whale-code 6.5.5 → 6.5.6

package/dist/shared/api-client.js

@@ -4,16 +4,20 @@
  * Replaces proxy+direct call paths in agent-loop, subagent, teammate.
  * CLI-only: always calls through server proxy (no direct Anthropic SDK usage).
  */
-import { getContextManagement, getMaxOutputTokens, getThinkingConfig, isRetryableError, addPromptCaching, } from "./agent-core.js";
+
+import { getContextManagement, getMaxOutputTokens, getThinkingConfig, isRetryableError, addPromptCaching } from "./agent-core.js";
 import { getProvider } from "./constants.js";
 // ============================================================================
 // CONSTANTS
 // ============================================================================
+
 const MAX_RETRIES = 3;
 const RETRY_BASE_DELAY_MS = 1000;
+
 // ============================================================================
 // API REQUEST CONFIG BUILDER
 // ============================================================================
+
 /**
  * Build all Anthropic API config from a simple profile.
  *
@@ -23,293 +27,359 @@ const RETRY_BASE_DELAY_MS = 1000;
  * - 'teammate': clear at 80K/keep 3, no compaction
  */
 export function buildAPIRequest(opts) {
-    const { model, contextProfile, thinkingEnabled = false, maxOutputTokens, contextOverrides } = opts;
-    // Context management config per profile
-    // Non-Anthropic models (Gemini, OpenAI, etc.) don't support Anthropic betas or context management
-    const provider = getProvider(model);
-    const isAnthropicModel = provider === "anthropic" || provider === "bedrock";
-    let betas = isAnthropicModel ? ["context-management-2025-06-27"] : [];
-    let edits = [];
-    switch (contextProfile) {
-        case "main": {
-            const ctxMgmt = getContextManagement(model, contextOverrides);
-            betas = [...ctxMgmt.betas];
-            edits = ctxMgmt.config.edits;
-            break;
+  const {
+    model,
+    contextProfile,
+    thinkingEnabled = false,
+    maxOutputTokens,
+    contextOverrides
+  } = opts;
+
+  // Context management config per profile
+  // Non-Anthropic models (Gemini, OpenAI, etc.) don't support Anthropic betas or context management
+  const provider = getProvider(model);
+  const isAnthropicModel = provider === "anthropic" || provider === "bedrock";
+  let betas = isAnthropicModel ? ["context-management-2025-06-27"] : [];
+  let edits = [];
+  switch (contextProfile) {
+    case "main":
+      {
+        const ctxMgmt = getContextManagement(model, contextOverrides);
+        betas = [...ctxMgmt.betas];
+        edits = ctxMgmt.config.edits;
+        break;
+      }
+    case "subagent":
+      if (isAnthropicModel) {
+        edits = [{
+          type: "clear_thinking_20251015",
+          keep: {
+            type: "thinking_turns",
+            value: 1
+          }
+        }, {
+          type: "clear_tool_uses_20250919",
+          trigger: {
+            type: "input_tokens",
+            value: 60_000
+          },
+          keep: {
+            type: "tool_uses",
+            value: 2
+          }
+        }];
+        // Compaction for sub-agents on models that support it (lower threshold than main agent).
+        // Trigger at 80K — above the 60K clear_tool_uses threshold so clearing fires first.
+        const supportsCompaction = model.includes("opus-4-6") || model.includes("sonnet-4-6");
+        if (supportsCompaction) {
+          edits.push({
+            type: "compact_20260112",
+            trigger: {
+              type: "input_tokens",
+              value: 80_000
+            },
+            pause_after_compaction: true,
+            instructions: "Summarize preserving: task goal, files found/modified, key findings, next steps."
+          });
+          betas.push("compact-2026-01-12");
         }
-        case "subagent":
-            if (isAnthropicModel) {
-                edits = [
-                    {
-                        type: "clear_thinking_20251015",
-                        keep: { type: "thinking_turns", value: 1 },
-                    },
-                    {
-                        type: "clear_tool_uses_20250919",
-                        trigger: { type: "input_tokens", value: 60_000 },
-                        keep: { type: "tool_uses", value: 2 },
-                    },
-                ];
-                // Compaction for sub-agents on models that support it (lower threshold than main agent).
-                // Trigger at 80K — above the 60K clear_tool_uses threshold so clearing fires first.
-                const supportsCompaction = model.includes("opus-4-6") || model.includes("sonnet-4-6");
-                if (supportsCompaction) {
-                    edits.push({
-                        type: "compact_20260112",
-                        trigger: { type: "input_tokens", value: 80_000 },
-                        pause_after_compaction: true,
-                        instructions: "Summarize preserving: task goal, files found/modified, key findings, next steps.",
-                    });
-                    betas.push("compact-2026-01-12");
-                }
-            }
-            break;
-        case "teammate": {
-            const ctxMgmt = getContextManagement(model);
-            betas = [...ctxMgmt.betas];
-            edits = ctxMgmt.config.edits;
-            break;
-        }
-    }
-    // Thinking config
-    const thinkingCfg = getThinkingConfig(model, thinkingEnabled);
-    if (thinkingCfg.beta)
-        betas.push(thinkingCfg.beta);
-    // Max tokens — caller is responsible for setting appropriate limits per context
-    const maxTokens = getMaxOutputTokens(model, maxOutputTokens);
-    // Build thinking param (ensure budget < maxTokens)
-    // Only include for Anthropic/Bedrock — Gemini/OpenAI handle thinking internally
-    let thinking;
-    if (isAnthropicModel && thinkingCfg.thinking.type !== "disabled") {
-        thinking = thinkingCfg.thinking.budget_tokens
-            ? { ...thinkingCfg.thinking, budget_tokens: Math.min(thinkingCfg.thinking.budget_tokens, maxTokens - 1) }
-            : thinkingCfg.thinking;
-    }
-    // Strip clear_thinking edits when thinking is disabled — API requires thinking
-    // to be enabled or adaptive for clear_thinking_20251015 to be accepted
-    if (!thinking) {
-        edits = edits.filter(e => e.type !== "clear_thinking_20251015");
-    }
-    return {
-        betas,
-        contextManagement: { edits },
-        thinking,
-        maxTokens,
-    };
+      }
+      break;
+    case "teammate":
+      {
+        const ctxMgmt = getContextManagement(model);
+        betas = [...ctxMgmt.betas];
+        edits = ctxMgmt.config.edits;
+        break;
+      }
+  }
+
+  // Thinking config
+  const thinkingCfg = getThinkingConfig(model, thinkingEnabled);
+  if (thinkingCfg.beta) betas.push(thinkingCfg.beta);
+
+  // Max tokens — caller is responsible for setting appropriate limits per context
+  const maxTokens = getMaxOutputTokens(model, maxOutputTokens);
+
+  // Build thinking param (ensure budget < maxTokens)
+  // Only include for Anthropic/Bedrock — Gemini/OpenAI handle thinking internally
+  let thinking;
+  if (isAnthropicModel && thinkingCfg.thinking.type !== "disabled") {
+    thinking = thinkingCfg.thinking.budget_tokens ? {
+      ...thinkingCfg.thinking,
+      budget_tokens: Math.min(thinkingCfg.thinking.budget_tokens, maxTokens - 1)
+    } : thinkingCfg.thinking;
+  }
+
+  // Strip clear_thinking edits when thinking is disabled — API requires thinking
+  // to be enabled or adaptive for clear_thinking_20251015 to be accepted
+  if (!thinking) {
+    edits = edits.filter(e => e.type !== "clear_thinking_20251015");
+  }
+  return {
+    betas,
+    contextManagement: {
+      edits
+    },
+    thinking,
+    maxTokens
+  };
 }
+
+// ============================================================================
+// SERVER PROXY CALLER
+// ============================================================================
+
 /**
  * Call server proxy endpoint and return raw SSE stream.
  * Retries with exponential backoff on 429/500/529.
  */
 export async function callServerProxy(config) {
-    const { proxyUrl, signal, timeoutMs } = config;
-    let tokenRefreshed = false;
-    for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
-        try {
-            const body = {
-                mode: "proxy",
-                messages: config.messages,
-                system: config.system,
-                tools: config.tools,
-                model: config.model,
-                max_tokens: config.apiConfig.maxTokens,
-                stream: true,
-                betas: config.apiConfig.betas,
-                ...(config.apiConfig.contextManagement?.edits?.length
-                    ? { context_management: config.apiConfig.contextManagement }
-                    : {}),
-            };
-            if (config.apiConfig.thinking) {
-                body.thinking = config.apiConfig.thinking;
-            }
-            if (config.storeId) {
-                body.store_id = config.storeId;
-            }
-            // Prune old base64 images from conversation history before serializing.
-            // Keep images only in the last 2 messages to avoid multi-MB request bodies
-            // that can cause "Invalid JSON" errors from body truncation on the proxy.
-            if (Array.isArray(body.messages) && body.messages.length > 4) {
-                const msgs = body.messages;
-                const keepFrom = Math.max(0, msgs.length - 2);
-                for (let i = 0; i < keepFrom; i++) {
-                    const msg = msgs[i];
-                    if (!Array.isArray(msg?.content))
-                        continue;
-                    for (const block of msg.content) {
-                        if (block.type === "image" && block.source?.type === "base64") {
-                            block.type = "text";
-                            block.text = "[image]";
-                            delete block.source;
-                        }
-                        if (block.type === "tool_result" && Array.isArray(block.content)) {
-                            for (const sub of block.content) {
-                                if (sub.type === "image" && sub.source?.type === "base64") {
-                                    sub.type = "text";
-                                    sub.text = "[image]";
-                                    delete sub.source;
-                                }
-                            }
-                        }
-                    }
-                }
-            }
-            const serialized = JSON.stringify(body);
-            const fetchOpts = {
-                method: "POST",
-                headers: {
-                    "Content-Type": "application/json",
-                    "Authorization": `Bearer ${config.token}`,
-                    "Content-Length": String(Buffer.byteLength(serialized)),
-                },
-                body: serialized,
-                signal,
-            };
-            // Apply timeout only when the caller did NOT already supply a signal
-            let controller;
-            let timeout;
-            if (timeoutMs && !signal) {
-                controller = new AbortController();
-                timeout = setTimeout(() => controller.abort(), timeoutMs);
-                fetchOpts.signal = controller.signal;
+  const {
+    proxyUrl,
+    signal,
+    timeoutMs
+  } = config;
+  let tokenRefreshed = false;
+  for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
+    try {
+      const body = {
+        mode: "proxy",
+        messages: config.messages,
+        system: config.system,
+        tools: config.tools,
+        model: config.model,
+        max_tokens: config.apiConfig.maxTokens,
+        stream: true,
+        betas: config.apiConfig.betas,
+        ...(config.apiConfig.contextManagement?.edits?.length ? {
+          context_management: config.apiConfig.contextManagement
+        } : {})
+      };
+      if (config.apiConfig.thinking) {
+        body.thinking = config.apiConfig.thinking;
+        // Anthropic requires temperature=1.0 when thinking is enabled — enforce client-side
+        if (config.apiConfig.thinking.type !== "disabled") {
+          body.temperature = 1.0;
+        }
+      }
+      if (config.storeId) {
+        body.store_id = config.storeId;
+      }
+
+      // Prune old base64 images from conversation history before serializing.
+      // Keep images only in the last 2 messages to avoid multi-MB request bodies
+      // that can cause "Invalid JSON" errors from body truncation on the proxy.
+      if (Array.isArray(body.messages) && body.messages.length > 4) {
+        const msgs = body.messages;
+        const keepFrom = Math.max(0, msgs.length - 2);
+        for (let i = 0; i < keepFrom; i++) {
+          const msg = msgs[i];
+          if (!Array.isArray(msg?.content)) continue;
+          for (const block of msg.content) {
+            if (block.type === "image" && block.source?.type === "base64") {
+              block.type = "text";
+              block.text = "[image]";
+              delete block.source;
             }
-            try {
-                const response = await fetch(proxyUrl, fetchOpts);
-                if (timeout)
-                    clearTimeout(timeout);
-                if (response.ok && response.body) {
-                    return response.body;
-                }
-                const errorBody = await response.text();
-                // Auto-refresh on 401 (token invalidated server-side before local expiry)
-                if (response.status === 401 && config.onTokenRefresh && !tokenRefreshed) {
-                    const newToken = await config.onTokenRefresh();
-                    if (newToken) {
-                        tokenRefreshed = true;
-                        config.token = newToken;
-                        continue; // retry with new token
-                    }
+            if (block.type === "tool_result" && Array.isArray(block.content)) {
+              for (const sub of block.content) {
+                if (sub.type === "image" && sub.source?.type === "base64") {
+                  sub.type = "text";
+                  sub.text = "[image]";
+                  delete sub.source;
                 }
-                throw Object.assign(new Error(`Proxy error (${response.status}): ${errorBody}`), { status: response.status });
-            }
-            catch (err) {
-                if (timeout)
-                    clearTimeout(timeout);
-                throw err;
+              }
             }
+          }
         }
-        catch (err) {
-            if (signal?.aborted)
-                throw err;
-            if (attempt < MAX_RETRIES && isRetryableError(err)) {
-                const errMsg = err instanceof Error ? err.message : String(err);
-                config.onRetry?.(attempt + 1, MAX_RETRIES, errMsg);
-                const delay = RETRY_BASE_DELAY_MS * Math.pow(2, attempt);
-                // Abort-aware backoff — ESC can interrupt the wait immediately
-                await new Promise((resolve, reject) => {
-                    if (signal?.aborted) {
-                        reject(signal.reason ?? new Error("Aborted"));
-                        return;
-                    }
-                    const timer = setTimeout(resolve, delay);
-                    const onAbort = () => { clearTimeout(timer); reject(signal.reason ?? new Error("Aborted")); };
-                    signal?.addEventListener("abort", onAbort, { once: true });
-                });
-                // Fallback model on last retry
-                if (attempt === MAX_RETRIES - 1 && config.fallbackModel) {
-                    const fromModel = config.model;
-                    config.model = config.fallbackModel;
-                    config.onFallback?.(fromModel, config.model);
-                }
-                continue;
-            }
-            throw err;
+      }
+      const serialized = JSON.stringify(body);
+      const fetchOpts = {
+        method: "POST",
+        headers: {
+          "Content-Type": "application/json",
+          "Authorization": `Bearer ${config.token}`,
+          "Content-Length": String(Buffer.byteLength(serialized))
+        },
+        body: serialized,
+        signal
+      };
+
+      // Apply timeout only when the caller did NOT already supply a signal
+      let controller;
+      let timeout;
+      if (timeoutMs && !signal) {
+        controller = new AbortController();
+        timeout = setTimeout(() => controller.abort(), timeoutMs);
+        fetchOpts.signal = controller.signal;
+      }
+      try {
+        const response = await fetch(proxyUrl, fetchOpts);
+        if (timeout) clearTimeout(timeout);
+        if (response.ok && response.body) {
+          return response.body;
+        }
+        const errorBody = await response.text();
+
+        // Auto-refresh on 401 (token invalidated server-side before local expiry)
+        if (response.status === 401 && config.onTokenRefresh && !tokenRefreshed) {
+          const newToken = await config.onTokenRefresh();
+          if (newToken) {
+            tokenRefreshed = true;
+            config.token = newToken;
+            continue; // retry with new token
+          }
+        }
+        throw Object.assign(new Error(`Proxy error (${response.status}): ${errorBody}`), {
+          status: response.status
+        });
+      } catch (err) {
+        if (timeout) clearTimeout(timeout);
+        throw err;
+      }
+    } catch (err) {
+      if (signal?.aborted) throw err;
+      if (attempt < MAX_RETRIES && isRetryableError(err)) {
+        const errMsg = err instanceof Error ? err.message : String(err);
+        config.onRetry?.(attempt + 1, MAX_RETRIES, errMsg);
+        const delay = RETRY_BASE_DELAY_MS * Math.pow(2, attempt);
+
+        // Abort-aware backoff — ESC can interrupt the wait immediately
+        await new Promise((resolve, reject) => {
+          if (signal?.aborted) {
+            reject(signal.reason ?? new Error("Aborted"));
+            return;
+          }
+          const timer = setTimeout(resolve, delay);
+          const onAbort = () => {
+            clearTimeout(timer);
+            reject(signal.reason ?? new Error("Aborted"));
+          };
+          signal?.addEventListener("abort", onAbort, {
+            once: true
+          });
+        });
+
+        // Fallback model on last retry
+        if (attempt === MAX_RETRIES - 1 && config.fallbackModel) {
+          const fromModel = config.model;
+          config.model = config.fallbackModel;
+          config.onFallback?.(fromModel, config.model);
         }
+        continue;
+      }
+      throw err;
     }
-    throw new Error("Failed to get response after retries");
+  }
+  throw new Error("Failed to get response after retries");
 }
+
 // ============================================================================
 // HELPERS — system prompt + caching setup
 // ============================================================================
+
 /**
  * Build system blocks with prompt caching.
  * Cached system prompt + optional dynamic cost context (after cache breakpoint).
  */
 export function buildSystemBlocks(systemPrompt, costContext, enableCaching = true) {
-    const system = [];
-    if (enableCaching) {
-        system.push({
-            type: "text",
-            text: systemPrompt,
-            cache_control: { type: "ephemeral" },
-        });
-    }
-    else {
-        system.push({ type: "text", text: systemPrompt });
-    }
-    if (costContext) {
-        system.push({ type: "text", text: costContext });
-    }
-    return system;
+  const system = [];
+  if (enableCaching) {
+    system.push({
+      type: "text",
+      text: systemPrompt,
+      cache_control: {
+        type: "ephemeral"
+      }
+    });
+  } else {
+    system.push({
+      type: "text",
+      text: systemPrompt
+    });
+  }
+  if (costContext) {
+    system.push({
+      type: "text",
+      text: costContext
+    });
+  }
+  return system;
 }
+
 /**
  * Prepare messages and tools with prompt caching.
  * Skips cache_control injection for Gemini (uses implicit caching).
  */
 export function prepareWithCaching(tools, messages, model) {
-    if (model) {
-        const provider = getProvider(model);
-        if (provider === "gemini" || provider === "openai") {
-            return { tools, messages };
-        }
+  if (model) {
+    const provider = getProvider(model);
+    if (provider === "gemini" || provider === "openai") {
+      return {
+        tools,
+        messages
+      };
     }
-    return addPromptCaching(tools, messages);
+  }
+  return addPromptCaching(tools, messages);
 }
+
 // ============================================================================
 // GEMINI CONTEXT TRIMMING
 // ============================================================================
+
 /**
  * Client-side context trimming for Gemini (no server-side context management).
  * When estimated input tokens exceed threshold, replaces old tool result contents
  * with "[trimmed]" to stay within Gemini's 1M context window.
  */
 export function trimGeminiContext(messages, estimatedTokens, threshold = 950_000, keepRecent = 5) {
-    if (estimatedTokens < threshold)
-        return messages;
-    // Find all tool_result blocks and trim oldest ones
-    let toolResultCount = 0;
-    for (const msg of messages) {
-        if (Array.isArray(msg.content)) {
-            for (const block of msg.content) {
-                if (block.type === "tool_result")
-                    toolResultCount++;
-            }
-        }
+  if (estimatedTokens < threshold) return messages;
+  let toolResultCount = 0;
+  for (const msg of messages) {
+    if (Array.isArray(msg.content)) {
+      for (const block of msg.content) {
+        if (block.type === "tool_result") toolResultCount++;
+      }
     }
-    if (toolResultCount <= keepRecent)
-        return messages;
-    const trimCount = toolResultCount - keepRecent;
-    let trimmed = 0;
-    return messages.map((msg) => {
-        if (!Array.isArray(msg.content))
-            return msg;
-        const content = msg.content.map((block) => {
-            if (block.type === "tool_result" && trimmed < trimCount) {
-                trimmed++;
-                if (typeof block.content === "string") {
-                    return { ...block, content: "[trimmed]" };
-                }
-                else if (Array.isArray(block.content)) {
-                    return { ...block, content: [{ type: "text", text: "[trimmed]" }] };
-                }
-            }
-            return block;
-        });
-        return { ...msg, content };
+  }
+  if (toolResultCount <= keepRecent) return messages;
+  const trimCount = toolResultCount - keepRecent;
+  let trimmed = 0;
+  return messages.map(msg => {
+    if (!Array.isArray(msg.content)) return msg;
+    const content = msg.content.map(block => {
+      if (block.type === "tool_result" && trimmed < trimCount) {
+        trimmed++;
+        if (typeof block.content === "string") {
+          return {
+            ...block,
+            content: "[trimmed]"
+          };
+        } else if (Array.isArray(block.content)) {
+          return {
+            ...block,
+            content: [{
+              type: "text",
+              text: "[trimmed]"
+            }]
+          };
+        }
+      }
+      return block;
     });
+    return {
+      ...msg,
+      content
+    };
+  });
 }
+
 // ============================================================================
 // OPENAI CONTEXT TRIMMING
 // ============================================================================
+
 /**
  * Client-side context trimming for OpenAI (no server-side context management).
  * When estimated input tokens exceed threshold, replaces old tool result contents
@@ -318,113 +388,135 @@ export function trimGeminiContext(messages, estimatedTokens, threshold = 950_000
 // ============================================================================
 // AUDIO TRANSCRIPTION CLIENT
 // ============================================================================
+
 /**
  * Call server transcription endpoint (OpenAI Whisper via our Fly.io proxy).
  * Used by tool-dispatch when __AUDIO__ marker is detected in tool results.
  */
 export async function callTranscribe(opts) {
-    const { proxyUrl, token, storeId, audioBase64, mediaType } = opts;
-    // Use same server URL as proxy, just different mode
-    const response = await fetch(proxyUrl, {
-        method: "POST",
-        headers: {
-            "Content-Type": "application/json",
-            Authorization: `Bearer ${token}`,
-        },
-        body: JSON.stringify({
-            mode: "transcribe",
-            audio_base64: audioBase64,
-            media_type: mediaType,
-            store_id: storeId,
-        }),
-    });
-    if (!response.ok) {
-        const errText = await response.text();
-        throw new Error(`Transcription failed (${response.status}): ${errText}`);
-    }
-    const data = (await response.json());
-    if (!data.success || !data.transcript) {
-        throw new Error(data.error || "Transcription returned no transcript");
-    }
-    return data.transcript;
+  const {
+    proxyUrl,
+    token,
+    storeId,
+    audioBase64,
+    mediaType
+  } = opts;
+
+  // Use same server URL as proxy, just different mode
+  const response = await fetch(proxyUrl, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      Authorization: `Bearer ${token}`
+    },
+    body: JSON.stringify({
+      mode: "transcribe",
+      audio_base64: audioBase64,
+      media_type: mediaType,
+      store_id: storeId
+    })
+  });
+  if (!response.ok) {
+    const errText = await response.text();
+    throw new Error(`Transcription failed (${response.status}): ${errText}`);
+  }
+  const data = await response.json();
+  if (!data.success || !data.transcript) {
+    throw new Error(data.error || "Transcription returned no transcript");
+  }
+  return data.transcript;
 }
+
 // ============================================================================
 // OPENAI CONTEXT TRIMMING
 // ============================================================================
+
 export function trimOpenAIContext(messages, estimatedTokens, threshold = 190_000, keepRecent = 5) {
-    if (estimatedTokens < threshold)
-        return messages;
-    // Find all tool_result blocks and trim oldest ones
-    let toolResultCount = 0;
-    for (const msg of messages) {
-        if (Array.isArray(msg.content)) {
-            for (const block of msg.content) {
-                if (block.type === "tool_result")
-                    toolResultCount++;
-            }
-        }
+  if (estimatedTokens < threshold) return messages;
+  let toolResultCount = 0;
+  for (const msg of messages) {
+    if (Array.isArray(msg.content)) {
+      for (const block of msg.content) {
+        if (block.type === "tool_result") toolResultCount++;
+      }
     }
-    if (toolResultCount <= keepRecent)
-        return messages;
-    const trimCount = toolResultCount - keepRecent;
-    let trimmed = 0;
-    return messages.map((msg) => {
-        if (!Array.isArray(msg.content))
-            return msg;
-        const content = msg.content.map((block) => {
-            if (block.type === "tool_result" && trimmed < trimCount) {
-                trimmed++;
-                if (typeof block.content === "string") {
-                    return { ...block, content: "[trimmed]" };
-                }
-                else if (Array.isArray(block.content)) {
-                    return { ...block, content: [{ type: "text", text: "[trimmed]" }] };
-                }
-            }
-            return block;
-        });
-        return { ...msg, content };
+  }
+  if (toolResultCount <= keepRecent) return messages;
+  const trimCount = toolResultCount - keepRecent;
+  let trimmed = 0;
+  return messages.map(msg => {
+    if (!Array.isArray(msg.content)) return msg;
+    const content = msg.content.map(block => {
+      if (block.type === "tool_result" && trimmed < trimCount) {
+        trimmed++;
+        if (typeof block.content === "string") {
+          return {
+            ...block,
+            content: "[trimmed]"
+          };
+        } else if (Array.isArray(block.content)) {
+          return {
+            ...block,
+            content: [{
+              type: "text",
+              text: "[trimmed]"
+            }]
+          };
+        }
+      }
+      return block;
     });
+    return {
+      ...msg,
+      content
+    };
+  });
 }
+
 // ============================================================================
 // PROVIDER COMPACTION CLIENT
 // ============================================================================
+
 const COMPACTION_TIMEOUT_MS = 30_000;
+
 /**
  * Request conversation compaction from the server (Haiku summarization).
  * Used by non-Anthropic providers (OpenAI, Gemini) that lack native compaction.
  * Returns summary string or null on failure (caller keeps unsummarized history).
  */
 export async function requestProviderCompaction(opts) {
-    const { proxyUrl, token, messages, systemPrompt } = opts;
+  const {
+    proxyUrl,
+    token,
+    messages,
+    systemPrompt
+  } = opts;
+  try {
+    const controller = new AbortController();
+    const timeout = setTimeout(() => controller.abort(), COMPACTION_TIMEOUT_MS);
     try {
-        const controller = new AbortController();
-        const timeout = setTimeout(() => controller.abort(), COMPACTION_TIMEOUT_MS);
-        try {
-            const response = await fetch(proxyUrl, {
-                method: "POST",
-                headers: {
-                    "Content-Type": "application/json",
-                    Authorization: `Bearer ${token}`,
-                },
-                body: JSON.stringify({
-                    mode: "compact",
-                    messages,
-                    system_prompt: systemPrompt,
-                }),
-                signal: controller.signal,
-            });
-            clearTimeout(timeout);
-            if (!response.ok)
-                return null;
-            const data = (await response.json());
-            return data.success && data.compaction_content ? data.compaction_content : null;
-        }
-        finally {
-            clearTimeout(timeout);
-        }
-    }
-    catch {
-        return null; // Graceful failure — caller keeps unsummarized history
+      const response = await fetch(proxyUrl, {
+        method: "POST",
+        headers: {
+          "Content-Type": "application/json",
+          Authorization: `Bearer ${token}`
+        },
+        body: JSON.stringify({
+          mode: "compact",
+          messages,
+          system_prompt: systemPrompt
+        }),
+        signal: controller.signal
+      });
+      clearTimeout(timeout);
+      if (!response.ok) return null;
+      const data = await response.json();
+      return data.success && data.compaction_content ? data.compaction_content : null;
+    } finally {
+      clearTimeout(timeout);
     }
+  } catch {
+    return null; // Graceful failure — caller keeps unsummarized history
+  }
 }
+//# sourceMappingURL=api-client.js.map