talon-agent 5.18.1 → 5.19.0

package/src/core/doctor/logs.ts

@@ -0,0 +1,124 @@
+/**
+ * Doctor's "what went wrong lately" probes: the last hour of talon.log
+ * summarised by component, and the operator alerts still open — so
+ * `talon doctor` points at the failing subsystem instead of only at the
+ * environment.
+ *
+ * Alerts come from the daemon's in-memory table when doctor runs inside
+ * the daemon (the chat /doctor command). `talon doctor` is its own
+ * process and sees an empty table, so it replays the alert lines in the
+ * log instead; those are reported as warnings without counting toward
+ * the issue total — the log can only say what was last announced.
+ */
+
+import {
+  readLogRecords,
+  replayAlerts,
+  type LogRecord,
+} from "../daemon/log-reader.js";
+import { activeAlerts } from "../frontend-runtime/alerts.js";
+import { files } from "../../util/paths.js";
+import type { DoctorCheck } from "./types.js";
+
+const HOUR_MS = 3_600_000;
+/** How far back the alert replay looks for a raise still unresolved. */
+const ALERT_WINDOW_MS = 24 * HOUR_MS;
+const READ = { limit: Number.MAX_SAFE_INTEGER, maxBytesPerFile: 16 << 20 };
+const DETAIL_CHARS = 160;
+
+export type RecentLogOptions = {
+  logPath?: string;
+  now?: number;
+  /** The live alert table; defaults to this process's `activeAlerts()`. */
+  liveAlerts?: ReturnType<typeof activeAlerts>;
+};
+
+function clip(text: string): string {
+  return text.length > DETAIL_CHARS ? `${text.slice(0, DETAIL_CHARS)}…` : text;
+}
+
+function byComponent(records: readonly LogRecord[]): string {
+  const counts = new Map<string, number>();
+  for (const r of records) {
+    const c = r.component ?? "?";
+    counts.set(c, (counts.get(c) ?? 0) + 1);
+  }
+  return [...counts]
+    .sort((a, b) => b[1] - a[1])
+    .map(([c, n]) => `${c} ${n}`)
+    .join(", ");
+}
+
+/** The last hour's warn/error lines, as one doctor check. */
+export function checkRecentErrors(opts: RecentLogOptions = {}): DoctorCheck {
+  const logPath = opts.logPath ?? files.log;
+  const now = opts.now ?? Date.now();
+  const recent = readLogRecords(logPath, {
+    ...READ,
+    filter: { since: now - HOUR_MS, minLevel: "warn" },
+  });
+  const errors = recent.filter(
+    (r) => r.level === "error" || r.level === "fatal",
+  );
+  const warnings = recent.length - errors.length;
+  if (errors.length === 0) {
+    return warnings === 0
+      ? { label: "Log: no errors or warnings in the last hour", status: "ok" }
+      : {
+          label: `Log: ${warnings} warning(s) in the last hour`,
+          status: "info",
+          detail: byComponent(recent),
+        };
+  }
+  const last = errors.at(-1)!;
+  const at = new Date(last.ts).toTimeString().slice(0, 5);
+  const what = clip(`${last.msg}${last.err ? ` (${last.err})` : ""}`);
+  return {
+    label: `Log: ${errors.length} error(s) in the last hour — ${byComponent(errors)}`,
+    status: "warn",
+    detail:
+      `latest ${at} ${last.component ?? "?"}: ${what} — ` +
+      `talon logs --errors --since 1h, or ~/.talon/ns/proc/errors`,
+  };
+}
+
+/** Open alerts, one check each (or one "none" line). */
+export function checkOpenAlerts(opts: RecentLogOptions = {}): DoctorCheck[] {
+  const now = opts.now ?? Date.now();
+  const live = opts.liveAlerts ?? activeAlerts();
+  const fromLog = live.length === 0;
+  const alerts = fromLog ? loggedAlerts(opts.logPath ?? files.log, now) : live;
+  if (alerts.length === 0)
+    return [{ label: "Alerts: none open", status: "ok" }];
+  return alerts.map((a) => {
+    const mins = Math.max(0, Math.round((now - a.since) / 60_000));
+    const label = `Alert ${a.key} (${a.severity}, ${mins} min)`;
+    if (fromLog) {
+      return {
+        label,
+        status: "warn",
+        detail: `last logged: ${clip(a.message)}`,
+      };
+    }
+    return {
+      label,
+      status: a.severity === "warn" ? "warn" : "fail",
+      detail: clip(a.message),
+      issue: true,
+    };
+  });
+}
+
+/** Replay the last day's alert lines (plus boot lines, which reset them). */
+function loggedAlerts(logPath: string, now: number) {
+  const since = now - ALERT_WINDOW_MS;
+  const alertLines = readLogRecords(logPath, {
+    ...READ,
+    filter: { since, component: "alert" },
+  });
+  const boots = readLogRecords(logPath, {
+    ...READ,
+    filter: { since, component: "bot", grep: "Starting Talon" },
+  });
+  return replayAlerts([...alertLines, ...boots].sort((a, b) => a.ts - b.ts));
+}

package/src/core/doctor/types.ts

@@ -1,7 +1,7 @@
 /**
  * The doctor's data shapes, on their own so a backend factory can declare
  * its checks (`BackendFactory.doctor`) without importing the collector —
- * `core/doctor.ts` reads the registry, so the types living there would
+ * `core/doctor/index.ts` reads the registry, so the types living there would
  * close a cycle. Checks are pure data (label / status / detail); renderers
  * decide presentation.
  */

package/src/core/engine/backend-controller/index.ts

@@ -10,7 +10,7 @@
  * Different roles, and different chats, have different cost/latency/
  * quality needs. Typical post-Anthropic-metering setup: chat default
  * on free-tier OpenAI Agents, heartbeats on Claude Sonnet, dream
- * shared with chat — plus Pandario stays on Claude while DMs use the
+ * shared with chat — plus one group chat stays on Claude while DMs use the
  * cheap default. A single-active model can't express that. The pool
  * lets each holder bind independently while deduplicating instances
  * when ids overlap.
@@ -24,7 +24,6 @@
  *   - `pool`     — init/teardown, role accessors, availability, snapshot,
  *                  listeners.
  *   - `rebind`   — rebind/release holders + per-chat accessors.
- *   - `legacy`   — single-active aliases routed to the chat role.
  */
 
 export * from "./types.js";
@@ -57,14 +56,3 @@ export {
   hasChatBackendOverride,
   resolveChatBackend,
 } from "./rebind.js";
-export {
-  initBackendController,
-  getActiveBackend,
-  hasActiveBackend,
-  getActiveBackendOrNull,
-  getActiveBackendId,
-  getActiveBackendLabel,
-  switchBackend,
-  cleanupBackendController,
-  resetBackendControllerForTest,
-} from "./legacy.js";

package/src/core/engine/backend-router/router.ts

@@ -2,7 +2,7 @@
  * Plan-aware backend routing for background work.
  *
  * Background work — `spawn_agent` sub-agents, cron `query` jobs, the
- * heartbeat — used to inherit whichever backend the chat happened to be on.
+ * heartbeat — would otherwise inherit whichever backend the chat is on.
  * On a multi-subscription install that is a good way to burn one plan to its
  * ceiling while another sits idle. When nothing is pinned, this picks the
  * backend with the most headroom instead.

package/src/core/engine/dispatcher.ts

@@ -12,8 +12,15 @@ import type { ExecuteParams, ExecuteResult } from "../types.js";
 import { formatRelayBlock, takePendingRelay } from "./cross-chat-relay.js";
 import { recordMessageSignal } from "../memory/taps.js";
 import { log } from "../../util/log.js";
+import { setStuckLoopListener } from "../../util/watchdog.js";
 import { taskTable, type KillOutcome } from "../tasks/index.js";
 import { initWeaver, type Weaver, type WeaverDeps } from "../weaver/index.js";
+import {
+  noteTurnFailed,
+  noteTurnSucceeded,
+  stuckLoopAlert,
+  type TurnBinding,
+} from "./turn-health.js";
 
 // ── Dependencies (injected at startup) ──────────────────────────────────────
 
@@ -21,7 +28,7 @@ import { initWeaver, type Weaver, type WeaverDeps } from "../weaver/index.js";
  * `getBackend` takes the string chat id so it can route per-chat —
  * a chat with a backend override returns its override backend, others
  * fall through to the global chat-role backend. Tests can pass a
- * stub that ignores the chat id. See `core/backend-controller.ts`.
+ * stub that ignores the chat id. See `core/engine/backend-controller/`.
  *
  * `resolveActiveModel` walks the 5-step active-model resolution
  * chain for the chat and returns both the resolved `ModelRef` and
@@ -33,9 +40,18 @@ import { initWeaver, type Weaver, type WeaverDeps } from "../weaver/index.js";
 type DispatcherDeps = WeaverDeps;
 
 let weaver: Weaver | null = null;
+let backendIdFor: (chatId: string) => string = () => "unknown";
 
 export function initDispatcher(d: DispatcherDeps): void {
   weaver = initWeaver(d);
+  backendIdFor = (chatId) => {
+    try {
+      return d.getBackend(chatId).id;
+    } catch {
+      return "unknown";
+    }
+  };
+  setStuckLoopListener(stuckLoopAlert);
   log("dispatcher", "Initialized (per-chat serial, cross-chat parallel)");
 }
 
@@ -94,5 +110,42 @@ export async function execute(params: ExecuteParams): Promise<ExecuteResult> {
   const prompt = relayed.length
     ? formatRelayBlock(relayed) + params.prompt
     : params.prompt;
-  return weaver.runTurn(relayed.length ? { ...params, prompt } : params);
+  const startedAt = Date.now();
+  try {
+    const result = await weaver.runTurn(
+      relayed.length ? { ...params, prompt } : params,
+    );
+    // Only a turn that reached a backend vouches for it — a refusal or a
+    // turn killed in the queue binds no warp this turn.
+    const bound = turnBinding(weaver, params, startedAt, false);
+    if (bound) noteTurnSucceeded(bound);
+    return result;
+  } catch (err) {
+    const bound = turnBinding(weaver, params, startedAt, true);
+    if (bound) noteTurnFailed(bound, err);
+    throw err;
+  }
+}
+
+/**
+ * The backend/model this turn ran on, for turn-health alerts: the warp
+ * the Weaver bound for it, or — when the turn failed before binding one
+ * — the backend currently serving the chat.
+ */
+function turnBinding(
+  w: Weaver,
+  params: ExecuteParams,
+  startedAt: number,
+  failed: boolean,
+): TurnBinding | null {
+  const common = {
+    chatId: params.chatId,
+    source: params.source,
+    durationMs: Date.now() - startedAt,
+  };
+  const warp = w.loom.get(params.chatId)?.warp;
+  if (warp && warp.boundAt >= startedAt) {
+    return { ...common, backendId: warp.backendId, model: warp.model };
+  }
+  return failed ? { ...common, backendId: backendIdFor(params.chatId) } : null;
 }

package/src/core/engine/fault-text.ts

@@ -0,0 +1,40 @@
+/**
+ * The error text an operator alert or a key=value log line may carry.
+ *
+ * Alerts land on a phone and log lines get pasted into issues, so the
+ * raw message is flattened to one line, clipped, and stripped of the
+ * credential shapes a provider error can echo back (bearer tokens, API
+ * keys, `token=` query parameters).
+ */
+
+const SECRET_PATTERNS: readonly RegExp[] = [
+  /\bBearer\s+[\w.~+/=-]+/gi,
+  /\b(?:sk|pk|rk)-[\w-]{8,}/g,
+  /\b(?:gh[pousr]|github_pat)_\w{8,}/g,
+  /\bxox[abprs]-[\w-]{8,}/g,
+  /\b(access_token|refresh_token|api_key|apikey|token|key|secret|password)=[^\s&"']+/gi,
+];
+
+/** One-line, clipped, secret-free text for `err`. */
+export function faultText(err: unknown, max = 200): string {
+  let raw: string;
+  if (err instanceof Error) raw = err.message || err.name;
+  else if (typeof err === "string") raw = err;
+  else {
+    try {
+      raw = String(err);
+    } catch {
+      raw = "[unprintable error]";
+    }
+  }
+  let text = raw.replace(/\s+/g, " ").trim();
+  for (const pattern of SECRET_PATTERNS) {
+    text = text.replace(pattern, (match, name?: string) =>
+      typeof name === "string" && match.includes("=")
+        ? `${name}=[redacted]`
+        : "[redacted]",
+    );
+  }
+  if (!text) return "unknown error";
+  return text.length > max ? `${text.slice(0, max - 1)}…` : text;
+}

package/src/core/engine/gateway-actions/agents/index.ts

@@ -3,8 +3,9 @@
  *
  *   - `control` — what a parent calls: spawn / list / status / wait / send /
  *     kill. Works from a chat and from inside another agent.
- *   - `report`  — what a sub-agent calls about itself: report_result /
- *     message_parent / check_inbox.
+ *   - `report`  — what a sub-agent calls about itself and its siblings:
+ *     report_result / message_parent / check_inbox / list_peers /
+ *     message_peer.
  *
  * Both sets are reachable from an `agent:<id>` context: the gateway routes
  * those chat keys straight here (see `Gateway.handleAction`), because a

package/src/core/engine/gateway-actions/agents/report.ts

@@ -16,6 +16,7 @@ import {
   agentIdFromContextLabel,
   agentRegistry,
   deliverMessage,
+  deliverToAgent,
   describeParent,
 } from "../../../agents/index.js";
 import { logError } from "../../../../util/log.js";
@@ -86,6 +87,67 @@ export const agentReportHandlers: SharedActionHandlers = {
     };
   },
 
+  list_peers: (_body, _chatId, _backend, chatKey) => {
+    const record = callingAgent(chatKey);
+    if (!record) return notAnAgent("list_peers");
+    const peers = agentRegistry.peersOf(record.id);
+    if (peers.length === 0) {
+      return {
+        ok: true,
+        text:
+          "No peers — you are the only agent your parent has running. " +
+          "Report to your parent as usual.",
+      };
+    }
+    const rendered = peers
+      .map(
+        (peer) =>
+          `- ${peer.label} [${peer.state}]\n  ID: ${peer.id}\n  Working on: ${peer.brief.slice(0, 160).replace(/\s+/g, " ")}…`,
+      )
+      .join("\n");
+    return {
+      ok: true,
+      text: `${peers.length} peer(s) running alongside you:\n\n${rendered}`,
+    };
+  },
+
+  message_peer: (body, _chatId, _backend, chatKey) => {
+    const record = callingAgent(chatKey);
+    if (!record) return notAnAgent("message_peer");
+    const id = String(body.agent_id ?? "").trim();
+    if (!id) return { ok: false, error: "Missing agent_id" };
+    const text = String(body.text ?? "").trim();
+    if (!text) return { ok: false, error: "Missing text" };
+    // Peers only. An agent may address a sibling, never an arbitrary id:
+    // resolving through peersOf is what stops one swarm reaching into
+    // another, and it means a wrong id fails closed rather than delivering.
+    const peer = agentRegistry
+      .peersOf(record.id)
+      .find((candidate) => candidate.id === id);
+    if (!peer) {
+      return {
+        ok: false,
+        error:
+          `No peer "${id}". You can only message agents spawned by the same ` +
+          `parent as you — call list_peers to see them.`,
+      };
+    }
+    if (!deliverToAgent(record.id, id, text)) {
+      return {
+        ok: false,
+        error:
+          `Could not deliver to "${peer.label}" (${id}): it has already ` +
+          `settled, or its inbox is full.`,
+      };
+    }
+    return {
+      ok: true,
+      text:
+        `Queued for peer "${peer.label}" (${id}). It reads its inbox at its ` +
+        `own milestones, so this is not an interrupt.`,
+    };
+  },
+
   check_inbox: (_body, _chatId, _backend, chatKey) => {
     const record = callingAgent(chatKey);
     if (!record) return notAnAgent("check_inbox");

package/src/core/engine/gateway-actions/history.ts

@@ -19,10 +19,8 @@ export const historyHandlers: SharedActionHandlers = {
     const limit = Math.min(100, Number(body.limit ?? 30));
     const cid = chatKey;
     // The tool schema advertises `offset_id` and `before` for paging back;
-    // this fallback used to ignore both and hand back the same newest
-    // window no matter what the model asked, so "go further back" was a
-    // silent no-op on frontends without a platform history API (WhatsApp,
-    // native).
+    // ignoring them here would make "go further back" a silent no-op on
+    // frontends without a platform history API (WhatsApp, native).
     const offsetId = Number(body.offset_id);
     if (Number.isFinite(offsetId) && offsetId > 0) {
       return { ok: true, text: getFormattedBefore(cid, offsetId, limit) };

package/src/core/engine/gateway-actions/native/exec.ts

@@ -61,23 +61,20 @@ async function bash(
     }
   }
 
-  const result = await (async () => {
-    if (background === true) {
-      if (active) {
-        return {
-          ok: false,
-          text:
-            "background:true runs on the daemon host only. On a teleported device, " +
-            "background it in-shell instead: `cmd > /tmp/out.log 2>&1 &`, then poll " +
-            "the log with read.",
-        };
-      }
-      return bashBackground(cmd, dir);
+  if (background === true) {
+    if (active) {
+      return {
+        ok: false,
+        text:
+          "background:true runs on the daemon host only. On a teleported device, " +
+          "background it in-shell instead: `cmd > /tmp/out.log 2>&1 &`, then poll " +
+          "the log with read.",
+      };
     }
-    if (active) return bashTeleported(chatId, active.deviceId, cmd, timeoutMs);
-    return bashLocal(cmd, dir, timeoutMs);
-  })();
-  return result;
+    return bashBackground(cmd, dir);
+  }
+  if (active) return bashTeleported(chatId, active.deviceId, cmd, timeoutMs);
+  return bashLocal(cmd, dir, timeoutMs);
 }
 
 function bashLocal(

package/src/core/engine/gateway.ts

@@ -14,8 +14,10 @@ import { classify } from "../errors.js";
 import { getActiveCount } from "./dispatcher.js";
 import { Loom, getActiveLoom, type ContextRegistry } from "../weaver/index.js";
 import { getHealthStatus } from "../../util/watchdog.js";
+import { activeAlerts } from "../frontend-runtime/alerts.js";
 import { getActiveSessionCount } from "../../storage/sessions.js";
 import { log, logError, logDebug } from "../../util/log.js";
+import { runInChatTurnScope } from "../../util/logging/turn-scope.js";
 import {
   handleSharedAction,
   handleChatFreeAction,
@@ -36,6 +38,38 @@ import type { FrontendActionHandler } from "../types.js";
 import type { Backend } from "../agent-runtime/capabilities.js";
 import { resolveOwnerFrontendId } from "../frontend-runtime/routing.js";
 
+/** Serialized size of an action result; -1 when it can't be serialized. */
+function actionResultBytes(result: unknown): number {
+  try {
+    return Buffer.byteLength(JSON.stringify(result) ?? "");
+  } catch {
+    return -1;
+  }
+}
+
+/** `tool.action name=… chat=… ms=… ok=… bytes=… [err=…]` for one bridge action. */
+function logActionOutcome(
+  body: Record<string, unknown>,
+  chat: string,
+  ms: number,
+  result: unknown,
+): void {
+  const action = typeof body.action === "string" ? body.action : "?";
+  const r = result as { ok?: unknown; error?: unknown } | null | undefined;
+  const failed = !!r && typeof r === "object" && r.ok === false;
+  const line =
+    `tool.action name=${action} chat=${chat || "-"} ms=${ms} ` +
+    `ok=${!failed} bytes=${actionResultBytes(result)}`;
+  if (!failed) {
+    logDebug("gateway", line);
+    return;
+  }
+  const err = String(r.error ?? "")
+    .replace(/\s+/g, " ")
+    .slice(0, 200);
+  log("gateway", `${line} err=${err}`);
+}
+
 // ── Retry helper (stateless — standalone export) ─────────────────────────────
 
 /**
@@ -290,7 +324,31 @@ export class Gateway {
     return null;
   }
 
-  private async handleAction(body: Record<string, unknown>): Promise<unknown> {
+  /**
+   * Run one bridge action inside the log scope of the turn that issued it
+   * (the MCP subprocess reaches us over HTTP, a fresh async root), so the
+   * action's own lines carry `turn=<id>`, and close it with one
+   * `tool.action` summary line — debug on success, info with the error
+   * text when the action answered `ok:false`.
+   */
+  private handleAction(body: Record<string, unknown>): Promise<unknown> {
+    const rawChatId = body._chatId ? String(body._chatId) : "";
+    const numericId = Number(rawChatId);
+    const chatKeys = [
+      rawChatId,
+      rawChatId !== "" && !isNaN(numericId)
+        ? this.loom.stringIdForNumeric(numericId)
+        : null,
+    ];
+    return runInChatTurnScope(chatKeys, async () => {
+      const t0 = Date.now();
+      const result = await this.routeAction(body);
+      logActionOutcome(body, rawChatId, Date.now() - t0, result);
+      return result;
+    });
+  }
+
+  private async routeAction(body: Record<string, unknown>): Promise<unknown> {
     // Route by _chatId from the MCP subprocess request.
     // _chatId may be a string (Teams: "teams_chat_19:...") or numeric string
     // (Telegram: "123456"). The context map is keyed by numeric chatId, so
@@ -492,6 +550,7 @@ export class Gateway {
         w.msSinceLastMessage < 60000
           ? "just now"
           : `${Math.round(w.msSinceLastMessage / 60000)}m ago`,
+      alerts: activeAlerts(),
     };
   }