@hsupu/copilot-api 0.7.20 → 0.7.21

package/config.example.yaml

@@ -56,7 +56,7 @@ model_overrides:
 stream_idle_timeout: 300       # Max seconds between SSE events (0 = no timeout).
                                # Applies to all streaming paths (Anthropic, Chat Completions, Responses).
 
-fetch_timeout: 300             # Seconds: request start → HTTP response headers (0 = no timeout).
+fetch_timeout: 600             # Seconds: request start → HTTP response headers (0 = no timeout).
                                # Applies to all upstream API clients.
 
 stale_request_max_age: 600     # Max seconds an active request can live before the stale reaper
@@ -69,8 +69,8 @@ stale_request_max_age: 600     # Max seconds an active request can live before t
 # Control graceful shutdown timing.
 
 shutdown:
-  graceful_wait: 60   # Phase 2: seconds to wait for in-flight requests to complete naturally (default: 60)
-  abort_wait: 120     # Phase 3: seconds to wait after abort signal for handlers to wrap up (default: 120)
+  graceful_wait: 300  # Phase 2: seconds to wait for in-flight requests to complete naturally (default: 60)
+  abort_wait: 600     # Phase 3: seconds to wait after abort signal for handlers to wrap up (default: 120)
 
 # ============================================================================
 # History
@@ -87,13 +87,21 @@ history:
 # Settings for Anthropic API tool handling and timeouts.
 
 anthropic:
-  strip_server_tools: false  # Strip server-side tools (web_search, etc.) from requests
-  dedup_tool_calls: false    # false | "input" | "result" (true = "input" for compat)
-                             # "input": dedup by (name, input); "result": also require identical result
-  strip_read_tool_result_tags: false  # Strip <system-reminder> tags from Read tool results
-  # rewrite_system_reminders: false   # false = keep all (default), true = remove all
-  rewrite_system_reminders:           # Or provide rewrite rules (first match wins, top-down).
-                                      # Note: `model` field is NOT supported here (only in system_prompt_overrides).
+  strip_server_tools: false               # Strip server-side tools (web_search, etc.) from requests
+  dedup_tool_calls: false                 # false | "input" | "result" (true = "input" for compat)
+                                          # "input": dedup by (name, input); "result": also require identical result
+  strip_read_tool_result_tags: false      # Strip <system-reminder> tags from Read tool results
+  context_editing: off                    # off | clear-thinking | clear-tooluse | clear-both
+                                          # Server-side context editing mode. Controls how Anthropic's
+                                          # context_management trims older context when input grows large.
+                                          # - off: disabled (default). No context_management sent.
+                                          # - clear-thinking: clear old thinking blocks.
+                                          # - clear-tooluse: clear old tool_use/result pairs.
+                                          # - clear-both: apply both clear-thinking and clear-tooluse.
+                                          # Only effective for models supporting context editing.
+  # rewrite_system_reminders: false       # false = keep all (default), true = remove all
+  rewrite_system_reminders:               # Or provide rewrite rules (first match wins, top-down).
+                                          # Note: `model` field is NOT supported here (only in system_prompt_overrides).
     - from: "^Whenever you read a file, you should consider whether it would be considered malware"
       to: ""                              # Empty = remove the tag
     # - from: ".*"                        # Catch-all: keep unchanged (gms flags are automatic)

package/dist/main.mjs

@@ -639,6 +639,7 @@ const state = {
 	accountType: "individual",
 	autoTruncate: true,
 	compressToolResultsBeforeTruncate: true,
+	contextEditingMode: "off",
 	stripServerTools: false,
 	dedupToolCalls: false,
 	fetchTimeout: 300,
@@ -772,12 +773,15 @@ async function ensurePaths() {
 	await ensureFile(PATHS.GITHUB_TOKEN_PATH);
 }
 async function ensureFile(filePath) {
+	const isWindows = process.platform === "win32";
 	try {
 		await fs.access(filePath, fs.constants.W_OK);
-		if (((await fs.stat(filePath)).mode & 511) !== 384) await fs.chmod(filePath, 384);
+		if (!isWindows) {
+			if (((await fs.stat(filePath)).mode & 511) !== 384) await fs.chmod(filePath, 384);
+		}
 	} catch {
 		await fs.writeFile(filePath, "");
-		await fs.chmod(filePath, 384);
+		if (!isWindows) await fs.chmod(filePath, 384);
 	}
 }
 
@@ -878,6 +882,7 @@ async function applyConfigToState() {
 		if (a.strip_server_tools !== void 0) state.stripServerTools = a.strip_server_tools;
 		if (a.dedup_tool_calls !== void 0) state.dedupToolCalls = a.dedup_tool_calls === true ? "input" : a.dedup_tool_calls;
 		if (a.strip_read_tool_result_tags !== void 0) state.stripReadToolResultTags = a.strip_read_tool_result_tags;
+		if (a.context_editing !== void 0) state.contextEditingMode = a.context_editing;
 		if (a.rewrite_system_reminders !== void 0) {
 			if (typeof a.rewrite_system_reminders === "boolean") state.rewriteSystemReminders = a.rewrite_system_reminders;
 			else if (Array.isArray(a.rewrite_system_reminders)) state.rewriteSystemReminders = compileRewriteRules(a.rewrite_system_reminders);
@@ -4723,7 +4728,7 @@ const setupClaudeCode = defineCommand({
 
 //#endregion
 //#region package.json
-var version = "0.7.20";
+var version = "0.7.21";
 
 //#endregion
 //#region src/lib/context/error-persistence.ts
@@ -4957,6 +4962,22 @@ function isEndpointSupported(model, endpoint) {
 	return model.supported_endpoints.includes(endpoint);
 }
 
+//#endregion
+//#region src/lib/ws.ts
+/** Create a shared WebSocket adapter for the given Hono app */
+async function createWebSocketAdapter(app) {
+	if (typeof globalThis.Bun !== "undefined") {
+		const { upgradeWebSocket } = await import("hono/bun");
+		return { upgradeWebSocket };
+	}
+	const { createNodeWebSocket } = await import("@hono/node-ws");
+	const nodeWs = createNodeWebSocket({ app });
+	return {
+		upgradeWebSocket: nodeWs.upgradeWebSocket,
+		injectWebSocket: (server) => nodeWs.injectWebSocket(server)
+	};
+}
+
 //#endregion
 //#region src/routes/history/api.ts
 function handleGetEntries(c) {
@@ -5066,25 +5087,12 @@ historyRoutes.get("/api/sessions/:id", handleGetSession);
 historyRoutes.delete("/api/sessions/:id", handleDeleteSession);
 /**
 * Initialize WebSocket support for history real-time updates.
-* Registers the /ws route on historyRoutes using the appropriate WebSocket
-* adapter for the current runtime (hono/bun for Bun, @hono/node-ws for Node.js).
+* Registers the /history/ws route on the root app using the shared WebSocket adapter.
 *
-* @param rootApp - The root Hono app instance (needed by @hono/node-ws to match upgrade requests)
-* @returns An `injectWebSocket` function that must be called with the Node.js HTTP server
-* after the server is created. Returns `undefined` under Bun (no injection needed).
+* @param rootApp - The root Hono app instance
+* @param upgradeWs - Shared WebSocket upgrade function from createWebSocketAdapter
 */
-async function initHistoryWebSocket(rootApp) {
-	let upgradeWs;
-	let injectFn;
-	if (typeof globalThis.Bun !== "undefined") {
-		const { upgradeWebSocket } = await import("hono/bun");
-		upgradeWs = upgradeWebSocket;
-	} else {
-		const { createNodeWebSocket } = await import("@hono/node-ws");
-		const nodeWs = createNodeWebSocket({ app: rootApp });
-		upgradeWs = nodeWs.upgradeWebSocket;
-		injectFn = (server) => nodeWs.injectWebSocket(server);
-	}
+function initHistoryWebSocket(rootApp, upgradeWs) {
 	rootApp.get("/history/ws", upgradeWs(() => ({
 		onOpen(_event, ws) {
 			addClient(ws.raw);
@@ -5098,7 +5106,6 @@ async function initHistoryWebSocket(rootApp) {
 			removeClient(ws.raw);
 		}
 	})));
-	return injectFn;
 }
 /**
 * Resolve a UI directory that exists at runtime.
@@ -6095,25 +6102,13 @@ async function handleResponseCreate(ws, payload) {
 * Initialize WebSocket routes for the Responses API.
 *
 * Registers GET /v1/responses and GET /responses on the root Hono app
-* with WebSocket upgrade handling. Follows the same pattern as
-* initHistoryWebSocket in src/routes/history/route.ts.
+* with WebSocket upgrade handling. Uses the shared WebSocket adapter
+* to avoid multiple upgrade listeners on the same HTTP server.
 *
-* @returns An inject function for Node.js HTTP server (undefined for Bun)
+* @param rootApp - The root Hono app instance
+* @param upgradeWs - Shared WebSocket upgrade function from createWebSocketAdapter
 */
-async function initResponsesWebSocket(rootApp) {
-	let upgradeWs;
-	let injectFn;
-	if (typeof globalThis.Bun !== "undefined") {
-		const { upgradeWebSocket } = await import("hono/bun");
-		upgradeWs = upgradeWebSocket;
-	} else {
-		const { createNodeWebSocket } = await import("@hono/node-ws");
-		const nodeWs = createNodeWebSocket({ app: rootApp });
-		upgradeWs = nodeWs.upgradeWebSocket;
-		injectFn = (server) => {
-			nodeWs.injectWebSocket(server);
-		};
-	}
+function initResponsesWebSocket(rootApp, upgradeWs) {
 	const wsHandler = upgradeWs(() => ({
 		onOpen(_event, _ws) {
 			consola.debug("[WS] Responses API WebSocket connected");
@@ -6147,7 +6142,6 @@ async function initResponsesWebSocket(rootApp) {
 	rootApp.get("/v1/responses", wsHandler);
 	rootApp.get("/responses", wsHandler);
 	consola.debug("[WS] Responses API WebSocket routes registered");
-	return injectFn;
 }
 
 //#endregion
@@ -6984,14 +6978,14 @@ function sanitizeOpenAIMessages(payload) {
 			content: filtered
 		};
 	});
-	const removedCount = originalCount - messages.length;
-	if (removedCount > 0) consola.info(`[Sanitizer:OpenAI] Filtered ${removedCount} orphaned tool messages`);
+	const blocksRemoved = originalCount - messages.length;
+	if (blocksRemoved > 0) consola.info(`[Sanitizer:OpenAI] Filtered ${blocksRemoved} orphaned tool messages`);
 	return {
 		payload: {
 			...payload,
 			messages: allMessages
 		},
-		blocksRemoved: removedCount,
+		blocksRemoved,
 		systemReminderRemovals
 	};
 }
@@ -8718,6 +8712,14 @@ function modelSupportsContextEditing(modelId) {
 	return normalized.startsWith("claude-haiku-4-5") || normalized.startsWith("claude-sonnet-4-5") || normalized.startsWith("claude-sonnet-4") || normalized.startsWith("claude-opus-4-5") || normalized.startsWith("claude-opus-4-6") || normalized.startsWith("claude-opus-4-1") || normalized.startsWith("claude-opus-4");
 }
 /**
+* Check if context editing is enabled for a model.
+* Requires both model support AND config mode != 'off'.
+* Mirrors VSCode Copilot Chat's isAnthropicContextEditingEnabled().
+*/
+function isContextEditingEnabled(modelId) {
+	return modelSupportsContextEditing(modelId) && state.contextEditingMode !== "off";
+}
+/**
 * Tool search is supported by:
 * - Claude Opus 4.5/4.6
 */
@@ -8756,7 +8758,7 @@ function buildAnthropicBetaHeaders(modelId, resolvedModel) {
 	const headers = {};
 	const betaFeatures = [];
 	if (!modelHasAdaptiveThinking(resolvedModel)) betaFeatures.push("interleaved-thinking-2025-05-14");
-	if (modelSupportsContextEditing(modelId)) betaFeatures.push("context-management-2025-06-27");
+	if (isContextEditingEnabled(modelId)) betaFeatures.push("context-management-2025-06-27");
 	if (modelSupportsToolSearch(modelId)) betaFeatures.push("advanced-tool-use-2025-11-20");
 	if (betaFeatures.length > 0) headers["anthropic-beta"] = betaFeatures.join(",");
 	return headers;
@@ -8767,22 +8769,28 @@ function buildAnthropicBetaHeaders(modelId, resolvedModel) {
 * From anthropic.ts:270-329 (buildContextManagement + getContextManagementFromConfig):
 * - clear_thinking: keep last N thinking turns
 * - clear_tool_uses: triggered by input_tokens threshold, keep last N tool uses
+*
+* Only builds edits matching the requested mode:
+* - "off" → undefined (no context management)
+* - "clear-thinking" → clear_thinking only (if thinking is enabled)
+* - "clear-tooluse" → clear_tool_uses only
+* - "clear-both" → both edits
 */
-function buildContextManagement(modelId, hasThinking) {
-	if (!modelSupportsContextEditing(modelId)) return;
+function buildContextManagement(mode, hasThinking) {
+	if (mode === "off") return;
 	const triggerType = "input_tokens";
 	const triggerValue = 1e5;
 	const keepCount = 3;
 	const thinkingKeepTurns = 1;
 	const edits = [];
-	if (hasThinking) edits.push({
+	if ((mode === "clear-thinking" || mode === "clear-both") && hasThinking) edits.push({
 		type: "clear_thinking_20251015",
 		keep: {
 			type: "thinking_turns",
 			value: Math.max(1, thinkingKeepTurns)
 		}
 	});
-	edits.push({
+	if (mode === "clear-tooluse" || mode === "clear-both") edits.push({
 		type: "clear_tool_uses_20250919",
 		trigger: {
 			type: triggerType,
@@ -8793,7 +8801,7 @@ function buildContextManagement(modelId, hasThinking) {
 			value: keepCount
 		}
 	});
-	return { edits };
+	return edits.length > 0 ? { edits } : void 0;
 }
 
 //#endregion
@@ -9108,8 +9116,9 @@ async function createAnthropicMessages(payload, opts) {
 		"anthropic-version": "2023-06-01",
 		...buildAnthropicBetaHeaders(model, opts?.resolvedModel)
 	};
-	if (!wire.context_management) {
-		const contextManagement = buildContextManagement(model, Boolean(thinking && thinking.type !== "disabled"));
+	if (!wire.context_management && isContextEditingEnabled(model)) {
+		const hasThinking = Boolean(thinking && thinking.type !== "disabled");
+		const contextManagement = buildContextManagement(state.contextEditingMode, hasThinking);
 		if (contextManagement) {
 			wire.context_management = contextManagement;
 			consola.debug("[DirectAnthropic] Added context_management:", JSON.stringify(contextManagement));
@@ -10343,6 +10352,7 @@ async function runServer(options) {
 	state.showGitHubToken = options.showGitHubToken;
 	state.autoTruncate = options.autoTruncate;
 	await ensurePaths();
+	consola.info(`Data directory: ${PATHS.APP_DIR}`);
 	const config = await applyConfigToState();
 	const proxyUrl = options.proxy ?? config.proxy;
 	initProxy({
@@ -10428,8 +10438,9 @@ async function runServer(options) {
 		if (runtime?.bun?.server) c.env = { server: runtime.bun.server };
 		await next();
 	});
-	const injectHistoryWs = await initHistoryWebSocket(server);
-	const injectResponsesWs = await initResponsesWebSocket(server);
+	const wsAdapter = await createWebSocketAdapter(server);
+	initHistoryWebSocket(server, wsAdapter.upgradeWebSocket);
+	initResponsesWebSocket(server, wsAdapter.upgradeWebSocket);
 	consola.box(`Web UI:\n🌐 Usage Viewer: https://ericc-ch.github.io/copilot-api?endpoint=${serverUrl}/usage\n📜 History UI:   ${serverUrl}/history`);
 	const bunWebSocket = typeof globalThis.Bun !== "undefined" ? (await import("hono/bun")).websocket : void 0;
 	let serverInstance;
@@ -10451,13 +10462,9 @@ async function runServer(options) {
 	}
 	setServerInstance(serverInstance);
 	setupShutdownHandlers();
-	if (injectHistoryWs || injectResponsesWs) {
+	if (wsAdapter.injectWebSocket) {
 		const nodeServer = serverInstance.node?.server;
-		if (nodeServer && "on" in nodeServer) {
-			const ns = nodeServer;
-			injectHistoryWs?.(ns);
-			injectResponsesWs?.(ns);
-		}
+		if (nodeServer && "on" in nodeServer) wsAdapter.injectWebSocket(nodeServer);
 	}
 	await waitForShutdown();
 }