@aitne/daemon 0.1.10 → 0.1.11

package/dist/db/schema.js

@@ -107,7 +107,7 @@ CREATE TABLE IF NOT EXISTS agent_actions (
     --   'trigger'      — fired by an automation_triggers row
     --   'message'      — owner DM / mention / dashboard chat
     --   'cron'         — built-in routine (morning, hourly, evening, ...)
-    --   'observation'  — hourly-check consumed pending observations
+    --   'observation'  — activity-scan consumed pending observations
     --   'manual'       — Run-now / dashboard-driven invocation
     --   NULL           — legacy / not yet classified
     -- source_ref is the id of the upstream entity (e.g. trigger id).
@@ -447,9 +447,9 @@ CREATE TABLE IF NOT EXISTS observations (
     -- cost-reduction-structural section A -- pre-summarization at insert time.
     -- The summarizer worker (observers/observation-summarizer.ts) drains
     -- pending rows asynchronously and populates these columns; the
-    -- hourly_check skill consumes the summary instead of fetching raw
+    -- activity_scan skill consumes the summary instead of fetching raw
     -- content unless novelty_score >= 2. NULL novelty_score + non-
-    -- 'done' status means hourly_check falls back to legacy fetch-on-doubt.
+    -- 'done' status means activity_scan falls back to legacy fetch-on-doubt.
     summary_text TEXT,
     novelty_score INTEGER CHECK (novelty_score IS NULL OR (novelty_score >= 0 AND novelty_score <= 3)),
     summary_at TEXT,
@@ -1217,7 +1217,16 @@ CREATE TABLE IF NOT EXISTS browser_research_clusters (
     last_wiki_offer_at INTEGER,
     research_offer_accepted_at INTEGER,
     wiki_summary_written_at INTEGER,
-    agent_summary_revision INTEGER DEFAULT 0
+    agent_summary_revision INTEGER DEFAULT 0,
+    -- RESEARCH_CLUSTER_COST_FIX_PLAN.md F1 — local agent-day label
+    -- ('YYYY-MM-DD', see getAgentDayDateStr) on which the day-boundary
+    -- fan-out last enqueued routine.research_cluster_update for this
+    -- cluster. Stamped BEFORE the EventBus put so a replayed
+    -- day-boundary callback (wake catch-up fires on every sleep gap
+    -- >= 5 min) cannot re-enqueue the same cluster within one agent
+    -- day. NULL = never enqueued. Pre-existing installs get the column
+    -- via migration 0011.
+    journal_update_enqueued_on TEXT
 );
 
 CREATE TABLE IF NOT EXISTS browser_pending_offers (
@@ -1651,7 +1660,12 @@ CREATE TABLE IF NOT EXISTS browser_task (
         CHECK (extract_chars_total >= 0),
     created_at                  INTEGER NOT NULL,
     started_at                  INTEGER,
-    finished_at                 INTEGER
+    finished_at                 INTEGER,
+    -- BACKGROUND_TASK_RUNNER_DESIGN.md Phase 1 — browser-task reports are
+    -- delivered through the shared task.delivery boundary. This timestamp
+    -- is the recovery key for completed reports whose worker stored the
+    -- report before the DM was sent/recorded.
+    delivered_at                INTEGER
 );
 CREATE INDEX IF NOT EXISTS idx_browser_task_created_at
     ON browser_task(created_at DESC);
@@ -1704,6 +1718,10 @@ CREATE TABLE IF NOT EXISTS browser_task_clarifications (
     -- sweeps overdue rows on the same 30 s tick that handles the
     -- pending-queue timeout.
     deadline_at     INTEGER NOT NULL,
+    -- Phase 1 task.delivery recovery key for ask_user rows. Null means the
+    -- clarification has not yet been surfaced through the gated delivery
+    -- boundary; the deadline scanner still owns answer expiry separately.
+    delivered_at    INTEGER,
     answer          TEXT,
     answered_at     INTEGER,
     resolved        INTEGER NOT NULL DEFAULT 0
@@ -1774,6 +1792,166 @@ CREATE INDEX IF NOT EXISTS idx_final_confirm_tokens_status_expires
 CREATE INDEX IF NOT EXISTS idx_final_confirm_tokens_task
     ON browser_task_final_confirm_tokens(task_id);
 
+-- ── Background-task runner (BACKGROUND_TASK_RUNNER_DESIGN.md §6) ──────────
+--
+-- Generic detached-task surface — the browser-task lifecycle skeleton
+-- (state machine, clarifications store, slot manager, parked-resume,
+-- delivery boundary) minus the Playwright driver. A worker is a fresh
+-- Agent-SDK session seeded with a self-contained 'brief'; on completion
+-- (or when it needs input) it WRITES AN ARTIFACT to this row via its
+-- 'finish' / 'ask_user' tools rather than DMing directly. The runner's
+-- delivery boundary reads the 'notify' disposition and, when true, the
+-- REAL DM agent delivers (active delivery turn) or a no-model 'draft'
+-- send (idle) through the shared 'task.delivery' path.
+--
+-- Separate table (Decision 7) — 'browser_task' carries a large browser-
+-- specific column set (site_key, allowlist regex, blocked-request /
+-- extract counters, final-confirm gate) that does not generalize, and
+-- browser_task is experimental. The artifact fields ('report', 'draft',
+-- 'notify', 'significance', 'artifact_path', 'notification_policy',
+-- 'delivered_at') are the genuinely new shape this design adds.
+--
+-- Boot-recovery (§10.2): unlike browser_task (whose in-memory
+-- BrowserContext is unrecoverable, so rows are force-failed), background
+-- tasks are RE-DISPATCHED FROM 'brief' on restart — they are precisely
+-- the long-lived tasks likely to span a restart, and the brief is
+-- self-contained. The event-pipeline boot hook resets non-terminal rows
+-- to 'pending' and re-runs them once the runner is wired.
+CREATE TABLE IF NOT EXISTS background_task (
+    -- uuid v4 (crypto.randomUUID).
+    id                  TEXT PRIMARY KEY,
+    -- Self-contained worker prompt: objective + scope + the inputs the
+    -- task needs + the output-language directive + persona hints for the
+    -- 'draft' + the notification policy / (if_significant) criteria. A
+    -- 'settingSources:["project"]' worker sees ONLY this, so an
+    -- under-specified brief → clarification ping-pong (§9). Capped
+    -- 1..16384 at the route layer.
+    brief               TEXT NOT NULL,
+    -- Short label for status / listing / delivery. Optional; the route
+    -- derives one from the brief when absent.
+    title               TEXT,
+    -- State machine — mirrors browser_task minus final_confirm/abandoned.
+    -- 'awaiting_user' parks for a clarification; the slot stays held.
+    state               TEXT NOT NULL DEFAULT 'pending'
+        CHECK (state IN (
+            'pending',
+            'running',
+            'awaiting_user',
+            'completed',
+            'failed',
+            'timeout',
+            'cancelled'
+        )),
+    -- The spawn-time notification policy the worker evaluates at finish
+    -- time (§4.3). 'always' (the common case) — notify even on a "0
+    -- issues" result. 'if_significant' — notify iff the concrete criteria
+    -- written into the brief are met. 'silent' — file only.
+    notification_policy TEXT NOT NULL DEFAULT 'always'
+        CHECK (notification_policy IN ('always', 'if_significant', 'silent')),
+    -- Phase 4 if_significant criteria DSL (§4.3): a JSON array of concrete,
+    -- atomic conditions the worker checks one-by-one against its result
+    -- ("any repo main build is red", "total spend > 100"). Authored by the
+    -- DM agent at spawn for 'if_significant' tasks; the driver injects them
+    -- as a numbered checklist and the worker sets notify=true iff ANY is
+    -- met (and records which in significance). NULL / empty array means the
+    -- worker falls back to the prose criteria in the brief (back-compat).
+    significance_criteria TEXT,
+    -- ── ARTIFACT (written by the worker's finish tool) ───────────────
+    -- Full verbatim result (findings, numbers, URLs, IDs) — the fidelity
+    -- anchor, never paraphrased by a weak model. Null until finished. On
+    -- the fail-loud path (worker died before finish) the runner
+    -- synthesizes a failure note here.
+    report              TEXT,
+    -- Worker-authored plain-language summary in the owner's language.
+    -- NOT the final DM — it is the body for the idle direct-send path,
+    -- grounding material for the active delivery turn, and the fallback
+    -- if a delivery turn errors. Null until finished.
+    draft               TEXT,
+    -- Worker disposition vs the policy: 1 ⇒ surface, 0 ⇒ file only. NULL
+    -- until finished. The recovery sweep selects 'notify = 1 AND
+    -- delivered_at IS NULL'.
+    notify              INTEGER,
+    -- Worker note: why notify is true/false (digest + audit). Free-form.
+    significance        TEXT,
+    -- Optional vault MD path for research-type output (reuses the
+    -- obsidian 30_outputs/ pattern). Null for quick tasks whose row
+    -- fields suffice.
+    artifact_path       TEXT,
+    -- ── lifecycle ────────────────────────────────────────────────────
+    -- Free-form categorical detail on non-success terminals. Examples:
+    -- 'daemon_restarted', 'budget_exceeded', 'max_turns_exceeded',
+    -- 'execute_timeout', 'clarification_deadline', 'queue_timeout',
+    -- 'sdk_error', 'backend_misconfigured', 'runner_unavailable'.
+    outcome_detail      TEXT,
+    -- "<platform>:<channel_id>". NULL when no DM channel is associable
+    -- (synthetic / channel-less runs) — the artifact is still filed but
+    -- there is nowhere to deliver it.
+    originating_channel TEXT,
+    -- Links to the spawning owner-DM turn / correlation chain.
+    correlation_id      TEXT,
+    -- FK to agent_schedule when the task was inserted via 'scheduleAt'.
+    -- ON DELETE SET NULL so deleting a schedule row leaves history intact.
+    schedule_row_id     INTEGER REFERENCES agent_schedule(id) ON DELETE SET NULL,
+    -- Budget envelope inputs. 'tier' (lite|medium|high) selects the
+    -- default turn/budget/timeout envelope; 'max_budget_usd' is an
+    -- optional per-task override, clamped to the hard cap at run time.
+    tier                TEXT
+        CHECK (tier IS NULL OR tier IN ('lite', 'medium', 'high')),
+    max_budget_usd      REAL,
+    -- Parked worker SDK session id (resume after /clarify). Captured from
+    -- the init message; null until the first turn streams.
+    backend_session_id  TEXT,
+    created_at          INTEGER NOT NULL,
+    started_at          INTEGER,
+    finished_at         INTEGER,
+    -- NULL until delivered; the recovery-sweep key (§10.2).
+    delivered_at        INTEGER
+);
+CREATE INDEX IF NOT EXISTS idx_background_task_state
+    ON background_task(state);
+CREATE INDEX IF NOT EXISTS idx_background_task_corr
+    ON background_task(correlation_id);
+-- Powers the delivery recovery sweep (completed & notify=1 &
+-- delivered_at IS NULL) without scanning every historical row.
+CREATE INDEX IF NOT EXISTS idx_background_task_delivery
+    ON background_task(state, notify, delivered_at);
+CREATE INDEX IF NOT EXISTS idx_background_task_created_at
+    ON background_task(created_at DESC);
+-- The non-terminal index powers the boot re-dispatch sweep + the
+-- "needs attention" / status queries.
+CREATE INDEX IF NOT EXISTS idx_background_task_non_terminal
+    ON background_task(state)
+    WHERE state IN ('pending', 'running', 'awaiting_user');
+
+-- Mirror of browser_task_clarifications (CAS-resolve, TTL). Clarifications
+-- are always notify=true (the task cannot proceed without an answer), so
+-- there is no policy column. The TTL default is longer than browser-task's
+-- 5 min (no browser resource is held while parked).
+CREATE TABLE IF NOT EXISTS background_task_clarifications (
+    -- uuid v4 == clarificationId.
+    id              TEXT PRIMARY KEY,
+    task_id         TEXT NOT NULL REFERENCES background_task(id) ON DELETE CASCADE,
+    -- Worker-authored; the delivery turn weaves it / the idle send DMs it.
+    question        TEXT NOT NULL,
+    context_summary TEXT,
+    asked_at        INTEGER NOT NULL,
+    -- asked_at + configurable TTL (backgroundTaskClarificationTtlMinutes).
+    deadline_at     INTEGER NOT NULL,
+    -- task.delivery recovery key for ask_user rows.
+    delivered_at    INTEGER,
+    answer          TEXT,
+    answered_at     INTEGER,
+    resolved        INTEGER NOT NULL DEFAULT 0
+        CHECK (resolved IN (0, 1))
+);
+CREATE INDEX IF NOT EXISTS idx_bgtask_clar_task
+    ON background_task_clarifications(task_id);
+-- Partial index over the unresolved set — the deadline scanner sweeps
+-- this on the same housekeeping tick that handles delivery recovery.
+CREATE INDEX IF NOT EXISTS idx_bgtask_clar_unresolved
+    ON background_task_clarifications(deadline_at)
+    WHERE resolved = 0;
+
 -- INTEGRATION-DRIFT-PHASE-7-PLAN.md §3.2 — persistent dedup for the
 -- 15-minute imminent-meeting reminder. Pre-Phase-7 the scheduler kept
 -- an in-memory Set, which was lost on every daemon restart and re-DMed
@@ -2066,7 +2244,7 @@ VALUES
 -- Per-process seeds. Two tiers are wired in at install time:
 --   - Sonnet (DEFAULT_CLAUDE_MEDIUM_MODEL) for "main agent work" surfaces
 --     where output quality drives the operator's daily experience: DMs,
---     dashboard chat, hourly check, daily/weekly/monthly review, morning
+--     dashboard chat, activity scan, daily/weekly/monthly review, morning
 --     routine, scheduled tasks, git.project.* one-shots. Standard
 --     50-turn / $1.00 envelope (git.project.* span 30–100 turns and
 --     $0.50–$2.00 — retemplate is widest because re-template work is
@@ -2164,7 +2342,7 @@ VALUES
     -- above).
     ('routine.morning_routine_today',   'claude', '${DEFAULT_CLAUDE_MEDIUM_MODEL}', 50, 1.50, 'preset'),
     ('routine.morning_routine_journal', 'claude', '${DEFAULT_CLAUDE_LITE_MODEL}',   20, 0.30, 'preset'),
-    ('routine.hourly_check',    'claude', '${DEFAULT_CLAUDE_MEDIUM_MODEL}',  50,  1.00, 'preset'),
+    ('routine.activity_scan',    'claude', '${DEFAULT_CLAUDE_MEDIUM_MODEL}',  50,  1.00, 'preset'),
     -- $0.50 budget: a typical drift-triggered refresh on Sonnet runs ~$0.10
     -- in 4 turns, but a busy-calendar drift (many/large pending calendar
     -- observations read via the task-flow GET limit=200) compounded by a
@@ -2265,24 +2443,38 @@ VALUES
     -- the seeded max_turns=1 caps every backend at one assistant turn,
     -- so this envelope is the absolute ceiling — defense-in-depth on
     -- top of the prompt's "no tools" contract.
-    ('routine.hourly_check.triage', 'claude', '${DEFAULT_CLAUDE_LITE_MODEL}', 1,  0.05, 'preset'),
+    ('routine.activity_scan.triage', 'claude', '${DEFAULT_CLAUDE_LITE_MODEL}', 1,  0.05, 'preset'),
     -- docs/design/appendices/routine-data-acquisition.md §6.2 / §6.9 — pre-pass window
     -- fetcher dispatched before each routine session. Lite tier
     -- (Haiku-class) per P3 ("Lite for Fetch"). Envelope sized for the
     -- worst-case routine (morning_routine: fan-out across 2 mail
-    -- providers × N accounts + 2 calendar providers + notion). 20
-    -- turns is enough headroom for ~6 partials × 3 tool calls each;
+    -- providers × N accounts + 2 calendar providers + notion);
     -- $0.50 caps the fan-out so a misconfigured account list cannot
     -- drain budget. The lite-tier nominal ($0.20) under-provisioned
     -- real morning fan-outs and tripped BackendQuotaError(max_budget_usd)
     -- mid-fetch, so this envelope is widened via the per-process
     -- override in plan-presets.ts (ENVELOPE_OVERRIDES_BY_PROCESS_KEY).
-    ('routine.fetch_window',    'claude', '${DEFAULT_CLAUDE_LITE_MODEL}', 20,  0.50, 'preset'),
+    -- max_turns 20 → 10 (PREPASS_COST_REDUCTION_PLAN.md N4): with the
+    -- per-integration fan-out each session handles ONE partial; live
+    -- P99 over 502 runs is 8 turns. Lock-step with plan-presets.ts.
+    -- Seed-only change — existing installs keep their row (INSERT is
+    -- ignore-on-conflict) until they re-apply defaults.
+    ('routine.fetch_window',    'claude', '${DEFAULT_CLAUDE_LITE_MODEL}', 10,  0.50, 'preset'),
     -- BROWSER_HISTORY_INTEGRATION_PLAN P3:
     --   research_cluster_update — nightly per-cluster journal append.
-    --     Lite tier (Haiku-class) — templated DM dispatcher with a
-    --     single PUT to context/research/<slug>.md. 5 turns / $0.05 is
-    --     enough headroom; the §10.3 safety floor refuses Codex
+    --     Lite tier (Haiku-class) — a small curl flow against the
+    --     daemon API ending in one append PATCH to research/<slug>.md.
+    --     $0.50 is a STOP-LOSS, not a per-run cost target
+    --     (RESEARCH_CLUSTER_COST_FIX_PLAN.md RC2/F3): the SDK budget
+    --     check only fires between turns, and a cold-prompt-cache run
+    --     writes the full session prefix (~$0.13-0.30 observed) before
+    --     the check can abort — the original $0.05 seed killed every
+    --     cold run after the money was already spent. The F1
+    --     per-agent-day enqueue stamp bounds the key to one run per
+    --     cluster per day, so this ceiling caps daily spend at
+    --     $0.50/cluster. Existing installs are bumped by migration
+    --     0012; keep in lock-step with ENVELOPE_OVERRIDES_BY_PROCESS_KEY
+    --     in plan-presets.ts. The §10.3 safety floor refuses Codex
     --     outright, so the seeded claude row is the only eligible
     --     binding until the operator widens via /settings/models.
     --   research_dispatch — accept path. Medium tier (Sonnet). Uses
@@ -2293,7 +2485,7 @@ VALUES
     --     tier; smaller envelope (30/$0.50) — the agent composes from
     --     the cluster journal it already wrote, so WebFetch fan-out is
     --     bounded.
-    ('routine.research_cluster_update', 'claude', '${DEFAULT_CLAUDE_LITE_MODEL}',    5,  0.05, 'preset'),
+    ('routine.research_cluster_update', 'claude', '${DEFAULT_CLAUDE_LITE_MODEL}',    5,  0.50, 'preset'),
     ('routine.research_dispatch',       'claude', '${DEFAULT_CLAUDE_MEDIUM_MODEL}', 50,  1.00, 'preset'),
     ('routine.research_wiki_summary',   'claude', '${DEFAULT_CLAUDE_MEDIUM_MODEL}', 30,  0.50, 'preset'),
     -- BROWSER_TASK_REDESIGN_PLAN.md §5 — open-ended browser sub-agent.
@@ -2307,7 +2499,16 @@ VALUES
     -- Keep this row's (max_turns, max_budget_usd) in sync with §5's
     -- envelope spec — a future widening to support multi-tab tasks
     -- needs both this seed and the spec to move together.
-    ('browser_task',                    'claude', '${DEFAULT_CLAUDE_MEDIUM_MODEL}', 30,  1.00, 'preset');
+    ('browser_task',                    'claude', '${DEFAULT_CLAUDE_MEDIUM_MODEL}', 30,  1.00, 'preset'),
+    -- BACKGROUND_TASK_RUNNER_DESIGN.md §6 — generic detached-task worker.
+    -- Medium tier (Sonnet) by default; the worker runs arbitrary
+    -- long-running work (deep research, multi-repo audit, monitoring).
+    -- The seed envelope is the FALLBACK + the per-tier base — the row's
+    -- 'tier'/'max_budget_usd' (POST body) scale within the hard caps in
+    -- background-task-budget.ts. 40 turns / $2.00 is the medium-tier base;
+    -- long tasks select tier='high' for the wider envelope. Operators can
+    -- pin model/backend per-row from /settings/models.
+    ('background_task',                 'claude', '${DEFAULT_CLAUDE_MEDIUM_MODEL}', 40,  2.00, 'preset');
 
 INSERT OR IGNORE INTO settings (key, value_json, updated_at)
 VALUES (

package/dist/db/voice-transcripts-store.d.ts

@@ -2,7 +2,7 @@ import type Database from "better-sqlite3";
 /**
  * Persistence helpers for `voice_transcripts` — local-Whisper transcripts
  * keyed 1:1 to `chat_attachments.id`. The cache lets a re-dispatch of the
- * same turn (or a Phase-9 hourly check that revisits an old voice
+ * same turn (or a Phase-9 activity scan that revisits an old voice
  * message) reuse a transcription instead of re-running inference.
  *
  * See `docs/design/appendices/voice-transcription.md`.

package/dist/index.js

@@ -5,7 +5,6 @@ import { getDegradedMode, getVaultRestructurePendingConsent, isSetupCompleted, i
 import { initDirectories } from "./init.js";
 import { EventBus } from "./core/event-bus.js";
 import { AgentScheduler } from "./core/scheduler.js";
-import { CustomRoutineScheduler } from "./core/custom-routine-scheduler.js";
 import { HealthMonitor } from "./core/health-monitor.js";
 import { Heartbeat } from "./core/heartbeat.js";
 import { MessageHub } from "./adapters/message-hub.js";
@@ -22,7 +21,7 @@ import { resolveUserSkillsRoot } from "./core/user-skills-root.js";
 import { bootstrapManagementMd, startManagementMdWatcher, } from "./core/management-md.js";
 import { bootstrapManagementRegistry, startManagementRegistryWatcher, } from "./core/management-registry.js";
 import { startDocsIndexer, } from "./core/docs/indexer.js";
-import { APP_NAME, EventPriority, getBackendIds, } from "@aitne/shared";
+import { APP_NAME, EventPriority, getAgentDayDateStr, getBackendIds, } from "@aitne/shared";
 import { getOwnerChannel, selectFirstPairedPlatform, } from "./messaging/owner-channels.js";
 import { SUPPORTED_MESSAGING_PLATFORMS, } from "./messaging/constants.js";
 import { AgentWriteTracker } from "./safety/agent-write-tracker.js";
@@ -30,6 +29,8 @@ import { InMemoryTodayWriteLockManager, getTodayWriteLockTimeoutMs, } from "./co
 import { InMemoryRoadmapWriteLockManager, getRoadmapWriteLockTimeoutMs, } from "./core/roadmap-write-lock.js";
 import { runRoadmapMechanicalMaintenance } from "./core/roadmap-maintenance.js";
 import { fanoutResearchClusterUpdates } from "./core/browser-history/research-cluster-fanout.js";
+import { runDayBoundaryTasks } from "./core/day-boundary.js";
+import { SleepInhibitor } from "./core/sleep-inhibitor.js";
 import { safeRunPreMorningDigestJob } from "./core/browser-history/pre-morning-digest-job.js";
 import { shouldStartObserversFor } from "./core/integration-lifecycle.js";
 import { sweepExpiredMigrationBackups } from "./api/routes/setup-migrate.js";
@@ -47,7 +48,7 @@ import { SecretBroker } from "./secrets/secret-broker.js";
 import { captureOriginalShellEnv, syncBackendApiKeyToEnv, } from "./secrets/backend-api-key-env.js";
 import { createLogger, toSafeErrorMessage } from "./logging.js";
 import { runCatchup, runPostMessagingCatchup, } from "./bootstrap/catchup.js";
-import { createAdapterReloaders, } from "./bootstrap/adapters.js";
+import { createAdapterReloaders, createAdapterWatchdog, } from "./bootstrap/adapters.js";
 import { createInitialSecretState, createServiceReloaders, } from "./bootstrap/services.js";
 import { initDatabase } from "./bootstrap/db.js";
 import { bootstrapAgents } from "./core/agents/loader-boot.js";
@@ -93,6 +94,16 @@ async function startup() {
     const { db, settingsStore, persistedSettings, attachmentStore } = initDatabase({
         config,
     });
+    // ── Keep-awake (sleep inhibitor) ──
+    // Host sleep freezes every daemon timer; on a sleeping macOS laptop the
+    // 04:00 day-boundary flow degrades into hours of wake-catchup replays
+    // riding maintenance DarkWakes, each at cold-prompt-cache cost. Hold a
+    // `caffeinate` assertion tied to our pid for the daemon's lifetime
+    // (AC-power-only by default — see `preventSleepMode` in
+    // settings/runtime-settings.ts). Started after initDatabase so a
+    // dashboard-persisted mode from the settings table is honored.
+    const sleepInhibitor = new SleepInhibitor({ mode: config.preventSleepMode });
+    sleepInhibitor.start();
     // ── Integration Delegation Framework (Phase 1) ──
     // Reconcile `<contextDir>/policies/integrations.md` with the DB integrations map.
     // Creates the file on first run, parses hand-edits if present, and
@@ -371,6 +382,12 @@ async function startup() {
         reloadSlackAdapter(false),
         reloadTelegramAdapter(false),
     ]);
+    // Connection watchdog for the library-managed adapters. Probes the live
+    // socket/poll state on an interval and forces a stop→start cycle through
+    // the reloaders when a transport stays dead (e.g. after machine sleep
+    // kills the TCP sockets and the library's own reconnect gave up).
+    // Started after messageHub.startAll() below; stopped in shutdown.
+    const adapterWatchdog = createAdapterWatchdog({ messageHub, state: adapterState }, { reloadDiscordAdapter, reloadSlackAdapter, reloadTelegramAdapter });
     if (config.whatsappEnabled) {
         if (!config.whatsappOwnerPhone) {
             throw new Error("PA_WHATSAPP_ENABLED=true but PA_WHATSAPP_OWNER_PHONE is not set");
@@ -819,15 +836,10 @@ async function startup() {
     const heartbeat = new Heartbeat();
     // ── 9. Scheduler ──
     const scheduler = new AgentScheduler(eventBus, db, config);
-    // ── 9.1 Custom routine scheduler (B-007 §5.8) ──
-    // Reads `policies/routines/custom/*.md` from the context dir and registers a
-    // cron job per enabled routine. Reloaded from the context API whenever
-    // the agent or dashboard edits a file under that directory.
-    const customRoutineScheduler = new CustomRoutineScheduler({
-        contextDir: getContextDir(config),
-        eventBus,
-        timezone: config.timezone || undefined,
-    });
+    // NB: the legacy CustomRoutineScheduler (B-007 §5.8 — per-file node-cron
+    // jobs over `policies/routines/custom/*.md`) was retired at the Agents-hub
+    // redesign; `bootstrapAgents` converts those files into user Agents once at
+    // boot (AGENTS_HUB_REDESIGN_PLAN.md §3).
     // ── 10. Event Processing Pipeline ──
     // §9.5 SignalDetector + §10 event-processing pipeline (agent cores,
     // BackendRouter, dispatcher + setters, NotificationManager,
@@ -885,7 +897,7 @@ async function startup() {
             emitRoadmapRefreshSink = cb;
         },
     });
-    const { dispatcher, sessionManager, notificationManager, signalDetector, eventBroadcaster, auditLogger, docsQAAdapter, agentBackends, opencodeServerManager, delegatedBackendInvoker, authHealthMonitor, authRecovery, authTelemetry, readTokenManager, migrationLock, contextWriteGate, buildDelegatedSyncWorker, getDelegatedSyncWorker, rematerializeActiveDmWorkdirs, handleSecretChange, handleGoogleServicesReady, handlePromptContextChanged, keepaliveTimer, browserTaskDeadlineTimer, } = eventPipeline;
+    const { dispatcher, sessionManager, notificationManager, signalDetector, eventBroadcaster, auditLogger, docsQAAdapter, agentBackends, opencodeServerManager, delegatedBackendInvoker, authHealthMonitor, authRecovery, authTelemetry, readTokenManager, migrationLock, contextWriteGate, buildDelegatedSyncWorker, getDelegatedSyncWorker, rematerializeActiveDmWorkdirs, handleSecretChange, handleGoogleServicesReady, handlePromptContextChanged, keepaliveTimer, browserTaskDeadlineTimer, backgroundTaskHousekeepingTimer, } = eventPipeline;
     // ── 10.5 Agent Definitions load (AGENT_DEFINITIONS_DESIGN.md §6.1 / §7.1) ──
     // Runs AFTER `runMigrations` (inside `initDatabase`) and BEFORE
     // `scheduler.start()`: scans the built-in + user `agent.md` roots into the
@@ -934,7 +946,6 @@ async function startup() {
         sessionManager,
         scheduler,
         agentEnabledCache,
-        customRoutineScheduler,
         healthMonitor,
         heartbeat,
         messageHub,
@@ -949,6 +960,10 @@ async function startup() {
         // same in-memory state as the runner's promote/release cascade.
         browserTaskRunner: eventPipeline.browserTaskRunner,
         browserTaskSlotStateRef: eventPipeline.browserTaskSlotStateRef,
+        // BACKGROUND_TASK_RUNNER_DESIGN.md §4 — generic runner + shared slot
+        // state, threaded through so the API routes share the runner's state.
+        backgroundTaskRunner: eventPipeline.backgroundTaskRunner,
+        backgroundTaskSlotStateRef: eventPipeline.backgroundTaskSlotStateRef,
         writeTracker,
         auditLogger,
         attachmentStore,
@@ -1007,16 +1022,24 @@ async function startup() {
     // (BROWSER_HISTORY_INTEGRATION_PLAN §10.6 step 3). The fan-out is
     // bounded at 25 clusters / cycle; backlog clusters surface on the
     // next day-boundary tick.
+    //
+    // RESEARCH_CLUSTER_COST_FIX_PLAN.md F2 — `runDayBoundaryTasks` gates
+    // the body behind a per-agent-day runtime_state marker. The scheduler
+    // invokes this callback from three sites (04:00 cron, wake catch-up,
+    // morning self-heal); wrapping the single composition site keeps a
+    // sleep-replay morning from re-running the body. Errors propagate
+    // WITHOUT writing the marker — each scheduler site catches + logs and
+    // the next fire retries the whole body.
     scheduler.setDayBoundaryCallback(async () => {
-        await dispatcher.summarizeDmSessions();
-        try {
-            const result = await fanoutResearchClusterUpdates(db, eventBus);
-            if (result.enqueuedSlugs.length > 0) {
-                logger.info({ enqueuedSlugs: result.enqueuedSlugs }, "Research cluster updates enqueued at day boundary");
-            }
-        }
-        catch (err) {
-            logger.error({ err }, "Research cluster update fan-out failed; will retry next day boundary");
+        const todayAgentDay = getAgentDayDateStr(config.timezone || undefined, config.dayBoundaryHour);
+        const result = await runDayBoundaryTasks({
+            db,
+            todayAgentDay,
+            summarizeDmSessions: () => dispatcher.summarizeDmSessions(),
+            fanoutResearchClusterUpdates: () => fanoutResearchClusterUpdates(db, eventBus, { todayAgentDay }),
+        });
+        if (result.ran && result.enqueuedSlugs.length > 0) {
+            logger.info({ enqueuedSlugs: result.enqueuedSlugs }, "Research cluster updates enqueued at day boundary");
         }
     });
     // Register direct DM callback: sends scheduled messages without running an agent
@@ -1030,7 +1053,7 @@ async function startup() {
         }
         return messageHub.sendToUser(message);
     });
-    scheduler.setHourlyCheckCallback((source) => dispatcher.triggerHourlyCheck(source));
+    scheduler.setActivityScanCallback((source) => dispatcher.triggerActivityScan(source));
     // B-004 Phase 2a — nightly context-index reconciler. The design doc
     // (§4.1, §5.3) originally proposed an `agent_schedule` row with
     // `task_type: "internal.reconcile_context_index"`, but the dispatcher's
@@ -1095,11 +1118,42 @@ async function startup() {
             },
         });
     });
-    // Phase 4 auth probe — runs BEFORE the hourly check on each cron
+    // Phase 4 auth probe — runs BEFORE the activity scan on each cron
     // tick so auth health detection happens independently of the
     // observation-threshold gate. checkAll() owns its own kill switch
     // (authProbeDisabled), morning-routine skip, and in-flight dedupe.
     scheduler.setAuthProbeCallback(() => authHealthMonitor.checkAll());
+    // SELF_TUNING_REVIEW_CYCLE_DESIGN.md §3.4 Phase 3 — self-tuning
+    // auto-revert monitor. Rides the hourly cron tick (P2 — no new scheduled
+    // session), throttles itself to one pass per day, and is a pure no-op
+    // until the Phase 3 actuator has written `runtime_state.self_tuning:*`
+    // ledger entries. Reverts go through the same `applyConfigUpdates`
+    // chokepoint the actuator used; the owner is DM'd on every auto-revert.
+    // Deliberately NOT gated on `selfTuningEnabled`: a safety rollback must
+    // keep working even if the owner turns the loop off after a change
+    // landed.
+    scheduler.setSelfTuningRevertMonitorCallback(async () => {
+        const { runSelfTuningRevertMonitor } = await import("./core/feedback/tuning-revert-monitor.js");
+        const { applyConfigUpdates } = await import("./api/env-writer.js");
+        const { SELF_TUNING_NOTIFICATION_TYPE } = await import("./core/feedback/tuning-recommender.js");
+        return runSelfTuningRevertMonitor({
+            db,
+            applyUpdates: (updates) => applyConfigUpdates(config, settingsStore, updates, { db }),
+            feedbackLearningEnabled: config.feedbackLearningEnabled,
+            sendDm: async (message) => {
+                await notificationManager.send(message, {
+                    // Logged to notification_log under this type; R2 excludes it
+                    // from demotion candidates (SELF_TUNING_NOTIFICATION_TYPE).
+                    type: SELF_TUNING_NOTIFICATION_TYPE,
+                    source: "self-tuning-revert-monitor",
+                    priority: EventPriority.NORMAL,
+                    timestamp: new Date(),
+                    data: {},
+                    correlationId: randomBytes(8).toString("hex"),
+                }, { priority: "high", destinationMode: "configured_only" });
+            },
+        });
+    });
     // Wire the autonomous-work gate: when policies/management.md is missing
     // or a setup conversation is active, the scheduler pauses cron routines
     // and ScheduleWatcher claims. This prevents any autonomous turn from
@@ -1108,7 +1162,7 @@ async function startup() {
     scheduler.setAutonomousGate(() => dispatcher.isAutonomousAllowed());
     // Pre-routine morning_routine gate (sleep-skip recovery). When the
     // current agent-day's morning_routine has not completed yet — typical
-    // cause: Mac slept through the 04:00 cron tick — hourly_check and the
+    // cause: Mac slept through the 04:00 cron tick — activity_scan and the
     // review routines enqueue a wake row instead of running on stale state.
     // Wired here after both `dispatcher` and `scheduler` exist so the
     // binding is a single, stable function reference for the duration of
@@ -1117,8 +1171,8 @@ async function startup() {
     const startupCatchup = await runCatchup(db, dispatcher, config);
     // ── 13. Start all components ──
     await messageHub.startAll();
+    adapterWatchdog.start();
     scheduler.start();
-    customRoutineScheduler.start();
     signalDetector.start();
     const registeredPlatforms = messageHub.getPlatforms();
     // Single-app installs (Telegram-only / Discord-only / etc.) would
@@ -1282,10 +1336,11 @@ async function startup() {
     const shutdown = async (signal) => {
         logger.info({ signal }, "Shutdown signal received");
         dispatcher.stop(); // Signals dispatcher to exit run() loop
+        adapterWatchdog.stop();
         scheduler.stop();
-        customRoutineScheduler.stop();
         healthMonitor.stop();
         heartbeat.stop();
+        sleepInhibitor.stop();
         signalDetector.stop();
         notificationManager.stop(); // Clear pending batch-flush timer
         authRecovery.shutdown(); // Kill any active recovery subprocesses
@@ -1299,6 +1354,8 @@ async function startup() {
         clearInterval(keepaliveTimer);
         // BROWSER_TASK_REDESIGN_PLAN.md §5 / §5.1 — 30 s deadline tick.
         clearInterval(browserTaskDeadlineTimer);
+        // BACKGROUND_TASK_RUNNER_DESIGN.md §6 — 30 s housekeeping tick.
+        clearInterval(backgroundTaskHousekeepingTimer);
         clearTimeout(migrationBackupSweepInitial);
         clearInterval(migrationBackupSweepTimer);
         if (managementMdWatcher) {
@@ -1340,7 +1397,7 @@ async function startup() {
     logger.info(`${APP_NAME} Daemon ready`);
 }
 // Catchup (`runCatchup` / `runPostMessagingCatchup`) and the pure schedule
-// predicates (`getDueCatchupRoutines`, `shouldCatchUpHourlyCheck`,
+// predicates (`getDueCatchupRoutines`, `shouldCatchUpActivityScan`,
 // `getProgressMinutesForHour`, `hasFreshAgentDayTodayMd`,
 // `readSkillCurationCadence`) live in `./bootstrap/` — see
 // `docs/design/appendices/file-split-plan.md` §10. Imports are at the top