wasper-cli 0.2.0 → 0.3.1

package/dist/index.js

@@ -3158,6 +3158,14 @@ var SCHEMA = `
     name TEXT NOT NULL DEFAULT '',
     created_at INTEGER NOT NULL DEFAULT (unixepoch())
   );
+
+  CREATE TABLE IF NOT EXISTS chat_memory (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    role TEXT NOT NULL,
+    content TEXT NOT NULL,
+    created_at INTEGER NOT NULL DEFAULT (unixepoch())
+  );
+  CREATE INDEX IF NOT EXISTS idx_memory_created ON chat_memory(created_at DESC);
 `;
 
 // src/db/index.ts
@@ -3337,6 +3345,19 @@ var init_db = __esm(() => {
     getSetting: (key) => (db.query("SELECT value FROM settings WHERE key = ?").get(key) ?? null)?.value ?? null,
     setSetting: (key, value) => {
       db.run("INSERT INTO settings(key,value) VALUES(?,?) ON CONFLICT(key) DO UPDATE SET value=excluded.value", [key, value]);
+    },
+    saveMemory: (role, content) => {
+      db.query("INSERT INTO chat_memory (role, content) VALUES (?, ?)").run(role, content);
+    },
+    getMemory: (limit = 20) => {
+      const rows = db.query("SELECT role, content FROM chat_memory ORDER BY created_at DESC LIMIT ?").all(limit);
+      return rows.reverse();
+    },
+    clearMemory: () => {
+      db.query("DELETE FROM chat_memory").run();
+    },
+    trimMemory: (keepLast = 40) => {
+      db.query("DELETE FROM chat_memory WHERE id NOT IN (SELECT id FROM chat_memory ORDER BY created_at DESC LIMIT ?)").run(keepLast);
     }
   };
 });
@@ -3586,7 +3607,7 @@ var package_default;
 var init_package = __esm(() => {
   package_default = {
     name: "wasper-cli",
-    version: "0.2.0",
+    version: "0.3.1",
     description: "Host an MCP server + API proxy from any OpenAPI spec. Like Drizzle Studio, but for APIs.",
     type: "module",
     homepage: "https://wasper.site",
@@ -4578,6 +4599,517 @@ var init_engine2 = __esm(() => {
   init_state();
 });
 
+// src/agent/harness.ts
+function mergeSignals(a, b) {
+  if (!a && !b)
+    return new AbortController().signal;
+  if (!a)
+    return b;
+  if (!b)
+    return a;
+  const ctrl = new AbortController;
+  const abort = () => ctrl.abort();
+  a.addEventListener("abort", abort, { once: true });
+  b.addEventListener("abort", abort, { once: true });
+  return ctrl.signal;
+}
+async function fetchWithRetry(url, opts, emit, signal, maxRetries = 4) {
+  for (let attempt = 0;attempt <= maxRetries; attempt++) {
+    const stepSignal = signal;
+    let res;
+    try {
+      res = await fetch(url, { ...opts, signal: stepSignal });
+    } catch (e) {
+      const msg = e instanceof Error ? e.message : String(e);
+      if (signal?.aborted)
+        throw e;
+      if (attempt === maxRetries)
+        throw e;
+      const isNetwork = msg.includes("ECONNREFUSED") || msg.includes("ENOTFOUND") || msg.includes("network") || msg.includes("fetch");
+      if (!isNetwork)
+        throw e;
+      const delay2 = Math.min(1000 * Math.pow(2, attempt) + Math.random() * 300, 15000);
+      emit({ type: "info", message: `Network error, retrying in ${Math.round(delay2 / 1000)}s\u2026` });
+      await new Promise((r) => setTimeout(r, delay2));
+      continue;
+    }
+    if (!RETRYABLE_STATUS.has(res.status) || attempt === maxRetries)
+      return res;
+    const retryAfter = parseInt(res.headers.get("retry-after") ?? "0", 10);
+    const delay = retryAfter > 0 ? retryAfter * 1000 : Math.min(1000 * Math.pow(2, attempt) + Math.random() * 300, 30000);
+    const label = res.status === 429 ? "Rate limited" : `Server error ${res.status}`;
+    emit({ type: "info", message: `${label} \u2014 retrying in ${Math.round(delay / 1000)}s\u2026 (attempt ${attempt + 1}/${maxRetries})` });
+    await new Promise((r) => setTimeout(r, delay));
+  }
+  return fetch(url, opts);
+}
+function trimContext(messages) {
+  if (JSON.stringify(messages).length <= MAX_CONTEXT_CHARS)
+    return { messages, trimmed: false };
+  const result = [...messages];
+  while (JSON.stringify(result).length > MAX_CONTEXT_CHARS && result.length > 2) {
+    let removed = false;
+    const toolIdx = result.findIndex((m) => m.role === "tool");
+    if (toolIdx !== -1) {
+      result.splice(toolIdx, 1);
+      if (toolIdx > 0) {
+        const prev = result[toolIdx - 1];
+        if (prev?.role === "assistant") {
+          const tc = prev.tool_calls;
+          if (Array.isArray(tc) && tc.length)
+            result.splice(toolIdx - 1, 1);
+        }
+      }
+      removed = true;
+    }
+    if (!removed) {
+      const anthropicIdx = result.findIndex((m) => {
+        if (m.role !== "user")
+          return false;
+        const c = m.content;
+        return Array.isArray(c) && c.some((b) => b.type === "tool_result");
+      });
+      if (anthropicIdx !== -1) {
+        result.splice(anthropicIdx, 1);
+        if (anthropicIdx > 0 && result[anthropicIdx - 1]?.role === "assistant") {
+          result.splice(anthropicIdx - 1, 1);
+        }
+        removed = true;
+      }
+    }
+    if (!removed)
+      break;
+  }
+  return { messages: result, trimmed: true };
+}
+async function* readSSE(body) {
+  const reader = body.getReader();
+  const decoder = new TextDecoder;
+  let buf = "";
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done)
+        break;
+      buf += decoder.decode(value, { stream: true });
+      const parts = buf.split(`
+
+`);
+      buf = parts.pop() ?? "";
+      for (const part of parts) {
+        let data = "";
+        for (const line of part.split(`
+`)) {
+          if (line.startsWith("data: ")) {
+            data = line.slice(6);
+            break;
+          }
+        }
+        if (!data || data === "[DONE]")
+          continue;
+        try {
+          yield JSON.parse(data);
+        } catch {}
+      }
+    }
+  } finally {
+    reader.releaseLock();
+  }
+}
+function buildAnthropicTools(schemas) {
+  return schemas.map((s) => ({
+    name: s.name,
+    description: s.description,
+    input_schema: { type: "object", properties: s.params, required: s.required }
+  }));
+}
+async function streamAnthropic(cfg, system, messages, tools, emit, signal) {
+  const base = (cfg.baseUrl || "https://api.anthropic.com").replace(/\/$/, "");
+  const systemContent = cfg.enablePromptCache ? [{ type: "text", text: system, cache_control: { type: "ephemeral" } }] : system;
+  const stepSignal = mergeSignals(signal, AbortSignal.timeout(cfg.stepTimeoutMs));
+  const res = await fetchWithRetry(`${base}/v1/messages`, {
+    method: "POST",
+    headers: {
+      "Content-Type": "application/json",
+      "x-api-key": cfg.apiKey,
+      "anthropic-version": "2023-06-01",
+      ...cfg.enablePromptCache ? { "anthropic-beta": "prompt-caching-1-0" } : {},
+      ...cfg.extraHeaders
+    },
+    body: JSON.stringify({
+      model: cfg.model,
+      max_tokens: cfg.maxTokens,
+      temperature: cfg.temperature,
+      ...cfg.topK > 0 ? { top_k: cfg.topK } : {},
+      system: systemContent,
+      messages,
+      tools,
+      stream: true
+    })
+  }, emit, stepSignal);
+  if (!res.ok) {
+    const body = await res.text();
+    const retryable = RETRYABLE_STATUS.has(res.status);
+    throw Object.assign(new Error(`Anthropic ${res.status}: ${body}`), { retryable });
+  }
+  const result = { text: "", thinking: "", stopReason: "", toolUses: [], usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } };
+  const blocks = [];
+  const inputAccum = {};
+  for await (const ev of readSSE(res.body)) {
+    if (signal.aborted)
+      break;
+    const evType = ev.type;
+    if (evType === "message_start") {
+      const usage = ev.message?.usage;
+      if (usage) {
+        result.usage.input = usage.input_tokens ?? 0;
+        result.usage.cacheRead = usage.cache_read_input_tokens ?? 0;
+        result.usage.cacheWrite = usage.cache_creation_input_tokens ?? 0;
+      }
+    } else if (evType === "content_block_start") {
+      const idx = ev.index;
+      const cb = ev.content_block;
+      blocks[idx] = { type: cb.type, id: cb.id, name: cb.name };
+      if (cb.type === "tool_use")
+        inputAccum[idx] = "";
+    } else if (evType === "content_block_delta") {
+      const idx = ev.index;
+      const delta = ev.delta;
+      if (delta.type === "text_delta" && delta.text) {
+        result.text += delta.text;
+        if (!blocks[idx])
+          blocks[idx] = { type: "text" };
+        blocks[idx].text = (blocks[idx].text ?? "") + delta.text;
+        emit({ type: "text_delta", text: delta.text });
+      } else if (delta.type === "thinking_delta" && delta.thinking) {
+        result.thinking += delta.thinking;
+        emit({ type: "thinking", text: delta.thinking });
+      } else if (delta.type === "input_json_delta" && delta.partial_json) {
+        inputAccum[idx] = (inputAccum[idx] ?? "") + delta.partial_json;
+      }
+    } else if (evType === "content_block_stop") {
+      const idx = ev.index;
+      if (blocks[idx]?.type === "tool_use") {
+        try {
+          blocks[idx].input = JSON.parse(inputAccum[idx] ?? "{}");
+        } catch {
+          blocks[idx].input = {};
+        }
+      }
+    } else if (evType === "message_delta") {
+      const delta = ev.delta;
+      const usage = ev.usage;
+      if (delta.stop_reason)
+        result.stopReason = delta.stop_reason;
+      if (usage?.output_tokens)
+        result.usage.output = usage.output_tokens;
+    }
+  }
+  for (const b of blocks) {
+    if (b.type === "tool_use" && b.id && b.name) {
+      result.toolUses.push({ id: b.id, name: b.name, input: b.input ?? {} });
+    }
+  }
+  result._anthropicBlocks = blocks;
+  return result;
+}
+function buildOpenAITools(schemas) {
+  return schemas.map((s) => ({
+    type: "function",
+    function: { name: s.name, description: s.description, parameters: { type: "object", properties: s.params, required: s.required } }
+  }));
+}
+async function streamOpenAI(cfg, system, messages, tools, emit, signal) {
+  const providerBases = {
+    openai: "https://api.openai.com",
+    mistral: "https://api.mistral.ai",
+    groq: "https://api.groq.com/openai",
+    "github-copilot": "https://api.githubcopilot.com"
+  };
+  const base = (cfg.baseUrl || providerBases[cfg.provider] || "https://api.openai.com").replace(/\/$/, "");
+  const providerHeaders = cfg.provider === "github-copilot" ? { "Copilot-Integration-Id": "vscode-chat", "Editor-Version": "vscode/1.85.0" } : {};
+  const authHeaders = cfg.apiKey ? { Authorization: `Bearer ${cfg.apiKey}` } : {};
+  const stepSignal = mergeSignals(signal, AbortSignal.timeout(cfg.stepTimeoutMs));
+  const res = await fetchWithRetry(`${base}/v1/chat/completions`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json", ...authHeaders, ...providerHeaders, ...cfg.extraHeaders },
+    body: JSON.stringify({
+      model: cfg.model,
+      max_tokens: cfg.maxTokens,
+      temperature: cfg.temperature,
+      messages: [{ role: "system", content: system }, ...messages],
+      tools,
+      tool_choice: "auto",
+      stream: true,
+      stream_options: { include_usage: true }
+    })
+  }, emit, stepSignal);
+  if (!res.ok) {
+    const body = await res.text();
+    const retryable = RETRYABLE_STATUS.has(res.status);
+    throw Object.assign(new Error(body), { retryable });
+  }
+  const result = { text: "", thinking: "", stopReason: "", toolUses: [], usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } };
+  const tcAccum = {};
+  for await (const ev of readSSE(res.body)) {
+    if (signal.aborted)
+      break;
+    if (ev.object === "error") {
+      const retryable = ev.code === "1300" || ev.raw_status_code === 429 || ev.raw_status_code === 503;
+      throw Object.assign(new Error(JSON.stringify(ev)), { retryable });
+    }
+    if (ev.usage) {
+      const u = ev.usage;
+      result.usage.input = u.prompt_tokens ?? 0;
+      result.usage.output = u.completion_tokens ?? 0;
+    }
+    const choices = ev.choices;
+    const choice = choices?.[0];
+    if (!choice)
+      continue;
+    const fr = choice.finish_reason;
+    if (fr)
+      result.stopReason = fr;
+    const delta = choice.delta;
+    if (!delta)
+      continue;
+    if (typeof delta.content === "string" && delta.content) {
+      result.text += delta.content;
+      emit({ type: "text_delta", text: delta.content });
+    }
+    const tcDeltas = delta.tool_calls;
+    if (tcDeltas) {
+      for (const tc of tcDeltas) {
+        if (!tcAccum[tc.index])
+          tcAccum[tc.index] = { id: "", name: "", args: "" };
+        const e = tcAccum[tc.index];
+        if (tc.id)
+          e.id += tc.id;
+        if (tc.function?.name)
+          e.name += tc.function.name;
+        if (tc.function?.arguments)
+          e.args += tc.function.arguments;
+      }
+    }
+  }
+  for (const tc of Object.values(tcAccum)) {
+    let input = {};
+    try {
+      input = JSON.parse(tc.args);
+    } catch {}
+    result.toolUses.push({ id: tc.id, name: tc.name, input });
+  }
+  result._openaiTcAccum = tcAccum;
+  return result;
+}
+async function callOllama(cfg, system, messages, emit, signal) {
+  const base = (cfg.baseUrl || "http://localhost:11434").replace(/\/$/, "");
+  const stepSignal = mergeSignals(signal, AbortSignal.timeout(cfg.stepTimeoutMs));
+  const res = await fetchWithRetry(`${base}/api/chat`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ model: cfg.model, messages: [{ role: "system", content: system }, ...messages], stream: false })
+  }, emit, stepSignal);
+  if (!res.ok)
+    throw new Error(`Ollama ${res.status}: ${await res.text()}`);
+  const d = await res.json();
+  const text = d.message?.content ?? "";
+  emit({ type: "text_delta", text });
+  return { text, thinking: "", stopReason: "end_turn", toolUses: [], usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } };
+}
+async function callGemini(cfg, system, messages, emit, signal) {
+  const base = (cfg.baseUrl || "https://generativelanguage.googleapis.com").replace(/\/$/, "");
+  const stepSignal = mergeSignals(signal, AbortSignal.timeout(cfg.stepTimeoutMs));
+  const res = await fetchWithRetry(`${base}/v1beta/models/${cfg.model}:generateContent?key=${cfg.apiKey}`, {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({
+      systemInstruction: { parts: [{ text: system }] },
+      contents: messages.map((m) => ({
+        role: m.role === "assistant" ? "model" : "user",
+        parts: [{ text: m.content }]
+      })),
+      generationConfig: { maxOutputTokens: cfg.maxTokens }
+    })
+  }, emit, stepSignal);
+  if (!res.ok)
+    throw new Error(`Gemini ${res.status}: ${await res.text()}`);
+  const d = await res.json();
+  const text = d.candidates?.[0]?.content?.parts?.[0]?.text ?? "";
+  emit({ type: "text_delta", text });
+  return { text, thinking: "", stopReason: "end_turn", toolUses: [], usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 } };
+}
+async function runAgentLoop(config2, system, initialMessages, toolSchemas, executeTool, emit, signal = new AbortController().signal, toolCache = new Map) {
+  const cfg = { ...DEFAULTS, ...config2 };
+  const isAnthropic = cfg.provider === "anthropic";
+  const isOllama = cfg.provider === "ollama";
+  const isGemini = cfg.provider === "gemini";
+  const anthropicTools = buildAnthropicTools(toolSchemas);
+  const openaiTools = buildOpenAITools(toolSchemas);
+  const messages = [...initialMessages];
+  const allToolCalls = [];
+  const totalTokens = { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 };
+  let totalToolsUsed = 0;
+  let consecutiveErrors = 0;
+  const endpointErrors = {};
+  for (let iter = 0;iter < cfg.maxIterations; iter++) {
+    if (signal.aborted) {
+      return { content: "", toolCalls: allToolCalls, stopReason: "cancelled", tokens: totalTokens };
+    }
+    const { messages: trimmed, trimmed: didTrim } = trimContext(messages);
+    if (didTrim) {
+      emit({ type: "info", message: "Context trimmed to fit within limits." });
+      messages.splice(0, messages.length, ...trimmed);
+    }
+    let turn;
+    try {
+      if (isAnthropic) {
+        turn = await streamAnthropic(cfg, system, messages, anthropicTools, emit, signal);
+      } else if (isOllama) {
+        turn = await callOllama(cfg, system, messages, emit, signal);
+      } else if (isGemini) {
+        turn = await callGemini(cfg, system, messages, emit, signal);
+      } else {
+        turn = await streamOpenAI(cfg, system, messages, openaiTools, emit, signal);
+      }
+    } catch (e) {
+      if (signal.aborted) {
+        return { content: "", toolCalls: allToolCalls, stopReason: "cancelled", tokens: totalTokens };
+      }
+      const msg = e instanceof Error ? e.message : String(e);
+      const retryable = e.retryable ?? false;
+      emit({ type: "error", message: msg, retryable });
+      throw e;
+    }
+    totalTokens.input += turn.usage.input;
+    totalTokens.output += turn.usage.output;
+    totalTokens.cacheRead += turn.usage.cacheRead;
+    totalTokens.cacheWrite += turn.usage.cacheWrite;
+    if (turn.usage.input || turn.usage.output) {
+      emit({ type: "token_usage", ...turn.usage });
+    }
+    const wantsTools = isAnthropic ? turn.stopReason === "tool_use" : turn.stopReason === "tool_calls";
+    if (!wantsTools || turn.toolUses.length === 0) {
+      return { content: turn.text, toolCalls: allToolCalls, stopReason: "end_turn", tokens: totalTokens };
+    }
+    if (totalToolsUsed >= cfg.maxTotalTools) {
+      return {
+        content: `Agent stopped: reached ${cfg.maxTotalTools} tool calls. Break your request into smaller steps.`,
+        toolCalls: allToolCalls,
+        stopReason: "max_tools",
+        tokens: totalTokens
+      };
+    }
+    const dedupeKey = (name, input) => `${name}:${JSON.stringify(input)}`;
+    const turnResults = [];
+    const executeOne = async (use) => {
+      totalToolsUsed++;
+      const key = dedupeKey(use.name, use.input);
+      const cachedResult = toolCache.get(key);
+      const isCached = !!cachedResult;
+      emit({ type: "tool_start", id: use.id, tool: use.name, input: use.input, cached: isCached });
+      let result;
+      let ms = 0;
+      if (isCached) {
+        result = cachedResult;
+      } else {
+        const t0 = Date.now();
+        result = await executeTool(use.name, use.input);
+        ms = Date.now() - t0;
+        if (!result.isError && use.name !== "execute_api_request" && use.name !== "fetch_url") {
+          toolCache.set(key, result);
+        }
+      }
+      emit({ type: "tool_done", id: use.id, tool: use.name, output: result.text, isError: result.isError, ms, cached: isCached });
+      return { id: use.id, name: use.name, text: result.text, isError: result.isError, ms, cached: isCached };
+    };
+    const toolUses = turn.toolUses;
+    if (cfg.parallelTools && toolUses.length > 1) {
+      const pure = toolUses.filter((u) => u.name !== "execute_api_request" && u.name !== "fetch_url");
+      const sideEffect = toolUses.filter((u) => u.name === "execute_api_request" || u.name === "fetch_url");
+      const pureResults = await Promise.all(pure.map((u) => executeOne(u)));
+      const sideEffectResults = [];
+      for (const u of sideEffect)
+        sideEffectResults.push(await executeOne(u));
+      const resultMap = new Map([...pureResults, ...sideEffectResults].map((r) => [r.id, r]));
+      for (const u of toolUses) {
+        const r = resultMap.get(u.id);
+        if (r)
+          turnResults.push(r);
+      }
+    } else {
+      for (const u of toolUses)
+        turnResults.push(await executeOne(u));
+    }
+    for (const r of turnResults) {
+      allToolCalls.push({ id: r.id, tool: r.name, input: turn.toolUses.find((u) => u.id === r.id)?.input ?? {}, output: r.text, isError: r.isError, ms: r.ms, cached: r.cached });
+      if (r.isError) {
+        consecutiveErrors++;
+        if (r.name === "execute_api_request") {
+          const eid = String(turn.toolUses.find((u) => u.id === r.id)?.input?.operationId ?? r.id);
+          endpointErrors[eid] = (endpointErrors[eid] ?? 0) + 1;
+          if (endpointErrors[eid] >= cfg.maxEndpointErrors) {
+            return {
+              content: `Endpoint "${eid}" failed ${cfg.maxEndpointErrors} times. Last error: ${r.text}`,
+              toolCalls: allToolCalls,
+              stopReason: "max_endpoint_errors",
+              tokens: totalTokens
+            };
+          }
+        }
+        if (consecutiveErrors >= cfg.maxConsecutiveErrors) {
+          return {
+            content: `Stopped after ${cfg.maxConsecutiveErrors} consecutive errors. Last: ${r.text}`,
+            toolCalls: allToolCalls,
+            stopReason: "max_errors",
+            tokens: totalTokens
+          };
+        }
+      } else {
+        consecutiveErrors = 0;
+      }
+    }
+    if (isAnthropic) {
+      const blocks = turn._anthropicBlocks ?? [];
+      messages.push({ role: "assistant", content: blocks });
+      messages.push({
+        role: "user",
+        content: turnResults.map((r) => ({ type: "tool_result", tool_use_id: r.id, content: r.text }))
+      });
+    } else {
+      const tc = turn._openaiTcAccum ?? {};
+      messages.push({
+        role: "assistant",
+        content: turn.text || null,
+        tool_calls: Object.values(tc).map((t) => ({ id: t.id, type: "function", function: { name: t.name, arguments: t.args } }))
+      });
+      for (const r of turnResults) {
+        messages.push({ role: "tool", tool_call_id: r.id, content: r.text });
+      }
+    }
+  }
+  return { content: "(max iterations reached)", toolCalls: allToolCalls, stopReason: "max_iterations", tokens: totalTokens };
+}
+var RETRYABLE_STATUS, MAX_CONTEXT_CHARS = 300000, DEFAULTS;
+var init_harness = __esm(() => {
+  RETRYABLE_STATUS = new Set([429, 500, 502, 503, 504]);
+  DEFAULTS = {
+    apiKey: "",
+    baseUrl: "",
+    extraHeaders: {},
+    maxTokens: 4096,
+    temperature: 1,
+    topK: 0,
+    maxIterations: 40,
+    maxTotalTools: 40,
+    maxConsecutiveErrors: 5,
+    maxEndpointErrors: 3,
+    stepTimeoutMs: 60000,
+    parallelTools: true,
+    enablePromptCache: true
+  };
+});
+
 // src/api/routes.ts
 import dns from "dns/promises";
 function json(data, status = 200) {
@@ -4641,6 +5173,12 @@ async function apiRouter(req) {
     return handleDeleteRule(path);
   if (path === "/api/ai/chat" && method === "POST")
     return handleAiChat(req);
+  if (path === "/api/ai/memory" && method === "GET")
+    return json({ memory: dbQueries.getMemory(40) });
+  if (path === "/api/ai/memory" && method === "DELETE") {
+    dbQueries.clearMemory();
+    return json({ success: true });
+  }
   if (path === "/api/debug/dns" && method === "GET")
     return handleDnsQuery(searchParams);
   if (path === "/api/debug/ping" && method === "GET")
@@ -4866,39 +5404,54 @@ async function handleSetSettings(req) {
   dbQueries.setSettings(body);
   return json(body);
 }
-async function executeTool(name, args) {
+async function executeTool(name, args, cache = new Map) {
   const { operations, spec } = getState();
   if (name === "search_endpoints") {
+    const cacheKey2 = `search:${String(args.query ?? "").toLowerCase()}`;
+    const hit = cache.get(cacheKey2);
+    if (hit)
+      return hit;
     const q = String(args.query ?? "").toLowerCase();
     const terms = q.split(/\s+/).filter(Boolean);
     const matches = operations.filter((op) => {
       const hay = [op.operationId, op.path, op.method, ...op.tags ?? [], op.summary ?? "", op.description ?? ""].join(" ").toLowerCase();
       return terms.every((t) => hay.includes(t));
     }).slice(0, 30).map((op) => ({ operationId: op.operationId, method: op.method.toUpperCase(), path: op.path, summary: op.summary ?? null, tags: op.tags }));
-    if (!matches.length)
-      return { text: `No endpoints found matching "${args.query}". Total: ${operations.length}.`, isError: false };
-    return { text: JSON.stringify({ count: matches.length, total: operations.length, endpoints: matches }, null, 2), isError: false };
+    const text = !matches.length ? `No endpoints found matching "${args.query}". Total: ${operations.length}.` : JSON.stringify({ count: matches.length, total: operations.length, endpoints: matches }, null, 2);
+    const result = { text, isError: false };
+    cache.set(cacheKey2, result);
+    return result;
   }
   if (name === "get_endpoint_schema") {
+    const cacheKey2 = `schema:${String(args.operationId ?? "")}`;
+    const hit = cache.get(cacheKey2);
+    if (hit)
+      return hit;
     const op = operations.find((o) => o.operationId === args.operationId);
     if (!op)
       return { text: `Endpoint not found: "${args.operationId}"`, isError: true };
-    return {
-      text: JSON.stringify({
-        operationId: op.operationId,
-        method: op.method.toUpperCase(),
-        path: op.path,
-        summary: op.summary ?? null,
-        description: op.description ?? null,
-        tags: op.tags,
-        parameters: op.parameters,
-        requestBody: op.requestBody ?? null,
-        responses: op.responses
-      }, null, 2),
-      isError: false
-    };
+    const text = JSON.stringify({
+      operationId: op.operationId,
+      method: op.method.toUpperCase(),
+      path: op.path,
+      summary: op.summary ?? null,
+      description: op.description ?? null,
+      tags: op.tags,
+      parameters: op.parameters,
+      requestBody: op.requestBody ?? null,
+      responses: op.responses
+    }, null, 2);
+    const result = { text, isError: false };
+    cache.set(cacheKey2, result);
+    return result;
   }
   if (name === "execute_api_request") {
+    const now = Date.now();
+    const gap = now - _lastApiCallMs;
+    if (gap < MIN_API_CALL_INTERVAL_MS) {
+      await new Promise((r) => setTimeout(r, MIN_API_CALL_INTERVAL_MS - gap));
+    }
+    _lastApiCallMs = Date.now();
     const op = operations.find((o) => o.operationId === args.operationId);
     if (!op)
       return { text: `Endpoint not found: "${args.operationId}"`, isError: true };
@@ -4929,20 +5482,81 @@ async function executeTool(name, args) {
     const bodyStr = reqBody !== undefined ? typeof reqBody === "string" ? reqBody : JSON.stringify(reqBody) : null;
     if (bodyStr !== null && op.requestBody?.contentType)
       authedHeaders["Content-Type"] = op.requestBody.contentType;
+    const logId = randomUUID2();
     try {
       const start = Date.now();
       const res = await fetch(authedUrl, { method: op.method.toUpperCase(), headers: authedHeaders, body: bodyStr ?? undefined });
-      const text = await res.text();
+      const responseText = await res.text();
       const latency = Date.now() - start;
-      let pretty = text;
+      const resHeaders = Object.fromEntries(res.headers.entries());
+      dbQueries.insertLog({
+        id: logId,
+        source: "ai",
+        tool_name: String(args.operationId ?? op.operationId),
+        method: op.method.toUpperCase(),
+        url: authedUrl,
+        request_headers: JSON.stringify(authedHeaders),
+        request_body: bodyStr,
+        status_code: res.status,
+        response_headers: JSON.stringify(resHeaders),
+        response_body: responseText.slice(0, 8192),
+        latency_ms: latency,
+        error: null
+      });
+      logBus.emit({
+        id: logId,
+        source: "ai",
+        tool_name: String(args.operationId ?? op.operationId),
+        method: op.method.toUpperCase(),
+        url: authedUrl,
+        request_headers: JSON.stringify(authedHeaders),
+        request_body: bodyStr,
+        status_code: res.status,
+        response_headers: JSON.stringify(resHeaders),
+        response_body: responseText.slice(0, 2048),
+        latency_ms: latency,
+        error: null,
+        created_at: Date.now()
+      });
+      let pretty = responseText;
       try {
-        pretty = JSON.stringify(JSON.parse(text), null, 2);
+        pretty = JSON.stringify(JSON.parse(responseText), null, 2);
       } catch {}
       return { text: `HTTP ${res.status} (${latency}ms)
 
 ${pretty}`, isError: !res.ok };
     } catch (e) {
-      return { text: `Network error: ${e instanceof Error ? e.message : String(e)}`, isError: true };
+      const errMsg = e instanceof Error ? e.message : String(e);
+      dbQueries.insertLog({
+        id: logId,
+        source: "ai",
+        tool_name: String(args.operationId ?? op.operationId),
+        method: op.method.toUpperCase(),
+        url: authedUrl,
+        request_headers: JSON.stringify(authedHeaders),
+        request_body: bodyStr,
+        status_code: null,
+        response_headers: null,
+        response_body: null,
+        latency_ms: null,
+        error: errMsg
+      });
+      logBus.emit({
+        id: logId,
+        source: "ai",
+        tool_name: String(args.operationId ?? op.operationId),
+        method: op.method.toUpperCase(),
+        url: authedUrl,
+        request_headers: null,
+        request_body: bodyStr,
+        status_code: null,
+        response_headers: null,
+        response_body: null,
+        latency_ms: null,
+        error: errMsg,
+        created_at: Date.now()
+      });
+      return { text: `Network error: ${errMsg}`, isError: true };
     }
   }
   if (name === "fetch_url") {
@@ -5092,10 +5706,25 @@ ${stripped}`, isError: !res.ok };
   }
   if (name === "save_auth_token") {
     const profileName = String(args.name ?? "AI Login").trim();
+    const tokenType = String(args.token_type ?? "bearer");
+    if (tokenType === "basic" || args.username && args.password) {
+      const username = String(args.username ?? "").trim();
+      const password = String(args.password ?? "").trim();
+      if (!username || !password)
+        return { text: "Error: username and password are required for basic auth", isError: true };
+      const authConfig2 = { type: "basic", username, password };
+      const profileId2 = randomUUID2();
+      try {
+        dbQueries.insertProfile({ id: profileId2, name: profileName, description: "Saved by AI", type: "basic", config: JSON.stringify(authConfig2), token_cache: null, is_active: 0 });
+        dbQueries.activateProfile(profileId2);
+        return { text: JSON.stringify({ success: true, message: `Saved and activated basic auth profile "${profileName}"`, id: profileId2 }), isError: false };
+      } catch (e) {
+        return { text: `Error saving profile: ${e instanceof Error ? e.message : String(e)}`, isError: true };
+      }
+    }
     const token = String(args.token ?? "").trim();
     if (!token)
-      return { text: "Error: token is required", isError: true };
-    const tokenType = String(args.token_type ?? "bearer");
+      return { text: "Error: token is required for bearer/apikey auth", isError: true };
     const headerName = String(args.header_name ?? "X-Api-Key");
     let authConfig;
     let type;
@@ -5120,274 +5749,6 @@ ${stripped}`, isError: !res.ok };
   }
   return { text: `Unknown tool: ${name}`, isError: true };
 }
-async function fetchWithRetry(url, opts, emit, maxRetries = 3) {
-  for (let attempt = 0;attempt <= maxRetries; attempt++) {
-    const res = await fetch(url, opts);
-    if (res.status !== 429 || attempt === maxRetries)
-      return res;
-    const retryAfter = parseInt(res.headers.get("retry-after") ?? "0", 10);
-    const delay = retryAfter > 0 ? retryAfter * 1000 : Math.min(1000 * Math.pow(2, attempt) + Math.random() * 500, 30000);
-    emit({ type: "info", message: `Rate limited \u2014 retrying in ${Math.round(delay / 1000)}s\u2026 (attempt ${attempt + 1}/${maxRetries})` });
-    await new Promise((r) => setTimeout(r, delay));
-  }
-  return fetch(url, opts);
-}
-async function anthropicAgentLoop(apiKey, model, system, initialMessages, emit) {
-  const msgs = [...initialMessages];
-  const toolCalls = [];
-  let totalTools = 0;
-  let consecutiveErrors = 0;
-  const endpointErrors = {};
-  for (let iter = 0;iter < 40; iter++) {
-    const res = await fetchWithRetry("https://api.anthropic.com/v1/messages", {
-      method: "POST",
-      headers: { "Content-Type": "application/json", "x-api-key": apiKey, "anthropic-version": "2023-06-01" },
-      body: JSON.stringify({ model, max_tokens: 4096, system, messages: msgs, tools: ANTHROPIC_TOOLS, stream: true })
-    }, emit);
-    if (!res.ok)
-      throw new Error(`Anthropic error: ${await res.text()}`);
-    let fullText = "";
-    let stopReason = "";
-    const contentBlocks = [];
-    const inputAccum = {};
-    const reader = res.body.getReader();
-    const decoder = new TextDecoder;
-    let buf = "";
-    while (true) {
-      const { done, value } = await reader.read();
-      if (done)
-        break;
-      buf += decoder.decode(value, { stream: true });
-      const parts = buf.split(`
-
-`);
-      buf = parts.pop() ?? "";
-      for (const part of parts) {
-        let dataLine = "";
-        for (const line of part.split(`
-`)) {
-          if (line.startsWith("data: ")) {
-            dataLine = line.slice(6);
-            break;
-          }
-        }
-        if (!dataLine || dataLine === "[DONE]")
-          continue;
-        let ev;
-        try {
-          ev = JSON.parse(dataLine);
-        } catch {
-          continue;
-        }
-        const type = ev.type;
-        if (type === "content_block_start") {
-          const idx = ev.index;
-          const cb = ev.content_block;
-          contentBlocks[idx] = { type: cb.type, id: cb.id, name: cb.name };
-          if (cb.type === "tool_use")
-            inputAccum[idx] = "";
-        } else if (type === "content_block_delta") {
-          const idx = ev.index;
-          const delta = ev.delta;
-          if (delta.type === "text_delta" && delta.text) {
-            fullText += delta.text;
-            if (!contentBlocks[idx])
-              contentBlocks[idx] = { type: "text", text: "" };
-            contentBlocks[idx].text = (contentBlocks[idx].text ?? "") + delta.text;
-            emit({ type: "text_delta", text: delta.text });
-          } else if (delta.type === "input_json_delta" && delta.partial_json) {
-            inputAccum[idx] = (inputAccum[idx] ?? "") + delta.partial_json;
-          }
-        } else if (type === "content_block_stop") {
-          const idx = ev.index;
-          if (contentBlocks[idx]?.type === "tool_use") {
-            try {
-              contentBlocks[idx].input = JSON.parse(inputAccum[idx] ?? "{}");
-            } catch {
-              contentBlocks[idx].input = {};
-            }
-          }
-        } else if (type === "message_delta") {
-          const delta = ev.delta;
-          if (delta.stop_reason)
-            stopReason = delta.stop_reason;
-        }
-      }
-    }
-    if (stopReason !== "tool_use")
-      return { content: fullText, toolCalls };
-    if (totalTools >= MAX_TOTAL_TOOLS) {
-      return { content: `Agent stopped: reached ${MAX_TOTAL_TOOLS} tool calls. Please break your request into smaller steps.`, toolCalls };
-    }
-    msgs.push({ role: "assistant", content: contentBlocks });
-    const toolResults = [];
-    for (const block of contentBlocks) {
-      if (block.type !== "tool_use" || !block.id || !block.name)
-        continue;
-      totalTools++;
-      emit({ type: "tool_start", tool: block.name, input: block.input ?? {} });
-      const result = await executeTool(block.name, block.input ?? {});
-      emit({ type: "tool_done", tool: block.name, input: block.input ?? {}, output: result.text, isError: result.isError });
-      toolCalls.push({ tool: block.name, input: block.input ?? {}, output: result.text, isError: result.isError });
-      if (result.isError) {
-        consecutiveErrors++;
-        if (block.name === "execute_api_request" && block.input?.operationId) {
-          const eid = String(block.input.operationId);
-          endpointErrors[eid] = (endpointErrors[eid] ?? 0) + 1;
-          if (endpointErrors[eid] >= MAX_SAME_ENDPOINT_ERRORS) {
-            const stopContent = result.text + `
-
-[AGENT LOOP STOPPED: endpoint "${eid}" failed ${MAX_SAME_ENDPOINT_ERRORS} times \u2014 stopping to avoid loop]`;
-            toolResults.push({ type: "tool_result", tool_use_id: block.id, content: stopContent });
-            msgs.push({ role: "user", content: toolResults });
-            return { content: `Endpoint "${eid}" failed ${MAX_SAME_ENDPOINT_ERRORS} times. Last error: ${result.text}`, toolCalls };
-          }
-        }
-        if (consecutiveErrors >= MAX_CONSECUTIVE_ERRORS) {
-          const stopMsg = `Stopped after ${MAX_CONSECUTIVE_ERRORS} consecutive errors. Last: ${result.text}`;
-          const stopContent = result.text + `
-
-[AGENT LOOP STOPPED: ${stopMsg}]`;
-          toolResults.push({ type: "tool_result", tool_use_id: block.id, content: stopContent });
-          msgs.push({ role: "user", content: toolResults });
-          return { content: stopMsg, toolCalls };
-        }
-      } else {
-        consecutiveErrors = 0;
-      }
-      toolResults.push({ type: "tool_result", tool_use_id: block.id, content: result.text });
-    }
-    if (!toolResults.length)
-      return { content: fullText, toolCalls };
-    msgs.push({ role: "user", content: toolResults });
-  }
-  return { content: "(max iterations reached)", toolCalls };
-}
-async function openaiCompatibleLoop(base, apiKey, model, extraHeaders, system, initialMessages, emit) {
-  const msgs = [{ role: "system", content: system }, ...initialMessages];
-  const toolCalls = [];
-  const authHeaders = {};
-  if (apiKey)
-    authHeaders["Authorization"] = `Bearer ${apiKey}`;
-  let totalTools = 0;
-  let consecutiveErrors = 0;
-  const endpointErrors = {};
-  for (let iter = 0;iter < 40; iter++) {
-    const res = await fetchWithRetry(`${base}/v1/chat/completions`, {
-      method: "POST",
-      headers: { "Content-Type": "application/json", ...authHeaders, ...extraHeaders },
-      body: JSON.stringify({ model, messages: msgs, tools: OPENAI_TOOLS, tool_choice: "auto", stream: true })
-    }, emit);
-    if (!res.ok)
-      throw new Error(await res.text());
-    let fullContent = "";
-    let finishReason = "";
-    const tcAccum = {};
-    const reader = res.body.getReader();
-    const dec = new TextDecoder;
-    let buf = "";
-    outer:
-      while (true) {
-        const { done, value } = await reader.read();
-        if (done)
-          break;
-        buf += dec.decode(value, { stream: true });
-        const parts = buf.split(`
-
-`);
-        buf = parts.pop() ?? "";
-        for (const part of parts) {
-          let data = "";
-          for (const line of part.split(`
-`)) {
-            if (line.startsWith("data: ")) {
-              data = line.slice(6);
-              break;
-            }
-          }
-          if (!data)
-            continue;
-          if (data === "[DONE]")
-            break outer;
-          let ev;
-          try {
-            ev = JSON.parse(data);
-          } catch {
-            continue;
-          }
-          if (ev.object === "error")
-            throw new Error(JSON.stringify(ev));
-          const choices = ev.choices;
-          const choice = choices?.[0];
-          if (!choice)
-            continue;
-          const fr = choice.finish_reason;
-          if (fr)
-            finishReason = fr;
-          const delta = choice.delta;
-          if (!delta)
-            continue;
-          if (typeof delta.content === "string" && delta.content) {
-            fullContent += delta.content;
-            emit({ type: "text_delta", text: delta.content });
-          }
-          const tcDeltas = delta.tool_calls;
-          if (tcDeltas) {
-            for (const tc of tcDeltas) {
-              if (!tcAccum[tc.index])
-                tcAccum[tc.index] = { id: "", name: "", args: "" };
-              const entry = tcAccum[tc.index];
-              if (tc.id)
-                entry.id += tc.id;
-              if (tc.function?.name)
-                entry.name += tc.function.name;
-              if (tc.function?.arguments)
-                entry.args += tc.function.arguments;
-            }
-          }
-        }
-      }
-    if (finishReason !== "tool_calls")
-      return { content: fullContent, toolCalls };
-    if (totalTools >= MAX_TOTAL_TOOLS) {
-      return { content: `Agent stopped: reached ${MAX_TOTAL_TOOLS} tool calls. Please break your request into smaller steps.`, toolCalls };
-    }
-    const msgToolCalls = Object.values(tcAccum).map((tc) => ({
-      id: tc.id,
-      type: "function",
-      function: { name: tc.name, arguments: tc.args }
-    }));
-    msgs.push({ role: "assistant", content: fullContent || null, tool_calls: msgToolCalls });
-    for (const tc of Object.values(tcAccum)) {
-      let args = {};
-      try {
-        args = JSON.parse(tc.args);
-      } catch {}
-      totalTools++;
-      emit({ type: "tool_start", tool: tc.name, input: args });
-      const result = await executeTool(tc.name, args);
-      emit({ type: "tool_done", tool: tc.name, input: args, output: result.text, isError: result.isError });
-      toolCalls.push({ tool: tc.name, input: args, output: result.text, isError: result.isError });
-      msgs.push({ role: "tool", tool_call_id: tc.id, content: result.text });
-      if (result.isError) {
-        consecutiveErrors++;
-        if (tc.name === "execute_api_request" && args.operationId) {
-          const eid = String(args.operationId);
-          endpointErrors[eid] = (endpointErrors[eid] ?? 0) + 1;
-          if (endpointErrors[eid] >= MAX_SAME_ENDPOINT_ERRORS) {
-            return { content: `Endpoint "${eid}" failed ${MAX_SAME_ENDPOINT_ERRORS} times. Last error: ${result.text}`, toolCalls };
-          }
-        }
-        if (consecutiveErrors >= MAX_CONSECUTIVE_ERRORS) {
-          return { content: `Stopped after ${MAX_CONSECUTIVE_ERRORS} consecutive errors. Last: ${result.text}`, toolCalls };
-        }
-      } else {
-        consecutiveErrors = 0;
-      }
-    }
-  }
-  return { content: "(max iterations reached)", toolCalls };
-}
 async function handleAiChat(req) {
   let body;
   try {
@@ -5398,43 +5759,58 @@ async function handleAiChat(req) {
   const settingsRow = dbQueries.getSettings();
   const settings = settingsRow ? JSON.parse(settingsRow.value) : {};
   const ai = settings.ai ?? {};
+  const provider = ai.provider ?? "anthropic";
+  const providerDefaults = PROVIDER_DEFAULTS[provider] ?? { model: "" };
+  const requiresKey = provider !== "ollama" && provider !== "custom";
+  if (requiresKey && !ai.apiKey) {
+    return json({ error: "No AI API key configured. Go to Settings \u2192 AI Provider to add one." }, 400);
+  }
+  if (!hasState())
+    return json({ error: "No spec loaded." }, 400);
   const { spec, operations } = getState();
   const preview = operations.slice(0, 40).map((op) => `- ${op.method.toUpperCase()} ${op.path}${op.summary ? `: ${op.summary}` : ""}`).join(`
 `);
   const activeAuth = dbQueries.getActiveProfile();
-  const authLine = activeAuth ? `Active auth: "${activeAuth.name}" (${activeAuth.type})` : "No active auth profile. If the API requires auth, call list_auth_profiles first, then set_active_auth, or login and call save_auth_token.";
+  const authLine = activeAuth ? `Active auth: "${activeAuth.name}" (${activeAuth.type})` : "No active auth profile. Call list_auth_profiles, then set_active_auth or save_auth_token.";
+  const memory = dbQueries.getMemory(20);
+  const memorySection = memory.length ? `
+## Memory from previous sessions
+${memory.map((m) => `${m.role === "user" ? "User" : "Assistant"}: ${m.content.slice(0, 300)}${m.content.length > 300 ? "\u2026" : ""}`).join(`
+`)}
+` : "";
   const system = `You are an AI assistant for the "${spec.title}" API (v${spec.version}). Base URL: ${spec.baseUrl}.
 Total endpoints: ${operations.length}. Sample:
 ${preview}${operations.length > 40 ? `
 ... and ${operations.length - 40} more` : ""}
 
 ${authLine}
+${memorySection}
+Tools:
+- search_endpoints / get_endpoint_schema \u2014 explore API structure (results cached; never repeat the same query)
+- execute_api_request \u2014 call an endpoint
+- list_auth_profiles / set_active_auth / save_auth_token \u2014 manage credentials
+  \u2022 save_auth_token supports token_type="basic" with username+password for HTTP Basic auth
+- fetch_url \u2014 external docs
+- dns_lookup \u2014 connectivity diagnostics
+- get_recent_logs \u2014 proxy traffic history
+- run_security_check \u2014 static security analysis
 
-Tools available:
-- search_endpoints / get_endpoint_schema: explore API structure
-- execute_api_request: call an endpoint
-- list_auth_profiles: list all saved auth profiles (name, type, active)
-- set_active_auth(name): switch to a saved profile before making requests
-- save_auth_token(name, token): IMMEDIATELY call this after a successful login that returns a token \u2014 saves the token as a named profile and activates it so subsequent requests are authenticated
-- fetch_url: fetch external docs
-- dns_lookup: DNS resolution / connectivity
-- get_recent_logs: recent request/response traffic
-- run_security_check: security analysis on an endpoint
+Auth workflow: 401/403 \u2192 list_auth_profiles \u2192 set_active_auth OR find login endpoint \u2192 save_auth_token \u2192 retry.
 
-Authentication workflow: if requests return 401/403, call list_auth_profiles first. If a profile exists, call set_active_auth. If none, find and call the login endpoint, extract the token from the response, then call save_auth_token immediately. After saving, retry the original request.
+Rules:
+- Never repeat a search you already ran \u2014 results are cached.
+- Diagnose errors before retrying. Three failures on the same endpoint stops the agent.
+- Do not fire rapid successive API requests.
 
-IMPORTANT: if an endpoint returns an error, diagnose it (check the schema, check auth) and fix the root cause before retrying. If the same endpoint fails 3 times the agent will be forcibly stopped. Do not retry without changing something.
+Be concise. Format code and JSON in fenced blocks.${ai.customInstructions ? `
 
-Be concise and practical. Format code and JSON in code blocks.${body.extra_context ? `
+---
+## Custom instructions
+${ai.customInstructions}` : ""}${body.extra_context ? `
 
 ---
-## Current context
+## Context
 ${body.extra_context}` : ""}`;
-  const provider = ai.provider ?? "anthropic";
-  const requiresKey = provider !== "ollama" && provider !== "custom";
-  if (requiresKey && !ai.apiKey) {
-    return json({ error: "No AI API key configured. Go to Settings \u2192 AI Provider to add one." }, 400);
-  }
   const { readable, writable } = new TransformStream;
   const writer = writable.getWriter();
   const enc = new TextEncoder;
@@ -5444,62 +5820,39 @@ ${body.extra_context}` : ""}`;
 `)).catch(() => {});
   };
   const msgs = body.messages;
+  const toolCache = new Map;
+  const abortCtrl = new AbortController;
+  const lastUserMsg = [...msgs].reverse().find((m) => m.role === "user");
+  const userMemoryContent = typeof lastUserMsg?.content === "string" ? lastUserMsg.content : null;
   (async () => {
     try {
-      let result;
-      if (provider === "anthropic") {
-        result = await anthropicAgentLoop(ai.apiKey, ai.model || "claude-haiku-4-5-20251001", system, msgs, emit);
-      } else if (provider === "openai") {
-        const base = (ai.baseUrl || "https://api.openai.com").replace(/\/$/, "");
-        result = await openaiCompatibleLoop(base, ai.apiKey, ai.model || "gpt-4o-mini", {}, system, msgs, emit);
-      } else if (provider === "mistral") {
-        const base = (ai.baseUrl || "https://api.mistral.ai").replace(/\/$/, "");
-        result = await openaiCompatibleLoop(base, ai.apiKey, ai.model || "mistral-small-latest", {}, system, msgs, emit);
-      } else if (provider === "github-copilot") {
-        const base = (ai.baseUrl || "https://api.githubcopilot.com").replace(/\/$/, "");
-        result = await openaiCompatibleLoop(base, ai.apiKey, ai.model || "gpt-4o", {
-          "Copilot-Integration-Id": "vscode-chat",
-          "Editor-Version": "vscode/1.85.0"
-        }, system, msgs, emit);
-      } else if (provider === "groq") {
-        const base = (ai.baseUrl || "https://api.groq.com/openai").replace(/\/$/, "");
-        result = await openaiCompatibleLoop(base, ai.apiKey, ai.model || "llama-3.1-70b-versatile", {}, system, msgs, emit);
-      } else if (provider === "custom") {
-        if (!ai.baseUrl) {
-          emit({ type: "error", message: "Custom provider requires a Base URL." });
-          await writer.close();
-          return;
-        }
-        result = await openaiCompatibleLoop(ai.baseUrl.replace(/\/$/, ""), ai.apiKey, ai.model || "", {}, system, msgs, emit);
-      } else if (provider === "ollama") {
-        const base = (ai.baseUrl || "http://localhost:11434").replace(/\/$/, "");
-        const res = await fetch(`${base}/api/chat`, {
-          method: "POST",
-          headers: { "Content-Type": "application/json" },
-          body: JSON.stringify({ model: ai.model || "llama3", messages: [{ role: "system", content: system }, ...msgs], stream: false })
-        });
-        const d = await res.json();
-        result = { content: d.message.content ?? "", toolCalls: [] };
-      } else if (provider === "gemini") {
-        const model = ai.model || "gemini-1.5-flash";
-        const base = (ai.baseUrl || "https://generativelanguage.googleapis.com").replace(/\/$/, "");
-        const res = await fetch(`${base}/v1beta/models/${model}:generateContent?key=${ai.apiKey}`, {
-          method: "POST",
-          headers: { "Content-Type": "application/json" },
-          body: JSON.stringify({
-            systemInstruction: { parts: [{ text: system }] },
-            contents: msgs.map((m) => ({ role: m.role === "assistant" ? "model" : "user", parts: [{ text: m.content }] })),
-            generationConfig: { maxOutputTokens: 4096 }
-          })
-        });
-        const d = await res.json();
-        result = { content: d.candidates[0]?.content.parts[0]?.text ?? "", toolCalls: [] };
-      } else {
-        emit({ type: "error", message: `Unknown provider: ${provider}` });
-        await writer.close();
-        return;
+      const result = await runAgentLoop({
+        provider,
+        apiKey: ai.apiKey,
+        model: ai.model || providerDefaults.model,
+        baseUrl: ai.baseUrl || providerDefaults.baseUrl,
+        maxTokens: ai.maxTokens ?? 4096,
+        stepTimeoutMs: ai.stepTimeoutMs ?? 60000,
+        temperature: ai.temperature,
+        topK: ai.topK && ai.topK > 0 ? ai.topK : undefined,
+        parallelTools: true,
+        enablePromptCache: true
+      }, system, msgs, TOOL_SCHEMAS, (name, args) => executeTool(name, args, toolCache), emit, abortCtrl.signal, toolCache);
+      if (result.content && result.stopReason !== "max_iterations") {
+        try {
+          if (userMemoryContent)
+            dbQueries.saveMemory("user", userMemoryContent.slice(0, 1000));
+          dbQueries.saveMemory("assistant", result.content.slice(0, 1000));
+          dbQueries.trimMemory(40);
+        } catch {}
       }
-      emit({ type: "done", content: result.content, toolCalls: result.toolCalls });
+      emit({
+        type: "done",
+        content: result.content,
+        toolCalls: result.toolCalls,
+        stopReason: result.stopReason,
+        tokens: result.tokens
+      });
     } catch (e) {
       emit({ type: "error", message: e instanceof Error ? e.message : String(e) });
     } finally {
@@ -5509,11 +5862,7 @@ ${body.extra_context}` : ""}`;
     }
   })();
   return new Response(readable, {
-    headers: {
-      "Content-Type": "text/event-stream",
-      "Cache-Control": "no-cache",
-      ...CORS4
-    }
+    headers: { "Content-Type": "text/event-stream", "Cache-Control": "no-cache", ...CORS4 }
   });
 }
 function handleGetProfiles() {
@@ -6120,7 +6469,7 @@ function handleDeleteCaptureBin(path) {
   dbQueries.deleteCaptureBin(id);
   return json({ ok: true });
 }
-var CORS4, TOOL_DEFS, ANTHROPIC_TOOLS, OPENAI_TOOLS, MAX_TOTAL_TOOLS = 40, MAX_CONSECUTIVE_ERRORS = 5, MAX_SAME_ENDPOINT_ERRORS = 3;
+var CORS4, TOOL_DEFS, _lastApiCallMs = 0, MIN_API_CALL_INTERVAL_MS = 400, TOOL_SCHEMAS, PROVIDER_DEFAULTS;
 var init_routes = __esm(() => {
   init_db();
   init_engine();
@@ -6130,6 +6479,7 @@ var init_routes = __esm(() => {
   init_config();
   init_version();
   init_engine2();
+  init_harness();
   CORS4 = {
     "Access-Control-Allow-Origin": "*",
     "Access-Control-Allow-Methods": "GET, POST, PUT, PATCH, DELETE, OPTIONS",
@@ -6198,25 +6548,34 @@ var init_routes = __esm(() => {
       required: ["name"]
     },
     save_auth_token: {
-      description: "Save a bearer token or API key as a named auth profile and immediately activate it. Call this right after a successful login endpoint returns a token so all subsequent API requests are authenticated.",
+      description: "Save a bearer token, API key, or basic auth credentials as a named auth profile and immediately activate it. Call this right after a successful login endpoint returns a token so all subsequent API requests are authenticated.",
       params: {
         name: { type: "string", description: 'Profile name, e.g. "user session" or the username' },
-        token: { type: "string", description: "The bearer token or API key value to save" },
-        token_type: { type: "string", enum: ["bearer", "apikey_header", "apikey_query"], description: "Token type (default: bearer)" },
-        header_name: { type: "string", description: "Header name for apikey_header type (default: X-Api-Key)" }
+        token: { type: "string", description: "The bearer token or API key value (omit for basic auth)" },
+        token_type: { type: "string", enum: ["bearer", "apikey_header", "apikey_query", "basic"], description: "Token type (default: bearer)" },
+        header_name: { type: "string", description: "Header name for apikey_header type (default: X-Api-Key)" },
+        username: { type: "string", description: "Username for basic auth" },
+        password: { type: "string", description: "Password for basic auth" }
       },
-      required: ["name", "token"]
+      required: ["name"]
     }
   };
-  ANTHROPIC_TOOLS = Object.entries(TOOL_DEFS).map(([name, def]) => ({
+  TOOL_SCHEMAS = Object.entries(TOOL_DEFS).map(([name, def]) => ({
     name,
     description: def.description,
-    input_schema: { type: "object", properties: def.params, required: def.required }
-  }));
-  OPENAI_TOOLS = Object.entries(TOOL_DEFS).map(([name, def]) => ({
-    type: "function",
-    function: { name, description: def.description, parameters: { type: "object", properties: def.params, required: def.required } }
+    params: def.params,
+    required: def.required
   }));
+  PROVIDER_DEFAULTS = {
+    anthropic: { model: "claude-haiku-4-5-20251001" },
+    openai: { model: "gpt-4o-mini", baseUrl: "https://api.openai.com" },
+    mistral: { model: "mistral-small-latest", baseUrl: "https://api.mistral.ai" },
+    groq: { model: "llama-3.1-70b-versatile", baseUrl: "https://api.groq.com/openai" },
+    "github-copilot": { model: "gpt-4o", baseUrl: "https://api.githubcopilot.com" },
+    ollama: { model: "llama3", baseUrl: "http://localhost:11434" },
+    gemini: { model: "gemini-1.5-flash" },
+    custom: { model: "" }
+  };
 });
 
 // src/daemon.ts