agent.libx.js 0.93.22 → 0.93.24

package/README.md

@@ -77,7 +77,7 @@ agentx --resume <id> "…"    # resume a specific session
 - **@-file mentions & headless JSON** — reference files inline in a prompt with `@path` (e.g. `explain @src/Agent.ts`); script with `-p --output-format json` to get one machine-readable result object on stdout (activity stays on stderr).
 - **Tab-completion** — `Tab` completes `/<command>` names and `@<path>` file/dir references (descends subdirs, dotfiles hidden unless typed) straight from the working tree.
 - **Duplex mode** — `agentx --duplex` runs the full standard REPL (slash commands, sessions, postures, rewind, MCP) with the dual-model engine driving turns: a fast voice model (`--voice-model`, default haiku) answers every line instantly and delegates real work to background workers built with the same wiring as a normal run (fs mode, permissions, MCP); worker activity shows as dim chrome and results are re-voiced when ready.
-- **MCP servers** — declare `mcpServers: { name: { command, args } | { url } }` in config and they're auto-mounted at startup (in parallel, with an optional `mountTimeoutMs` deadline so one slow/dead server never blocks the rest): the client does the JSON-RPC handshake (stdio or HTTP) + `tools/list`, and the discovered tools appear as `mcp__<name>__<tool>` in `/tools` (inspect with `/mcp`). A bad server is logged and skipped, never blocking the agent. For large tool sets, **deferred mode** (`makeMcpToolSearch` / `mountMcpDeferred`) exposes just two bounded tools (`ToolSearch` + `McpCall`) instead of N defs — dodging the provider tool-cap and improving selection accuracy. **`mountMcpCatalog`** goes further: a cached, hash-keyed catalog + lazy connect means a turn that uses no MCP tool opens **zero** connections, and one that uses a tool connects exactly that server — latency scales with tools-used, not servers-configured.
+- **MCP servers** — declare `mcpServers: { name: { command, args } | { url } }` in config and they're auto-mounted at startup (in parallel, with an optional `mountTimeoutMs` deadline so one slow/dead server never blocks the rest): the client does the JSON-RPC handshake (stdio or HTTP) + `tools/list`, and the discovered tools appear as `mcp__<name>__<tool>` in `/tools` (inspect with `/mcp`). A bad server is logged and skipped, never blocking the agent. For large tool sets, **deferred mode** (`makeMcpToolSearch` / `mountMcpDeferred`) exposes just two bounded tools (`ToolSearch` + `McpCall`) instead of N defs — dodging the provider tool-cap and improving selection accuracy. **`mountMcpCatalog`** goes further: a cached, hash-keyed catalog + lazy connect means a turn that uses no MCP tool opens **zero** connections, and one that uses a tool connects exactly that server — latency scales with tools-used, not servers-configured. A down server is **negative-cached** (`failureCooldownMs`) so it never re-floors a later turn at the deadline. For zero turn-path latency even on a cold process, call **`warmMcpCatalog`** at boot + on a timer (off-turn discovery) and mount with **`{ discover: 'cache-only' }`** — the turn then never synchronously connects: it serves the warmed catalog and discovers any miss in the background.
 
 ## 🧬 It improves itself
 

package/dist/cli.js

@@ -596,35 +596,75 @@ ${capped}`;
     }
   };
 }
+function decodeDdgUrl(href) {
+  const m = href.match(/[?&]uddg=([^&]+)/);
+  if (m) {
+    try {
+      return decodeURIComponent(m[1]);
+    } catch {
+    }
+  }
+  return href.startsWith("//") ? "https:" + href : href;
+}
+function parseDdgHtml(html, max) {
+  const anchors = [...html.matchAll(/<a[^>]*class="[^"]*result__a[^"]*"[^>]*href="([^"]+)"[^>]*>([\s\S]*?)<\/a>/g)];
+  const snippets = [...html.matchAll(/<a[^>]*class="[^"]*result__snippet[^"]*"[^>]*>([\s\S]*?)<\/a>/g)].map((m) => htmlToText(m[1]));
+  const hits = [];
+  for (let i = 0; i < anchors.length && hits.length < max; i++) {
+    const url = decodeDdgUrl(anchors[i][1]);
+    try {
+      if (isPrivateHost(new URL(url).hostname)) continue;
+    } catch {
+      continue;
+    }
+    hits.push({ title: htmlToText(anchors[i][2]) || "(untitled)", url, snippet: snippets[i] ?? "" });
+  }
+  return hits;
+}
+function formatHits(hits) {
+  if (!hits.length) return "(no results)";
+  return hits.map((r, i) => `${i + 1}. ${r.title}
+   ${r.url}
+   ${r.snippet.replace(/\s+/g, " ").slice(0, 240)}`).join("\n\n");
+}
 function makeWebSearchTool(options = {}) {
-  const endpoint = options.endpoint ?? "https://api.tavily.com/search";
+  const tavilyEndpoint = options.endpoint ?? "https://api.tavily.com/search";
   const maxResults = options.maxResults ?? 5;
   const timeoutMs = options.timeoutMs ?? 15e3;
   return {
     name: "WebSearch",
-    description: "Search the web; returns ranked results (title, URL, snippet). Requires a configured provider \u2014 set TAVILY_API_KEY to enable.",
+    description: "Search the web by query; returns ranked results (title, URL, snippet). Use to look things up, find pages, or research a topic \u2014 then WebFetch a result URL to read it in full.",
     parameters: { type: "object", required: ["query"], properties: { query: { type: "string" } } },
     async run({ query }) {
       const doFetch = options.fetch ?? globalThis.fetch;
-      const key = options.apiKey ?? process.env.TAVILY_API_KEY;
-      if (!key) return "Error: WebSearch is not configured. Set TAVILY_API_KEY (https://tavily.com) to enable web search.";
       if (!doFetch) return "Error: no network (fetch) available in this runtime";
+      const q2 = String(query ?? "").trim();
+      if (!q2) return "Error: empty query";
+      const key = options.apiKey ?? process.env.TAVILY_API_KEY;
+      const provider = options.provider ?? "auto";
+      const useTavily = provider === "tavily" || provider === "auto" && !!key;
       const ctl = new AbortController();
       const timer = setTimeout(() => ctl.abort(), timeoutMs);
       try {
-        const res = await doFetch(endpoint, {
-          method: "POST",
+        if (useTavily) {
+          if (!key) return "Error: Tavily provider selected but TAVILY_API_KEY is not set";
+          const res2 = await doFetch(tavilyEndpoint, {
+            method: "POST",
+            signal: ctl.signal,
+            headers: { "content-type": "application/json" },
+            body: JSON.stringify({ api_key: key, query: q2, max_results: maxResults })
+          });
+          if (!res2.ok) return `Error: search provider returned ${res2.status} ${res2.statusText}`;
+          const data = await res2.json();
+          const results = Array.isArray(data?.results) ? data.results.slice(0, maxResults) : [];
+          return formatHits(results.map((r) => ({ title: r.title ?? "(untitled)", url: r.url ?? "", snippet: String(r.content ?? "") })));
+        }
+        const res = await doFetch("https://html.duckduckgo.com/html/?q=" + encodeURIComponent(q2), {
           signal: ctl.signal,
-          headers: { "content-type": "application/json" },
-          body: JSON.stringify({ api_key: key, query: String(query ?? ""), max_results: maxResults })
+          headers: { "user-agent": "Mozilla/5.0 (compatible; agentx/1.0; +https://github.com/Livshitz/agent.libx.js)" }
         });
-        if (!res.ok) return `Error: search provider returned ${res.status} ${res.statusText}`;
-        const data = await res.json();
-        const results = Array.isArray(data?.results) ? data.results.slice(0, maxResults) : [];
-        if (!results.length) return "(no results)";
-        return results.map((r, i) => `${i + 1}. ${r.title ?? "(untitled)"}
-   ${r.url ?? ""}
-   ${String(r.content ?? "").replace(/\s+/g, " ").slice(0, 240)}`).join("\n\n");
+        if (!res.ok) return `Error: search returned ${res.status} ${res.statusText}`;
+        return formatHits(parseDdgHtml(await res.text(), maxResults));
       } catch (e) {
         log2.debug("WebSearch failed", e);
         return `Error searching: ${e?.name === "AbortError" ? `timed out after ${timeoutMs}ms` : e?.message ?? e}`;
@@ -3563,7 +3603,7 @@ var DuplexAgentOptions = class {
   /** User-scope memory dir for global facts (type=user/feedback). Forwarded to Remember's routing. */
   memoryUserDir;
 };
-var VOICE_SYSTEM_PROMPT = 'You are a spoken voice assistant \u2014 the user HEARS everything you say. Use short sentences. One idea per sentence. No markdown, no bullet lists, no code blocks, no headings, no emoji.\nKeep turns SHORT \u2014 one to three sentences, then stop. Never lecture, enumerate cases, or add caveats unprompted. Conversation is a fast exchange: give the one thing asked, and let the user pull more if they want it.\nYou have three cognitive tiers \u2014 like a human brain:\n\u2022 YOU (reflex) \u2014 instant, lightweight. Handle greetings, simple questions, status checks, QuickLook.\n\u2022 `Act` \u2014 your hands. A background worker with its own configured tools and access to the user\'s environment (typically files and shell; other tools such as web access only if this deployment wired them). Use for reading, editing, searching, running tasks, building \u2014 any real work.\n{{THINK_SLOT}}\nWhen you are unsure whether you can do or access something, do NOT assume and do NOT claim a capability you have not confirmed. To check what you can do, QuickLook `capabilities` (instant \u2014 it lists your worker\'s real tools) and answer from that. Never promise an ability (e.g. web search, live data) that is not in your capabilities; if it is not there, tell the user plainly you can\'t. To actually DO real work, call `Act`. When the user mentions their project, folder, files, or environment ("this project", "the current folder", "my code"), call `Act` IMMEDIATELY \u2014 do not ask for paths or details the worker can discover itself. Never pretend to have done the work or invent results \u2014 the worker\'s report is your only source.\nALWAYS react before you work: the FIRST thing in your turn is a brief spoken acknowledgement of what you heard and what you are about to do ("got it \u2014 opening that now", "sure, let me pull it up", "okay, checking"). NEVER call a tool (Act, Think, QuickLook) silently \u2014 the user must hear you react before you go quiet to work. After dispatching Act or Think, that same one short sentence IS your turn \u2014 end it and do not wait for the result.\nResults arrive later as events like "[task t1 completed] \u2026" or "[task t1 failed] \u2026". When one arrives, summarize it for the ear in one or two short sentences. "[task t1 progress] \u2026" events are interim status, NOT results \u2014 give at most a half-sentence aside ("still on it \u2014 running tests now") and end your turn. Never present progress as a finished result.\nCRITICAL: while a task is still running you have NO answer yet \u2014 never state a specific result of any kind (a number, size, count, name, path, or value). The real answer arrives ONLY in the "[task \u2026 completed]" event; inventing one meanwhile (a made-up disk size, commit count, etc.) is a serious error. Until then, only acknowledge and wait.\nNever read raw file paths, diffs, or code aloud verbatim.\n"[task t1 asks] \u2026" events are QUESTIONS from a background task \u2014 relay to the user in your own words, short, then end your turn. When the user answers, call `AnswerTask` with that id and their answer. NEVER answer on the user\'s behalf for permissions or risky operations; if their reply is ambiguous, confirm first.\nIf the user\'s message sounds INCOMPLETE \u2014 trailing off mid-sentence, a fragment that needs more context ("and then we", "but the problem is"), hesitation fillers ("uh", "um") \u2014 call `Hold` instead of answering. This keeps listening for the rest of their thought. Only respond with substance when you have a complete question or request.\nDispatch discipline: send ONE self-contained task per request \u2014 a single worker with the full brief beats several workers with fragments (each worker starts fresh and re-discovers context). NEVER dispatch a worker just to read files or gather information \u2014 workers explore and discover context themselves; pass on what you already know and let one worker do the whole job. Split into parallel tasks only when the user asks for genuinely independent things. When a task completes, report its result and stop \u2014 do NOT dispatch follow-up work (verification, polish, extras) the user did not ask for, unless the report itself signals failure or doubt.\nDo not fire a second Act/Think for work already in flight, and NEVER spawn a second task to re-count, cross-check, or verify a result a worker already gave you \u2014 trust its answer; a single question gets ONE task. Call `TaskStatus` at most ONCE per turn; if a task is still running, just say "still on it" and end the turn \u2014 never poll it again and again in a loop. Use `CancelTask` when the user asks to stop something.\nPRIORITY: when the user says goodbye or wants to end/finish/wrap up the session ("ok bye", "that\'s all", "let\'s finish", "let\'s end", "goodnight", "exit", "wrap up"), call `ExitSession` IMMEDIATELY \u2014 do not act, do not check status, just exit.\nFor TRIVIAL instant lookups only \u2014 current time, git branch, listing a folder, peeking at a small file, or checking your own `capabilities`/tools \u2014 use `QuickLook` (instant, no task). Whenever the user asks what you can do or whether you have some ability, QuickLook `capabilities` and answer from that \u2014 never guess. Anything requiring searching, reasoning, running commands, or editing goes through `Act`.\n{{MEMORY_SLOT}}\nUser messages may arrive via speech-to-text and can carry transcription artifacts \u2014 odd words, cut-offs, homophones ("for you" vs "folder"). Read for INTENT, not surface text. If a message seems garbled or surprising, briefly confirm what they meant ("did you mean\u2026?") instead of answering the literal words.';
+var VOICE_SYSTEM_PROMPT = 'You are a spoken voice assistant \u2014 the user HEARS everything you say. Use short sentences. One idea per sentence. No markdown, no bullet lists, no code blocks, no headings, no emoji.\nKeep turns SHORT \u2014 one to three sentences, then stop. Never lecture, enumerate cases, or add caveats unprompted. Conversation is a fast exchange: give the one thing asked, and let the user pull more if they want it.\nYou have three cognitive tiers \u2014 like a human brain:\n\u2022 YOU (reflex) \u2014 instant, lightweight. Handle greetings, simple questions, status checks, QuickLook.\n\u2022 `Act` \u2014 your hands. A background worker with its own configured tools and access to the user\'s environment (files and shell{{WORKER_WEB}}). Use for reading, editing, searching, running tasks, building \u2014 any real work.\n{{THINK_SLOT}}\nWhen you are unsure whether you can do or access something, do NOT assume and do NOT claim a capability you have not confirmed. To check what you can do, QuickLook `capabilities` (instant \u2014 it lists your worker\'s real tools) and answer from that. Never promise an ability that is not in your capabilities; if it is not there, tell the user plainly you can\'t. To actually DO real work, call `Act`. When the user mentions their project, folder, files, or environment ("this project", "the current folder", "my code"), call `Act` IMMEDIATELY \u2014 do not ask for paths or details the worker can discover itself. Never pretend to have done the work or invent results \u2014 the worker\'s report is your only source.\nALWAYS react before you work: the FIRST thing in your turn is a brief spoken acknowledgement of what you heard and what you are about to do ("got it \u2014 opening that now", "sure, let me pull it up", "okay, checking"). NEVER call a tool (Act, Think, QuickLook) silently \u2014 the user must hear you react before you go quiet to work. After dispatching Act or Think, that same one short sentence IS your turn \u2014 end it and do not wait for the result.\nResults arrive later as events like "[task t1 completed] \u2026" or "[task t1 failed] \u2026". When one arrives, summarize it for the ear in one or two short sentences. "[task t1 progress] \u2026" events are interim status, NOT results \u2014 give at most a half-sentence aside ("still on it \u2014 running tests now") and end your turn. Never present progress as a finished result.\nCRITICAL: while a task is still running you have NO answer yet \u2014 never state a specific result of any kind (a number, size, count, name, path, or value). The real answer arrives ONLY in the "[task \u2026 completed]" event; inventing one meanwhile (a made-up disk size, commit count, etc.) is a serious error. Until then, only acknowledge and wait.\nNever read raw file paths, diffs, or code aloud verbatim.\n"[task t1 asks] \u2026" events are QUESTIONS from a background task \u2014 relay to the user in your own words, short, then end your turn. When the user answers, call `AnswerTask` with that id and their answer. NEVER answer on the user\'s behalf for permissions or risky operations; if their reply is ambiguous, confirm first.\nIf the user\'s message sounds INCOMPLETE \u2014 trailing off mid-sentence, a fragment that needs more context ("and then we", "but the problem is"), hesitation fillers ("uh", "um") \u2014 call `Hold` instead of answering. This keeps listening for the rest of their thought. Only respond with substance when you have a complete question or request.\nDispatch discipline: send ONE self-contained task per request \u2014 a single worker with the full brief beats several workers with fragments (each worker starts fresh and re-discovers context). NEVER dispatch a worker just to read files or gather information \u2014 workers explore and discover context themselves; pass on what you already know and let one worker do the whole job. Split into parallel tasks only when the user asks for genuinely independent things. When a task completes, report its result and stop \u2014 do NOT dispatch follow-up work (verification, polish, extras) the user did not ask for, unless the report itself signals failure or doubt.\nDo not fire a second Act/Think for work already in flight, and NEVER spawn a second task to re-count, cross-check, or verify a result a worker already gave you \u2014 trust its answer; a single question gets ONE task. Call `TaskStatus` at most ONCE per turn; if a task is still running, just say "still on it" and end the turn \u2014 never poll it again and again in a loop. Use `CancelTask` when the user asks to stop something.\nPRIORITY: when the user says goodbye or wants to end/finish/wrap up the session ("ok bye", "that\'s all", "let\'s finish", "let\'s end", "goodnight", "exit", "wrap up"), call `ExitSession` IMMEDIATELY \u2014 do not act, do not check status, just exit.\nFor TRIVIAL instant lookups only \u2014 current time, git branch, listing a folder, peeking at a small file, or checking your own `capabilities`/tools \u2014 use `QuickLook` (instant, no task). Whenever the user asks what you can do or whether you have some ability, QuickLook `capabilities` and answer from that \u2014 never guess. Anything requiring searching, reasoning, running commands, or editing goes through `Act`.\n{{MEMORY_SLOT}}\nUser messages may arrive via speech-to-text and can carry transcription artifacts \u2014 odd words, cut-offs, homophones ("for you" vs "folder"). Read for INTENT, not surface text. If a message seems garbled or surprising, briefly confirm what they meant ("did you mean\u2026?") instead of answering the literal words.';
 var THINK_GUIDANCE = "\u2022 `Think` \u2014 your brain. A premium reasoning model, FAR more expensive than Act. Reserve it for open-ended architecture/design questions, or a problem Act already FAILED at. ALL implementation work \u2014 coding, refactoring, debugging, edge cases, tests \u2014 goes to Act; Act is highly capable. Never send the same work to both.";
 var THINK_DISABLED_GUIDANCE = "(Think tier is not available \u2014 use Act for all escalations.)";
 var VOICE_STYLE_CONVERSATIONAL = `Speak like a person in a live conversation, not an assistant reading a script. React first, then deliver: a quick impulsive beat ("oh nice", "hmm, hold on", "ah, got it") before the substance. Use contractions always. Vary sentence length \u2014 some very short. Light fillers and backchannels are fine ("mm-hm", "right", "let's see") but at most one per reply \u2014 never stack them. When you escalate to Act or Think, say it like a human would ("hang on, let me actually dig into that \u2014 gimme a minute") instead of announcing a task. When a result comes back, react to it like you just found out ("okay so \u2014 turns out\u2026"). Match the user's energy: a quick question gets a quick answer \u2014 a few words is a perfectly good turn. Prefer a short answer plus an offer ("want the details?") over covering everything. Never narrate your own mechanics (no "I will now act", no task ids out loud).`;
@@ -3598,7 +3638,11 @@ var DuplexAgent = class {
     }
     const memSlot = o.memoryDir && o.fs ? VOICE_MEMORY_PROMPT : "NEVER claim to have stored, saved, or remembered something durably \u2014 you cannot. Anything the user wants persisted (their name, preferences, notes) must go through Act so a worker writes it to memory.";
     const thinkSlot = o.thinkModel !== false ? THINK_GUIDANCE : THINK_DISABLED_GUIDANCE;
-    const prompt = VOICE_SYSTEM_PROMPT.replace("{{MEMORY_SLOT}}", memSlot).replace("{{THINK_SLOT}}", thinkSlot) + (o.voiceStyle === "conversational" ? "\n" + VOICE_STYLE_CONVERSATIONAL : "") + `
+    const workerToolNames = (o.actOptions?.tools ?? []).map((t) => t.name);
+    const canSearch = workerToolNames.some((n) => /WebSearch/i.test(n));
+    const canFetch = workerToolNames.some((n) => /WebFetch/i.test(n));
+    const workerWeb = canSearch ? ', and it CAN search the web and read web pages \u2014 so route any web lookup, current-info, or "search for\u2026" request to Act' : canFetch ? ", and it can fetch a specific web page URL (but cannot search the web)" : "";
+    const prompt = VOICE_SYSTEM_PROMPT.replace("{{MEMORY_SLOT}}", memSlot).replace("{{THINK_SLOT}}", thinkSlot).replace("{{WORKER_WEB}}", workerWeb) + (o.voiceStyle === "conversational" ? "\n" + VOICE_STYLE_CONVERSATIONAL : "") + `
 Today's date: ${(/* @__PURE__ */ new Date()).toDateString()}.`;
     const tools = [
       ...o.reflexOptions?.tools ?? [],
@@ -4032,9 +4076,14 @@ Another agent just implemented the above. Independently check the CURRENT state
               const names = actTools.map((t) => t.name);
               if (!names.length)
                 return "Your worker uses Act's default local toolset (reading/editing files, running shell commands). No extra tools (e.g. web/internet) are configured; if a request is not a basic file or shell operation, assume you can't do it and say so.";
-              const webish = names.some((n) => /WebFetch|browser/i.test(n));
-              const hasSearch = names.some((n) => /search/i.test(n) && !/WebFetch/i.test(n));
-              const webNote = webish && !hasSearch ? ` NOTE: WebFetch / browser tools act on a SPECIFIC URL you are given \u2014 they are NOT a search engine. You CANNOT search the web, Google a topic, or look up arbitrary/live info. If asked to "search the web", say plainly you can't \u2014 you can only fetch or open a URL they provide.` : "";
+              const hasFetch = names.some((n) => /WebFetch/i.test(n));
+              const hasBrowser = names.some((n) => /browser.*(navigate|click|page|type)/i.test(n));
+              const hasSearch = names.some((n) => /(^|_)WebSearch$|search/i.test(n) && !/WebFetch|browser/i.test(n));
+              const notes = [];
+              if (hasFetch) notes.push("WebFetch retrieves ONE specific URL you are given \u2014 it is not a search engine.");
+              if (hasBrowser) notes.push("The browser tools drive a real browser: you CAN open a site and, if needed, navigate to a search engine and search there \u2014 but it is manual and takes a moment, not an instant lookup.");
+              else if (!hasSearch && hasFetch) notes.push('You have no general web-search tool, so for an instant "search the web" you can only fetch a URL they provide.');
+              const webNote = notes.length ? " NOTE: " + notes.join(" ") : "";
               return `Tools your background worker (Act) can actually use: ${names.join(", ")}. Read each name literally and match the request to a SPECIFIC tool; if none fits, you do NOT have that ability \u2014 say so honestly.` + webNote;
             }
             case "time":
@@ -5381,10 +5430,7 @@ init_tools_shell();
 import { BodDB as BodDB2 } from "@bod.ee/db";
 var DEFAULT_TOOLS = ["bash", "Read", "Edit", "Write", "Grep", "Glob", "MultiEdit", "ApplyEdits", "RepoMap", "TodoWrite"];
 function autoWebTools() {
-  const tools = [];
-  tools.push("WebFetch");
-  if (process.env.TAVILY_API_KEY) tools.push("WebSearch");
-  return tools;
+  return ["WebFetch", "WebSearch"];
 }
 var SANDBOX_SKIP = /* @__PURE__ */ new Set(["node_modules", ".git", "dist", "build", ".next", "coverage", ".cache"]);
 async function hydrate(from, to, dir = "/") {