scream-code 0.11.2 → 0.11.3

package/dist/{app-8_tNuOds.mjs → app-D6B25mE5.mjs}

@@ -76030,6 +76030,52 @@ function canSplitAfterContext(messages, index) {
 /** Max consecutive compaction failures before auto-compaction is
 *  disabled for the remainder of the turn. Resets each turn. */
 const MAX_CONSECUTIVE_FAILURES = 3;
+/** Truncate tool call arguments to a brief summary for the compaction history. */
+function truncateArgsForSummary(args) {
+	if (args === null || args === void 0) return "";
+	try {
+		const parsed = JSON.parse(args);
+		if (typeof parsed !== "object" || parsed === null) return "";
+		const entries = Object.entries(parsed);
+		const parts = [];
+		for (const [key, value] of entries) {
+			const valStr = typeof value === "string" ? value.length > 50 ? `"${value.slice(0, 50)}…"` : `"${value}"` : String(value);
+			parts.push(`${key}=${valStr}`);
+		}
+		return parts.slice(0, 3).join(", ");
+	} catch {
+		return args.slice(0, 80);
+	}
+}
+/** Extract a brief tool-call history from compacted messages so the model
+*  knows what was already tried after compaction. Prevents repeating
+*  failed approaches. */
+function formatToolCallHistory(messages) {
+	const resultStatus = /* @__PURE__ */ new Map();
+	for (const msg of messages) if (msg.role === "tool" && msg.toolCallId !== void 0) resultStatus.set(msg.toolCallId, msg.isError === true);
+	const entries = [];
+	for (const msg of messages) {
+		if (msg.role !== "assistant" || msg.toolCalls.length === 0) continue;
+		for (const tc of msg.toolCalls) {
+			const isError = resultStatus.get(tc.id) ?? false;
+			const argsSummary = truncateArgsForSummary(tc.arguments);
+			entries.push({
+				name: tc.name,
+				argsSummary,
+				isError
+			});
+		}
+	}
+	if (entries.length === 0) return "";
+	return [
+		"## Recent Tool Calls",
+		"",
+		...entries.slice(-15).map((e) => {
+			const status = e.isError ? "error" : "success";
+			return `- ${e.name}(${e.argsSummary}) -> ${status}`;
+		})
+	].join("\n");
+}
 /** Minimal system prompt used during compaction. The full agent system
 *  prompt contains tool descriptions and runtime injections that contradict
 *  the compaction instruction ("DO NOT CALL ANY TOOLS"). This compact prompt
@@ -76284,7 +76330,8 @@ var FullCompaction = class {
 			const messagesToCompactForOps = originalHistory.slice(0, compactedCount);
 			const fileOps = createFileOps();
 			for (const msg of messagesToCompactForOps) extractFileOpsFromMessage(msg, fileOps);
-			const processedSummary = this.postProcessSummary(summary, fileOps);
+			const toolCallHistory = formatToolCallHistory(messagesToCompactForOps);
+			const processedSummary = this.postProcessSummary(summary, fileOps, toolCallHistory);
 			const tokensAfter = estimateTokens$1(processedSummary) + estimateTokensForMessages(recent);
 			const result = {
 				summary: processedSummary,
@@ -76397,7 +76444,7 @@ var FullCompaction = class {
 	* compression. Without this, both are lost after compaction because the
 	* original messages containing them are removed from the context window.
 	*/
-	postProcessSummary(summary, fileOps) {
+	postProcessSummary(summary, fileOps, toolCallHistory) {
 		const todos = this.agent.tools.storeData()["todo"] ?? [];
 		const sections = [summary.trim()];
 		if (todos.length > 0) {
@@ -76412,6 +76459,7 @@ var FullCompaction = class {
 		}
 		const filesSection = formatFileOperations(fileOps);
 		if (filesSection.length > 0) sections.push(filesSection);
+		if (toolCallHistory.length > 0) sections.push(toolCallHistory);
 		return sections.join("\n\n");
 	}
 };
@@ -76518,7 +76566,7 @@ const DEFAULT_CONFIG = {
 	keepRecentTokens: 4e4,
 	pruneMinReclaimTokens: 2e4,
 	minContentTokens: 100,
-	minContextUsageRatio: .5,
+	minContextUsageRatio: .3,
 	truncatedMarker: "[Old tool result content cleared]",
 	uselessMarker: "[Uneventful result elided]",
 	noMatchesMarker: "[no matches]"
@@ -77865,8 +77913,15 @@ const TOOL_OUTPUT_EMPTY_TEXT = "Tool output is empty.";
 /** Maximum token count for tool results persisted in conversation history.
 *  Results exceeding this limit are truncated to avoid bloating every
 *  subsequent API request with stale data.  The model can re-read the
-*  full content via read_file when needed. */
-const MAX_TOOL_RESULT_TOKENS = 8e3;
+*  full content via read_file when needed.
+*
+*  Lowered from 8000 to 5000 to reduce input-context pressure on long
+*  sessions. When input grows large, the available output space
+*  (context_window - input_tokens) shrinks, causing the model to hit
+*  max_tokens before emitting a tool call. Smaller tool-result footprints
+*  leave more room for output. The current turn always sees the full
+*  result via streaming; only the history copy is truncated. */
+const MAX_TOOL_RESULT_TOKENS = 5e3;
 const TOOL_TRUNCATION_NOTICE = "\n[content truncated — use read_file to re-read if needed]";
 var ContextMemory = class {
 	agent;
@@ -96038,10 +96093,25 @@ function parseEnvBudget(raw) {
 }
 /**
 * Compute the effective `max_completion_tokens` cap.
-*/
+*
+* Aligned with oh-my-pi's approach: use a reasonable output-token cap
+* instead of the full context window. `max_context_tokens` is the total
+* (input + output) budget, not the output-only budget. Using it as
+* max_tokens tells the API "generate up to 128K output tokens", which is
+* incorrect - the actual max output is much smaller. The API server
+* clamps it to context_window - input_tokens, which shrinks as context
+* grows, causing "max_tokens limit - no tool call" errors on long
+* sessions.
+*
+* The 64K cap matches oh-my-pi's OUTPUT_CAP_WHEN_UNKNOWN and is well
+* above any current model's real output limit (typically 8K-32K), so it
+* never artificially limits output while preventing the context-window
+* value from being used directly.
+*/
+const OUTPUT_TOKEN_CAP = 64e3;
 function computeCompletionBudgetCap(args) {
 	const maxCtx = args.capability?.max_context_tokens ?? 0;
-	const cap = args.budget.hardCap ?? (maxCtx > 0 ? maxCtx : args.budget.fallback ?? DEFAULT_UNKNOWN_CONTEXT_FALLBACK);
+	const cap = args.budget.hardCap ?? (maxCtx > 0 ? Math.min(maxCtx, OUTPUT_TOKEN_CAP) : args.budget.fallback ?? DEFAULT_UNKNOWN_CONTEXT_FALLBACK);
 	return Math.max(MIN_FLOOR, cap);
 }
 /**
@@ -120218,7 +120288,7 @@ function optionalBuildString(value) {
 	return typeof value === "string" && value.length > 0 ? value : void 0;
 }
 const SCREAM_BUILD_INFO = {
-	version: optionalBuildString("0.11.2"),
+	version: optionalBuildString("0.11.3"),
 	channel: optionalBuildString(""),
 	commit: optionalBuildString(""),
 	buildTarget: optionalBuildString("darwin-arm64")
@@ -125398,13 +125468,17 @@ function findProjectRoot(startDir) {
 	let dir = resolve(startDir);
 	let markerDir;
 	while (true) {
-		if (existsSync(resolve(dir, ".git"))) return dir;
+		const hasGit = existsSync(resolve(dir, ".git"));
 		if (markerDir === void 0) {
 			for (const marker of PROJECT_ROOT_MARKERS) if (existsSync(resolve(dir, marker))) {
 				markerDir = dir;
 				break;
 			}
 		}
+		if (hasGit) {
+			if (markerDir === void 0 || markerDir === dir) return dir;
+			return PROJECT_ROOT_MARKERS.some((m) => existsSync(resolve(dir, m))) ? dir : markerDir;
+		}
 		const parent = dirname$1(dir);
 		if (parent === dir) break;
 		dir = parent;
@@ -126958,7 +127032,7 @@ var ThinkingComponent = class {
 		this.stopSpinner();
 	}
 	dispose() {
-		this.stopSpinner();
+		this.finalize();
 	}
 	setExpanded(expanded) {
 		if (this.expanded === expanded) return;
@@ -148086,7 +148160,11 @@ async function runWebServer(opts) {
 		}
 		ws.send(JSON.stringify({ type: "server_empty" }));
 	});
-	await new Promise((resolve) => {
+	await new Promise((resolve, reject) => {
+		httpServer.once("error", (err) => {
+			if (err.code === "EADDRINUSE") reject(/* @__PURE__ */ new Error(`端口 ${opts.port} 已被占用，请先关闭占用该端口的进程，或使用 --port <port> 指定其他端口。`));
+			else reject(err);
+		});
 		httpServer.listen(opts.port, "127.0.0.1", resolve);
 	});
 	const url = `http://localhost:${opts.port}`;

package/dist/main.mjs

@@ -6,7 +6,7 @@ const __dirname = __cjsShimDirname(__filename);
 import "./suppress-sqlite-warning-C2VB0doZ.mjs";
 //#region src/main.ts
 try {
-	(await import("./app-8_tNuOds.mjs")).main();
+	(await import("./app-D6B25mE5.mjs")).main();
 } catch (error) {
 	process.stderr.write(`${error instanceof Error ? error.stack ?? error.message : String(error)}\n`);
 	process.exit(1);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "scream-code",
-  "version": "0.11.2",
+  "version": "0.11.3",
   "description": "A terminal-native AI agent for builders",
   "license": "MIT",
   "author": "ScreamCli",