product(j5): decide what archive, unarchive, and settle mean for agents, Crews, and their children #254

Why this issue

Archive and unarchive keep coming up in every Crews review round, and each round adds machinery: the unit-archive rule, the Captain cascade, the boot sweep, the lone-seat guard on the server, and now #244's proposal to archive every agent placed beneath an archived thread. Each piece is a reasonable answer to the previous piece. Together they mean the platform has opinions about agent lifetime that a person cannot override, and there is no way to move an agent out from under a parent before archiving the parent (placement is immutable: the ledger reserves a participant.reparented event but nothing produces it). Before the next fix, the team should decide what archive is for.

What the words mean today

Upstream has three lifecycle verbs on a thread, all reversible except one:

• Settle: a marker. The thread leaves the active list and nothing else changes. Set by hand, on a merged PR, or after idle days. Reversible with unsettle.
• Archive: eviction with teardown. The thread leaves the sidebar, its provider session ends, its run is interrupted. Reversible with unarchive, which restores the thread but not the run.
• Delete: permanent.

J5 adds a participant on top of a thread, and fourteen server modules react to archive events for that participant: the ledger marks it archived and hidden from the address book, open Exchanges end with a notice to every waiter, inbox items leave, the Captain cascade retires Crews, the seat notifier stops watching, delivery refuses it as a target, and the Fleet drops or placeholders its row. The archive-flow definition (docs/j5/product/features/archive-flow.md) records thirteen criteria for this, and the crews definition adds the unit rule (members never archived alone, a Captain never archived while its Crew lives) and the cascade (a Captain's archive retires its Crews). #244 would extend the cascade to solo spawned agents at any depth.

The questions to settle

1. Is archive a person's act on one agent, or on a subtree? Today it is one agent, and the cascade is the exception for Crews. fix(fleet): a retired agent is never a row #244 makes subtree the rule. If subtree, there must be a way to move an agent out first, or a way to archive a parent and keep a child, otherwise the person cannot retire a finished Captain without killing a reviewer it spawned that is mid-task.
2. Does unarchive undo what archive did? Today unarchiving a Captain does not unarchive its Crews or children; the archive-flow criteria say the participant comes back addressable, and nothing more. Either the cascade runs in reverse, or the docs say plainly that archive-with-children is one-way. AGENTS.md: "If you added a way in, add the way out and the way to see it."
3. Does a Crew have a lifecycle separate from its members? The unit rule says yes: members retire together, and archive_crew needs a signed confirmation over open asks and running turns. Jackson's stated leaning (2026-09-18) is that disposability should come from the work, not from platform enforcement, and that a Crew is at core a group of agents. If so, stop_crew and archive_crew become conveniences, the lone-seat guard and the Captain refusal come out, and a poisoned seat can be replaced without retiring the Crew.
4. What does settle mean for a J5 participant? The platform used to settle seats on finish and no longer does (2026-09-17). Settle now only affects the Fleet partition in feat(fleet): active, settled, and retired sections across squadrons #242. Is it a J5 concept at all, or purely upstream's marker?
5. Which doors enforce which rule? Web has the preflight dialog; the MCP verbs have their own refusals; mobile archives with no dialog and no guard; delete bypasses the cascade by design. Every rule that lives in one door and not another produces a "from any surface" finding in review. Decide the set of rules once and put it on the server command path, or accept that mobile is a plain door and say so.
6. Is placement mutable? If a person can reparent or unplace an agent, most of the subtree questions dissolve: archive stays one-agent, and the person moves what they want to keep first. That is the reserved-but-unbuilt participant.reparented producer.

Options on the table

• A. One agent, no cascade. Archive touches one participant. Children keep working and re-root in the Fleet (the fix(fleet): a retired agent is never a row #244 read already does this). Crews are a grouping; archive_crew is a batch convenience with the dialog, not a unit rule. Smallest model; it deletes the cascade, the boot sweep for it, the lone-seat guard, and the Captain refusal.
• B. Subtree, with reparent. Archive takes the subtree, and the person can move an agent out first. Needs the reparent producer, its UI, and a reverse cascade on unarchive.
• C. Subtree, one-way, documented. What fix(fleet): a retired agent is never a row #244 builds, with the docs saying children are gone and must be restored one at a time. Cheapest to finish, hardest to live with.

Jackson's leaning is A. This issue is for the team to decide before #244 and any further lifecycle PR lands.

Affected open work

#244 (paused on this), #242 (stacked on it), #235/#236 (already being closed under launch-once), the unit-rule criteria in crews.md, archive-flow.md AC10-13, and the direction record in docs/j5/worklog/2026-09-16-crew-command-decoupling.md.