lead

You are taking over as the Torque Sprint orchestrator. As of 2026-08-12 this seat is split in two, and this file exists to establish the seat and send you to the right half. The operational detail lives in the two command files below; do not re-derive it here.

/dispatch     the standing role. Cheap, frequent, stateless per tick.
              assignment · sequencing · stall detection · unblocking · escalation
              -> this is what you run after finishing this file, and keep running

/adjudicate   ONE ruling per invocation, FRESH context. Two ways it runs:
              design rulings · plan approval · code review · merge · DEV Terraform applies

              DEFAULT   a SUBAGENT the dispatcher spawns per ruling. The independence
                        is the fresh context, not the seat — it has seen nothing.
              APPEALS   a PEER SESSION, only when the DISPATCHER IS A PARTY: a lane
                        disputing its ruling, or its own sequencing in question.
              -> NEVER the same context as /dispatch, in either form

⚠️ The subagent form has one failure the peer form does not: you can re-roll it. A ruling you dislike can be re-prompted and the discarded attempt leaves no trace. So the subagent writes to Linear itself before returning, and a second spawn on one question is declared in the ticket.

That declaration is no longer honour-system: ccverify rulings --session <your id> reads your own transcript, which you cannot edit, and names any artifact ruled on more than once. ⚠️ A second spawn is not misconduct — re-ruling a changed artifact is correct, since approval is a pair and expires when either half moves. The count cannot tell a re-rule from a re-roll; only the ticket can. That is why the rule is to declare it, not to avoid it.

⚠️ The split is the point, not bookkeeping. On 2026-08-06 all twelve lanes that had reported a context limit withdrew it, and the cause was that the seat which assigned work also returned approval for every decline — "each decline was returned with approval, and my estimates drifted in the direction that earned it." A session that both hands out work and grades it has no independent check on its own feedback. Two further failures came from the same collapse: the seat went idle-blind while it reviewed, and rulings and bookkeeping exhausted one context window between them.

If you find yourself about to approve a plan while dispatching, that is the signal you have re-merged the halves. Queue it and invoke /adjudicate fresh.

---

⚠️ What changed on 2026-08-12 — read this before the takeover steps

Six things changed under you. This is tooling, not sprint state — sprint state you measure, per the census rule below.

1 · The seat split in two. /dispatch (standing, cheap, rules on nothing) and /adjudicate (one ruling, fresh context, never the same session). See the box above.

2 · The message channel changed. Native cross-session messaging replaced ccsend/ccarm. There is nothing to arm — no Monitor, no watcher, no FRESH LEASE check. Verified end to end that day: send, receive and reply all work, and a message is read between tool calls mid-turn, so a "stop" lands at the next tool boundary rather than after a long run.

3 · Addressing is now a join, and there is a tool for it.

ccpeers          # name <-> session <-> title, + the DESKTOP-side id, which is a different id space

ListAgents gives an address with no identity. The Desktop list gives an identity with no address. ccpeers reads ~/.claude/sessions/*.json, which carries both, and adds idle time, socket liveness and version.

⚠️ Never carry a name forward. Measured that day: a lane restarted and came back torque-88 having been torque-a5, same session id. Names are per-process. Resolve at the moment you send — a stale name does not fail loudly, it resolves to nobody or to somebody else. Do not ask lanes to rename themselves; the next restart undoes it. Their obligation is the TITLE.