@opencow-ai/opencow-agent-sdk 0.4.12 → 0.4.14

package/dist/cli.mjs

@@ -83882,6 +83882,48 @@ var init_proxy = __esm(() => {
   });
 });
 
+// src/providers/shared/httpObservability.ts
+function classifyModelHttpEndpoint(url3) {
+  try {
+    const pathname = new URL(url3).pathname.replace(/\/+$/, "");
+    if (pathname.endsWith("/messages/count_tokens"))
+      return "count_tokens";
+    if (pathname.endsWith("/chat/completions"))
+      return "messages";
+    if (pathname.endsWith("/messages"))
+      return "messages";
+    if (pathname.endsWith("/responses"))
+      return "responses";
+    return pathname || "unknown";
+  } catch {
+    return "unknown";
+  }
+}
+function errorToFailureReason(error41) {
+  if (error41 instanceof Error)
+    return error41.message;
+  return String(error41);
+}
+function responseBodyToFailureReason(status, body) {
+  let detail = body.trim();
+  try {
+    const parsed = JSON.parse(body);
+    const message = typeof parsed.error?.message === "string" ? parsed.error.message : typeof parsed.message === "string" ? parsed.message : undefined;
+    const code = typeof parsed.error?.code === "string" ? parsed.error.code : typeof parsed.error?.type === "string" ? parsed.error.type : undefined;
+    detail = [code, message].filter(Boolean).join(": ") || detail;
+  } catch {}
+  const compact = detail.replace(/\s+/g, " ").slice(0, 240);
+  return compact ? `HTTP ${status}: ${compact}` : `HTTP ${status}`;
+}
+function emitSdkHttp(event) {
+  try {
+    emitHttp(event);
+  } catch {}
+}
+var init_httpObservability = __esm(() => {
+  init_observer();
+});
+
 // src/providers/shared/geminiAuth.ts
 import { homedir as homedir10 } from "node:os";
 import { join as join23 } from "node:path";
@@ -84582,6 +84624,44 @@ var init_capabilities2 = __esm(() => {
   ]);
 });
 
+// src/providers/shared/requestSideChannels.ts
+function isRecord(value) {
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function parseJsonObject(raw) {
+  if (!raw)
+    return;
+  try {
+    const parsed = JSON.parse(raw);
+    return isRecord(parsed) ? parsed : undefined;
+  } catch {
+    return;
+  }
+}
+function readExplicitExtraBodyParams() {
+  return parseJsonObject(getQueryEnvVar(QUERY_ENV_KEY_EXTRA_BODY)) ?? parseJsonObject(resolveEnvVar("EXTRA_BODY"));
+}
+function mergeOpenAICompatibleExtraBodyParams(body, params) {
+  const extra = {};
+  const shouldMerge = (key, value) => value !== undefined && !(key in body);
+  const explicitExtraBody = readExplicitExtraBodyParams();
+  if (explicitExtraBody) {
+    Object.assign(extra, explicitExtraBody);
+  }
+  const nestedExtraBody = params.extra_body;
+  if (isRecord(nestedExtraBody)) {
+    Object.assign(extra, nestedExtraBody);
+  }
+  for (const [key, value] of Object.entries(extra)) {
+    if (shouldMerge(key, value))
+      body[key] = value;
+  }
+}
+var init_requestSideChannels = __esm(() => {
+  init_state2();
+  init_config3();
+});
+
 // src/providers/codex/shim.ts
 function parseSseChunk(chunk) {
   const lines = chunk.split(`
@@ -84930,6 +85010,7 @@ async function performCodexRequest(options) {
       body.top_p = options.params.top_p;
     }
   }
+  mergeOpenAICompatibleExtraBodyParams(body, options.params);
   const headers = {
     "Content-Type": "application/json",
     ...options.defaultHeaders,
@@ -85301,6 +85382,7 @@ var init_shim = __esm(() => {
   init_sdk();
   init_schema();
   init_capabilities2();
+  init_requestSideChannels();
   init_canonical();
 });
 
@@ -85423,7 +85505,7 @@ function convertContentBlocks(content) {
     return parts[0].text ?? "";
   return parts;
 }
-function convertMessages(messages, system) {
+function convertMessages(messages, system, options = {}) {
   const result = [];
   const sysText = convertSystemPrompt2(system);
   if (sysText) {
@@ -85486,7 +85568,7 @@ function convertMessages(messages, system) {
             return text || null;
           })()
         };
-        if (thinkingBlocks.length > 0) {
+        if (options.replayReasoningContent !== false && thinkingBlocks.length > 0) {
           const reasoningText = thinkingBlocks.map((b) => b.thinking ?? "").filter(Boolean).join(`
 `);
           if (reasoningText) {
@@ -85561,7 +85643,7 @@ function convertChunkUsage(usage) {
   return openaiUsageToAnthropicUsage(usage);
 }
 function toOpenAIChatReasoningEffort(effort) {
-  return effort === "xhigh" ? "high" : effort;
+  return effort;
 }
 function getOpenAIChatProviderCapabilities(model) {
   return {
@@ -85619,7 +85701,8 @@ function serializeParallelToolCalls(messages, capabilities) {
 }
 function buildOpenAIRequestBody(params, ctx) {
   const capabilities = getOpenAIChatProviderCapabilities(ctx.resolvedModel);
-  const openaiMessages = serializeParallelToolCalls(convertMessages(params.messages, params.system), capabilities);
+  const replayReasoningContent = !isGeminiTarget(ctx.resolvedModel);
+  const openaiMessages = serializeParallelToolCalls(convertMessages(params.messages, params.system, { replayReasoningContent }), capabilities);
   const body = {
     model: ctx.resolvedModel,
     messages: openaiMessages,
@@ -85670,6 +85753,7 @@ function buildOpenAIRequestBody(params, ctx) {
       }
     }
   }
+  mergeOpenAICompatibleExtraBodyParams(body, params);
   return body;
 }
 async function* openaiStreamToAnthropic(response, model) {
@@ -85996,17 +86080,26 @@ class OpenAIShimMessages {
   defaultHeaders;
   reasoningEffort;
   providerOverride;
-  constructor(defaultHeaders, reasoningEffort, providerOverride) {
+  source;
+  constructor(defaultHeaders, reasoningEffort, providerOverride, source) {
     this.defaultHeaders = defaultHeaders;
     this.reasoningEffort = reasoningEffort;
     this.providerOverride = providerOverride;
+    this.source = source;
   }
   create(params, options) {
     const self2 = this;
     let httpResponse;
     const promise2 = (async () => {
       const overrideTransport = self2.providerOverride?.transport === "anthropic" ? undefined : self2.providerOverride?.transport;
-      const request = resolveProviderRequest({ model: self2.providerOverride?.model ?? params.model, baseUrl: self2.providerOverride?.baseURL, reasoningEffortOverride: self2.reasoningEffort, transportOverride: overrideTransport });
+      const hasProviderReasoningEffortOverride = !!self2.providerOverride && Object.prototype.hasOwnProperty.call(self2.providerOverride, "reasoningEffort");
+      const reasoningEffortOverride = hasProviderReasoningEffortOverride ? self2.providerOverride?.reasoningEffort ?? null : self2.reasoningEffort;
+      const request = resolveProviderRequest({
+        model: self2.providerOverride?.model ?? params.model,
+        baseUrl: self2.providerOverride?.baseURL,
+        ...reasoningEffortOverride !== undefined ? { reasoningEffortOverride } : {},
+        transportOverride: overrideTransport
+      });
       const response = await self2._doRequest(request, params, options);
       httpResponse = response;
       if (params.stream) {
@@ -86127,28 +86220,93 @@ class OpenAIShimMessages {
       signal: options?.signal
     };
     const maxAttempts = isGithub ? GITHUB_429_MAX_RETRIES : 1;
+    const endpoint = classifyModelHttpEndpoint(chatCompletionsUrl);
     let response;
     for (let attempt = 0;attempt < maxAttempts; attempt++) {
+      const attemptNumber = attempt + 1;
+      const requestId = `req_${crypto.randomUUID().replace(/-/g, "")}`;
+      const requestStartedAt = Date.now();
+      emitSdkHttp({
+        type: "request_start",
+        requestId,
+        method: fetchInit.method,
+        url: chatCompletionsUrl,
+        source: this.source,
+        endpoint,
+        attempt: attemptNumber,
+        maxAttempts,
+        timestamp: requestStartedAt
+      });
       try {
         response = await fetch(chatCompletionsUrl, fetchInit);
       } catch (fetchErr) {
         const msg = fetchErr instanceof Error ? `${fetchErr.message}${fetchErr.cause ? ` (${fetchErr.cause?.code ?? fetchErr.cause?.message ?? ""})` : ""}` : String(fetchErr);
+        emitSdkHttp({
+          type: "error",
+          requestId,
+          error: fetchErr instanceof Error ? fetchErr : new Error(String(fetchErr)),
+          durationMs: Date.now() - requestStartedAt,
+          source: this.source,
+          endpoint,
+          attempt: attemptNumber,
+          maxAttempts,
+          failureReason: msg,
+          timestamp: Date.now()
+        });
         throw new APIConnectionError({
           message: `OpenAI-compat connection failed: ${msg}`,
           cause: fetchErr
         });
       }
       if (response.ok) {
+        emitSdkHttp({
+          type: "response_start",
+          requestId,
+          status: response.status,
+          durationMs: Date.now() - requestStartedAt,
+          source: this.source,
+          endpoint,
+          attempt: attemptNumber,
+          maxAttempts,
+          timestamp: Date.now()
+        });
         return response;
       }
       if (isGithub && response.status === 429 && attempt < maxAttempts - 1) {
         await response.text().catch(() => {});
         const delaySec = Math.min(GITHUB_429_BASE_DELAY_SEC * 2 ** attempt, GITHUB_429_MAX_DELAY_SEC);
+        emitSdkHttp({
+          type: "response_start",
+          requestId,
+          status: response.status,
+          durationMs: Date.now() - requestStartedAt,
+          source: this.source,
+          endpoint,
+          attempt: attemptNumber,
+          maxAttempts,
+          willRetry: true,
+          retryDelayMs: delaySec * 1000,
+          failureReason: `HTTP ${response.status}${formatRetryAfterHint(response)}`,
+          timestamp: Date.now()
+        });
         await sleepMs(delaySec * 1000);
         continue;
       }
       const errorBody = await response.text().catch(() => "unknown error");
       const rateHint = isGithub && response.status === 429 ? formatRetryAfterHint(response) : "";
+      const failureReason = `${responseBodyToFailureReason(response.status, errorBody)}${rateHint}`;
+      emitSdkHttp({
+        type: "response_start",
+        requestId,
+        status: response.status,
+        durationMs: Date.now() - requestStartedAt,
+        source: this.source,
+        endpoint,
+        attempt: attemptNumber,
+        maxAttempts,
+        failureReason,
+        timestamp: Date.now()
+      });
       let errorResponse;
       try {
         errorResponse = JSON.parse(errorBody);
@@ -86237,8 +86395,8 @@ function convertOpenAIResponseToAnthropic(data, model) {
 class OpenAIShimBeta {
   messages;
   reasoningEffort;
-  constructor(defaultHeaders, reasoningEffort, providerOverride) {
-    this.messages = new OpenAIShimMessages(defaultHeaders, reasoningEffort, providerOverride);
+  constructor(defaultHeaders, reasoningEffort, providerOverride, source) {
+    this.messages = new OpenAIShimMessages(defaultHeaders, reasoningEffort, providerOverride, source);
     this.reasoningEffort = reasoningEffort;
   }
 }
@@ -86260,7 +86418,7 @@ function createOpenAIShimClient(options) {
   }
   const beta = new OpenAIShimBeta({
     ...options.defaultHeaders ?? {}
-  }, options.reasoningEffort, options.providerOverride);
+  }, options.reasoningEffort, options.providerOverride, options.source);
   return {
     beta,
     messages: beta.messages
@@ -86279,7 +86437,9 @@ var init_shim2 = __esm(() => {
   init_schemaSanitizer();
   init_providerProfile();
   init_capabilities2();
+  init_httpObservability();
   init_canonical();
+  init_requestSideChannels();
   OpenAIShimStream = class OpenAIShimStream {
     generator;
     controller = new AbortController;
@@ -86376,7 +86536,8 @@ async function getNormalizedClient({
       defaultHeaders: safeHeaders,
       maxRetries,
       timeout: parseInt(process.env.API_TIMEOUT_MS || String(600000), 10),
-      providerOverride
+      providerOverride,
+      source
     });
   }
   const transportOverride = getQueryEnvVar(QUERY_ENV_KEY_TRANSPORT_OVERRIDE);
@@ -86386,7 +86547,8 @@ async function getNormalizedClient({
     return createOpenAIShimClient2({
       defaultHeaders,
       maxRetries,
-      timeout: parseInt(process.env.API_TIMEOUT_MS || String(600000), 10)
+      timeout: parseInt(process.env.API_TIMEOUT_MS || String(600000), 10),
+      source
     });
   }
   const stagingOauthBaseUrl = process.env.USER_TYPE === "ant" && isEnvTruthy(process.env.USE_STAGING_OAUTH) ? getOauthConfig().BASE_API_URL : undefined;
@@ -86438,42 +86600,49 @@ function buildFetch(fetchOverride, source) {
     const url3 = input instanceof Request ? input.url : String(input);
     const method = init?.method ?? (input instanceof Request ? input.method : "GET");
     const requestId = headers.get(CLIENT_REQUEST_ID_HEADER) ?? randomUUID();
+    const endpoint = classifyModelHttpEndpoint(url3);
     try {
       logForDebugging2(`[API REQUEST] ${new URL(url3).pathname} ${CLIENT_REQUEST_ID_HEADER}=${requestId} source=${source ?? "unknown"}`);
     } catch {}
     const requestStartedAt = Date.now();
+    emitSdkHttp({
+      type: "request_start",
+      requestId,
+      method,
+      url: url3,
+      source,
+      endpoint,
+      attempt: 1,
+      maxAttempts: 1,
+      timestamp: requestStartedAt
+    });
     try {
-      emitHttp({
-        type: "request_start",
+      const response = await inner(input, { ...init, headers });
+      emitSdkHttp({
+        type: "response_start",
         requestId,
-        method,
-        url: url3,
+        status: response.status,
+        durationMs: Date.now() - requestStartedAt,
         source,
-        timestamp: requestStartedAt
+        endpoint,
+        attempt: 1,
+        maxAttempts: 1,
+        timestamp: Date.now()
       });
-    } catch {}
-    try {
-      const response = await inner(input, { ...init, headers });
-      try {
-        emitHttp({
-          type: "response_start",
-          requestId,
-          status: response.status,
-          durationMs: Date.now() - requestStartedAt,
-          timestamp: Date.now()
-        });
-      } catch {}
       return response;
     } catch (err) {
-      try {
-        emitHttp({
-          type: "error",
-          requestId,
-          error: err instanceof Error ? err : new Error(String(err)),
-          durationMs: Date.now() - requestStartedAt,
-          timestamp: Date.now()
-        });
-      } catch {}
+      emitSdkHttp({
+        type: "error",
+        requestId,
+        error: err instanceof Error ? err : new Error(String(err)),
+        durationMs: Date.now() - requestStartedAt,
+        source,
+        endpoint,
+        attempt: 1,
+        maxAttempts: 1,
+        failureReason: errorToFailureReason(err),
+        timestamp: Date.now()
+      });
       throw err;
     }
   };
@@ -86492,7 +86661,7 @@ var init_clientFactory = __esm(() => {
   init_oauth();
   init_debug();
   init_envUtils();
-  init_observer();
+  init_httpObservability();
 });
 
 // src/providers/anthropic/index.ts
@@ -92628,7 +92797,7 @@ function parseReasoningEffort(value) {
   if (!value)
     return;
   const normalized = value.trim().toLowerCase();
-  if (normalized === "low" || normalized === "medium" || normalized === "high" || normalized === "xhigh") {
+  if (normalized === "none" || normalized === "minimal" || normalized === "low" || normalized === "medium" || normalized === "high" || normalized === "xhigh") {
     return normalized;
   }
   return;
@@ -92718,7 +92887,8 @@ function normalizeGithubModelsApiModel(requestedModel) {
   return segment;
 }
 function resolveProviderTransport(options) {
-  const rawBaseUrl = asEnvUrl(options?.baseUrl) ?? asEnvUrl(getQueryEnvVar("OPENAI_BASE_URL")) ?? asEnvUrl(getQueryEnvVar("OPENAI_API_BASE"));
+  const optionBaseUrl = asEnvUrl(options?.baseUrl);
+  const rawBaseUrl = optionBaseUrl ?? asEnvUrl(getQueryEnvVar("OPENAI_BASE_URL")) ?? asEnvUrl(getQueryEnvVar("OPENAI_API_BASE"));
   const envOverride = options?.transportOverride ?? getQueryEnvVar(QUERY_ENV_KEY_TRANSPORT_OVERRIDE);
   if (envOverride && envOverride !== "auto") {
     let normalized;
@@ -92728,7 +92898,8 @@ function resolveProviderTransport(options) {
     } else {
       normalized = envOverride;
     }
-    const baseUrl2 = (rawBaseUrl ?? (normalized === "openai_responses" ? DEFAULT_CODEX_BASE_URL : DEFAULT_OPENAI_BASE_URL)).replace(/\/+$/, "");
+    const baseUrlSource = rawBaseUrl ?? (normalized === "openai_responses" ? DEFAULT_CODEX_BASE_URL : DEFAULT_OPENAI_BASE_URL);
+    const baseUrl2 = baseUrlSource.replace(/\/+$/, "");
     return { transport: normalized, baseUrl: baseUrl2 };
   }
   const modelForDetection = options?.model?.trim() ?? "";
@@ -92749,7 +92920,10 @@ function resolveProviderRequest(options) {
     transportOverride: options?.transportOverride
   });
   const resolvedModel = transport === "chat_completions" && isEnvTruthy(getQueryEnvVar("CLAUDE_CODE_USE_GITHUB")) ? normalizeGithubModelsApiModel(requestedModel) : descriptor.baseModel;
-  const reasoning = options?.reasoningEffortOverride ? { effort: options.reasoningEffortOverride } : descriptor.reasoning;
+  const hasReasoningEffortOverride = !!options && Object.prototype.hasOwnProperty.call(options, "reasoningEffortOverride");
+  const rawEnvReasoningEffortOverride = getQueryEnvVar(QUERY_ENV_KEY_REASONING_EFFORT_OVERRIDE);
+  const envReasoningEffortOverride = parseReasoningEffort(rawEnvReasoningEffortOverride);
+  const reasoning = hasReasoningEffortOverride ? options?.reasoningEffortOverride ? { effort: options.reasoningEffortOverride } : undefined : rawEnvReasoningEffortOverride === QUERY_ENV_VALUE_REASONING_EFFORT_CLEAR ? undefined : envReasoningEffortOverride ? { effort: envReasoningEffortOverride } : descriptor.reasoning;
   return {
     transport,
     requestedModel,
@@ -92860,7 +93034,7 @@ function getReasoningEffortForModel(model) {
   const aliasConfig = CODEX_ALIAS_MODELS[alias];
   return aliasConfig?.reasoningEffort;
 }
-var DEFAULT_OPENAI_BASE_URL = "https://api.openai.com/v1", DEFAULT_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex", DEFAULT_GITHUB_MODELS_API_MODEL = "openai/gpt-4.1", CODEX_ALIAS_MODELS, QUERY_ENV_KEY_TRANSPORT_OVERRIDE = "__OPENCOW_TRANSPORT_OVERRIDE", QUERY_ENV_KEY_PROVIDER_SPECIFIC_OPENAI_RESPONSES = "__OPENCOW_PROVIDER_SPECIFIC_OPENAI_RESPONSES", LOCALHOST_HOSTNAMES, warnedCodexAliasOnce = false, MissingProviderModelError;
+var DEFAULT_OPENAI_BASE_URL = "https://api.openai.com/v1", DEFAULT_CODEX_BASE_URL = "https://chatgpt.com/backend-api/codex", DEFAULT_GITHUB_MODELS_API_MODEL = "openai/gpt-4.1", CODEX_ALIAS_MODELS, QUERY_ENV_KEY_TRANSPORT_OVERRIDE = "__OPENCOW_TRANSPORT_OVERRIDE", QUERY_ENV_KEY_REASONING_EFFORT_OVERRIDE = "__OPENCOW_REASONING_EFFORT_OVERRIDE", QUERY_ENV_VALUE_REASONING_EFFORT_CLEAR = "__OPENCOW_CLEAR_REASONING_EFFORT__", QUERY_ENV_KEY_PROVIDER_SPECIFIC_OPENAI_RESPONSES = "__OPENCOW_PROVIDER_SPECIFIC_OPENAI_RESPONSES", QUERY_ENV_KEY_EXTRA_BODY = "__OPENCOW_EXTRA_BODY", LOCALHOST_HOSTNAMES, warnedCodexAliasOnce = false, MissingProviderModelError;
 var init_config3 = __esm(() => {
   init_envUtils();
   init_state2();
@@ -94524,7 +94698,7 @@ function printStartupScreen() {
   const sLen = ` ● ${sL}    Ready — type /help to begin`.length;
   out.push(boxRow(sRow, W2, sLen));
   out.push(`${rgb(...BORDER)}╚${"═".repeat(W2 - 2)}╝${RESET}`);
-  out.push(`  ${DIM}${rgb(...DIMCOL)}opencow ${RESET}${rgb(...ACCENT)}v${"0.4.12"}${RESET}`);
+  out.push(`  ${DIM}${rgb(...DIMCOL)}opencow ${RESET}${rgb(...ACCENT)}v${"0.4.14"}${RESET}`);
   out.push("");
   process.stdout.write(out.join(`
 `) + `
@@ -96190,6 +96364,11 @@ var FINGERPRINT_SALT = "59cf53e54c78";
 var init_fingerprint = () => {};
 
 // src/session/sideQuery.ts
+function headersForBetas(betas) {
+  if (!betas || betas.length === 0)
+    return;
+  return { "anthropic-beta": betas.join(",") };
+}
 function extractFirstUserMessageText(messages) {
   const firstUserMessage = messages.find((m) => m.role === "user");
   if (!firstUserMessage)
@@ -96252,7 +96431,7 @@ async function sideQuery(opts) {
   }
   const normalizedModel = normalizeModelStringForAPI(model);
   const start = Date.now();
-  const response = await client.beta.messages.create({
+  const response = await client.messages.create({
     model: normalizedModel,
     max_tokens,
     system: systemBlocks,
@@ -96263,9 +96442,12 @@ async function sideQuery(opts) {
     ...temperature !== undefined && { temperature },
     ...stop_sequences && { stop_sequences },
     ...thinkingConfig && { thinking: thinkingConfig },
-    ...betas.length > 0 && { betas },
-    metadata: getAPIMetadata()
-  }, { signal });
+    metadata: getAPIMetadata(),
+    ...getExtraBodyParams()
+  }, {
+    signal,
+    ...betas.length > 0 ? { headers: headersForBetas(betas) } : {}
+  });
   const requestId = response._request_id ?? undefined;
   const now = Date.now();
   const lastCompletion = getLastApiCompletionTimestamp();
@@ -113069,15 +113251,16 @@ function isMaxTokensCapEnabled() {
 }
 function getMaxOutputTokensForModel(model, opts) {
   const maxOutputTokens = getModelMaxOutputTokens(model);
-  const defaultTokens = isMaxTokensCapEnabled() ? Math.min(maxOutputTokens.default, CAPPED_DEFAULT_MAX_TOKENS) : maxOutputTokens.default;
+  const upperLimit = opts?.upperLimitOverride !== undefined && Number.isFinite(opts.upperLimitOverride) && opts.upperLimitOverride >= 1 ? Math.floor(opts.upperLimitOverride) : maxOutputTokens.upperLimit;
+  const defaultTokens = isMaxTokensCapEnabled() ? Math.min(maxOutputTokens.default, CAPPED_DEFAULT_MAX_TOKENS, upperLimit) : Math.min(maxOutputTokens.default, upperLimit);
   if (opts?.override !== undefined) {
-    if (!Number.isFinite(opts.override) || opts.override < 1 || opts.override > maxOutputTokens.upperLimit) {
-      console.warn(`[opencow] Options.maxOutputTokens=${opts.override} out of range ` + `[1, ${maxOutputTokens.upperLimit}] for model ${model}; clamping.`);
+    if (!Number.isFinite(opts.override) || opts.override < 1 || opts.override > upperLimit) {
+      console.warn(`[opencow] Options.maxOutputTokens=${opts.override} out of range ` + `[1, ${upperLimit}] for model ${model}; clamping.`);
     }
-    const clamped = Math.min(Math.max(1, Math.floor(opts.override)), maxOutputTokens.upperLimit);
+    const clamped = Math.min(Math.max(1, Math.floor(opts.override)), upperLimit);
     return clamped;
   }
-  const result = validateBoundedIntEnvVar("CLAUDE_CODE_MAX_OUTPUT_TOKENS", resolveEnvVar("MAX_OUTPUT_TOKENS"), defaultTokens, maxOutputTokens.upperLimit);
+  const result = validateBoundedIntEnvVar("CLAUDE_CODE_MAX_OUTPUT_TOKENS", resolveEnvVar("MAX_OUTPUT_TOKENS"), defaultTokens, upperLimit);
   return result.effective;
 }
 var init_maxTokens = __esm(() => {
@@ -118590,6 +118773,13 @@ function getDisableExtglobCommand(shellPath) {
   }
   return null;
 }
+function getHostProvidedPathCommand() {
+  const path11 = getHostProvidedEnvVar("PATH");
+  if (!path11) {
+    return null;
+  }
+  return `export PATH=${quote([path11])}`;
+}
 async function createBashShellProvider(shellPath, options2) {
   let currentSandboxTmpDir;
   const snapshotPromise = options2?.skipSnapshot ? Promise.resolve(undefined) : createAndSaveSnapshot(shellPath).catch((error41) => {
@@ -118630,6 +118820,10 @@ async function createBashShellProvider(shellPath, options2) {
         const finalPath = getPlatform() === "windows" ? windowsPathToPosixPath(snapshotFilePath) : snapshotFilePath;
         commandParts.push(`source ${quote([finalPath])} 2>/dev/null || true`);
       }
+      const hostPathCommand = getHostProvidedPathCommand();
+      if (hostPathCommand) {
+        commandParts.push(hostPathCommand);
+      }
       const sessionEnvScript2 = await getSessionEnvironmentScript();
       if (sessionEnvScript2) {
         commandParts.push(sessionEnvScript2);
@@ -147231,12 +147425,13 @@ async function makeTestQuery() {
   });
   const messages = [{ role: "user", content: "quota" }];
   const betas = getModelBetas(model);
-  return anthropic.beta.messages.create({
+  return anthropic.messages.create({
     model,
     max_tokens: 1,
     messages,
-    metadata: getAPIMetadata(),
-    ...betas.length > 0 ? { betas } : {}
+    metadata: getAPIMetadata()
+  }, {
+    ...betas.length > 0 ? { headers: { "anthropic-beta": betas.join(",") } } : {}
   }).asResponse();
 }
 async function checkQuotaStatus() {
@@ -148393,6 +148588,17 @@ async function* withRetry(getClient, operation, options2) {
         status: error41.status,
         provider: getAPIProviderForStatsig()
       });
+      emitSdkHttp({
+        type: "retry",
+        source: options2.querySource,
+        endpoint: "messages",
+        attempt: reportedAttempt,
+        ...persistent ? {} : { maxAttempts: maxRetries + 1 },
+        retryDelayMs: delayMs,
+        status: error41 instanceof APIError ? error41.status : undefined,
+        failureReason: errorToFailureReason(error41),
+        timestamp: Date.now()
+      });
       if (persistent) {
         if (delayMs > 60000) {
           logEvent("tengu_api_persistent_retry_wait", {
@@ -148646,6 +148852,7 @@ var init_retry = __esm(() => {
   init_mocking();
   init_errors8();
   init_errorUtils();
+  init_httpObservability();
   FOREGROUND_529_RETRY_SOURCES = new Set([
     "repl_main_thread",
     "repl_main_thread:outputStyle:custom",
@@ -243655,7 +243862,11 @@ async function* queryLoop(params, consumedCommandUuids) {
       if (false) {}
       const mediaRecoveryEnabled = reactiveCompact?.isReactiveCompactEnabled() ?? false;
       if (!compactionResult && querySource !== "compact" && querySource !== "session_memory" && !(reactiveCompact?.isReactiveCompactEnabled() && isAutoCompactEnabled()) && !collapseOwnsIt) {
-        const { isAtBlockingLimit } = calculateTokenWarningState(tokenCountWithEstimation(messagesForQuery) - snipTokensFreed, toolUseContext.options.mainLoopModel);
+        const { isAtBlockingLimit } = calculateTokenWarningState(tokenCountWithEstimation(messagesForQuery) - snipTokensFreed, toolUseContext.options.mainLoopModel, {
+          contextWindow: toolUseContext.options.contextWindow,
+          maxOutputTokens: toolUseContext.options.maxOutputTokens,
+          maxOutputTokensLimit: toolUseContext.options.maxOutputTokensLimit
+        });
         if (isAtBlockingLimit) {
           yield createAssistantAPIErrorMessage({
             content: PROMPT_TOO_LONG_ERROR_MESSAGE,
@@ -243698,6 +243909,7 @@ async function* queryLoop(params, consumedCommandUuids) {
                 allowedAgentTypes: toolUseContext.options.agentDefinitions.allowedAgentTypes,
                 hasAppendSystemPrompt: !!toolUseContext.options.appendSystemPrompt,
                 maxOutputTokensOverride,
+                maxOutputTokensLimitOverride: params.maxOutputTokensLimitOverride,
                 fetchOverride: dumpPromptsFetch,
                 mcpTools: appState.mcp.tools,
                 hasPendingMcpServers: appState.mcp.clients.some((c6) => c6.type === "pending"),
@@ -244407,7 +244619,7 @@ function getAnthropicEnvMetadata() {
 function getBuildAgeMinutes() {
   if (false)
     ;
-  const buildTime = new Date("2026-06-24T10:02:32.669Z").getTime();
+  const buildTime = new Date("2026-07-07T07:10:57.155Z").getTime();
   if (isNaN(buildTime))
     return;
   return Math.floor((Date.now() - buildTime) / 60000);
@@ -245008,6 +245220,7 @@ async function runForkedAgent({
       toolUseContext: isolatedToolUseContext,
       querySource,
       maxOutputTokensOverride: maxOutputTokens,
+      maxOutputTokensLimitOverride: isolatedToolUseContext.options.maxOutputTokensLimit,
       maxTurns,
       skipCacheWrite
     })) {
@@ -264874,49 +265087,36 @@ assistant: Still waiting on the audit — that's one of the things it's checking
 
 <example>
 user: "Can you get a second opinion on whether this migration is safe?"
-assistant: <thinking>I'll ask the code-reviewer agent — it won't see my analysis, so it can give an independent read.</thinking>
+assistant: <thinking>Forking this review keeps the detailed audit out of my context while I keep steering the work.</thinking>
 <commentary>
-A subagent_type is specified, so the agent starts fresh. It needs full context in the prompt. The briefing explains what to assess and why.
+No subagent_type is specified, so this uses the fork path. If you do specify subagent_type, copy an exact type from the available agent list; never invent one from an example.
 </commentary>
 ${AGENT_TOOL_NAME2}({
   name: "migration-review",
   description: "Independent migration review",
-  subagent_type: "code-reviewer",
   prompt: "Review migration 0042_user_schema.sql for safety. Context: we're adding a NOT NULL column to a 50M-row table. Existing rows get a backfill default. I want a second opinion on whether the backfill approach is safe under concurrent writes — I've checked locking behavior but want independent verification. Report: is this safe, and if not, what specifically breaks?"
 })
 </example>
 `;
   const currentExamples = `Example usage:
 
-<example_agent_descriptions>
-"test-runner": use this agent after you are done writing code to run tests
-"greeting-responder": use this agent to respond to user greetings with a friendly joke
-</example_agent_descriptions>
-
 <example>
-user: "Please write a function that checks if a number is prime"
-assistant: I'm going to use the ${FILE_WRITE_TOOL_NAME2} tool to write the following code:
-<code>
-function isPrime(n) {
-  if (n <= 1) return false
-  for (let i = 2; i * i <= n; i++) {
-    if (n % i === 0) return false
-  }
-  return true
-}
-</code>
+user: "Hello"
 <commentary>
-Since a significant piece of code was written and the task was completed, now use the test-runner agent to run the tests
+Do not use the ${AGENT_TOOL_NAME2} tool for a simple greeting or a short reply. Answer directly.
 </commentary>
-assistant: Uses the ${AGENT_TOOL_NAME2} tool to launch the test-runner agent
+assistant: "Hi! Nice to see you. What shall we work on today?"
 </example>
 
 <example>
-user: "Hello"
+user: "Can you investigate whether this migration is safe?"
 <commentary>
-Since the user is greeting, use the greeting-responder agent to respond with a friendly joke
+This is a multi-step review, so the ${AGENT_TOOL_NAME2} tool can help. Omit subagent_type to use the default general-purpose agent, or specify a subagent_type only by copying an exact type from the available agent list.
 </commentary>
-assistant: "I'm going to use the ${AGENT_TOOL_NAME2} tool to launch the greeting-responder agent"
+${AGENT_TOOL_NAME2}({
+  description: "Review migration safety",
+  prompt: "Review migration 0042_user_schema.sql for safety. Context: we're adding a NOT NULL column to a 50M-row table. Existing rows get a backfill default. Check whether the backfill approach is safe under concurrent writes. Report: is this safe, and if not, what specifically breaks?"
+})
 </example>
 `;
   const listViaAttachment = shouldInjectAgentListInMessages();
@@ -264929,7 +265129,7 @@ The ${AGENT_TOOL_NAME2} tool launches specialized agents (subprocesses) that aut
 
 ${agentListSection}
 
-${forkEnabled ? `When using the ${AGENT_TOOL_NAME2} tool, specify a subagent_type to use a specialized agent, or omit it to fork yourself — a fork inherits your full conversation context.` : `When using the ${AGENT_TOOL_NAME2} tool, specify a subagent_type parameter to select which agent type to use. If omitted, the general-purpose agent is used.`}`;
+${forkEnabled ? `When using the ${AGENT_TOOL_NAME2} tool, specify a subagent_type to use a specialized agent, or omit it to fork yourself — a fork inherits your full conversation context. Only use subagent_type values that appear in the available agent types; never invent an agent type from an example.` : `When using the ${AGENT_TOOL_NAME2} tool, specify a subagent_type parameter to select which agent type to use. If omitted, the general-purpose agent is used. Only use subagent_type values that appear in the available agent types; never invent an agent type from an example.`}`;
   if (isCoordinator) {
     return shared2;
   }
@@ -264957,7 +265157,7 @@ Usage notes:
 - The agent's outputs should generally be trusted
 - Clearly tell the agent whether you expect it to write code or just to do research (search, file reads, web fetches, etc.)${forkEnabled ? "" : ", since it is not aware of the user's intent"}
 - If the agent description mentions that it should be used proactively, then you should try your best to use it without the user having to ask for it first. Use your judgement.
-- If the user specifies that they want you to run agents "in parallel", you MUST send a single message with multiple ${AGENT_TOOL_NAME2} tool use content blocks. For example, if you need to launch both a build-validator agent and a test-runner agent in parallel, send a single message with both tool calls.
+- If the user specifies that they want you to run agents "in parallel", you MUST send a single message with multiple ${AGENT_TOOL_NAME2} tool use content blocks. Each call must either omit subagent_type or use an exact type from the available agent list.
 - You can optionally set \`isolation: "worktree"\` to run the agent in a temporary git worktree, giving it an isolated copy of the repository. The worktree is automatically cleaned up if the agent makes no changes; if changes are made, the worktree path and branch are returned in the result.${process.env.USER_TYPE === "ant" ? `
 - You can set \`isolation: "remote"\` to run the agent in a remote CCR environment. This is always a background task; you'll be notified when it completes. Use for long-running tasks that need a fresh sandbox.` : ""}${isInProcessTeammate() ? `
 - The run_in_background, name, team_name, and mode parameters are not available in this context. Only synchronous subagents are supported.` : isTeammate() ? `
@@ -264972,7 +265172,6 @@ var init_prompt8 = __esm(() => {
   init_teammate();
   init_teammateContext();
   init_prompt();
-  init_prompt2();
   init_constants4();
   init_forkSubagent();
   init_state2();
@@ -288389,7 +288588,7 @@ async function loadPluginSettings(pluginPath, manifest) {
   try {
     const content = await getFsImplementation().readFile(settingsJsonPath, { encoding: "utf-8" });
     const parsed = jsonParse(content);
-    if (isRecord(parsed)) {
+    if (isRecord2(parsed)) {
       const filtered = parsePluginSettings(parsed);
       if (filtered) {
         logForDebugging2(`Loaded settings from settings.json for plugin ${manifest.name}`);
@@ -289079,7 +289278,7 @@ function cachePluginSettings(plugins) {
     logForDebugging2(`Cached plugin settings with keys: ${Object.keys(settings).join(", ")}`);
   }
 }
-function isRecord(value) {
+function isRecord2(value) {
   return typeof value === "object" && value !== null && !Array.isArray(value);
 }
 var PluginSettingsSchema, loadAllPlugins, loadAllPluginsCacheOnly;
@@ -302297,7 +302496,17 @@ ${formattedSummary}`;
   if (transcriptPath) {
     baseSummary += `
 
-If you need specific details from before compaction (like exact code snippets, error messages, or content you generated), read the full transcript at: ${transcriptPath}`;
+IMPORTANT — Transcript recovery protocol:
+The full pre-compaction conversation is preserved at: ${transcriptPath}
+When you encounter ANY of these situations, you MUST use the Read tool to search the transcript for the missing context BEFORE responding:
+- You are unsure about a specific detail (file path, function name, error message, code snippet, user preference)
+- The user references something you cannot find in the summary above
+- You need exact code that was previously read or written
+- You are about to make a decision but feel uncertain whether the user already gave guidance on it
+- A tool name, skill name, or configuration was discussed but is not in the summary
+
+The transcript is a JSONL file. Read its tail (last 500–2000 lines) first for the most recent context; if that doesn't resolve the gap, read earlier sections.
+Do NOT guess or hallucinate details that might have been discussed — read the transcript instead.`;
   }
   if (recentMessagesPreserved) {
     baseSummary += `
@@ -302364,6 +302573,7 @@ Your summary should include the following sections:
 8. Current Work: Describe in detail precisely what was being worked on immediately before this summary request, paying special attention to the most recent messages from both user and assistant. Include file names and code snippets where applicable.
 9. Optional Next Step: List the next step that you will take that is related to the most recent work you were doing. IMPORTANT: ensure that this step is DIRECTLY in line with the user's most recent explicit requests, and the task you were working on immediately before this summary request. If your last task was concluded, then only list next steps if they are explicitly in line with the users request. Do not start on tangential requests or really old requests that were already completed without confirming with the user first.
                        If there is a next step, include direct quotes from the most recent conversation showing exactly what task you were working on and where you left off. This should be verbatim to ensure there's no drift in task interpretation.
+10. Uncertain or Incomplete Context: List any items where you are not fully confident in the details — e.g., a tool/skill name you vaguely recall but cannot confirm, a user preference you think was stated but cannot pinpoint, or a decision whose rationale is unclear. Mark each with [NEEDS_TRANSCRIPT_LOOKUP] so the post-compaction assistant knows to read the transcript file for these specific items before acting on them.
 
 Here's an example of how your output should be structured:
 
@@ -302414,10 +302624,14 @@ Here's an example of how your output should be structured:
 9. Optional Next Step:
    [Optional Next step to take]
 
+10. Uncertain or Incomplete Context:
+   - [Item with unclear details] [NEEDS_TRANSCRIPT_LOOKUP]
+   - [...]
+
 </summary>
 </example>
 
-Please provide your summary based on the conversation so far, following this structure and ensuring precision and thoroughness in your response. 
+Please provide your summary based on the conversation so far, following this structure and ensuring precision and thoroughness in your response.
 
 There may be additional summarization instructions provided in the included context. If so, remember to follow these instructions when creating the above summary. Examples of instructions include:
 <example>
@@ -302445,6 +302659,7 @@ Your summary should include the following sections:
 7. Pending Tasks: Outline any pending tasks from the recent messages.
 8. Current Work: Describe precisely what was being worked on immediately before this summary request.
 9. Optional Next Step: List the next step related to the most recent work. Include direct quotes from the most recent conversation.
+10. Uncertain or Incomplete Context: List any items where details are unclear or potentially missing from the recent messages. Mark each with [NEEDS_TRANSCRIPT_LOOKUP].
 
 Here's an example of how your output should be structured:
 
@@ -302485,6 +302700,9 @@ Here's an example of how your output should be structured:
 9. Optional Next Step:
    [Optional Next step to take]
 
+10. Uncertain or Incomplete Context:
+   - [Item with unclear details] [NEEDS_TRANSCRIPT_LOOKUP]
+
 </summary>
 </example>
 
@@ -302505,6 +302723,7 @@ Your summary should include the following sections:
 7. Pending Tasks: Outline any pending tasks.
 8. Work Completed: Describe what was accomplished by the end of this portion.
 9. Context for Continuing Work: Summarize any context, decisions, or state that would be needed to understand and continue the work in subsequent messages.
+10. Uncertain or Incomplete Context: List any items where details are unclear or potentially incomplete. Mark each with [NEEDS_TRANSCRIPT_LOOKUP] so the continuing session knows to look them up in the transcript before acting.
 
 Here's an example of how your output should be structured:
 
@@ -302545,6 +302764,9 @@ Here's an example of how your output should be structured:
 9. Context for Continuing Work:
    [Key context, decisions, or state needed to continue the work]
 
+10. Uncertain or Incomplete Context:
+   - [Item with unclear details] [NEEDS_TRANSCRIPT_LOOKUP]
+
 </summary>
 </example>
 
@@ -303199,7 +303421,10 @@ async function streamCompactSummary({
           toolChoice: undefined,
           isNonInteractiveSession: context4.options.isNonInteractiveSession,
           hasAppendSystemPrompt: !!context4.options.appendSystemPrompt,
-          maxOutputTokensOverride: Math.min(COMPACT_MAX_OUTPUT_TOKENS, getMaxOutputTokensForModel(context4.options.mainLoopModel)),
+          maxOutputTokensOverride: Math.min(COMPACT_MAX_OUTPUT_TOKENS, getMaxOutputTokensForModel(context4.options.mainLoopModel, {
+            upperLimitOverride: context4.options.maxOutputTokensLimit
+          })),
+          maxOutputTokensLimitOverride: context4.options.maxOutputTokensLimit,
           querySource: "compact",
           agents: context4.options.agentDefinitions.activeAgents,
           mcpTools: [],
@@ -304096,7 +304321,10 @@ var init_sessionMemoryCompact = __esm(() => {
 
 // src/controller/compact/autoCompact.ts
 function getEffectiveContextWindowSize(model, opts) {
-  const reservedTokensForSummary = Math.min(getMaxOutputTokensForModel(model, { override: opts?.maxOutputTokens }), MAX_OUTPUT_TOKENS_FOR_SUMMARY);
+  const reservedTokensForSummary = Math.min(getMaxOutputTokensForModel(model, {
+    override: opts?.maxOutputTokens,
+    upperLimitOverride: opts?.maxOutputTokensLimit
+  }), MAX_OUTPUT_TOKENS_FOR_SUMMARY);
   let contextWindow = getContextWindowForModel(model, getSdkBetas(), {
     override: opts?.contextWindow
   });
@@ -304182,7 +304410,8 @@ async function autoCompactIfNeeded(messages, toolUseContext, cacheSafeParams, qu
   const model = toolUseContext.options.mainLoopModel;
   const opts = {
     contextWindow: toolUseContext.options.contextWindow,
-    maxOutputTokens: toolUseContext.options.maxOutputTokens
+    maxOutputTokens: toolUseContext.options.maxOutputTokens,
+    maxOutputTokensLimit: toolUseContext.options.maxOutputTokensLimit
   };
   const shouldCompact = await shouldAutoCompact(messages, model, querySource, snipTokensFreed, opts);
   if (!shouldCompact) {
@@ -305287,6 +305516,11 @@ var init_toolSearch = __esm(() => {
 });
 
 // src/providers/shared/tokenCount.ts
+function headersForBetas2(betas) {
+  if (!betas || betas.length === 0)
+    return;
+  return { "anthropic-beta": betas.join(",") };
+}
 function hasThinkingBlocks(messages) {
   for (const message of messages) {
     if (message.role === "assistant" && Array.isArray(message.content)) {
@@ -305361,17 +305595,18 @@ async function countMessagesTokensWithAPI(messages, tools) {
         model,
         source: "count_tokens"
       });
-      const response = await anthropic.beta.messages.countTokens({
+      const response = await anthropic.messages.countTokens({
         model: normalizeModelStringForAPI(model),
         messages: messages.length > 0 ? messages : [{ role: "user", content: "foo" }],
         tools,
-        ...betas.length > 0 && { betas },
         ...containsThinking && {
           thinking: {
             type: "enabled",
             budget_tokens: TOKEN_COUNT_THINKING_BUDGET
           }
         }
+      }, {
+        ...betas.length > 0 ? { headers: headersForBetas2(betas) } : {}
       });
       if (typeof response.input_tokens !== "number") {
         return null;
@@ -305410,12 +305645,11 @@ async function countTokensViaHaikuFallback(messages, tools) {
   const normalizedMessages = stripToolSearchFieldsFromMessages(messages);
   const messagesToSend = normalizedMessages.length > 0 ? normalizedMessages : [{ role: "user", content: "count" }];
   const betas = getModelBetas(model);
-  const response = await anthropic.beta.messages.create({
+  const response = await anthropic.messages.create({
     model: normalizeModelStringForAPI(model),
     max_tokens: containsThinking ? TOKEN_COUNT_MAX_TOKENS : 1,
     messages: messagesToSend,
     tools: tools.length > 0 ? tools : undefined,
-    ...betas.length > 0 && { betas },
     metadata: getAPIMetadata(),
     ...getExtraBodyParams(),
     ...containsThinking && {
@@ -305424,6 +305658,8 @@ async function countTokensViaHaikuFallback(messages, tools) {
         budget_tokens: TOKEN_COUNT_THINKING_BUDGET
       }
     }
+  }, {
+    ...betas.length > 0 ? { headers: headersForBetas2(betas) } : {}
   });
   const usage = response.usage;
   const inputTokens = usage.input_tokens;
@@ -307224,6 +307460,11 @@ var init_apiMicrocompact = __esm(() => {
 
 // src/controller/loop.ts
 import { randomUUID as randomUUID17 } from "crypto";
+function headersForBetas3(betas) {
+  if (!betas || betas.length === 0)
+    return;
+  return { "anthropic-beta": betas.join(",") };
+}
 function adjustParamsForNonStreaming(params, maxTokensCap) {
   const cappedMaxTokens = Math.min(params.max_tokens, maxTokensCap);
   const adjustedParams = { ...params };
@@ -307301,13 +307542,16 @@ async function* executeNonStreamingRequest(clientOptions, retryOptions, paramsFr
     captureRequest(retryParams);
     onAttempt(attempt, start, retryParams.max_tokens);
     const adjustedParams = adjustParamsForNonStreaming(retryParams, MAX_NON_STREAMING_TOKENS);
+    const { betas: requestBetas, ...standardParams } = adjustedParams;
+    const headers = headersForBetas3(requestBetas);
     try {
-      return await anthropic.beta.messages.create({
-        ...adjustedParams,
-        model: normalizeModelStringForAPI(adjustedParams.model)
+      return await anthropic.messages.create({
+        ...standardParams,
+        model: normalizeModelStringForAPI(standardParams.model)
       }, {
         signal: retryOptions.signal,
-        timeout: fallbackTimeoutMs
+        timeout: fallbackTimeoutMs,
+        ...headers ? { headers } : {}
       });
     } catch (err2) {
       if (err2 instanceof APIUserAbortError)
@@ -307630,7 +307874,10 @@ ${deferredToolList}
         betasParams.push(STRUCTURED_OUTPUTS_BETA_HEADER);
       }
     }
-    const maxOutputTokens2 = retryContext?.maxTokensOverride || options2.maxOutputTokensOverride || getMaxOutputTokensForModel(options2.model);
+    const maxOutputTokens2 = retryContext?.maxTokensOverride || getMaxOutputTokensForModel(options2.model, {
+      override: options2.maxOutputTokensOverride,
+      upperLimitOverride: options2.maxOutputTokensLimitOverride
+    });
     const hasThinking = thinkingConfig.type !== "disabled" && !isEnvTruthy(resolveEnvVar("DISABLE_THINKING"));
     let thinking = undefined;
     if (hasThinking && modelSupportsThinking(options2.model)) {
@@ -307786,11 +308033,19 @@ ${deferredToolList}
         headlessProfilerCheckpoint("api_request_sent");
       }
       clientRequestId = shouldSendClientRequestId() ? randomUUID17() : undefined;
-      const result = await anthropic.beta.messages.create({ ...params, stream: true }, {
-        signal,
+      const { betas: requestBetas, ...standardParams } = params;
+      const headers = {
+        ...headersForBetas3(requestBetas) ?? {},
         ...clientRequestId && {
-          headers: { [CLIENT_REQUEST_ID_HEADER]: clientRequestId }
+          [CLIENT_REQUEST_ID_HEADER]: clientRequestId
         }
+      };
+      const result = await anthropic.messages.create({
+        ...standardParams,
+        stream: true
+      }, {
+        signal,
+        ...Object.keys(headers).length > 0 ? { headers } : {}
       }).withResponse();
       queryCheckpoint("query_response_headers_received");
       streamRequestId = result.request_id;
@@ -312103,6 +312358,11 @@ function stripToolReferenceBlocksFromUserMessage(message) {
     }
   };
 }
+function omitToolSearchCaller(block2) {
+  const blockWithoutCaller = { ...block2 };
+  delete blockWithoutCaller.caller;
+  return blockWithoutCaller;
+}
 function stripCallerFieldFromAssistantMessage(message) {
   const hasCallerField = message.message.content.some((block2) => block2.type === "tool_use" && ("caller" in block2) && block2.caller !== null);
   if (!hasCallerField) {
@@ -312116,12 +312376,7 @@ function stripCallerFieldFromAssistantMessage(message) {
         if (block2.type !== "tool_use") {
           return block2;
         }
-        return {
-          type: "tool_use",
-          id: block2.id,
-          name: block2.name,
-          input: block2.input
-        };
+        return omitToolSearchCaller(block2);
       })
     }
   };
@@ -312398,9 +312653,9 @@ function normalizeMessagesForAPI(messages, tools = []) {
                     input: normalizedInput
                   };
                 }
+                const blockWithoutCaller = omitToolSearchCaller(block2);
                 return {
-                  type: "tool_use",
-                  id: block2.id,
+                  ...blockWithoutCaller,
                   name: canonicalName,
                   input: normalizedInput
                 };
@@ -312597,6 +312852,15 @@ function mergeUserContentBlocks(a2, b) {
   }
   return [...a2.slice(0, -1), smooshed, ...toolResults];
 }
+function normalizeToolUseIdFromAPI(contentBlock) {
+  if (typeof contentBlock.id === "string" && contentBlock.id.trim().length > 0) {
+    return contentBlock.id;
+  }
+  logEvent("tengu_tool_use_missing_id_repaired", {
+    toolName: sanitizeToolNameForAnalytics(contentBlock.name)
+  });
+  return `toolu_${randomUUID18().replace(/-/g, "")}`;
+}
 function normalizeContentFromAPI(contentBlocks, tools, agentId) {
   if (!contentBlocks) {
     return [];
@@ -312635,6 +312899,7 @@ function normalizeContentFromAPI(contentBlocks, tools, agentId) {
         }
         return {
           ...contentBlock,
+          id: normalizeToolUseIdFromAPI(contentBlock),
           input: normalizedInput
         };
       }
@@ -458141,6 +458406,11 @@ var init_useDeferredHookMessages = __esm(() => {
 });
 
 // src/providers/anthropic/verifyApiKey.ts
+function headersForBetas4(betas) {
+  if (!betas || betas.length === 0)
+    return;
+  return { "anthropic-beta": betas.join(",") };
+}
 async function verifyApiKey(apiKey, isNonInteractiveSession) {
   if (isNonInteractiveSession) {
     return true;
@@ -458155,14 +458425,15 @@ async function verifyApiKey(apiKey, isNonInteractiveSession) {
       source: "verify_api_key"
     }), async (anthropic) => {
       const messages = [{ role: "user", content: "test" }];
-      await anthropic.beta.messages.create({
+      await anthropic.messages.create({
         model,
         max_tokens: 1,
         messages,
         temperature: 1,
-        ...betas.length > 0 && { betas },
         metadata: getAPIMetadata(),
         ...getExtraBodyParams()
+      }, {
+        ...betas.length > 0 ? { headers: headersForBetas4(betas) } : {}
       });
       return true;
     }, { maxRetries: 2, model, thinkingConfig: { type: "disabled" } }));
@@ -460063,8 +460334,10 @@ async function processUserInputBase(input, mode, setToolJSX, context6, pastedCon
       }
     }
   }
+  const processSlashCommand = getProcessSlashCommand();
   if (false) {}
-  const shouldExtractAttachments = !skipAttachments && inputString !== null && (mode !== "prompt" || effectiveSkipSlash || !inputString.startsWith("/"));
+  const shouldFallbackSlashToPrompt = inputString !== null && mode === "prompt" && !effectiveSkipSlash && inputString.startsWith("/") && !processSlashCommand;
+  const shouldExtractAttachments = !skipAttachments && inputString !== null && (mode !== "prompt" || effectiveSkipSlash || !inputString.startsWith("/") || shouldFallbackSlashToPrompt);
   queryCheckpoint("query_attachment_loading_start");
   const attachmentMessages = shouldExtractAttachments ? await toArray2(getAttachmentMessages(inputString, context6, ideSelection ?? null, [], messages, querySource)) : [];
   queryCheckpoint("query_attachment_loading_end");
@@ -460075,11 +460348,10 @@ async function processUserInputBase(input, mode, setToolJSX, context6, pastedCon
     return addImageMetadataMessage(await processBashCommand(inputString, precedingInputBlocks, attachmentMessages, context6, setToolJSX), imageMetadataTexts);
   }
   if (inputString !== null && !effectiveSkipSlash && inputString.startsWith("/")) {
-    const processSlashCommand = getProcessSlashCommand();
-    if (!processSlashCommand)
-      throw new SlashCommandHandlerNotRegisteredError("slash");
-    const slashResult = await processSlashCommand(inputString, precedingInputBlocks, imageContentBlocks, attachmentMessages, context6, setToolJSX, uuid3, isAlreadyProcessing, canUseTool);
-    return addImageMetadataMessage(slashResult, imageMetadataTexts);
+    if (processSlashCommand) {
+      const slashResult = await processSlashCommand(inputString, precedingInputBlocks, imageContentBlocks, attachmentMessages, context6, setToolJSX, uuid3, isAlreadyProcessing, canUseTool);
+      return addImageMetadataMessage(slashResult, imageMetadataTexts);
+    }
   }
   if (inputString !== null && mode === "prompt") {
     const trimmedInput = inputString.trim();
@@ -468371,21 +468643,21 @@ function extractLspInfoFromManifest(lspServers) {
   }
   return extractFromServerConfigRecord(lspServers);
 }
-function isRecord2(value) {
+function isRecord3(value) {
   return typeof value === "object" && value !== null;
 }
 function extractFromServerConfigRecord(serverConfigs) {
   const extensions = new Set;
   let command = null;
   for (const [_serverName, config2] of Object.entries(serverConfigs)) {
-    if (!isRecord2(config2)) {
+    if (!isRecord3(config2)) {
       continue;
     }
     if (!command && typeof config2.command === "string") {
       command = config2.command;
     }
     const extMapping = config2.extensionToLanguage;
-    if (isRecord2(extMapping)) {
+    if (isRecord3(extMapping)) {
       for (const ext of Object.keys(extMapping)) {
         extensions.add(ext.toLowerCase());
       }
@@ -479782,7 +480054,7 @@ function buildPrimarySection() {
   }, undefined, false, undefined, this);
   return [{
     label: "Version",
-    value: "0.4.12"
+    value: "0.4.14"
   }, {
     label: "Session name",
     value: nameValue
@@ -483063,7 +483335,7 @@ var init_OverageCreditUpsell = __esm(() => {
 });
 
 // src/providers/codex/usage.ts
-function isRecord3(value) {
+function isRecord4(value) {
   return typeof value === "object" && value !== null;
 }
 function asString(value) {
@@ -483082,7 +483354,7 @@ function toIsoFromUnixSeconds(value) {
   return new Date(seconds * 1000).toISOString();
 }
 function normalizeWindow(value) {
-  if (!isRecord3(value))
+  if (!isRecord4(value))
     return;
   const usedPercent = asNumber(value.used_percent) ?? asNumber(value.usedPercent);
   if (usedPercent === undefined)
@@ -483099,7 +483371,7 @@ function normalizeWindow(value) {
   };
 }
 function normalizeCredits(value) {
-  if (!isRecord3(value))
+  if (!isRecord4(value))
     return;
   const hasCredits = asBoolean(value.has_credits) ?? asBoolean(value.hasCredits) ?? false;
   const unlimited = asBoolean(value.unlimited) ?? false;
@@ -483114,7 +483386,7 @@ function normalizeCredits(value) {
   };
 }
 function normalizeSnapshot(value, fallbackLimitName) {
-  if (!isRecord3(value))
+  if (!isRecord4(value))
     return;
   const limitName = asString(value.limit_name) ?? asString(value.limitName) ?? asString(value.limit_id) ?? asString(value.limitId) ?? fallbackLimitName;
   const primary = normalizeWindow(value.primary) ?? normalizeWindow(value.primary_window);
@@ -483134,7 +483406,7 @@ function normalizeSnapshotsFromCollection(value, defaultLimitName = "codex") {
   if (Array.isArray(value)) {
     return value.map((item, index) => normalizeSnapshot(item, index === 0 ? defaultLimitName : `${defaultLimitName}-${index + 1}`)).filter((item) => item !== undefined);
   }
-  if (!isRecord3(value))
+  if (!isRecord4(value))
     return [];
   return Object.entries(value).map(([key, entry]) => normalizeSnapshot(entry, key)).filter((item) => item !== undefined);
 }
@@ -483168,7 +483440,7 @@ function normalizeCodexUsagePayload(payload) {
       snapshots: normalizeSnapshotsFromCollection(payload)
     };
   }
-  if (!isRecord3(payload)) {
+  if (!isRecord4(payload)) {
     return { snapshots: [] };
   }
   if ("rate_limit" in payload || "code_review_rate_limit" in payload || "additional_rate_limits" in payload || "credits" in payload) {
@@ -533101,8 +533373,8 @@ var AGENT_CREATION_SYSTEM_PROMPT, AGENT_MEMORY_INSTRUCTIONS = `
    - [domain-specific item 3]"
 
    Examples of domain-specific memory instructions:
-   - For a code-reviewer: "Update your agent memory as you discover code patterns, style conventions, common issues, and architectural decisions in this codebase."
-   - For a test-runner: "Update your agent memory as you discover test patterns, common failure modes, flaky tests, and testing best practices."
+   - For a code review agent: "Update your agent memory as you discover code patterns, style conventions, common issues, and architectural decisions in this codebase."
+   - For a test investigation agent: "Update your agent memory as you discover test patterns, common failure modes, flaky tests, and testing best practices."
    - For an architect: "Update your agent memory as you discover codepaths, library locations, key architectural decisions, and component relationships."
    - For a documentation writer: "Update your agent memory as you discover documentation patterns, API structures, and terminology conventions."
 
@@ -533151,29 +533423,30 @@ When a user describes what they want an agent to do, you will:
   - in the 'whenToUse' field of the JSON object, you should include examples of when this agent should be used.
   - examples should be of the form:
     - <example>
-      Context: The user is creating a test-runner agent that should be called after a logical chunk of code is written.
+      Context: The user is creating an agent that should review recently changed code after a logical chunk of implementation work.
       user: "Please write a function that checks if a number is prime"
       assistant: "Here is the relevant function: "
       <function call omitted for brevity only for this example>
       <commentary>
-      Since a significant piece of code was written, use the ${AGENT_TOOL_NAME} tool to launch the test-runner agent to run the tests.
+      Since a significant piece of code was written, use the ${AGENT_TOOL_NAME} tool with this newly created agent's exact identifier to review the change.
       </commentary>
-      assistant: "Now let me use the test-runner agent to run the tests"
+      assistant: "Now let me use the review agent to inspect the change."
     </example>
     - <example>
-      Context: User is creating an agent to respond to the word "hello" with a friendly jok.
-      user: "Hello"
-      assistant: "I'm going to use the ${AGENT_TOOL_NAME} tool to launch the greeting-responder agent to respond with a friendly joke"
+      Context: The user is creating an agent that should investigate flaky CI failures.
+      user: "The integration test failed again, can you look into the pattern?"
+      assistant: "I'm going to use the ${AGENT_TOOL_NAME} tool with this newly created agent's exact identifier to investigate the flaky CI failure pattern."
       <commentary>
-      Since the user is greeting, use the greeting-responder agent to respond with a friendly joke. 
+      Since the request matches the new agent's investigation scope, delegate the focused analysis to that exact agent type.
       </commentary>
     </example>
+  - Do not use fictional agent names in examples. Refer to "this newly created agent's exact identifier" or to the actual identifier you output in the JSON.
   - If the user mentioned or implied that the agent should be used proactively, you should include examples of this.
 - NOTE: Ensure that in the examples, you are making the assistant use the Agent tool and not simply respond directly to the task.
 
 Your output must be a valid JSON object with exactly these fields:
 {
-  "identifier": "A unique, descriptive identifier using lowercase letters, numbers, and hyphens (e.g., 'test-runner', 'api-docs-writer', 'code-formatter')",
+  "identifier": "A unique, descriptive identifier using lowercase letters, numbers, and hyphens. Use an identifier that matches the agent being created.",
   "whenToUse": "A precise, actionable description starting with 'Use this agent when...' that clearly defines the triggering conditions and use cases. Ensure you include examples as described above.",
   "systemPrompt": "The complete system prompt that will govern the agent's behavior, written in second person ('You are...', 'You will...') and structured for maximum clarity and effectiveness"
 }
@@ -536104,7 +536377,7 @@ var init_bridge_kick = __esm(() => {
 var call58 = async () => {
   return {
     type: "text",
-    value: `${"99.0.0"} (built ${"2026-06-24T10:02:32.669Z"})`
+    value: `${"99.0.0"} (built ${"2026-07-07T07:10:57.155Z"})`
   };
 }, version2, version_default;
 var init_version = __esm(() => {
@@ -558214,7 +558487,7 @@ function WelcomeV2() {
             dimColor: true,
             children: [
               "v",
-              "0.4.12",
+              "0.4.14",
               " "
             ]
           }, undefined, true, undefined, this)
@@ -558414,7 +558687,7 @@ function WelcomeV2() {
           dimColor: true,
           children: [
             "v",
-            "0.4.12",
+            "0.4.14",
             " "
           ]
         }, undefined, true, undefined, this)
@@ -558640,7 +558913,7 @@ function AppleTerminalWelcomeV2(t0) {
         dimColor: true,
         children: [
           "v",
-          "0.4.12",
+          "0.4.14",
           " "
         ]
       }, undefined, true, undefined, this);
@@ -558894,7 +559167,7 @@ function AppleTerminalWelcomeV2(t0) {
       dimColor: true,
       children: [
         "v",
-        "0.4.12",
+        "0.4.14",
         " "
       ]
     }, undefined, true, undefined, this);
@@ -569513,6 +569786,7 @@ class QueryEngine {
       maxTurns,
       maxBudgetUsd,
       maxOutputTokens,
+      maxOutputTokensLimit,
       contextWindow,
       compact: compact2,
       taskBudget,
@@ -569602,6 +569876,7 @@ class QueryEngine {
         theme: resolveThemeSetting(getGlobalConfig().theme),
         maxBudgetUsd,
         maxOutputTokens,
+        maxOutputTokensLimit,
         contextWindow,
         modelProviders: this.config.modelProviders,
         subagentDisallowedTools: this.config.subagentDisallowedTools
@@ -569707,6 +569982,7 @@ class QueryEngine {
         agentDefinitions: { activeAgents: agents2, allAgents: [] },
         maxBudgetUsd,
         maxOutputTokens,
+        maxOutputTokensLimit,
         contextWindow,
         modelProviders: this.config.modelProviders,
         subagentDisallowedTools: this.config.subagentDisallowedTools
@@ -569827,6 +570103,7 @@ class QueryEngine {
       querySource: "sdk",
       maxTurns,
       maxOutputTokensOverride: maxOutputTokens,
+      maxOutputTokensLimitOverride: maxOutputTokensLimit,
       compactRequest: compact2,
       taskBudget
     })) {
@@ -570197,6 +570474,7 @@ async function* ask({
   maxTurns,
   maxBudgetUsd,
   maxOutputTokens,
+  maxOutputTokensLimit,
   contextWindow,
   compact: compact2,
   taskBudget,
@@ -570242,6 +570520,7 @@ async function* ask({
     maxTurns,
     maxBudgetUsd,
     maxOutputTokens,
+    maxOutputTokensLimit,
     contextWindow,
     compact: compact2,
     taskBudget,
@@ -579756,7 +580035,7 @@ Usage: claude --remote "your task description"`, () => gracefulShutdown(1));
         pendingHookMessages
       }, renderAndRun);
     }
-  }).version("0.4.12 (OpenCow)", "-v, --version", "Output the version number");
+  }).version("0.4.14 (OpenCow)", "-v, --version", "Output the version number");
   program2.option("-w, --worktree [name]", "Create a new git worktree for this session (optionally specify a name)");
   program2.option("--tmux", "Create a tmux session for the worktree (requires --worktree). Uses iTerm2 native panes when available; use --tmux=classic for traditional tmux.");
   if (canUserConfigureAdvisor()) {
@@ -580402,7 +580681,7 @@ if (false) {}
 async function main2() {
   const args = process.argv.slice(2);
   if (args.length === 1 && (args[0] === "--version" || args[0] === "-v" || args[0] === "-V")) {
-    console.log(`${"0.4.12"} (OpenCow)`);
+    console.log(`${"0.4.14"} (OpenCow)`);
     return;
   }
   if (args.includes("--provider")) {
@@ -580520,4 +580799,4 @@ async function main2() {
 }
 main2();
 
-//# debugId=61239C1856C7488B64756E2164756E21
+//# debugId=C329783BF76928B664756E2164756E21