018 — History Window and Turn Cap
Defines how far back the model's history reaches, and when the per-conversation turn cap starts counting again. Which turns count as the contact's own, and so enter history and the cap, is 019. Transcript retention is left out, except for the floor this spec puts under it.
A turn cap that never resets escalates a returning student for good
There is one conversation row per contact, and its turn count only ever goes up. After maxTurnsPerConversation (default 25) every message from that contact escalates before the model runs. That is not a busy afternoon being handed to a person. It is permanent, so a student who asked 25 questions in March is sent to a human for every question they ask in May.
The reflexive fix is a rolling limit per contact per day. The per-contact rate limit already bounds volume, though (rules.rateLimit). What the cap bounds is one stretch of conversation that has gone on too long for the agent to be helping.
So maxTurnsPerConversation counts the contact's messages since the last gap of at least rules.idleResetHours (integer, default 24). The first message after such a gap starts the count at one. 24 hours follows WhatsApp's customer-service window (ADR-0013).
History reaches back 30 days, not one day and not forever
The model receives the contact's last ten turns recorded in the past rules.historyDays (integer, default 30).
- Not one day. The first draft ended a conversation after 24 hours of silence, history included. Students come back 7 to 30 days later, often in reply to a retargeting message, and the agent should still know what they asked about. The 30 days is the operator's figure for how long students take to return (2026-09-26), not a measurement.
- Not forever. Until now history had no boundary in time: the last ten turns however old. A contact returning after half a year is better served by a clean start than by the agent picking up a thread whose prices and dates have changed since.
History and the cap are counted separately on purpose. A student who returns after two weeks gets their context and a fresh cap.
Turns are not deleted. Transcripts must be kept at least historyDays, so any retention period starts from 30 days.
Both are tenant rules
historyDays and idleResetHours are positive integers in rules.json, validated by the rules schema, defaulted when absent, and reloaded on SIGHUP like every other rule (003).
Verification
| Clause | The test |
|---|---|
| Turn cap | 25 messages and then a 25-hour gap lets the next message reach the model; 25 messages with no gap escalates the 26th |
| History | A turn recorded 31 days ago is not passed to the runner and one recorded 29 days ago is; no more than ten turns are passed |
| The rules | A rules.json without either field loads with 30 and 24; zero or a fraction is refused |
What these cannot catch:
- Whether 30 days and 24 hours are right. Both are configurable, and neither is measured.
- Old turns misleading the agent, such as a price quoted a month ago being repeated after it changed. Only an eval case that seeds such a turn would show it, and ADR-0013 names it as the reason to revisit.