0sec-cli 0.22.0 → 0.22.4

package/chunks/{chunk-2GV4HQKE.js → chunk-QVIRWJ45.js}

@@ -1,9 +1,19 @@
 #!/usr/bin/env node
 import { createRequire as __0CreateRequire } from "node:module";
 const require = __0CreateRequire(import.meta.url);
+import {
+  isAdmittedSmolvmWorkbench
+} from "./chunk-2WWAPKAW.js";
+import {
+  MESSAGE_CACHE_BREAKPOINTS,
+  planMessageBreakpoints,
+  providerSupportsPromptCache,
+  readCacheUsage,
+  withCacheControl
+} from "./chunk-RS4MAN3Q.js";
 import {
   VERSION
-} from "./chunk-76KSBWOQ.js";
+} from "./chunk-6KFB6AYN.js";
 
 // packages/core/dist/agent/feature-presets.js
 var FP_MOAT_FLAGS = [
@@ -1029,66 +1039,25 @@ import { appendFileSync, existsSync, readFileSync, renameSync, writeFileSync } f
 import { homedir } from "node:os";
 import { join } from "node:path";
 
-// packages/core/dist/runtime/prompt-cache.js
-var MAX_CACHE_BREAKPOINTS = 4;
-var MESSAGE_CACHE_BREAKPOINTS = MAX_CACHE_BREAKPOINTS - 1;
-var BREAKPOINT_SPACING_BLOCKS = 15;
-var NATIVE_CACHE_PROVIDERS = /* @__PURE__ */ new Set(["anthropic"]);
-var OPT_IN_CACHE_PROVIDERS = /* @__PURE__ */ new Set(["z-ai", "kimi"]);
-function providerSupportsPromptCache(provider) {
-  if (NATIVE_CACHE_PROVIDERS.has(provider))
-    return true;
-  if (!OPT_IN_CACHE_PROVIDERS.has(provider))
-    return false;
-  return readExtraCacheProviders().has(provider);
-}
-function readExtraCacheProviders() {
-  const raw = process.env["ZERO_PROMPT_CACHE_EXTRA_PROVIDERS"];
+// packages/core/dist/runtime/workbench-provider-proxy.js
+function resolveWorkbenchProviderProxy(env2) {
+  const raw = env2.ZERO_WORKBENCH_PROVIDER_PROXY;
   if (!raw)
-    return /* @__PURE__ */ new Set();
-  return new Set(raw.split(",").map((entry) => entry.trim().toLowerCase()).filter((entry) => entry.length > 0));
-}
-function planMessageBreakpoints(messages, budget = MESSAGE_CACHE_BREAKPOINTS) {
-  if (messages.length === 0 || budget <= 0)
-    return [];
-  const lastIndex = messages.length - 1;
-  const picks = /* @__PURE__ */ new Set([lastIndex]);
-  const rollingBudget = Math.max(1, budget - 1);
-  let blocksSincePick = blockCount(messages[lastIndex]);
-  for (let i = lastIndex - 1; i >= 0 && picks.size < rollingBudget; i--) {
-    if (blocksSincePick >= BREAKPOINT_SPACING_BLOCKS) {
-      picks.add(i);
-      blocksSincePick = 0;
-    }
-    blocksSincePick += blockCount(messages[i]);
-  }
-  if (picks.size < budget)
-    picks.add(0);
-  return [...picks].sort((a, b) => a - b).slice(0, budget);
-}
-function blockCount(message) {
-  return message?.content.length ?? 0;
-}
-function withCacheControl(block) {
-  return { ...block, cache_control: { type: "ephemeral" } };
-}
-function readCacheUsage(raw) {
-  if (typeof raw !== "object" || raw === null)
     return void 0;
-  const usage = raw;
-  const uncachedInput = numberOr(usage.input_tokens, 0);
-  const outputTokens = numberOr(usage.output_tokens, 0);
-  const cacheRead = numberOr(usage.cache_read_input_tokens, void 0);
-  const cacheWrite = numberOr(usage.cache_creation_input_tokens, void 0);
-  return {
-    inputTokens: uncachedInput + (cacheRead ?? 0) + (cacheWrite ?? 0),
-    outputTokens,
-    ...cacheRead !== void 0 ? { cachedInputTokens: cacheRead } : {},
-    ...cacheWrite !== void 0 ? { cacheWriteTokens: cacheWrite } : {}
-  };
-}
-function numberOr(value, fallback) {
-  return typeof value === "number" && Number.isFinite(value) ? value : fallback;
+  if (!isAdmittedSmolvmWorkbench())
+    throw new Error("Provider proxy requires authenticated SmolVM workbench admission");
+  let url;
+  try {
+    url = new URL(raw);
+  } catch {
+    throw new Error("Invalid guest provider proxy address");
+  }
+  if (url.protocol !== "http:" || url.hostname !== "127.0.0.1" || !url.port || url.pathname !== "/provider/request" || url.username || url.password || url.search || url.hash)
+    throw new Error("Provider proxy must be the fixed guest loopback route");
+  const model = env2.ZERO_WORKBENCH_PROVIDER_MODEL;
+  if (!model || !/^[A-Za-z0-9._-]{1,128}$/.test(model))
+    throw new Error("Missing exact workbench provider model");
+  return { url: url.href, model };
 }
 
 // packages/core/dist/runtime/llm-api.js
@@ -1098,6 +1067,31 @@ function readResponsesCachedTokens(usage) {
   const cached = Number(details?.cached_tokens ?? 0);
   return Number.isFinite(cached) && cached > 0 ? { cachedInputTokens: cached } : {};
 }
+function readChatUsage(raw) {
+  if (!raw || typeof raw !== "object")
+    return void 0;
+  const usage = raw;
+  if (usage.prompt_tokens === void 0 && usage.completion_tokens === void 0)
+    return void 0;
+  return {
+    inputTokens: Number(usage.prompt_tokens ?? 0),
+    outputTokens: Number(usage.completion_tokens ?? 0),
+    ...readResponsesCachedTokens({ input_tokens_details: usage.prompt_tokens_details })
+  };
+}
+function readGoogleUsage(raw) {
+  if (!raw || typeof raw !== "object")
+    return void 0;
+  const usage = raw;
+  if (usage.promptTokenCount === void 0 && usage.candidatesTokenCount === void 0)
+    return void 0;
+  const cached = Number(usage.cachedContentTokenCount ?? 0);
+  return {
+    inputTokens: Number(usage.promptTokenCount ?? 0),
+    outputTokens: Number(usage.candidatesTokenCount ?? 0) + Number(usage.thoughtsTokenCount ?? 0),
+    ...cached > 0 ? { cachedInputTokens: cached } : {}
+  };
+}
 function safeParseJson(raw) {
   if (!raw)
     return {};
@@ -1107,6 +1101,18 @@ function safeParseJson(raw) {
     return { _raw: raw };
   }
 }
+function parseResponsesFunctionCall(item) {
+  if (item.status !== void 0 && item.status !== "completed" || typeof item.call_id !== "string" || !item.call_id.trim() || typeof item.name !== "string" || !item.name.trim() || typeof item.arguments !== "string")
+    return void 0;
+  try {
+    const input = JSON.parse(item.arguments);
+    if (!input || typeof input !== "object" || Array.isArray(input))
+      return void 0;
+    return { type: "tool_use", id: item.call_id, name: item.name, input };
+  } catch {
+    return void 0;
+  }
+}
 function isWireBlockArray(blocks) {
   return blocks.every((block) => block !== null && typeof block === "object" && !Array.isArray(block) && typeof block.type === "string");
 }
@@ -2187,6 +2193,13 @@ var DEFAULT_PROVIDER_MODELS = {
 };
 var AUTO_MODEL_SENTINEL = "auto";
 function detectProvider(configApiKey, preferredModel, env2, configProvider) {
+  const proxy = resolveWorkbenchProviderProxy(env2);
+  if (proxy) {
+    if ([configProvider, env2.ZERO_SELECTED_PROVIDER, env2.ZERO_FORCE_PROVIDER].some((provider) => provider && provider !== "chatgpt-codex")) {
+      throw new Error("Workbench provider broker supports chatgpt-codex only; no fallback is permitted");
+    }
+    return { provider: "chatgpt-codex", apiKey: "", baseUrl: proxy.url, defaultModel: proxy.model, wireApi: "responses" };
+  }
   if (configProvider !== void 0 && !Object.hasOwn(DEFAULT_PROVIDER_MODELS, configProvider)) {
     throw new Error(`RuntimeConfig.provider is unsupported: ${configProvider}`);
   }
@@ -2204,8 +2217,10 @@ function detectProvider(configApiKey, preferredModel, env2, configProvider) {
       throw new Error(`${source} is unsupported: ${pinnedProviderRaw}`);
     }
     const provider = pinnedProviderRaw;
-    const model = preferredModel ?? env2["ZERO_MODEL"] ?? (configProvider !== void 0 ? DEFAULT_PROVIDER_MODELS[provider] : void 0);
+    const model = preferredModel ?? env2["ZERO_MODEL"] ?? (provider === "azure" ? env2.AZURE_OPENAI_MODEL ?? parseCodexAzureConfig(env2).model : void 0) ?? (configProvider !== void 0 ? DEFAULT_PROVIDER_MODELS[provider] : void 0);
     if (model === void 0 || model === "") {
+      if (provider === "azure")
+        throw new Error("Choose an Azure deployment model in Connections to finish setup.");
       throw new Error(`${source} requires an explicit model`);
     }
     if (configApiKey && provider === "chatgpt-codex") {
@@ -2520,6 +2535,7 @@ var LlmApiRuntime = class _LlmApiRuntime {
   // Not readonly: a live provider switch re-freezes env from the new account.
   env;
   codexAuthState;
+  workbenchProviderProxy;
   geminiAuthState;
   provider;
   apiKey;
@@ -2552,6 +2568,7 @@ var LlmApiRuntime = class _LlmApiRuntime {
         timeout: Math.min(timeout, inherited.config.timeout || 12e4)
       };
       this.env = inherited.env;
+      this.workbenchProviderProxy = inherited.workbenchProviderProxy;
       this.provider = inherited.provider;
       this.apiKey = inherited.apiKey;
       this.baseUrl = inherited.baseUrl;
@@ -2591,8 +2608,9 @@ var LlmApiRuntime = class _LlmApiRuntime {
       ...config.agentModels ? { agentModels: Object.freeze({ ...config.agentModels }) } : {}
     };
     this.env = Object.freeze({ ...process.env, ...config.env });
+    this.workbenchProviderProxy = resolveWorkbenchProviderProxy(this.env);
     this.azureConfig = parseCodexAzureConfig(this.env);
-    this.fallbackChain = getFallbackChain(this.env).map((entry) => ({
+    this.fallbackChain = (this.workbenchProviderProxy ? [] : getFallbackChain(this.env)).map((entry) => ({
       ...entry,
       credentials: resolveFailoverProvider(entry.provider, entry.model, this.env)
     }));
@@ -2603,7 +2621,7 @@ var LlmApiRuntime = class _LlmApiRuntime {
     this.baseUrl = detected.baseUrl;
     this.wireApi = detected.wireApi;
     this.codexAuthState = void 0;
-    if (this.provider === "chatgpt-codex" || this.fallbackChain.some((entry) => entry.provider === "chatgpt-codex")) {
+    if (!this.workbenchProviderProxy && (this.provider === "chatgpt-codex" || this.fallbackChain.some((entry) => entry.provider === "chatgpt-codex"))) {
       if (readChatGptCodexEnv(this.env) || readChatGptCodexAuthFile(this.env)) {
         this.codexAuthState = resolveChatGptCodexAuthState(this.env);
       }
@@ -2661,14 +2679,50 @@ var LlmApiRuntime = class _LlmApiRuntime {
     }
     return createHash("sha256").update(JSON.stringify(identity)).digest("hex");
   }
+  /** Documented effort controls only on routes which send Responses reasoning. */
+  reasoningConfiguration() {
+    if (this.wireApi !== "responses" || !["openai", "azure", "chatgpt-codex"].includes(this.provider))
+      return null;
+    const model = this.model.replace(/^openai\//, "");
+    let options;
+    if (/^gpt-(?:6-astra|6\.1-sol)(?:-20\d\d-\d\d-\d\d)?$/.test(model))
+      options = ["low", "medium", "high", "xhigh", "max"];
+    else if (/^gpt-(?:6-sol|6-luna|5\.6(?:-sol|-terra|-luna)?)(?:-20\d\d-\d\d-\d\d)?$/.test(model))
+      options = ["none", "low", "medium", "high", "xhigh", "max"];
+    else
+      return null;
+    const effort = this.reasoningEffort ?? defaultReasoningEffort(this.model) ?? "medium";
+    return options.includes(effort) ? { effort, options } : null;
+  }
+  /** Apply an effort for subsequent requests and same-model worker forks. */
+  setReasoningEffort(effort) {
+    const configuration = this.reasoningConfiguration();
+    if (!configuration || !configuration.options.includes(effort))
+      throw new Error("This reasoning effort is not supported by the selected model and connection.");
+    this.reasoningEffort = effort;
+  }
   /** Discover models using this runtime's captured account, including after a separate login changes. */
   async codexModelCatalog(signal) {
+    if (this.workbenchProviderProxy) {
+      signal?.throwIfAborted();
+      return [{ id: this.workbenchProviderProxy.model }];
+    }
     const state = this.codexAuthState;
     if (this.provider !== "chatgpt-codex" || !state)
       throw new Error("No active Codex subscription");
-    const { loadCodexModelCatalog } = await import("./codex-models-7TLJSY3L.js");
+    const { loadCodexModelCatalog } = await import("./codex-models-Z43V3427.js");
     return loadCodexModelCatalog({ signal, resolveCredentials: () => refreshChatGptCodexAuthState(state) });
   }
+  /** Host-only inference grant bound to this captured account, including after
+   * later logins or runtime reconfiguration. Never serialize the closure or
+   * its result into guest admission, workspace files or environment. */
+  workbenchCredentialResolver() {
+    const state = this.codexAuthState;
+    if (this.workbenchProviderProxy || this.provider !== "chatgpt-codex" || !state) {
+      throw new Error("Host workbench credential broker requires a captured chatgpt-codex account");
+    }
+    return () => refreshChatGptCodexAuthState(state);
+  }
   /**
    * Mutate the live selection in place so the NEXT turn (the engine reads
    * `config.runtime` per turn) and the NEXT `forkForSubagent` pick up the new
@@ -2685,6 +2739,7 @@ var LlmApiRuntime = class _LlmApiRuntime {
         model: sel.model,
         ...sel.agentModels !== void 0 ? { agentModels: sel.agentModels } : {},
         ...sel.singleModel !== void 0 ? { singleModel: sel.singleModel } : {},
+        ...sel.autoRoute !== void 0 ? { autoRoute: sel.autoRoute } : {},
         ...sel.env !== void 0 ? { env: sel.env } : {}
       };
       this.applyConfiguration(merged);
@@ -2696,6 +2751,9 @@ var LlmApiRuntime = class _LlmApiRuntime {
     if (sel.singleModel !== void 0) {
       this.config = { ...this.config, singleModel: sel.singleModel };
     }
+    if (sel.autoRoute !== void 0) {
+      this.config = { ...this.config, autoRoute: sel.autoRoute };
+    }
     if (sel.model !== void 0) {
       const model = this.provider === "opencode" ? opencodeModelId(sel.model) : this.provider === "copilot" ? copilotModelId(sel.model) : sel.model;
       if (model !== this.model) {
@@ -2727,6 +2785,8 @@ var LlmApiRuntime = class _LlmApiRuntime {
    * failover chain uses — including auth-only providers (chatgpt-codex OAuth).
    */
   accessibleProviders() {
+    if (this.workbenchProviderProxy)
+      return ["chatgpt-codex"];
     const out = [];
     for (const provider of Object.keys(DEFAULT_PROVIDER_MODELS)) {
       const probeModel = DEFAULT_PROVIDER_MODELS[provider] || "probe";
@@ -2743,6 +2803,8 @@ var LlmApiRuntime = class _LlmApiRuntime {
    * provider has creds, so this is a helpful starting set, not the whole bound.
    */
   accessibleModels() {
+    if (this.workbenchProviderProxy)
+      return [this.workbenchProviderProxy.model];
     const models = /* @__PURE__ */ new Set();
     if (this.model)
       models.add(this.model);
@@ -2760,6 +2822,8 @@ var LlmApiRuntime = class _LlmApiRuntime {
    * accessible defaults as a fallback for ids the router does not pattern-match.
    */
   isModelAccessible(model) {
+    if (this.workbenchProviderProxy)
+      return this.accessibleModels().includes(model);
     if (providerForModel(model, this.env) !== void 0)
       return true;
     return this.accessibleModels().includes(model);
@@ -2849,6 +2913,17 @@ var LlmApiRuntime = class _LlmApiRuntime {
   resolvedModel() {
     return this.model;
   }
+  /** JSON-safe routing policy, not the runtime's private environment or configuration. */
+  modelSelection() {
+    return {
+      agentModels: { ...this.config.agentModels },
+      singleModel: this.config.singleModel === true,
+      autoRoute: this.config.autoRoute === true
+    };
+  }
+  resolvedProvider() {
+    return this.provider;
+  }
   /** Price discovered IDs on their actual route without changing the wire model. */
   resolvedPricingModel() {
     if (this.provider === "chatgpt-codex" || this.provider === "copilot" || this.provider === "google")
@@ -3252,7 +3327,13 @@ Missing: ${missing.join("; ")}
       abort?.throwIfCancelled();
       let res;
       try {
-        res = await fetch(this.buildUrl(), {
+        res = this.workbenchProviderProxy ? await fetch(this.workbenchProviderProxy.url, {
+          method: "POST",
+          redirect: "error",
+          headers: { "Content-Type": "application/json" },
+          body: JSON.stringify({ provider: this.provider, model: this.model, body: bodyFactory() }),
+          signal
+        }) : await fetch(this.buildUrl(), {
           method: "POST",
           headers: await this.ensureFreshHeaders(),
           body: bodyFactory(),
@@ -3377,6 +3458,22 @@ Missing: ${missing.join("; ")}
   // ── Legacy Runtime interface (single-prompt) ──
   async execute(prompt, context) {
     const start = Date.now();
+    if (context?.signal?.aborted) {
+      return { output: "", exitCode: null, timedOut: false, durationMs: 0, error: "API execution cancelled." };
+    }
+    if (this.workbenchProviderProxy) {
+      const result = await this.executeNative(context?.systemPrompt ?? "", [
+        { role: "user", content: [{ type: "text", text: prompt }] }
+      ], [], void 0, context?.signal);
+      return {
+        output: result.content.filter((block) => block.type === "text").map((block) => block.text).join("\n"),
+        exitCode: result.cancelled ? null : result.stopReason === "error" ? 1 : 0,
+        timedOut: false,
+        durationMs: Date.now() - start,
+        usage: result.usage,
+        error: result.error
+      };
+    }
     if (!this.apiKey && this.provider !== "chatgpt-codex" && this.provider !== "google") {
       return {
         output: "",
@@ -3389,9 +3486,11 @@ Missing: ${missing.join("; ")}
     const systemPrompt = context?.systemPrompt ?? "";
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(), this.config.timeout || 12e4);
+    const abort = composeCallAbort(controller.signal, context?.signal);
     try {
       let res;
       do {
+        abort.throwIfCancelled();
         if (this.isOpenAICompat && this.wireApi === "chat_completions") {
           const messages = [];
           if (systemPrompt) {
@@ -3404,7 +3503,7 @@ Missing: ${missing.join("; ")}
             messages,
             // See executeNative: explicit reasoning_effort passthrough only.
             ...this.reasoningEffort ? { reasoning_effort: this.reasoningEffort } : {}
-          }), controller.signal);
+          }), abort.signal, abort);
         } else if (this.isOpenAICompat && this.wireApi === "responses") {
           const input = [];
           if (systemPrompt) {
@@ -3422,7 +3521,7 @@ Missing: ${missing.join("; ")}
             model: this.model,
             input,
             ...isCodex ? { store: false } : { max_output_tokens: NATIVE_COMPLETION_TOKEN_LIMIT }
-          }), controller.signal);
+          }), abort.signal, abort);
         } else if (this.isGeminiCodeAssist) {
           const request = {
             ...systemPrompt ? { systemInstruction: { parts: [{ text: systemPrompt }] } } : {},
@@ -3430,13 +3529,13 @@ Missing: ${missing.join("; ")}
             generationConfig: { maxOutputTokens: NATIVE_COMPLETION_TOKEN_LIMIT }
           };
           const geminiBody = await this.wrapGeminiCodeAssistBody(request);
-          res = await this.postWithRetry(() => JSON.stringify(geminiBody), controller.signal);
+          res = await this.postWithRetry(() => JSON.stringify(geminiBody), abort.signal, abort);
         } else if (this.isGoogleWire) {
           res = await this.postWithRetry(() => JSON.stringify({
             ...systemPrompt ? { systemInstruction: { parts: [{ text: systemPrompt }] } } : {},
             contents: [{ role: "user", parts: [{ text: prompt }] }],
             generationConfig: { maxOutputTokens: NATIVE_COMPLETION_TOKEN_LIMIT }
-          }), controller.signal);
+          }), abort.signal, abort);
         } else if (this.isAnthropicWire) {
           res = await this.postWithRetry(() => JSON.stringify({
             model: this.model,
@@ -3444,12 +3543,11 @@ Missing: ${missing.join("; ")}
             ...this.anthropicThinkingField(),
             ...systemPrompt ? { system: systemPrompt } : {},
             messages: [{ role: "user", content: prompt }]
-          }), controller.signal);
+          }), abort.signal, abort);
         } else {
           throw new Error(`execute: provider ${this.provider} is not mapped to a wire`);
         }
       } while (!res);
-      clearTimeout(timer);
       const body = await res.text();
       if (!res.ok) {
         appendNativeTrace({
@@ -3483,12 +3581,9 @@ Missing: ${missing.join("; ")}
       if (this.isAnthropicWire) {
         usage = readCacheUsage(json.usage);
       } else if ((this.isGoogleWire || this.isGeminiCodeAssist) && json.usageMetadata) {
-        usage = {
-          inputTokens: json.usageMetadata.promptTokenCount ?? 0,
-          outputTokens: json.usageMetadata.candidatesTokenCount ?? 0
-        };
+        usage = readGoogleUsage(json.usageMetadata);
       } else if (this.isOpenAICompat && json.usage) {
-        usage = this.wireApi === "chat_completions" ? { inputTokens: json.usage.prompt_tokens ?? 0, outputTokens: json.usage.completion_tokens ?? 0 } : { inputTokens: json.usage.input_tokens ?? 0, outputTokens: json.usage.output_tokens ?? 0 };
+        usage = this.wireApi === "chat_completions" ? readChatUsage(json.usage) : { inputTokens: json.usage.input_tokens ?? 0, outputTokens: json.usage.output_tokens ?? 0, ...readResponsesCachedTokens(json.usage) };
       }
       return {
         output: text,
@@ -3499,6 +3594,9 @@ Missing: ${missing.join("; ")}
       };
     } catch (err) {
       clearTimeout(timer);
+      if (abort.operatorAborted()) {
+        return { output: "", exitCode: null, timedOut: false, durationMs: Date.now() - start, error: "API execution cancelled." };
+      }
       if (err instanceof QuotaExhaustedError) {
         return {
           output: "",
@@ -3517,6 +3615,9 @@ Missing: ${missing.join("; ")}
         durationMs: Date.now() - start,
         error: timedOut ? `${this.providerLabel} API request timed out` : `${this.providerLabel} API error: ${msg}`
       };
+    } finally {
+      clearTimeout(timer);
+      abort.dispose();
     }
   }
   // ── Native Runtime interface (structured messages + tool_use) ──
@@ -3903,62 +4004,7 @@ Missing: ${missing.join("; ")}
         const finishReason = choice?.finish_reason;
         stopReason = finishReason === "tool_calls" || finishReason === "function_call" ? "tool_use" : finishReason === "length" ? "max_tokens" : "end_turn";
         if (json.usage) {
-          usage = {
-            inputTokens: json.usage.prompt_tokens ?? 0,
-            outputTokens: json.usage.completion_tokens ?? 0
-          };
-        }
-      } else if (this.isOpenAICompat && this.wireApi === "responses") {
-        content = [];
-        const reasoningSummaries = [];
-        const rawOutput = json.output ?? [];
-        if (rawOutput.length > 0) {
-          providerRaw = {
-            provider: this.provider,
-            model: this.model,
-            wireApi: this.wireApi,
-            output: rawOutput
-          };
-        }
-        for (const item of json.output ?? []) {
-          if (item.type === "function_call") {
-            content.push({
-              type: "tool_use",
-              id: item.call_id,
-              name: item.name,
-              input: safeParseJson(item.arguments)
-            });
-            continue;
-          }
-          if (item.type === "reasoning") {
-            const summaryParts = Array.isArray(item.summary) ? item.summary.map((block) => typeof block.text === "string" ? block.text : "").filter((text) => text.trim().length > 0) : [];
-            const reasoningText = summaryParts.join("\n").trim();
-            if (reasoningText)
-              reasoningSummaries.push(reasoningText);
-            continue;
-          }
-          for (const block of item.content ?? []) {
-            if (block.type === "output_text") {
-              content.push({ type: "text", text: block.text });
-            } else if (block.type === "summary_text" || block.type === "reasoning_text") {
-              const text = typeof block.text === "string" ? block.text : "";
-              if (text.trim())
-                reasoningSummaries.push(text);
-            }
-          }
-        }
-        if (callbacks?.onThinking && reasoningSummaries.length > 0) {
-          callbacks.onThinking(reasoningSummaries.join("\n"));
-        }
-        stopReason = content.some((block) => block.type === "tool_use") ? "tool_use" : "end_turn";
-        if (json.usage) {
-          usage = {
-            inputTokens: json.usage.input_tokens ?? 0,
-            outputTokens: json.usage.output_tokens ?? 0,
-            // Responses `input_tokens` already includes the cached span, so
-            // this is instrumentation only — see the streaming path.
-            ...readResponsesCachedTokens(json.usage)
-          };
+          usage = readChatUsage(json.usage);
         }
       } else if (this.isGoogleWire || this.isGeminiCodeAssist) {
         const candidate = json.candidates?.[0];
@@ -3990,10 +4036,7 @@ Missing: ${missing.join("; ")}
         const finishReason = String(candidate?.finishReason ?? "").toUpperCase();
         stopReason = content.some((block) => block.type === "tool_use") ? "tool_use" : finishReason === "MAX_TOKENS" ? "max_tokens" : "end_turn";
         if (json.usageMetadata) {
-          usage = {
-            inputTokens: json.usageMetadata.promptTokenCount ?? 0,
-            outputTokens: json.usageMetadata.candidatesTokenCount ?? 0
-          };
+          usage = readGoogleUsage(json.usageMetadata);
         }
       } else {
         const rawBlocks = json.content ?? [];
@@ -4135,6 +4178,9 @@ Missing: ${missing.join("; ")}
     const eventTypes = /* @__PURE__ */ new Set();
     const identifier = (value) => typeof value === "string" && /^[A-Za-z0-9_.:-]{1,96}$/.test(value) ? value : null;
     let completedResponse = null;
+    let outputCapResponse;
+    let assistantText = "";
+    const pendingFunctionCalls = /* @__PURE__ */ new Set();
     let streamFailure;
     let responseUsage;
     const streamedOutputItems = [];
@@ -4239,9 +4285,13 @@ Missing: ${missing.join("; ")}
             };
             callbacks?.onUsage?.(responseUsage);
           }
+          const incomplete = response?.incomplete_details;
+          if (type === "response.incomplete" && response?.status === "incomplete" && incomplete?.reason === "max_output_tokens" && response.error == null && event.error == null && event.code == null && !completedResponse && malformedEvents === 0) {
+            outputCapResponse = response;
+            break responses;
+          }
           if (!terminal || !response || response.error != null || (type === "response.done" || response.status !== void 0) && response.status !== "completed") {
             const error = response?.error ?? event.error;
-            const incomplete = response?.incomplete_details;
             streamFailure = {
               event: type,
               status: identifier(response?.status),
@@ -4257,6 +4307,7 @@ Missing: ${missing.join("; ")}
         if (type === "response.output_text.delta" || this.provider === "openrouter" && type === "response.content_part.delta") {
           const delta = typeof event.delta === "string" ? event.delta : "";
           if (delta) {
+            assistantText += delta;
             callbacks?.onDelta?.("assistant_response", delta);
           }
           continue;
@@ -4278,9 +4329,18 @@ Missing: ${missing.join("; ")}
           }
           continue;
         }
+        if (type === "response.output_item.added") {
+          const item = event.item;
+          if (item?.type === "function_call") {
+            pendingFunctionCalls.add(String(item.id ?? item.call_id ?? event.output_index ?? pendingFunctionCalls.size));
+          }
+          continue;
+        }
         if (type === "response.output_item.done") {
           const item = event.item;
           if (item && typeof item.type === "string") {
+            if (item.type === "function_call")
+              pendingFunctionCalls.add(String(item.id ?? item.call_id ?? event.output_index ?? pendingFunctionCalls.size));
             streamedOutputItems.push(item);
           }
           continue;
@@ -4288,6 +4348,46 @@ Missing: ${missing.join("; ")}
       }
     }
     emitThinking(true);
+    if (outputCapResponse) {
+      void reader.cancel().catch(() => {
+      });
+      const finalOutput = Array.isArray(outputCapResponse.output) ? outputCapResponse.output : [];
+      const output = finalOutput.length > 0 ? finalOutput : streamedOutputItems;
+      let finalText = "";
+      for (const item of output) {
+        if (item.type === "function_call")
+          pendingFunctionCalls.add(String(item.id ?? item.call_id ?? pendingFunctionCalls.size));
+        if (!Array.isArray(item.content))
+          continue;
+        for (const block of item.content) {
+          if (block.type === "output_text" && typeof block.text === "string")
+            finalText += block.text;
+        }
+      }
+      const text = finalText.startsWith(assistantText) ? finalText : assistantText || finalText;
+      if (operatorSignal?.aborted) {
+        return { ...this.cancelledResult(start), content: [{ type: "text", text }], usage: responseUsage };
+      }
+      const checkpoint = {
+        reason: "max_output_tokens",
+        ...identifier(outputCapResponse.id) ? { responseId: String(outputCapResponse.id) } : {},
+        discardedToolCalls: pendingFunctionCalls.size
+      };
+      appendNativeTrace({
+        kind: "native-response-checkpoint",
+        provider: this.providerLabel,
+        wireApi: this.wireApi,
+        checkpoint,
+        usage: responseUsage ?? null
+      });
+      return {
+        content: [{ type: "text", text }],
+        stopReason: "max_tokens",
+        checkpoint,
+        usage: responseUsage,
+        durationMs: Date.now() - start
+      };
+    }
     if (!completedResponse || streamFailure) {
       if (streamFailure) {
         void reader.cancel().catch(() => {
@@ -4333,18 +4433,25 @@ Missing: ${missing.join("; ")}
       })),
       topLevelKeys: Object.keys(completedResponse)
     });
-    const completedOutput = completedResponse.output ?? [];
-    const outputItems = streamedOutputItems.length > 0 ? streamedOutputItems : completedOutput;
+    const completedOutput = Array.isArray(completedResponse.output) ? completedResponse.output : [];
+    const outputItems = completedOutput.length > 0 ? completedOutput : streamedOutputItems;
     const content = [];
     const reasoningSummaries = [];
+    const callIds = /* @__PURE__ */ new Set();
     for (const item of outputItems) {
       if (item.type === "function_call") {
-        content.push({
-          type: "tool_use",
-          id: String(item.call_id),
-          name: String(item.name),
-          input: safeParseJson(String(item.arguments ?? "{}"))
-        });
+        const call = parseResponsesFunctionCall(item);
+        if (!call || callIds.has(call.id) || malformedEvents > 0) {
+          return {
+            content: [{ type: "text", text: assistantText }],
+            stopReason: "error",
+            usage: responseUsage,
+            durationMs: Date.now() - start,
+            error: `${this.providerLabel} API error: Responses completed with an incomplete or malformed function call`
+          };
+        }
+        callIds.add(call.id);
+        content.push(call);
         continue;
       }
       if (item.type === "reasoning") {
@@ -4360,6 +4467,9 @@ Missing: ${missing.join("; ")}
         }
       }
     }
+    const observedTextFallback = !!assistantText && !content.some((block) => block.type === "text");
+    if (observedTextFallback)
+      content.unshift({ type: "text", text: assistantText });
     if (lastThinkingEmit === 0 && reasoningSummaries.length > 0) {
       thinkingText = reasoningSummaries.join("\n");
       emitThinking(true);
@@ -4373,7 +4483,7 @@ Missing: ${missing.join("; ")}
       // reasoning items with their `encrypted_content` still attached, each
       // immediately followed by the item it produced. Handing it back lets the
       // next turn replay it verbatim instead of re-deriving the reasoning.
-      ...outputItems.length > 0 ? {
+      ...!observedTextFallback && outputItems.length > 0 ? {
         providerRaw: {
           provider: this.provider,
           model: this.model,
@@ -4384,6 +4494,8 @@ Missing: ${missing.join("; ")}
     };
   }
   async isAvailable() {
+    if (this.workbenchProviderProxy)
+      return true;
     if (this.provider === "chatgpt-codex") {
       return this.codexAuthState !== void 0;
     }