capability[UNKNOWN_CAPABILITY_MARKER] === true) return true;
+	return !capability.image_in && !capability.video_in && !capability.audio_in && !capability.thinking && !capability.tool_use && !capability.thinking_effort && !capability.thinking_xhigh && !capability.thinking_max && capability.max_context_tokens === 0;
+}
 //#endregion
 //#region ../../packages/kosong/src/providers/capability-registry.ts
+const CACHE_CAPABILITY = Object.freeze({
+	strategy: "explicit-block",
+	maxCacheableBlocks: 4,
+	supportedScopes: [
+		"global",
+		"project",
+		"session",
+		"none"
+	]
+});
+const OPENAI_CACHE_CAPABILITY = Object.freeze({
+	strategy: "prompt-cache-key",
+	supportedScopes: ["global"]
+});
+const NO_CACHE_CAPABILITY = Object.freeze({
+	strategy: "none",
+	supportedScopes: []
+});
 const OPENAI_RESPONSES_DEVELOPER_ROLE_MODELS = new Set([
 	"gpt-4.1",
 	"gpt-4.1-mini",
@@ -8105,7 +8131,8 @@ const OPENAI_REASONING_CAPABILITY = Object.freeze({
 	thinking_effort: true,
 	thinking_xhigh: false,
 	thinking_max: false,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: OPENAI_CACHE_CAPABILITY
 });
 const OPENAI_REASONING_XHIGH_CAPABILITY = Object.freeze({
 	image_in: false,
@@ -8116,7 +8143,8 @@ const OPENAI_REASONING_XHIGH_CAPABILITY = Object.freeze({
 	thinking_effort: true,
 	thinking_xhigh: true,
 	thinking_max: false,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: OPENAI_CACHE_CAPABILITY
 });
 const OPENAI_VISION_TOOL_CAPABILITY = Object.freeze({
 	image_in: true,
@@ -8127,7 +8155,8 @@ const OPENAI_VISION_TOOL_CAPABILITY = Object.freeze({
 	thinking_effort: false,
 	thinking_xhigh: false,
 	thinking_max: false,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: OPENAI_CACHE_CAPABILITY
 });
 const OPENAI_TEXT_TOOL_CAPABILITY = Object.freeze({
 	image_in: false,
@@ -8138,7 +8167,8 @@ const OPENAI_TEXT_TOOL_CAPABILITY = Object.freeze({
 	thinking_effort: false,
 	thinking_xhigh: false,
 	thinking_max: false,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: OPENAI_CACHE_CAPABILITY
 });
 const ANTHROPIC_VISION_TOOL_CAPABILITY = Object.freeze({
 	image_in: true,
@@ -8149,7 +8179,8 @@ const ANTHROPIC_VISION_TOOL_CAPABILITY = Object.freeze({
 	thinking_effort: false,
 	thinking_xhigh: false,
 	thinking_max: false,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: CACHE_CAPABILITY
 });
 const ANTHROPIC_THINKING_VISION_TOOL_CAPABILITY = Object.freeze({
 	image_in: true,
@@ -8160,7 +8191,8 @@ const ANTHROPIC_THINKING_VISION_TOOL_CAPABILITY = Object.freeze({
 	thinking_effort: true,
 	thinking_xhigh: false,
 	thinking_max: true,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: CACHE_CAPABILITY
 });
 const ANTHROPIC_THINKING_XHIGH_VISION_TOOL_CAPABILITY = Object.freeze({
 	image_in: true,
@@ -8171,7 +8203,8 @@ const ANTHROPIC_THINKING_XHIGH_VISION_TOOL_CAPABILITY = Object.freeze({
 	thinking_effort: true,
 	thinking_xhigh: true,
 	thinking_max: true,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: CACHE_CAPABILITY
 });
 const GEMINI_MULTIMODAL_TOOL_CAPABILITY = Object.freeze({
 	image_in: true,
@@ -8182,7 +8215,8 @@ const GEMINI_MULTIMODAL_TOOL_CAPABILITY = Object.freeze({
 	thinking_effort: false,
 	thinking_xhigh: false,
 	thinking_max: false,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: NO_CACHE_CAPABILITY
 });
 const GEMINI_THINKING_MULTIMODAL_TOOL_CAPABILITY = Object.freeze({
 	image_in: true,
@@ -8193,7 +8227,8 @@ const GEMINI_THINKING_MULTIMODAL_TOOL_CAPABILITY = Object.freeze({
 	thinking_effort: false,
 	thinking_xhigh: false,
 	thinking_max: false,
-	max_context_tokens: 0
+	max_context_tokens: 0,
+	cache: NO_CACHE_CAPABILITY
 });
 const OPENAI_LEGACY_CAPABILITY_CATALOG = [
 	{
@@ -8262,7 +8297,12 @@ function capabilityFromCatalog(modelName, catalog) {
 	return UNKNOWN_CAPABILITY;
 }
 function getOpenAILegacyModelCapability(modelName) {
-	return capabilityFromCatalog(modelName, OPENAI_LEGACY_CAPABILITY_CATALOG);
+	const capability = capabilityFromCatalog(modelName, OPENAI_LEGACY_CAPABILITY_CATALOG);
+	if (!capability.cache) return {
+		...capability,
+		cache: OPENAI_CACHE_CAPABILITY
+	};
+	return capability;
 }
 function getOpenAIResponsesModelCapability(modelName) {
 	return capabilityFromCatalog(modelName, OPENAI_RESPONSES_CAPABILITY_CATALOG);
@@ -8277,6 +8317,23 @@ function getGoogleGenAIModelCapability(modelName) {
 	if (normalized.startsWith("gemini-2.5-") || normalized.includes("thinking")) return GEMINI_THINKING_MULTIMODAL_TOOL_CAPABILITY;
 	return GEMINI_MULTIMODAL_TOOL_CAPABILITY;
 }
+/**
+* Tries all provider-specific capability registries and returns the first
+* non-UNKNOWN match. Used when the provider context is unknown (e.g., catalog
+* enrichment) to fill in accurate thinking capability flags.
+*/
+function resolveCapabilityFromRegistry(modelName) {
+	const registries = [
+		getAnthropicModelCapability,
+		getOpenAIResponsesModelCapability,
+		getOpenAILegacyModelCapability,
+		getGoogleGenAIModelCapability
+	];
+	for (const fn of registries) {
+		const cap = fn(modelName);
+		if (!isUnknownCapability(cap)) return cap;
+	}
+}
 function usesOpenAIResponsesDeveloperRole(modelName) {
 	const normalized = normalizeModelName(modelName);
 	if (OPENAI_RESPONSES_DEVELOPER_ROLE_MODELS.has(normalized)) return true;
@@ -8511,31 +8568,34 @@ function clampEffort(effort, model) {
 	return effort;
 }
 const CACHE_CONTROL = { type: "ephemeral" };
-function splitSystemPrompt(systemPrompt, breakpoints) {
-	if (!breakpoints || breakpoints.length === 0) return systemPrompt ? [{
-		type: "text",
-		text: systemPrompt,
-		cache_control: CACHE_CONTROL
-	}] : [];
+/**
+* Content block types that support cache_control injection.
+*/
+const CACHEABLE_TYPES = new Set([
+	"text",
+	"image",
+	"document",
+	"search_result",
+	"tool_use",
+	"tool_result",
+	"server_tool_use",
+	"web_search_tool_result"
+]);
+/**
+* Convert a PromptPlan to Anthropic TextBlockParam[] with cache control.
+*
+* Injects cache_control on blocks with cacheScope other than 'none'.
+*/
+function promptPlanToSystemBlocks(promptPlan) {
 	const blocks = [];
-	let remaining = systemPrompt;
-	for (const marker of breakpoints) {
-		const idx = remaining.indexOf(marker);
-		if (idx === -1) continue;
-		const before = remaining.slice(0, idx).trim();
-		if (before) blocks.push({
+	for (const block of promptPlan.blocks) {
+		const textBlock = {
 			type: "text",
-			text: before,
-			cache_control: CACHE_CONTROL
-		});
-		remaining = remaining.slice(idx + marker.length);
+			text: block.text
+		};
+		if (block.cacheScope !== "none") textBlock.cache_control = CACHE_CONTROL;
+		blocks.push(textBlock);
 	}
-	const final = remaining.trim();
-	if (final) blocks.push({
-		type: "text",
-		text: final,
-		cache_control: CACHE_CONTROL
-	});
 	return blocks;
 }
 /**
@@ -8617,9 +8677,11 @@ function convertMessage$2(message) {
 	};
 	if (role === "tool") {
 		if (message.toolCallId === void 0) throw new ChatProviderError("Tool message missing `toolCallId`.");
+		const block = toolResultToBlock(message.toolCallId, message.content);
+		if (message.cacheHint?.isLastTurnEnd || message.cacheHint?.isSuddenLargeContext) block.cache_control = CACHE_CONTROL;
 		return {
 			role: "user",
-			content: [toolResultToBlock(message.toolCallId, message.content)]
+			content: [block]
 		};
 	}
 	const blocks = [];
@@ -8653,6 +8715,10 @@ function convertMessage$2(message) {
 			input: toolInput
 		});
 	}
+	if (message.cacheHint?.isLastTurnEnd || message.cacheHint?.isSuddenLargeContext) {
+		const lastBlock = blocks.at(-1);
+		if (lastBlock !== void 0 && CACHEABLE_TYPES.has(lastBlock.type)) lastBlock.cache_control = CACHE_CONTROL;
+	}
 	return {
 		role,
 		content: blocks
@@ -8901,8 +8967,16 @@ var AnthropicChatProvider = class {
 		return getAnthropicModelCapability(model ?? this._model);
 	}
 	async generate(systemPrompt, tools, history, options) {
-		const systemBlocks = splitSystemPrompt(systemPrompt, options?.cacheBreakpoints);
-		const system = systemBlocks.length > 0 ? systemBlocks : void 0;
+		let system;
+		if (options?.promptPlan) {
+			const systemBlocks = promptPlanToSystemBlocks(options.promptPlan);
+			system = systemBlocks.length > 0 ? systemBlocks : void 0;
+		} else if (systemPrompt) system = [{
+			type: "text",
+			text: systemPrompt,
+			cache_control: CACHE_CONTROL
+		}];
+		else system = void 0;
 		const messages = [];
 		for (const msg of history) {
 			const converted = convertMessage$2(msg);
@@ -45335,6 +45409,22 @@ function isEffectivelyEmptyContent(parts) {
 	}
 	return true;
 }
+/**
+* Derive a stable SHA256 hash from cacheable blocks in a PromptPlan.
+*
+* Only blocks with cacheScope 'global' are included in the hash, as OpenAI
+* only supports caching the prefix (global scope).
+*
+* @param promptPlan - The prompt plan containing cacheable blocks.
+* @returns A hexadecimal SHA256 hash string.
+*/
+function deriveCacheKeyFromPromptPlan$1(promptPlan) {
+	if (!promptPlan || promptPlan.blocks.length === 0) return createHash("sha256").digest("hex");
+	const cacheableTexts = [];
+	for (const block of promptPlan.blocks) if (block.cacheScope === "global") cacheableTexts.push(block.text);
+	const concatenated = cacheableTexts.join("");
+	return createHash("sha256").update(concatenated).digest("hex");
+}
 function convertMessage$1(message, reasoningKey, toolMessageConversion) {
 	let reasoningContent = "";
 	const nonThinkParts = [];
@@ -45569,6 +45659,10 @@ var OpenAICompletionsChatProvider = class {
 		};
 		if (tools.length > 0) createParams["tools"] = tools.map((t) => convertTool$1(t));
 		if (this._stream) createParams["stream_options"] = { include_usage: true };
+		if (options?.promptPlan) {
+			const cacheKey = deriveCacheKeyFromPromptPlan$1(options.promptPlan);
+			if (cacheKey) createParams["prompt_cache_key"] = cacheKey;
+		}
 		try {
 			return new OpenAICompletionsStreamedMessage(await this._createClient(options?.auth).chat.completions.create(createParams, options?.signal ? { signal: options.signal } : void 0), this._stream, this._reasoningKey);
 		} catch (error) {
@@ -45688,6 +45782,19 @@ function normalizeResponsesFinishReason(status, incompleteReason) {
 		rawFinishReason: null
 	};
 }
+/**
+* Derive a stable cache key from cacheable blocks in a PromptPlan.
+*
+* Only blocks with cacheScope 'global' are included in the hash, as OpenAI
+* only supports caching the prefix (global scope).
+*/
+function deriveCacheKeyFromPromptPlan(promptPlan) {
+	if (!promptPlan || promptPlan.blocks.length === 0) return void 0;
+	const cacheableTexts = [];
+	for (const block of promptPlan.blocks) if (block.cacheScope === "global") cacheableTexts.push(block.text);
+	if (cacheableTexts.length === 0) return void 0;
+	return createHash("sha256").update(cacheableTexts.join("")).digest("hex");
+}
 function asRawObject(value) {
 	if (value === null || typeof value !== "object" || Array.isArray(value)) return null;
 	return value;
@@ -46217,6 +46324,10 @@ var OpenAIResponsesChatProvider = class {
 				stream: this._stream,
 				...kwargs
 			};
+			if (options?.promptPlan) {
+				const cacheKey = deriveCacheKeyFromPromptPlan(options.promptPlan);
+				if (cacheKey) createParams["prompt_cache_key"] = cacheKey;
+			}
 			if (!("responses" in client) || typeof client.responses?.create !== "function") throw new Error("OpenAI SDK version does not support Responses API. Upgrade to >=4.x with responses support.");
 			return new OpenAIResponsesStreamedMessage(await client.responses.create(createParams, options?.signal ? { signal: options.signal } : void 0), this._stream);
 		} catch (error) {
@@ -46335,22 +46446,29 @@ function catalogModelToCapability(model) {
 	if (!isUsableChatModel(model)) return void 0;
 	const inputs = model.modalities?.input ?? [];
 	const output = model.limit?.output;
+	const base = {
+		image_in: inputs.includes("image"),
+		video_in: inputs.includes("video"),
+		audio_in: inputs.includes("audio"),
+		thinking: Boolean(model.reasoning),
+		tool_use: model.tool_call ?? true,
+		thinking_effort: false,
+		thinking_xhigh: false,
+		thinking_max: false,
+		max_context_tokens: context
+	};
+	const registry = resolveCapabilityFromRegistry(model.id);
+	const capability = registry !== void 0 ? {
+		...base,
+		...registry,
+		max_context_tokens: context
+	} : base;
 	return {
 		id: model.id,
 		name: typeof model.name === "string" && model.name.length > 0 ? model.name : void 0,
 		maxOutputSize: typeof output === "number" && output > 0 ? output : void 0,
 		reasoningKey: catalogReasoningKey(model.interleaved),
-		capability: {
-			image_in: inputs.includes("image"),
-			video_in: inputs.includes("video"),
-			audio_in: inputs.includes("audio"),
-			thinking: Boolean(model.reasoning),
-			tool_use: model.tool_call ?? true,
-			thinking_effort: false,
-			thinking_xhigh: false,
-			thinking_max: false,
-			max_context_tokens: context
-		}
+		capability
 	};
 }
 function catalogReasoningKey(interleaved) {
@@ -46535,6 +46653,17 @@ function addUsage(a, b) {
 		inputCacheCreation: a.inputCacheCreation + b.inputCacheCreation
 	};
 }
+/**
+* Compute the cache hit rate as a branded number between 0 and 1.
+*
+* Returns `undefined` when no input tokens were processed (inputTotal === 0),
+* so callers can distinguish "no data" from "zero hits".
+*/
+function cacheHitRate(usage) {
+	const total = inputTotal(usage);
+	if (total === 0) return void 0;
+	return usage.inputCacheRead / total;
+}
 //#endregion
 //#region ../../packages/kosong/src/index.ts
 /**
@@ -46695,7 +46824,7 @@ function redactCtx(ctx) {
 	};
 	return walk(ctx, 0);
 }
-function truncate$2(value, max) {
+function truncate$1(value, max) {
 	return value.length <= max ? value : value.slice(0, max - 1) + ELLIPSIS$3;
 }
 function serializeValue(raw) {
@@ -46719,7 +46848,7 @@ function quote(value) {
 	return `"${value.replaceAll("\\", "\\\\").replaceAll("\"", "\\\"").replaceAll("\n", "\\n")}"`;
 }
 function formatPair(key, raw) {
-	const limited = truncate$2(serializeValue(raw), CTX_VALUE_MAX_CHARS);
+	const limited = truncate$1(serializeValue(raw), CTX_VALUE_MAX_CHARS);
 	return `${SAFE_KEY_RE.test(key) ? key : quote(key)}=${/[\s="\\]/.test(limited) || limited.length === 0 ? quote(limited) : limited}`;
 }
 function clipBytes(text, maxBytes) {
@@ -46743,7 +46872,7 @@ function indentStack(stack) {
 function formatEntry(entry, options = {}) {
 	const ctx = entry.ctx ? redactCtx(entry.ctx) : void 0;
 	const omitContextKeys = new Set(options.omitContextKeys ?? []);
-	const msg = truncate$2(entry.msg, 200);
+	const msg = truncate$1(entry.msg, 200);
 	const pairs = [];
 	if (ctx) for (const [k, v] of Object.entries(ctx)) {
 		if (omitContextKeys.has(k)) continue;
@@ -54524,13 +54653,11 @@ function createSystemPromptRenderer(merged) {
 }
 function buildTemplateVars(context, promptVars) {
 	const skills = typeof context.skills === "string" ? context.skills : context.skills?.getModelSkillListing() ?? "";
-	const now = context.now instanceof Date ? context.now.toISOString() : context.now ?? (/* @__PURE__ */ new Date()).toISOString();
 	const agentsMd = context.agentsMd ?? "";
 	return {
 		...promptVars,
 		BYF_OS: context.osEnv.osKind,
 		BYF_SHELL: `${context.osEnv.shellName} (\`${context.osEnv.shellPath}\`)`,
-		BYF_NOW: now,
 		BYF_WORK_DIR: context.cwd,
 		BYF_AGENTS_MD: agentsMd,
 		BYF_AGENTS_MD_TOO_LONG: estimateTokens(agentsMd) > 4e3 ? "true" : "",
@@ -54611,10 +54738,10 @@ function normalizeSourcePath(path) {
 var agent_default$1 = "name: agent\ndescription: Default BYF agent\n\nsystemPromptPath: ./system.md\npromptVars:\n  roleAdditional: ''\n\ntools:\n  - Read\n  - Write\n  - Edit\n  - Grep\n  - Glob\n  - Bash\n  - TaskList\n  - TaskOutput\n  - TaskStop\n  - ReadMediaFile\n  - TodoList\n  - Skill\n  - WebSearch\n  - Agent\n  - FetchURL\n  - AskUserQuestion\n  - mcp__*\n\nsubagents:\n  coder:\n    description: Good at general software engineering tasks.\n  explore:\n    description: Fast codebase exploration with prompt-enforced read-only behavior.\n";
 //#endregion
 //#region ../../packages/agent-core/src/profile/default/coder.yaml
-var coder_default = "extends: agent\nname: coder\npromptVars:\n  roleAdditional: |\n    You are now running as a subagent. All the `user` messages are sent by the main agent. The main agent cannot see your context, it can only see your last message when you finish the task. You must treat the parent agent as your caller. Do not directly ask the end user questions. If something is unclear, explain the ambiguity in your final summary to the parent agent.\nwhenToUse: |\n  Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\ntools:\n  - Bash\n  - Read\n  - ReadMediaFile\n  - Glob\n  - Grep\n  - Write\n  - Edit\n  - WebSearch\n  - FetchURL\n  - mcp__*\n";
+var coder_default = "extends: agent\nname: coder\npromptVars:\n  roleAdditional: |\n    You are operating as a subagent instance. The main BYF agent spawned you to handle a specific task. All user messages come from the main agent — it cannot see your context and will only see your final result when you finish. Do not address the end user directly. If something is unclear, explain the ambiguity in your summary to the parent agent.\nwhenToUse: |\n  Use this agent for non-trivial software engineering work that may require reading files, editing code, running commands, and returning a compact but technically complete summary to the parent agent.\ntools:\n  - Bash\n  - Read\n  - ReadMediaFile\n  - Glob\n  - Grep\n  - Write\n  - Edit\n  - WebSearch\n  - FetchURL\n  - mcp__*\n";
 //#endregion
 //#region ../../packages/agent-core/src/profile/default/explore.yaml
-var explore_default = "extends: agent\nname: explore\npromptVars:\n  roleAdditional: |\n    You are now running as a subagent. All the `user` messages are sent by the main agent. The main agent cannot see your context, it can only see your last message when you finish the task. You must treat the parent agent as your caller. Do not directly ask the end user questions. If something is unclear, explain the ambiguity in your final summary to the parent agent.\n\n    You are a codebase exploration specialist. Your role is EXCLUSIVELY to search, read, and analyze existing code and resources. You do NOT have access to file editing tools.\n\n    Your strengths:\n    - Rapidly finding files using glob patterns\n    - Searching code and text with powerful regex patterns\n    - Reading and analyzing file contents\n    - Running read-only shell commands (git log, git diff, ls, find, etc.)\n\n    Guidelines:\n    - Use Glob for broad file pattern matching. Patterns MUST contain a literal anchor (extension or subdirectory); pure wildcards like `*` or `**/*` are rejected by the tool.\n    - Use Grep for searching file contents with regex\n    - Use Read when you know the specific file path\n    - Use Bash ONLY for read-only operations (ls, git status, git log, git diff, find)\n    - NEVER use Bash for any file creation or modification commands\n    - Adapt your search depth based on the thoroughness level specified by the caller\n    - Wherever possible, spawn multiple parallel tool calls for grepping and reading files to maximize speed\n\n    If the prompt includes a <git-context> block, use it to orient yourself about the repository state before starting your investigation.\n\n    You are meant to be a fast agent. Complete the search request efficiently and report your findings clearly in a structured format.\nwhenToUse: |\n  Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \"src/**/*.yaml\"), search code for keywords (e.g. \"database connection\"), or answer questions about the codebase (e.g. \"how does the auth module work?\"). When calling this agent, specify the desired thoroughness level: \"quick\" for basic searches, \"medium\" for moderate exploration, or \"thorough\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\ntools:\n  - Bash\n  - Read\n  - ReadMediaFile\n  - Glob\n  - Grep\n  - WebSearch\n  - FetchURL\n";
+var explore_default = "extends: agent\nname: explore\npromptVars:\n  roleAdditional: |\n    You are operating as a subagent instance. The main BYF agent spawned you to handle a specific task. All user messages come from the main agent — it cannot see your context and will only see your final result when you finish. Do not address the end user directly. If something is unclear, explain the ambiguity in your summary to the parent agent.\n\n    You are a codebase exploration specialist. Your role is EXCLUSIVELY to search, read, and analyze existing code and resources. You do NOT have access to file editing tools.\n\n    Your strengths:\n    - Rapidly finding files using glob patterns\n    - Searching code and text with powerful regex patterns\n    - Reading and analyzing file contents\n    - Running read-only shell commands (git log, git diff, ls, find, etc.)\n\n    Guidelines:\n    - Use Glob for broad file pattern matching. Patterns MUST contain a literal anchor (extension or subdirectory); pure wildcards like `*` or `**/*` are rejected by the tool.\n    - Use Grep for searching file contents with regex\n    - Use Read when you know the specific file path\n    - Use Bash ONLY for read-only operations (ls, git status, git log, git diff, find)\n    - NEVER use Bash for any file creation or modification commands\n    - Adapt your search depth based on the thoroughness level specified by the caller\n    - Wherever possible, spawn multiple parallel tool calls for grepping and reading files to maximize speed\n\n    If the prompt includes a <git-context> block, use it to orient yourself about the repository state before starting your investigation.\n\n    You are meant to be a fast agent. Complete the search request efficiently and report your findings clearly in a structured format.\nwhenToUse: |\n  Fast agent specialized for exploring codebases. Use this when you need to quickly find files by patterns (e.g. \"src/**/*.yaml\"), search code for keywords (e.g. \"database connection\"), or answer questions about the codebase (e.g. \"how does the auth module work?\"). When calling this agent, specify the desired thoroughness level: \"quick\" for basic searches, \"medium\" for moderate exploration, or \"thorough\" for comprehensive analysis across multiple locations and naming conventions. Use this agent for any read-only exploration that will clearly require more than 3 search queries. Prefer launching multiple explore agents concurrently when investigating independent questions.\ntools:\n  - Bash\n  - Read\n  - ReadMediaFile\n  - Glob\n  - Grep\n  - WebSearch\n  - FetchURL\n";
 //#endregion
 //#region ../../packages/agent-core/src/profile/default/init.md
 var init_default = "You are a software engineering expert with many years of programming experience. Please explore the current project directory to understand the project's architecture and main details.\n\nTask requirements:\n1. Analyze the project structure and identify key configuration files (such as pyproject.toml, package.json, Cargo.toml, etc.).\n2. Understand the project's technology stack, build process and runtime architecture.\n3. Identify how the code is organized and main module divisions.\n4. Discover project-specific development conventions, testing strategies, and deployment processes.\n\nAfter the exploration, you should do a thorough summary of your findings and overwrite it into `AGENTS.md` file in the project root. You need to refer to what is already in the file when you do so.\n\nFor your information, `AGENTS.md` is a file intended to be read by AI coding agents. Expect the reader of this file know nothing about the project.\n\nYou should compose this file according to the actual project content. Do not make any assumptions or generalizations. Ensure the information is accurate and useful. You must use the natural language that is mainly used in the project's comments and documentation.\n\nPopular sections that people usually write in `AGENTS.md` are:\n\n- Project overview\n- Build and test commands\n- Code style guidelines\n- Testing instructions\n- Security considerations\n";
@@ -54624,7 +54751,7 @@ const PROFILE_SOURCES = {
 	"profile/default/agent.yaml": agent_default$1,
 	"profile/default/coder.yaml": coder_default,
 	"profile/default/explore.yaml": explore_default,
-	"profile/default/system.md": "You are BYF, an interactive general AI agent running on a user's computer.\n\nYour primary goal is to help users with software engineering tasks by taking action — use the tools available to you to make real changes on the user's system. You should also answer questions when asked. Always adhere strictly to the following system instructions and the user's requirements.\n\n{{ ROLE_ADDITIONAL }}\n\n# Prompt and Tool Use\n\nThe user's messages may contain questions and/or task descriptions in natural language, code snippets, logs, file paths, or other forms of information. Read them, understand them and do what the user requested. For simple questions/greetings that do not involve any information in the working directory or on the internet, you may simply reply directly. For anything else, default to taking action with tools. When the request could be interpreted as either a question to answer or a task to complete, treat it as a task.\n\nWhen handling the user's request, if it involves creating, modifying, or running code or files, you MUST use the appropriate tools (e.g., `Write`, `Bash`) to make actual changes — do not just describe the solution in text. For questions that only need an explanation, you may reply in text directly. When calling tools, do not provide explanations because the tool calls themselves should be self-explanatory. You MUST follow the description of each tool and its parameters when calling tools.\n\nIf the `Agent` tool is available, you can use it to delegate a focused subtask to a subagent instance. The tool can either start a new instance or resume an existing one by its agent id. Subagent instances are persistent session objects with their own context history. When delegating, provide a complete prompt with all necessary context — a new subagent instance does not see your current context. If an existing subagent already has useful context or the task clearly continues its prior work, prefer resuming it over creating a new instance. Default to foreground subagents; use `run_in_background=true` only when there is a clear benefit to letting the conversation continue before the subagent finishes and you do not need the result immediately.\n\nYou have the capability to output any number of tool calls in a single response. If you anticipate making multiple non-interfering tool calls, you are HIGHLY RECOMMENDED to make them in parallel to significantly improve efficiency. This is very important to your performance.\n\nThe results of the tool calls will be returned to you in a tool message. You must determine your next action based on the tool call results, which could be one of the following: 1. Continue working on the task, 2. Inform the user that the task is completed or has failed, or 3. Ask the user for more information.\n\nThe system may insert information wrapped in `<system>` tags within user or tool messages. This information provides supplementary context relevant to the current task — take it into consideration when determining your next action.\n\nTool results and user messages may also include `<system-reminder>` tags. Unlike `<system>` tags, these are **authoritative system directives** that you MUST follow. They bear no direct relation to the specific tool results or user messages in which they appear. Always read them carefully and comply with their instructions — they may override or constrain your normal behavior.\n\nIf the `Bash`, `TaskList`, `TaskOutput`, and `TaskStop` tools are available and you are the root agent, you can use background `Bash` for long-running shell commands. Launch it via `Bash` with `run_in_background=true` and a short `description`. The system will notify you when the background task reaches a terminal state. Use `TaskList` to re-enumerate active tasks when needed, especially after context compaction. Use `TaskOutput` for non-blocking status/output snapshots; only set `block=true` when you intentionally want to wait for completion. After starting a background task, default to returning control to the user instead of immediately waiting on it. Use `TaskStop` only when you need to cancel the task. For human users in the interactive shell, the only task-management slash command is `/tasks`. Do not tell users to run `/task`, `/tasks list`, `/tasks output`, `/tasks stop`, or any other invented slash subcommands. If you are a subagent or these tools are not available, do not assume you can create or control background tasks.\n\nIf a foreground tool call or a background agent requests approval, the approval is coordinated through the unified approval runtime and surfaced through the root UI channel. Do not assume approvals are local to a single subagent turn.\n\nWhen responding to the user, you MUST use the SAME language as the user, unless explicitly instructed to do otherwise.\n\n# Tool Efficiency Guidelines\n\nThe following common command categories are usually available in Bash. Availability depends on the host, so when in doubt run `which <command>` first to confirm a command exists before relying on it.\n- Navigation and inspection: `ls`, `pwd`, `cd`, `stat`, `file`, `du`, `df`, `tree`\n- File and directory management: `cp`, `mv`, `rm`, `mkdir`, `touch`, `ln`, `chmod`, `chown`\n- Text and data processing: `wc`, `sort`, `uniq`, `cut`, `tr`, `diff`, `xargs`\n- Archives and compression: `tar`, `gzip`, `gunzip`, `zip`, `unzip`\n- Networking and transfer: `curl`, `wget`, `ping`, `ssh`, `scp`\n- Version control: `git`\n- Process and system: `ps`, `kill`, `top`, `env`, `date`, `uname`, `whoami`\n- Language and package toolchains: `node`, `npm`, `pnpm`, `yarn`, `python`, `pip` (use whichever the project actually relies on)\n\nWhen using Bash:\n- For multiple related commands, use `&&` to chain them in a single call, e.g. `cd /path && ls -la`\n- Use `;` to run commands sequentially regardless of success/failure\n- Use `||` for conditional execution (run second command only if first fails)\n- Use pipe operations (`|`) and redirections (`>`, `>>`) to chain input and output between commands\n- Always quote file paths containing spaces with double quotes (e.g., cd \"/path with spaces/\")\n- Compose multi-step logic in a single call with `if` / `case` / `for` / `while` control flows.\n\nWhen using Grep, ALWAYS use the Grep tool instead of running `grep` or `rg` from a shell — direct shell calls bypass workspace policy, output limits, and sensitive-file filtering. Use Grep only when the task is to search for unknown content or locations.\n\nWhen reading files:\n- If the user provides a concrete file path to a text file, call Read directly. Do not `Glob`, `ls`, or otherwise pre-check known text file paths; missing or invalid file paths return errors you can handle.\n- Do not use Read for directories; use `ls` via Bash for a known directory, or Glob when you need files/directories matching a pattern.\n\nWhen editing files:\n- Use Edit for targeted changes to existing files; use Write only for new files or complete overwrites.\n- To modify a file, always use Edit; do not run a Shell `sed` command for edits.\n\n# First Principles\n\nThink from first principles. Start from real requirements, code facts, and verification results; if the goal is unclear, discuss it with the user first. Treat code, not documentation, as the source of truth.\n\n# General Guidelines for Coding\n\nWhen building something from scratch, you should:\n\n- Understand the user's requirements.\n- Ask the user for clarification if there is anything unclear.\n- Design the architecture and make a plan for the implementation.\n- Write the code in a modular and maintainable way.\n\nAlways use tools to implement your code changes:\n\n- Use `Write` to create or overwrite source files. Code that only appears in your text response is NOT saved to the file system and will not take effect.\n- Use `Bash` to run and test your code after writing it.\n- Iterate: if tests fail, read the error, fix the code with `Write` or `Edit`, and re-test with `Bash`.\n\nWhen working on an existing codebase, you should:\n\n- Understand the codebase by reading it with tools (`Read`, `Glob`, `Grep`) before making changes. Identify the ultimate goal and the most important criteria to achieve the goal.\n- When using `Glob`, include a literal anchor (file extension or subdirectory) in the pattern. Pure wildcards like `*` or `**/*` are rejected by the tool.\n- For a bug fix, you typically need to check error logs or failed tests, scan over the codebase to find the root cause, and figure out a fix. If user mentioned any failed tests, you should make sure they pass after the changes.\n- For a feature, you typically need to design the architecture, and write the code in a modular and maintainable way, with minimal intrusions to existing code. Add new tests if the project already has tests.\n- For a code refactoring, you typically need to update all the places that call the code you are refactoring if the interface changes. DO NOT change any existing logic especially in tests, focus only on fixing any errors caused by the interface changes.\n- Make MINIMAL changes to achieve the goal. This is very important to your performance.\n- Follow the coding style of existing code in the project.\n- For broader codebase exploration and deep research, use `Agent` with `subagent_type=\"explore\"` — a fast, read-only agent specialized for searching and understanding codebases. Reach for it when your task will clearly require more than 3 search queries, or when you need to investigate multiple files and patterns. Launch multiple explore agents concurrently when investigating independent questions.\n\nDO NOT run `git commit`, `git push`, `git reset`, `git rebase` and/or do any other git mutations unless explicitly asked to do so. Ask for confirmation each time when you need to do git mutations, even if the user has confirmed in earlier conversations.\n\n# General Guidelines for Research and Data Processing\n\n- Understand the user's requirements thoroughly, ask for clarification before you start if needed.\n- Make plans before doing deep or wide research, to ensure you are always on track.\n- Search on the Internet if possible, with carefully-designed search queries to improve efficiency and accuracy.\n- Use proper tools or shell commands or Python packages to process or generate images, videos, PDFs, docs, spreadsheets, presentations, or other multimedia files. Detect if there are already such tools in the environment. If you have to install third-party tools/packages, you MUST ensure that they are installed in a virtual/isolated environment.\n- Avoid installing or deleting anything to/from outside of the current working directory. If you have to do so, ask the user for confirmation.\n\n# Working Environment\n\n## Operating System\n\nYou are running on **{{ BYF_OS }}**. The Bash tool executes commands using **{{ BYF_SHELL }}**.\n{% if BYF_OS == \"Windows\" %}\n\nIMPORTANT: You are on Windows. The Bash tool runs through Git Bash, so use Unix shell syntax inside Bash commands — `/dev/null` not `NUL`, and forward slashes in paths. For file operations, always prefer the built-in tools (Read, Write, Edit, Glob, Grep) over Bash commands — they work reliably across all platforms.\n{% endif %}\n\nThe operating environment is not in a sandbox. Any actions you do will immediately affect the user's system. So you MUST be extremely cautious. Unless being explicitly instructed to do so, you should never access (read/write/execute) files outside of the working directory.\n\n## Date and Time\n\nThe current date and time in ISO format is `{{ BYF_NOW }}`. This is only a reference for you when searching the web, or checking file modification time, etc. If you need the exact time, use Bash tool with proper command.\n\n## Working Directory\n\nThe current working directory is `{{ BYF_WORK_DIR }}`. This should be considered as the project root if you are instructed to perform tasks on the project. Every file system operation will be relative to the working directory if you do not explicitly specify the absolute path. Tools may require absolute paths for some parameters, IF SO, YOU MUST use absolute paths for these parameters.\n{% if BYF_ADDITIONAL_DIRS_INFO %}\n\n## Additional Directories\n\nThe following directories have been added to the workspace. You can read, write, search, and glob files in these directories as part of your workspace scope.\n\n{{ BYF_ADDITIONAL_DIRS_INFO }}\n{% endif %}\n\n__CACHE_BOUNDARY__\n\n# Project Information\n\nMarkdown files named `AGENTS.md` usually contain the background, structure, coding styles, user preferences and other relevant information about the project. You should use this information to understand the project and the user's preferences. `AGENTS.md` files may exist at different locations in the project, but typically there is one in the project root.\n\n> Why `AGENTS.md`?\n>\n> `README.md` files are for humans: quick starts, project descriptions, and contribution guidelines. `AGENTS.md` complements this by containing the extra, sometimes detailed context coding agents need: build steps, tests, and conventions that might clutter a README or aren’t relevant to human contributors.\n>\n> We intentionally kept it separate to:\n>\n> - Give agents a clear, predictable place for instructions.\n> - Keep `README`s concise and focused on human contributors.\n> - Provide precise, agent-focused guidance that complements existing `README` and docs.\n\n{% if BYF_AGENTS_MD_TOO_LONG %}\n> ⚠️ The merged AGENTS.md content exceeds 4,000 tokens. Consider compressing project instructions to reduce context usage.\n{% endif %}\n\nThe `AGENTS.md` instructions (merged from all applicable directories):\n\n`````````\n{{ BYF_AGENTS_MD }}\n`````````\n\n`AGENTS.md` files can appear at any level of the project directory tree, including inside `.byf/` directories. Each file governs the directory it resides in and all subdirectories beneath it. When multiple `AGENTS.md` files apply to a file you are modifying, instructions in deeper directories take precedence over those in parent directories. User instructions given directly in the conversation always take the highest precedence.\n\nWhen working on files in subdirectories, always check whether those directories contain their own `AGENTS.md` with more specific guidance that supplements or overrides the instructions above. You may also check `README`/`README.md` files for more information about the project.\n\nIf you modified any files/styles/structures/configurations/workflows/... mentioned in `AGENTS.md` files, you MUST update the corresponding `AGENTS.md` files to keep them up-to-date.\n\n# Skills\n\nSkills are reusable capabilities. When a skill from the listing matches the user's request, you MUST call the `Skill` tool (not free-form text).\n\n{{ BYF_SKILLS }}\n\n# Ultimate Reminders\n\nAt any time, you should be HELPFUL, CONCISE, and ACCURATE. Be thorough in your actions — test what you build, verify what you change — not in your explanations.\n\n- Never give the user more than what they want.\n- Try your best to avoid any hallucination. Do fact checking before providing any factual information.\n- Do not give up too early.\n- ALWAYS, keep it stupidly simple. Do not overcomplicate things.\n"
+	"profile/default/system.md": "You are BYF, an AI agent running on the user's computer. Your job is to help\nusers accomplish tasks by taking action — read, write, search, and execute to\nmake real changes on the user's system. Answer questions when asked; otherwise,\nact.\n\nWhen responding, use the same language as the user unless explicitly instructed\notherwise.\n\n{{ ROLE_ADDITIONAL }}\n\n# First Principles\n\nThink from first principles. Strip away assumptions and conventions; every\naction must be traceable to a verifiable fact — the actual file contents,\ncommand output, data, or the user's explicit words. When in doubt, read\nbefore guessing, ask before assuming, verify before claiming.\n\n# Tool Use\n\nUse tools only when the task requires them. If the request can be answered\nwithout reading files, running commands, or searching the web, reply in text\ndirectly. When a request is ambiguous, prefer action — the user can see your\noutput and correct course.\n\nCode that only appears in your text response is NOT saved to the file system\nand will not take effect. To create or modify files, use `Write` or `Edit`.\nTo run commands, use `Bash`.\n\n# Protocol\n\n<system> tags in user or tool messages provide supplementary context. Treat\nthem as background information.\n\n<system-reminder> tags are authoritative directives that override default\nbehavior. They are unrelated to the messages they appear in. Always comply.\n\n# Safety\n\nThe environment is not a sandbox — your actions immediately affect the user's\nsystem.\n\n- Stay within the working directory unless explicitly instructed otherwise.\n- Git operations are destructive and may affect remote repositories. Never\n  execute git mutations unless explicitly asked; confirm each time.\n- Avoid installing or deleting anything outside the working directory. If\n  necessary, ask for confirmation first.\n\n# Working Environment\n\n## Operating System\n\nYou are running on **{{ BYF_OS }}**. The Bash tool executes commands using **{{ BYF_SHELL }}**.\n{% if BYF_OS == \"Windows\" %}\n\nIMPORTANT: You are on Windows. The Bash tool runs through Git Bash, so use Unix shell syntax inside Bash commands — `/dev/null` not `NUL`, and forward slashes in paths. For file operations, always prefer the built-in tools (Read, Write, Edit, Glob, Grep) over Bash commands — they work reliably across all platforms.\n{% endif %}\n\n## Working Directory\n\nThe current working directory is `{{ BYF_WORK_DIR }}`. This should be considered as the project root if you are instructed to perform tasks on the project. Every file system operation will be relative to the working directory if you do not explicitly specify the absolute path. Tools may require absolute paths for some parameters, IF SO, YOU MUST use absolute paths for these parameters.\n{% if BYF_ADDITIONAL_DIRS_INFO %}\n\n## Additional Directories\n\nThe following directories have been added to the workspace. You can read, write, search, and glob files in these directories as part of your workspace scope.\n\n{{ BYF_ADDITIONAL_DIRS_INFO }}\n{% endif %}\n\n# Project Information\n\n`AGENTS.md` files contain project-specific context, styles, and conventions for agents. They may exist at different locations in the project — each file governs its directory and all subdirectories beneath it. Deeper files take precedence over parent files.\n\nIf instructions conflict:\n- `<system-reminder>` directives override all other instructions, including user messages.\n- Safety rules are hard constraints and must never be violated, even if a user message or AGENTS.md says otherwise.\n- Beyond those two, user messages > AGENTS.md > default system instructions.\n\n{% if BYF_AGENTS_MD_TOO_LONG %}\n> ⚠️ The merged AGENTS.md content exceeds 4,000 tokens. Consider compressing project instructions to reduce context usage.\n{% endif %}\n\nThe `AGENTS.md` instructions (merged from all applicable directories):\n\n`````````\n{{ BYF_AGENTS_MD }}\n`````````\n\nIf you modified anything mentioned in `AGENTS.md` files, update the corresponding files to keep them up-to-date.\n\n# Skills\n\nSkills are reusable capabilities. When a skill from the listing matches the user's request, you MUST call the `Skill` tool (not free-form text).\n\n{{ BYF_SKILLS }}\n"
 };
 const DEFAULT_INIT_PROMPT = init_default;
 const DEFAULT_AGENT_PROFILES = loadAgentProfilesFromSources([
@@ -58257,7 +58384,6 @@ function isWithin(child, parent) {
 }
 //#endregion
 //#region ../../packages/agent-core/src/skill/registry.ts
-const LISTING_DESC_MAX = 100;
 var SkillRegistry = class {
 	byName = /* @__PURE__ */ new Map();
 	roots = [];
@@ -58359,10 +58485,7 @@ function formatFullSkill(skill) {
 	];
 }
 function formatModelSkill(skill) {
-	return [`- ${skill.name}: ${truncate$1(skill.description, LISTING_DESC_MAX)}`];
-}
-function truncate$1(value, max) {
-	return value.length > max ? value.slice(0, max) : value;
+	return [`- ${skill.name}: ${skill.description}`];
 }
 //#endregion
 //#region ../../packages/agent-core/src/tools/builtin/collaboration/skill-tool.md
@@ -65765,9 +65888,6 @@ function notificationKey(origin) {
 }
 //#endregion
 //#region ../../node_modules/.pnpm/@antfu+utils@9.3.0/node_modules/@antfu/utils/dist/index.mjs
-function uniq(array) {
-	return Array.from(new Set(array));
-}
 function notNullish(v) {
 	return v != null;
 }
@@ -66612,6 +66732,19 @@ var FullCompaction = class {
 	computeCompactableCount(history) {
 		return sliceCompleteMessages(history, this.strategy.computeCompactCount(history, this.maxContextSize));
 	}
+	restoreRecord(record) {
+		switch (record.type) {
+			case "full_compaction.begin":
+				this.begin(record);
+				break;
+			case "full_compaction.cancel":
+				this.cancel();
+				break;
+			case "full_compaction.complete":
+				this.complete(record);
+				break;
+		}
+	}
 };
 function extractCompactionSummary(response) {
 	const summary = typeof response.message.content === "string" ? response.message.content : response.message.content.map((part) => part.type === "text" ? part.text : "").join("");
@@ -66753,6 +66886,13 @@ var ConfigState = class {
 			return;
 		}
 	}
+	restoreRecord(record) {
+		switch (record.type) {
+			case "config.update":
+				this.update(record);
+				break;
+		}
+	}
 };
 //#endregion
 //#region ../../packages/agent-core/src/agent/context/output-offloading.ts
@@ -67123,6 +67263,83 @@ var ContextMemory = class {
 			});
 		}
 	}
+	restoreRecord(record) {
+		switch (record.type) {
+			case "context.append_message":
+				this.appendMessage(record.message);
+				break;
+			case "context.clear":
+				this.restoreClear();
+				break;
+			case "context.apply_compaction":
+				this.restoreApplyCompaction(record);
+				break;
+			case "context.mark_last_user_prompt_blocked":
+				this.restoreMarkLastUserPromptBlocked(record);
+				break;
+			case "context.append_loop_event":
+				this.restoreAppendLoopEvent(record);
+				break;
+			case "context.observation_masking":
+				this.restoreObservationMasking();
+				break;
+		}
+	}
+	restoreClear() {
+		this._history = [];
+		this._tokenCount = 0;
+		this.tokenCountCoveredMessageCount = 0;
+		this.openSteps.clear();
+		this.pendingToolResultIds.clear();
+		this.deferredMessages = [];
+		this.toolCallInfo.clear();
+		this.agent.injection.onContextClear();
+		this.agent.emitStatusUpdated();
+	}
+	restoreApplyCompaction(record) {
+		const compactedCount = record.compactedCount;
+		const summary = record.summary;
+		const tokensAfter = record.tokensAfter;
+		this._history = [{
+			role: "assistant",
+			content: [{
+				type: "text",
+				text: summary
+			}],
+			toolCalls: [],
+			origin: { kind: "compaction_summary" }
+		}, ...this._history.slice(compactedCount)];
+		this.openSteps.clear();
+		this.flushDeferredMessagesIfToolExchangeClosed();
+		this._tokenCount = tokensAfter;
+		this.tokenCountCoveredMessageCount = this._history.length;
+		this.agent.injection.onContextCompacted(compactedCount);
+		this.agent.emitStatusUpdated();
+	}
+	restoreMarkLastUserPromptBlocked(record) {
+		const hookEvent = record.hookEvent;
+		for (let i = this._history.length - 1; i >= 0; i--) {
+			const message = this._history[i];
+			if (message?.role !== "user" || message.origin?.kind !== "user") continue;
+			this._history[i] = {
+				...message,
+				origin: {
+					...message.origin,
+					blockedByHook: hookEvent
+				}
+			};
+			return;
+		}
+	}
+	async restoreAppendLoopEvent(record) {
+		await this.appendLoopEvent(record.event);
+	}
+	restoreObservationMasking() {
+		const maxContextSize = this.agent.config.modelCapabilities.max_context_tokens;
+		const { history } = applyObservationMasking(this._history, maxContextSize, this.toolCallInfo);
+		this._history = history;
+		this.agent.emitStatusUpdated();
+	}
 };
 function toolResultOutputForModel(result) {
 	const output = result.output;
@@ -67563,13 +67780,16 @@ var DirectoryTreeInjector = class extends DynamicInjector {
 	injectionVariant = "directory_tree";
 	lastTree;
 	hasInjected = false;
+	capturedTimestamp;
 	async getInjection() {
 		const kaos = this.agent.runtime.kaos;
-		const tree = await buildTree(kaos, this.agent.config.cwd || kaos.getcwd());
+		const workDir = this.agent.config.cwd || kaos.getcwd();
+		const tree = await buildTree(kaos, workDir);
 		if (this.hasInjected && tree === this.lastTree) return;
 		this.lastTree = tree;
 		this.hasInjected = true;
-		return tree;
+		if (this.capturedTimestamp === void 0) this.capturedTimestamp = (/* @__PURE__ */ new Date()).toISOString();
+		return `Current working directory structure (${workDir}):\n${tree}\n\nThe current date and time in ISO format is \`${this.capturedTimestamp}\`. This is only a reference for you when searching the web or checking file modification time, etc. If you need the exact time, use Bash tool with proper command.`;
 	}
 };
 async function buildTree(kaos, workDir) {
@@ -70181,16 +70401,20 @@ var PermissionManager = class {
 			block: true,
 			reason: this.formatMessage(name, matchedRule?.reason)
 		};
-		const policyResult = await this.evaluatePolicies(context, matchedRule);
-		if (policyResult !== void 0) return this.permissionPolicyResultToPrepare(policyResult, context);
 		if (mode === "auto") {
+			const policyResult = await this.evaluatePolicies(context, matchedRule);
+			if (policyResult !== void 0) return this.permissionPolicyResultToPrepare(policyResult, context);
 			if (this.wouldAskInManualMode(name, args)) this.trackToolApproved(name, "afk");
 			return;
 		}
 		if (mode === "yolo") {
+			const policyResult = await this.evaluatePolicies(context, matchedRule);
+			if (policyResult !== void 0) return this.permissionPolicyResultToPrepare(policyResult, context);
 			if (this.wouldAskInManualMode(name, args)) this.trackToolApproved(name, "yolo");
 			return;
 		}
+		const policyResult = await this.evaluatePolicies(context, matchedRule);
+		if (policyResult !== void 0) return this.permissionPolicyResultToPrepare(policyResult, context);
 		if (decision === "allow") {
 			if (matchedRule?.scope === "session-runtime") this.trackToolApproved(name, "auto_session", "session");
 			return;
@@ -70309,6 +70533,16 @@ var PermissionManager = class {
 		if (scope !== void 0) properties["scope"] = scope;
 		this.agent.telemetry.track("tool_approved", properties);
 	}
+	restoreRecord(record) {
+		switch (record.type) {
+			case "permission.set_mode":
+				this.setMode(record.mode);
+				break;
+			case "permission.record_approval_result":
+				this.recordApprovalResult(record);
+				break;
+		}
+	}
 };
 function approvalTelemetryMode(mode) {
 	return mode === "auto" ? "afk" : mode;
@@ -70500,80 +70734,12 @@ function parseRecordLine(line, lineNumber, filePath, allowTruncated) {
 }
 //#endregion
 //#region ../../packages/agent-core/src/agent/records/index.ts
-async function restoreAgentRecord(agent, input) {
-	switch (input.type) {
-		case "metadata": return;
-		case "turn.prompt":
-			agent.turn.restorePrompt();
-			return;
-		case "turn.steer":
-			agent.turn.restoreSteer(input.input, input.origin);
-			return;
-		case "turn.cancel":
-			agent.turn.cancel(input.turnId);
-			return;
-		case "background.stop": return;
-		case "config.update":
-			agent.config.update(input);
-			return;
-		case "permission.set_mode":
-			agent.permission.setMode(input.mode);
-			return;
-		case "permission.record_approval_result":
-			agent.permission.recordApprovalResult(input);
-			return;
-		case "usage.record":
-			agent.usage.record(input.model, input.usage, "session");
-			return;
-		case "full_compaction.begin":
-			agent.fullCompaction.begin(input);
-			return;
-		case "full_compaction.cancel":
-			agent.fullCompaction.cancel();
-			return;
-		case "full_compaction.complete":
-			agent.fullCompaction.complete(input);
-			return;
-		case "plan_mode.enter":
-		case "plan_mode.cancel":
-		case "plan_mode.exit": return;
-		case "context.append_message":
-			agent.context.appendMessage(input.message);
-			return;
-		case "context.mark_last_user_prompt_blocked":
-			agent.context.markLastUserPromptBlocked(input.hookEvent);
-			return;
-		case "context.append_loop_event":
-			await agent.context.appendLoopEvent(input.event);
-			return;
-		case "context.clear":
-			agent.context.clear();
-			return;
-		case "context.apply_compaction":
-			agent.context.applyCompaction(input);
-			return;
-		case "context.observation_masking":
-			agent.context.applyObservationMasking();
-			return;
-		case "tools.register_user_tool":
-			agent.tools.registerUserTool(input);
-			return;
-		case "tools.unregister_user_tool":
-			agent.tools.unregisterUserTool(input.name);
-			return;
-		case "tools.set_active_tools":
-			agent.tools.setActiveTools(input.names);
-			return;
-		case "tools.update_store":
-			agent.tools.updateStore(input.key, input.value);
-			return;
-	}
-}
 var AgentRecords = class {
 	agent;
 	persistence;
 	_restoring = false;
 	metadataInitialized = false;
+	handlers = {};
 	constructor(agent, persistence) {
 		this.agent = agent;
 		this.persistence = persistence;
@@ -70581,6 +70747,9 @@ var AgentRecords = class {
 	get restoring() {
 		return this._restoring;
 	}
+	registerHandlers(handlers) {
+		this.handlers = { ...handlers };
+	}
 	logRecord(record) {
 		if (this._restoring) return;
 		const stamped = record.time !== void 0 ? record : {
@@ -70601,11 +70770,30 @@ var AgentRecords = class {
 	restore(record) {
 		this._restoring = true;
 		try {
-			restoreAgentRecord(this.agent, record);
+			this.routeToHandler(record);
 		} finally {
 			this._restoring = false;
 		}
 	}
+	routeToHandler(record) {
+		const handlerKey = this.getHandlerKey(record.type);
+		if (handlerKey === null || this.handlers[handlerKey] === void 0) return;
+		this.handlers[handlerKey].restoreRecord(record);
+	}
+	getHandlerKey(recordType) {
+		if (recordType === "metadata") return null;
+		return {
+			context: "context",
+			config: "config",
+			turn: "turn",
+			permission: "permission",
+			tools: "tools",
+			usage: "usage",
+			background: "background",
+			full_compaction: "fullCompaction",
+			plan_mode: "planMode"
+		}[recordType.split(".")[0]] ?? null;
+	}
 	async replay() {
 		if (!this.persistence) throw new Error("No persistence provided for AgentRecords");
 		let migrations = [];
@@ -82867,11 +83055,87 @@ var ToolManager = class {
 		return (input) => withProviderRequestAuth(resolveAuth, (auth) => uploadVideo(input, { auth }));
 	}
 	get loopTools() {
+		const builtinNames = [...this.builtinTools.keys()].filter((name) => this.enabledTools.has(name)).sort();
+		const userNames = [...this.userTools.keys()].filter((name) => this.enabledTools.has(name)).sort();
 		const mcpNames = [...this.mcpTools.keys()].filter((name) => this.isMcpToolEnabled(name));
-		return uniq([...this.enabledTools, ...mcpNames]).toSorted((a, b) => a.localeCompare(b)).map((name) => this.userTools.get(name) ?? this.mcpTools.get(name)?.tool ?? this.builtinTools.get(name)).filter((tool) => !!tool);
+		return [
+			...builtinNames.map((name) => this.builtinTools.get(name)),
+			...userNames.map((name) => this.userTools.get(name)),
+			...mcpNames.map((name) => this.mcpTools.get(name)?.tool)
+		].filter((tool) => !!tool);
+	}
+	restoreRecord(record) {
+		switch (record.type) {
+			case "tools.register_user_tool":
+				this.registerUserTool(record);
+				break;
+			case "tools.unregister_user_tool":
+				this.unregisterUserTool(record.name);
+				break;
+			case "tools.set_active_tools":
+				this.setActiveTools(record.names);
+				break;
+			case "tools.update_store":
+				this.updateStore(record.key, record.value);
+				break;
+		}
 	}
 };
 //#endregion
+//#region ../../packages/agent-core/src/agent/cache-staking/index.ts
+const DEFAULT_SIZE_THRESHOLD = 2e3;
+/**
+* Apply cache staking hints to a message array based on turn boundaries.
+*
+* - **Stake 3**: Tags the last assistant message of the previous turn with
+*   `cacheHint.isLastTurnEnd = true`.
+* - **Stake 4** (conditional): Tags the largest content block in the current
+*   turn that exceeds `sizeThreshold` with `cacheHint.isSuddenLargeContext = true`.
+*
+* Returns a new array with shallow-copied messages (original messages are
+* not mutated).
+*/
+function applyCacheStaking(messages, context) {
+	const { previousTurnMessageCount, sizeThreshold = DEFAULT_SIZE_THRESHOLD } = context;
+	if (previousTurnMessageCount <= 0 || messages.length === 0) return messages;
+	const result = messages.map((msg) => ({ ...msg }));
+	const lastTurnIndex = previousTurnMessageCount - 1;
+	if (lastTurnIndex < result.length) {
+		const lastTurnMsg = result[lastTurnIndex];
+		if (lastTurnMsg.role === "assistant") {
+			const existingHint = lastTurnMsg.cacheHint ?? {};
+			result[lastTurnIndex] = {
+				...lastTurnMsg,
+				cacheHint: {
+					...existingHint,
+					isLastTurnEnd: true
+				}
+			};
+		}
+	}
+	let largestIndex = -1;
+	let largestSize = 0;
+	for (let i = previousTurnMessageCount; i < result.length; i++) {
+		const contentLength = result[i].content.filter((p) => p.type === "text").reduce((sum, p) => sum + p.text.length, 0);
+		if (contentLength >= sizeThreshold && contentLength > largestSize) {
+			largestSize = contentLength;
+			largestIndex = i;
+		}
+	}
+	if (largestIndex >= 0) {
+		const target = result[largestIndex];
+		const existingHint = target.cacheHint ?? {};
+		result[largestIndex] = {
+			...target,
+			cacheHint: {
+				...existingHint,
+				isSuddenLargeContext: true
+			}
+		};
+	}
+	return result;
+}
+//#endregion
 //#region ../../packages/agent-core/src/agent/turn/canonical-args.ts
 /**
 * JSON canonicalization used by tool-call telemetry and dedup.
@@ -82893,6 +83157,231 @@ function isPlainRecord(value) {
 	return proto === Object.prototype || proto === null;
 }
 //#endregion
+//#region ../../packages/agent-core/src/prompt-plan/builder.ts
+/**
+* Cache boundary marker used to split the system prompt into cacheable blocks.
+*
+* @deprecated This marker is deprecated in favor of implicit boundaries based on section headers.
+* It is still supported for backward compatibility.
+*/
+const CACHE_BOUNDARY_MARKER = "__CACHE_BOUNDARY__";
+/**
+* Section headers that define implicit cache boundaries.
+*
+* These headers mark natural breaks in the system prompt where cache boundaries should be placed:
+* - "# Project Information" marks the start of project-specific content
+* - "# Skills" marks the start of session-specific skills listing
+*/
+const IMPLICIT_BOUNDARY_HEADERS = ["# Project Information", "# Skills"];
+/**
+* Block names by position.
+*
+* - First block (before first marker): 'base'
+* - Last block (after last marker): 'sessionContext'
+* - Intermediate blocks: Sequential names from 'projectInstructions', 'skillsListing', etc.
+*/
+const BLOCK_NAMES = [
+	"base",
+	"projectInstructions",
+	"skillsListing",
+	"sessionContext"
+];
+/**
+* Get the cache scope for a block by its position.
+*
+* @param position - The block position (0-indexed)
+* @param totalBlocks - Total number of blocks
+* @returns The default cache scope for this position
+*/
+function getDefaultScopeForPosition(position, totalBlocks) {
+	if (position === 0) return "global";
+	if (position === totalBlocks - 1) return "session";
+	if (position === 1) return "project";
+	return "session";
+}
+/**
+* Filter cache scope based on provider's supported scopes.
+*
+* @param scope - The desired cache scope
+* @param capability - The provider's cache capability
+* @returns The filtered scope (or 'none' if not supported)
+*/
+function filterScopeByCapability(scope, capability) {
+	if (capability.strategy === "none") return "none";
+	if (capability.supportedScopes === void 0) return scope;
+	if (capability.supportedScopes.includes(scope)) return scope;
+	return "none";
+}
+/**
+* Find implicit boundary positions in the system prompt.
+*
+* Searches for section headers that mark natural cache boundaries.
+* Returns sorted indices of where each boundary header starts.
+*
+* @param prompt - The system prompt to search
+* @returns Array of character positions where implicit boundaries occur
+*/
+function findImplicitBoundaries(prompt) {
+	const boundaries = [];
+	for (const header of IMPLICIT_BOUNDARY_HEADERS) {
+		const index = prompt.indexOf(header);
+		if (index !== -1) boundaries.push(index);
+	}
+	return boundaries.sort((a, b) => a - b);
+}
+/**
+* Split prompt by implicit boundaries into blocks.
+*
+* Creates blocks based on the position of section headers that mark natural boundaries.
+* The block before the first header is the base block.
+* Blocks between headers are intermediate blocks.
+* The block after the last header is the session context block.
+*
+* @param prompt - The system prompt to split
+* @param boundaryPositions - Sorted array of boundary positions
+* @returns Array of text blocks
+*/
+function splitByImplicitBoundaries(prompt, boundaryPositions) {
+	if (boundaryPositions.length === 0) return [prompt];
+	const blocks = [];
+	let previousPosition = 0;
+	for (const position of boundaryPositions) {
+		blocks.push(prompt.slice(previousPosition, position));
+		previousPosition = position;
+	}
+	blocks.push(prompt.slice(previousPosition));
+	return blocks;
+}
+/**
+* Detect if a prompt contains implicit cache boundaries.
+*
+* A prompt has implicit boundaries if it contains any of the known boundary headers.
+*
+* @param prompt - The system prompt to check
+* @returns true if implicit boundaries are detected
+*/
+function hasImplicitBoundaries(prompt) {
+	return IMPLICIT_BOUNDARY_HEADERS.some((header) => prompt.includes(header));
+}
+/**
+* Build a prompt plan from a rendered system prompt and provider cache capability.
+*
+* This function parses cache boundary markers from the system prompt and creates
+* a structured plan with named blocks, each with an appropriate cache scope.
+*
+* @param renderedSystemPrompt - The fully rendered system prompt (may contain `__CACHE_BOUNDARY__` markers)
+* @param providerCacheCapability - The provider's cache capability (for scope filtering)
+* @returns A prompt plan with cacheable blocks
+*
+* @example
+* ```ts
+* const prompt = `Base instructions
+* __CACHE_BOUNDARY__
+* Session context`;
+*
+* const capability = {
+*   strategy: 'explicit-block',
+*   supportedScopes: ['global', 'session'],
+* };
+*
+* const plan = buildPromptPlan(prompt, capability);
+* // {
+* //   blocks: [
+* //     { name: 'base', text: 'Base instructions\n', cacheScope: 'global' },
+* //     { name: 'sessionContext', text: 'Session context', cacheScope: 'session' },
+* //   ],
+* // }
+* ```
+*/
+/**
+* Normalize block text by handling newlines around cache boundaries.
+*
+* After splitting by `__CACHE_BOUNDARY__`, the pattern is typically:
+* - `[text]\n__CACHE_BOUNDARY__\n[next text]`
+*
+* Rules:
+* - First block: Keep as-is (preserves any trailing newline before marker)
+* - Middle blocks: Keep the newline that was between markers
+* - Last block: Trim leading newline (the one immediately after the last marker)
+*
+* @param text - The raw block text after splitting
+* @param index - The block index
+* @param totalBlocks - Total number of blocks
+* @returns Normalized block text
+*/
+function normalizeBlockText(text, index, totalBlocks) {
+	if (index === 0) return text;
+	if (index === totalBlocks - 1) {
+		if (text.startsWith("\n")) return text.slice(1);
+		return text;
+	}
+	return text;
+}
+function buildPromptPlan(renderedSystemPrompt, providerCacheCapability) {
+	const parts = renderedSystemPrompt.split(CACHE_BOUNDARY_MARKER);
+	if (parts.length > 1) return createPlanFromParts(parts, providerCacheCapability);
+	if (hasImplicitBoundaries(renderedSystemPrompt)) return createPlanFromParts(splitByImplicitBoundaries(renderedSystemPrompt, findImplicitBoundaries(renderedSystemPrompt)), providerCacheCapability);
+	return { blocks: [{
+		name: "base",
+		text: renderedSystemPrompt,
+		cacheScope: "none"
+	}] };
+}
+/**
+* Create a prompt plan from pre-split parts.
+*
+* This is shared logic for both explicit and implicit boundaries.
+*
+* @param parts - The pre-split text parts
+* @param providerCacheCapability - The provider's cache capability
+* @returns A prompt plan with cacheable blocks
+*/
+function createPlanFromParts(parts, providerCacheCapability) {
+	const blocks = parts.map((part, index) => {
+		const totalBlocks = parts.length;
+		const filteredScope = filterScopeByCapability(getDefaultScopeForPosition(index, totalBlocks), providerCacheCapability);
+		return {
+			name: getBlockNameForPart(part, index, totalBlocks),
+			text: normalizeBlockText(part, index, totalBlocks),
+			cacheScope: filteredScope
+		};
+	});
+	const maxCacheable = providerCacheCapability.maxCacheableBlocks;
+	if (maxCacheable !== void 0 && maxCacheable > 0) {
+		let cacheableCount = 0;
+		for (const block of blocks) if (block.cacheScope !== "none") {
+			cacheableCount++;
+			if (cacheableCount > maxCacheable) block.cacheScope = "none";
+		}
+	}
+	return { blocks };
+}
+/**
+* Get the block name for a given part, considering its content.
+*
+* This is a smarter version of getBlockName that looks at the content
+* to determine the appropriate name for the last block.
+*
+* @param part - The text content of this block
+* @param position - The block position (0-indexed)
+* @param totalBlocks - Total number of blocks
+* @returns The block name
+*/
+function getBlockNameForPart(part, position, totalBlocks) {
+	if (position === 0) return "base";
+	if (position === totalBlocks - 1) {
+		if (part.includes("# Skills")) return "sessionContext";
+		if (part.includes("# Project Information")) return "projectInstructions";
+		return "sessionContext";
+	}
+	const nameIndex = position - 1 + 1;
+	if (nameIndex < BLOCK_NAMES.length && nameIndex < BLOCK_NAMES.length - 1) {
+		const name = BLOCK_NAMES[nameIndex];
+		if (name !== void 0) return name;
+	}
+	return "sessionContext";
+}
+//#endregion
 //#region ../../packages/agent-core/src/agent/turn/kosong-llm.ts
 /**
 * Kosong-backed implementation of the loop `LLM` interface.
@@ -82939,7 +83428,7 @@ var KosongLLM = class {
 			systemPrompt: this.systemPrompt,
 			tools: params.tools
 		});
-		const result = await this.generate(effectiveProvider, this.systemPrompt, [...params.tools], [...params.messages], callbacks, generateOptions(params));
+		const result = await this.generate(effectiveProvider, this.systemPrompt, [...params.tools], [...params.messages], callbacks, generateOptions(params, this.systemPrompt, effectiveProvider));
 		if (params.onTextPart !== void 0 || params.onThinkPart !== void 0) {
 			for (const part of result.message.content) if (part.type === "text" && params.onTextPart !== void 0) await params.onTextPart(part);
 			else if (part.type === "think" && params.onThinkPart !== void 0) await params.onThinkPart(part);
@@ -82965,10 +83454,11 @@ var KosongLLM = class {
 		].includes(error.statusCode);
 	}
 };
-function generateOptions(params) {
+function generateOptions(params, systemPrompt, provider) {
+	const promptPlan = buildPromptPlan(systemPrompt, getProviderCacheCapability(provider));
 	const options = {
 		signal: params.signal,
-		cacheBreakpoints: ["__CACHE_BOUNDARY__"]
+		promptPlan
 	};
 	if (params.requestLogContext !== void 0) return {
 		...options,
@@ -82976,6 +83466,18 @@ function generateOptions(params) {
 	};
 	return options;
 }
+/**
+* Get the cache capability from a provider.
+*
+* Safely handles providers that don't implement getCapability or don't have cache.
+* Returns a default capability with 'none' strategy for non-caching providers.
+*/
+function getProviderCacheCapability(provider) {
+	if (typeof provider.getCapability !== "function") return { strategy: "none" };
+	const capability = provider.getCapability();
+	if (capability?.cache !== void 0) return capability.cache;
+	return { strategy: "none" };
+}
 function buildKosongCallbacks(params) {
 	const toolCallIdentities = /* @__PURE__ */ new Map();
 	const pendingIndexedToolCallDeltas = /* @__PURE__ */ new Map();
@@ -83222,6 +83724,7 @@ var TurnFlow = class {
 	agent;
 	steerBuffer = [];
 	turnId = -1;
+	_previousTurnMessageCount = 0;
 	activeTurn = null;
 	toolCallStartedAt = /* @__PURE__ */ new Map();
 	toolCallDupType = /* @__PURE__ */ new Map();
@@ -83266,6 +83769,7 @@ var TurnFlow = class {
 			return null;
 		}
 		this.turnId += 1;
+		this._previousTurnMessageCount = this.agent.context.messages.length;
 		this.currentStep = 0;
 		this.stepToolCallKeys.clear();
 		this.toolCallDupType.clear();
@@ -83502,7 +84006,10 @@ var TurnFlow = class {
 						generate: this.agent.generate,
 						completionBudgetConfig
 					}),
-					buildMessages: () => this.agent.context.messages,
+					buildMessages: () => {
+						const messages = this.agent.context.messages;
+						return applyCacheStaking(messages, { previousTurnMessageCount: this._previousTurnMessageCount });
+					},
 					dispatchEvent: this.buildDispatchEvent(turnId),
 					tools: this.agent.tools.loopTools,
 					log: this.agent.log,
@@ -83700,6 +84207,19 @@ var TurnFlow = class {
 		const failure = this.stepFailureByTurn.get(turnId);
 		return failure?.reason === "error" && failure.activeStep !== void 0;
 	}
+	restoreRecord(record) {
+		switch (record.type) {
+			case "turn.prompt":
+				if (this.activeTurn === "resuming") this.activeTurn = null;
+				this.restorePrompt();
+				break;
+			case "turn.steer":
+				if (this.activeTurn === "resuming") this.activeTurn = null;
+				this.restoreSteer(record.input, record.origin);
+				break;
+			case "turn.cancel": break;
+		}
+	}
 };
 function mapLoopEvent(event, turnId) {
 	switch (event.type) {
@@ -83936,10 +84456,12 @@ var UsageRecorder = class {
 		const byModel = this.byModelSnapshot();
 		const hasByModel = Object.keys(byModel).length > 0;
 		const currentTurn = this.currentTurn;
+		const total = hasByModel ? totalUsage(byModel) : void 0;
 		return {
 			byModel: hasByModel ? byModel : void 0,
-			total: hasByModel ? totalUsage(byModel) : void 0,
-			currentTurn: currentTurn === void 0 ? void 0 : copyUsage(currentTurn)
+			total,
+			currentTurn: currentTurn === void 0 ? void 0 : copyUsage(currentTurn),
+			cacheHitRate: total !== void 0 ? cacheHitRate(total) : void 0
 		};
 	}
 	status() {
@@ -83950,6 +84472,13 @@ var UsageRecorder = class {
 	byModelSnapshot() {
 		return Object.fromEntries(Object.entries(this.byModel).map(([model, usage]) => [model, copyUsage(usage)]));
 	}
+	restoreRecord(record) {
+		switch (record.type) {
+			case "usage.record":
+				this.record(record.model, record.usage, "session");
+				break;
+		}
+	}
 };
 function totalUsage(byModel) {
 	let total;
@@ -84012,6 +84541,15 @@ var Agent = class {
 			sessionDir: config.backgroundSessionDir
 		});
 		this.replayBuilder = new ReplayBuilder(this);
+		this.records.registerHandlers({
+			context: this.context,
+			config: this.config,
+			usage: this.usage,
+			turn: this.turn,
+			permission: this.permission,
+			tools: this.tools,
+			fullCompaction: this.fullCompaction
+		});
 	}
 	get generate() {
 		return async (provider, systemPrompt, tools, history, callbacks, options) => {
@@ -84032,7 +84570,7 @@ var Agent = class {
 	}
 	logLlmRequest(provider, systemPrompt, tools, history, options) {
 		const context = buildLlmRequestContext(options);
-		const configMetadata = buildLlmConfigMetadata(provider, this.config.modelAlias, systemPrompt, tools);
+		const configMetadata = buildLlmConfigMetadata(provider, this.config.modelAlias, systemPrompt, tools, options);
 		this.logLlmConfigIfChanged(context, configMetadata, buildLlmConfigSignature(configMetadata, systemPrompt, tools));
 		this.log.info("llm request", {
 			...context,
@@ -84062,11 +84600,15 @@ var Agent = class {
 		this.tools.setActiveTools(profile.tools);
 	}
 	async resume() {
-		const result = await this.records.replay();
-		await this.background.loadFromDisk();
-		await this.background.reconcile();
-		this.turn.finishResume();
-		return result;
+		try {
+			const result = await this.records.replay();
+			await this.background.loadFromDisk();
+			await this.background.reconcile();
+			this.turn.finishResume();
+			return result;
+		} catch (error) {
+			return { error: error instanceof Error ? error : new Error(String(error)) };
+		}
 	}
 	get rpcMethods() {
 		return {
@@ -84222,8 +84764,8 @@ function buildLlmRequestMetadata(systemPrompt, tools, history) {
 	if (partialMessageCount > 0) metadata.partialMessageCount = partialMessageCount;
 	return metadata;
 }
-function buildLlmConfigMetadata(provider, modelAlias, systemPrompt, tools) {
-	return {
+function buildLlmConfigMetadata(provider, modelAlias, systemPrompt, tools, options) {
+	const metadata = {
 		provider: provider.name,
 		model: provider.modelName,
 		modelAlias,
@@ -84231,6 +84773,30 @@ function buildLlmConfigMetadata(provider, modelAlias, systemPrompt, tools) {
 		systemPromptChars: systemPrompt.length,
 		toolCount: tools.length
 	};
+	const providerCacheStrategy = getProviderCacheStrategy(provider);
+	if (providerCacheStrategy !== void 0) metadata.providerCacheStrategy = providerCacheStrategy;
+	const promptPlan = options?.promptPlan;
+	if (promptPlan !== void 0 && promptPlan.blocks.length > 0) metadata.cacheBlockHashes = extractCacheBlockHashes(promptPlan);
+	return metadata;
+}
+/**
+* Get the cache strategy from a provider's capability.
+*
+* Safely handles providers that don't implement getCapability or don't have cache.
+*/
+function getProviderCacheStrategy(provider) {
+	if (typeof provider.getCapability !== "function") return;
+	return provider.getCapability()?.cache?.strategy;
+}
+/**
+* Extract cache block hashes from a PromptPlan.
+*
+* Returns a Record mapping block names to their SHA256 hashes.
+*/
+function extractCacheBlockHashes(promptPlan) {
+	const hashes = {};
+	for (const block of promptPlan.blocks) hashes[block.name] = fingerprint(block.text);
+	return hashes;
 }
 function buildLlmConfigSignature(metadata, systemPrompt, tools) {
 	const toolsForSignature = tools.map(({ name, description, parameters }) => ({
@@ -84276,6 +84842,7 @@ const OAuthRefSchema = z.object({
 	key: z.string().min(1)
 });
 const StringRecordSchema = z.record(z.string(), z.string());
+const ExtraBodySchema = z.record(z.string(), z.unknown());
 const ProviderConfigSchema = z.object({
 	type: ProviderTypeSchema,
 	apiKey: z.string().optional(),
@@ -84284,7 +84851,8 @@ const ProviderConfigSchema = z.object({
 	thinkingEffortKey: z.string().optional(),
 	oauth: OAuthRefSchema.optional(),
 	env: StringRecordSchema.optional(),
-	customHeaders: StringRecordSchema.optional()
+	customHeaders: StringRecordSchema.optional(),
+	extraBody: ExtraBodySchema.optional()
 });
 const ModelAliasSchema = z.object({
 	provider: z.string(),
@@ -88750,11 +89318,19 @@ var Session$1 = class {
 		const { agents } = await this.readMetadata();
 		this.agents.clear();
 		let warning;
+		const failedAgents = [];
 		const resumeTasks = Object.keys(agents).map(async (id) => {
 			const result = await this.ensureResumeAgentInstantiated(id, agents).resume();
 			if (result.warning !== void 0 && warning === void 0) warning = result.warning;
+			if (result.error !== void 0) failedAgents.push({
+				id,
+				error: result.error
+			});
 		});
 		await Promise.all(resumeTasks);
+		const mainAgentFailure = failedAgents.find(({ id }) => id === "main");
+		if (mainAgentFailure !== void 0) throw mainAgentFailure.error;
+		for (const { id, error } of failedAgents) this.log.warn(`Sub agent "${id}" failed to resume: ${error.message}`);
 		const resumeWarning = warning;
 		const main = this.agents.get("main");
 		const profile = DEFAULT_AGENT_PROFILES["agent"];
@@ -104097,13 +104673,17 @@ function toKosongProviderConfig(provider, model, byfRequestHeaders, maxOutputSiz
 				...byfRequestHeaders,
 				...provider.customHeaders
 			};
+			const generationKwargs = {
+				prompt_cache_key: promptCacheKey,
+				extra_body: provider.extraBody
+			};
 			if (Object.keys(defaultHeaders).length === 0) return {
 				type: "openai-completions",
 				model,
 				baseUrl: providerValue(provider.baseUrl, provider.env, "BYF_BASE_URL"),
 				reasoningKey,
 				thinkingEffortKey: provider.thinkingEffortKey,
-				generationKwargs: { prompt_cache_key: promptCacheKey },
+				generationKwargs,
 				apiKey: providerApiKey(provider)
 			};
 			return {
@@ -104112,7 +104692,7 @@ function toKosongProviderConfig(provider, model, byfRequestHeaders, maxOutputSiz
 				baseUrl: providerValue(provider.baseUrl, provider.env, "BYF_BASE_URL"),
 				reasoningKey,
 				thinkingEffortKey: provider.thinkingEffortKey,
-				generationKwargs: { prompt_cache_key: promptCacheKey },
+				generationKwargs,
 				defaultHeaders,
 				apiKey: providerApiKey(provider)
 			};
@@ -105949,6 +106529,9 @@ function capabilityToStrings(capability) {
 	if (capability.video_in) caps.push("video_in");
 	if (capability.audio_in) caps.push("audio_in");
 	if (capability.thinking) caps.push("thinking");
+	if (capability.thinking_effort) caps.push("thinking_effort");
+	if (capability.thinking_xhigh) caps.push("thinking_xhigh");
+	if (capability.thinking_max) caps.push("thinking_max");
 	if (capability.tool_use) caps.push("tool_use");
 	return caps.length > 0 ? caps : void 0;
 }
@@ -106004,6 +106587,40 @@ function applyCatalogProvider(config, options) {
 	config.defaultThinking = options.thinking;
 	return { defaultModel };
 }
+/**
+* Tests whether `candidate` is a prefix of `modelId` followed by either
+* end-of-string or a `-` separator. This matches `gpt-5.5` against
+* `gpt-5.5-2025-06-01` but not against `gpt-5.5-turbo` (different segment).
+*/
+function catalogIdMatchesModelId(candidate, modelId) {
+	if (modelId === candidate) return true;
+	if (modelId.startsWith(candidate) && modelId[candidate.length] === "-") return true;
+	return false;
+}
+/**
+* Searches all providers in the catalog for a model whose ID matches
+* `modelId` (prefix + separator boundary). Returns the first match.
+*/
+function findCatalogModel(catalog, modelId) {
+	for (const entry of Object.values(catalog)) {
+		const models = catalogProviderModels(entry);
+		for (const model of models) if (catalogIdMatchesModelId(model.id, modelId)) return model;
+	}
+}
+/**
+* Merges catalog metadata (priority) with provider-supplied values (fallback).
+* Catalog provides: capabilities, maxContextSize, maxOutputSize, reasoningKey.
+* Provider provides: displayName (user chose this provider, keep its naming).
+*/
+function enrichWithCatalog(providerModel, catalogModel) {
+	return {
+		maxContextSize: catalogModel.capability.max_context_tokens || providerModel.contextLength,
+		maxOutputSize: catalogModel.maxOutputSize,
+		capabilities: capabilityToStrings(catalogModel.capability),
+		displayName: providerModel.displayName,
+		reasoningKey: catalogModel.reasoningKey
+	};
+}
 //#endregion
 //#region ../../packages/oauth/src/utils.ts
 function isRecord$2(value) {
@@ -106363,6 +106980,27 @@ function validateOptions(opts) {
 	};
 }
 //#endregion
+//#region src/tui/utils/dead-terminal.ts
+/**
+* Detects errors that mean the controlling terminal (stdout/stderr pty) is
+* effectively gone — for example after the parent shell crashed, the tmux
+* server vanished, or an SSH connection dropped without delivering SIGHUP.
+*
+* Continuing to write to a dead terminal would re-fire the same error on every
+* render tick and pin a CPU core. Callers should respond by skipping any
+* cleanup that touches stdout/stderr and exiting immediately.
+*/
+const DEAD_TERMINAL_ERROR_CODES = new Set([
+	"EIO",
+	"EPIPE",
+	"ENOTCONN"
+]);
+function isDeadTerminalError(error) {
+	if (error === null || typeof error !== "object" || !("code" in error)) return false;
+	const code = error.code;
+	return code !== void 0 && DEAD_TERMINAL_ERROR_CODES.has(code);
+}
+//#endregion
 //#region src/cli/run-prompt.ts
 const PROMPT_UI_MODE = "print";
 const PROMPT_MAIN_AGENT_ID = "main";
@@ -106480,6 +107118,9 @@ function installHeadlessHandlers(session) {
 }
 function installPromptTerminationCleanup(promptProcess, cleanup) {
 	let terminating = false;
+	const emergencyExit = () => {
+		promptProcess.exit(129);
+	};
 	const exitAfterCleanup = async (signal) => {
 		if (terminating) return;
 		terminating = true;
@@ -106493,9 +107134,22 @@ function installPromptTerminationCleanup(promptProcess, cleanup) {
 	const onSigterm = () => exitAfterCleanup("SIGTERM");
 	promptProcess.once("SIGINT", onSigint);
 	promptProcess.once("SIGTERM", onSigterm);
+	let onSighup;
+	if (process.platform !== "win32") {
+		onSighup = () => emergencyExit();
+		process.prependListener("SIGHUP", onSighup);
+	}
+	const terminalErrorHandler = (error) => {
+		if (isDeadTerminalError(error)) emergencyExit();
+	};
+	process.stdout.on("error", terminalErrorHandler);
+	process.stderr.on("error", terminalErrorHandler);
 	return () => {
 		promptProcess.off("SIGINT", onSigint);
 		promptProcess.off("SIGTERM", onSigterm);
+		if (onSighup !== void 0) process.off("SIGHUP", onSighup);
+		process.stdout.off("error", terminalErrorHandler);
+		process.stderr.off("error", terminalErrorHandler);
 	};
 }
 function signalExitCode(signal) {
@@ -116119,27 +116773,6 @@ function hasDispose(value) {
 	return typeof value === "object" && value !== null && "dispose" in value && typeof value.dispose === "function";
 }
 //#endregion
-//#region src/tui/utils/dead-terminal.ts
-/**
-* Detects errors that mean the controlling terminal (stdout/stderr pty) is
-* effectively gone — for example after the parent shell crashed, the tmux
-* server vanished, or an SSH connection dropped without delivering SIGHUP.
-*
-* Continuing to write to a dead terminal would re-fire the same error on every
-* render tick and pin a CPU core. Callers should respond by skipping any
-* cleanup that touches stdout/stderr and exiting immediately.
-*/
-const DEAD_TERMINAL_ERROR_CODES = new Set([
-	"EIO",
-	"EPIPE",
-	"ENOTCONN"
-]);
-function isDeadTerminalError(error) {
-	if (error === null || typeof error !== "object" || !("code" in error)) return false;
-	const code = error.code;
-	return code !== void 0 && DEAD_TERMINAL_ERROR_CODES.has(code);
-}
-//#endregion
 //#region src/tui/utils/errors.ts
 function isAbortMessage(message) {
 	return message === "Aborted" || message.endsWith(": Aborted");
@@ -117558,7 +118191,6 @@ var LoginFlow = class {
 		const baseUrl = await promptTextInput(dialogHost, colors, {
 			title: "Base URL",
 			subtitle: "The OpenAI-compatible API endpoint",
-			initialValue: "https://api.openai.com/v1",
 			placeholder: "https://api.openai.com/v1"
 		});
 		if (baseUrl === void 0) return;
@@ -117585,20 +118217,29 @@ var LoginFlow = class {
 			this.deps.showStatus("No models found at this endpoint. Enter model ID manually.");
 			return this.handleManualModelEntry(name, baseUrl, apiKey);
 		}
+		const catalog = await this.fetchCatalogWithFallback();
+		const enriched = {};
 		const modelDict = {};
-		for (const m of models) modelDict[`${name}/${m.id}`] = {
-			provider: name,
-			model: m.id,
-			maxContextSize: m.contextLength,
-			capabilities: capabilitiesForModel(m),
-			displayName: m.displayName
-		};
+		for (const m of models) {
+			const aliasKey = `${name}/${m.id}`;
+			const enrichedData = catalog !== void 0 ? this.enrichModelFromCatalog(m, catalog) : void 0;
+			if (enrichedData !== void 0) enriched[aliasKey] = enrichedData;
+			modelDict[aliasKey] = {
+				provider: name,
+				model: m.id,
+				maxContextSize: enrichedData?.maxContextSize ?? m.contextLength,
+				capabilities: enrichedData?.capabilities ?? capabilitiesForModel(m),
+				displayName: m.displayName,
+				reasoningKey: enrichedData?.reasoningKey,
+				maxOutputSize: enrichedData?.maxOutputSize
+			};
+		}
 		const selection = await promptModelSelector(dialogHost, colors, modelDict);
 		if (selection === void 0) return;
 		const selectedId = selection.alias.split("/").slice(1).join("/");
 		const selectedModel = models.find((m) => m.id === selectedId);
 		if (selectedModel === void 0) return;
-		await this.applyConfig(name, baseUrl, apiKey, models, selectedModel, selection.thinkingEffort !== "off");
+		await this.applyConfig(name, baseUrl, apiKey, models, selectedModel, selection.thinkingEffort !== "off", enriched);
 		this.deps.track("login", {
 			provider: name,
 			model: selectedModel.id
@@ -117631,14 +118272,14 @@ var LoginFlow = class {
 			supportsImageIn: false,
 			supportsVideoIn: false
 		};
-		await this.applyConfig(name, baseUrl, apiKey, [manualModelInfo], manualModelInfo, false);
+		await this.applyConfig(name, baseUrl, apiKey, [manualModelInfo], manualModelInfo, false, {});
 		this.deps.track("login", {
 			provider: name,
 			model: manualModel
 		});
 		this.deps.showStatus(`Connected: ${name} · ${manualModel}`);
 	}
-	async applyConfig(name, baseUrl, apiKey, models, selectedModel, thinking) {
+	async applyConfig(name, baseUrl, apiKey, models, selectedModel, thinking, enriched) {
 		const config = await this.deps.getConfig();
 		this.deps.applyProviderConfig(config, {
 			name,
@@ -117648,6 +118289,13 @@ var LoginFlow = class {
 			selectedModel,
 			thinking
 		});
+		if (enriched !== void 0) for (const [key, patch] of Object.entries(enriched)) {
+			const existing = config.models?.[key];
+			if (existing !== void 0) config.models[key] = {
+				...existing,
+				...patch
+			};
+		}
 		await this.deps.setConfig({
 			providers: config.providers,
 			models: config.models,
@@ -117656,6 +118304,18 @@ var LoginFlow = class {
 		});
 		await this.deps.refreshConfigAfterLogin();
 	}
+	async fetchCatalogWithFallback() {
+		try {
+			return await fetchCatalog(DEFAULT_CATALOG_URL);
+		} catch {
+			return loadBuiltInCatalog(this.deps.builtInCatalogJson);
+		}
+	}
+	enrichModelFromCatalog(model, catalog) {
+		const catalogModel = findCatalogModel(catalog, model.id);
+		if (catalogModel === void 0) return void 0;
+		return enrichWithCatalog(model, catalogModel);
+	}
 };
 function capabilitiesForModel(m) {
 	const caps = [];
@@ -120747,7 +121407,8 @@ var ByfTui = class {
 			showLoginProgressSpinner: (label) => this.showLoginProgressSpinner(label),
 			track: (event, props) => {
 				this.track(event, props);
-			}
+			},
+			builtInCatalogJson: void 0
 		}).run();
 	}
 	async handleLogoutCommand(_args) {