Siblings stay — the two delegation lanes under obot-prime

Decisions

Siblings came off Remote Control

Keeping siblings had an immediate consequence for how they are set up. Since July every sibling had started bridged to Remote Control, so @jwildfire could open any of them from claude.ai/code or his phone. That was his instruction at the time. It stopped matching how he works the moment prime became the single front door: on the evening of the same day he said “I think we can turn off the Remote control default for subagents” — meaning siblings, since the in-conversation lane never had a bridge at all. He does not open sibling sessions any more, so roughly twenty remote-control sessions a day were being registered for a window nobody looked through.

The obvious fix turned out to be the wrong one. Removing the remote-control flag from the spawn command changes nothing, because a separate account-wide setting turned on in July bridges every session on its own; the flag and the setting are alternatives, not a chain. The spawn now switches that setting off for the one process it is launching, which is something an agent is allowed to do without touching any settings file. Three throwaway probe sessions confirmed all of it on the current release: unflagged still bridges, the switched-off spawn does not, and switching it off disturbs nothing else about how the session runs.

Two things were checked rather than assumed, because the decision above depends on them. Prime can still message and resume a sibling that is not bridged — that lane is local to the machine and never went through Remote Control — and unbridged siblings still appear in the terminal’s own agents list. So the resumability that this page cites as one of the four legs the decision stands on is intact. One small benefit came free: every bridged sibling used to leave a permanent entry in the account’s remote-control session list, and that list had grown to forty-seven dead sessions crowding out the four that were actually running.

The sessions @jwildfire actually reaches from his phone are untouched. Prime still bridges, and so does the unattended overnight lead. An interactive reviewer — the one kind of sibling he talks to directly — is opted back in by name. What is genuinely lost is a canary: the bridge check used to run after every spawn, about twenty times a day, which quietly monitored a Claude Code feature combination that is undocumented and could break in an update. It now runs only on the rare opt-in spawn, so a future break would most likely show up as @jwildfire not finding prime on his phone rather than as an agent noticing first.

Filed and shipped as Turn off the Remote Control default for background siblings and the pull request that closed it, merged 2026-08-15. Full mechanism, probe evidence and triage table in the Remote Control runbook.

The context cost is structural, not stylistic

When prime runs work as an in-conversation subagent, the subagent's final report is delivered back into prime's own conversation — it becomes part of the memory prime carries forward. When prime spawns a sibling, the work lands on disk and on GitHub, and all prime carries is a short report message. On 2026-08-14 roughly twenty siblings ran in one day, and again about as many on 2026-08-15 — several with very large outputs: a sixty-artifact description backfill, a milestone sweep across seventeen published releases, a cross-repo widget parity matrix. Routed through subagents, all of that would have flowed through the one session that is designed to run indefinitely under a strict memory budget — the same budget an entire decision artifact (how prime remembers) was written to protect. The front door cannot also be the loading dock.

One honest caveat belongs in the record: the context cost alone is partly mitigable — a subagent can be told to write its output to disk and return a one-line pointer, the same discipline the sibling briefing already enforces. The decision is safer than its single stated reason, because the comparison this page grew out of found three more legs it stands on:

• Durability bit twice on the deciding day alone. A sibling died mid-task on an unrecoverable API error, and its work was recoverable precisely because it was a separate session: a replacement sibling reconciled against GitHub and the dead session's own record, then finished the job without double-posting anything. Separately, a background watcher inside a session was killed by the harness without notice. Subagents die with the session that spawned them — and a standing front-door session is exactly the kind of session that restarts.
• Siblings are resumable across hours. The governance sibling was resumed three separate times across ten hours of real work — briefed once, then re-engaged with its context intact each time a new approval landed. That capability was used heavily by prime on the deciding day even though @jwildfire himself never attached to a sibling.
• Siblings take a per-task model and reasoning-effort choice. The judgment-heavy work of the deciding day ran on the strongest model at high effort while mechanical chores ran cheap; the in-conversation lane can choose a model but not an effort level.

One test, no deliberating