Status: Accepted · Recorded: 2026-11

Amended by ADR 0075 (a mains settling read makes up to nine attempts, early ones listening a shorter window) and by ADR 0076 (the read at the predicted end of a move is a settling read; a stopped report never schedules from its count), and by ADR 0079 (a settling status read to a low-power cover in an unknown state leads with one short attempt before its three).

Amends ADR 0073 (which reads take the plan), ADR 0066 (the refresh takes the settling plan), the "three tries" of ADR 0040 (the poll after a stop) and the worst-case blocking arithmetic of ADR 0013.

Context

ADR 0073 gave the engine an attempt plan and a guard that keeps every attempt inside the exchange budget, but left every caller at today's attempt counts. A status read that decides where a cover ended up is the read worth repeating: after a stop nothing moves, the entity shows the last reported position until the read is answered, and a lost frame costs the whole backoff slot (5 s or more) before the next look. A read during a move is the opposite: the device is busy and its silence is expected, so a repeat changes nothing and only fills the air.

Decision

Which reads settle. A status read is a settling read, and takes attempt_plan() with ReadKind::SETTLING, when it is
- the poll after an accepted stop,
- the poll that confirms a cover that reported stopped short of its target (the hub marks it when it sees that report, as it marks the poll after a stop),
- a scheduled poll in the middle band of the backoff ladder (the three slots after the first miss), or
- a read with no ladder behind it: a one-off poll and the opt-in status refresh.
Progress reads stay single tries. Every other scheduled poll of a cover (the first slot after a command, the slots past the band, a challenge-answering device) keeps one attempt and the backoff ladder as its retry. This is a deliberate airtime and responsiveness choice: the device is probably busy, and a stop pressed during a move must not wait behind a burst of reads.
Heaters keep their counts. Heater and water heater polls, rain, sun and battery polls, and commands keep the count-based budget (scheduled_poll_max_tries(), EXCHANGE_RETRY_COUNT). Heating-level heaters share the cover status path but branch before the plan.
Callers ask the engine for the plan. ExchangeEngine::plan_read() builds the plan for a request, with the transmit time the engine itself would plan for it (the radio's estimate for the request's start preamble), so the plan and the budget guard agree. The hub passes a read kind where it used to pass a count.
The post-stop poll, the refresh and the one-off poll no longer say "three tries". They make as many attempts as the plan gives: up to six at the defaults for a mains-powered cover (a 32-byte start preamble and a 400 ms window), three for a low-power one. On the bench boards (SX1276, SX1262 and LR1121) a read to an address no device answers made five attempts 422 ms apart and gave up after about 2.15 s: the guard counts the time that has really passed, and each attempt's overhead on the radio leaves no room for a sixth.

Consequences

Worst case on the loop. A settling read whose every attempt is lost blocks the loop for nearly the whole budget (about 2.5 s at the default), where three attempts with gaps blocked it for about 1.6 s. A stop queued during it runs as soon as the read gives up; the operation queue runs a user command ahead of a queued poll, so at most one settling read stands in front of it. The loop stays below ESPHome's 2550 ms warning because the guard bounds every attempt (ADR 0073).
Airtime. A settling read of a silent mains cover sends up to six requests instead of three. The refresh pays this once per interval for a cover that does not answer. The status read of a cover that does answer is unchanged: it returns at the first reply.
Low-power targets. A settling read to a low-power device makes three attempts, each listening 550 ms after its wake-up preamble, with no gap, instead of three attempts listening 400 ms with a 250 ms gap. A target that challenges has fewer: two challenged attempts leave no room for a third inside 2500 ms, as ADR 0073 states for every challenged low-power exchange. Raise exchange_total_budget_ms for a low-power device that challenges every poll and needs the third.
Stale marks. The mark that makes the next poll a settling read is dropped by a new command, by clearing the tracking window and by the poll that takes it, so a later poll never inherits one.
ADR 0066 and 0040 wording. Where they say the refresh or the poll after a stop has three tries, it reads: the settling plan's attempts.