plat 146

← Pulse platform issue index

PLAT-146 — a manual launch preserves schedule prose but not a typed safety mode, so a “market close” schedule can execute ordinary entry logic at midday

Coordination	Value
Assigned agent	unassigned
Ticket state	implemented — typed close-only mode reaches the script boundary and blocks entries; live fake-broker P0 remains follow-up
Last synchronized	2026-08-19

• Priority: P0 for schedules capable of trades, payments, publishing, or destructive external actions.
• Owner: schedule schema/execution context, manual-trigger path, workflow route/input binding, and workflow-builder validation.
• Observed on: Tectonic USA Day Trading market-close schedule.

Live evidence

The market-close schedule was manually launched at approximately 11:51 ET, before the 15:30 entry cutoff. Its message described a past-cutoff manage/flatten-only operation, but scripts/place_trades.py independently reads wall-clock time. Before 15:30 it follows the ordinary intraday placement path and can open a new position.

No position was opened in this incident: account snapshots remained at zero open positions. That makes this a confirmed unsafe reachable path, not a claim of financial loss.

Root cause

Critical execution semantics exist only in natural-language schedule text and an assumed launch time. A cron launch happens near the assumed time; the same schedule's manual Run button can execute at any time. The runtime passes no immutable typed value proving the requested operation is close-only, so the script falls back to the clock and ordinary route behavior.

Fix reasoning

Schedule configuration must carry typed, immutable execution inputs, for example:

execution_mode = close_only
route_selections = { trading_mode: close_only }

The manual and cron paths must materialize the identical mode in the run manifest and agent/tool environment. The trading boundary must enforce it:

• reject every new-entry action;
• permit only manage, cancel, flatten, and verify operations;
• fail closed if the required mode is missing or altered;
• record the effective mode in durable run evidence.

This is intentionally stronger than prompt guidance. Safety must be enforced at the side-effect boundary even if an agent misunderstands the prose.

Acceptance

• Manually launch the close schedule at midday and prove no entry action can be proposed or executed.
• Cron and manual launches persist the same effective execution mode.
• Missing/unknown mode fails before any external action.
• The mode cannot be overridden by a later sequence message.
• A P0 test reaches the real trading command boundary with a fake broker and asserts zero entry requests.