rovecode 0.3.3 → 0.4.0-beta.3

package/src/tools/websearch.ts

@@ -1,359 +0,0 @@
-/** web_search tool (PORT #56): query → up to N ranked web results as compact text.
- *  Ported from opencode (MIT, snapshot ebece6e, packages/opencode/src/tool/websearch.ts +
- *  mcp-websearch.ts). Backend = Exa's HOSTED MCP endpoint https://mcp.exa.ai/mcp, called KEYLESS by
- *  default with one JSON-RPC `tools/call` of `web_search_exa` (mcp-websearch.ts:66-96 `call`: POST, Accept
- *  json + event-stream, {jsonrpc:"2.0", id:1, method:"tools/call", params:{name, arguments}});
- *  arguments {query, type:"auto", numResults (default 8), livecrawl:"fallback", contextMaxCharacters
- *  (optional — omitted unless the model passes one)} (websearch.ts:83-96); 25s timeout (:95); the
- *  reader takes the first `result.content[].text` from
- *  a JSON body or an SSE `data:` line (mcp-websearch.ts:20-42 parsePayload/parseResponse). No key
- *  is needed; opencode appends an optional EXA_API_KEY as `?exaApiKey=` (mcp-websearch.ts:4-6) —
- *  mirrored by the optional `apiKey` dep, which the ONE registration line in cli/runtime.ts reads
- *  from the environment; this module reads no env at all, every URL it echoes into an output is
- *  shown without its query string, and error text that arrives through a seam (a rejecting fetch or
- *  resolver, a body stream, a non-2xx body, a JSON-RPC error) is scrubbed — the keyed URL → its
- *  display form, the key itself → `[key]` — so a key never leaks into a result the model (or a log) sees.
- *  Deviations: opencode hands the model Exa's context string verbatim; here it is PARSED into ranked
- *  results (Title/URL/Published/Author/Highlights blocks separated by `---`, or the raw {results:[…]}
- *  JSON older servers return) and re-rendered as `N. title · url · date` + a bounded snippet, with
- *  result-count, snippet, title/url and total-output caps; the Parallel provider, the per-session
- *  provider coin flip and the livecrawl/type/contextMaxCharacters args are dropped: livecrawl/type are
- *  fixed to opencode's defaults ("fallback"/"auto"), while contextMaxCharacters is ALWAYS sent as
- *  max(10000, n*1500) — opencode omits the key unless the model passes one, so Exa applies its own
- *  10000 default there; rovecode's value is pinned in websearch.test.ts (12000 at n=8, 30000 at n=20).
- *  Every request — the first and every same-host redirect hop (re-POSTed, ≤5) — passes the SAME SSRF
- *  guard as web_fetch (webfetch.ts ssrfDenyReason: private/loopback/link-local/reserved refused,
- *  names resolved through the injectable resolver, raced against the timeout) with NO allow-private
- *  door: the endpoint is fixed, so a private answer for it is a broken resolver or an attack. A
- *  cross-host redirect is refused. Non-2xx, timeout, abort, malformed body and backend errors are
- *  tool error results — the tool never throws and never writes files. Text in neither known result
- *  shape is passed through verbatim, bounded and flagged (format drift is not a dead tool).
- *  Policy: kind "network" → action net.fetch; the schema declares no url/path/command, so the
- *  resource is the tool name (core/tools.ts describeResource): `allow net.fetch web_search`
- *  auto-runs it and the runtime's default `prompt net.fetch *` makes it PROMPT by default, exactly
- *  like web_fetch. */
-
-import type { Tool, ToolContext, ToolOutput } from "../core/types.ts";
-import { abortable, canonicalHost, dnsResolver, readBounded, ssrfDenyReason, type FetchLike, type Resolver } from "./webfetch.ts";
-import pkg from "../../package.json";
-
-// ---------- bounds (advertised in the schema; args are clamped, never trusted) ----------
-
-export const ENDPOINT = "https://mcp.exa.ai/mcp";
-export const RESULTS_DEFAULT = 8;          // opencode numResults default
-export const RESULTS_CAP = 20;
-export const QUERY_CHARS = 1000;           // longer queries are refused, not sent
-export const SNIPPET_CHARS = 500;          // per result
-export const TITLE_CHARS = 200;
-export const URL_CHARS = 2000;
-export const OUTPUT_CHARS_CAP = 16_000;    // whole output; trailing results are dropped whole to fit
-export const MAX_BYTES = 1024 * 1024;      // response body read cap
-export const MAX_REDIRECTS = 5;
-export const TIMEOUT_DEFAULT_MS = 25_000;  // opencode "25 seconds"
-const USER_AGENT = `Mozilla/5.0 (compatible; rovecode/${pkg.version} web_search)`; // port #48: was aion/0.2
-const ACCEPT = "application/json, text/event-stream";
-const REDIRECT_STATUSES = new Set([301, 302, 303, 307, 308]);
-const MARKER_RESERVE = 96;                 // room kept for the truncation marker under the cap
-
-/** files.ts clampLimit contract: ceiling RESULTS_CAP; absent/NaN/non-positive → default; fractions
- *  floor; numeric strings (non-native tool-call parsing) count. */
-export function clampResults(v: unknown): number {
-  const n = typeof v === "string" && v.trim() !== "" ? Number(v) : v;
-  return typeof n === "number" && Number.isFinite(n) && n > 0 ? Math.min(Math.floor(n), RESULTS_CAP) : RESULTS_DEFAULT;
-}
-
-/** Surrogate-safe clip with an ellipsis. */
-function clip(s: string, max: number): string {
-  return s.length <= max ? s : s.slice(0, max - 1).replace(/[\uD800-\uDBFF]$/, "") + "…";
-}
-
-// ---------- MCP envelope ----------
-
-export type McpPayload = { text: string } | { error: string };
-
-/** One JSON-RPC envelope → its first text content, or the error it carries; null when it is
- *  neither (opencode parsePayload: `result.content.find(item => item.text)`). */
-function envelopePayload(raw: string): McpPayload | null {
-  let j: unknown;
-  try { j = JSON.parse(raw); } catch { return null; }
-  if (!j || typeof j !== "object") return null;
-  const env = j as { error?: unknown; result?: unknown };
-  if (env.error && typeof env.error === "object") {
-    const m = (env.error as { message?: unknown }).message;
-    return { error: typeof m === "string" && m !== "" ? m : JSON.stringify(env.error) };
-  }
-  if (!env.result || typeof env.result !== "object") return null;
-  const r = env.result as { content?: unknown; isError?: unknown };
-  const items: unknown[] = Array.isArray(r.content) ? r.content : [];
-  const text = items
-    .map((c) => (c && typeof c === "object" ? (c as { text?: unknown }).text : undefined))
-    .find((t): t is string => typeof t === "string" && t !== "");
-  if (r.isError === true) return { error: text ?? "backend reported an error without a message" };
-  return text === undefined ? null : { text };
-}
-
-/** Body → payload: a JSON object body, else the first SSE `data:` line that carries one (opencode
- *  parseResponse). null = no envelope with text or error anywhere (malformed). */
-export function parseMcpBody(body: string): McpPayload | null {
-  const trimmed = body.trim();
-  if (trimmed.startsWith("{")) { const p = envelopePayload(trimmed); if (p) return p; }
-  for (const line of body.split("\n")) {
-    if (!line.startsWith("data:")) continue;
-    const p = envelopePayload(line.slice(5).trim());
-    if (p) return p;
-  }
-  return null;
-}
-
-// ---------- result text ----------
-
-export interface SearchResult { title: string; url: string; snippet: string; publishedDate?: string }
-
-/** Exa highlight text → one line: `...` gap-marker lines become ` … `, whitespace collapses. */
-function cleanSnippet(raw: string): string {
-  const out: string[] = [];
-  for (const line of raw.split(/\r?\n/)) {
-    const l = line.trim();
-    if (l === "") continue;
-    if (l === "...") { if (out.length > 0 && out[out.length - 1] !== "…") out.push("…"); continue; }
-    out.push(l);
-  }
-  while (out.length > 0 && out[out.length - 1] === "…") out.pop();
-  return out.join(" ").replace(/\s+/g, " ").trim();
-}
-
-const KEY_LINE = /^(Title|URL|Published|Author|Score|ID|Image|Favicon):\s*(.*)$/;
-const BODY_MARK = /^(Highlights|Text|Summary):\s*$/;
-
-/** Exa's LLM context string — `Title: / URL: / Published: / Author: / Highlights:` blocks separated
- *  by `---` — → results in rank order. Blocks without an http(s) URL are skipped; `N/A` dates are
- *  dropped; the title falls back to the url. */
-export function parseExaContext(text: string): SearchResult[] {
-  const results: SearchResult[] = [];
-  for (const block of text.split(/\r?\n\s*---\s*\r?\n/)) {
-    const lines = block.split(/\r?\n/);
-    const fields: Record<string, string> = {};
-    let i = 0;
-    for (; i < lines.length; i++) {
-      const line = lines[i]!.trim();
-      if (line === "") { if (Object.keys(fields).length > 0) break; continue; }
-      const m = KEY_LINE.exec(line);
-      if (!m) break;
-      fields[m[1]!] = m[2]!.trim();
-    }
-    while (i < lines.length && lines[i]!.trim() === "") i++;
-    if (i < lines.length && BODY_MARK.test(lines[i]!.trim())) i++;
-    const url = fields["URL"] ?? "";
-    if (!/^https?:\/\//i.test(url)) continue;
-    const published = fields["Published"];
-    results.push({
-      title: fields["Title"] || url,
-      url,
-      snippet: cleanSnippet(lines.slice(i).join("\n")),
-      ...(published && published.toUpperCase() !== "N/A" ? { publishedDate: published } : {}),
-    });
-  }
-  return results;
-}
-
-/** Raw Exa search JSON ({results:[{title,url,publishedDate,text|highlights|summary}]}) → results;
- *  null when the text is not that shape. */
-function parseExaJson(text: string): SearchResult[] | null {
-  let j: unknown;
-  try { j = JSON.parse(text); } catch { return null; }
-  const list = j && typeof j === "object" ? (j as { results?: unknown }).results : undefined;
-  if (!Array.isArray(list)) return null;
-  const out: SearchResult[] = [];
-  for (const r of list as unknown[]) {
-    if (!r || typeof r !== "object") continue;
-    const o = r as Record<string, unknown>;
-    const url = typeof o["url"] === "string" ? o["url"] : "";
-    if (!/^https?:\/\//i.test(url)) continue;
-    const highlights = Array.isArray(o["highlights"]) ? (o["highlights"] as unknown[]).filter((h): h is string => typeof h === "string") : [];
-    const body = typeof o["text"] === "string" ? o["text"] : highlights.length > 0 ? highlights.join("\n...\n") : typeof o["summary"] === "string" ? o["summary"] : "";
-    const published = typeof o["publishedDate"] === "string" && o["publishedDate"] !== "" ? o["publishedDate"] : undefined;
-    out.push({ title: typeof o["title"] === "string" && o["title"] !== "" ? o["title"] : url, url, snippet: cleanSnippet(body), ...(published ? { publishedDate: published } : {}) });
-  }
-  return out;
-}
-
-/** Backend text → results ([] = a recognized no-result answer); null when the text is in neither
- *  known shape (the tool then renders it verbatim, bounded). */
-export function parseResultsText(text: string): SearchResult[] | null {
-  const t = text.trim();
-  if (t === "" || /^No search results found/i.test(t)) return [];
-  if (t.startsWith("{")) return parseExaJson(t);
-  const ctx = parseExaContext(t);
-  return ctx.length > 0 ? ctx : null;
-}
-
-export interface Rendered { output: string; shown: number; truncated: boolean }
-
-/** Header line, then `N. title · url[ · date]` + an indented snippet per result, blank-line
- *  separated. Whole trailing results are dropped (never split) so the output stays under `cap`. */
-export function formatResults(query: string, results: SearchResult[], cap = OUTPUT_CHARS_CAP): Rendered {
-  const header = `${results.length} result${results.length === 1 ? "" : "s"} for "${clip(query, TITLE_CHARS)}" (Exa web search)`;
-  const entries = results.map((r, i) => {
-    const head = `${i + 1}. ${clip(r.title.replace(/\s+/g, " ").trim(), TITLE_CHARS)} · ${clip(r.url, URL_CHARS)}${r.publishedDate ? ` · ${r.publishedDate}` : ""}`;
-    const snippet = clip(r.snippet, SNIPPET_CHARS);
-    return snippet === "" ? head : `${head}\n   ${snippet}`;
-  });
-  let output = header;
-  let shown = 0;
-  for (const e of entries) {
-    if (output.length + 2 + e.length + MARKER_RESERVE > cap) break;
-    output += `\n\n${e}`;
-    shown++;
-  }
-  const truncated = shown < entries.length;
-  if (truncated) output += `\n\n(Showing ${shown} of ${results.length} results: output capped at ${cap} characters.)`;
-  return { output, shown, truncated };
-}
-
-// ---------- the tool ----------
-
-/** A URL as outputs show it: origin + path only — the query may carry the api key. */
-function displayUrl(u: URL): string { return u.origin + u.pathname; }
-
-export interface WebSearchDeps {
-  fetch?: FetchLike; resolve?: Resolver; timeoutMs?: number;
-  /** optional Exa key, sent as opencode does (`?exaApiKey=`); blank/undefined = keyless */
-  apiKey?: string;
-}
-
-/** Builds the tool; `deps` are test seams (fixture-mapping fetch, scripted resolver, short timeout)
- *  plus the optional apiKey the runtime registration injects. The module-level webSearchTool is the
- *  keyless instance (listing surfaces). */
-export function createWebSearchTool(deps: WebSearchDeps = {}): Tool {
-  const fetchImpl: FetchLike = deps.fetch ?? ((u, i) => fetch(u, i));
-  const resolve: Resolver = deps.resolve ?? dnsResolver;
-  const limitMs = typeof deps.timeoutMs === "number" && Number.isFinite(deps.timeoutMs) && deps.timeoutMs > 0 ? Math.floor(deps.timeoutMs) : TIMEOUT_DEFAULT_MS;
-  const key = typeof deps.apiKey === "string" ? deps.apiKey.trim() : "";
-  const endpoint = key === "" ? ENDPOINT : `${ENDPOINT}?exaApiKey=${encodeURIComponent(key)}`;
-  // Error text that arrives through a seam may embed the request URL or the key: a wrapped/injected
-  // fetch or proxy that puts the URL in its message (Bun's own errors keep it in err.path today), a
-  // resolver, a body stream, a non-2xx body, a JSON-RPC error. Scrubbed before it reaches an output:
-  // the keyed endpoint and the CURRENT request URL → their display form (origin+path), then the key
-  // itself (raw and URL-encoded) → `[key]`. Keyless, this rewrites nothing.
-  const scrub = (text: string, u: URL): string => {
-    let out = text.split(endpoint).join(ENDPOINT).split(u.href).join(displayUrl(u));
-    if (key !== "") for (const k of new Set([key, encodeURIComponent(key)])) out = out.split(k).join("[key]");
-    return out;
-  };
-
-  async function run(a: { query?: unknown; max_results?: unknown }, ctx: ToolContext): Promise<ToolOutput> {
-    const query = typeof a.query === "string" ? a.query.trim() : "";
-    if (query === "") return { ok: false, output: "web_search: query is required" };
-    if (query.length > QUERY_CHARS) return { ok: false, output: `web_search: query too long (${query.length} characters; max ${QUERY_CHARS})` };
-    const n = clampResults(a.max_results);
-    if (ctx.signal.aborted) return { ok: false, output: "web_search: aborted" };
-    const rpc = JSON.stringify({
-      jsonrpc: "2.0", id: 1, method: "tools/call",
-      params: { name: "web_search_exa", arguments: { query, type: "auto", numResults: n, livecrawl: "fallback", contextMaxCharacters: Math.max(10_000, n * 1_500) } },
-    });
-
-    // Timeout: a REF'D setTimeout drives the controller (Bun's AbortSignal.timeout timer is unref'd
-    // and never fires on an idle loop); the caller's signal is chained so Esc aborts the socket too.
-    const ac = new AbortController();
-    let timedOut = false;
-    let current = new URL(endpoint); // the URL in flight: the endpoint, then each redirect target
-    const timer = setTimeout(() => { timedOut = true; ac.abort(); }, limitMs);
-    (timer as unknown as { ref?: () => void }).ref?.();
-    const onAbort = (): void => ac.abort();
-    ctx.signal.addEventListener("abort", onAbort, { once: true });
-    // a rejection from the resolver, the fetch or the body read; its message is scrubbed against the URL in flight
-    const fail = (e: unknown): ToolOutput => {
-      if (ac.signal.aborted) return { ok: false, output: timedOut ? `web_search: timed out after ${limitMs}ms` : "web_search: aborted" };
-      return { ok: false, output: `web_search: request failed: ${scrub(e instanceof Error ? e.message : String(e), current)}` };
-    };
-
-    try {
-      const origin = canonicalHost(current.hostname);
-      let hops = 0;
-      for (;;) {
-        // scheme + SSRF gate on the endpoint AND every redirect target; no allow-private door
-        if (current.protocol !== "http:" && current.protocol !== "https:") {
-          return { ok: false, output: `web_search: backend redirected to unsupported URL scheme ${current.protocol} (http/https only)` };
-        }
-        let reason: string | null;
-        try { reason = await abortable(ssrfDenyReason(current.hostname, resolve), ac.signal); } catch (e) { return fail(e); }
-        if (reason !== null) return { ok: false, output: `web_search: refused ${displayUrl(current)}: ${scrub(reason, current)}` };
-        // the policy decision covered the fixed backend; another host never gets the request
-        if (canonicalHost(current.hostname) !== origin) {
-          return { ok: false, output: `web_search: backend redirected to ${displayUrl(current)} (${canonicalHost(current.hostname)} is not ${origin}); not followed` };
-        }
-        let res: Response;
-        try {
-          res = await fetchImpl(current.href, {
-            method: "POST", redirect: "manual", signal: ac.signal, body: rpc,
-            headers: { "User-Agent": USER_AGENT, Accept: ACCEPT, "Content-Type": "application/json" },
-          });
-        } catch (e) { return fail(e); }
-
-        if (REDIRECT_STATUSES.has(res.status)) {
-          const location = res.headers.get("location");
-          await res.body?.cancel().catch(() => {});
-          if (!location) return { ok: false, output: `web_search: HTTP ${res.status} redirect from ${displayUrl(current)} without a Location header` };
-          if (++hops > MAX_REDIRECTS) return { ok: false, output: `web_search: too many redirects (more than ${MAX_REDIRECTS}) from ${ENDPOINT}` };
-          try { current = new URL(location, current); } catch { return { ok: false, output: `web_search: invalid redirect target ${location}` }; }
-          continue;
-        }
-
-        let body: { bytes: Uint8Array; truncated: boolean };
-        try { body = await readBounded(res, res.ok ? MAX_BYTES : 2048); } catch (e) { return fail(e); }
-        const text = new TextDecoder().decode(body.bytes);
-        if (!res.ok) {
-          const detail = text.split(/\r?\n/).map((l) => l.trim()).find((l) => l !== "") ?? "";
-          return { ok: false, output: `web_search: backend HTTP ${res.status}${detail === "" ? "" : `: ${clip(scrub(detail, current), 200)}`}` };
-        }
-        const payload = parseMcpBody(text);
-        if (payload === null) {
-          return { ok: false, output: `web_search: malformed backend response (${body.truncated ? `body truncated at ${MAX_BYTES} bytes` : `${body.bytes.length} bytes, no JSON-RPC result with text`})` };
-        }
-        if ("error" in payload) return { ok: false, output: `web_search: backend error: ${clip(scrub(payload.error, current).replace(/\s+/g, " ").trim(), 300)}` };
-        const parsed = parseResultsText(payload.text);
-        if (parsed === null) {
-          const raw = payload.text.trim();
-          const shown = clip(raw, OUTPUT_CHARS_CAP);
-          return {
-            ok: true,
-            output: `web_search "${clip(query, TITLE_CHARS)}": the backend answered in an unrecognized format; shown verbatim${shown.length < raw.length ? ` (capped at ${OUTPUT_CHARS_CAP} characters)` : ""}:\n\n${shown}`,
-            data: { query, count: 0, total: 0, unstructured: true, truncated: shown.length < raw.length },
-          };
-        }
-        if (parsed.length === 0) {
-          return { ok: true, output: `No results for "${clip(query, TITLE_CHARS)}" (Exa web search). Try different or fewer terms.`, data: { query, count: 0, total: 0, results: [], truncated: false } };
-        }
-        const r = formatResults(query, parsed.slice(0, n));
-        return {
-          ok: true,
-          output: r.output,
-          data: { query, count: r.shown, total: parsed.length, results: parsed.slice(0, r.shown), truncated: r.shown < parsed.length },
-        };
-      }
-    } finally {
-      clearTimeout(timer);
-      ctx.signal.removeEventListener("abort", onAbort);
-    }
-  }
-
-  return {
-    schema: {
-      name: "web_search",
-      description: `Search the web and return up to max_results ranked results (default ${RESULTS_DEFAULT}, cap ${RESULTS_CAP}), each as "N. title · url · published date (when known)" followed by a snippet of at most ${SNIPPET_CHARS} characters; the whole output is capped at ${OUTPUT_CHARS_CAP} characters (trailing results dropped, with a marker). Results are search snippets, not pages: call web_fetch on a result's url to read it. Backend: Exa via its hosted MCP endpoint (${ENDPOINT}); no API key is required. Times out after ${TIMEOUT_DEFAULT_MS / 1000}s; a backend that redirects elsewhere or resolves to a private address is refused. The current year is ${new Date().getFullYear()}: put it in queries about recent events.`,
-      args: {
-        type: "object",
-        properties: {
-          query: { type: "string", description: `search query in plain words (at most ${QUERY_CHARS} characters)` },
-          max_results: { type: "integer", description: `results to return (default ${RESULTS_DEFAULT}, cap ${RESULTS_CAP})` },
-        },
-        required: ["query"],
-      },
-    },
-    kind: "network",
-    sequential: false,
-    execute: (args, ctx) => run((args ?? {}) as { query?: unknown; max_results?: unknown }, ctx),
-  };
-}
-
-export const webSearchTool: Tool = createWebSearchTool();

package/src/tui/agents-cmd.ts

@@ -1,41 +0,0 @@
-/** `/agents list` (port #62): the custom subagent definitions the runtime loaded at boot (core/agents.ts —
- *  `.rovecode/agents/<name>.md`, project shadows `<home>/agents`), one row each with its SOURCE (scope + path),
- *  model, mode and tool allow-list, after the built-in `main` row; the files skipped at boot follow under
- *  "skipped:". Read-only over the runtime snapshot — definitions are loaded once per runtime (restart rovecode
- *  to pick up edits). In the sextant surface the bare `/agents` is the crew board (its renderer-local command,
- *  local-commands.ts); only `/agents list` reaches handleSlash here — and the pass-through is sextant/keys.ts's
- *  `if (arg) return false`, already landed. */
-
-import type { CustomAgent, DiscoveredAgents } from "../core/agents.ts";
-import { AGENTS_DIR, PROJECT_STATE_DIR } from "../core/agents.ts";
-import type { Renderer } from "./renderer.ts";
-
-export interface AgentsCmdCtx {
-  renderer: Renderer;
-  /** the runtime's boot snapshot (Runtime.agents) */
-  agents: DiscoveredAgents;
-}
-
-/** the built-in definition every runtime has — `task start` without `agent` runs it */
-export const MAIN_AGENT_ROW = "main — built-in: the starting run's model and system prompt, every child tool (default for `task start`)";
-
-/** `name — description (scope: path; model …; mode …; tools …)` */
-export function formatAgentRow(a: CustomAgent): string {
-  const tools = a.tools.includes("*") ? "all child tools" : a.tools.join(", ");
-  return `${a.name} — ${a.description} (${a.scope}: ${a.path}; model ${a.model ?? "inherited"}; mode ${a.mode ?? "act"}; tools ${tools})`;
-}
-
-/** The whole note: main + custom rows, a hint when there are none, then the skipped files. */
-export function formatAgentList(d: DiscoveredAgents): string {
-  const rows = [MAIN_AGENT_ROW, ...d.agents.map(formatAgentRow)];
-  const hint = d.agents.length === 0
-    ? `\n(no custom definitions — add ${PROJECT_STATE_DIR}/${AGENTS_DIR}/<name>.md: frontmatter description: / model: / mode: plan|act / tools: <allow-list>, body = system prompt)`
-    : "";
-  const skipped = d.warnings.length > 0 ? "\nskipped:\n" + d.warnings.map((w) => `  ${w}`).join("\n") : "";
-  return rows.join("\n") + hint + skipped;
-}
-
-/** `/agents [list]` — the argument is accepted (the sextant's pass-through spelling) and ignored. */
-export function cmdAgents(ctx: AgentsCmdCtx): void {
-  ctx.renderer.addSystemNote(formatAgentList(ctx.agents), ctx.agents.warnings.length > 0 ? "warn" : "info");
-}