@llmist/cli 12.3.6 → 13.0.0

package/dist/cli.js

@@ -40,7 +40,18 @@ var OPTION_FLAGS = {
   voice: "--voice <name>",
   speechFormat: "--format <format>",
   speechSpeed: "--speed <value>",
-  speechOutput: "-o, --output <path>"
+  speechOutput: "-o, --output <path>",
+  // Rate limiting options
+  rateLimitRpm: "--rate-limit-rpm <count>",
+  rateLimitTpm: "--rate-limit-tpm <count>",
+  rateLimitDaily: "--rate-limit-daily <count>",
+  rateLimitSafetyMargin: "--rate-limit-safety-margin <value>",
+  noRateLimit: "--no-rate-limit",
+  // Retry options
+  maxRetries: "--max-retries <count>",
+  retryMinTimeout: "--retry-min-timeout <ms>",
+  retryMaxTimeout: "--retry-max-timeout <ms>",
+  noRetry: "--no-retry"
 };
 var OPTION_DESCRIPTIONS = {
   model: "Model identifier, e.g. openai:gpt-5-nano or anthropic:claude-sonnet-4-5.",
@@ -66,7 +77,18 @@ var OPTION_DESCRIPTIONS = {
   voice: "Voice name for speech generation, e.g. 'nova', 'alloy', 'Zephyr'.",
   speechFormat: "Audio format: 'mp3', 'opus', 'aac', 'flac', 'wav', 'pcm'.",
   speechSpeed: "Speech speed multiplier (0.25 to 4.0, default 1.0).",
-  speechOutput: "Output path for audio file. Defaults to stdout if not specified."
+  speechOutput: "Output path for audio file. Defaults to stdout if not specified.",
+  // Rate limiting descriptions
+  rateLimitRpm: "Maximum requests per minute (RPM). Overrides config and defaults.",
+  rateLimitTpm: "Maximum tokens per minute (TPM). Overrides config and defaults.",
+  rateLimitDaily: "Maximum tokens per day. Useful for Gemini free tier.",
+  rateLimitSafetyMargin: "Safety margin (0-1). Start throttling at this percentage of limit.",
+  noRateLimit: "Disable rate limiting entirely.",
+  // Retry descriptions
+  maxRetries: "Maximum retry attempts for failed API calls.",
+  retryMinTimeout: "Initial retry delay in milliseconds.",
+  retryMaxTimeout: "Maximum retry delay in milliseconds.",
+  noRetry: "Disable retry logic for API calls."
 };
 var SUMMARY_PREFIX = "[llmist]";
 
@@ -76,7 +98,7 @@ import { Command, InvalidArgumentError as InvalidArgumentError2 } from "commande
 // package.json
 var package_default = {
   name: "@llmist/cli",
-  version: "12.3.6",
+  version: "13.0.0",
   description: "CLI for llmist - run LLM agents from the command line",
   type: "module",
   main: "dist/cli.js",
@@ -132,7 +154,7 @@ var package_default = {
     node: ">=22.0.0"
   },
   dependencies: {
-    llmist: "^12.3.6",
+    llmist: "^13.0.0",
     "@unblessed/node": "^1.0.0-alpha.23",
     chalk: "^5.6.2",
     commander: "^12.1.0",
@@ -146,7 +168,7 @@ var package_default = {
     zod: "^4.1.12"
   },
   devDependencies: {
-    "@llmist/testing": "^12.3.6",
+    "@llmist/testing": "^13.0.0",
     "@types/diff": "^8.0.0",
     "@types/js-yaml": "^4.0.9",
     "@types/marked-terminal": "^6.1.1",
@@ -399,6 +421,8 @@ var COMPLETE_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "inherits",
   "log-level",
   "log-llm-requests",
+  "rate-limits",
+  "retry",
   "type"
   // Allowed for inheritance compatibility, ignored for built-in commands
 ]);
@@ -429,6 +453,8 @@ var AGENT_CONFIG_KEYS = /* @__PURE__ */ new Set([
   "inherits",
   "log-level",
   "log-llm-requests",
+  "rate-limits",
+  "retry",
   "type"
   // Allowed for inheritance compatibility, ignored for built-in commands
 ]);
@@ -554,7 +580,110 @@ function validateSubagentConfigMap(value, section) {
   }
   const result = {};
   for (const [subagentName, config] of Object.entries(value)) {
-    result[subagentName] = validateSingleSubagentConfig(config, subagentName, `${section}.subagents`);
+    result[subagentName] = validateSingleSubagentConfig(
+      config,
+      subagentName,
+      `${section}.subagents`
+    );
+  }
+  return result;
+}
+var RATE_LIMITS_CONFIG_KEYS = /* @__PURE__ */ new Set([
+  "requests-per-minute",
+  "tokens-per-minute",
+  "tokens-per-day",
+  "safety-margin",
+  "enabled"
+]);
+function validateRateLimitsConfig(value, section) {
+  if (typeof value !== "object" || value === null || Array.isArray(value)) {
+    throw new ConfigError(`[${section}] must be a table`);
+  }
+  const raw = value;
+  const result = {};
+  for (const [key, val] of Object.entries(raw)) {
+    if (!RATE_LIMITS_CONFIG_KEYS.has(key)) {
+      throw new ConfigError(`[${section}] has unknown key: ${key}`);
+    }
+    switch (key) {
+      case "requests-per-minute":
+        result["requests-per-minute"] = validateNumber(val, key, section, {
+          integer: true,
+          min: 1
+        });
+        break;
+      case "tokens-per-minute":
+        result["tokens-per-minute"] = validateNumber(val, key, section, { integer: true, min: 1 });
+        break;
+      case "tokens-per-day":
+        result["tokens-per-day"] = validateNumber(val, key, section, { integer: true, min: 1 });
+        break;
+      case "safety-margin":
+        result["safety-margin"] = validateNumber(val, key, section, { min: 0, max: 1 });
+        break;
+      case "enabled":
+        result.enabled = validateBoolean(val, key, section);
+        break;
+    }
+  }
+  if (result["requests-per-minute"] && result["requests-per-minute"] > 1e4) {
+    console.warn(
+      `\u26A0\uFE0F  Warning: [${section}].requests-per-minute is very high (${result["requests-per-minute"]}). Make sure your API tier supports this rate.`
+    );
+  }
+  if (result["tokens-per-minute"] && result["tokens-per-minute"] > 5e6) {
+    console.warn(
+      `\u26A0\uFE0F  Warning: [${section}].tokens-per-minute is very high (${result["tokens-per-minute"]}). Make sure your API tier supports this rate.`
+    );
+  }
+  return result;
+}
+var RETRY_CONFIG_KEYS = /* @__PURE__ */ new Set([
+  "enabled",
+  "retries",
+  "min-timeout",
+  "max-timeout",
+  "factor",
+  "randomize",
+  "respect-retry-after",
+  "max-retry-after-ms"
+]);
+function validateRetryConfig(value, section) {
+  if (typeof value !== "object" || value === null || Array.isArray(value)) {
+    throw new ConfigError(`[${section}] must be a table`);
+  }
+  const raw = value;
+  const result = {};
+  for (const [key, val] of Object.entries(raw)) {
+    if (!RETRY_CONFIG_KEYS.has(key)) {
+      throw new ConfigError(`[${section}] has unknown key: ${key}`);
+    }
+    switch (key) {
+      case "enabled":
+        result.enabled = validateBoolean(val, key, section);
+        break;
+      case "retries":
+        result.retries = validateNumber(val, key, section, { integer: true, min: 0 });
+        break;
+      case "min-timeout":
+        result["min-timeout"] = validateNumber(val, key, section, { integer: true, min: 0 });
+        break;
+      case "max-timeout":
+        result["max-timeout"] = validateNumber(val, key, section, { integer: true, min: 0 });
+        break;
+      case "factor":
+        result.factor = validateNumber(val, key, section, { min: 1 });
+        break;
+      case "randomize":
+        result.randomize = validateBoolean(val, key, section);
+        break;
+      case "respect-retry-after":
+        result["respect-retry-after"] = validateBoolean(val, key, section);
+        break;
+      case "max-retry-after-ms":
+        result["max-retry-after-ms"] = validateNumber(val, key, section, { integer: true, min: 0 });
+        break;
+    }
   }
   return result;
 }
@@ -610,7 +739,9 @@ function validateInitialGadgets(value, section) {
     }
     const entryObj = entry;
     if (!("gadget" in entryObj)) {
-      throw new ConfigError(`[${section}].initial-gadgets[${i}] is missing required field 'gadget'`);
+      throw new ConfigError(
+        `[${section}].initial-gadgets[${i}] is missing required field 'gadget'`
+      );
     }
     if (typeof entryObj.gadget !== "string") {
       throw new ConfigError(`[${section}].initial-gadgets[${i}].gadget must be a string`);
@@ -624,7 +755,9 @@ function validateInitialGadgets(value, section) {
       throw new ConfigError(`[${section}].initial-gadgets[${i}].parameters must be a table`);
     }
     if (!("result" in entryObj)) {
-      throw new ConfigError(`[${section}].initial-gadgets[${i}] is missing required field 'result'`);
+      throw new ConfigError(
+        `[${section}].initial-gadgets[${i}] is missing required field 'result'`
+      );
     }
     if (typeof entryObj.result !== "string") {
       throw new ConfigError(`[${section}].initial-gadgets[${i}].result must be a string`);
@@ -705,7 +838,20 @@ function validateCompleteConfig(raw, section) {
     result.quiet = validateBoolean(rawObj.quiet, "quiet", section);
   }
   if ("log-llm-requests" in rawObj) {
-    result["log-llm-requests"] = validateBoolean(rawObj["log-llm-requests"], "log-llm-requests", section);
+    result["log-llm-requests"] = validateBoolean(
+      rawObj["log-llm-requests"],
+      "log-llm-requests",
+      section
+    );
+  }
+  if ("rate-limits" in rawObj) {
+    result["rate-limits"] = validateRateLimitsConfig(
+      rawObj["rate-limits"],
+      `${section}.rate-limits`
+    );
+  }
+  if ("retry" in rawObj) {
+    result.retry = validateRetryConfig(rawObj.retry, `${section}.retry`);
   }
   return result;
 }
@@ -789,7 +935,20 @@ function validateAgentConfig(raw, section) {
     result.quiet = validateBoolean(rawObj.quiet, "quiet", section);
   }
   if ("log-llm-requests" in rawObj) {
-    result["log-llm-requests"] = validateBoolean(rawObj["log-llm-requests"], "log-llm-requests", section);
+    result["log-llm-requests"] = validateBoolean(
+      rawObj["log-llm-requests"],
+      "log-llm-requests",
+      section
+    );
+  }
+  if ("rate-limits" in rawObj) {
+    result["rate-limits"] = validateRateLimitsConfig(
+      rawObj["rate-limits"],
+      `${section}.rate-limits`
+    );
+  }
+  if ("retry" in rawObj) {
+    result.retry = validateRetryConfig(rawObj.retry, `${section}.retry`);
   }
   return result;
 }
@@ -959,7 +1118,11 @@ function validateCustomConfig(raw, section) {
     result.quiet = validateBoolean(rawObj.quiet, "quiet", section);
   }
   if ("log-llm-requests" in rawObj) {
-    result["log-llm-requests"] = validateBoolean(rawObj["log-llm-requests"], "log-llm-requests", section);
+    result["log-llm-requests"] = validateBoolean(
+      rawObj["log-llm-requests"],
+      "log-llm-requests",
+      section
+    );
   }
   Object.assign(result, validateLoggingConfig(rawObj, section));
   return result;
@@ -999,6 +1162,10 @@ function validateConfig(raw, configPath) {
         result.prompts = validatePromptsConfig(value, key);
       } else if (key === "subagents") {
         result.subagents = validateGlobalSubagentConfig(value, key);
+      } else if (key === "rate-limits") {
+        result["rate-limits"] = validateRateLimitsConfig(value, key);
+      } else if (key === "retry") {
+        result.retry = validateRetryConfig(value, key);
       } else {
         result[key] = validateCustomConfig(value, key);
       }
@@ -1046,7 +1213,9 @@ function getCustomCommandNames(config) {
     "image",
     "speech",
     "prompts",
-    "subagents"
+    "subagents",
+    "rate-limits",
+    "retry"
   ]);
   return Object.keys(config).filter((key) => !reserved.has(key));
 }
@@ -1191,7 +1360,7 @@ function resolveInheritance(config, configPath) {
     if (resolvedGadgets.length > 0) {
       merged.gadgets = resolvedGadgets;
     }
-    delete merged["gadget"];
+    delete merged.gadget;
     delete merged["gadget-add"];
     delete merged["gadget-remove"];
     resolving.delete(name);
@@ -3483,7 +3652,42 @@ function addCompleteOptions(cmd, defaults) {
     OPTION_FLAGS.logLlmRequests,
     OPTION_DESCRIPTIONS.logLlmRequests,
     defaults?.["log-llm-requests"]
-  ).option(OPTION_FLAGS.inputImage, OPTION_DESCRIPTIONS.inputImage).option(OPTION_FLAGS.inputAudio, OPTION_DESCRIPTIONS.inputAudio);
+  ).option(OPTION_FLAGS.inputImage, OPTION_DESCRIPTIONS.inputImage).option(OPTION_FLAGS.inputAudio, OPTION_DESCRIPTIONS.inputAudio).option(
+    OPTION_FLAGS.rateLimitRpm,
+    OPTION_DESCRIPTIONS.rateLimitRpm,
+    createNumericParser({ label: "RPM", integer: true, min: 1 }),
+    defaults?.["rate-limits"]?.["requests-per-minute"]
+  ).option(
+    OPTION_FLAGS.rateLimitTpm,
+    OPTION_DESCRIPTIONS.rateLimitTpm,
+    createNumericParser({ label: "TPM", integer: true, min: 1 }),
+    defaults?.["rate-limits"]?.["tokens-per-minute"]
+  ).option(
+    OPTION_FLAGS.rateLimitDaily,
+    OPTION_DESCRIPTIONS.rateLimitDaily,
+    createNumericParser({ label: "Daily tokens", integer: true, min: 1 }),
+    defaults?.["rate-limits"]?.["tokens-per-day"]
+  ).option(
+    OPTION_FLAGS.rateLimitSafetyMargin,
+    OPTION_DESCRIPTIONS.rateLimitSafetyMargin,
+    createNumericParser({ label: "Safety margin", min: 0, max: 1 }),
+    defaults?.["rate-limits"]?.["safety-margin"]
+  ).option(OPTION_FLAGS.noRateLimit, OPTION_DESCRIPTIONS.noRateLimit).option(
+    OPTION_FLAGS.maxRetries,
+    OPTION_DESCRIPTIONS.maxRetries,
+    createNumericParser({ label: "Max retries", integer: true, min: 0 }),
+    defaults?.retry?.retries
+  ).option(
+    OPTION_FLAGS.retryMinTimeout,
+    OPTION_DESCRIPTIONS.retryMinTimeout,
+    createNumericParser({ label: "Min timeout", integer: true, min: 0 }),
+    defaults?.retry?.["min-timeout"]
+  ).option(
+    OPTION_FLAGS.retryMaxTimeout,
+    OPTION_DESCRIPTIONS.retryMaxTimeout,
+    createNumericParser({ label: "Max timeout", integer: true, min: 0 }),
+    defaults?.retry?.["max-timeout"]
+  ).option(OPTION_FLAGS.noRetry, OPTION_DESCRIPTIONS.noRetry);
 }
 function addAgentOptions(cmd, defaults) {
   const gadgetAccumulator = (value, previous = []) => [
@@ -3511,7 +3715,42 @@ function addAgentOptions(cmd, defaults) {
     OPTION_FLAGS.logLlmRequests,
     OPTION_DESCRIPTIONS.logLlmRequests,
     defaults?.["log-llm-requests"]
-  ).option(OPTION_FLAGS.inputImage, OPTION_DESCRIPTIONS.inputImage).option(OPTION_FLAGS.inputAudio, OPTION_DESCRIPTIONS.inputAudio);
+  ).option(OPTION_FLAGS.inputImage, OPTION_DESCRIPTIONS.inputImage).option(OPTION_FLAGS.inputAudio, OPTION_DESCRIPTIONS.inputAudio).option(
+    OPTION_FLAGS.rateLimitRpm,
+    OPTION_DESCRIPTIONS.rateLimitRpm,
+    createNumericParser({ label: "RPM", integer: true, min: 1 }),
+    defaults?.["rate-limits"]?.["requests-per-minute"]
+  ).option(
+    OPTION_FLAGS.rateLimitTpm,
+    OPTION_DESCRIPTIONS.rateLimitTpm,
+    createNumericParser({ label: "TPM", integer: true, min: 1 }),
+    defaults?.["rate-limits"]?.["tokens-per-minute"]
+  ).option(
+    OPTION_FLAGS.rateLimitDaily,
+    OPTION_DESCRIPTIONS.rateLimitDaily,
+    createNumericParser({ label: "Daily tokens", integer: true, min: 1 }),
+    defaults?.["rate-limits"]?.["tokens-per-day"]
+  ).option(
+    OPTION_FLAGS.rateLimitSafetyMargin,
+    OPTION_DESCRIPTIONS.rateLimitSafetyMargin,
+    createNumericParser({ label: "Safety margin", min: 0, max: 1 }),
+    defaults?.["rate-limits"]?.["safety-margin"]
+  ).option(OPTION_FLAGS.noRateLimit, OPTION_DESCRIPTIONS.noRateLimit).option(
+    OPTION_FLAGS.maxRetries,
+    OPTION_DESCRIPTIONS.maxRetries,
+    createNumericParser({ label: "Max retries", integer: true, min: 0 }),
+    defaults?.retry?.retries
+  ).option(
+    OPTION_FLAGS.retryMinTimeout,
+    OPTION_DESCRIPTIONS.retryMinTimeout,
+    createNumericParser({ label: "Min timeout", integer: true, min: 0 }),
+    defaults?.retry?.["min-timeout"]
+  ).option(
+    OPTION_FLAGS.retryMaxTimeout,
+    OPTION_DESCRIPTIONS.retryMaxTimeout,
+    createNumericParser({ label: "Max timeout", integer: true, min: 0 }),
+    defaults?.retry?.["max-timeout"]
+  ).option(OPTION_FLAGS.noRetry, OPTION_DESCRIPTIONS.noRetry);
 }
 function configToCompleteOptions(config) {
   const result = {};
@@ -3521,6 +3760,21 @@ function configToCompleteOptions(config) {
   if (config["max-tokens"] !== void 0) result.maxTokens = config["max-tokens"];
   if (config.quiet !== void 0) result.quiet = config.quiet;
   if (config["log-llm-requests"] !== void 0) result.logLlmRequests = config["log-llm-requests"];
+  if (config["rate-limits"]) {
+    const rl = config["rate-limits"];
+    if (rl["requests-per-minute"] !== void 0) result.rateLimitRpm = rl["requests-per-minute"];
+    if (rl["tokens-per-minute"] !== void 0) result.rateLimitTpm = rl["tokens-per-minute"];
+    if (rl["tokens-per-day"] !== void 0) result.rateLimitDaily = rl["tokens-per-day"];
+    if (rl["safety-margin"] !== void 0) result.rateLimitSafetyMargin = rl["safety-margin"];
+    if (rl.enabled === false) result.noRateLimit = true;
+  }
+  if (config.retry) {
+    const r = config.retry;
+    if (r.retries !== void 0) result.maxRetries = r.retries;
+    if (r["min-timeout"] !== void 0) result.retryMinTimeout = r["min-timeout"];
+    if (r["max-timeout"] !== void 0) result.retryMaxTimeout = r["max-timeout"];
+    if (r.enabled === false) result.noRetry = true;
+  }
   return result;
 }
 function configToAgentOptions(config) {
@@ -3545,9 +3799,190 @@ function configToAgentOptions(config) {
   if (config["log-llm-requests"] !== void 0) result.logLlmRequests = config["log-llm-requests"];
   if (config.subagents !== void 0) result.subagents = config.subagents;
   if (config["initial-gadgets"] !== void 0) result.initialGadgets = config["initial-gadgets"];
+  if (config["rate-limits"]) {
+    const rl = config["rate-limits"];
+    if (rl["requests-per-minute"] !== void 0) result.rateLimitRpm = rl["requests-per-minute"];
+    if (rl["tokens-per-minute"] !== void 0) result.rateLimitTpm = rl["tokens-per-minute"];
+    if (rl["tokens-per-day"] !== void 0) result.rateLimitDaily = rl["tokens-per-day"];
+    if (rl["safety-margin"] !== void 0) result.rateLimitSafetyMargin = rl["safety-margin"];
+    if (rl.enabled === false) result.noRateLimit = true;
+  }
+  if (config.retry) {
+    const r = config.retry;
+    if (r.retries !== void 0) result.maxRetries = r.retries;
+    if (r["min-timeout"] !== void 0) result.retryMinTimeout = r["min-timeout"];
+    if (r["max-timeout"] !== void 0) result.retryMaxTimeout = r["max-timeout"];
+    if (r.enabled === false) result.noRetry = true;
+  }
   return result;
 }
 
+// src/rate-limit-resolver.ts
+import { getProvider, resolveModel } from "llmist";
+var PROVIDER_DEFAULTS = {
+  anthropic: {
+    requestsPerMinute: 50,
+    // Tier 1 safe
+    tokensPerMinute: 4e4,
+    // Tier 1 for claude-3-5-sonnet
+    safetyMargin: 0.8
+  },
+  openai: {
+    requestsPerMinute: 3,
+    // Free tier minimum (very conservative)
+    tokensPerMinute: 4e4,
+    safetyMargin: 0.8
+  },
+  gemini: {
+    requestsPerMinute: 15,
+    // Free tier documented
+    tokensPerMinute: 1e6,
+    // Free tier documented
+    tokensPerDay: 15e5,
+    // Free tier daily limit
+    safetyMargin: 0.8
+  }
+};
+function detectProvider(model) {
+  try {
+    const resolved = resolveModel(model, { silent: true });
+    return getProvider(resolved) ?? null;
+  } catch {
+    return null;
+  }
+}
+function resolveRateLimitConfig(options, globalConfig, profileConfig, model) {
+  if (options.noRateLimit === true) {
+    return { enabled: false, safetyMargin: 0.8 };
+  }
+  let resolved;
+  if (model) {
+    const provider = detectProvider(model);
+    if (provider && PROVIDER_DEFAULTS[provider]) {
+      resolved = { ...PROVIDER_DEFAULTS[provider] };
+    }
+  }
+  if (!resolved) {
+    resolved = {};
+  }
+  if (globalConfig) {
+    if (globalConfig["requests-per-minute"] !== void 0) {
+      resolved.requestsPerMinute = globalConfig["requests-per-minute"];
+    }
+    if (globalConfig["tokens-per-minute"] !== void 0) {
+      resolved.tokensPerMinute = globalConfig["tokens-per-minute"];
+    }
+    if (globalConfig["tokens-per-day"] !== void 0) {
+      resolved.tokensPerDay = globalConfig["tokens-per-day"];
+    }
+    if (globalConfig["safety-margin"] !== void 0) {
+      resolved.safetyMargin = globalConfig["safety-margin"];
+    }
+    if (globalConfig.enabled !== void 0) {
+      resolved.enabled = globalConfig.enabled;
+    }
+  }
+  if (profileConfig) {
+    if (profileConfig["requests-per-minute"] !== void 0) {
+      resolved.requestsPerMinute = profileConfig["requests-per-minute"];
+    }
+    if (profileConfig["tokens-per-minute"] !== void 0) {
+      resolved.tokensPerMinute = profileConfig["tokens-per-minute"];
+    }
+    if (profileConfig["tokens-per-day"] !== void 0) {
+      resolved.tokensPerDay = profileConfig["tokens-per-day"];
+    }
+    if (profileConfig["safety-margin"] !== void 0) {
+      resolved.safetyMargin = profileConfig["safety-margin"];
+    }
+    if (profileConfig.enabled !== void 0) {
+      resolved.enabled = profileConfig.enabled;
+    }
+  }
+  if (options.rateLimitRpm !== void 0) {
+    resolved.requestsPerMinute = options.rateLimitRpm;
+  }
+  if (options.rateLimitTpm !== void 0) {
+    resolved.tokensPerMinute = options.rateLimitTpm;
+  }
+  if (options.rateLimitDaily !== void 0) {
+    resolved.tokensPerDay = options.rateLimitDaily;
+  }
+  if (options.rateLimitSafetyMargin !== void 0) {
+    resolved.safetyMargin = options.rateLimitSafetyMargin;
+  }
+  const hasLimits = resolved.requestsPerMinute !== void 0 || resolved.tokensPerMinute !== void 0 || resolved.tokensPerDay !== void 0;
+  if (resolved.enabled === false) {
+    return { enabled: false, safetyMargin: resolved.safetyMargin ?? 0.8 };
+  }
+  if (!hasLimits) {
+    return void 0;
+  }
+  return {
+    ...resolved,
+    enabled: true,
+    safetyMargin: resolved.safetyMargin ?? 0.8
+  };
+}
+var DEFAULT_RETRY_CONFIG = {
+  enabled: true,
+  retries: 3,
+  minTimeout: 1e3,
+  maxTimeout: 3e4,
+  factor: 2,
+  randomize: true,
+  respectRetryAfter: true,
+  maxRetryAfterMs: 12e4
+};
+function resolveRetryConfig(options, globalConfig, profileConfig) {
+  const resolved = { ...DEFAULT_RETRY_CONFIG };
+  if (globalConfig) {
+    if (globalConfig.enabled !== void 0) resolved.enabled = globalConfig.enabled;
+    if (globalConfig.retries !== void 0) resolved.retries = globalConfig.retries;
+    if (globalConfig["min-timeout"] !== void 0)
+      resolved.minTimeout = globalConfig["min-timeout"];
+    if (globalConfig["max-timeout"] !== void 0)
+      resolved.maxTimeout = globalConfig["max-timeout"];
+    if (globalConfig.factor !== void 0) resolved.factor = globalConfig.factor;
+    if (globalConfig.randomize !== void 0) resolved.randomize = globalConfig.randomize;
+    if (globalConfig["respect-retry-after"] !== void 0) {
+      resolved.respectRetryAfter = globalConfig["respect-retry-after"];
+    }
+    if (globalConfig["max-retry-after-ms"] !== void 0) {
+      resolved.maxRetryAfterMs = globalConfig["max-retry-after-ms"];
+    }
+  }
+  if (profileConfig) {
+    if (profileConfig.enabled !== void 0) resolved.enabled = profileConfig.enabled;
+    if (profileConfig.retries !== void 0) resolved.retries = profileConfig.retries;
+    if (profileConfig["min-timeout"] !== void 0)
+      resolved.minTimeout = profileConfig["min-timeout"];
+    if (profileConfig["max-timeout"] !== void 0)
+      resolved.maxTimeout = profileConfig["max-timeout"];
+    if (profileConfig.factor !== void 0) resolved.factor = profileConfig.factor;
+    if (profileConfig.randomize !== void 0) resolved.randomize = profileConfig.randomize;
+    if (profileConfig["respect-retry-after"] !== void 0) {
+      resolved.respectRetryAfter = profileConfig["respect-retry-after"];
+    }
+    if (profileConfig["max-retry-after-ms"] !== void 0) {
+      resolved.maxRetryAfterMs = profileConfig["max-retry-after-ms"];
+    }
+  }
+  if (options.maxRetries !== void 0) {
+    resolved.retries = options.maxRetries;
+  }
+  if (options.retryMinTimeout !== void 0) {
+    resolved.minTimeout = options.retryMinTimeout;
+  }
+  if (options.retryMaxTimeout !== void 0) {
+    resolved.maxTimeout = options.retryMaxTimeout;
+  }
+  if (options.noRetry === true) {
+    resolved.enabled = false;
+  }
+  return resolved;
+}
+
 // src/subagent-config.ts
 var INHERIT_MODEL = "inherit";
 function resolveSubagentConfig(subagentName, parentModel, profileConfig, globalConfig) {
@@ -4142,6 +4577,33 @@ var BlockRenderer = class _BlockRenderer {
     this.rebuildBlocks();
     return id;
   }
+  /**
+   * Add a system message block (for rate limiting, retry notifications, etc.).
+   *
+   * Displays immediately with an icon and color based on category.
+   * Non-selectable like text blocks.
+   *
+   * @param message - The system message text
+   * @param category - Message category for styling
+   * @returns The block ID
+   */
+  addSystemMessage(message, category) {
+    const id = this.generateId("system");
+    const node = {
+      id,
+      type: "system_message",
+      depth: 0,
+      parentId: null,
+      sessionId: this.currentSessionId,
+      message,
+      category,
+      children: []
+    };
+    this.nodes.set(id, node);
+    this.rootIds.push(id);
+    this.rebuildBlocks();
+    return id;
+  }
   /**
    * Add a user message block (for REPL mid-session input).
    *
@@ -4524,6 +4986,50 @@ ${fullContent}
         }
         return this.abbreviateToLines(fullContent, 2, selected);
       }
+      case "system_message": {
+        const icon = this.getSystemMessageIcon(node.category);
+        const color = this.getSystemMessageColor(node.category);
+        const RESET2 = "\x1B[0m";
+        return `${indent}${color}${icon} ${node.message}${RESET2}`;
+      }
+    }
+  }
+  /**
+   * Get icon for system message category.
+   */
+  getSystemMessageIcon(category) {
+    switch (category) {
+      case "throttle":
+        return "\u23F8";
+      case "retry":
+        return "\u{1F504}";
+      case "info":
+        return "\u2139\uFE0F";
+      case "warning":
+        return "\u26A0\uFE0F";
+      case "error":
+        return "\u274C";
+    }
+  }
+  /**
+   * Get ANSI color code for system message category.
+   */
+  getSystemMessageColor(category) {
+    const YELLOW2 = "\x1B[33m";
+    const BLUE = "\x1B[34m";
+    const GRAY = "\x1B[90m";
+    const RED2 = "\x1B[31m";
+    switch (category) {
+      case "throttle":
+        return YELLOW2;
+      case "retry":
+        return BLUE;
+      case "info":
+        return GRAY;
+      case "warning":
+        return YELLOW2;
+      case "error":
+        return RED2;
     }
   }
   /**
@@ -6268,6 +6774,10 @@ var StatusBar = class {
   nodeIdToGadgetName = /* @__PURE__ */ new Map();
   /** Tree subscription unsubscribe function */
   treeUnsubscribe = null;
+  /** Rate limiting state */
+  rateLimitState = null;
+  /** Retry state */
+  retryState = null;
   constructor(statusBox, model, renderCallback, renderNowCallback) {
     this.statusBox = statusBox;
     this.renderCallback = renderCallback;
@@ -6407,6 +6917,37 @@ var StatusBar = class {
     this.stopSpinner();
     this.render();
   }
+  /**
+   * Show rate limiting throttle indicator.
+   * @param delayMs - Delay in milliseconds before next request
+   */
+  showThrottling(delayMs) {
+    this.rateLimitState = { isThrottling: true, delayMs };
+    this.render(true);
+  }
+  /**
+   * Clear rate limiting throttle indicator.
+   */
+  clearThrottling() {
+    this.rateLimitState = null;
+    this.render(true);
+  }
+  /**
+   * Show retry attempt indicator.
+   * @param attemptNumber - Current attempt number (1-based)
+   * @param retriesLeft - Number of retries remaining after this attempt
+   */
+  showRetry(attemptNumber, retriesLeft) {
+    this.retryState = { attemptNumber, retriesLeft };
+    this.render(true);
+  }
+  /**
+   * Clear retry attempt indicator.
+   */
+  clearRetry() {
+    this.retryState = null;
+    this.render(true);
+  }
   // ─────────────────────────────────────────────────────────────────────────────
   // Tree Subscription (for tree-only block creation)
   // ─────────────────────────────────────────────────────────────────────────────
@@ -6634,6 +7175,15 @@ var StatusBar = class {
       const typeStr = debug.nodeType ? ` [${debug.nodeType}]` : "";
       parts.push(`${GRAY}${debugStr}${typeStr}${RESET2}`);
     }
+    if (this.rateLimitState?.isThrottling) {
+      const seconds = Math.ceil(this.rateLimitState.delayMs / 1e3);
+      parts.push(`${YELLOW2}\u23F8 Throttled ${seconds}s${RESET2}`);
+    }
+    if (this.retryState) {
+      const { attemptNumber, retriesLeft } = this.retryState;
+      const totalAttempts = attemptNumber + retriesLeft;
+      parts.push(`${BLUE}\u{1F504} Retry ${attemptNumber}/${totalAttempts}${RESET2}`);
+    }
     if (this.activeLLMCalls.size > 0 || this.activeGadgets.size > 0) {
       const spinner = SPINNER_FRAMES2[this.spinnerFrame];
       if (this.activeLLMCalls.size > 0) {
@@ -7018,6 +7568,42 @@ var TUIApp = class _TUIApp {
   flushText() {
     this.statusBar.clearActivity();
   }
+  /**
+   * Show rate limiting throttle indicator in status bar.
+   * @param delayMs - Delay in milliseconds before next request
+   */
+  showThrottling(delayMs) {
+    this.statusBar.showThrottling(delayMs);
+  }
+  /**
+   * Clear rate limiting throttle indicator from status bar.
+   */
+  clearThrottling() {
+    this.statusBar.clearThrottling();
+  }
+  /**
+   * Show retry attempt indicator in status bar.
+   * @param attemptNumber - Current attempt number (1-based)
+   * @param retriesLeft - Number of retries remaining after this attempt
+   */
+  showRetry(attemptNumber, retriesLeft) {
+    this.statusBar.showRetry(attemptNumber, retriesLeft);
+  }
+  /**
+   * Clear retry attempt indicator from status bar.
+   */
+  clearRetry() {
+    this.statusBar.clearRetry();
+  }
+  /**
+   * Add a system message to the conversation (for rate limiting, retry notifications, etc.).
+   * @param message - The system message text
+   * @param category - Message category for styling
+   * @returns The block ID
+   */
+  addSystemMessage(message, category) {
+    return this.blockRenderer.addSystemMessage(message, category);
+  }
   // ─────────────────────────────────────────────────────────────────────────────
   // Profile Management
   // ─────────────────────────────────────────────────────────────────────────────
@@ -7223,7 +7809,7 @@ async function executeAgent(promptArg, options, env, commandName) {
     }
   }
   const abortController = new AbortController();
-  let wasCancelled = false;
+  let _wasCancelled = false;
   const handleQuit = () => {
     if (tui) {
       tui.destroy();
@@ -7233,7 +7819,7 @@ async function executeAgent(promptArg, options, env, commandName) {
   if (tui) {
     tui.onQuit(handleQuit);
     tui.onCancel(() => {
-      wasCancelled = true;
+      _wasCancelled = true;
       abortController.abort();
     });
   }
@@ -7258,12 +7844,12 @@ async function executeAgent(promptArg, options, env, commandName) {
     gadgetApprovals,
     defaultMode: "allowed"
   };
-  let usage;
+  let _usage;
   let iterations = 0;
   const llmLogsEnabled = options.logLlmRequests === true;
   const llmLogDir = llmLogsEnabled ? env.session?.logDir : void 0;
   let llmCallCounter = 0;
-  const countGadgetOutputTokens = async (output) => {
+  const _countGadgetOutputTokens = async (output) => {
     if (!output) return void 0;
     try {
       const messages = [{ role: "assistant", content: output }];
@@ -7306,12 +7892,46 @@ async function executeAgent(promptArg, options, env, commandName) {
       // onLLMCallComplete: Capture metadata for final summary and file logging
       onLLMCallComplete: async (context) => {
         if (context.subagentContext) return;
-        usage = context.usage;
+        _usage = context.usage;
         iterations = Math.max(iterations, context.iteration + 1);
         if (llmLogDir) {
           const filename = `${formatCallNumber(llmCallCounter)}.response`;
           await writeLogFile(llmLogDir, filename, context.rawResponse);
         }
+        if (tui) {
+          tui.clearRetry();
+        }
+      },
+      // onRateLimitThrottle: Show throttling delay in status bar and conversation
+      onRateLimitThrottle: async (context) => {
+        if (context.subagentContext) return;
+        if (tui) {
+          const seconds = Math.ceil(context.delayMs / 1e3);
+          tui.showThrottling(context.delayMs);
+          const statsMsg = [];
+          if (context.stats.rpm > 0) statsMsg.push(`${context.stats.rpm} RPM`);
+          if (context.stats.tpm > 0)
+            statsMsg.push(`${Math.round(context.stats.tpm / 1e3)}K TPM`);
+          const statsStr = statsMsg.length > 0 ? ` (${statsMsg.join(", ")})` : "";
+          tui.addSystemMessage(
+            `Rate limit approaching${statsStr}, waiting ${seconds}s...`,
+            "throttle"
+          );
+          setTimeout(() => tui.clearThrottling(), context.delayMs);
+        }
+      },
+      // onRetryAttempt: Show retry attempt in status bar and conversation
+      onRetryAttempt: async (context) => {
+        if (context.subagentContext) return;
+        if (tui) {
+          const totalAttempts = context.attemptNumber + context.retriesLeft;
+          tui.showRetry(context.attemptNumber, context.retriesLeft);
+          const retryAfterInfo = context.retryAfterMs ? ` (server requested ${Math.ceil(context.retryAfterMs / 1e3)}s wait)` : "";
+          tui.addSystemMessage(
+            `Request failed (attempt ${context.attemptNumber}/${totalAttempts}), retrying...${retryAfterInfo}`,
+            "retry"
+          );
+        }
       }
     },
     // SHOWCASE: Controller-based approval gating for gadgets
@@ -7380,6 +8000,19 @@ ${ctx.gadgetName} requires interactive approval. Run in a terminal to approve.`
       }
     }
   });
+  const rateLimitConfig = resolveRateLimitConfig(
+    options,
+    options.globalRateLimits,
+    options.profileRateLimits,
+    options.model
+  );
+  const retryConfig = resolveRetryConfig(options, options.globalRetry, options.profileRetry);
+  if (rateLimitConfig) {
+    builder.withRateLimits(rateLimitConfig);
+  }
+  if (retryConfig) {
+    builder.withRetry(retryConfig);
+  }
   if (options.system) {
     builder.withSystem(options.system);
   }
@@ -7530,7 +8163,7 @@ ${ctx.gadgetName} requires interactive approval. Run in a terminal to approve.`
     }
   }
 }
-function registerAgentCommand(program, env, config, globalSubagents) {
+function registerAgentCommand(program, env, config, globalSubagents, globalRateLimits, globalRetry) {
   const cmd = program.command(COMMANDS.agent).description("Run the llmist agent loop with optional gadgets.").argument("[prompt]", "Prompt for the agent loop. Falls back to stdin when available.");
   addAgentOptions(cmd, config);
   cmd.action(
@@ -7540,7 +8173,11 @@ function registerAgentCommand(program, env, config, globalSubagents) {
         gadgetApproval: config?.["gadget-approval"],
         subagents: config?.subagents,
         globalSubagents,
-        initialGadgets: config?.["initial-gadgets"]
+        initialGadgets: config?.["initial-gadgets"],
+        globalRateLimits,
+        globalRetry,
+        profileRateLimits: config?.["rate-limits"],
+        profileRetry: config?.retry
       };
       return executeAgent(prompt, mergedOptions, env, "agent");
     }, env)
@@ -7548,14 +8185,11 @@ function registerAgentCommand(program, env, config, globalSubagents) {
 }
 
 // src/complete-command.ts
-import { text as text2 } from "llmist";
-import { LLMMessageBuilder } from "llmist";
-import { resolveModel } from "llmist";
-import { FALLBACK_CHARS_PER_TOKEN as FALLBACK_CHARS_PER_TOKEN2 } from "llmist";
+import { FALLBACK_CHARS_PER_TOKEN as FALLBACK_CHARS_PER_TOKEN2, LLMMessageBuilder, resolveModel as resolveModel2, text as text2 } from "llmist";
 async function executeComplete(promptArg, options, env) {
   const prompt = await resolvePrompt(promptArg, env);
   const client = env.createClient();
-  const model = resolveModel(options.model);
+  const model = resolveModel2(options.model);
   const builder = new LLMMessageBuilder();
   if (options.system) {
     builder.addSystem(options.system);
@@ -7629,11 +8263,18 @@ async function executeComplete(promptArg, options, env) {
     }
   }
 }
-function registerCompleteCommand(program, env, config) {
+function registerCompleteCommand(program, env, config, globalRateLimits, globalRetry) {
   const cmd = program.command(COMMANDS.complete).description("Stream a single completion from a specified model.").argument("[prompt]", "Prompt to send to the LLM. If omitted, stdin is used when available.");
   addCompleteOptions(cmd, config);
   cmd.action(
-    (prompt, options) => executeAction(() => executeComplete(prompt, options, env), env)
+    (prompt, options) => executeAction(() => {
+      const mergedOptions = {
+        ...options,
+        globalRateLimits,
+        globalRetry
+      };
+      return executeComplete(prompt, mergedOptions, env);
+    }, env)
   );
 }
 
@@ -7760,104 +8401,6 @@ System Prompt (${chars.toLocaleString()} chars, ${lines} lines):
   }
 }
 
-// src/init-command.ts
-import { existsSync as existsSync3, mkdirSync, writeFileSync } from "fs";
-import { dirname as dirname2 } from "path";
-var STARTER_CONFIG = `# ~/.llmist/cli.toml
-# llmist CLI configuration file
-#
-# This is a minimal starter config. For a comprehensive example with all options:
-#   https://github.com/zbigniewsobiecki/llmist/blob/main/examples/cli.example.toml
-#
-# Key concepts:
-#   - Any section can inherit from others using: inherits = "section-name"
-#   - Prompts can use templates with Eta syntax: <%~ include("@prompt-name") %>
-#   - Custom sections become CLI commands: [my-command] -> llmist my-command
-
-#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
-# GLOBAL OPTIONS
-# These apply to all commands. CLI flags override these settings.
-#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
-[global]
-# log-level = "info"              # silly, trace, debug, info, warn, error, fatal
-# log-file = "/tmp/llmist.log"    # Enable file logging (JSON format)
-
-#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
-# COMPLETE COMMAND DEFAULTS
-# For single LLM responses: llmist complete "prompt"
-# Model format: provider:model (e.g., openai:gpt-4o, anthropic:claude-sonnet-4-5)
-#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
-[complete]
-# model = "openai:gpt-4o"
-# temperature = 0.7               # 0-2, higher = more creative
-# max-tokens = 4096               # Maximum response length
-
-#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
-# AGENT COMMAND DEFAULTS
-# For tool-using agents: llmist agent "prompt"
-#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
-[agent]
-# model = "anthropic:claude-sonnet-4-5"
-# max-iterations = 15             # Max tool-use loops before stopping
-# gadgets = [                     # Tools the agent can use
-#   "ListDirectory",
-#   "ReadFile",
-#   "WriteFile",
-# ]
-
-#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
-# CUSTOM COMMANDS
-# Any other section becomes a new CLI command!
-# Uncomment below to create: llmist summarize "your text"
-#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
-# [summarize]
-# type = "complete"               # "complete" or "agent"
-# description = "Summarize text concisely."
-# system = "Summarize the following text in 2-3 bullet points."
-# temperature = 0.3
-`;
-async function executeInit(_options, env) {
-  const configPath = getConfigPath();
-  const configDir = dirname2(configPath);
-  if (existsSync3(configPath)) {
-    env.stderr.write(`Configuration already exists at ${configPath}
-`);
-    env.stderr.write("\n");
-    env.stderr.write(`To view it:  cat ${configPath}
-`);
-    env.stderr.write(`To reset:    rm ${configPath} && llmist init
-`);
-    return;
-  }
-  if (!existsSync3(configDir)) {
-    mkdirSync(configDir, { recursive: true });
-  }
-  writeFileSync(configPath, STARTER_CONFIG, "utf-8");
-  env.stderr.write(`Created ${configPath}
-`);
-  env.stderr.write("\n");
-  env.stderr.write("Next steps:\n");
-  env.stderr.write("  1. Set your API key:\n");
-  env.stderr.write("       export OPENAI_API_KEY=sk-...\n");
-  env.stderr.write("       export ANTHROPIC_API_KEY=sk-...\n");
-  env.stderr.write("       export GEMINI_API_KEY=...\n");
-  env.stderr.write("\n");
-  env.stderr.write(`  2. Customize your config:
-`);
-  env.stderr.write(`       $EDITOR ${configPath}
-`);
-  env.stderr.write("\n");
-  env.stderr.write("  3. See all options:\n");
-  env.stderr.write(
-    "       https://github.com/zbigniewsobiecki/llmist/blob/main/examples/cli.example.toml\n"
-  );
-  env.stderr.write("\n");
-  env.stderr.write('Try it: llmist complete "Hello, world!"\n');
-}
-function registerInitCommand(program, env) {
-  program.command(COMMANDS.init).description("Initialize llmist configuration at ~/.llmist/cli.toml").action((options) => executeAction(() => executeInit(options, env), env));
-}
-
 // src/environment.ts
 import { join as join3 } from "path";
 import readline from "readline";
@@ -7956,7 +8499,7 @@ function createCommandEnvironment(baseEnv, config) {
     createLogger: createLoggerFactory(loggerConfig, baseEnv.session?.logDir)
   };
 }
-function registerCustomCommand(program, name, config, env, globalSubagents) {
+function registerCustomCommand(program, name, config, env, globalSubagents, globalRateLimits, globalRetry) {
   const type = config.type ?? "agent";
   const description = config.description ?? `Custom ${type} command`;
   const cmd = program.command(name).description(description).argument("[prompt]", "Prompt for the command. Falls back to stdin when available.");
@@ -7968,7 +8511,9 @@ function registerCustomCommand(program, name, config, env, globalSubagents) {
         const configDefaults = configToCompleteOptions(config);
         const options = {
           ...configDefaults,
-          ...cliOptions
+          ...cliOptions,
+          globalRateLimits,
+          globalRetry
         };
         await executeComplete(prompt, options, cmdEnv);
       }, cmdEnv);
@@ -7982,7 +8527,9 @@ function registerCustomCommand(program, name, config, env, globalSubagents) {
         const options = {
           ...configDefaults,
           ...cliOptions,
-          globalSubagents
+          globalSubagents,
+          globalRateLimits,
+          globalRetry
         };
         await executeAgent(prompt, options, cmdEnv, name);
       }, cmdEnv);
@@ -8425,7 +8972,7 @@ function registerGadgetCommand(program, env) {
 }
 
 // src/image-command.ts
-import { writeFileSync as writeFileSync2 } from "fs";
+import { writeFileSync } from "fs";
 var DEFAULT_IMAGE_MODEL = "dall-e-3";
 async function executeImage(promptArg, options, env) {
   const prompt = await resolvePrompt(promptArg, env);
@@ -8449,7 +8996,7 @@ async function executeImage(promptArg, options, env) {
     const imageData = result.images[0];
     if (imageData.b64Json) {
       const buffer = Buffer.from(imageData.b64Json, "base64");
-      writeFileSync2(options.output, buffer);
+      writeFileSync(options.output, buffer);
       if (!options.quiet) {
         env.stderr.write(`${SUMMARY_PREFIX} Image saved to ${options.output}
 `);
@@ -8487,6 +9034,104 @@ function registerImageCommand(program, env, config) {
   );
 }
 
+// src/init-command.ts
+import { existsSync as existsSync3, mkdirSync, writeFileSync as writeFileSync2 } from "fs";
+import { dirname as dirname2 } from "path";
+var STARTER_CONFIG = `# ~/.llmist/cli.toml
+# llmist CLI configuration file
+#
+# This is a minimal starter config. For a comprehensive example with all options:
+#   https://github.com/zbigniewsobiecki/llmist/blob/main/examples/cli.example.toml
+#
+# Key concepts:
+#   - Any section can inherit from others using: inherits = "section-name"
+#   - Prompts can use templates with Eta syntax: <%~ include("@prompt-name") %>
+#   - Custom sections become CLI commands: [my-command] -> llmist my-command
+
+#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+# GLOBAL OPTIONS
+# These apply to all commands. CLI flags override these settings.
+#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+[global]
+# log-level = "info"              # silly, trace, debug, info, warn, error, fatal
+# log-file = "/tmp/llmist.log"    # Enable file logging (JSON format)
+
+#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+# COMPLETE COMMAND DEFAULTS
+# For single LLM responses: llmist complete "prompt"
+# Model format: provider:model (e.g., openai:gpt-4o, anthropic:claude-sonnet-4-5)
+#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+[complete]
+# model = "openai:gpt-4o"
+# temperature = 0.7               # 0-2, higher = more creative
+# max-tokens = 4096               # Maximum response length
+
+#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+# AGENT COMMAND DEFAULTS
+# For tool-using agents: llmist agent "prompt"
+#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+[agent]
+# model = "anthropic:claude-sonnet-4-5"
+# max-iterations = 15             # Max tool-use loops before stopping
+# gadgets = [                     # Tools the agent can use
+#   "ListDirectory",
+#   "ReadFile",
+#   "WriteFile",
+# ]
+
+#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+# CUSTOM COMMANDS
+# Any other section becomes a new CLI command!
+# Uncomment below to create: llmist summarize "your text"
+#\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500
+# [summarize]
+# type = "complete"               # "complete" or "agent"
+# description = "Summarize text concisely."
+# system = "Summarize the following text in 2-3 bullet points."
+# temperature = 0.3
+`;
+async function executeInit(_options, env) {
+  const configPath = getConfigPath();
+  const configDir = dirname2(configPath);
+  if (existsSync3(configPath)) {
+    env.stderr.write(`Configuration already exists at ${configPath}
+`);
+    env.stderr.write("\n");
+    env.stderr.write(`To view it:  cat ${configPath}
+`);
+    env.stderr.write(`To reset:    rm ${configPath} && llmist init
+`);
+    return;
+  }
+  if (!existsSync3(configDir)) {
+    mkdirSync(configDir, { recursive: true });
+  }
+  writeFileSync2(configPath, STARTER_CONFIG, "utf-8");
+  env.stderr.write(`Created ${configPath}
+`);
+  env.stderr.write("\n");
+  env.stderr.write("Next steps:\n");
+  env.stderr.write("  1. Set your API key:\n");
+  env.stderr.write("       export OPENAI_API_KEY=sk-...\n");
+  env.stderr.write("       export ANTHROPIC_API_KEY=sk-...\n");
+  env.stderr.write("       export GEMINI_API_KEY=...\n");
+  env.stderr.write("\n");
+  env.stderr.write(`  2. Customize your config:
+`);
+  env.stderr.write(`       $EDITOR ${configPath}
+`);
+  env.stderr.write("\n");
+  env.stderr.write("  3. See all options:\n");
+  env.stderr.write(
+    "       https://github.com/zbigniewsobiecki/llmist/blob/main/examples/cli.example.toml\n"
+  );
+  env.stderr.write("\n");
+  env.stderr.write('Try it: llmist complete "Hello, world!"\n');
+}
+function registerInitCommand(program, env) {
+  program.command(COMMANDS.init).description("Initialize llmist configuration at ~/.llmist/cli.toml").action((options) => executeAction(() => executeInit(options, env), env));
+}
+
 // src/models-command.ts
 import chalk7 from "chalk";
 import { MODEL_ALIASES } from "llmist";
@@ -9105,10 +9750,10 @@ function registerSpeechCommand(program, env, config) {
 }
 
 // src/vision-command.ts
-import { resolveModel as resolveModel2 } from "llmist";
+import { resolveModel as resolveModel3 } from "llmist";
 async function executeVision(imagePath, options, env) {
   const client = env.createClient();
-  const model = resolveModel2(options.model);
+  const model = resolveModel3(options.model);
   const imageBuffer = await readFileBuffer(imagePath);
   const prompt = options.prompt ?? "Describe this image in detail.";
   const stderrTTY = env.stderr.isTTY === true;
@@ -9154,8 +9799,15 @@ function createProgram(env, config) {
     writeOut: (str) => env.stdout.write(str),
     writeErr: (str) => env.stderr.write(str)
   });
-  registerCompleteCommand(program, env, config?.complete);
-  registerAgentCommand(program, env, config?.agent, config?.subagents);
+  registerCompleteCommand(program, env, config?.complete, config?.["rate-limits"], config?.retry);
+  registerAgentCommand(
+    program,
+    env,
+    config?.agent,
+    config?.subagents,
+    config?.["rate-limits"],
+    config?.retry
+  );
   registerImageCommand(program, env, config?.image);
   registerSpeechCommand(program, env, config?.speech);
   registerVisionCommand(program, env);
@@ -9167,7 +9819,15 @@ function createProgram(env, config) {
     const customNames = getCustomCommandNames(config);
     for (const name of customNames) {
       const cmdConfig = config[name];
-      registerCustomCommand(program, name, cmdConfig, env, config.subagents);
+      registerCustomCommand(
+        program,
+        name,
+        cmdConfig,
+        env,
+        config.subagents,
+        config["rate-limits"],
+        config.retry
+      );
     }
   }
   return program;