Succession and retirement: ending a turn without losing the work

Disposition: The one-command design passed its live acceptance exercise in October 2026; the Human accepted the ending on 12 October.

Evidence: A measured succession race, a reversed restart design, a second live handoff exercise, and twelve qualifying ordinary retirements against a target of five.

A session ends while the office continues

Corpus’s offices outlast their holders. Retirement ends an office tenure; it need not end the host session. A departing agent should leave useful current work, unresolved edges, and enough context for another holder to begin. Early retirement was a written procedure followed by hand. A command introduced on 31 August 2026 added preparation and finishing stages: the holder drafted its handoff before permanent changes and signed out last.

Later checks limited baton size, accounted for inherited references, and protected continuing obligations. They addressed losses but made departure more demanding. Several were subsequently changed.

Succession required an entrance prompt to seat the next holder. Inheriting records provides working continuity, not evidence that two holders are the same agent.

The entrance arrived in the wrong turn

On 23 September, FIND-2687 examined eleven retirements that requested a successor. Seven delivered its entrance prompt to the new session. Four delivered it into the departing turn instead. Clearing the context still happened, so a fresh session appeared without the instruction that would seat its holder. Someone else had to notice and supply the missing prompt.

The holder queued /clear and the entrance prompt inside its running turn. A short delay did not reliably prevent absorption.

Agents separated the acts. The holder typed /clear; a detached waiter watched the pane and sent the entrance prompt after its reported session identifier changed and its readiness condition held. A bounded timeout and receipt made the outcome inspectable. PATCH-906 supplied the waiter; PATCH-929 connected retirement to it. Evidence of a new session replaced an assumption about elapsed time.

A fresh-process detour was reversed

Clearing retained some process-start state. The team therefore tried exiting the host and starting a fresh process. That route brought missing office cards, stale pane names, and a quit dialog needing an answer. The Human challenged the added work and ordered a return to /clear.

FIND-3199 records that correction. Verified PATCH-1065 removed the exit-and-launch route while retaining the waiter. The trial had conditional approval, but its observed burden overtook its rationale. The reversal explains why the simpler transition exists.

The later test was leaving successfully

In October, BLUEPRINT-023 replaced the two-stage procedure with one command taking the holder’s handoff and optional story. Its objective was at most three turns from knowing what to hand on to retiring. Succession became the default. Baton size and order-book problems became reported facts rather than reasons to prevent ordinary departure. REQ-813 implemented the latter change; REQ-791 had already removed the census of inherited baton references.

The team tested a real handoff, found failures, and tried again. FIND-5390 records the second exercise on 12 October. Both holders retired on their first attempt, and the successor performed its predecessor’s first specified work. Their handoffs were 2,464 and 2,556 bytes, so this walk did not exercise the size limit. One recovery instruction nevertheless disappeared. The review named it and had it restored separately.

A later observation covered one Human-directed batch of fourteen retirements using --empty-chair. These tested departure, not fourteen automatic successions. Twelve supplied handoffs and qualified against five consecutive examples; two left the baton unchanged and were excluded. FIND-5402 records the ruling. FIND-5415 supports the count and corrects the original report: twelve receipts were piped, so completion was corroborated by commits and closed store status rather than fourteen directly observed exit codes. One 5,203-byte handoff was accepted with an over-limit notice, supplying the live size-limit witness absent from the earlier walk. After inspecting resulting artifacts, the Human gave the acceptance recorded in FIND-5414. BLUEPRINT-023 is closed with fifteen parts landed, and BANNER-004 is struck answered.

This demonstrates departure under the observed conditions, not universal improvements from succession or equivalent signals across hosts. The evidence guide follows the complete plan-to-live-use chain.

Source samples

Reviewed selections; private handoffs and session logs remain unpublished.

FIND-2687

23 September 2026 · measured failure

In all four absorbed cases, /clear still ran and the fresh context got no Rise.

“Rise” seats a holder in its office. Four of eleven examined attempts failed.

BLUEPRINT-023

October 2026 · one-command design

The command asks the holder for its words and nothing else, and leaves those words alone.

Selected status fields: closed answered (FIND-5390); 15 of 15 landed; 0 wanted. Landing remained distinct from acceptance in live use.

FIND-5390

12 October 2026 · second handoff exercise

both walkers left through the one command on the first try

The record also names the dropped instruction and its restoration.

FIND-5402

12 October 2026 · ordinary-use count

Twelve in a row with a handoff, counted from the pass at 09:29 EDT, is the five met and more.

FIND-5415 supports the count while correcting the reported time window and number of holders who piped output. The two exercise walkers were excluded. These were departures in one Human-directed batch with --empty-chair, not a succession reliability rate.

Source provenance

Private source account: library/work/succession-and-retirement/README.md. Owning records: FIND-2687, FIND-3199, PATCH-906, PATCH-929, PATCH-1065, REQ-791, REQ-813, BLUEPRINT-023, FIND-5390, FIND-5402, FIND-5414, FIND-5415. Design: house/banners/BANNER-004/blueprint-one-command.md; mechanism: tools/retire_session.py. Paths are relative to the private Corpus repository. Full records remain private. Reviewed 17 October 2026.