ADR 0163: show turn elapsed time in the status box

Date: 2026-09-03 Status: Accepted

Context

The closing status box reported session, a cumulative clock, and nothing about the turn it was closing. An operator reading “session 9m” could not tell whether the work just handed to them took twelve seconds or eight minutes, and comparing two boxes in the same session told them only that time had passed. The one number they most often want, the cost of the turn in front of them, was the number the box did not carry.

The evidence already existed. A prompt records turn.started and the Stop hook records turn.completed, and the Stop hook’s own policy already resolves a turn window from those events.

Alternatives considered

• Store a turn-start timestamp on the heartbeat. That adds a second writer for a fact the ledger already holds authoritatively, and the two would drift after a crash or a resume.
• Reuse the Stop hook’s resolveTurnStart directly. It answers a policy question, including a Claude recovery-continuation rule tied to stop_hook_active, which the status command has no input for.
• Measure from the last tool call instead of the prompt. That reports the tail of a turn rather than the turn, and reads as a much cheaper turn than the operator actually paid for.
• Report the previous completed turn instead of the current one. The box is issued inside the turn being measured, so the current turn is both available and the one the operator is about to read.

Decision

The box carries a turn row directly under session, and --json carries turn_elapsed_secs with a turn_complete flag. The measurement runs from the newest turn.started for this owner to the moment the box renders, or to a turn.completed once one has closed the turn.

A Stop-hook bounce ends one turn and opens another carrying stop_remediation, so the walk back folds those restarts into the turn they retried and reports elapsed time since the operator’s prompt. The walk stops at a generation boundary, so a resumed session never absorbs the previous generation’s turn. A terminal counts only when it follows the newest start, which keeps a bounce’s own turn.completed from ending the measurement mid-remediation.

The row is best-effort in the same sense as the qa row. An unreadable ledger, a ledger that fails validation, or a session with no turn evidence renders no row rather than failing the status command.

The format keeps seconds visible below an hour, because most turns land there and a bare 4m cannot be compared with the turn before it.

Result

session and turn now read as two different clocks, and a long session with a short turn is legible at a glance. The row costs one ledger read, which the reader already serves incrementally from its cache. Because remediation restarts are folded, a turn that the Stop hook bounced reports the whole cost of the retry cycle rather than the last attempt, which is the number worth seeing.