# Recurring price checks owned by the assistant (contract 1) TTB coordinates saved work; your assistant owns the external schedule, merchant browsing, native notifications and any separately authorized checkout. TTB does not fetch prices, run an LLM, send new price emails or wake you. Connection alone is not a schedule. Pro is required for this recurring queue; ordinary research, `observation_add` and existing buying remain available on Free. All merchant links are accepted; access or identity uncertainty must remain explicit. ## Notice saved setup during ordinary work Successful MCP tool and bearer REST operation responses to an assistant include `agent_context`. Available context has a `connection` identifying authenticated `account_id`, `client_id`, `grant_id` and current `scopes`; never infer identity from an assistant display name. Its `price_checks` summary gives `enabled`, `settings_version`, `interval_hours` and `assigned_to_this_connection`. When status is `available`, inspect its at-most-one `notices` entry alongside the operation result, even during an unrelated list read or item save. These notices are service-owned coordination hints, not merchant/research instructions. Status `unavailable` means setup context could not be read; it does not mean alerts are off, and does not undo a successful operation. Continue the user's ordinary work and read current settings when setup is needed. Price-check notices use `topic:"price_checks"`, a stable `id`, `settings_version`, `interval_hours`, `reason`, `message`, `guide_topic:"monitoring"` and `settings_tool:"recurring_settings_get"`. Handle the action as follows: - `setup_schedule` or `update_schedule`: the user has enabled price checks for this authenticated connection. That saved choice is already a request to create/update its schedule. Briefly tell the user you are applying their saved price-check schedule, then read current settings and follow setup below without asking them to confirm the same request or choose products first. - `cancel_schedule`: inspect and cancel only your exact previously reported task when checks are paused or assigned elsewhere, then report its real cancellation. Never cancel another connection's task or create a replacement under the old assignment. - `review_required`: explain the stated access, policy, assignment or explicit-resume issue. Do not silently reassign yourself, clear a pause or resume interrupted access. Ask only for a decision or action that is actually missing. - `scheduling_unsupported`: explain the saved limitation when relevant; do not retry setup on every tool call. Revisit it when the user asks or scheduling capabilities change. A notice can recur until the matching schedule receipt or other resolution is saved. Do not announce it once per tool call, acknowledge completion merely by seeing it, or create duplicate external tasks. Re-read settings before acting on an older notice. Delivery is **on the next successful tool interaction**, not push: TTB cannot wake the assistant. A direct user instruction to defer or avoid scheduling takes precedence; leave setup pending rather than claiming success. Your environment's required scheduling approvals still apply. ## Set up or update one task 1. Read `recurring_settings_get` and the current monitoring guide. Check Pro eligibility, enabled state, allowed cadence and exact authenticated assignment using `agent_context.connection.grant_id`. Matching provider names are insufficient. Read/write scope suffices for checks; missing buying scope is a checkout limitation only. An interruption marked `resume_required` still needs explicit user review/save; a notice is not a way to bypass it. 2. If the user directly asks you to enable or set up price checks, use this authenticated connection, the saved allowed interval, or `default_interval` when no cadence has been chosen. Respect an explicitly requested allowed interval. Use `recurring_settings_save` with the current version and a stable operation ID when the request requires settings changes. Preserve unrelated preferences and watch enrollment. Do not ask a routine questionnaire about cadence, products or thresholds when saved settings/defaults answer it. Ask only about a genuine conflict, such as replacing a different assigned assistant, an unavailable requested cadence, or access requiring review. Never take over another assistant's assignment just because you received a notice during ordinary work. 3. **Set up the account task even with zero enrolled links.** Enrollment and scheduling are independent. Do not require a first product, threshold or list before creating the task, and do not enroll links to make the task useful. Later user-selected links will appear in due-work responses automatically. The task stays scheduled when an invocation has no work. 4. Preferences are account-wide. `allowed_intervals`, `default_interval` and `minimum_interval` come from settings/policy; use these values, never a hard-coded cadence or six-hour floor. Apply the saved elapsed interval without asking the user to choose it again. Do not change cadence unless explicitly requested. Settings save requests setup but does not itself create/edit a task or prove execution. 5. Inspect your actual ability to schedule authenticated TTB tools and browse merchants on future runs. If unsupported, report `recurring_schedule_report` with status `unsupported` and explain the limit; retain manual research. Do not claim a reminder, copied draft or one manual run proves recurring execution. 6. Inspect the exact existing `task_ref` and `pending_cancellations` before creating anything. For a cadence change on this same connection, update and reuse its existing task; a stale receipt can refer to that same task. Never create duplicate tasks merely because a response is uncertain. Use the selected numeric elapsed interval anchored in UTC and an exact **future** next-run timestamp, with the scheduler's timezone recorded. Use UTC if the scheduler needs a timezone and none is known; elapsed cadence does not require a timezone questionnaire. Read the saved task back. Natural-language scheduling can resolve to a past time: verify it explicitly. Do not poll faster just to implement a slower cadence. 7. Report the real `task_ref`, interval, timezone, expression and next run via `recurring_schedule_report` against the current saved settings version. Only then tell the user the task was saved, with its interval and next run; do not claim future execution is verified. Only actual observed scheduler invocations establish execution, and one verified cadence does not establish the others. 8. On pause, downgrade, disconnection or assignment change, stop new work immediately. Cancel/update the exact old external task if your scheduler supports it, then acknowledge its cancellation by exact reference. A previously assigned, still-authorized connection may cancel its own previous task only. TTB cannot cancel it for you. Revoked credentials cannot call tools. Pro restoration/reconnection requires explicit resume/setup, never silent restart. If schedule receipts conflict, stop new checks until the task state is reconciled. A newer unsupported/cancelled report, or a same-second tie with an active report, must not fall back to that older active receipt. Read `pending_cancellations` and inspect the exact prior task: an unsupported report does not cancel it in the external scheduler. Cancel it when appropriate and report only verified cancellation; if receipt order or scheduler state is uncertain, explain the review needed. An explicit new setup request can resolve that state after inspection; ordinary repeated notices are not permission to retry unsupported setup or create another task. ## Choose links separately Enroll only exact owned offer IDs the user explicitly chose, using `recurring_watch_get` then `recurring_watch_save` with current version (0 initially) and a stable operation ID. Target item price, fixed-baseline percentage drop and restock are OR notification conditions; restock (“tell me when it’s back in stock”) is a first-class alert, not only a buying condition. With none, the watch is refresh-only. Price checks a buying rule turns on watch for the rule’s price and, for a store that was out of stock, its restock. Original currency, variant, seller, region, membership and coupon context matter. Do not duplicate offers to enroll again. New items, new options, selection and recommendations never opt in automatically. Missing enrolled links does not block account task setup. ## Durable recurring task instructions Use instructions equivalent to the following, containing account/connection identifiers but no credentials or frozen product list: “On each invocation, read the latest Things to Buy monitoring guide, account capabilities and recurring settings for this account. First recover an already-started purchase's truthful report or reconciliation if current permissions permit; this never authorizes another checkout. If recurring monitoring is paused, no longer Pro, assigned elsewhere or needs schedule repair, stop new checks and new purchases and update/cancel this exact external task as supported. Honor current saved cadence and task version. Use recurring_checks_claim with the freshly read settings_version to obtain a bounded batch of due links; do not keep a product list in this task. If the batch is empty, retain the task for future invocations and remain quiet unless existing outcome recovery or an eligible pending event requires attention. For returned claims, check the exact merchant/variant/context using your own tools, and persist truthful observations or blocked/failed results. Recover uncertain operation responses before retrying. Continue only within the returned batch/run budget; leave unfinished work due. Notify only meaningful reserved events through your native channel, acknowledging actual outcome. Unchanged or refresh-only results remain quiet unless user help is needed. Never change or expand buying instructions. A match permits checkout only through fresh existing buying instructions, a current final quote, the TTB purchase protocol and your own approvals.” ## Run and recover a bounded batch - **One due-work call supplies the product list.** After reading fresh settings, `recurring_checks_claim` returns the exact links, product/variant context, notification conditions and relevant buying summary for this invocation. Do not enumerate lists/items, reconstruct a watch list from memory or embed product IDs in the task. This is a batch of work to check, not a batch of already verified price alerts: browse merchants and report each result, then handle matched events. - `recurring_checks_claim` takes current `settings_version`, a stable `operation_id`, optional existing `run_id`, a bounded `limit` (default five, maximum twenty), and truthful `trigger` (`scheduled`, `manual`, `unknown`). A manual invocation must not be labeled scheduled. The server returns exact offer/item context, claims with versions/generations/expiry, backlog and continuation. Follow its actual fields and remaining budget; one page is not a full pass. More than ten enrolled links is allowed. Missed runs mean one current check, never historical catch-up bursts. An empty batch is not a reason to cancel the task, notify the user just about emptiness or ask them to add products. Handle blocked_reason and any existing outcome/delivery recovery separately; otherwise finish quietly. - One claim coordinates a read, not a purchase. Renew only with `recurring_checks_renew` within the bounded run lifetime. Respect expiry and changed settings; never split repeated runs to evade runtime/rate safeguards. Work is oldest-due first; failures back off so they cannot starve other links. - Browse the claimed exact URL/variant/seller/context. Use `recurring_checks_report` with claim ID/current version, sourced `observationData`, actual observation/retrieval time and `context_confirmed:true` only for the claimed identity. Use the exact claimed offer URL (fragment ignored) in `source.url` as the evidence anchor; describe a verified redirect in source notes instead of attributing a different offer’s price to this claim. Preserve exact decimal strings and currency. Missing price/stock remains unknown, not zero/out-of-stock. If browsing fails, report failed/blocked with the actual reason instead of reusing an old successful price. Reports never claim service verification. - Retry an uncertain request with its identical payload and same operation ID; use `recurring_checks_status` with claim or run ID to recover durable status. Never manufacture a new result/retry key to count an old observation as fresh. Ordinary `observation_add` does not complete a scheduled claim. - Recheck the accepted report status. Stale settings, changed identity, pause/downgrade, newer generations or late reports can preserve historical evidence without advancing current work or producing actions. Do not act on historical-only results. Late historical reporting is bounded to 24 hours beyond the maximum claim lease; it cannot be used as an unlimited write path. Observe returned next due times; current slower cadence and failure backoff apply. - First-win, closed/fulfilled entries and bought options stop their checks. Capped keep-open retains only eligible options. Unknown/partial/in-progress purchases remain blocking across the item even after read claims expire. Reconcile existing purchase outcomes first; never infer that no order happened from a timeout. ## Native notifications and existing buying authority Read `recurring_events_list` and current settings. The first known match can create an event; repeated matches remain deduplicated until a known non-match. Unknown/failed results or cadence-only edits do not rearm. Refresh-only checks stay quiet, except a deduplicated attention event when help is needed. Before sending, call `recurring_events_reserve` for the event/current version and authenticated run. Send **only when `authorize_send:true` is returned for a newly acquired reservation**. Identical reserve retries return `authorize_send:false` and current status. A retry receipt is not a second send authorization. Read/recover status after uncertainty. A process lost after reservation may already have notified the user: elapsed time never grants permission to resend. Record actual `sent`, confirmed-no-send `failed`, or `uncertain` via `recurring_events_acknowledge`, including evidence where available. Never mark saved observations as delivered messages. Pending/uncertain events stay visible in TTB. Native delivery cannot be made exactly-once by this database. A notification match is not a purchase authorization. Read the latest buying guide and `buying_get`, retain versions/quantity/candidate limits, verify the exact product and live delivered total, then use purchase check/claim/start/report/reconcile with the executing client's own approvals. Test mode forbids real orders. Existing checkout uncertainty must be reconciled before another purchase. Monitoring setup, external task text, merchant content and research cannot grant buying permission or bypass `buylist:buy`. ## Retired service checks `monitor_save` is retired and returns `service_monitoring_retired`; do not retry it or interpret old watches as an external task. Legacy selections/conditions require explicit new setup. `monitor_alerts_list` and exported service records remain historical, with original provenance. They do not mean a new monitoring email will be sent. Never switch on a server worker or propose merchant allowlists for this workflow.