@kernelonpanic/kitcode 1.4.2 → 1.4.3

package/dist/index.js

@@ -170,7 +170,10 @@ function describeProviderFailure(error, providerId, knownSecrets = []) {
 }
 function statusOf(error) {
   const value = prop(error, "status") ?? prop(error, "statusCode");
-  return typeof value === "number" && value >= 100 ? value : void 0;
+  if (typeof value === "number" && value >= 100) return value;
+  const code = prop(error, "code") ?? prop(prop(error, "error"), "code");
+  const numeric = typeof code === "string" && /^\d{3}$/.test(code) ? Number(code) : code;
+  return typeof numeric === "number" && numeric >= 400 && numeric <= 599 ? numeric : void 0;
 }
 function isUserAbort(error, signal) {
   return signal?.aborted === true || abortKind(error) === "user";
@@ -1440,13 +1443,20 @@ function resolveEffort(effort, messages) {
   return text.length > 500 || HARD_TASK.test(text) ? "high" : "medium";
 }
 var rejected = /* @__PURE__ */ new Set();
+var capped = /* @__PURE__ */ new Map();
+var RANK = { low: 0, medium: 1, high: 2, xhigh: 3, max: 4 };
 function markEffortRejected(providerId, model) {
   rejected.add(`${providerId}/${model}`);
 }
+function markEffortCapped(providerId, model, level) {
+  capped.set(`${providerId}/${model}`, level);
+}
 function effectiveEffort(kind, model, effort, messages, providerId) {
   if (providerId !== void 0 && rejected.has(`${providerId}/${model}`)) return void 0;
   const requested = resolveEffort(effort, messages);
-  return kind === "openai" ? openAiEffort(model, requested) : anthropicEffort(model, requested);
+  const level = kind === "openai" ? openAiEffort(model, requested) : anthropicEffort(model, requested);
+  const cap = providerId === void 0 ? void 0 : capped.get(`${providerId}/${model}`);
+  return level && cap && RANK[level] > RANK[cap] ? cap : level;
 }
 function anthropicEffort(model, effort) {
   if (!effort) return void 0;
@@ -2137,18 +2147,18 @@ function sleep(ms, signal) {
     signal.addEventListener("abort", onAbort, { once: true });
   });
 }
+var CHARS_PER_TOKEN = 4;
 function estimateRequestTokens(cfg, messages) {
   let characters3 = cfg.system.length;
   try {
     characters3 += JSON.stringify(
       messages,
-      (_key, value) => typeof value === "object" && value !== null && value.type === "image" ? `[image:${"#".repeat(2e3)}]` : value
+      (_key, value) => typeof value === "object" && value !== null && value.type === "image" ? `[image:${"#".repeat(6e3)}]` : value
     ).length;
     characters3 += JSON.stringify(cfg.tools.schemas()).length;
   } catch {
-    return Math.max(1, characters3);
   }
-  return Math.max(1, characters3);
+  return Math.max(1, Math.ceil(characters3 / CHARS_PER_TOKEN));
 }
 async function runToolCalls(cfg, content, hooks, signal) {
   const calls = content.filter((block) => block.type === "tool_use");
@@ -3789,6 +3799,7 @@ function withoutThinkingReplay(history) {
   });
   return changed ? out : history;
 }
+var CHARS_PER_TOKEN2 = 4;
 function estimateHistoryTokens(history) {
   let chars = 0;
   for (const message of history) {
@@ -3800,7 +3811,7 @@ function estimateHistoryTokens(history) {
       else if (block.type === "image") chars += 6e3;
     }
   }
-  return Math.round(chars / 4);
+  return Math.round(chars / CHARS_PER_TOKEN2);
 }
 function shouldAutoCompact(used, window) {
   return window !== null && window > 0 && used / window >= 0.8;
@@ -3820,7 +3831,9 @@ function estimateCompactBudget(history, maxTokens = MERGE_MAX_TOKENS) {
   const rendered = renderedMessages.join("\n\n");
   if (rendered.length <= SINGLE_REQUEST_MAX_CHARS) {
     return {
-      inputTokens: Math.max(1, rendered.length + 1e3),
+      // Tokens, not characters: counting characters made a turn budget refuse
+      // compaction (four times too expensive) just when the window was full.
+      inputTokens: Math.max(1, Math.ceil(rendered.length / CHARS_PER_TOKEN2) + 1e3),
       maxOutputTokens: Math.max(1, Math.min(SINGLE_MAX_TOKENS, Math.floor(maxTokens)))
     };
   }
@@ -3838,7 +3851,7 @@ function estimateCompactBudget(history, maxTokens = MERGE_MAX_TOKENS) {
   const intermediateOutput = chunks.length * chunkOutput + Math.max(0, mergeRequests - 1) * mergeOutput;
   const requestCount = chunks.length + mergeRequests;
   return {
-    inputTokens: Math.max(1, chunkInput + intermediateOutput * 8 + requestCount * 1e3),
+    inputTokens: Math.max(1, Math.ceil(chunkInput / CHARS_PER_TOKEN2) + intermediateOutput * 2 + requestCount * 1e3),
     maxOutputTokens: totalOutput
   };
 }
@@ -4400,7 +4413,7 @@ function clip(value) {
 // package.json
 var package_default = {
   name: "@kernelonpanic/kitcode",
-  version: "1.4.2",
+  version: "1.4.3",
   description: "Terminal coding agent with a config you never have to write by hand",
   type: "module",
   license: "MIT",
@@ -4468,7 +4481,22 @@ var package_default = {
 
 // src/version.ts
 var KITCODE_VERSION = package_default.version;
-var KITCODE_COMMIT = true ? "6b1ec0a742c44892447e71f477c068d8ba514c6a" : "development";
+var KITCODE_COMMIT = true ? "586eb4def64a8ef824bf2fe2f58a9f71e18090f0" : "development";
+function kitcodeUserAgent(sdkUserAgent2) {
+  return `KitCode/${KITCODE_VERSION} ${sdkUserAgent2}`.trim();
+}
+function sdkUserAgent(client) {
+  const read = client?.getUserAgent;
+  try {
+    return typeof read === "function" ? String(read.call(client)) : "";
+  } catch {
+    return "";
+  }
+}
+function withKitcodeUserAgent(sdkUserAgent2, configured = {}) {
+  const overridden = Object.keys(configured).some((name) => name.toLowerCase() === "user-agent");
+  return overridden ? { ...configured } : { "User-Agent": kitcodeUserAgent(sdkUserAgent2), ...configured };
+}
 
 // src/mcp/client.ts
 var clientInfo = { name: "kitcode", version: KITCODE_VERSION };
@@ -5047,13 +5075,14 @@ var KNOWN_MODELS = [
   knownModel("claude-fable-5", "Claude Fable 5", 1e6, 128e3)
 ];
 function createAnthropicProvider(args) {
-  const client = new Anthropic2({
+  const options = {
     apiKey: args.apiKey,
     baseURL: args.baseUrl,
-    defaultHeaders: args.headers,
     maxRetries: MAX_RETRIES,
     timeout: args.timeoutMs ?? REQUEST_TIMEOUT_MS
-  });
+  };
+  const userAgent = sdkUserAgent(new Anthropic2(options));
+  const client = new Anthropic2({ ...options, defaultHeaders: withKitcodeUserAgent(userAgent, args.headers) });
   return {
     id: args.id,
     kind: "anthropic",
@@ -5062,7 +5091,35 @@ function createAnthropicProvider(args) {
     knownModels: () => KNOWN_MODELS
   };
 }
+var thinkingRejected = /* @__PURE__ */ new Set();
 async function* streamTurn(client, providerId, apiKey2, req) {
+  const key = `${providerId}/${req.model}`;
+  let thinking = req.thinking === true && !thinkingRejected.has(key);
+  for (; ; ) {
+    let streamed = false;
+    try {
+      for await (const event of streamOnce(client, providerId, apiKey2, req, thinking)) {
+        if (event.type !== "rate_limits" && event.type !== "usage") streamed = true;
+        yield event;
+      }
+      return;
+    } catch (error) {
+      if (!streamed && thinking && refusesThinking(error)) {
+        thinkingRejected.add(key);
+        thinking = false;
+        continue;
+      }
+      throw error;
+    }
+  }
+}
+function refusesThinking(error) {
+  if (!(error instanceof ProviderError)) return false;
+  if (error.status !== void 0 && error.status !== 400 && error.status !== 422) return false;
+  const cause = error.cause instanceof Error ? error.cause.message : "";
+  return /thinking|budget_tokens|adaptive/i.test(`${error.message} ${cause}`);
+}
+async function* streamOnce(client, providerId, apiKey2, req, thinkingOn) {
   const params = {
     model: req.model,
     max_tokens: req.maxTokens,
@@ -5070,7 +5127,7 @@ async function* streamTurn(client, providerId, apiKey2, req) {
     messages: toMessageParams(req.messages),
     tools: toToolParams(req.tools)
   };
-  if (req.thinking) {
+  if (thinkingOn) {
     const thinking2 = thinkingParam(req.model, req.maxTokens);
     if (thinking2) params.thinking = thinking2;
   }
@@ -5232,7 +5289,7 @@ var THINKING_BUDGET = 16e3;
 var MIN_THINKING_BUDGET = 1024;
 function thinkingParam(model, maxTokens) {
   const id = (model.split("/").at(-1) ?? model).replace(/^(?:[a-z]{2}\.)?anthropic\./, "");
-  if (!BUDGET_THINKING.test(id)) return { type: "adaptive", display: "summarized" };
+  if (id.startsWith("claude-") && !BUDGET_THINKING.test(id)) return { type: "adaptive", display: "summarized" };
   const budget = Math.min(THINKING_BUDGET, maxTokens - 1);
   return budget >= MIN_THINKING_BUDGET ? { type: "enabled", budget_tokens: budget } : void 0;
 }
@@ -5252,13 +5309,14 @@ var FINISH_REASONS = {
   content_filter: "refusal"
 };
 function createOpenAiProvider(args) {
-  const client = new OpenAI2({
+  const options = {
     baseURL: args.baseUrl,
     apiKey: args.apiKey,
-    defaultHeaders: args.headers,
     maxRetries: MAX_RETRIES,
     timeout: args.timeoutMs ?? REQUEST_TIMEOUT_MS
-  });
+  };
+  const userAgent = sdkUserAgent(new OpenAI2(options));
+  const client = new OpenAI2({ ...options, defaultHeaders: withKitcodeUserAgent(userAgent, args.headers) });
   return {
     id: args.id,
     kind: "openai",
@@ -5275,7 +5333,7 @@ async function* streamTurn2(client, providerId, apiKey2, req) {
   for (; ; ) {
     let streamed = false;
     try {
-      for await (const event of streamOnce(client, providerId, apiKey2, req, effort, thinkingOff, tokenField)) {
+      for await (const event of streamOnce2(client, providerId, apiKey2, req, effort, thinkingOff, tokenField)) {
         if (event.type !== "rate_limits" && event.type !== "usage") streamed = true;
         yield event;
       }
@@ -5288,8 +5346,14 @@ async function* streamTurn2(client, providerId, apiKey2, req) {
         continue;
       }
       if (effort && rejectsField(error, /reasoning|effort/i)) {
-        markEffortRejected(providerId, req.model);
-        effort = void 0;
+        const lower = effort === "xhigh" || effort === "max" ? "high" : effort === "high" ? "medium" : void 0;
+        if (lower) {
+          markEffortCapped(providerId, req.model, lower);
+          effort = lower;
+        } else {
+          markEffortRejected(providerId, req.model);
+          effort = void 0;
+        }
         continue;
       }
       if (thinkingOff && rejectsField(error, /thinking/i)) {
@@ -5314,11 +5378,12 @@ function thinkingOffParams(model) {
   return void 0;
 }
 function rejectsField(error, field2) {
-  if (!(error instanceof ProviderError) || error.status !== 400 && error.status !== 422) return false;
+  if (!(error instanceof ProviderError)) return false;
+  if (error.status !== void 0 && error.status !== 400 && error.status !== 422) return false;
   const cause = error.cause instanceof Error ? error.cause.message : "";
   return field2.test(`${error.message} ${cause}`);
 }
-async function* streamOnce(client, providerId, apiKey2, req, effort, thinkingOff, tokenField) {
+async function* streamOnce2(client, providerId, apiKey2, req, effort, thinkingOff, tokenField) {
   let text = "";
   let thinking = "";
   let finishReason = null;
@@ -8041,7 +8106,7 @@ ${recovered.removed === 1 ? "The image was" : `${recovered.removed} images were`
 }
 function imageRejection(error, history) {
   const status = error?.status;
-  if (typeof status !== "number" || ![400, 413, 415, 422].includes(status)) return null;
+  if (status !== void 0 && (typeof status !== "number" || ![400, 413, 415, 422].includes(status))) return null;
   const message = error instanceof Error ? error.message : String(error);
   if (status !== 413 && !/image|vision|multi-?modal/i.test(message)) return null;
   const result = withoutImages(history);
@@ -8387,7 +8452,7 @@ function splitCommandLine(raw) {
 }
 
 // src/ui/speed.ts
-var CHARS_PER_TOKEN = 4;
+var CHARS_PER_TOKEN3 = 4;
 var MIN_LIVE_MS = 1e3;
 function createSpeedMeter() {
   return { streamedChars: 0, firstTokenAt: null, last: null };
@@ -8414,7 +8479,7 @@ function readSpeed(meter, busy, now = Date.now()) {
   if (busy && meter.firstTokenAt !== null && meter.streamedChars > 0) {
     const ms = now - meter.firstTokenAt;
     if (ms >= MIN_LIVE_MS) {
-      return { tokensPerSecond: meter.streamedChars / CHARS_PER_TOKEN / (ms / 1e3), live: true };
+      return { tokensPerSecond: meter.streamedChars / CHARS_PER_TOKEN3 / (ms / 1e3), live: true };
     }
   }
   return meter.last === null || !Number.isFinite(meter.last) ? null : { tokensPerSecond: meter.last, live: false };