@sleep2agi/commhub-server 0.9.0-preview.3 → 0.9.0-preview.31

package/src/db.ts

@@ -89,6 +89,8 @@ for (const col of [
   { name: "process_cpu_pct", def: "REAL" },
   { name: "process_uptime_seconds", def: "REAL" },
   { name: "process_in_flight_count", def: "INTEGER" },
+  { name: "external_schedules", def: "TEXT" },
+  { name: "peer_reply_inbox_capable", def: "INTEGER NOT NULL DEFAULT 0" },
 ]) {
   try { db.exec(`ALTER TABLE sessions ADD COLUMN ${col.name} ${col.def}`); } catch {}
 }
@@ -162,6 +164,11 @@ for (const col of [
   { name: "scope", def: "TEXT DEFAULT 'single'" },
   { name: "meta_json", def: "TEXT" },
   { name: "node_id", def: "TEXT" },
+  // #520 — explicit logical task identity.  Initial deliveries historically
+  // used inbox.id == tasks.task_id, while retry/reassign generate a fresh
+  // inbox row id.  Keep transport-row identity and logical-task identity
+  // separate so runtime evidence/replies survive those redeliveries.
+  { name: "task_id", def: "TEXT" },
 ]) {
   try { db.exec(`ALTER TABLE inbox ADD COLUMN ${col.name} ${col.def}`); } catch {}
 }
@@ -191,6 +198,8 @@ db.exec(`
     created_at        TEXT NOT NULL DEFAULT (datetime('now')),
     delivered_at      TEXT,
     started_at        TEXT,
+    runtime_submitted_at TEXT,
+    consumed_at       TEXT,
     completed_at      TEXT,
     expires_at        TEXT
   );
@@ -237,6 +246,8 @@ db.exec(`
     task_id       TEXT NOT NULL,
     from_status   TEXT,
     to_status     TEXT NOT NULL,
+    event_type    TEXT,
+    event_key     TEXT,
     actor         TEXT NOT NULL DEFAULT 'system',
     detail        TEXT,
     created_at    TEXT NOT NULL DEFAULT (datetime('now'))
@@ -301,6 +312,124 @@ try {
   if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
 }
 
+// RFC-036 / B4 — immutable per-node owner. New nodes receive this value in
+// the same transaction that mints their node token. Heartbeats may verify the
+// binding but must never first-claim or replace it. Legacy NULL rows remain
+// read-only for external-schedule writes.
+try {
+  db.exec(`ALTER TABLE nodes ADD COLUMN owner_user_id 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;
+}
+
+// #462 — user-set custom avatar (nullable http/https image URL). Pure
+// display attribute: deliberately NOT part of the RFC-024 config-apply
+// pipeline (nothing for agent-node to consume, no revision/ack). Written
+// only via PUT /api/nodes/:ref/avatar which validates the URL shape.
+try {
+  db.exec(`ALTER TABLE nodes ADD COLUMN avatar_url TEXT`);
+} catch (e: any) {
+  if (!/duplicate column|already exists/i.test(e?.message || "")) throw e;
+}
+
+// Node DISPLAY attributes (Vincent 目标第 3 条 — dashboard/app 多维度编辑).
+// Same posture as avatar_url above: hub-side presentation only, NOT part of
+// the RFC-024 config-apply pipeline (agent-node consumes none of them, so a
+// doorbell/ack round-trip would be meaningless and could wedge the config
+// revision). Written only via PUT /api/nodes/:ref/attrs.
+//
+//   display_name   — human label shown in the UI. Distinct from `alias`,
+//                    which is the ROUTING key (messages address the alias)
+//                    and from `node_name`, which is an addressing key
+//                    (`WHERE node_id OR node_name OR alias`). Renaming
+//                    either of those must go through the rename 2PC
+//                    (/api/node-rename/*), never through an attrs write.
+//   team, tags     — organisation metadata; tags stored as JSON array text.
+//   attrs_revision — optimistic-lock counter for these display fields ONLY.
+//                    Deliberately separate from config_revision so that
+//                    editing a tag never bumps the node's *config* revision
+//                    (which would perturb the config ack protocol).
+//
+// All nullable / defaulted so the ~100 rows already in flight keep reading.
+for (const ddl of [
+  `ALTER TABLE nodes ADD COLUMN display_name TEXT`,
+  `ALTER TABLE nodes ADD COLUMN team TEXT`,
+  `ALTER TABLE nodes ADD COLUMN tags TEXT`,
+  `ALTER TABLE nodes ADD COLUMN attrs_revision INTEGER NOT NULL DEFAULT 0`,
+]) {
+  try {
+    db.exec(ddl);
+  } 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).
@@ -387,6 +516,7 @@ 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",
+  "ALTER TABLE api_tokens ADD COLUMN bound_node_id TEXT",
 ]) {
   try { db.exec(ddl); }
   catch (e: any) {
@@ -395,6 +525,8 @@ for (const ddl of [
 }
 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; }
+try { db.exec(`CREATE INDEX IF NOT EXISTS idx_tokens_bound_node ON api_tokens(bound_node_id)`); }
+catch (e: any) { 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
@@ -436,6 +568,76 @@ try {
   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 (
@@ -456,6 +658,37 @@ db.exec(`
   CREATE INDEX IF NOT EXISTS idx_audit_network ON audit_log(network_id);
 `);
 
+// SkillHub — immutable, network-scoped skill submissions. A (network, slug,
+// version) tuple is write-once: retries with the same content are idempotent,
+// while different content must use a new version. Node submissions remain
+// pending until an owner/admin explicitly publishes them.
+db.exec(`
+  CREATE TABLE IF NOT EXISTS skillhub_skills (
+    skill_id          TEXT PRIMARY KEY,
+    network_id        TEXT NOT NULL,
+    slug              TEXT NOT NULL,
+    name              TEXT NOT NULL,
+    description       TEXT NOT NULL DEFAULT '',
+    version           TEXT NOT NULL,
+    content           TEXT NOT NULL,
+    content_hash      TEXT NOT NULL,
+    status            TEXT NOT NULL DEFAULT 'pending' CHECK(status IN ('pending', 'published', 'rejected', 'archived')),
+    source_type       TEXT NOT NULL CHECK(source_type IN ('node', 'user')),
+    source_alias      TEXT,
+    created_by_user   TEXT,
+    reviewed_by_user  TEXT,
+    review_note       TEXT,
+    created_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at        TEXT NOT NULL DEFAULT (datetime('now')),
+    reviewed_at       TEXT,
+    UNIQUE(network_id, slug, version)
+  );
+  CREATE INDEX IF NOT EXISTS idx_skillhub_network_status
+    ON skillhub_skills(network_id, status, updated_at DESC);
+  CREATE INDEX IF NOT EXISTS idx_skillhub_network_slug
+    ON skillhub_skills(network_id, slug, version);
+`);
+
 // ── V3: licenses table ──
 db.exec(`
   CREATE TABLE IF NOT EXISTS licenses (
@@ -566,6 +799,14 @@ for (const table of ["sessions", "nodes", "tasks", "inbox", "task_events", "comp
   try { db.exec(`ALTER TABLE ${table} ADD COLUMN network_id TEXT`); } catch {}
 }
 
+// #167 — stable lifecycle event names plus an internal idempotency key.
+// Ordinary status transitions keep event_key NULL, so retries and repeated
+// transitions remain auditable. Watcher warnings set a stable key; the unique
+// index makes every (task, threshold) write-once across patrol rounds/workers.
+try { db.exec("ALTER TABLE task_events ADD COLUMN event_type TEXT"); } catch {}
+try { db.exec("ALTER TABLE task_events ADD COLUMN event_key TEXT"); } catch {}
+try { db.exec("CREATE UNIQUE INDEX IF NOT EXISTS idx_task_events_task_key ON task_events(task_id, event_key)"); } catch {}
+
 // PR-1 (#146): durable node identity on historical inbox/tasks rows. Some
 // existing databases created `tasks` before from_node_id/to_node_id were in the
 // CREATE TABLE statement, so keep these ALTERs explicit and idempotent.
@@ -852,6 +1093,167 @@ try { db.exec("CREATE INDEX IF NOT EXISTS idx_completions_network ON completions
 // the reply up the chain via parent_task_id.
 try { db.exec("ALTER TABLE tasks ADD COLUMN parent_task_id TEXT"); } catch {}
 try { db.exec("ALTER TABLE tasks ADD COLUMN meta_json TEXT"); } catch {}
+// #520 — two monotonic evidence levels for this exact task:
+// runtime_submitted_at means agent-node handed the body to the vendor runtime;
+// consumed_at requires an attributable turn-start/first-activity signal.
+// Both stay separate from delivered_at (enqueue), acked (process receipt), and
+// started_at (legacy report_status/content matching). Old binaries ignore them.
+try { db.exec("ALTER TABLE tasks ADD COLUMN runtime_submitted_at TEXT"); } catch {}
+try { db.exec("ALTER TABLE tasks ADD COLUMN consumed_at TEXT"); } catch {}
+
+// #1181 durable outbound terminal cursor.  A task's created_at cannot be used
+// as a recovery watermark because an old task may become terminal after newer
+// tasks.  Record terminal transitions in their own monotonically increasing
+// journal instead.  The journal is append-only for a logical task; retry keeps
+// the existing task identity and therefore the same idempotency domain.
+db.exec(`
+  CREATE TABLE IF NOT EXISTS task_terminal_events (
+    terminal_seq INTEGER PRIMARY KEY AUTOINCREMENT,
+    task_id TEXT NOT NULL UNIQUE,
+    network_id TEXT,
+    from_node_id TEXT,
+    completed_at TEXT NOT NULL DEFAULT (datetime('now'))
+  );
+  CREATE INDEX IF NOT EXISTS idx_task_terminal_events_sender_seq
+    ON task_terminal_events(network_id, from_node_id, terminal_seq);
+`);
+if (db.dialect === "sqlite") {
+  db.exec(`
+    CREATE TRIGGER IF NOT EXISTS tasks_terminal_event_insert
+    AFTER INSERT ON tasks
+    WHEN NEW.status IN ('replied', 'failed', 'cancelled', 'expired')
+    BEGIN
+      INSERT OR IGNORE INTO task_terminal_events (task_id, network_id, from_node_id, completed_at)
+      VALUES (NEW.task_id, NEW.network_id, NEW.from_node_id, COALESCE(NEW.completed_at, datetime('now')));
+    END;
+    CREATE TRIGGER IF NOT EXISTS tasks_terminal_event_update
+    AFTER UPDATE OF status ON tasks
+    WHEN NEW.status IN ('replied', 'failed', 'cancelled', 'expired')
+      AND OLD.status NOT IN ('replied', 'failed', 'cancelled', 'expired')
+    BEGIN
+      INSERT OR IGNORE INTO task_terminal_events (task_id, network_id, from_node_id, completed_at)
+      VALUES (NEW.task_id, NEW.network_id, NEW.from_node_id, COALESCE(NEW.completed_at, datetime('now')));
+    END;
+  `);
+  db.run(`INSERT OR IGNORE INTO task_terminal_events (task_id, network_id, from_node_id, completed_at)
+          SELECT task_id, network_id, from_node_id, COALESCE(completed_at, created_at)
+            FROM tasks
+           WHERE status IN ('replied', 'failed', 'cancelled', 'expired')
+           ORDER BY COALESCE(completed_at, created_at), task_id`);
+} else {
+  db.exec(`
+    CREATE OR REPLACE FUNCTION record_task_terminal_event() RETURNS trigger AS $$
+    BEGIN
+      IF NEW.status IN ('replied', 'failed', 'cancelled', 'expired')
+         AND (TG_OP = 'INSERT' OR OLD.status NOT IN ('replied', 'failed', 'cancelled', 'expired')) THEN
+        INSERT INTO task_terminal_events (task_id, network_id, from_node_id, completed_at)
+        VALUES (NEW.task_id, NEW.network_id, NEW.from_node_id, COALESCE(NEW.completed_at, NOW()))
+        ON CONFLICT (task_id) DO NOTHING;
+      END IF;
+      RETURN NEW;
+    END;
+    $$ LANGUAGE plpgsql;
+    DROP TRIGGER IF EXISTS tasks_terminal_event_insert ON tasks;
+    DROP TRIGGER IF EXISTS tasks_terminal_event_update ON tasks;
+    CREATE TRIGGER tasks_terminal_event_insert AFTER INSERT ON tasks
+      FOR EACH ROW EXECUTE FUNCTION record_task_terminal_event();
+    CREATE TRIGGER tasks_terminal_event_update AFTER UPDATE OF status ON tasks
+      FOR EACH ROW EXECUTE FUNCTION record_task_terminal_event();
+    INSERT INTO task_terminal_events (task_id, network_id, from_node_id, completed_at)
+    SELECT task_id, network_id, from_node_id, COALESCE(completed_at, created_at)
+      FROM tasks WHERE status IN ('replied', 'failed', 'cancelled', 'expired')
+    ON CONFLICT (task_id) DO NOTHING;
+  `);
+}
+
+// ── Hub scheduled tasks ──────────────────────────────────────────────
+// Scheduling is a Hub concern: clients manage these rows, while agents only
+// receive the ordinary inbox/tasks rows produced for each occurrence.  Keep
+// the occurrence table separate so retries, skips and failures remain
+// auditable without overloading the task lifecycle.
+db.exec(`
+  CREATE TABLE IF NOT EXISTS scheduled_tasks (
+    schedule_id      TEXT PRIMARY KEY,
+    network_id       TEXT NOT NULL,
+    created_by       TEXT,
+    name             TEXT NOT NULL,
+    target_node_id   TEXT NOT NULL,
+    target_alias     TEXT NOT NULL,
+    task_content     TEXT NOT NULL,
+    priority         TEXT NOT NULL DEFAULT 'normal',
+    schedule_type    TEXT NOT NULL,
+    schedule_json    TEXT NOT NULL,
+    timezone         TEXT NOT NULL DEFAULT 'UTC',
+    overlap_policy   TEXT NOT NULL DEFAULT 'skip',
+    misfire_policy   TEXT NOT NULL DEFAULT 'catch_up_once',
+    status           TEXT NOT NULL DEFAULT 'active',
+    next_run_at      TEXT,
+    last_run_at      TEXT,
+    revision         INTEGER NOT NULL DEFAULT 1,
+    created_at       TEXT NOT NULL DEFAULT (datetime('now')),
+    updated_at       TEXT NOT NULL DEFAULT (datetime('now'))
+  );
+
+  CREATE INDEX IF NOT EXISTS idx_scheduled_tasks_due
+    ON scheduled_tasks(status, next_run_at);
+  CREATE INDEX IF NOT EXISTS idx_scheduled_tasks_network
+    ON scheduled_tasks(network_id, updated_at DESC);
+  CREATE INDEX IF NOT EXISTS idx_scheduled_tasks_target
+    ON scheduled_tasks(network_id, target_node_id);
+
+  CREATE TABLE IF NOT EXISTS scheduled_task_runs (
+    run_id           TEXT PRIMARY KEY,
+    schedule_id      TEXT NOT NULL,
+    network_id       TEXT NOT NULL,
+    scheduled_for    TEXT NOT NULL,
+    task_id          TEXT,
+    status           TEXT NOT NULL,
+    error_code       TEXT,
+    error_message    TEXT,
+    created_at       TEXT NOT NULL DEFAULT (datetime('now')),
+    completed_at     TEXT,
+    UNIQUE(schedule_id, scheduled_for)
+  );
+
+  CREATE INDEX IF NOT EXISTS idx_scheduled_runs_schedule
+    ON scheduled_task_runs(schedule_id, created_at DESC);
+  CREATE INDEX IF NOT EXISTS idx_scheduled_runs_network
+    ON scheduled_task_runs(network_id, created_at DESC);
+`);
+try { db.exec("ALTER TABLE scheduled_tasks ADD COLUMN misfire_policy TEXT NOT NULL DEFAULT 'catch_up_once'"); } catch {}
+
+// RFC-036 / B4 — owner-authorized edits of node-host managed schedules.
+// This is intentionally separate from Hub scheduled_tasks: these rows are
+// one-shot control intents consumed by the exact token-bound node. The partial
+// unique index makes the single-flight promise survive multiple Hub workers.
+db.exec(`
+  CREATE TABLE IF NOT EXISTS external_schedule_edits (
+    intent_id          TEXT PRIMARY KEY,
+    network_id         TEXT NOT NULL,
+    node_id            TEXT NOT NULL,
+    schedule_id        TEXT NOT NULL,
+    base_revision      INTEGER NOT NULL,
+    patch_json         TEXT NOT NULL,
+    status             TEXT NOT NULL DEFAULT 'pending',
+    expires_at         INTEGER NOT NULL,
+    created_at         INTEGER NOT NULL,
+    delivered_at       INTEGER,
+    acked_at           INTEGER,
+    created_by_user    TEXT NOT NULL,
+    created_by_token   TEXT NOT NULL,
+    consumed_by_token  TEXT,
+    result_revision    INTEGER,
+    error_code         TEXT
+  );
+
+  CREATE INDEX IF NOT EXISTS idx_external_schedule_edits_node
+    ON external_schedule_edits(network_id, node_id, status, created_at);
+  CREATE INDEX IF NOT EXISTS idx_external_schedule_edits_owner
+    ON external_schedule_edits(created_by_user, created_at DESC);
+  CREATE UNIQUE INDEX IF NOT EXISTS idx_external_schedule_edits_singleflight
+    ON external_schedule_edits(node_id, schedule_id)
+    WHERE status IN ('pending', 'delivered');
+`);
 try { db.exec("CREATE INDEX IF NOT EXISTS idx_tasks_parent ON tasks(parent_task_id)"); } catch {}
 
 // Helpers
@@ -1160,38 +1562,40 @@ export function chainReplyToParent(
     const marker = `\n\n[via ${childAlias} 子任务结果]\n${currentReply}`;
     const newResult = parent.result ? parent.result + marker : `[via ${childAlias} 子任务结果]\n${currentReply}`;
 
-    // Bump parent status to replied if still open. If it was already replied
-    // (e.g. 指挥室 sent an early status update), we still want to update the
-    // result so the dashboard can render the final chained answer — but keep
-    // status as replied (it was already terminal).
-    if (parent.status === "delivered" || parent.status === "acked" || parent.status === "running" || parent.status === "created") {
-      db.run(
-        "UPDATE tasks SET status = ?1, result = ?2, completed_at = datetime('now') WHERE task_id = ?3",
-        [replyStatus, newResult.slice(0, 8000), parent.task_id]
-      );
-      logTaskEvent(parent.task_id, parent.status, replyStatus, "auto-chain", `from ${childAlias}`);
-    } else {
-      db.run(
-        "UPDATE tasks SET result = ?1, completed_at = datetime('now') WHERE task_id = ?2",
-        [newResult.slice(0, 8000), parent.task_id]
-      );
-      logTaskEvent(parent.task_id, parent.status, parent.status, "auto-chain-append", `from ${childAlias}`);
-    }
-
-    if (parent.from_name && parent.from_name !== "hub" && parent.from_name !== "api") {
-      try {
-        const notifyId = `chain_${crypto.randomUUID().replace(/-/g, "").slice(0, 16)}`;
-        const notifyNode = db.get<{ node_id: string | null }>(
-          "SELECT node_id FROM sessions WHERE alias = ?1 AND COALESCE(network_id, 'default') = COALESCE(?2, 'default') ORDER BY updated_at DESC LIMIT 1",
-          [parent.from_name, parent.network_id ?? null]
+    db.transaction(() => {
+      // Bump parent status to replied if still open. The task transition and
+      // its scheduler-run mirror share this transaction, so a crash cannot
+      // leave one terminal while the other remains delivered.
+      if (parent.status === "delivered" || parent.status === "acked" || parent.status === "running" || parent.status === "created") {
+        db.run(
+          "UPDATE tasks SET status = ?1, result = ?2, completed_at = datetime('now') WHERE task_id = ?3",
+          [replyStatus, newResult.slice(0, 8000), parent.task_id]
         );
+        syncScheduledRunForTask(parent.task_id, parent.network_id);
+        logTaskEvent(parent.task_id, parent.status, replyStatus, "auto-chain", `from ${childAlias}`);
+      } else {
         db.run(
-          `INSERT INTO inbox (id, session_name, node_id, type, priority, content, from_session, in_reply_to, requires_response, network_id)
-           VALUES (?1, ?2, ?3, 'reply', 'normal', ?4, ?5, ?6, 'none', ?7)`,
-          [notifyId, parent.from_name, notifyNode?.node_id ?? null, `[${childAlias} 子任务完成]\n${currentReply.slice(0, 4000)}`, parent.to_name, parent.task_id, parent.network_id ?? null]
+          "UPDATE tasks SET result = ?1, completed_at = datetime('now') WHERE task_id = ?2",
+          [newResult.slice(0, 8000), parent.task_id]
         );
-      } catch {}
-    }
+        logTaskEvent(parent.task_id, parent.status, parent.status, "auto-chain-append", `from ${childAlias}`);
+      }
+
+      if (parent.from_name && parent.from_name !== "hub" && parent.from_name !== "api") {
+        try {
+          const notifyId = `chain_${crypto.randomUUID().replace(/-/g, "").slice(0, 16)}`;
+          const notifyNode = db.get<{ node_id: string | null }>(
+            "SELECT node_id FROM sessions WHERE alias = ?1 AND COALESCE(network_id, 'default') = COALESCE(?2, 'default') ORDER BY updated_at DESC LIMIT 1",
+            [parent.from_name, parent.network_id ?? null]
+          );
+          db.run(
+            `INSERT INTO inbox (id, session_name, node_id, type, priority, content, from_session, in_reply_to, requires_response, network_id)
+             VALUES (?1, ?2, ?3, 'reply', 'normal', ?4, ?5, ?6, 'none', ?7)`,
+            [notifyId, parent.from_name, notifyNode?.node_id ?? null, `[${childAlias} 子任务完成]\n${currentReply.slice(0, 4000)}`, parent.to_name, parent.task_id, parent.network_id ?? null]
+          );
+        } catch {}
+      }
+    });
 
     // This iteration actually wrote a parent row (and possibly an
     // inbox notification). Record that so the caller can gate its
@@ -1205,12 +1609,98 @@ export function chainReplyToParent(
   return { chained };
 }
 
+const SCHEDULED_RUN_TERMINAL_STATUSES = new Set(["replied", "failed", "cancelled", "expired"]);
+
+type ScheduledTaskLifecycleRow = {
+  task_id: string;
+  network_id: string | null;
+  status: string;
+  completed_at: string | null;
+};
+
+type ScheduledTaskBinding = { run_id: string; schedule_id: string };
+
+/**
+ * Mirror an exact scheduler-created task's lifecycle into its run row.
+ *
+ * The status is always read back from `tasks`; callers cannot supply one.
+ * The scheduler's storage binding is the four-part tuple
+ * (run_id, schedule_id, task_id, network_id). It is read from
+ * `scheduled_task_runs`, not caller-controlled task metadata (which may also
+ * legitimately be replaced by reply attachment metadata).
+ *
+ * Terminal states close the run. `retry_task` calls this after resetting the
+ * same logical task_id to delivered, which deliberately reopens that run so
+ * the history reflects the current retry rather than the first failed attempt.
+ */
+export function syncScheduledRunForTask(taskId: string, expectedNetworkId?: string | null): { matched: boolean; status?: string } {
+  const task = db.get<ScheduledTaskLifecycleRow>(
+    `SELECT task_id, network_id, status, completed_at
+       FROM tasks
+      WHERE task_id = ?1
+        AND (?2 IS NULL OR network_id = ?2)`,
+    taskId,
+    expectedNetworkId ?? null,
+  );
+  if (!task?.network_id) return { matched: false };
+  const bindings = db.all<ScheduledTaskBinding>(
+    `SELECT run_id, schedule_id FROM scheduled_task_runs
+      WHERE task_id = ?1 AND network_id = ?2`,
+    task.task_id,
+    task.network_id,
+  );
+  // One task belongs to at most one scheduled occurrence. Fail closed if a
+  // corrupt/legacy database violates that invariant rather than fanning one
+  // lifecycle transition into multiple runs.
+  if (bindings.length !== 1) return { matched: false };
+  const binding = bindings[0];
+
+  if (SCHEDULED_RUN_TERMINAL_STATUSES.has(task.status)) {
+    const errorCode = task.status === "replied" ? null : `task_${task.status}`;
+    const updated = db.run(
+      `UPDATE scheduled_task_runs
+          SET status = ?1,
+              error_code = ?2,
+              error_message = NULL,
+              completed_at = COALESCE(?3, datetime('now'))
+        WHERE run_id = ?4 AND task_id = ?5 AND network_id = ?6 AND schedule_id = ?7`,
+      [task.status, errorCode, task.completed_at, binding.run_id, task.task_id, task.network_id, binding.schedule_id],
+    );
+    return { matched: updated.changes === 1, status: task.status };
+  }
+
+  if (task.status === "delivered") {
+    const updated = db.run(
+      `UPDATE scheduled_task_runs
+          SET status = 'delivered', error_code = NULL, error_message = NULL, completed_at = NULL
+        WHERE run_id = ?1 AND task_id = ?2 AND network_id = ?3 AND schedule_id = ?4`,
+      [binding.run_id, task.task_id, task.network_id, binding.schedule_id],
+    );
+    return { matched: updated.changes === 1, status: task.status };
+  }
+
+  return { matched: false };
+}
+
+function taskEventTypeForStatus(toStatus: string): string {
+  switch (toStatus) {
+    case "delivered": return "task.send.delivered";
+    case "acked": return "task.ack";
+    case "running": return "task.started";
+    case "replied": return "task.replied";
+    case "expired": return "task.expired";
+    case "failed":
+    case "cancelled": return "task.failed";
+    default: return `task.status.${toStatus}`;
+  }
+}
+
 export function logTaskEvent(taskId: string, fromStatus: string | null, toStatus: string, actor: string, detail?: string) {
   try {
     db.run(
-      `INSERT INTO task_events (task_id, from_status, to_status, actor, detail, network_id)
-       VALUES (?1, ?2, ?3, ?4, ?5, (SELECT network_id FROM tasks WHERE task_id = ?1))`,
-      [taskId, fromStatus, toStatus, actor, detail ?? null]
+      `INSERT INTO task_events (task_id, from_status, to_status, event_type, actor, detail, network_id)
+       VALUES (?1, ?2, ?3, ?4, ?5, ?6, (SELECT network_id FROM tasks WHERE task_id = ?1))`,
+      [taskId, fromStatus, toStatus, taskEventTypeForStatus(toStatus), actor, detail ?? null]
     );
   } catch {}
 }