@cryer/star-cli 0.3.3 → 0.3.4

package/README.md

@@ -52,6 +52,8 @@ Config file: `~/.star-cli/config.toml` (project-level override: `.star/config.to
 defaultModel = "gpt"
 permissionMode = "ask"   # ask | auto | readonly | yolo | plan
 contextMaxTokens = 100000
+# compactThresholdTokens = 80000   # optional auto-compaction trigger in tokens; defaults to the
+                                  # context window. /compact always compacts, regardless of this.
 contextCompaction = "summary"   # summary | truncate — how over-budget history is compacted
 # Seconds with no stream output before a stalled response is ended gracefully
 # (some relays never close the stream). The first token gets a fixed 120s allowance.
@@ -148,7 +150,7 @@ star --clear-sessions all     delete every stored session
 | `/connect` | interactive provider onboarding wizard: pick a preset or custom endpoint, paste the API key (masked), name a model — appends `[[providers]]`/`[[models]]` to the config, stores the key in `~/.star-cli/.env` (never in `config.toml`), and can set the new model as default + switch to it immediately |
 | `/plan` | toggle plan mode: read-only research, then approve the generated plan before it executes (session-only) |
 | `/memory [add <text>]` | show the long-term memory file (`~/.star-cli/MEMORY.md`, injected into every session's system prompt), or append a line to it — see [Long-term memory](#long-term-memory) |
-| `/compact` | compact conversation history to free up context |
+| `/compact` | compact conversation history now — always runs, even below the auto-compact threshold |
 | `/export [path]` | export the current session to a Markdown file |
 | `/undo` | undo the last conversation turn: revert its file changes (write_file/edit_file) and retract its messages — earlier turns are never touched (previews the message count and per-file revert diffs, then asks for confirmation) |
 | `/rewind [n]` | list file-change checkpoints, or rewind to just before checkpoint `n`: restore every file changed since then and retract the matching conversation messages (asks for confirmation first) |

package/README.zh-CN.md

@@ -52,6 +52,8 @@ star -p "hi"             # non-interactive print mode
 defaultModel = "gpt"
 permissionMode = "ask"   # ask | auto | readonly | yolo | plan
 contextMaxTokens = 100000
+# compactThresholdTokens = 80000   # optional auto-compaction trigger in tokens; defaults to the
+                                  # context window. /compact always compacts, regardless of this.
 contextCompaction = "summary"   # summary | truncate — how over-budget history is compacted
 # Seconds with no stream output before a stalled response is ended gracefully
 # (some relays never close the stream). The first token gets a fixed 120s allowance.
@@ -148,7 +150,7 @@ star --clear-sessions all     delete every stored session
 | `/connect` | 交互式服务商接入向导：选择预设或自定义端点，粘贴 API 密钥（掩码显示），命名模型——把 `[[providers]]`/`[[models]]` 追加到配置，密钥存入 `~/.star-cli/.env`（绝不写入 `config.toml`），还可将新模型设为默认并立即切换 |
 | `/plan` | 切换计划模式：先只读调研，再批准生成的计划后才执行（仅会话内有效） |
 | `/memory [add <text>]` | 查看长期记忆文件（`~/.star-cli/MEMORY.md`，注入每个会话的系统提示词），或向其中追加一行——见[长期记忆](#长期记忆) |
-| `/compact` | 压缩对话历史以释放上下文 |
+| `/compact` | 立即压缩对话历史——低于自动压缩阈值时也会执行 |
 | `/export [path]` | 将当前会话导出为 Markdown 文件 |
 | `/undo` | 撤销上一轮对话：还原其文件改动（write_file/edit_file）并撤回其消息——更早的回合不受影响（先预览消息数量与逐文件还原 diff，再要求确认） |
 | `/rewind [n]` | 列出文件改动检查点，或回滚到检查点 `n` 之前：还原此后改动的所有文件并撤回对应的对话消息（先要求确认） |

package/dist/{chunk-3L6QKJM4.js → chunk-K3R42LOX.js}

@@ -12,6 +12,123 @@ import {
 // src/agent/loop.ts
 import { tool as aiTool, generateText as generateText3 } from "ai";
 
+// src/config/schema.ts
+import { z } from "zod";
+var ProviderConfigSchema = z.object({
+  name: z.string(),
+  // openai-responses talks to the Responses API (/responses) instead of chat
+  // completions — the wire format codex uses, and the native API of
+  // reasoning-model relays that otherwise hide reasoning behind empty chat
+  // completions. streamChat sends strictSchemas:false + store:false with it.
+  protocol: z.enum(["openai-compatible", "anthropic", "openai-responses"]).default("openai-compatible"),
+  baseURL: z.string(),
+  apiKeyEnv: z.string().optional(),
+  apiKey: z.string().optional(),
+  headers: z.record(z.string()).optional()
+});
+var ModelConfigSchema = z.object({
+  name: z.string(),
+  provider: z.string(),
+  model: z.string(),
+  maxTokens: z.number().int().positive().optional(),
+  // Per-model context window; overrides the top-level contextMaxTokens for
+  // compaction and the status bar when set.
+  contextMaxTokens: z.number().int().positive().optional(),
+  // Cost estimation needs BOTH prices (USD per 1M tokens); with only one set
+  // the model is treated as unpriced.
+  promptPrice: z.number().nonnegative().optional(),
+  completionPrice: z.number().nonnegative().optional(),
+  // Optional price (USD per 1M tokens) for cache-read prompt tokens. When
+  // unset, OpenAI-style cached tokens bill at promptPrice and Anthropic-style
+  // cache reads are not billed — set this for accurate cache pricing.
+  cacheReadPrice: z.number().nonnegative().optional()
+});
+var PermissionsConfigSchema = z.object({
+  allow: z.array(z.string()).default([]),
+  deny: z.array(z.string()).default([])
+});
+var HookConfigSchema = z.object({
+  event: z.enum(["PreToolUse", "PostToolUse", "Stop"]),
+  matcher: z.string().refine(
+    (pattern) => {
+      try {
+        new RegExp(pattern);
+        return true;
+      } catch {
+        return false;
+      }
+    },
+    { message: "invalid regular expression" }
+  ).optional(),
+  command: z.string().min(1),
+  timeoutSec: z.number().positive().default(30)
+});
+var ConfigSchema = z.object({
+  defaultModel: z.string().default(""),
+  permissionMode: z.enum(["auto", "ask", "readonly", "yolo", "plan"]).default("ask"),
+  providers: z.array(ProviderConfigSchema).default([]),
+  models: z.array(ModelConfigSchema).default([]),
+  maxSteps: z.number().int().positive().default(50),
+  contextMaxTokens: z.number().int().positive().default(1e5),
+  // Token threshold at which the agent loop auto-compacts the history.
+  // Defaults to the effective context window (compact only when full); set a
+  // lower value to compact earlier and keep headroom. Manual /compact ignores
+  // this — it always compacts.
+  compactThresholdTokens: z.number().int().positive().optional(),
+  // Seconds without any stream part before a stalled response is ended
+  // gracefully (some relays never send the terminal chunks). Applies once
+  // streaming has started; the first part gets a longer, fixed allowance.
+  streamIdleTimeoutSec: z.number().positive().default(20),
+  // Seconds to wait for the very first stream part before giving up; slow
+  // thinking-model relays can buffer for minutes before sending anything.
+  streamFirstChunkTimeoutSec: z.number().positive().default(300),
+  // Extra attempts per model request when a stream fails transiently
+  // (network error, 429/5xx, idle watchdog cutoff) or comes back empty, so a
+  // relay hiccup does not silently end a half-finished turn. 0 disables.
+  // The wait between attempts honors Retry-After headers and otherwise backs
+  // off exponentially with jitter (llm/retry.ts); retries after a
+  // timeout/empty failure also scale both stream timeouts up (attempt number
+  // ×, capped at 3x) since the relay is likely overloaded. Default 5 matches
+  // opencode/Codex retry budgets and rides out flaky relays (中转站) better
+  // than a minimal budget.
+  streamMaxRetries: z.number().int().min(0).default(5),
+  // How many consecutive unproductive replies may be nudged before the turn
+  // is handed back: a text-only reply that announces pending work ("我将…",
+  // "I will…") or answers a nudge with more words, or an empty reply that
+  // survived the stream retry budget (steered with a nudge — a changed
+  // request — since identical resends already failed). Tool calls reset the
+  // count, so a productive task keeps going; exhaustion surfaces a notice
+  // instead of stopping silently. Never fires in plan mode. 0 disables.
+  maxAutoContinues: z.number().int().min(0).default(2),
+  contextCompaction: z.enum(["summary", "truncate"]).default("summary"),
+  // Optional per-session dollar budget; when set, the REPL stops the session
+  // once the estimated cost for the session reaches this amount.
+  sessionBudgetUsd: z.number().positive().optional(),
+  // Terminal bell (REPL only): ring when a turn takes longer than
+  // notifyBellThresholdSec, and when a background task finishes.
+  // STAR_NO_NOTIFY=1 disables without touching the config.
+  notifyBell: z.boolean().default(true),
+  notifyBellThresholdSec: z.number().positive().default(10),
+  permissions: PermissionsConfigSchema.default({ allow: [], deny: [] }),
+  hooks: z.array(HookConfigSchema).default([]),
+  // Consecutive identical tool calls (same tool + same arguments) allowed
+  // before the loop refuses to execute the repeat and tells the model to
+  // change approach — a guard against agents stuck retrying a failing call
+  // (opencode's doom_loop). 0 disables.
+  doomLoopThreshold: z.number().int().min(0).default(3),
+  // Track the whole working tree with an internal git repo (outside the
+  // user's .git) at every turn boundary, so /undo and /redo can restore
+  // changes made by ANY tool (bash included), not just write_file/edit_file.
+  // Falls back silently to per-file snapshots when git is unavailable.
+  gitSnapshots: z.boolean().default(true)
+});
+function contextWindowTokens(config, modelName) {
+  return config.models.find((m) => m.name === modelName)?.contextMaxTokens ?? config.contextMaxTokens;
+}
+function resolveCompactThreshold(config, windowTokens) {
+  return Math.min(config.compactThresholdTokens ?? windowTokens, windowTokens);
+}
+
 // src/context/compaction.ts
 import { generateText } from "ai";
 
@@ -106,8 +223,9 @@ ${transcript}`,
   });
   return text.trim();
 }
-function compactMessages(messages, maxTokens) {
-  if (estimateTokens(messages) <= maxTokens) {
+function compactMessages(messages, maxTokens, opts = {}) {
+  const limit = opts.force ? 0 : maxTokens;
+  if (estimateTokens(messages) <= limit) {
     return { messages, compacted: false, droppedCount: 0 };
   }
   const first = messages[0];
@@ -133,7 +251,7 @@ function compactMessages(messages, maxTokens) {
     droppedCount += turn.length;
     turnIndex += 1;
     const candidate = [...head, placeholderMessage(droppedCount), ...rest.slice(droppedCount)];
-    if (estimateTokens(candidate) <= maxTokens) {
+    if (estimateTokens(candidate) <= limit) {
       break;
     }
   }
@@ -1091,7 +1209,7 @@ async function rewindToSnapshot(id, includeSubagent = false) {
 import { spawn as spawn3 } from "child_process";
 import { existsSync as existsSync2 } from "fs";
 import path4 from "path";
-import { z } from "zod";
+import { z as z2 } from "zod";
 
 // src/tasks/manager.ts
 import { spawn as spawn2 } from "child_process";
@@ -1270,11 +1388,11 @@ function resolveShell() {
   }
   return cachedSpec;
 }
-var schema = z.object({
-  command: z.string().describe("Shell command to execute"),
-  timeout: z.number().positive().max(MAX_TIMEOUT).optional().describe("Timeout in seconds (default 120, max 600)"),
-  description: z.string().optional().describe("Short description of what the command does"),
-  run_in_background: z.boolean().optional().describe(
+var schema = z2.object({
+  command: z2.string().describe("Shell command to execute"),
+  timeout: z2.number().positive().max(MAX_TIMEOUT).optional().describe("Timeout in seconds (default 120, max 600)"),
+  description: z2.string().optional().describe("Short description of what the command does"),
+  run_in_background: z2.boolean().optional().describe(
     "Run the command in the background and return a task id immediately (default false). Use task_output/task_kill to inspect or stop it."
   )
 });
@@ -1408,12 +1526,12 @@ Command timed out after ${timeoutSeconds}s`,
 // src/tools/fs/edit.ts
 import { readFile as readFile2, stat as stat2, writeFile as writeFile2 } from "fs/promises";
 import path5 from "path";
-import { z as z2 } from "zod";
-var schema2 = z2.object({
-  path: z2.string().describe("File path, absolute or relative to the working directory"),
-  old_string: z2.string().min(1).describe("Exact text to replace"),
-  new_string: z2.string().describe("Replacement text"),
-  replace_all: z2.boolean().optional().describe("Replace every occurrence instead of requiring a unique match")
+import { z as z3 } from "zod";
+var schema2 = z3.object({
+  path: z3.string().describe("File path, absolute or relative to the working directory"),
+  old_string: z3.string().min(1).describe("Exact text to replace"),
+  new_string: z3.string().describe("Replacement text"),
+  replace_all: z3.boolean().optional().describe("Replace every occurrence instead of requiring a unique match")
 });
 var editFileTool = {
   name: "edit_file",
@@ -1477,7 +1595,7 @@ var editFileTool = {
 
 // src/tools/fs/glob.ts
 import path7 from "path";
-import { z as z3 } from "zod";
+import { z as z4 } from "zod";
 
 // src/tools/fs/util.ts
 import { readFileSync, statSync } from "fs";
@@ -1708,11 +1826,11 @@ function literalDirPrefix(pattern) {
   }
   return out;
 }
-var schema3 = z3.object({
-  pattern: z3.string().describe(
+var schema3 = z4.object({
+  pattern: z4.string().describe(
     "Glob pattern, e.g. 'src/**/*.ts'. Supports **, *, ?, {a,b} braces and [abc] character classes. A bare pattern like '*.ts' matches basenames recursively."
   ),
-  path: z3.string().optional().describe("Directory to search, defaults to the working directory")
+  path: z4.string().optional().describe("Directory to search, defaults to the working directory")
 });
 var globTool = {
   name: "glob",
@@ -1736,7 +1854,7 @@ var globTool = {
 // src/tools/fs/grep.ts
 import { readFile as readFile3, stat as stat4 } from "fs/promises";
 import path8 from "path";
-import { z as z4 } from "zod";
+import { z as z5 } from "zod";
 var MAX_MATCHES = 250;
 var MAX_LINE_CHARS = 500;
 var MAX_PATTERN_CHARS = 200;
@@ -1744,11 +1862,11 @@ var MAX_FILE_BYTES = 5 * 1024 * 1024;
 function isTooComplexPattern(pattern) {
   return pattern.length > MAX_PATTERN_CHARS || /\)[+*]/.test(pattern) || /\)\{\d+(,\d*)?\}/.test(pattern);
 }
-var schema4 = z4.object({
-  pattern: z4.string().describe("Regular expression to search for"),
-  path: z4.string().optional().describe("File or directory to search, defaults to the working directory"),
-  glob: z4.string().optional().describe("Glob pattern to filter which files are searched, e.g. '*.ts'"),
-  ignoreCase: z4.boolean().optional().describe("Case-insensitive matching")
+var schema4 = z5.object({
+  pattern: z5.string().describe("Regular expression to search for"),
+  path: z5.string().optional().describe("File or directory to search, defaults to the working directory"),
+  glob: z5.string().optional().describe("Glob pattern to filter which files are searched, e.g. '*.ts'"),
+  ignoreCase: z5.boolean().optional().describe("Case-insensitive matching")
 });
 var grepTool = {
   name: "grep",
@@ -1847,7 +1965,7 @@ var grepTool = {
 // src/tools/fs/read.ts
 import { open, stat as stat5 } from "fs/promises";
 import path9 from "path";
-import { z as z5 } from "zod";
+import { z as z6 } from "zod";
 var DEFAULT_LIMIT = 2e3;
 var MAX_CHARS = 100 * 1024;
 var MAX_READ_BYTES = 1024 * 1024;
@@ -1873,10 +1991,10 @@ function decodeTextBuffer(buf) {
   }
   return { text: buf.toString("utf8") };
 }
-var schema5 = z5.object({
-  path: z5.string().describe("File path, absolute or relative to the working directory"),
-  offset: z5.number().int().positive().optional().describe("1-based line number to start reading from"),
-  limit: z5.number().int().positive().optional().describe("Maximum number of lines to read (default 2000)")
+var schema5 = z6.object({
+  path: z6.string().describe("File path, absolute or relative to the working directory"),
+  offset: z6.number().int().positive().optional().describe("1-based line number to start reading from"),
+  limit: z6.number().int().positive().optional().describe("Maximum number of lines to read (default 2000)")
 });
 var readFileTool = {
   name: "read_file",
@@ -1949,10 +2067,10 @@ var readFileTool = {
 // src/tools/fs/write.ts
 import { mkdir as mkdir3, writeFile as writeFile3 } from "fs/promises";
 import path10 from "path";
-import { z as z6 } from "zod";
-var schema6 = z6.object({
-  path: z6.string().describe("File path, absolute or relative to the working directory"),
-  content: z6.string().describe("Full content to write to the file")
+import { z as z7 } from "zod";
+var schema6 = z7.object({
+  path: z7.string().describe("File path, absolute or relative to the working directory"),
+  content: z7.string().describe("Full content to write to the file")
 });
 var writeFileTool = {
   name: "write_file",
@@ -1974,7 +2092,7 @@ var writeFileTool = {
 };
 
 // src/tools/tasks.ts
-import { z as z7 } from "zod";
+import { z as z8 } from "zod";
 
 // src/agent/agent-tasks.ts
 import { EventEmitter as EventEmitter2 } from "events";
@@ -2086,6 +2204,20 @@ var AgentTaskManager = class extends EventEmitter2 {
   }
 };
 var defaultAgentTasks = new AgentTaskManager();
+function formatAgentTaskUpdate(task) {
+  const label = task.description ? `: ${task.description}` : "";
+  if (task.status === "running") {
+    return `Background subagent ${task.id} started${label}`;
+  }
+  let text = `Background subagent ${task.id} ${task.status}${label}`;
+  if (task.status === "failed") {
+    const reason = (task.result.split("\n")[0] ?? "").replace(/^Subagent failed:\s*/, "");
+    if (reason) {
+      text += ` \u2014 ${reason.length > 140 ? `${reason.slice(0, 140)}\u2026` : reason}`;
+    }
+  }
+  return text;
+}
 
 // src/tasks/format.ts
 function formatDuration(startedAt, endedAt) {
@@ -2131,8 +2263,8 @@ function formatTaskFinished(task) {
 }
 
 // src/tools/tasks.ts
-var idSchema = z7.object({
-  id: z7.string().describe("Task id, e.g. task-1 or agent-1")
+var idSchema = z8.object({
+  id: z8.string().describe("Task id, e.g. task-1 or agent-1")
 });
 function truncateOutput(s) {
   if (s.length <= MAX_OUTPUT) return s;
@@ -2149,7 +2281,7 @@ var taskListTool = {
   name: "task_list",
   description: "List background tasks (shell commands and subagents) with status and runtime.",
   permission: "read",
-  parameters: z7.object({}),
+  parameters: z8.object({}),
   execute() {
     const shell = formatTaskList(defaultTaskManager.list());
     const agents = defaultAgentTasks.list();
@@ -2210,16 +2342,16 @@ var taskKillTool = {
 // src/tools/todo.ts
 import { mkdir as mkdir4, readFile as readFile4, rename, writeFile as writeFile4 } from "fs/promises";
 import path11 from "path";
-import { z as z8 } from "zod";
-var todoItemSchema = z8.object({
-  id: z8.number().int(),
-  title: z8.string(),
-  status: z8.enum(["pending", "in_progress", "done"])
+import { z as z9 } from "zod";
+var todoItemSchema = z9.object({
+  id: z9.number().int(),
+  title: z9.string(),
+  status: z9.enum(["pending", "in_progress", "done"])
 });
-var writeSchema = z8.object({
-  todos: z8.array(todoItemSchema).describe("The full todo list, replacing the current one")
+var writeSchema = z9.object({
+  todos: z9.array(todoItemSchema).describe("The full todo list, replacing the current one")
 });
-var readSchema = z8.object({});
+var readSchema = z9.object({});
 function fileFor(cwd) {
   return path11.join(cwd, ".star", "todos.json");
 }
@@ -2335,13 +2467,13 @@ function createTodoTools(store = defaultStore) {
 }
 
 // src/tools/web/fetch.ts
-import { z as z9 } from "zod";
+import { z as z10 } from "zod";
 var DEFAULT_MAX_CHARS = 2e4;
 var MAX_BYTES = 2 * 1024 * 1024;
 var TIMEOUT_MS = 3e4;
-var schema7 = z9.object({
-  url: z9.string().describe("The http(s) URL to fetch"),
-  maxChars: z9.number().int().positive().optional().describe("Maximum characters of text to return (default 20000)")
+var schema7 = z10.object({
+  url: z10.string().describe("The http(s) URL to fetch"),
+  maxChars: z10.number().int().positive().optional().describe("Maximum characters of text to return (default 20000)")
 });
 function safeCodePoint(n, raw) {
   if (!Number.isInteger(n) || n < 0 || n > 1114111) {
@@ -2475,15 +2607,15 @@ var webFetchTool = {
 };
 
 // src/tools/web/search.ts
-import { z as z10 } from "zod";
+import { z as z11 } from "zod";
 var TIMEOUT_MS2 = 15e3;
 var DEFAULT_MAX_RESULTS = 5;
 var MAX_RESULTS_LIMIT = 10;
 var SEARCH_URL = "https://html.duckduckgo.com/html/?q=";
 var USER_AGENT = "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/124.0.0.0 Safari/537.36";
-var schema8 = z10.object({
-  query: z10.string().min(1).describe("The search query"),
-  maxResults: z10.number().int().min(1).max(MAX_RESULTS_LIMIT).optional().describe(
+var schema8 = z11.object({
+  query: z11.string().min(1).describe("The search query"),
+  maxResults: z11.number().int().min(1).max(MAX_RESULTS_LIMIT).optional().describe(
     `Maximum number of results to return (default ${DEFAULT_MAX_RESULTS}, max ${MAX_RESULTS_LIMIT})`
   )
 });
@@ -2664,7 +2796,7 @@ ${body}`;
 // src/agent/skills.ts
 import { readFileSync as readFileSync3, readdirSync, statSync as statSync3 } from "fs";
 import path13 from "path";
-import { z as z11 } from "zod";
+import { z as z12 } from "zod";
 var SKILL_FILE = "SKILL.md";
 var SKILL_MAX_CHARS = 32 * 1024;
 var NAME_PATTERN = /^[a-z0-9-]+$/;
@@ -2749,8 +2881,8 @@ function createSkillTool(deps) {
     name: "skill",
     description: "Load the full instructions of a skill by name. Available skills are listed in the system prompt when present; call this before starting work that matches one of them.",
     permission: "read",
-    parameters: z11.object({
-      name: z11.string().describe("Name of the skill to load, as listed in the system prompt.")
+    parameters: z12.object({
+      name: z12.string().describe("Name of the skill to load, as listed in the system prompt.")
     }),
     async execute(args) {
       const skills = discoverSkills(deps.cwd, deps.home);
@@ -2780,7 +2912,7 @@ function createSkillTool(deps) {
 }
 
 // src/agent/subagent.ts
-import { z as z12 } from "zod";
+import { z as z13 } from "zod";
 
 // src/tools/index.ts
 function createDefaultRegistry(todoStore) {
@@ -2792,7 +2924,7 @@ var MAX_SUBAGENT_DEPTH = 1;
 var MAX_REPORT_CHARS = 2e4;
 var SUBAGENT_PROMPT = "You are a subagent spawned to handle a focused subtask on behalf of the main agent. Work autonomously with the tools available to you, then finish with a concise report of what you found or did. Your final text is returned to the main agent as the tool result \u2014 make it self-contained.";
 async function runSubagent(deps, args, cwd, signal) {
-  const { AgentLoop: AgentLoop2 } = await import("./loop-TQ7GFTTE.js");
+  const { AgentLoop: AgentLoop2 } = await import("./loop-SDZVEQND.js");
   const child = new AgentLoop2({
     model: deps.model,
     registry: createDefaultRegistry(),
@@ -2850,10 +2982,10 @@ function createSubagentTool(deps) {
     name: "subagent",
     description: "Spawn a subagent with its own agent loop to handle a focused subtask (research, exploration, or an isolated change). The subagent has the same tools but cannot spawn further subagents. Returns the subagent's final report. Prefer this for self-contained work that would otherwise clutter the main conversation. Set run_in_background to start it without blocking: independent subtasks then run concurrently and their reports are delivered to you automatically when they finish (also inspectable via task_list/task_output/task_kill).",
     permission: "exec",
-    parameters: z12.object({
-      prompt: z12.string().describe("Complete, self-contained instructions for the subagent."),
-      description: z12.string().optional().describe("Short label for the subtask, shown in the result header."),
-      run_in_background: z12.boolean().optional().describe(
+    parameters: z13.object({
+      prompt: z13.string().describe("Complete, self-contained instructions for the subagent."),
+      description: z13.string().optional().describe("Short label for the subtask, shown in the result header."),
+      run_in_background: z13.boolean().optional().describe(
         "Run concurrently in the background and return a task id immediately (default false). Use for independent subtasks that can run in parallel."
       )
     }),
@@ -2881,7 +3013,7 @@ function createSubagentTool(deps) {
 // src/agent/user-memory.ts
 import { appendFileSync, mkdirSync, readFileSync as readFileSync4, statSync as statSync4 } from "fs";
 import path14 from "path";
-import { z as z13 } from "zod";
+import { z as z14 } from "zod";
 var USER_MEMORY_MAX_CHARS = 32 * 1024;
 var REMEMBER_TEXT_MAX_CHARS = 500;
 var cache3 = /* @__PURE__ */ new Map();
@@ -2926,8 +3058,8 @@ function createRememberTool(deps = {}) {
     name: "remember",
     description: `Save a note to the user's long-term memory (${userMemoryPath()}), injected into every future session. ONLY use when the user explicitly asks to remember or save something. Never use autonomously.`,
     permission: "write",
-    parameters: z13.object({
-      text: z13.string().max(REMEMBER_TEXT_MAX_CHARS).describe("The note to save, as a single concise sentence.")
+    parameters: z14.object({
+      text: z14.string().max(REMEMBER_TEXT_MAX_CHARS).describe("The note to save, as a single concise sentence.")
     }),
     async execute(args) {
       appendUserMemory(args.text, deps.home);
@@ -3357,7 +3489,10 @@ ${finished.result}`
         }
       }
       const maxTokens = this.opts.contextMaxTokens ?? config.contextMaxTokens;
-      const compacted = compactMessages([...this.messages], maxTokens);
+      const compacted = compactMessages(
+        [...this.messages],
+        resolveCompactThreshold(config, maxTokens)
+      );
       if (compacted.compacted) {
         this.messages = await this.applyCompactionSummary(compacted, signal);
         this.turnMarkers = [];
@@ -3425,7 +3560,7 @@ ${finished.result}`
         if (attempt >= maxRetries) break;
         if (!failure && text.length === 0 && toolCalls.length === 0 && attempt >= 1 && (isFilteredEmptyFinish(lastFinishReason) || sawReasoning || lastUsage !== void 0 && lastUsage.completionTokens > 0))
           break;
-        const delayMs = computeRetryDelayMs(attempt, baseDelay, failure);
+        const delayMs = computeRetryDelayMs(attempt, failure ? baseDelay : baseDelay * 2, failure);
         yield {
           type: "retry",
           attempt: attempt + 2,
@@ -3772,6 +3907,9 @@ ${summary}`
 
 export {
   defaultAgentTasks,
+  formatAgentTaskUpdate,
+  ConfigSchema,
+  contextWindowTokens,
   estimateTokens,
   summarizeMessages,
   compactMessages,
@@ -3814,4 +3952,4 @@ export {
   PLAN_MODE_PROMPT,
   AgentLoop
 };
-//# sourceMappingURL=chunk-3L6QKJM4.js.map
+//# sourceMappingURL=chunk-K3R42LOX.js.map