@sleep2agi/commhub-server 0.9.0-preview.0 → 0.9.0-preview.12

package/src/db.ts

@@ -117,12 +117,43 @@ db.exec(`
     created_at              TEXT NOT NULL DEFAULT (datetime('now'))
   );
 
-  CREATE INDEX IF NOT EXISTS idx_agent_telemetry_host_time
-    ON agent_telemetry(network_id, hostname, ip, created_at);
   CREATE INDEX IF NOT EXISTS idx_agent_telemetry_alias_time
     ON agent_telemetry(network_id, alias, created_at);
 `);
 
+// Round-2/4 review ② — server-health index split.
+//
+// The /api/server-health/:host query (index.ts) filters
+//   WHERE (hostname = ?1 OR ip = ?2) AND created_at >= ?3
+// The old combined index (network_id, hostname, ip, created_at) only
+// covers an equality on hostname AND ip (not OR), so SQLite couldn't
+// use it for the disjunction and fell back to a full scan of
+// agent_telemetry on every server-health page load. At a 30s heartbeat
+// × N agents, the table grows fast and the scan dominates.
+//
+// Split into two narrow indexes so SQLite's OR optimizer can run an
+// index UNION on (hostname-path ∪ ip-path), each path covered by an
+// index that also lets the created_at range be range-scanned.
+//
+// **Honest trade-off** (corrected after reviewer flag): secondary
+// index count goes from 2 (host_time + alias_time) to 3 (hostname_time
+// + ip_time + alias_time). Every agent_telemetry insert now maintains
+// ONE MORE B-tree, so per-insert write-amp is +50%. The trade-off is
+// deliberate: write-amp is bounded by the heartbeat cadence (30s ×
+// N agents) and is small in absolute terms, while the read-path win
+// is going from O(n) full-scan-per-dashboard-poll to O(log n) two-
+// branch index union. For the public-facing hub's expected load
+// shape (many readers, few writers), that's a clear net gain.
+try { db.exec("DROP INDEX IF EXISTS idx_agent_telemetry_host_time"); } catch {}
+try {
+  db.exec(`CREATE INDEX IF NOT EXISTS idx_agent_telemetry_hostname_time
+           ON agent_telemetry(network_id, hostname, created_at)`);
+} catch {}
+try {
+  db.exec(`CREATE INDEX IF NOT EXISTS idx_agent_telemetry_ip_time
+           ON agent_telemetry(network_id, ip, created_at)`);
+} catch {}
+
 // inbox: add in_reply_to, requires_response, expires_at, scope, node identity
 for (const col of [
   { name: "in_reply_to", def: "TEXT" },
@@ -246,6 +277,137 @@ try {
   if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
 }
 
+// RFC-024 (2026-06-28): node config-apply foundation.
+//
+// `nodes` table gains two columns:
+//   - `config_revision`: monotonically-incrementing per-node revision
+//     that gates write conflicts (dashboard sends base_revision, hub
+//     rejects 409 if it doesn't match current). Promoted by hub when
+//     the node ACKs `applied` from a successful apply.
+//   - `config_snapshot`: masked JSON of the node's current effective
+//     config, posted by the node via report_status. Dashboard reads
+//     this for the GET snapshot path (never touches per-node files).
+//
+// Both ALTERs wrapped in try/catch for the same reason as
+// must_change_password above (no IF NOT EXISTS on SQLite < 3.35).
+try {
+  db.exec(`ALTER TABLE nodes ADD COLUMN config_revision INTEGER DEFAULT 0`);
+} catch (e: any) {
+  if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
+}
+try {
+  db.exec(`ALTER TABLE nodes ADD COLUMN config_snapshot TEXT`);
+} catch (e: any) {
+  if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
+}
+
+// RFC-026 §9 / #338 PR2 — daemon self-declare fields promoted from
+// config_snapshot to first-class indexable columns. The hub
+// `list_host_supervisors` MCP tool reads these directly without parsing
+// the snapshot JSON on every list call.
+//
+// Both fields are JSON arrays of strings:
+//   runtimes_supported    e.g. ["claude-agent-sdk","codex-sdk","grok-build-acp"]
+//   allowed_secret_keys   e.g. ["ANTHROPIC_API_KEY","OPENAI_API_KEY"]
+//
+// Stored as TEXT (JSON-encoded) — SQLite has no native array type and
+// the wrapper code is small. Both nullable for pre-PR2 daemons; the
+// list endpoint falls back to [] when null.
+try {
+  db.exec(`ALTER TABLE nodes ADD COLUMN runtimes_supported TEXT`);
+} catch (e: any) {
+  if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
+}
+try {
+  db.exec(`ALTER TABLE nodes ADD COLUMN allowed_secret_keys TEXT`);
+} catch (e: any) {
+  if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
+}
+
+// RFC-027 §2.3 — state machine for stop/delete (D5). 'active' default
+// matches pre-RFC behavior for every existing row. Transitions:
+//   active → stopping → stopped → (back to active via restart_node)
+//   active → deleting → (row gone)
+//   stopped → deleting → (row gone)
+// Used as the inbox-enqueue gate per §2.3 race-free invariant:
+// pushEvent / INSERT INTO inbox must refuse a non-active target so the
+// SIGTERM-in-flight window never gets a new task.
+try {
+  db.exec(`ALTER TABLE nodes ADD COLUMN lifecycle_state TEXT DEFAULT 'active'`);
+} catch (e: any) {
+  if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
+}
+
+// RFC-027 §2 — stop/delete request envelope. Mirrors
+// node_create_requests structure so the daemon-side pull/ack pattern
+// is symmetric with create_node. action='stop'|'delete' picks which
+// branch (delete adds backup_path + revokes ntok on ack).
+db.exec(`
+  CREATE TABLE IF NOT EXISTS node_stop_requests (
+    request_id TEXT PRIMARY KEY,
+    network_id TEXT NOT NULL,
+    daemon_node_id TEXT NOT NULL,
+    child_node_id TEXT NOT NULL,
+    child_alias TEXT NOT NULL,
+    action TEXT NOT NULL CHECK (action IN ('stop', 'delete')),
+    delete_config INTEGER NOT NULL DEFAULT 1,
+    grace_seconds INTEGER NOT NULL DEFAULT 10,
+    force INTEGER NOT NULL DEFAULT 0,
+    in_flight_at_dispatch INTEGER NOT NULL DEFAULT 0,
+    created_by_token TEXT NOT NULL,
+    status TEXT NOT NULL DEFAULT 'pending',
+    error TEXT,
+    backup_path TEXT,
+    exit_signal TEXT,
+    created_at INTEGER NOT NULL,
+    delivered_at INTEGER,
+    acked_at INTEGER
+  )
+`);
+db.exec(`CREATE INDEX IF NOT EXISTS idx_stop_req_daemon ON node_stop_requests(daemon_node_id, status)`);
+db.exec(`CREATE INDEX IF NOT EXISTS idx_stop_req_child ON node_stop_requests(child_node_id)`);
+
+// `node_config_updates` — pending + history. One row per dashboard write.
+// At most one row per node may be in a non-terminal state at a time
+// (single-flight enforced at the tool layer; this table just stores).
+//
+// network_id is denormalized from nodes(network_id) at insert time so
+// queries don't need to JOIN through nodes every time + so a node
+// deletion doesn't orphan the history (audit retention).
+db.exec(`
+  CREATE TABLE IF NOT EXISTS node_config_updates (
+    update_id        TEXT PRIMARY KEY,
+    node_id          TEXT NOT NULL,
+    network_id       TEXT NOT NULL,
+    patch_json       TEXT NOT NULL,
+    apply_mode       TEXT NOT NULL,
+    base_revision    INTEGER NOT NULL,
+    status           TEXT NOT NULL,
+    error            TEXT,
+    created_at       INTEGER NOT NULL,
+    created_by_token TEXT NOT NULL,
+    acked_at         INTEGER,
+    new_revision     INTEGER
+  );
+
+  CREATE INDEX IF NOT EXISTS idx_ncu_node_status ON node_config_updates(node_id, status);
+  CREATE INDEX IF NOT EXISTS idx_ncu_network ON node_config_updates(network_id);
+`);
+
+// F-C (CHANGE_REQ): partial unique index on non-terminal rows. The
+// single-flight check in tools.ts is an app-level check-then-INSERT —
+// safe under single-process Bun, but two hub workers could race and
+// double-insert. This index makes that race impossible at the DB
+// layer regardless of process count. Wrapped in try/catch because
+// SQLite < 3.8.0 doesn't support partial indexes (rare on modern
+// systems; if missing, app-level check is the fallback).
+try {
+  db.exec(`CREATE UNIQUE INDEX IF NOT EXISTS uniq_ncu_node_inflight ON node_config_updates(node_id) WHERE status IN ('pending', 'restarting')`);
+} catch (e: any) {
+  // Older SQLite — fall back to app-level single-flight only.
+  console.warn(`[commhub] partial unique index uniq_ncu_node_inflight skipped: ${e?.message || e}`);
+}
+
 // ── V3: networks table ──
 db.exec(`
   CREATE TABLE IF NOT EXISTS networks (
@@ -280,6 +442,136 @@ db.exec(`
   CREATE INDEX IF NOT EXISTS idx_tokens_user ON api_tokens(user_id);
 `);
 
+// RFC-026 v4 §4.4.8 C4 — api_tokens metadata so sweeper can revoke
+// orphan child-ntoks without ever touching plaintext tokens. ALTER in
+// try/catch (idempotent for already-migrated DBs).
+//   request_id  — links a child token to the create_node request that
+//                 minted it; sweeper joins on this
+//   revoked_at  — explicit revoke; resolveToken treats non-NULL as
+//                 invalid even before expiry
+for (const ddl of [
+  "ALTER TABLE api_tokens ADD COLUMN request_id TEXT",
+  "ALTER TABLE api_tokens ADD COLUMN revoked_at TEXT",
+  "ALTER TABLE api_tokens ADD COLUMN role TEXT",
+]) {
+  try { db.exec(ddl); }
+  catch (e: any) {
+    if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
+  }
+}
+try { db.exec(`CREATE INDEX IF NOT EXISTS idx_tokens_request ON api_tokens(request_id)`); }
+catch (e: any) { /* index may already exist */ void e; }
+
+// RFC-026 v4 §2.5 + §4.4 — node_create_requests. Metadata only;
+// env_blob NEVER lives in this table (F1 mint-stream-evict — secret
+// blob is in-memory Map keyed by request_id, drained on daemon get,
+// TTL 60s GC).
+//
+// env_keys is a JSON array of secret KEY NAMES only (for audit). The
+// actual values flow through pendingEnvBlobs Map and never touch
+// disk.
+db.exec(`
+  CREATE TABLE IF NOT EXISTS node_create_requests (
+    request_id        TEXT PRIMARY KEY,
+    daemon_node_id    TEXT NOT NULL,
+    child_name        TEXT NOT NULL,
+    network_id        TEXT NOT NULL,
+    runtime           TEXT NOT NULL,
+    model             TEXT NOT NULL,
+    flags_json        TEXT NOT NULL,
+    env_keys          TEXT NOT NULL,
+    status            TEXT NOT NULL,
+    error             TEXT,
+    child_token_id    TEXT,
+    created_at        INTEGER NOT NULL,
+    created_by_token  TEXT NOT NULL,
+    delivered_at      INTEGER,
+    acked_at          INTEGER,
+    child_node_id     TEXT
+  );
+
+  CREATE INDEX IF NOT EXISTS idx_ncr_daemon_status ON node_create_requests(daemon_node_id, status);
+  CREATE INDEX IF NOT EXISTS idx_ncr_network ON node_create_requests(network_id);
+  CREATE INDEX IF NOT EXISTS idx_ncr_child_name ON node_create_requests(child_name);
+`);
+try {
+  // Single-flight per (daemon, child_name) — prevents racing dashboard
+  // create on same name. Mirrors RFC-024 uniq_ncu_node_inflight.
+  db.exec(`CREATE UNIQUE INDEX IF NOT EXISTS uniq_ncr_inflight ON node_create_requests(daemon_node_id, child_name) WHERE status IN ('pending', 'delivered')`);
+} catch (e: any) {
+  console.warn(`[commhub] partial unique index uniq_ncr_inflight skipped: ${e?.message || e}`);
+}
+
+// RFC-028 P1 v4 §2.2 — providers + provider_models + network_secrets +
+// probe_results. All 4 tables idempotent CREATE TABLE IF NOT EXISTS. F2
+// lazy gate: tables created on hub boot regardless; vault master key
+// (ANET_HUB_SECRET_VAULT_KEY env) only required when these tables get
+// their first non-empty row (lazy in vault.ts). Empty tables = hub
+// boots fine without env (critical for existing prod hub升级 不 brick).
+db.exec(`
+  CREATE TABLE IF NOT EXISTS providers (
+    provider_id     TEXT PRIMARY KEY,
+    network_id      TEXT NOT NULL,
+    name            TEXT NOT NULL,
+    vendor          TEXT NOT NULL,
+    base_url        TEXT NOT NULL,
+    secret_key_ref  TEXT NOT NULL,
+    created_at      INTEGER NOT NULL,
+    created_by      TEXT NOT NULL,
+    enabled         INTEGER NOT NULL DEFAULT 1,
+    UNIQUE(network_id, name)
+  );
+  CREATE INDEX IF NOT EXISTS idx_providers_network ON providers(network_id);
+`);
+db.exec(`
+  CREATE TABLE IF NOT EXISTS provider_models (
+    model_id        TEXT PRIMARY KEY,
+    provider_id     TEXT NOT NULL,
+    model_name      TEXT NOT NULL,
+    display_name    TEXT,
+    context_window  INTEGER,
+    supports_vision INTEGER NOT NULL DEFAULT 0,
+    enabled         INTEGER NOT NULL DEFAULT 1,
+    created_at      INTEGER NOT NULL,
+    UNIQUE(provider_id, model_name)
+  );
+  CREATE INDEX IF NOT EXISTS idx_models_provider ON provider_models(provider_id);
+`);
+// network_secrets — encrypted-at-rest vault. ciphertext + iv + tag are
+// AES-GCM output; plaintext NEVER touches this table. master key from
+// ANET_HUB_SECRET_VAULT_KEY env (lazy gate in vault.ts).
+db.exec(`
+  CREATE TABLE IF NOT EXISTS network_secrets (
+    network_id   TEXT NOT NULL,
+    key          TEXT NOT NULL,
+    ciphertext   BLOB NOT NULL,
+    iv           BLOB NOT NULL,
+    tag          BLOB NOT NULL,
+    created_at   INTEGER NOT NULL,
+    updated_at   INTEGER NOT NULL,
+    PRIMARY KEY (network_id, key)
+  );
+`);
+// probe_results — connectivity matrix history. v3 R3 lock: error_label
+// is HUB-DERIVED ONLY from (status, raw_status_code) — daemon CANNOT
+// submit arbitrary error string. status enum mirrors ProbeAckPayload.
+db.exec(`
+  CREATE TABLE IF NOT EXISTS probe_results (
+    probe_id        TEXT PRIMARY KEY,
+    provider_id     TEXT NOT NULL,
+    model_name      TEXT NOT NULL,
+    daemon_node_id  TEXT NOT NULL,
+    network_id      TEXT NOT NULL,
+    status          TEXT NOT NULL,
+    latency_ms      INTEGER,
+    error_label     TEXT,
+    probed_at       INTEGER NOT NULL,
+    probed_by_user  TEXT,
+    raw_status_code INTEGER
+  );
+  CREATE INDEX IF NOT EXISTS idx_probe_matrix ON probe_results(network_id, provider_id, model_name, daemon_node_id, probed_at DESC);
+`);
+
 // ── V3: audit_log table ──
 db.exec(`
   CREATE TABLE IF NOT EXISTS audit_log (
@@ -707,8 +999,179 @@ export function generateId(prefix: string): string {
   return `${prefix}_${crypto.randomUUID().replace(/-/g, "").slice(0, 12)}`;
 }
 
-export function hashPassword(password: string): string {
-  return new Bun.CryptoHasher("sha256").update(`anet:${password}`).digest("hex");
+// ──────────────────────────────────────────────────────────────────
+// Round-6 A1 — Password hashing: salted scrypt with lazy migration
+//
+// Format: `scrypt$<N>$<salt-b64>$<hash-b64>`
+//   N        = log2(cost). E.g. 14 → 2^14 = 16384 (~50ms on modern hw).
+//   salt-b64 = base64 of 16 random bytes.
+//   hash-b64 = base64 of 64-byte scrypt output.
+//
+// Why scrypt: Node built-in crypto.scryptSync — zero new deps.
+//   Memory-hard, ASIC-resistant. Better than bcrypt under modern GPU
+//   attack. argon2id would be marginally better but pulls a native
+//   dep (argon2 npm package) we don't want on the hub.
+//
+// Why "$" delimiter: legacy hashes are bare hex (0-9a-f only). Any
+//   "$" in stored hash means new format. Trivial detect with zero
+//   false-positives.
+//
+// Lazy migration:
+//   - register/bootstrap/changePassword/resetUserPassword always
+//     write the new format.
+//   - login + changePassword (old-verify) accept BOTH formats.
+//   - on successful login against a legacy hash, the caller rehashes
+//     in place. Zero downtime, no forced password change.
+//
+// See feedback_assume_unit_before_threshold.md: explicit-set
+// crypto parameters; don't rely on library defaults.
+// ──────────────────────────────────────────────────────────────────
+
+import { scryptSync, randomBytes, timingSafeEqual } from "node:crypto";
+
+// Default cost parameter (N = 2^14 = 16384). Tunable via env for
+// test suites that don't want to spend ~50ms per scrypt call.
+//   N=10 → 2^10 = 1024 iter, ~5ms (test-only)
+//   N=14 → 2^14 = 16384 iter, ~50ms (production)
+//   N=15 → 2^15 = 32768 iter, ~100ms (if hw improves and ops want headroom)
+const DEFAULT_SCRYPT_N = 14;
+const SCRYPT_R = 8;
+const SCRYPT_P = 1;
+const SCRYPT_KEYLEN = 64;
+const SCRYPT_SALT_BYTES = 16;
+// Bun's scrypt defaults maxmem to ~32 MiB which is too tight for
+// N ≥ 14 (128 * r * 2^N = ~16 MiB at r=8, N=14; default has no
+// headroom for the internal buffers). Set explicitly so a higher-N
+// upgrade later doesn't suddenly throw at runtime.
+const SCRYPT_MAXMEM = 128 * 1024 * 1024; // 128 MiB
+
+function getScryptN(): number {
+  const raw = process.env.COMMHUB_SCRYPT_N;
+  if (!raw) return DEFAULT_SCRYPT_N;
+  const n = parseInt(raw, 10);
+  // Bound: 8 (way-too-weak, only for unit tests) — 20 (DoS risk).
+  if (!Number.isFinite(n) || n < 8 || n > 20) return DEFAULT_SCRYPT_N;
+  return n;
+}
+
+export function hashPassword(plain: string): string {
+  const N = getScryptN();
+  const salt = randomBytes(SCRYPT_SALT_BYTES);
+  const hash = scryptSync(plain, salt, SCRYPT_KEYLEN, {
+    N: 1 << N,
+    r: SCRYPT_R,
+    p: SCRYPT_P,
+    maxmem: SCRYPT_MAXMEM,
+  });
+  return `scrypt$${N}$${salt.toString("base64")}$${hash.toString("base64")}`;
+}
+
+/**
+ * Verify a plaintext password against the stored hash. Accepts BOTH
+ * the new scrypt format and the legacy bare-hex SHA-256 format
+ * (back-compat for pre-A1 deployments).
+ *
+ * Returns:
+ *   ok           — true iff the password matches.
+ *   needsRehash  — true iff the stored hash was legacy format AND the
+ *                  password matched. The caller MUST rehash and write
+ *                  the new format back to disk so the legacy hash
+ *                  stops existing. Returned false for already-new
+ *                  hashes (no rehash needed even on N upgrade —
+ *                  follow-up if/when we bump default N).
+ */
+export function verifyPassword(plain: string, stored: string): { ok: boolean; needsRehash: boolean } {
+  // Format detection: new format always contains "$". Legacy bare
+  // sha256 hex never does (0-9a-f only).
+  if (stored.includes("$")) {
+    const parts = stored.split("$");
+    if (parts.length !== 4 || parts[0] !== "scrypt") {
+      // Unknown new-style scheme — be conservative, reject.
+      return { ok: false, needsRehash: false };
+    }
+    const N = parseInt(parts[1], 10);
+    if (!Number.isFinite(N) || N < 8 || N > 20) return { ok: false, needsRehash: false };
+    let saltBuf: Buffer;
+    let expectedBuf: Buffer;
+    try {
+      saltBuf = Buffer.from(parts[2], "base64");
+      expectedBuf = Buffer.from(parts[3], "base64");
+    } catch {
+      return { ok: false, needsRehash: false };
+    }
+    if (expectedBuf.length === 0) return { ok: false, needsRehash: false };
+    let actualBuf: Buffer;
+    try {
+      actualBuf = scryptSync(plain, saltBuf, expectedBuf.length, {
+        N: 1 << N,
+        r: SCRYPT_R,
+        p: SCRYPT_P,
+        maxmem: SCRYPT_MAXMEM,
+      });
+    } catch {
+      return { ok: false, needsRehash: false };
+    }
+    // timingSafeEqual requires equal-length buffers; we sized actual
+    // to expected.length above so this is safe.
+    const ok = timingSafeEqual(actualBuf, expectedBuf);
+    return { ok, needsRehash: false };
+  }
+
+  // Legacy bare-hex SHA-256 format (pre-A1).
+  const legacy = new Bun.CryptoHasher("sha256").update(`anet:${plain}`).digest("hex");
+  // Use timingSafeEqual on byte buffers (hex strings would be
+  // compared via JS === which is variable-time on first-differing
+  // char — the timing leak is tiny for fixed-format hex but pin the
+  // habit). Both are exactly 64 hex chars when sha256 produced.
+  if (stored.length !== legacy.length) {
+    // Length mismatch is itself indistinguishable from a parse miss
+    // — but we still pay the scrypt to keep this branch's timing
+    // peer to the verify branches below.
+    burnEqualTimeScrypt(plain);
+    return { ok: false, needsRehash: false };
+  }
+  let ok = false;
+  try {
+    ok = timingSafeEqual(Buffer.from(stored, "hex"), Buffer.from(legacy, "hex"));
+  } catch {
+    burnEqualTimeScrypt(plain);
+    return { ok: false, needsRehash: false };
+  }
+  // Timing-oracle close (round-6 round-2 — independent reviewer flag):
+  // legacy-OK   path callers (auth.ts:login) do `hashPassword(password)`
+  //             immediately after to rehash → one scrypt cost on top of
+  //             this function's return. ~50ms.
+  // legacy-WRONG path callers reject and return → ZERO additional cost.
+  //
+  // That asymmetry leaks "is this user on the legacy hash" via wall-
+  // clock difference (sub-ms wrong vs ~50ms ok). To close it, burn an
+  // equal-time scrypt on the wrong-password branch too. Now both
+  // legacy paths cost SHA-256 + one scrypt, matching new-format paths
+  // and the missing-user dummy-scrypt path.
+  if (!ok) {
+    burnEqualTimeScrypt(plain);
+  }
+  return { ok, needsRehash: ok };
+}
+
+// Round-6 round-2 hardening — throwaway scrypt to equalize timing on
+// failure paths. Burns the same wall-clock cost as a real
+// hashPassword() call. Result discarded; salt is a fresh random so
+// the work is real (not optimised away). Cost ~N=14 → 50ms,
+// COMMHUB_SCRYPT_N controlled like the real path.
+function burnEqualTimeScrypt(plain: string): void {
+  try {
+    const N = getScryptN();
+    scryptSync(plain, randomBytes(SCRYPT_SALT_BYTES), SCRYPT_KEYLEN, {
+      N: 1 << N,
+      r: SCRYPT_R,
+      p: SCRYPT_P,
+      maxmem: SCRYPT_MAXMEM,
+    });
+  } catch {
+    // Even on failure to scrypt (shouldn't happen), don't surface —
+    // this is a timing pacer, not a correctness path.
+  }
 }
 
 export function hashToken(token: string): string {
@@ -762,17 +1225,39 @@ export function logAudit(userId: string | null, username: string | null, action:
  * (e.g. server-internal bookkeeping that doesn't have a tenant
  * context) can omit `callerNetId`, but every MCP tool path MUST
  * supply it.
+ *
+ * Returns `{ chained, stoppedReason? }` so the caller can gate its
+ * follow-up SSE push on whether the chain actually wrote (round-2
+ * fix per #275 follow-up, 通信牛 catch): even when the WRITE was
+ * refused, callers were still running `SELECT parent.from_name;
+ * pushEvent(parent.from_name, { parent_task_id }, callerNetId)` —
+ * leaking the foreign parent's task_id into the caller's network
+ * via the SSE payload. Now the caller checks `result.chained` before
+ * pushing the chained_reply event.
  */
+export type ChainReplyResult = {
+  /** True iff this call actually wrote into at least one parent row.
+   *  When false, callers MUST NOT push SSE events for parent
+   *  listeners — the chain would otherwise leak the foreign parent's
+   *  identifiers into the caller's network. */
+  chained: boolean;
+  /** Optional reason the chain stopped short. Today the only enforced
+   *  reason is `cross_network`; future reasons (max-depth, parent-
+   *  archived, etc.) can land here without breaking callers. */
+  stoppedReason?: "cross_network";
+};
+
 export function chainReplyToParent(
   childTaskId: string,
   replyText: string,
   replyStatus: "replied" | "failed" | "cancelled" = "replied",
   maxDepth = 5,
   callerNetId?: string | null,
-): void {
+): ChainReplyResult {
   let currentChildId: string | null = childTaskId;
   let currentReply = replyText;
   let depth = 0;
+  let chained = false;
   // Normalize so undefined ("don't enforce") stays distinct from
   // null ("explicit default network"). Caller passes undefined to
   // opt out of the cross-tenant check; null to require the parent
@@ -788,12 +1273,12 @@ export function chainReplyToParent(
       "SELECT parent_task_id, to_name, from_name, content FROM tasks WHERE task_id = ?1",
       currentChildId
     );
-    if (!child?.parent_task_id) return;
+    if (!child?.parent_task_id) return { chained };
     const parent: ParentRow | null = db.get<ParentRow>(
       "SELECT task_id, from_name, to_name, status, result, network_id, parent_task_id FROM tasks WHERE task_id = ?1",
       child.parent_task_id
     );
-    if (!parent) return;
+    if (!parent) return { chained };
 
     // round5 F2 fix: refuse to write into a parent that lives in a
     // different network than the caller. Stops chain-reply being a
@@ -803,7 +1288,7 @@ export function chainReplyToParent(
       const parentNet = parent.network_id ?? null;
       if (parentNet !== callerNorm) {
         console.log(`[commhub] 🚫 chainReplyToParent cross-network blocked: parent=${parent.task_id.slice(0, 8)} parent-net=${parentNet ?? "null"} caller-net=${callerNorm ?? "null"}`);
-        return;
+        return { chained, stoppedReason: "cross_network" };
       }
     }
 
@@ -844,10 +1329,16 @@ export function chainReplyToParent(
       } catch {}
     }
 
+    // This iteration actually wrote a parent row (and possibly an
+    // inbox notification). Record that so the caller can gate its
+    // SSE push on a real chain.
+    chained = true;
+
     // Recurse up the chain.
     currentChildId = parent.task_id;
     currentReply = newResult;
   }
+  return { chained };
 }
 
 export function logTaskEvent(taskId: string, fromStatus: string | null, toStatus: string, actor: string, detail?: string) {