[bug] The idle-reap message teaches the user what a session is, then tells them to do something the platform already does by itself

[serge-ivo]

Two issues (#257, #408) went to real trouble to stop "session" being a word the user has to know. One string undoes both, and it is the last thing a returning user reads.

The string

lib/coding-session-sweeper.ts:120-123:

const head = `${IDLE_REAP_PREFIX} ${idleHours} hours with no activity`;
return engineStopped
	? `${head} — the engine process was released. Start a new session to pick the work back up.`
	: `${head}, but the engine could not be stopped on your machine — it may still be running. Check it there if you need the process gone.`;

with IDLE_REAP_PREFIX = "Session closed automatically after" (:109).

This is written into coding_timeline as the run's outcome row, so it is the final line of the repo history the console renders when the user comes back. #257's own comment (agents/coder/web/src/CodingTab.tsx:154) says the history exists so "the user must not be shown an empty screen with a 'Start a session' button where their work used to be" — and then the history's last line tells them to start a session.

It is also not true

The user does not need to start anything, and starting a new session is not how they get their work back — the platform reattaches with continuity on the next engagement. Verified 2026-08-17 on instance bd43f4de: POST /v1/instances/:id/coding/sessions after this exact reap returned "resumed": true with resumeFrom pointing at the reaped session, and the engine recalled its previous closing question verbatim.

So the sentence teaches a concept the user should never have needed, in order to instruct an action the platform performs by itself, using a word ("new") that describes the opposite of what happens.

Owner report, 2026-08-17

"I just tried to ask it to continue, but it said the session has ended and all output pane is gone."

and then:

"why do users have to know about sessions? why can't it just be seamless?"

The reap message is, as far as I can trace, where that knowledge came from.

Fix

The reap itself is correct and should stay — IDLE_SESSION_MS is 6h (:63) and its header explains it is deliberately set below one night so an abandoned tab does not hold a child process on someone's laptop until morning. That is a good decision about a real cost.

What changes is that the reap stops being an announcement:

• Drop the instruction entirely. The user has nothing to do.
• Prefer sleep vocabulary to death vocabulary, or emit no user-visible row at all and let the history simply continue at the next turn. "Asleep" is what actually happened; the work is intact.
• The second branch (engine could not be stopped) IS worth telling a human — a stray process on their machine is their problem to know about. Keep that one, minus the session noun.

Related: #697, which is the sentence that should be there instead — resolveSessionContinuity already returns "the previous conversation on this repo was last touched 11 hours ago" and nothing displays it.

Grep for the noun elsewhere before closing: any user-facing string containing "session" in the coding surface is in scope, not just this one.

[serge-ivo]

Shipped in cfea06ac. CI green, API Worker deployed.

The reap message (lib/coding-session-sweeper.ts) — the instruction is gone and so is the noun:

• was: Session closed automatically after 6 hours with no activity — the engine process was released. Start a new session to pick the work back up.
• now: Went to sleep after 6 hours with no activity — the engine on your machine was released.

The engineStopped: false branch keeps its ask verbatim, per the issue: a stray process on the user's own hardware is the one thing here only they can act on.

It deliberately does not promise the conversation survives. Whether the next open continues it is resolveSessionContinuity's decision, taken later against facts the reap does not have — which engine gets launched, and how long the user stays away. That reassurance lands in #697 (console, live) and #698 (notification).

The marker rename was not free, and this is the part worth recording. IDLE_REAP_PREFIX is not a constant the code merely agrees on — it is the literal text of rows already in coding_timeline, read back by lastIdleReapForRepo with a LIKE. Rewording it without carrying the old spelling would not have renamed the marker; it would have hidden every reap written before the rename, silently, and the chat surface would have gone quiet again for exactly the users #407 was about. LEGACY_IDLE_REAP_PREFIXES is append-only and asserted.

The grep found two more, both fixed:

• sessionOpenedNotice restated the reap verbatim ("The previous session for this repo had been closed automatically after 6 hours…"). One event, two spellings.
• The same notice opened Started a coding session for X … (session csess_…) and then instructs the model to repeat the sentence to the user. That was the widest channel the platform had for teaching the noun, firing on every agent-initiated open. Now Started work on X … (id csess_…) — the disclosure it exists for is intact (a process appeared on their machine; the id is still the support handle) without naming a thing the user owns.
• continuity.reason had four of seven branches carrying "session". Not diagnostics: the console renders them on open ([bug] continuity.reason is computed, phrased for a human, returned on every session open, and displayed nowhere #697) and MCP returns them ([bug] No MCP tool can open a resuming session — the only opener hardcodes fresh:true, so #408's continuity is console-only #696), so they are read by the user at the exact moment they are wondering whether their work survived. All seven now say "conversation" or "attempt", pinned by a test that enumerates every branch, so adding one trips it.

Left deliberately, and reported rather than folded in: lib/end-coding-session-tool.ts still says "session" throughout. There the noun is the user's own — they asked to end a session — and substituting our word into the reply to their own request is a different change needing a different justification. Say the word if you want it done and I will file it.

Verified: pnpm typecheck, pnpm vitest run, biome, all nine script gates, docs:drift, docs:build, console build. Not verified live — the reap fires six hours after a real idle session, which cannot be forced from here; what is deployed is the code path, tested.