@dianshuv/copilot-api 0.6.3 → 0.7.1

package/dist/main.mjs

@@ -190,10 +190,6 @@ const GITHUB_APP_SCOPES = ["read:user"].join(" ");
 
 //#endregion
 //#region src/lib/auto-truncate-common.ts
-/**
-* Common types and configuration for auto-truncate modules.
-* Shared between OpenAI and Anthropic format handlers.
-*/
 const DEFAULT_AUTO_TRUNCATE_CONFIG = {
 	safetyMarginPercent: 2,
 	maxRequestBodyBytes: Infinity,
@@ -234,6 +230,99 @@ function onTokenLimitExceeded(modelId, reportedLimit) {
 function getEffectiveTokenLimit(modelId) {
 	return dynamicTokenLimits.get(modelId) ?? null;
 }
+const LARGE_TOOL_RESULT_THRESHOLD = 1e4;
+const COMPRESSED_SUMMARY_LENGTH = 500;
+function getMessageBytes(msg) {
+	return JSON.stringify(msg).length;
+}
+function compressToolResultContent(content) {
+	if (content.length <= LARGE_TOOL_RESULT_THRESHOLD) return content;
+	const halfLen = Math.floor(COMPRESSED_SUMMARY_LENGTH / 2);
+	const start = content.slice(0, halfLen);
+	const end = content.slice(-halfLen);
+	return `${start}\n\n[... ${(content.length - COMPRESSED_SUMMARY_LENGTH).toLocaleString()} characters omitted for brevity ...]\n\n${end}`;
+}
+function calculateLimits(model, config, defaultContextWindow) {
+	const rawTokenLimit = getEffectiveTokenLimit(model.id) ?? model.capabilities?.limits?.max_context_window_tokens ?? model.capabilities?.limits?.max_prompt_tokens ?? defaultContextWindow;
+	return {
+		tokenLimit: Math.floor(rawTokenLimit * (1 - config.safetyMarginPercent / 100)),
+		byteLimit: getEffectiveByteLimitBytes()
+	};
+}
+function ensureStartsWithUser(messages, logTag) {
+	let startIndex = 0;
+	while (startIndex < messages.length && messages[startIndex].role !== "user") startIndex++;
+	if (startIndex > 0) consola.debug(`[AutoTruncate:${logTag}] Skipped ${startIndex} leading non-user messages`);
+	return messages.slice(startIndex);
+}
+function findOptimalPreserveIndex(params) {
+	const { messages, systemBytes, systemTokens, payloadOverhead, tokenLimit, byteLimit, estimateTokens } = params;
+	if (messages.length === 0) return 0;
+	const markerBytes = 200;
+	const availableTokens = tokenLimit - systemTokens - 50;
+	const availableBytes = byteLimit - payloadOverhead - systemBytes - markerBytes;
+	if (availableTokens <= 0 || availableBytes <= 0) return messages.length;
+	const n = messages.length;
+	const cumTokens = Array.from({ length: n + 1 }, () => 0);
+	const cumBytes = Array.from({ length: n + 1 }, () => 0);
+	for (let i = n - 1; i >= 0; i--) {
+		const msg = messages[i];
+		cumTokens[i] = cumTokens[i + 1] + estimateTokens(msg);
+		cumBytes[i] = cumBytes[i + 1] + getMessageBytes(msg) + 1;
+	}
+	let left = 0;
+	let right = n;
+	while (left < right) {
+		const mid = left + right >>> 1;
+		if (cumTokens[mid] <= availableTokens && cumBytes[mid] <= availableBytes) right = mid;
+		else left = mid + 1;
+	}
+	return left;
+}
+function generateRemovedMessagesSummary(removedMessages, getToolCallNames) {
+	const toolCalls = [];
+	let userMessageCount = 0;
+	let assistantMessageCount = 0;
+	for (const msg of removedMessages) {
+		if (msg.role === "user") userMessageCount++;
+		else if (msg.role === "assistant") assistantMessageCount++;
+		for (const name of getToolCallNames(msg)) toolCalls.push(name);
+	}
+	const parts = [];
+	if (userMessageCount > 0 || assistantMessageCount > 0) {
+		const breakdown = [];
+		if (userMessageCount > 0) breakdown.push(`${userMessageCount} user`);
+		if (assistantMessageCount > 0) breakdown.push(`${assistantMessageCount} assistant`);
+		parts.push(`Messages: ${breakdown.join(", ")}`);
+	}
+	if (toolCalls.length > 0) {
+		const uniqueTools = [...new Set(toolCalls)];
+		const displayTools = uniqueTools.length > 5 ? [...uniqueTools.slice(0, 5), `+${uniqueTools.length - 5} more`] : uniqueTools;
+		parts.push(`Tools used: ${displayTools.join(", ")}`);
+	}
+	return parts.join(". ");
+}
+function findCompressThreshold(messages, tokenLimit, byteLimit, preservePercent, estimateTokens) {
+	const n = messages.length;
+	const cumTokens = Array.from({ length: n + 1 }, () => 0);
+	const cumBytes = Array.from({ length: n + 1 }, () => 0);
+	for (let i = n - 1; i >= 0; i--) {
+		const msg = messages[i];
+		cumTokens[i] = cumTokens[i + 1] + estimateTokens(msg);
+		cumBytes[i] = cumBytes[i + 1] + getMessageBytes(msg) + 1;
+	}
+	const preserveTokenLimit = Math.floor(tokenLimit * preservePercent);
+	const preserveByteLimit = Math.floor(byteLimit * preservePercent);
+	let thresholdIndex = n;
+	for (let i = n - 1; i >= 0; i--) {
+		if (cumTokens[i] > preserveTokenLimit || cumBytes[i] > preserveByteLimit) {
+			thresholdIndex = i + 1;
+			break;
+		}
+		thresholdIndex = i;
+	}
+	return thresholdIndex;
+}
 
 //#endregion
 //#region src/lib/error.ts
@@ -431,6 +520,9 @@ const sleep = (ms) => new Promise((resolve) => {
 	setTimeout(resolve, ms);
 });
 const isNullish = (value) => value === null || value === void 0;
+function findModelById(modelId) {
+	return state.models?.data.find((m) => m.id === modelId);
+}
 async function cacheModels() {
 	state.models = await getModels();
 }
@@ -1121,7 +1213,7 @@ const patchClaude = defineCommand({
 
 //#endregion
 //#region package.json
-var version = "0.6.3";
+var version = "0.7.1";
 
 //#endregion
 //#region src/lib/adaptive-rate-limiter.ts
@@ -1729,12 +1821,14 @@ const historyState = {
 	maxEntries: 1e3,
 	sessionTimeoutMs: 1800 * 1e3
 };
+const entryIndex = /* @__PURE__ */ new Map();
 function initHistory(enabled, maxEntries) {
 	historyState.enabled = enabled;
 	historyState.maxEntries = maxEntries;
 	historyState.entries = [];
 	historyState.sessions = /* @__PURE__ */ new Map();
 	historyState.currentSessionId = enabled ? generateId$1() : "";
+	entryIndex.clear();
 }
 function isHistoryEnabled() {
 	return historyState.enabled;
@@ -1783,6 +1877,7 @@ function recordRequest(endpoint, request) {
 		}
 	};
 	historyState.entries.push(entry);
+	entryIndex.set(entry.id, entry);
 	session.requestCount++;
 	if (!session.models.includes(request.model)) session.models.push(request.model);
 	if (request.tools && request.tools.length > 0) {
@@ -1792,6 +1887,7 @@ function recordRequest(endpoint, request) {
 	while (historyState.maxEntries > 0 && historyState.entries.length > historyState.maxEntries) {
 		const removed = historyState.entries.shift();
 		if (removed) {
+			entryIndex.delete(removed.id);
 			if (historyState.entries.filter((e) => e.sessionId === removed.sessionId).length === 0) historyState.sessions.delete(removed.sessionId);
 		}
 	}
@@ -1806,7 +1902,7 @@ function recordRequest(endpoint, request) {
 }
 function recordResponse(id, response, durationMs) {
 	if (!historyState.enabled || !id) return;
-	const entry = historyState.entries.find((e) => e.id === id);
+	const entry = entryIndex.get(id);
 	if (entry) {
 		entry.response = response;
 		entry.durationMs = durationMs;
@@ -1885,7 +1981,7 @@ function getHistory(options = {}) {
 	};
 }
 function getEntry(id) {
-	return historyState.entries.find((e) => e.id === id);
+	return entryIndex.get(id);
 }
 function getSessions() {
 	const sessions = Array.from(historyState.sessions.values()).sort((a, b) => b.lastActivity - a.lastActivity);
@@ -1904,11 +2000,14 @@ function clearHistory() {
 	historyState.entries = [];
 	historyState.sessions = /* @__PURE__ */ new Map();
 	historyState.currentSessionId = generateId$1();
+	entryIndex.clear();
 	notifyHistoryCleared();
 }
 function deleteSession(sessionId) {
 	if (!historyState.sessions.has(sessionId)) return false;
+	const removedEntries = historyState.entries.filter((e) => e.sessionId === sessionId);
 	historyState.entries = historyState.entries.filter((e) => e.sessionId !== sessionId);
+	for (const e of removedEntries) entryIndex.delete(e.id);
 	historyState.sessions.delete(sessionId);
 	if (historyState.currentSessionId === sessionId) historyState.currentSessionId = generateId$1();
 	notifySessionDeleted(sessionId);
@@ -2010,6 +2109,7 @@ function evictOldestEntries(count) {
 	if (count <= 0) return 0;
 	const actual = Math.min(count, historyState.entries.length);
 	const removed = historyState.entries.splice(0, actual);
+	for (const e of removed) entryIndex.delete(e.id);
 	for (const entry of removed) if (!historyState.entries.some((e) => e.sessionId === entry.sessionId)) historyState.sessions.delete(entry.sessionId);
 	return actual;
 }
@@ -2797,6 +2897,147 @@ const awaitApproval = async () => {
 	if (!await consola.prompt(`Accept incoming request?`, { type: "confirm" })) throw new HTTPError("Request rejected", 403, JSON.stringify({ message: "Request rejected" }));
 };
 
+//#endregion
+//#region src/lib/message-sanitizer.ts
+const startPattern = /^\s*<system-reminder>[\s\S]*?<\/system-reminder>\n*/;
+const endPatternWithNewline = /\n+<system-reminder>[\s\S]*?<\/system-reminder>\s*$/;
+const endPatternOnly = /^\s*<system-reminder>[\s\S]*?<\/system-reminder>\s*$/;
+function removeSystemReminderTags(text) {
+	let result = text;
+	let prev;
+	do {
+		prev = result;
+		result = result.replace(startPattern, "");
+	} while (result !== prev);
+	do {
+		prev = result;
+		result = result.replace(endPatternWithNewline, "");
+	} while (result !== prev);
+	result = result.replace(endPatternOnly, "");
+	return result;
+}
+
+//#endregion
+//#region src/lib/repetition-detector.ts
+/**
+* Stream repetition detector.
+*
+* Uses the KMP failure function (prefix function) to detect repeated patterns
+* in streaming text output. When a model gets stuck in a repetitive loop,
+* it wastes tokens producing the same content over and over. This detector
+* identifies such loops early so the caller can take action (log warning,
+* abort stream, etc.).
+*
+* The algorithm works by maintaining a sliding buffer of recent text and
+* computing the longest proper prefix that is also a suffix — if this
+* length exceeds `(text.length - period) >= minRepetitions * period`,
+* it means a pattern of length `period` has repeated enough times.
+*/
+const DEFAULT_CONFIG = {
+	minPatternLength: 10,
+	minRepetitions: 3,
+	maxBufferSize: 5e3
+};
+var RepetitionDetector = class {
+	buffer = "";
+	config;
+	detected = false;
+	constructor(config) {
+		this.config = {
+			...DEFAULT_CONFIG,
+			...config
+		};
+	}
+	/**
+	* Feed a text chunk into the detector.
+	* Returns `true` if repetition has been detected (now or previously).
+	* Once detected, subsequent calls return `true` without further analysis.
+	*/
+	feed(text) {
+		if (this.detected) return true;
+		if (!text) return false;
+		this.buffer += text;
+		if (this.buffer.length > this.config.maxBufferSize) this.buffer = this.buffer.slice(-this.config.maxBufferSize);
+		const minRequired = this.config.minPatternLength * this.config.minRepetitions;
+		if (this.buffer.length < minRequired) return false;
+		this.detected = detectRepetition(this.buffer, this.config.minPatternLength, this.config.minRepetitions);
+		return this.detected;
+	}
+	/** Reset detector state for a new stream */
+	reset() {
+		this.buffer = "";
+		this.detected = false;
+	}
+	/** Whether repetition has been detected */
+	get isDetected() {
+		return this.detected;
+	}
+};
+/**
+* Detect if the tail of `text` contains a repeating pattern.
+*
+* Uses the KMP prefix function: for a string S, the prefix function π[i]
+* gives the length of the longest proper prefix of S[0..i] that is also
+* a suffix. If π[n-1] ≥ (n - period) where period = n - π[n-1], then
+* the string is composed of a repeating unit of length `period`.
+*
+* We check the suffix of the buffer (last `checkLength` chars) to detect
+* if a pattern of at least `minPatternLength` chars repeats at least
+* `minRepetitions` times.
+*/
+function detectRepetition(text, minPatternLength, minRepetitions) {
+	const minWindow = minPatternLength * minRepetitions;
+	const maxWindow = Math.min(text.length, 2e3);
+	const windowSizes = [
+		minWindow,
+		Math.floor(maxWindow * .5),
+		maxWindow
+	].filter((w) => w >= minWindow && w <= text.length);
+	for (const windowSize of windowSizes) {
+		const window = text.slice(-windowSize);
+		const period = findRepeatingPeriod(window);
+		if (period >= minPatternLength) {
+			if (Math.floor(window.length / period) >= minRepetitions) return true;
+		}
+	}
+	return false;
+}
+/**
+* Find the shortest repeating period in a string using KMP prefix function.
+* Returns the period length, or the string length if no repetition found.
+*/
+function findRepeatingPeriod(s) {
+	const n = s.length;
+	if (n === 0) return 0;
+	const pi = new Int32Array(n);
+	for (let i = 1; i < n; i++) {
+		let j = pi[i - 1] ?? 0;
+		while (j > 0 && s[i] !== s[j]) j = pi[j - 1] ?? 0;
+		if (s[i] === s[j]) j++;
+		pi[i] = j;
+	}
+	const period = n - pi[n - 1];
+	if (period < n && n % period === 0) return period;
+	if (period < n && pi[n - 1] >= period) return period;
+	return n;
+}
+/**
+* Create a repetition detector callback for use in stream processing.
+* Returns a function that accepts text deltas and logs a warning on first detection.
+*/
+function createStreamRepetitionChecker(label, config) {
+	const detector = new RepetitionDetector(config);
+	let warned = false;
+	return (textDelta) => {
+		const isRepetitive = detector.feed(textDelta);
+		if (isRepetitive && !warned) {
+			warned = true;
+			consola.warn(`[RepetitionDetector] ${label}: Repetitive output detected in stream`);
+		}
+		return isRepetitive;
+	};
+}
+
 //#endregion
 //#region src/lib/tokenizer.ts
 const ENCODING_MAP = {
@@ -3009,6 +3250,32 @@ const getTokenCount = async (payload, model) => {
 	};
 };
 
+//#endregion
+//#region src/services/copilot/create-chat-completions.ts
+const createChatCompletions = async (payload, options) => {
+	if (!state.copilotToken) throw new Error("Copilot token not found");
+	const enableVision = payload.messages.some((x) => typeof x.content !== "string" && x.content?.some((x) => x.type === "image_url"));
+	const isAgentCall = payload.messages.some((msg) => ["assistant", "tool"].includes(msg.role));
+	const headers = {
+		...copilotHeaders(state, {
+			vision: enableVision,
+			intent: isAgentCall ? "conversation-agent" : "conversation-panel"
+		}),
+		"X-Initiator": options?.initiator ?? (isAgentCall ? "agent" : "user")
+	};
+	const response = await fetch(`${copilotBaseUrl(state)}/chat/completions`, {
+		method: "POST",
+		headers,
+		body: JSON.stringify(payload)
+	});
+	if (!response.ok) {
+		consola.error("Failed to create chat completions", response);
+		throw await HTTPError.fromResponse("Failed to create chat completions", response, payload.model);
+	}
+	if (payload.stream) return events(response);
+	return await response.json();
+};
+
 //#endregion
 //#region src/lib/auto-truncate-openai.ts
 /**
@@ -3023,13 +3290,6 @@ const getTokenCount = async (payload, model) => {
 * - Dynamic byte limit adjustment on 413 errors
 * - Optional smart compression of old tool_result content
 */
-function calculateLimits$1(model, config) {
-	const rawTokenLimit = getEffectiveTokenLimit(model.id) ?? model.capabilities?.limits?.max_context_window_tokens ?? model.capabilities?.limits?.max_prompt_tokens ?? 128e3;
-	return {
-		tokenLimit: Math.floor(rawTokenLimit * (1 - config.safetyMarginPercent / 100)),
-		byteLimit: getEffectiveByteLimitBytes()
-	};
-}
 /** Estimate tokens for a single message (fast approximation) */
 function estimateMessageTokens$1(msg) {
 	let charCount = 0;
@@ -3041,10 +3301,6 @@ function estimateMessageTokens$1(msg) {
 	if (msg.tool_calls) charCount += JSON.stringify(msg.tool_calls).length;
 	return Math.ceil(charCount / 4) + 10;
 }
-/** Get byte size of a message */
-function getMessageBytes$1(msg) {
-	return JSON.stringify(msg).length;
-}
 /** Extract system/developer messages from the beginning */
 function extractSystemMessages(messages) {
 	let splitIndex = 0;
@@ -3116,28 +3372,6 @@ function filterOrphanedToolUse$1(messages) {
 	if (removedCount > 0) consola.debug(`[AutoTruncate:OpenAI] Filtered ${removedCount} orphaned tool_use`);
 	return result;
 }
-/** Ensure messages start with a user message */
-function ensureStartsWithUser$1(messages) {
-	let startIndex = 0;
-	while (startIndex < messages.length && messages[startIndex].role !== "user") startIndex++;
-	if (startIndex > 0) consola.debug(`[AutoTruncate:OpenAI] Skipped ${startIndex} leading non-user messages`);
-	return messages.slice(startIndex);
-}
-/** Threshold for large tool message content (bytes) */
-const LARGE_TOOL_RESULT_THRESHOLD$1 = 1e4;
-/** Maximum length for compressed tool_result summary */
-const COMPRESSED_SUMMARY_LENGTH$1 = 500;
-/**
-* Compress a large tool message content to a summary.
-* Keeps the first and last portions with a note about truncation.
-*/
-function compressToolResultContent$1(content) {
-	if (content.length <= LARGE_TOOL_RESULT_THRESHOLD$1) return content;
-	const halfLen = Math.floor(COMPRESSED_SUMMARY_LENGTH$1 / 2);
-	const start = content.slice(0, halfLen);
-	const end = content.slice(-halfLen);
-	return `${start}\n\n[... ${(content.length - COMPRESSED_SUMMARY_LENGTH$1).toLocaleString()} characters omitted for brevity ...]\n\n${end}`;
-}
 /**
 * Smart compression strategy for OpenAI format:
 * 1. Calculate tokens/bytes from the end until reaching preservePercent of limit
@@ -3147,37 +3381,20 @@ function compressToolResultContent$1(content) {
 * @param preservePercent - Percentage of context to preserve uncompressed (0.0-1.0)
 */
 function smartCompressToolResults$1(messages, tokenLimit, byteLimit, preservePercent) {
-	const n = messages.length;
-	const cumTokens = Array.from({ length: n + 1 }, () => 0);
-	const cumBytes = Array.from({ length: n + 1 }, () => 0);
-	for (let i = n - 1; i >= 0; i--) {
-		const msg = messages[i];
-		cumTokens[i] = cumTokens[i + 1] + estimateMessageTokens$1(msg);
-		cumBytes[i] = cumBytes[i + 1] + getMessageBytes$1(msg) + 1;
-	}
-	const preserveTokenLimit = Math.floor(tokenLimit * preservePercent);
-	const preserveByteLimit = Math.floor(byteLimit * preservePercent);
-	let thresholdIndex = n;
-	for (let i = n - 1; i >= 0; i--) {
-		if (cumTokens[i] > preserveTokenLimit || cumBytes[i] > preserveByteLimit) {
-			thresholdIndex = i + 1;
-			break;
-		}
-		thresholdIndex = i;
-	}
-	if (thresholdIndex >= n) return {
+	const thresholdIndex = findCompressThreshold(messages, tokenLimit, byteLimit, preservePercent, estimateMessageTokens$1);
+	if (thresholdIndex >= messages.length) return {
 		messages,
 		compressedCount: 0,
-		compressThresholdIndex: n
+		compressThresholdIndex: messages.length
 	};
 	const result = [];
 	let compressedCount = 0;
 	for (const [i, msg] of messages.entries()) {
-		if (i < thresholdIndex && msg.role === "tool" && typeof msg.content === "string" && msg.content.length > LARGE_TOOL_RESULT_THRESHOLD$1) {
+		if (i < thresholdIndex && msg.role === "tool" && typeof msg.content === "string" && msg.content.length > LARGE_TOOL_RESULT_THRESHOLD) {
 			compressedCount++;
 			result.push({
 				...msg,
-				content: compressToolResultContent$1(msg.content)
+				content: compressToolResultContent(msg.content)
 			});
 			continue;
 		}
@@ -3190,42 +3407,13 @@ function smartCompressToolResults$1(messages, tokenLimit, byteLimit, preservePer
 	};
 }
 /**
-* Find the optimal index from which to preserve messages.
-* Uses binary search with pre-calculated cumulative sums.
-* Returns the smallest index where the preserved portion fits within limits.
-*/
-function findOptimalPreserveIndex$1(params) {
-	const { messages, systemBytes, systemTokens, payloadOverhead, tokenLimit, byteLimit } = params;
-	if (messages.length === 0) return 0;
-	const markerBytes = 200;
-	const availableTokens = tokenLimit - systemTokens - 50;
-	const availableBytes = byteLimit - payloadOverhead - systemBytes - markerBytes;
-	if (availableTokens <= 0 || availableBytes <= 0) return messages.length;
-	const n = messages.length;
-	const cumTokens = Array.from({ length: n + 1 }, () => 0);
-	const cumBytes = Array.from({ length: n + 1 }, () => 0);
-	for (let i = n - 1; i >= 0; i--) {
-		const msg = messages[i];
-		cumTokens[i] = cumTokens[i + 1] + estimateMessageTokens$1(msg);
-		cumBytes[i] = cumBytes[i + 1] + getMessageBytes$1(msg) + 1;
-	}
-	let left = 0;
-	let right = n;
-	while (left < right) {
-		const mid = left + right >>> 1;
-		if (cumTokens[mid] <= availableTokens && cumBytes[mid] <= availableBytes) right = mid;
-		else left = mid + 1;
-	}
-	return left;
-}
-/**
 * Check if payload needs compaction based on model limits or byte size.
 */
 async function checkNeedsCompactionOpenAI(payload, model, config = {}) {
-	const { tokenLimit, byteLimit } = calculateLimits$1(model, {
+	const { tokenLimit, byteLimit } = calculateLimits(model, {
 		...DEFAULT_AUTO_TRUNCATE_CONFIG,
 		...config
-	});
+	}, 128e3);
 	const currentTokens = (await getTokenCount(payload, model)).input;
 	const currentBytes = JSON.stringify(payload).length;
 	const exceedsTokens = currentTokens > tokenLimit;
@@ -3244,35 +3432,6 @@ async function checkNeedsCompactionOpenAI(payload, model, config = {}) {
 	};
 }
 /**
-* Generate a summary of removed messages for context.
-* Extracts key information like tool calls and topics.
-*/
-function generateRemovedMessagesSummary$1(removedMessages) {
-	const toolCalls = [];
-	let userMessageCount = 0;
-	let assistantMessageCount = 0;
-	for (const msg of removedMessages) {
-		if (msg.role === "user") userMessageCount++;
-		else if (msg.role === "assistant") assistantMessageCount++;
-		if (msg.tool_calls) {
-			for (const tc of msg.tool_calls) if (tc.function.name) toolCalls.push(tc.function.name);
-		}
-	}
-	const parts = [];
-	if (userMessageCount > 0 || assistantMessageCount > 0) {
-		const breakdown = [];
-		if (userMessageCount > 0) breakdown.push(`${userMessageCount} user`);
-		if (assistantMessageCount > 0) breakdown.push(`${assistantMessageCount} assistant`);
-		parts.push(`Messages: ${breakdown.join(", ")}`);
-	}
-	if (toolCalls.length > 0) {
-		const uniqueTools = [...new Set(toolCalls)];
-		const displayTools = uniqueTools.length > 5 ? [...uniqueTools.slice(0, 5), `+${uniqueTools.length - 5} more`] : uniqueTools;
-		parts.push(`Tools used: ${displayTools.join(", ")}`);
-	}
-	return parts.join(". ");
-}
-/**
 * Add a compression notice to the system message.
 * Informs the model that some tool content has been compressed.
 */
@@ -3326,7 +3485,7 @@ async function autoTruncateOpenAI(payload, model, config = {}) {
 		...DEFAULT_AUTO_TRUNCATE_CONFIG,
 		...config
 	};
-	const { tokenLimit, byteLimit } = calculateLimits$1(model, cfg);
+	const { tokenLimit, byteLimit } = calculateLimits(model, cfg, 128e3);
 	const originalBytes = JSON.stringify(payload).length;
 	const originalTokens = (await getTokenCount(payload, model)).input;
 	if (originalTokens <= tokenLimit && originalBytes <= byteLimit) return {
@@ -3371,16 +3530,17 @@ async function autoTruncateOpenAI(payload, model, config = {}) {
 		...payload,
 		messages: workingMessages
 	}).length - messagesJson.length;
-	const systemBytes = systemMessages.reduce((sum, m) => sum + getMessageBytes$1(m) + 1, 0);
+	const systemBytes = systemMessages.reduce((sum, m) => sum + getMessageBytes(m) + 1, 0);
 	const systemTokens = systemMessages.reduce((sum, m) => sum + estimateMessageTokens$1(m), 0);
 	consola.debug(`[AutoTruncate:OpenAI] overhead=${Math.round(payloadOverhead / 1024)}KB, system=${systemMessages.length} msgs (${Math.round(systemBytes / 1024)}KB)`);
-	const preserveIndex = findOptimalPreserveIndex$1({
+	const preserveIndex = findOptimalPreserveIndex({
 		messages: conversationMessages,
 		systemBytes,
 		systemTokens,
 		payloadOverhead,
 		tokenLimit,
-		byteLimit
+		byteLimit,
+		estimateTokens: estimateMessageTokens$1
 	});
 	if (preserveIndex === 0) {
 		consola.warn("[AutoTruncate:OpenAI] Cannot truncate, system messages too large");
@@ -3392,246 +3552,70 @@ async function autoTruncateOpenAI(payload, model, config = {}) {
 			removedMessageCount: 0
 		};
 	}
-	if (preserveIndex >= conversationMessages.length) {
-		consola.warn("[AutoTruncate:OpenAI] Would need to remove all messages");
-		return {
-			payload,
-			wasCompacted: false,
-			originalTokens,
-			compactedTokens: originalTokens,
-			removedMessageCount: 0
-		};
-	}
-	let preserved = conversationMessages.slice(preserveIndex);
-	preserved = filterOrphanedToolResults$1(preserved);
-	preserved = filterOrphanedToolUse$1(preserved);
-	preserved = ensureStartsWithUser$1(preserved);
-	preserved = filterOrphanedToolResults$1(preserved);
-	preserved = filterOrphanedToolUse$1(preserved);
-	if (preserved.length === 0) {
-		consola.warn("[AutoTruncate:OpenAI] All messages filtered out after cleanup");
-		return {
-			payload,
-			wasCompacted: false,
-			originalTokens,
-			compactedTokens: originalTokens,
-			removedMessageCount: 0
-		};
-	}
-	const removedMessages = conversationMessages.slice(0, preserveIndex);
-	const removedCount = conversationMessages.length - preserved.length;
-	const summary = generateRemovedMessagesSummary$1(removedMessages);
-	let newSystemMessages = systemMessages;
-	let newMessages = preserved;
-	if (systemMessages.length > 0) {
-		const truncationContext = createTruncationSystemContext$1(removedCount, compressedCount, summary);
-		const lastSystemIdx = systemMessages.length - 1;
-		const lastSystem = systemMessages[lastSystemIdx];
-		const updatedSystem = {
-			...lastSystem,
-			content: typeof lastSystem.content === "string" ? lastSystem.content + truncationContext : lastSystem.content
-		};
-		newSystemMessages = [...systemMessages.slice(0, lastSystemIdx), updatedSystem];
-	} else newMessages = [createTruncationMarker$2(removedCount, compressedCount, summary), ...preserved];
-	const newPayload = {
-		...payload,
-		messages: [...newSystemMessages, ...newMessages]
-	};
-	const newBytes = JSON.stringify(newPayload).length;
-	const newTokenCount = await getTokenCount(newPayload, model);
-	let reason = "tokens";
-	if (exceedsTokens && exceedsBytes) reason = "tokens+size";
-	else if (exceedsBytes) reason = "size";
-	const actions = [];
-	if (removedCount > 0) actions.push(`removed ${removedCount} msgs`);
-	if (compressedCount > 0) actions.push(`compressed ${compressedCount} tool_results`);
-	const actionInfo = actions.length > 0 ? ` (${actions.join(", ")})` : "";
-	consola.info(`[AutoTruncate:OpenAI] ${reason}: ${originalTokens}→${newTokenCount.input} tokens, ${Math.round(originalBytes / 1024)}→${Math.round(newBytes / 1024)}KB${actionInfo}`);
-	if (newBytes > byteLimit) consola.warn(`[AutoTruncate:OpenAI] Result still over byte limit (${Math.round(newBytes / 1024)}KB > ${Math.round(byteLimit / 1024)}KB)`);
-	return {
-		payload: newPayload,
-		wasCompacted: true,
-		originalTokens,
-		compactedTokens: newTokenCount.input,
-		removedMessageCount: removedCount
-	};
-}
-/**
-* Create a marker to prepend to responses indicating auto-truncation occurred.
-*/
-function createTruncationResponseMarkerOpenAI(result) {
-	if (!result.wasCompacted) return "";
-	const reduction = result.originalTokens - result.compactedTokens;
-	const percentage = Math.round(reduction / result.originalTokens * 100);
-	return `\n\n---\n[Auto-truncated: ${result.removedMessageCount} messages removed, ${result.originalTokens} → ${result.compactedTokens} tokens (${percentage}% reduction)]`;
-}
-
-//#endregion
-//#region src/lib/message-sanitizer.ts
-const startPattern = /^\s*<system-reminder>[\s\S]*?<\/system-reminder>\n*/;
-const endPatternWithNewline = /\n+<system-reminder>[\s\S]*?<\/system-reminder>\s*$/;
-const endPatternOnly = /^\s*<system-reminder>[\s\S]*?<\/system-reminder>\s*$/;
-function removeSystemReminderTags(text) {
-	let result = text;
-	let prev;
-	do {
-		prev = result;
-		result = result.replace(startPattern, "");
-	} while (result !== prev);
-	do {
-		prev = result;
-		result = result.replace(endPatternWithNewline, "");
-	} while (result !== prev);
-	result = result.replace(endPatternOnly, "");
-	return result;
-}
-
-//#endregion
-//#region src/lib/repetition-detector.ts
-/**
-* Stream repetition detector.
-*
-* Uses the KMP failure function (prefix function) to detect repeated patterns
-* in streaming text output. When a model gets stuck in a repetitive loop,
-* it wastes tokens producing the same content over and over. This detector
-* identifies such loops early so the caller can take action (log warning,
-* abort stream, etc.).
-*
-* The algorithm works by maintaining a sliding buffer of recent text and
-* computing the longest proper prefix that is also a suffix — if this
-* length exceeds `(text.length - period) >= minRepetitions * period`,
-* it means a pattern of length `period` has repeated enough times.
-*/
-const DEFAULT_CONFIG = {
-	minPatternLength: 10,
-	minRepetitions: 3,
-	maxBufferSize: 5e3
-};
-var RepetitionDetector = class {
-	buffer = "";
-	config;
-	detected = false;
-	constructor(config) {
-		this.config = {
-			...DEFAULT_CONFIG,
-			...config
-		};
-	}
-	/**
-	* Feed a text chunk into the detector.
-	* Returns `true` if repetition has been detected (now or previously).
-	* Once detected, subsequent calls return `true` without further analysis.
-	*/
-	feed(text) {
-		if (this.detected) return true;
-		if (!text) return false;
-		this.buffer += text;
-		if (this.buffer.length > this.config.maxBufferSize) this.buffer = this.buffer.slice(-this.config.maxBufferSize);
-		const minRequired = this.config.minPatternLength * this.config.minRepetitions;
-		if (this.buffer.length < minRequired) return false;
-		this.detected = detectRepetition(this.buffer, this.config.minPatternLength, this.config.minRepetitions);
-		return this.detected;
-	}
-	/** Reset detector state for a new stream */
-	reset() {
-		this.buffer = "";
-		this.detected = false;
-	}
-	/** Whether repetition has been detected */
-	get isDetected() {
-		return this.detected;
-	}
-};
-/**
-* Detect if the tail of `text` contains a repeating pattern.
-*
-* Uses the KMP prefix function: for a string S, the prefix function π[i]
-* gives the length of the longest proper prefix of S[0..i] that is also
-* a suffix. If π[n-1] ≥ (n - period) where period = n - π[n-1], then
-* the string is composed of a repeating unit of length `period`.
-*
-* We check the suffix of the buffer (last `checkLength` chars) to detect
-* if a pattern of at least `minPatternLength` chars repeats at least
-* `minRepetitions` times.
-*/
-function detectRepetition(text, minPatternLength, minRepetitions) {
-	const minWindow = minPatternLength * minRepetitions;
-	const maxWindow = Math.min(text.length, 2e3);
-	const windowSizes = [
-		minWindow,
-		Math.floor(maxWindow * .5),
-		maxWindow
-	].filter((w) => w >= minWindow && w <= text.length);
-	for (const windowSize of windowSizes) {
-		const window = text.slice(-windowSize);
-		const period = findRepeatingPeriod(window);
-		if (period >= minPatternLength) {
-			if (Math.floor(window.length / period) >= minRepetitions) return true;
-		}
-	}
-	return false;
-}
-/**
-* Find the shortest repeating period in a string using KMP prefix function.
-* Returns the period length, or the string length if no repetition found.
-*/
-function findRepeatingPeriod(s) {
-	const n = s.length;
-	if (n === 0) return 0;
-	const pi = new Int32Array(n);
-	for (let i = 1; i < n; i++) {
-		let j = pi[i - 1] ?? 0;
-		while (j > 0 && s[i] !== s[j]) j = pi[j - 1] ?? 0;
-		if (s[i] === s[j]) j++;
-		pi[i] = j;
-	}
-	const period = n - pi[n - 1];
-	if (period < n && n % period === 0) return period;
-	if (period < n && pi[n - 1] >= period) return period;
-	return n;
-}
-/**
-* Create a repetition detector callback for use in stream processing.
-* Returns a function that accepts text deltas and logs a warning on first detection.
-*/
-function createStreamRepetitionChecker(label, config) {
-	const detector = new RepetitionDetector(config);
-	let warned = false;
-	return (textDelta) => {
-		const isRepetitive = detector.feed(textDelta);
-		if (isRepetitive && !warned) {
-			warned = true;
-			consola.warn(`[RepetitionDetector] ${label}: Repetitive output detected in stream`);
-		}
-		return isRepetitive;
+	if (preserveIndex >= conversationMessages.length) {
+		consola.warn("[AutoTruncate:OpenAI] Would need to remove all messages");
+		return {
+			payload,
+			wasCompacted: false,
+			originalTokens,
+			compactedTokens: originalTokens,
+			removedMessageCount: 0
+		};
+	}
+	let preserved = conversationMessages.slice(preserveIndex);
+	preserved = filterOrphanedToolResults$1(preserved);
+	preserved = filterOrphanedToolUse$1(preserved);
+	preserved = ensureStartsWithUser(preserved, "OpenAI");
+	preserved = filterOrphanedToolResults$1(preserved);
+	preserved = filterOrphanedToolUse$1(preserved);
+	if (preserved.length === 0) {
+		consola.warn("[AutoTruncate:OpenAI] All messages filtered out after cleanup");
+		return {
+			payload,
+			wasCompacted: false,
+			originalTokens,
+			compactedTokens: originalTokens,
+			removedMessageCount: 0
+		};
+	}
+	const removedMessages = conversationMessages.slice(0, preserveIndex);
+	const removedCount = conversationMessages.length - preserved.length;
+	const summary = generateRemovedMessagesSummary(removedMessages, (msg) => msg.tool_calls?.map((tc) => tc.function.name).filter(Boolean) ?? []);
+	let newSystemMessages = systemMessages;
+	let newMessages = preserved;
+	if (systemMessages.length > 0) {
+		const truncationContext = createTruncationSystemContext$1(removedCount, compressedCount, summary);
+		const lastSystemIdx = systemMessages.length - 1;
+		const lastSystem = systemMessages[lastSystemIdx];
+		const updatedSystem = {
+			...lastSystem,
+			content: typeof lastSystem.content === "string" ? lastSystem.content + truncationContext : lastSystem.content
+		};
+		newSystemMessages = [...systemMessages.slice(0, lastSystemIdx), updatedSystem];
+	} else newMessages = [createTruncationMarker$2(removedCount, compressedCount, summary), ...preserved];
+	const newPayload = {
+		...payload,
+		messages: [...newSystemMessages, ...newMessages]
 	};
-}
-
-//#endregion
-//#region src/services/copilot/create-chat-completions.ts
-const createChatCompletions = async (payload, options) => {
-	if (!state.copilotToken) throw new Error("Copilot token not found");
-	const enableVision = payload.messages.some((x) => typeof x.content !== "string" && x.content?.some((x) => x.type === "image_url"));
-	const isAgentCall = payload.messages.some((msg) => ["assistant", "tool"].includes(msg.role));
-	const headers = {
-		...copilotHeaders(state, {
-			vision: enableVision,
-			intent: isAgentCall ? "conversation-agent" : "conversation-panel"
-		}),
-		"X-Initiator": options?.initiator ?? (isAgentCall ? "agent" : "user")
+	const newBytes = JSON.stringify(newPayload).length;
+	const newTokenCount = await getTokenCount(newPayload, model);
+	let reason = "tokens";
+	if (exceedsTokens && exceedsBytes) reason = "tokens+size";
+	else if (exceedsBytes) reason = "size";
+	const actions = [];
+	if (removedCount > 0) actions.push(`removed ${removedCount} msgs`);
+	if (compressedCount > 0) actions.push(`compressed ${compressedCount} tool_results`);
+	const actionInfo = actions.length > 0 ? ` (${actions.join(", ")})` : "";
+	consola.info(`[AutoTruncate:OpenAI] ${reason}: ${originalTokens}→${newTokenCount.input} tokens, ${Math.round(originalBytes / 1024)}→${Math.round(newBytes / 1024)}KB${actionInfo}`);
+	if (newBytes > byteLimit) consola.warn(`[AutoTruncate:OpenAI] Result still over byte limit (${Math.round(newBytes / 1024)}KB > ${Math.round(byteLimit / 1024)}KB)`);
+	return {
+		payload: newPayload,
+		wasCompacted: true,
+		originalTokens,
+		compactedTokens: newTokenCount.input,
+		removedMessageCount: removedCount
 	};
-	const response = await fetch(`${copilotBaseUrl(state)}/chat/completions`, {
-		method: "POST",
-		headers,
-		body: JSON.stringify(payload)
-	});
-	if (!response.ok) {
-		consola.error("Failed to create chat completions", response);
-		throw await HTTPError.fromResponse("Failed to create chat completions", response, payload.model);
-	}
-	if (payload.stream) return events(response);
-	return await response.json();
-};
+}
 
 //#endregion
 //#region src/routes/shared.ts
@@ -3838,7 +3822,7 @@ async function handleCompletion$1(c) {
 		trackingId,
 		startTime
 	};
-	const selectedModel = state.models?.data.find((model) => model.id === originalPayload.model);
+	const selectedModel = findModelById(originalPayload.model);
 	await logTokenCount(originalPayload, selectedModel);
 	const { finalPayload, truncateResult } = await buildFinalPayload(originalPayload, selectedModel);
 	if (truncateResult) ctx.truncateResult = truncateResult;
@@ -3882,6 +3866,7 @@ async function executeRequest(opts) {
 	}
 }
 async function logTokenCount(payload, selectedModel) {
+	if (consola.level < 4) return;
 	try {
 		if (selectedModel) {
 			const tokenCount = await getTokenCount(payload, selectedModel);
@@ -3895,7 +3880,7 @@ function handleNonStreamingResponse$1(c, originalResponse, ctx, payload) {
 	consola.debug("Non-streaming response:", JSON.stringify(originalResponse));
 	let response = originalResponse;
 	if (state.verbose && ctx.truncateResult?.wasCompacted && response.choices[0]?.message.content) {
-		const marker = createTruncationResponseMarkerOpenAI(ctx.truncateResult);
+		const marker = createTruncationMarker$1(ctx.truncateResult);
 		response = {
 			...response,
 			choices: response.choices.map((choice, i) => i === 0 ? {
@@ -3981,7 +3966,7 @@ async function handleStreamingResponse$1(opts) {
 	const checkRepetition = createStreamRepetitionChecker(`openai:${payload.model}`);
 	try {
 		if (state.verbose && ctx.truncateResult?.wasCompacted) {
-			const marker = createTruncationResponseMarkerOpenAI(ctx.truncateResult);
+			const marker = createTruncationMarker$1(ctx.truncateResult);
 			const markerChunk = {
 				id: `compact-marker-${Date.now()}`,
 				object: "chat.completion.chunk",
@@ -4340,7 +4325,7 @@ function isFileDataPart(part) {
 async function handleGeminiCountTokens(c, model) {
 	try {
 		const { payload } = translateGeminiToOpenAI(await c.req.json(), model);
-		const selectedModel = state.models?.data.find((m) => m.id === model);
+		const selectedModel = findModelById(model);
 		if (!selectedModel) {
 			consola.warn("Model not found for count_tokens, returning estimate");
 			return c.json({ totalTokens: 1 });
@@ -4500,7 +4485,7 @@ async function handleGeminiGenerate(c, model, isStream) {
 		updateTrackerModel(trackingId, model);
 		const { payload } = translateGeminiToOpenAI(geminiRequest, model);
 		payload.stream = isStream;
-		const selectedModel = state.models?.data.find((m) => m.id === model);
+		const selectedModel = findModelById(model);
 		if (isNullish(payload.max_tokens) && selectedModel) payload.max_tokens = selectedModel.capabilities?.limits?.max_output_tokens;
 		const ctx = {
 			historyId: recordRequest("gemini", {
@@ -4602,25 +4587,27 @@ function handleNonStreamResponse(c, response, model, ctx, payload) {
 //#endregion
 //#region src/routes/gemini/model-alias.ts
 /**
-* Maps Gemini model names that aren't available on GitHub Copilot
-* to equivalent models that are.
+* Maps Gemini model names to equivalent models available on GitHub Copilot.
+*
+* Two types of aliases:
 *
-* The Gemini CLI's routing classifier requests gemini-2.5-flash-lite
-* and gemini-2.5-flash, which Copilot doesn't serve. We map them to
-* the closest available flash model.
+* - **Forced**: Always applied regardless of Copilot model availability.
+*   Use when the old model name should never reach the backend.
 *
-* Aliases are only applied when the requested model is absent from
-* the Copilot model list, so if Copilot adds support for these models
-* natively, requests will go through unchanged.
+* - **Conditional**: Only applied when the requested model is absent from
+*   the Copilot model list, so if Copilot adds native support the request
+*   goes through unchanged.
 */
-const GEMINI_MODEL_ALIASES = {
+const GEMINI_FORCED_ALIASES = { "gemini-2.5-pro": "gemini-3.1-pro-preview" };
+const GEMINI_CONDITIONAL_ALIASES = {
 	"gemini-2.5-flash-lite": "gemini-3-flash-preview",
 	"gemini-2.5-flash": "gemini-3-flash-preview"
 };
 function resolveGeminiModelAlias(model) {
-	if (!(model in GEMINI_MODEL_ALIASES)) return model;
-	if (state.models?.data.some((m) => m.id === model)) return model;
-	return GEMINI_MODEL_ALIASES[model];
+	if (model in GEMINI_FORCED_ALIASES) return GEMINI_FORCED_ALIASES[model];
+	if (!(model in GEMINI_CONDITIONAL_ALIASES)) return model;
+	if (findModelById(model)) return model;
+	return GEMINI_CONDITIONAL_ALIASES[model];
 }
 
 //#endregion
@@ -6220,9 +6207,6 @@ async function countTotalTokens(payload, model) {
 	}
 	return total;
 }
-function getMessageBytes(msg) {
-	return JSON.stringify(msg).length;
-}
 /**
 * Get tool_use IDs from an assistant message.
 */
@@ -6307,30 +6291,6 @@ function filterOrphanedToolUse(messages) {
 	return result;
 }
 /**
-* Ensure messages start with a user message.
-*/
-function ensureStartsWithUser(messages) {
-	let startIndex = 0;
-	while (startIndex < messages.length && messages[startIndex].role !== "user") startIndex++;
-	if (startIndex > 0) consola.debug(`[AutoTruncate:Anthropic] Skipped ${startIndex} leading non-user messages`);
-	return messages.slice(startIndex);
-}
-/** Threshold for large tool_result content (bytes) */
-const LARGE_TOOL_RESULT_THRESHOLD = 1e4;
-/** Maximum length for compressed tool_result summary */
-const COMPRESSED_SUMMARY_LENGTH = 500;
-/**
-* Compress a large tool_result content to a summary.
-* Keeps the first and last portions with a note about truncation.
-*/
-function compressToolResultContent(content) {
-	if (content.length <= LARGE_TOOL_RESULT_THRESHOLD) return content;
-	const halfLen = Math.floor(COMPRESSED_SUMMARY_LENGTH / 2);
-	const start = content.slice(0, halfLen);
-	const end = content.slice(-halfLen);
-	return `${start}\n\n[... ${(content.length - COMPRESSED_SUMMARY_LENGTH).toLocaleString()} characters omitted for brevity ...]\n\n${end}`;
-}
-/**
 * Compress a tool_result block in an Anthropic message.
 */
 function compressToolResultBlock(block) {
@@ -6349,28 +6309,11 @@ function compressToolResultBlock(block) {
 * @param preservePercent - Percentage of context to preserve uncompressed (0.0-1.0)
 */
 function smartCompressToolResults(messages, tokenLimit, byteLimit, preservePercent) {
-	const n = messages.length;
-	const cumTokens = Array.from({ length: n + 1 }, () => 0);
-	const cumBytes = Array.from({ length: n + 1 }, () => 0);
-	for (let i = n - 1; i >= 0; i--) {
-		const msg = messages[i];
-		cumTokens[i] = cumTokens[i + 1] + estimateMessageTokens(msg);
-		cumBytes[i] = cumBytes[i + 1] + getMessageBytes(msg) + 1;
-	}
-	const preserveTokenLimit = Math.floor(tokenLimit * preservePercent);
-	const preserveByteLimit = Math.floor(byteLimit * preservePercent);
-	let thresholdIndex = n;
-	for (let i = n - 1; i >= 0; i--) {
-		if (cumTokens[i] > preserveTokenLimit || cumBytes[i] > preserveByteLimit) {
-			thresholdIndex = i + 1;
-			break;
-		}
-		thresholdIndex = i;
-	}
-	if (thresholdIndex >= n) return {
+	const thresholdIndex = findCompressThreshold(messages, tokenLimit, byteLimit, preservePercent, estimateMessageTokens);
+	if (thresholdIndex >= messages.length) return {
 		messages,
 		compressedCount: 0,
-		compressThresholdIndex: n
+		compressThresholdIndex: messages.length
 	};
 	const result = [];
 	let compressedCount = 0;
@@ -6399,68 +6342,6 @@ function smartCompressToolResults(messages, tokenLimit, byteLimit, preservePerce
 		compressThresholdIndex: thresholdIndex
 	};
 }
-/** Default fallback for when model capabilities are not available */
-const DEFAULT_CONTEXT_WINDOW = 2e5;
-function calculateLimits(model, config) {
-	const rawTokenLimit = getEffectiveTokenLimit(model.id) ?? model.capabilities?.limits?.max_context_window_tokens ?? model.capabilities?.limits?.max_prompt_tokens ?? DEFAULT_CONTEXT_WINDOW;
-	return {
-		tokenLimit: Math.floor(rawTokenLimit * (1 - config.safetyMarginPercent / 100)),
-		byteLimit: getEffectiveByteLimitBytes()
-	};
-}
-function findOptimalPreserveIndex(params) {
-	const { messages, systemBytes, systemTokens, payloadOverhead, tokenLimit, byteLimit } = params;
-	if (messages.length === 0) return 0;
-	const markerBytes = 200;
-	const availableTokens = tokenLimit - systemTokens - 50;
-	const availableBytes = byteLimit - payloadOverhead - systemBytes - markerBytes;
-	if (availableTokens <= 0 || availableBytes <= 0) return messages.length;
-	const n = messages.length;
-	const cumTokens = Array.from({ length: n + 1 }, () => 0);
-	const cumBytes = Array.from({ length: n + 1 }, () => 0);
-	for (let i = n - 1; i >= 0; i--) {
-		const msg = messages[i];
-		cumTokens[i] = cumTokens[i + 1] + estimateMessageTokens(msg);
-		cumBytes[i] = cumBytes[i + 1] + getMessageBytes(msg) + 1;
-	}
-	let left = 0;
-	let right = n;
-	while (left < right) {
-		const mid = left + right >>> 1;
-		if (cumTokens[mid] <= availableTokens && cumBytes[mid] <= availableBytes) right = mid;
-		else left = mid + 1;
-	}
-	return left;
-}
-/**
-* Generate a summary of removed messages for context.
-* Extracts key information like tool calls and topics.
-*/
-function generateRemovedMessagesSummary(removedMessages) {
-	const toolCalls = [];
-	let userMessageCount = 0;
-	let assistantMessageCount = 0;
-	for (const msg of removedMessages) {
-		if (msg.role === "user") userMessageCount++;
-		else assistantMessageCount++;
-		if (Array.isArray(msg.content)) {
-			for (const block of msg.content) if (block.type === "tool_use") toolCalls.push(block.name);
-		}
-	}
-	const parts = [];
-	if (userMessageCount > 0 || assistantMessageCount > 0) {
-		const breakdown = [];
-		if (userMessageCount > 0) breakdown.push(`${userMessageCount} user`);
-		if (assistantMessageCount > 0) breakdown.push(`${assistantMessageCount} assistant`);
-		parts.push(`Messages: ${breakdown.join(", ")}`);
-	}
-	if (toolCalls.length > 0) {
-		const uniqueTools = [...new Set(toolCalls)];
-		const displayTools = uniqueTools.length > 5 ? [...uniqueTools.slice(0, 5), `+${uniqueTools.length - 5} more`] : uniqueTools;
-		parts.push(`Tools used: ${displayTools.join(", ")}`);
-	}
-	return parts.join(". ");
-}
 /**
 * Add a compression notice to the system prompt.
 * Informs the model that some tool_result content has been compressed.
@@ -6512,7 +6393,7 @@ async function autoTruncateAnthropic(payload, model, config = {}) {
 		...DEFAULT_AUTO_TRUNCATE_CONFIG,
 		...config
 	};
-	const { tokenLimit, byteLimit } = calculateLimits(model, cfg);
+	const { tokenLimit, byteLimit } = calculateLimits(model, cfg, 2e5);
 	const originalBytes = JSON.stringify(payload).length;
 	const originalTokens = await countTotalTokens(payload, model);
 	if (originalTokens <= tokenLimit && originalBytes <= byteLimit) return {
@@ -6565,7 +6446,8 @@ async function autoTruncateAnthropic(payload, model, config = {}) {
 		systemTokens,
 		payloadOverhead,
 		tokenLimit,
-		byteLimit
+		byteLimit,
+		estimateTokens: estimateMessageTokens
 	});
 	if (preserveIndex === 0) {
 		consola.warn("[AutoTruncate:Anthropic] Cannot truncate, system messages too large");
@@ -6590,7 +6472,7 @@ async function autoTruncateAnthropic(payload, model, config = {}) {
 	let preserved = workingMessages.slice(preserveIndex);
 	preserved = filterOrphanedToolResults(preserved);
 	preserved = filterOrphanedToolUse(preserved);
-	preserved = ensureStartsWithUser(preserved);
+	preserved = ensureStartsWithUser(preserved, "Anthropic");
 	preserved = filterOrphanedToolResults(preserved);
 	preserved = filterOrphanedToolUse(preserved);
 	if (preserved.length === 0) {
@@ -6605,7 +6487,10 @@ async function autoTruncateAnthropic(payload, model, config = {}) {
 	}
 	const removedMessages = payload.messages.slice(0, preserveIndex);
 	const removedCount = workingMessages.length - preserved.length;
-	const summary = generateRemovedMessagesSummary(removedMessages);
+	const summary = generateRemovedMessagesSummary(removedMessages, (msg) => {
+		if (!Array.isArray(msg.content)) return [];
+		return msg.content.filter((block) => block.type === "tool_use").map((block) => block.name);
+	});
 	let newSystem = payload.system;
 	let newMessages = preserved;
 	if (payload.system !== void 0) {
@@ -6647,7 +6532,7 @@ async function checkNeedsCompactionAnthropic(payload, model, config = {}) {
 	const { tokenLimit, byteLimit } = calculateLimits(model, {
 		...DEFAULT_AUTO_TRUNCATE_CONFIG,
 		...config
-	});
+	}, 2e5);
 	const currentTokens = await countTotalTokens(payload, model);
 	const currentBytes = JSON.stringify(payload).length;
 	const exceedsTokens = currentTokens > tokenLimit;
@@ -6887,7 +6772,7 @@ async function createAnthropicMessages(payload, options) {
 	if (!state.copilotToken) throw new Error("Copilot token not found");
 	let filteredPayload = filterPayloadForCopilot(payload);
 	filteredPayload = adjustMaxTokensForThinking(filteredPayload);
-	const resolvedModel = state.models?.data.find((m) => m.id === filteredPayload.model);
+	const resolvedModel = findModelById(filteredPayload.model);
 	const enableVision = filteredPayload.messages.some((msg) => {
 		if (typeof msg.content === "string") return false;
 		return msg.content.some((block) => block.type === "image");
@@ -6958,7 +6843,7 @@ function stripServerToolsFromPayload(tools) {
 */
 function supportsDirectAnthropicApi(modelId) {
 	if (state.redirectAnthropic) return false;
-	return (state.models?.data.find((m) => m.id === modelId))?.vendor === "Anthropic";
+	return findModelById(modelId)?.vendor === "Anthropic";
 }
 
 //#endregion
@@ -7002,15 +6887,6 @@ function extractSystemPrompt(system) {
 	return system.map((block) => block.text).join("\n");
 }
 function extractToolCallsFromContent(content) {
-	const tools = [];
-	for (const block of content) if (typeof block === "object" && block !== null && "type" in block && block.type === "tool_use" && "id" in block && "name" in block && "input" in block) tools.push({
-		id: String(block.id),
-		name: String(block.name),
-		input: JSON.stringify(block.input)
-	});
-	return tools.length > 0 ? tools : void 0;
-}
-function extractToolCallsFromAnthropicContent(content) {
 	const tools = [];
 	for (const block of content) if (block.type === "tool_use") tools.push({
 		id: block.id,
@@ -7028,6 +6904,25 @@ function mapOpenAIStopReasonToAnthropic(finishReason) {
 		content_filter: "end_turn"
 	}[finishReason];
 }
+function prependMarkerToResponse(response, marker) {
+	if (!marker) return response;
+	const content = [...response.content];
+	const firstTextIndex = content.findIndex((block) => block.type === "text");
+	if (firstTextIndex !== -1) {
+		const textBlock = content[firstTextIndex];
+		if (textBlock.type === "text") content[firstTextIndex] = {
+			...textBlock,
+			text: marker + (textBlock.text ?? "")
+		};
+	} else content.unshift({
+		type: "text",
+		text: marker
+	});
+	return {
+		...response,
+		content
+	};
+}
 
 //#endregion
 //#region src/routes/messages/stream-accumulator.ts
@@ -7589,7 +7484,7 @@ function translateErrorToAnthropicErrorEvent() {
 */
 async function handleDirectAnthropicCompletion(c, anthropicPayload, ctx, initiatorOverride) {
 	consola.debug("Using direct Anthropic API path for model:", anthropicPayload.model);
-	const selectedModel = state.models?.data.find((m) => m.id === anthropicPayload.model);
+	const selectedModel = findModelById(anthropicPayload.model);
 	let effectivePayload = anthropicPayload;
 	let truncateResult;
 	if (state.autoTruncate && selectedModel) {
@@ -7672,7 +7567,7 @@ function handleDirectAnthropicNonStreamingResponse(c, response, ctx, truncateRes
 				}
 			})
 		},
-		toolCalls: extractToolCallsFromAnthropicContent(response.content)
+		toolCalls: extractToolCallsFromContent(response.content)
 	}, Date.now() - ctx.startTime);
 	if (ctx.trackingId) requestTracker.updateRequest(ctx.trackingId, {
 		inputTokens: response.usage.input_tokens,
@@ -7690,34 +7585,12 @@ function handleDirectAnthropicNonStreamingResponse(c, response, ctx, truncateRes
 		stopReason: response.stop_reason ?? void 0
 	});
 	let finalResponse = response;
-	if (state.verbose && truncateResult?.wasCompacted) finalResponse = prependMarkerToAnthropicResponse$1(response, createTruncationMarker$1(truncateResult));
+	if (state.verbose && truncateResult?.wasCompacted) finalResponse = prependMarkerToResponse(response, createTruncationMarker$1(truncateResult));
 	logServerToolBlocks(finalResponse.content);
 	finalResponse = filterServerToolBlocksFromResponse(finalResponse);
 	return c.json(finalResponse);
 }
 /**
-* Prepend marker to Anthropic response content (at the beginning of first text block)
-*/
-function prependMarkerToAnthropicResponse$1(response, marker) {
-	if (!marker) return response;
-	const content = [...response.content];
-	const firstTextIndex = content.findIndex((block) => block.type === "text");
-	if (firstTextIndex !== -1) {
-		const textBlock = content[firstTextIndex];
-		if (textBlock.type === "text") content[firstTextIndex] = {
-			...textBlock,
-			text: marker + textBlock.text
-		};
-	} else content.unshift({
-		type: "text",
-		text: marker
-	});
-	return {
-		...response,
-		content
-	};
-}
-/**
 * Handle streaming direct Anthropic response (passthrough SSE events)
 */
 async function handleDirectAnthropicStreamingResponse(opts) {
@@ -7826,7 +7699,7 @@ async function handleTranslatedCompletion(c, anthropicPayload, ctx, initiatorOve
 	const { payload: translatedPayload, toolNameMapping } = translateToOpenAI(anthropicPayload);
 	consola.debug("Translated OpenAI request payload:", JSON.stringify(translatedPayload));
 	updateTrackerResolvedModel(ctx.trackingId, translatedPayload.model);
-	const selectedModel = state.models?.data.find((model) => model.id === translatedPayload.model);
+	const selectedModel = findModelById(translatedPayload.model);
 	const { finalPayload: openAIPayload, truncateResult } = await buildFinalPayload(translatedPayload, selectedModel);
 	if (truncateResult) ctx.truncateResult = truncateResult;
 	if (state.manualApprove) await awaitApproval();
@@ -7863,8 +7736,8 @@ function handleNonStreamingResponse(opts) {
 	let anthropicResponse = translateToAnthropic(response, toolNameMapping);
 	consola.debug("Translated Anthropic response:", JSON.stringify(anthropicResponse));
 	if (state.verbose && ctx.truncateResult?.wasCompacted) {
-		const marker = createTruncationResponseMarkerOpenAI(ctx.truncateResult);
-		anthropicResponse = prependMarkerToAnthropicResponse(anthropicResponse, marker);
+		const marker = createTruncationMarker$1(ctx.truncateResult);
+		anthropicResponse = prependMarkerToResponse(anthropicResponse, marker);
 	}
 	recordResponse(ctx.historyId, {
 		success: true,
@@ -7906,24 +7779,6 @@ function handleNonStreamingResponse(opts) {
 	});
 	return c.json(anthropicResponse);
 }
-function prependMarkerToAnthropicResponse(response, marker) {
-	const content = [...response.content];
-	const firstTextIndex = content.findIndex((block) => block.type === "text");
-	if (firstTextIndex !== -1) {
-		const textBlock = content[firstTextIndex];
-		if (textBlock.type === "text") content[firstTextIndex] = {
-			...textBlock,
-			text: marker + textBlock.text
-		};
-	} else content.unshift({
-		type: "text",
-		text: marker
-	});
-	return {
-		...response,
-		content
-	};
-}
 async function handleStreamingResponse(opts) {
 	const { stream, response, toolNameMapping, anthropicPayload, ctx } = opts;
 	const streamState = {
@@ -7936,7 +7791,7 @@ async function handleStreamingResponse(opts) {
 	const checkRepetition = createStreamRepetitionChecker(`translated:${anthropicPayload.model}`);
 	try {
 		if (ctx.truncateResult?.wasCompacted) {
-			const marker = createTruncationResponseMarkerOpenAI(ctx.truncateResult);
+			const marker = createTruncationMarker$1(ctx.truncateResult);
 			await sendTruncationMarkerEvent(stream, streamState, marker);
 			acc.content += marker;
 		}
@@ -8110,7 +7965,7 @@ async function handleCountTokens(c) {
 		const anthropicPayload = await c.req.json();
 		anthropicPayload.model = resolveModelFromBetaHeader(anthropicPayload.model, anthropicBeta);
 		const { payload: openAIPayload } = translateToOpenAI(anthropicPayload);
-		const selectedModel = state.models?.data.find((model) => model.id === openAIPayload.model);
+		const selectedModel = findModelById(openAIPayload.model);
 		if (!selectedModel) {
 			consola.warn("Model not found, returning default token count");
 			return c.json({ input_tokens: 1 });
@@ -8469,7 +8324,7 @@ const handleResponses = async (c) => {
 		trackingId,
 		startTime
 	};
-	if (!((state.models?.data.find((m) => m.id === payload.model))?.supported_endpoints?.includes(RESPONSES_ENDPOINT) ?? false)) {
+	if (!(findModelById(payload.model)?.supported_endpoints?.includes(RESPONSES_ENDPOINT) ?? false)) {
 		recordErrorResponse(ctx, model, /* @__PURE__ */ new Error("This model does not support the responses endpoint."));
 		return c.json({ error: {
 			message: "This model does not support the responses endpoint. Please choose a different model.",
@@ -8809,6 +8664,17 @@ function parseTimezoneOffset(value) {
 	if (!Number.isFinite(n)) return 8;
 	return n;
 }
+const validContextEditingModes = [
+	"off",
+	"clear-thinking",
+	"clear-tooluse",
+	"clear-both"
+];
+function parseContextEditing(value) {
+	if (validContextEditingModes.includes(value)) return value;
+	consola.warn(`Invalid context editing mode: "${value}", using "off". Valid: ${validContextEditingModes.join(", ")}`);
+	return "off";
+}
 const start = defineCommand({
 	meta: {
 		name: "start",
@@ -8956,7 +8822,7 @@ const start = defineCommand({
 			compressToolResults: args["compress-tool-results"],
 			redirectAnthropic: args["redirect-anthropic"],
 			stripServerTools: args["strip-server-tools"],
-			contextEditing: args["context-editing"],
+			contextEditing: parseContextEditing(args["context-editing"]),
 			timezoneOffset: parseTimezoneOffset(args["timezone-offset"]),
 			posthogKey: args["posthog-key"]
 		});