@openparachute/agent 0.1.2 → 0.2.0

package/src/cron.ts

@@ -0,0 +1,380 @@
+/**
+ * A tiny, dependency-free cron evaluator for the runner (design
+ * `2026-06-17-runner-scheduled-agent-turns.md`).
+ *
+ * Scope (v1, deliberately narrow):
+ *   - FIVE fields only: `minute hour day-of-month month day-of-week`.
+ *   - Each field supports `*`, `*​/n` (step), `a-b` (range), `a-b/n` (range+step),
+ *     and `a,b,c` (comma list, each element any of the above).
+ *   - NO seconds field, NO macros (`@daily`), NO names (`MON`, `JAN`) — numeric only.
+ *   - day-of-week: 0-6, Sunday = 0. (7 is NOT accepted as Sunday in v1 — keep the
+ *     accepted set tight and unambiguous.)
+ *
+ * Day-of-month / day-of-week semantics follow standard cron: if BOTH `dom` and
+ * `dow` are restricted (neither is `*`), a day matches when EITHER matches (the
+ * union). If only one is restricted, only that one constrains. This is the Vixie
+ * cron rule and the one operators expect from "0 9 * * 1-5".
+ *
+ * Timezone correctness — the #1 risk. `nextRunAfter` evaluates the cron fields
+ * against the WALL CLOCK in a given IANA timezone (default: the daemon's local
+ * tz). It does this by walking forward minute-by-minute from `from`, projecting
+ * each candidate instant into the target tz via `Intl.DateTimeFormat` (Bun ships
+ * full ICU), and testing the projected wall-clock fields against the cron sets.
+ * This is O(minutes-until-next-match); bounded by a hard cap (~366 days of
+ * minutes) so a pathological/never-matching spec returns null instead of looping
+ * forever. The minute-walk is simple and DST-honest: because we read the wall
+ * clock the OS/ICU reports for each real instant, a spring-forward gap simply has
+ * no matching instant (the wall time never occurs → that fire is skipped that
+ * day), and a fall-back repeat can match twice (both 01:30s exist as distinct
+ * instants → both fire). v1 accepts that behavior rather than inventing a policy;
+ * see the comment at `nextRunAfter`.
+ */
+
+/** A parsed cron expression: the allowed-value Sets per field, plus restriction flags. */
+export interface ParsedCron {
+  minute: Set<number>;
+  hour: Set<number>;
+  /** day-of-month (1-31). */
+  dom: Set<number>;
+  /** month (1-12). */
+  month: Set<number>;
+  /** day-of-week (0-6, Sun=0). */
+  dow: Set<number>;
+  /** Whether `dom` was `*` (unrestricted) — drives the dom/dow union rule. */
+  domStar: boolean;
+  /** Whether `dow` was `*` (unrestricted) — drives the dom/dow union rule. */
+  dowStar: boolean;
+}
+
+/** Inclusive numeric bounds for each of the five fields (v1: numeric only). */
+const FIELD_BOUNDS: ReadonlyArray<{ min: number; max: number; name: string }> = [
+  { min: 0, max: 59, name: "minute" },
+  { min: 0, max: 23, name: "hour" },
+  { min: 1, max: 31, name: "day-of-month" },
+  { min: 1, max: 12, name: "month" },
+  { min: 0, max: 6, name: "day-of-week" },
+];
+
+/** Thrown by `parseCron` on a malformed expression — callers map it to a 400. */
+export class CronParseError extends Error {}
+
+/**
+ * Parse ONE field token (between the spaces) into the set of allowed integers in
+ * `[min, max]`. Supports `*`, `*​/n`, `a`, `a-b`, `a-b/n`, and comma lists of any
+ * of those. Throws `CronParseError` on anything out of range or malformed.
+ */
+function parseField(token: string, min: number, max: number, fieldName: string): { set: Set<number>; star: boolean } {
+  const trimmed = token.trim();
+  if (trimmed.length === 0) {
+    throw new CronParseError(`cron ${fieldName}: empty field`);
+  }
+  const set = new Set<number>();
+  // A field is `*` only if EVERY comma-element is `*` / `*​/n` over the full range.
+  // We track whether the literal token was a bare `*` (no list, no step) for the
+  // dom/dow union rule — a stepped `*​/2` is a restriction, not "unrestricted".
+  const star = trimmed === "*";
+
+  for (const part of trimmed.split(",")) {
+    const elem = part.trim();
+    if (elem.length === 0) {
+      throw new CronParseError(`cron ${fieldName}: empty list element in "${token}"`);
+    }
+
+    // Split off an optional step (`/n`).
+    let rangePart = elem;
+    let step = 1;
+    const slash = elem.indexOf("/");
+    if (slash >= 0) {
+      rangePart = elem.slice(0, slash);
+      const stepStr = elem.slice(slash + 1);
+      if (!/^\d+$/.test(stepStr)) {
+        throw new CronParseError(`cron ${fieldName}: bad step "/${stepStr}" in "${token}"`);
+      }
+      step = parseInt(stepStr, 10);
+      if (step <= 0) {
+        throw new CronParseError(`cron ${fieldName}: step must be >= 1 in "${token}"`);
+      }
+    }
+
+    // Resolve the range the step applies over.
+    let lo: number;
+    let hi: number;
+    if (rangePart === "*") {
+      lo = min;
+      hi = max;
+    } else if (rangePart.includes("-")) {
+      const [aStr, bStr, ...rest] = rangePart.split("-");
+      if (rest.length > 0 || aStr === undefined || bStr === undefined) {
+        throw new CronParseError(`cron ${fieldName}: bad range "${rangePart}" in "${token}"`);
+      }
+      if (!/^\d+$/.test(aStr) || !/^\d+$/.test(bStr)) {
+        throw new CronParseError(`cron ${fieldName}: non-numeric range "${rangePart}" in "${token}"`);
+      }
+      lo = parseInt(aStr, 10);
+      hi = parseInt(bStr, 10);
+      if (lo > hi) {
+        throw new CronParseError(`cron ${fieldName}: descending range "${rangePart}" in "${token}"`);
+      }
+    } else {
+      // A single number, optionally with a step (`5/2` means 5,7,9,… to max — the
+      // step extends a bare number to the field max, matching common cron impls).
+      if (!/^\d+$/.test(rangePart)) {
+        throw new CronParseError(`cron ${fieldName}: non-numeric value "${rangePart}" in "${token}"`);
+      }
+      lo = parseInt(rangePart, 10);
+      hi = slash >= 0 ? max : lo;
+    }
+
+    if (lo < min || hi > max) {
+      throw new CronParseError(
+        `cron ${fieldName}: value out of range (${lo}-${hi}); allowed ${min}-${max}`,
+      );
+    }
+    for (let v = lo; v <= hi; v += step) set.add(v);
+  }
+
+  if (set.size === 0) {
+    throw new CronParseError(`cron ${fieldName}: no values matched in "${token}"`);
+  }
+  return { set, star };
+}
+
+/**
+ * Parse a 5-field cron expression into a {@link ParsedCron}. Throws
+ * `CronParseError` (message names the offending field) on anything malformed.
+ * Whitespace between fields is any run of spaces/tabs.
+ */
+export function parseCron(expr: string): ParsedCron {
+  if (typeof expr !== "string") {
+    throw new CronParseError("cron expression must be a string");
+  }
+  const fields = expr.trim().split(/\s+/);
+  if (fields.length !== 5) {
+    throw new CronParseError(
+      `cron expression must have exactly 5 fields (min hour dom mon dow); got ${fields.length}: "${expr}"`,
+    );
+  }
+  const minute = parseField(fields[0]!, FIELD_BOUNDS[0]!.min, FIELD_BOUNDS[0]!.max, "minute");
+  const hour = parseField(fields[1]!, FIELD_BOUNDS[1]!.min, FIELD_BOUNDS[1]!.max, "hour");
+  const dom = parseField(fields[2]!, FIELD_BOUNDS[2]!.min, FIELD_BOUNDS[2]!.max, "day-of-month");
+  const month = parseField(fields[3]!, FIELD_BOUNDS[3]!.min, FIELD_BOUNDS[3]!.max, "month");
+  const dow = parseField(fields[4]!, FIELD_BOUNDS[4]!.min, FIELD_BOUNDS[4]!.max, "day-of-week");
+  return {
+    minute: minute.set,
+    hour: hour.set,
+    dom: dom.set,
+    month: month.set,
+    dow: dow.set,
+    domStar: dom.star,
+    dowStar: dow.star,
+  };
+}
+
+/** The wall-clock fields of an instant projected into a given IANA timezone. */
+interface WallClock {
+  minute: number;
+  hour: number;
+  /** day-of-month (1-31). */
+  dom: number;
+  /** month (1-12). */
+  month: number;
+  /** day-of-week (0-6, Sun=0). */
+  dow: number;
+}
+
+const WEEKDAY_INDEX: Record<string, number> = {
+  Sun: 0,
+  Mon: 1,
+  Tue: 2,
+  Wed: 3,
+  Thu: 4,
+  Fri: 5,
+  Sat: 6,
+};
+
+/**
+ * Project a real instant (`Date`) into its wall-clock fields IN `tz` via
+ * `Intl.DateTimeFormat`. Throws if `tz` is not a valid IANA zone (the formatter
+ * throws a RangeError — we let it propagate so `nextRunAfter`'s caller surfaces
+ * "bad timezone" rather than silently using UTC).
+ */
+function wallClockInTz(date: Date, tz: string): WallClock {
+  // `en-US` with explicit numeric parts + `weekday: short` gives stable,
+  // locale-independent token shapes we can parse back to integers.
+  const parts = new Intl.DateTimeFormat("en-US", {
+    timeZone: tz,
+    hour12: false,
+    weekday: "short",
+    year: "numeric",
+    month: "numeric",
+    day: "numeric",
+    hour: "numeric",
+    minute: "numeric",
+  }).formatToParts(date);
+
+  const get = (type: string): string => {
+    const p = parts.find((x) => x.type === type);
+    return p ? p.value : "";
+  };
+
+  // `hour12: false` can emit "24" for midnight in some ICU versions; normalize.
+  let hour = parseInt(get("hour"), 10);
+  if (hour === 24) hour = 0;
+
+  return {
+    minute: parseInt(get("minute"), 10),
+    hour,
+    dom: parseInt(get("day"), 10),
+    month: parseInt(get("month"), 10),
+    dow: WEEKDAY_INDEX[get("weekday")] ?? 0,
+  };
+}
+
+/** Does a wall clock satisfy the parsed cron? (applies the dom/dow union rule). */
+function matches(parsed: ParsedCron, wc: WallClock): boolean {
+  if (!parsed.minute.has(wc.minute)) return false;
+  if (!parsed.hour.has(wc.hour)) return false;
+  if (!parsed.month.has(wc.month)) return false;
+
+  // Standard cron dom/dow rule: if BOTH are restricted, match on EITHER (union).
+  // If only one is restricted, only that one constrains. If both are `*`, both
+  // pass trivially.
+  const domOk = parsed.dom.has(wc.dom);
+  const dowOk = parsed.dow.has(wc.dow);
+  if (!parsed.domStar && !parsed.dowStar) {
+    return domOk || dowOk;
+  }
+  if (!parsed.domStar) return domOk;
+  if (!parsed.dowStar) return dowOk;
+  return true; // both `*`
+}
+
+/** Does ONLY the date part (month + dom/dow union) match? Used for coarse day-skip. */
+function dateMatches(parsed: ParsedCron, wc: WallClock): boolean {
+  if (!parsed.month.has(wc.month)) return false;
+  const domOk = parsed.dom.has(wc.dom);
+  const dowOk = parsed.dow.has(wc.dow);
+  if (!parsed.domStar && !parsed.dowStar) return domOk || dowOk;
+  if (!parsed.domStar) return domOk;
+  if (!parsed.dowStar) return dowOk;
+  return true;
+}
+
+/**
+ * The hard cap for the forward search, expressed in DAYS. A sparse spec like
+ * "Feb 29" can legitimately be up to ~4 years out; cap at 5 years so it resolves
+ * while a truly never-matching spec (impossible with numeric-only fields, but
+ * cheap insurance) returns null instead of looping. Day-coarse skipping keeps the
+ * search O(days-until-match) + O(minutes-within-the-day), not O(total minutes).
+ */
+const MAX_LOOKAHEAD_DAYS = 5 * 366;
+
+/**
+ * Return the next instant STRICTLY AFTER `from` whose wall-clock-in-`tz` matches
+ * `expr`, or `null` if no match within ~5 years (effectively a never-firing spec).
+ *
+ * STRICTLY AFTER is load-bearing: the runner persists `nextRunAt` and re-derives
+ * forward from a fired instant, so returning `from` itself would double-fire on
+ * the same minute. We advance to the start of the NEXT minute first, then search
+ * forward at cron's one-minute resolution.
+ *
+ * Search strategy (so a sparse spec like Feb 29 doesn't take a million minute
+ * steps): a single FORWARD-ONLY minute cursor with a DAY-SKIP fast path. On each
+ * step we read the cursor's wall clock; if the cursor's wall DATE (month + dom/dow
+ * union) can't match, we skip the cursor forward to the next wall-midnight in one
+ * jump instead of crawling minute-by-minute through the dead day. On a date that
+ * DOES qualify, we walk its minutes testing the full predicate. So the cost is
+ * O(days-to-match) date-checks + O(minutes-in-the-few-matching-days) minute-checks.
+ *
+ * Forward-only is what keeps strictly-after honest: the cursor starts at the first
+ * minute AFTER `from` and never rewinds, so the first match it reaches is the
+ * earliest instant > `from`. (We deliberately do NOT rewind to wall-midnight on
+ * the matching day — that could surface a match earlier than `from`.)
+ *
+ * Timezone + DST: each candidate is a real instant; we read the wall clock the
+ * target tz reports for it. So a spring-forward gap (e.g. 02:00→03:00) has no
+ * matching instant for a 02:30 spec that day — it's simply skipped (the next
+ * day fires). A fall-back repeat (01:30 occurring twice) yields two distinct
+ * matching instants — both fire. v1 documents and accepts this rather than
+ * inventing skip/dedup policy; jobs are coarse ("daily 8am"), and the
+ * fire-once-on-miss catch-up in the runner bounds any practical surprise.
+ *
+ * The day-skip advances to the next WALL-midnight (00:00 of the next wall-day IN
+ * `zone`), computed from the wall clock we already read — NOT to UTC-midnight. This
+ * is load-bearing in a negative-offset zone: UTC-midnight there is ~17:00 of the
+ * wall-day, so jumping to it would strand the forward-only cursor in the *evening*
+ * and the crawl could never reach that wall-day's MORNING (a `0 9 * * *` would be
+ * missed; a sparse-dom morning job would never fire). Instead we step the cursor by
+ * the minutes from this wall-instant to the next wall-midnight
+ * (`(23-hour)*60 + (60-minute)`), which lands at/just-before the next wall-day's
+ * 00:00. A DST transition can make the landing 23:00 or 01:00 of the wall-day — the
+ * main loop self-corrects with a few minute steps, and it never lands in the
+ * evening. The minute-walk remains the source of truth, so no offset arithmetic can
+ * desync it. (Each skip advances ≥1 minute, so the search always terminates.)
+ *
+ * `from` defaults to now; `tz` defaults to the daemon's local timezone (resolved
+ * from `Intl.DateTimeFormat().resolvedOptions().timeZone`).
+ */
+export function nextRunAfter(expr: string | ParsedCron, tz?: string, from: Date = new Date()): Date | null {
+  const parsed = typeof expr === "string" ? parseCron(expr) : expr;
+  const zone = tz ?? Intl.DateTimeFormat().resolvedOptions().timeZone;
+
+  // Validate the zone once (and fail loudly) by projecting `from`. An invalid IANA
+  // zone throws RangeError here, which propagates to the caller as a clear error.
+  wallClockInTz(from, zone);
+
+  // Start at the top of the NEXT minute after `from` (strictly-after + minute
+  // resolution: zero the seconds/ms and step one minute past `from`'s minute).
+  const cursor = new Date(from.getTime());
+  cursor.setUTCSeconds(0, 0);
+  cursor.setUTCMinutes(cursor.getUTCMinutes() + 1);
+
+  // Bound the search by DATE-CHECKS (one per distinct wall-day we touch), so a
+  // sparse "Feb 29" still resolves while a never-matching spec terminates. We
+  // count days touched, not minutes, since the day-skip collapses dead days.
+  let daysTouched = 0;
+  let lastSkipKey = "";
+
+  // Cap the total minute steps generously: matching days are few, and each gets a
+  // bounded (~25h) minute scan; a hard ceiling is belt-and-suspenders against an
+  // infinite loop. 5y of days × ~1500 min is the theoretical worst case, but the
+  // day-skip means we never get near it for real specs.
+  const MAX_MINUTE_STEPS = MAX_LOOKAHEAD_DAYS * 24 * 60;
+
+  for (let i = 0; i < MAX_MINUTE_STEPS; i++) {
+    const wc = wallClockInTz(cursor, zone);
+    if (dateMatches(parsed, wc)) {
+      if (matches(parsed, wc)) return new Date(cursor.getTime());
+      // Date qualifies but this minute/hour doesn't — crawl one minute.
+      cursor.setUTCMinutes(cursor.getUTCMinutes() + 1);
+      continue;
+    }
+    // Date does NOT qualify — DAY-SKIP: jump to the next WALL-midnight (00:00 of
+    // the next wall-day IN `zone`), computed from the wall clock we already have.
+    // Each distinct dead wall-day counts once against the lookahead bound.
+    //
+    // CRITICAL: we must advance to the next *wall*-midnight, NOT UTC-midnight. In a
+    // negative-offset zone (e.g. America/Los_Angeles, UTC-7/8) UTC-midnight is
+    // ~17:00 of the wall-day — so zeroing the UTC clock would land the cursor in
+    // the *evening* of a wall-day and the forward-only crawl could never reach that
+    // wall-day's MORNING (a `0 9 * * *` would be missed every cycle; a sparse-dom
+    // morning job would exhaust the search → null). Stepping by the minutes from
+    // THIS wall-instant to the next wall-midnight lands the cursor AT/BEFORE the
+    // morning of the next wall-day, so the crawl reaches it.
+    //
+    // `mins` is the minutes remaining in this wall-day plus one (to roll into the
+    // next day's 00:00): (23 - hour)*60 covers the whole hours left, +(60 - minute)
+    // covers the rest of this minute's hour AND the +1 minute to cross midnight.
+    // A DST transition can make the landing 23:00 or 01:00 of the wall-day rather
+    // than exactly 00:00 — harmless: the main loop self-corrects with a few minute
+    // steps, and it never lands at the evening (17:00) the UTC bump produced.
+    const skipKey = `${wc.month}-${wc.dom}`;
+    if (skipKey !== lastSkipKey) {
+      lastSkipKey = skipKey;
+      if (++daysTouched > MAX_LOOKAHEAD_DAYS) return null;
+    }
+    const mins = (23 - wc.hour) * 60 + (60 - wc.minute);
+    cursor.setUTCMinutes(cursor.getUTCMinutes() + mins);
+  }
+  return null;
+}

package/src/daemon-agent-def-api.test.ts

@@ -0,0 +1,166 @@
+/**
+ * Daemon route test for the vault-native agent-def RELOAD webhook
+ * (POST /api/vault/agent-def, design 2026-06-17-vault-native-agents Phase 4a).
+ *
+ * Auth mirrors /api/vault/inbound: hub JWT, scope agent:send, uniform-401. The
+ * AgentDefRegistry is injected with a recorder for its `reload`, so we assert the
+ * route's dispatch (auth → parse → route to vault → reload) without a real vault.
+ * Uses the same sentinel-token `mock.module("./hub-jwt.ts")` harness as the other
+ * daemon route tests (file-scoped).
+ */
+import { describe, test, expect, mock } from "bun:test";
+
+const SEND_TOKEN = "test-send-token"; // agent:send
+import { HubJwtError, looksLikeJwt } from "@openparachute/scope-guard";
+mock.module("./hub-jwt.ts", () => ({
+  AGENT_AUDIENCE: "agent",
+  CHANNEL_AUDIENCE: "channel",
+  async validateHubJwt(token: string) {
+    const base = { sub: "test", aud: "agent", jti: undefined, clientId: undefined, vaultScope: undefined };
+    if (token === SEND_TOKEN) return { ...base, scopes: ["agent:read", "agent:send"] };
+    throw new HubJwtError("issuer", "invalid token");
+  },
+  HubJwtError,
+  looksLikeJwt,
+  resetJwksCache() {},
+  resetRevocationCache() {},
+}));
+
+import { createFetchHandler } from "./daemon.ts";
+import { ClientRegistry } from "./routing.ts";
+import { AgentDefRegistry, type InstantiateDeps } from "./agent-defs.ts";
+import type { Channel } from "./registry.ts";
+
+/** A registry whose `reload` is recorded; one bound def-vault by default. */
+function recordingRegistry(opts?: { vaults?: string[] }) {
+  const reloads: Array<{ vault: string; noteId: string; event?: string }> = [];
+  const noopDeps: InstantiateDeps = {
+    ensureChannel: async () => {},
+    setupAndRegister: async () => {},
+    deregister: async () => true,
+    removeChannel: async () => true,
+  };
+  const reg = new AgentDefRegistry(noopDeps, {
+    bindings: (opts?.vaults ?? ["default"]).map((v) => ({ vault: v, token: "t" })),
+  });
+  // Override reload to record (avoid any vault I/O).
+  reg.reload = (async (vault: string, noteId: string, event?: "created" | "updated" | "deleted") => {
+    reloads.push({ vault, noteId, event });
+    return "instantiated";
+  }) as typeof reg.reload;
+  return { reg, reloads };
+}
+
+function serverWith(channels: Map<string, Channel>, agentDefs?: AgentDefRegistry) {
+  const registry = new ClientRegistry();
+  const srv = Bun.serve({
+    port: 0,
+    hostname: "127.0.0.1",
+    idleTimeout: 0,
+    fetch: createFetchHandler(channels, registry, agentDefs ? { agentDefs } : undefined),
+  });
+  return { srv, base: `http://127.0.0.1:${srv.port}` };
+}
+
+const emptyChannels = () => new Map<string, Channel>();
+const auth = { authorization: `Bearer ${SEND_TOKEN}`, "content-type": "application/json" };
+
+describe("POST /api/vault/agent-def", () => {
+  test("no Authorization → 401", async () => {
+    const { reg } = recordingRegistry();
+    const { srv, base } = serverWith(emptyChannels(), reg);
+    const res = await fetch(`${base}/api/vault/agent-def`, {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({ note: { id: "n" } }),
+    });
+    expect(res.status).toBe(401);
+    srv.stop();
+  });
+
+  test("authed reload routes to the registry with vault + noteId + event (single-vault default)", async () => {
+    const { reg, reloads } = recordingRegistry();
+    const { srv, base } = serverWith(emptyChannels(), reg);
+    const res = await fetch(`${base}/api/vault/agent-def`, {
+      method: "POST",
+      headers: auth,
+      body: JSON.stringify({ event: "updated", note: { id: "Agents/uni-dev" } }),
+    });
+    expect(res.status).toBe(200);
+    const body = (await res.json()) as { ok: boolean; reloaded: string };
+    expect(body.ok).toBe(true);
+    expect(body.reloaded).toBe("instantiated");
+    // vault defaulted to the sole bound def-vault.
+    expect(reloads).toEqual([{ vault: "default", noteId: "Agents/uni-dev", event: "updated" }]);
+    srv.stop();
+  });
+
+  // Connector 1: the two def-reload triggers (note.created + note.updated) the
+  // hub provisions send `event: "created"` / `event: "updated"` here. Both must
+  // route through to reload() — the webhook coercion accepts BOTH, not just one.
+  test("a note.created trigger payload routes to reload (created → instantiate)", async () => {
+    const { reg, reloads } = recordingRegistry();
+    const { srv, base } = serverWith(emptyChannels(), reg);
+    const res = await fetch(`${base}/api/vault/agent-def`, {
+      method: "POST",
+      headers: auth,
+      body: JSON.stringify({ event: "created", note: { id: "Agents/new-agent" } }),
+    });
+    expect(res.status).toBe(200);
+    const body = (await res.json()) as { ok: boolean; reloaded: string };
+    expect(body.ok).toBe(true);
+    expect(reloads).toEqual([{ vault: "default", noteId: "Agents/new-agent", event: "created" }]);
+    srv.stop();
+  });
+
+  test("explicit body.vault is honored", async () => {
+    const { reg, reloads } = recordingRegistry({ vaults: ["default", "research"] });
+    const { srv, base } = serverWith(emptyChannels(), reg);
+    const res = await fetch(`${base}/api/vault/agent-def`, {
+      method: "POST",
+      headers: auth,
+      body: JSON.stringify({ event: "deleted", vault: "research", note: { id: "Agents/r" } }),
+    });
+    expect(res.status).toBe(200);
+    expect(reloads).toEqual([{ vault: "research", noteId: "Agents/r", event: "deleted" }]);
+    srv.stop();
+  });
+
+  test("multiple def-vaults + no explicit vault → 400 (ambiguous)", async () => {
+    const { reg, reloads } = recordingRegistry({ vaults: ["default", "research"] });
+    const { srv, base } = serverWith(emptyChannels(), reg);
+    const res = await fetch(`${base}/api/vault/agent-def`, {
+      method: "POST",
+      headers: auth,
+      body: JSON.stringify({ note: { id: "n" } }),
+    });
+    expect(res.status).toBe(400);
+    expect(reloads).toHaveLength(0);
+    srv.stop();
+  });
+
+  test("missing note.id → 400", async () => {
+    const { reg } = recordingRegistry();
+    const { srv, base } = serverWith(emptyChannels(), reg);
+    const res = await fetch(`${base}/api/vault/agent-def`, {
+      method: "POST",
+      headers: auth,
+      body: JSON.stringify({ event: "created" }),
+    });
+    expect(res.status).toBe(400);
+    srv.stop();
+  });
+
+  test("no agentDefs configured → clean no-op ack (200, reloaded: skipped)", async () => {
+    const { srv, base } = serverWith(emptyChannels()); // no registry
+    const res = await fetch(`${base}/api/vault/agent-def`, {
+      method: "POST",
+      headers: auth,
+      body: JSON.stringify({ note: { id: "n" } }),
+    });
+    expect(res.status).toBe(200);
+    const body = (await res.json()) as { ok: boolean; reloaded: string };
+    expect(body.reloaded).toBe("skipped");
+    srv.stop();
+  });
+});