talon-agent 3.0.5 → 3.1.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "talon-agent",
-  "version": "3.0.5",
+  "version": "3.1.1",
   "description": "Multi-frontend AI agent with full tool access, streaming, cron jobs, and plugin system",
   "author": "Dylan Neve",
   "license": "MIT",
@@ -42,6 +42,7 @@
     "src/native/",
     "src/plugins/",
     "src/storage/",
+    "src/types/",
     "src/util/",
     "src/app.ts",
     "src/bootstrap.ts",
@@ -70,22 +71,22 @@
     "test:openai-agents:backend": "vitest run --reporter=verbose --reporter=json --outputFile=openai-agents-backend-results.json src/__tests__/integration/openai-agents-live-discovery.test.ts",
     "build:sql": "tsx scripts/embed-sql.ts",
     "build:prompts": "tsx scripts/embed-prompts.ts",
-    "build:wasm": "node native/blake3-wasm/build.mjs",
+    "build:wasm": "node native/blake3-wasm/build.mjs && node native/strsim-wasm/build.mjs && node native/sqlguard-wasm/build.mjs && node native/htmlents-wasm/build.mjs",
     "tarball:check": "node .github/scripts/tarball-check.mjs",
     "build:stub-sea": "node src/__tests__/integration/stub-claude/build-sea.mjs",
     "test:watch": "vitest",
     "test:coverage": "vitest run --coverage",
     "typecheck": "tsc --noEmit",
     "lint": "oxlint src/",
+    "depcruise": "depcruise src",
+    "ratchets": "node scripts/check-ratchets.mjs",
     "knip": "knip",
     "format": "prettier --write src/ prompts/",
     "format:check": "prettier --check src/ prompts/",
     "ci:protect": "node .github/scripts/enforce-ci-gate.mjs",
     "build:gleam": "cd native/scheduler-core && gleam build --target javascript && node embed.mjs",
     "build:zig": "node native/textops-wasm/build.mjs",
-    "build:c": "node native/strsim-c/build.mjs && node native/sqlguard-c/build.mjs",
-    "build:cpp": "node native/htmlents-cpp/build.mjs",
-    "build:native": "npm run build:wasm && npm run build:zig && npm run build:c && npm run build:cpp && npm run build:gleam",
+    "build:native": "npm run build:wasm && npm run build:zig && npm run build:gleam",
     "build:driver": "node native/talon-driver/build.mjs",
     "build:driver:all": "node native/talon-driver/build.mjs --all",
     "build:warden": "node native/talon-warden/build.mjs",
@@ -130,9 +131,11 @@
     "zod": "^4.3.6"
   },
   "devDependencies": {
+    "@swc/core": "^1.15.46",
     "@types/node": "^26.0.0",
     "@types/write-file-atomic": "^4.0.3",
     "@vitest/coverage-v8": "^4.1.3",
+    "dependency-cruiser": "^18.1.0",
     "fast-check": "^4.6.0",
     "knip": "^6.3.1",
     "oxlint": "^1.59.0",

package/src/backend/kilo/models/index.ts

@@ -13,7 +13,7 @@
  * Kilo ids routinely contain "/" and ":" (e.g. "inclusionai/ling-2.6-1t:free").
  */
 
-import { ensureServer } from "../server.js";
+import { ensureServer, onServerStop } from "../server.js";
 import { createRemoteModelCatalogModule } from "../../remote-server/model-catalog/index.js";
 
 export type {
@@ -41,6 +41,9 @@ const kiloModels = createRemoteModelCatalogModule({
   quickPickLimit: 24,
 });
 
+// A stopped server invalidates the catalog it served.
+onServerStop(kiloModels.clearCache);
+
 export const getOpenCodeModelCatalog = kiloModels.getCatalog;
 export const clearModelCatalogCache = kiloModels.clearCache;
 export const getOpenCodeModelInfo = kiloModels.getModelInfo;

package/src/backend/kilo/server.ts

@@ -25,13 +25,12 @@ import type { TalonConfig } from "../../util/config.js";
 import type { FrontendName } from "../../core/agent-runtime/backend-registry.js";
 import { logWarn } from "../../util/log.js";
 import { buildDeliveryContract } from "../shared/delivery-contract.js";
-import { clearModelCatalogCache } from "./models/index.js";
 import {
   guessProviderID,
   getBucketPriority,
   normalizeModelLookup,
-  parseOpenCodeModelQuery,
-} from "./models/index.js";
+  parseRemoteModelQuery as parseOpenCodeModelQuery,
+} from "../remote-server/model-catalog/index.js";
 import {
   createRemoteServerState,
   ensureRemoteServer as ensureRemoteServerShared,
@@ -153,6 +152,18 @@ async function prewarmPluginMcpServers(): Promise<void> {
   await ensurePluginMcpServers(client, "prewarm");
 }
 
+/**
+ * Callbacks run when the server stops — cache invalidation lives with the
+ * caches. The models module registers its clear here at load time, so the
+ * server never has to import it (which would be a cycle).
+ */
+const stopHooks = new Set<() => void>();
+
+/** Register a callback to run whenever the Kilo server is stopped. */
+export function onServerStop(hook: () => void): void {
+  stopHooks.add(hook);
+}
+
 /**
  * Stop the local Kilo server (if we own it) and clear caches.
  *
@@ -160,7 +171,9 @@ async function prewarmPluginMcpServers(): Promise<void> {
  * server, this leaves it running — we don't own it.
  */
 export function stopKiloServer(): void {
-  stopRemoteServer(state, clearModelCatalogCache);
+  stopRemoteServer(state, () => {
+    for (const hook of stopHooks) hook();
+  });
 }
 
 // ── Server lifecycle ────────────────────────────────────────────────────────

package/src/backend/openai-agents/session.ts

@@ -251,7 +251,7 @@ export class TalonSession extends MemorySession {
     return current;
   }
 
-  async addItems(items: AgentInputItem[]): Promise<void> {
+  override async addItems(items: AgentInputItem[]): Promise<void> {
     await super.addItems(items);
     await this.enforceCap();
   }

package/src/backend/opencode/models/index.ts

@@ -12,7 +12,7 @@
  * stay at 4 and only short separator-free ids are embedded raw.
  */
 
-import { ensureServer } from "../server.js";
+import { ensureServer, onServerStop } from "../server.js";
 import { createRemoteModelCatalogModule } from "../../remote-server/model-catalog/index.js";
 
 export type {
@@ -40,6 +40,9 @@ const opencodeModels = createRemoteModelCatalogModule({
   quickPickLimit: 4,
 });
 
+// A stopped server invalidates the catalog it served.
+onServerStop(opencodeModels.clearCache);
+
 export const getOpenCodeModelCatalog = opencodeModels.getCatalog;
 export const clearModelCatalogCache = opencodeModels.clearCache;
 export const getOpenCodeModelInfo = opencodeModels.getModelInfo;

package/src/backend/opencode/server.ts

@@ -25,13 +25,12 @@ import type { TalonConfig } from "../../util/config.js";
 import type { FrontendName } from "../../core/agent-runtime/backend-registry.js";
 import { logWarn } from "../../util/log.js";
 import { buildDeliveryContract } from "../shared/delivery-contract.js";
-import { clearModelCatalogCache } from "./models/index.js";
 import {
   guessProviderID,
   getBucketPriority,
   normalizeModelLookup,
-  parseOpenCodeModelQuery,
-} from "./models/index.js";
+  parseRemoteModelQuery as parseOpenCodeModelQuery,
+} from "../remote-server/model-catalog/index.js";
 import {
   createRemoteServerState,
   ensureRemoteServer as ensureRemoteServerShared,
@@ -108,8 +107,22 @@ async function prewarmPluginMcpServers(): Promise<void> {
   await ensurePluginMcpServers(client, "prewarm");
 }
 
+/**
+ * Callbacks run when the server stops — cache invalidation lives with the
+ * caches. The models module registers its clear here at load time, so the
+ * server never has to import it (which would be a cycle).
+ */
+const stopHooks = new Set<() => void>();
+
+/** Register a callback to run whenever the OpenCode server is stopped. */
+export function onServerStop(hook: () => void): void {
+  stopHooks.add(hook);
+}
+
 export function stopOpenCodeServer(): void {
-  stopRemoteServer(state, clearModelCatalogCache);
+  stopRemoteServer(state, () => {
+    for (const hook of stopHooks) hook();
+  });
 }
 
 // ── Server lifecycle ────────────────────────────────────────────────────────

package/src/backend/shared/system-prompt.ts

@@ -41,6 +41,7 @@ import {
   type TalonConfig,
 } from "../../util/config.js";
 import { getPluginPromptAdditions } from "../../core/plugin/index.js";
+import { onPromptInputsChanged } from "../../core/prompt/invalidation.js";
 import { frontendForChatId, nonTerminalFrontends } from "./frontends.js";
 
 // ── Types ───────────────────────────────────────────────────────────────────
@@ -112,6 +113,10 @@ export function clearSystemPromptSnapshots(): void {
   snapshots.clear();
 }
 
+// Core and frontends invalidate through the core-side seam — they never
+// import this module (layer rule: nothing above backend/ reaches into it).
+onPromptInputsChanged(clearSystemPromptSnapshots);
+
 // ── Public API ──────────────────────────────────────────────────────────────
 
 /**

package/src/cli/events.ts

@@ -52,7 +52,7 @@ async function showHistory(limit: number): Promise<void> {
   const { readJournal } = await import("../storage/journal.js");
   let entries;
   try {
-    entries = readJournal({ limit });
+    entries = readJournal<TalonEvent & { at: number }>({ limit });
   } catch (err) {
     console.log(
       `  ${pc.red("✖")} Could not read the journal: ${err instanceof Error ? err.message : err}\n`,

package/src/cli/index.ts

@@ -172,7 +172,7 @@ export async function runCli(): Promise<void> {
       mainMenu();
       break;
     default: {
-      // "did you mean ...?" via the C similarity core (native/strsim-c).
+      // "did you mean ...?" via the native similarity core (native/strsim-wasm).
       const { closestMatch } = await import("../native/strsim.js");
       const suggestion = closestMatch(command, CLI_COMMANDS);
       const hint = suggestion

package/src/cli/tasks.ts

@@ -16,6 +16,7 @@ import type {
   TaskRecord,
   TaskState,
 } from "../core/tasks/index.js";
+import type { PublishedEvent } from "../core/bus/index.js";
 
 function formatDuration(ms: number): string {
   const seconds = Math.floor(ms / 1000);
@@ -120,7 +121,10 @@ async function journalTasks(
   const { readJournal } = await import("../storage/journal.js");
   const seen = new Set(liveTasks.map((t) => `${t.id}:${t.queuedAt}`));
   const historical: TaskRecord[] = [];
-  for (const entry of readJournal({ type: "task.settled", limit })) {
+  for (const entry of readJournal<PublishedEvent>({
+    type: "task.settled",
+    limit,
+  })) {
     if (entry.event.type !== "task.settled") continue;
     const task = entry.event.task;
     // (id, queuedAt) identifies a run across surfaces — per-process ids

package/src/core/background/triggers/pid.ts

@@ -0,0 +1,28 @@
+/**
+ * PID-starttime probe — the /proc read shared by spawn (capture at fork)
+ * and resume (compare after restart). Lives alone so neither module has
+ * to import the other for it.
+ */
+
+import { readFileSync } from "node:fs";
+
+/**
+ * Read field 22 (start time in jiffies since boot) from /proc/<pid>/stat.
+ * Returns undefined if /proc isn't available (non-Linux) or the read fails.
+ *
+ * Parsing note: the `comm` field (2nd) is wrapped in parens and may itself
+ * contain ')' — the safe parse finds the LAST ')' and splits the rest on
+ * space. After that split, index 19 corresponds to field 22.
+ */
+export function readPidStarttimeSync(pid: number): number | undefined {
+  try {
+    const stat = readFileSync(`/proc/${pid}/stat`, "utf-8");
+    const lastParen = stat.lastIndexOf(")");
+    if (lastParen < 0) return undefined;
+    const tail = stat.slice(lastParen + 2).split(" ");
+    const starttime = Number(tail[19]);
+    return Number.isFinite(starttime) ? starttime : undefined;
+  } catch {
+    return undefined;
+  }
+}

package/src/core/background/triggers/resume.ts

@@ -1,10 +1,10 @@
 /**
  * Post-restart resume — respawn persistent triggers, fire late wakes for
- * recently-terminated ones, plus the /proc-based PID-starttime probe and the
- * orphan-kill used to avoid duplicate spawns after an unclean crash.
+ * recently-terminated ones, plus the orphan-kill used to avoid duplicate
+ * spawns after an unclean crash. The /proc PID-starttime probe both this
+ * and spawn use lives in ./pid.ts.
  */
 
-import { readFileSync } from "node:fs";
 import {
   getAllTriggers,
   updateTrigger,
@@ -15,6 +15,7 @@ import { log, logError } from "../../../util/log.js";
 import { depsHolder } from "./state.js";
 import { fireWake } from "./output.js";
 import { spawnTrigger } from "./spawn.js";
+import { readPidStarttimeSync } from "./pid.js";
 
 /**
  * After the dispatcher is wired, walk the store and clean up leftover state
@@ -68,27 +69,6 @@ export async function resumeAfterRestart(): Promise<void> {
   }
 }
 
-/**
- * Read field 22 (start time in jiffies since boot) from /proc/<pid>/stat.
- * Returns undefined if /proc isn't available (non-Linux) or the read fails.
- *
- * Parsing note: the `comm` field (2nd) is wrapped in parens and may itself
- * contain ')' — the safe parse finds the LAST ')' and splits the rest on
- * space. After that split, index 19 corresponds to field 22.
- */
-export function readPidStarttimeSync(pid: number): number | undefined {
-  try {
-    const stat = readFileSync(`/proc/${pid}/stat`, "utf-8");
-    const lastParen = stat.lastIndexOf(")");
-    if (lastParen < 0) return undefined;
-    const tail = stat.slice(lastParen + 2).split(" ");
-    const starttime = Number(tail[19]);
-    return Number.isFinite(starttime) ? starttime : undefined;
-  } catch {
-    return undefined;
-  }
-}
-
 /**
  * Probe a stored PID from a previous Talon run and SIGKILL it if it's still
  * alive AND really is our former child (not a recycled PID). Used by

package/src/core/background/triggers/spawn.ts

@@ -25,7 +25,7 @@ import {
 import { commandForLanguage } from "./command.js";
 import { handleStdoutLine, handleStderrLine } from "./output.js";
 import { handleTimeout, finalizeExit, failTrigger } from "./exit.js";
-import { readPidStarttimeSync } from "./resume.js";
+import { readPidStarttimeSync } from "./pid.js";
 
 /**
  * Spawn a trigger's script as a supervised child process.

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

@@ -22,8 +22,8 @@ export async function performPluginReload(
   const { reloadPlugins, getPluginPromptAdditions } =
     await import("../../plugin/index.js");
   const { rebuildSystemPrompt } = await import("../../../util/config.js");
-  const { clearSystemPromptSnapshots } =
-    await import("../../../backend/shared/system-prompt.js");
+  const { notifyPromptInputsChanged } =
+    await import("../../prompt/invalidation.js");
 
   // reloadPlugins reads + validates config internally — no double read.
   // Frontends are derived from config if not explicitly provided.
@@ -37,7 +37,7 @@ export async function performPluginReload(
   // Plugin prompt additions changed — drop per-session prompt
   // snapshots so every chat's next turn picks up the new prompt
   // (deliberate one-time cache re-write per live session).
-  clearSystemPromptSnapshots();
+  notifyPromptInputsChanged();
 
   return { names };
 }

package/src/core/prompt/invalidation.ts

@@ -0,0 +1,23 @@
+/**
+ * Prompt-input invalidation — the core-side seam callers use when prompt
+ * inputs change out from under live sessions on purpose (plugin reload,
+ * skill toggle).
+ *
+ * Whoever caches assembled prompts registers a hook here (today: the
+ * per-session snapshot store in backend/shared/system-prompt.ts, at its
+ * module load). Core and frontends call `notifyPromptInputsChanged()`
+ * and never import the cache — the dependency points backend → core,
+ * as the layer rule requires.
+ */
+
+const hooks = new Set<() => void>();
+
+/** Register a callback run whenever prompt inputs are invalidated. */
+export function onPromptInputsChanged(hook: () => void): void {
+  hooks.add(hook);
+}
+
+/** Invalidate: every registered prompt cache drops its state. */
+export function notifyPromptInputsChanged(): void {
+  for (const hook of hooks) hook();
+}

package/src/core/types.ts

@@ -29,8 +29,8 @@ export type UnifiedModelInfo = {
   unavailableReason?: string;
 };
 
-export type ReasoningEffortLevel =
-  "off" | "minimal" | "low" | "medium" | "high" | "max" | "xhigh";
+export type { ReasoningEffortLevel } from "../types/effort.js";
+import type { ReasoningEffortLevel } from "../types/effort.js";
 
 /** Result of resolving a user's model query. */
 export type UnifiedModelResolution =

package/src/frontend/discord/callbacks/components.ts

@@ -34,9 +34,9 @@ import {
   setChatModelForBackend,
   setChatBackend,
   setChatEffort,
-  resolveModelName,
   type EffortLevel,
 } from "../../../storage/chat-settings.js";
+import { resolveModelId as resolveModelName } from "../../../core/models/catalog.js";
 import {
   registerChat,
   disablePulse,

package/src/frontend/discord/commands/settings.ts

@@ -18,9 +18,9 @@ import {
   setChatBackend,
   setChatEffort,
   setChatPulseInterval,
-  resolveModelName,
   type EffortLevel,
 } from "../../../storage/chat-settings.js";
+import { resolveModelId as resolveModelName } from "../../../core/models/catalog.js";
 import {
   registerChat,
   disablePulse,

package/src/frontend/native/auth.ts

@@ -0,0 +1,52 @@
+/**
+ * Bridge auth token — the shared secret behind every non-/health route.
+ *
+ * The operator can set one explicitly (`native.token`); this module covers
+ * the case where they didn't but the bridge is about to bind a non-loopback
+ * host. Serving the full agent API unauthenticated to the LAN is never an
+ * acceptable default, so a token is minted once and persisted under
+ * ~/.talon/keys/ — stable across restarts so paired clients keep working.
+ *
+ * The token value itself never goes to the log (SECURITY.md treats
+ * credentials in logs as a vulnerability); same-machine clients pick it up
+ * from the 0600 discovery file, remote clients read it from the key file.
+ */
+
+import { randomBytes } from "node:crypto";
+import { chmodSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { resolve } from "node:path";
+import { dirs } from "../../util/paths.js";
+import { log } from "../../util/log.js";
+
+const TOKEN_FILE = "bridge-token";
+/** 32 random bytes → 43 base64url chars; comfortably beyond brute force. */
+const TOKEN_BYTES = 32;
+
+/** Path of the persisted token, for operator-facing messages. */
+export function bridgeTokenPath(dir: string = dirs.keys): string {
+  return resolve(dir, TOKEN_FILE);
+}
+
+/**
+ * Load the persisted auto-generated bridge token, minting it on first use.
+ * Lives under ~/.talon/keys/ with owner-only permissions, like the TLS
+ * identity next to it.
+ */
+export function loadOrCreateBridgeToken(dir: string = dirs.keys): string {
+  const path = bridgeTokenPath(dir);
+  try {
+    const existing = readFileSync(path, "utf-8").trim();
+    if (existing) return existing;
+  } catch {
+    // first boot — nothing persisted yet
+  }
+  const token = randomBytes(TOKEN_BYTES).toString("base64url");
+  mkdirSync(dir, { recursive: true, mode: 0o700 });
+  writeFileSync(path, `${token}\n`, { mode: 0o600 });
+  chmodSync(path, 0o600); // mode above is ignored when the file exists
+  log(
+    "native",
+    `Minted bridge auth token (non-loopback bind with no native.token) — pair remote clients with the value in ${path}`,
+  );
+  return token;
+}

package/src/frontend/native/extensions.ts

@@ -21,7 +21,7 @@ import type { Backend } from "../../core/agent-runtime/capabilities.js";
 import { performPluginReload } from "../../core/engine/gateway-actions/plugins.js";
 import { getPluginPromptAdditions } from "../../core/plugin/index.js";
 import { listPluginItems, setPluginEnabled } from "../../core/plugin/manage.js";
-import { clearSystemPromptSnapshots } from "../../backend/shared/system-prompt.js";
+import { notifyPromptInputsChanged } from "../../core/prompt/invalidation.js";
 import { listSkills, setSkillEnabled } from "../../storage/skill-store.js";
 import { rebuildSystemPrompt, type TalonConfig } from "../../util/config.js";
 import { log } from "../../util/log.js";
@@ -88,6 +88,6 @@ export function toggleSkill(
   log("native", `Skill ${name} ${enabled ? "enabled" : "disabled"} via bridge`);
   rebuildSystemPrompt(config, getPluginPromptAdditions());
   backend?.control?.updateSystemPrompt?.(config.systemPrompt);
-  clearSystemPromptSnapshots();
+  notifyPromptInputsChanged();
   return { ok: true };
 }

package/src/frontend/native/index.ts

@@ -75,6 +75,7 @@ import { createNativeActionHandler } from "./actions.js";
 import { getMeshService } from "../../core/mesh/index.js";
 import { removeBridgeDiscovery, writeBridgeDiscovery } from "./discovery.js";
 import { isLoopbackHost, loadOrCreateBridgeTlsIdentity } from "./tls.js";
+import { loadOrCreateBridgeToken } from "./auth.js";
 import { readLogEntries } from "./logs.js";
 import {
   BRIDGE_PROTOCOL_VERSION,
@@ -1200,11 +1201,16 @@ export function createNativeFrontend(
   // Encrypted by default the moment the bridge leaves the machine; loopback
   // stays plain HTTP unless explicitly opted in (`native.tls`).
   const bridgeTls = nativeCfg.tls ?? !isLoopbackHost(bridgeHost);
+  // Never serve the agent API to the network unauthenticated: a non-loopback
+  // bind with no configured token gets a persistent auto-minted one instead.
+  const bridgeToken =
+    nativeCfg.token ??
+    (isLoopbackHost(bridgeHost) ? undefined : loadOrCreateBridgeToken());
   const server = new BridgeServer(
     {
       host: bridgeHost,
       port: nativeCfg.port ?? 19880,
-      token: nativeCfg.token,
+      token: bridgeToken,
       startedAt,
       ...(bridgeTls ? { tls: () => loadOrCreateBridgeTlsIdentity() } : {}),
     },
@@ -1274,7 +1280,7 @@ export function createNativeFrontend(
       const fingerprint = server.getFingerprint();
       await writeBridgeDiscovery({
         port: server.getPort(),
-        token: nativeCfg.token,
+        token: bridgeToken,
         scheme: server.getScheme(),
         ...(fingerprint ? { fingerprint } : {}),
         startedAt: Date.parse(startedAt),