v1.23.0 — agent-sync

SessionStart says only what this run must act on (AS-25). The hook printed the full
status at every startup, resume and /clear: 834 characters in the hub checkout
(measured 2026-10-20), mostly the setup and other runs' residue, closing on "NEXT: acquire
a lease", which the family's routing block already says and a read-only session never
needs. It now runs status --brief: one line naming the leases this run holds, a lease
held past scheduledIntervalSeconds, this run's own expired locks (with reap) and claim
tags out of step with the leases, or nothing at all. The brief reads local state only (no
record plane, no remote, no vault) and leaves the signal watermark alone, so the next
status still shows what is new. It prints on startup only; resume and clear still
stamp the session (the matcher keeps both, ASY-08) and print nothing, without starting
python. AGENT_SYNC_SESSION_VERBOSE=1 restores the full block; status is unchanged.
Measured interleaved on this machine at load average ~470: startup 834 to 0 characters,
median 1037 to 533 ms (p90 1399 to 853 ms, n=15 each); resume 920 to 63 ms median.

The PreToolUse guard answers inside its budget (AS-26). In git lease mode the first
write of a guarded file took that file's claim on the lease remote, synchronously: one or
two ls-remote calls (1.6 s each to github.com) and a push with no timeout, plus a cloud
record-plane append where one is configured. Measured against github.com, the 1.22.2 guard
took 8.6, 8.9 and 10.9 s for three first writes; the 2026-10-20 audit saw 10 to 20 s
against the hook's 20 s timeout, past which the host lets the call through anyway. The
claim now runs in a forked, detached process and the guard waits for it only until
AGENT_SYNC_GUARD_BUDGET (1.5 s from the coordinator's start by default; 0 restores the
synchronous claim). Past the budget the write proceeds under the task lease; the claim
then lands and leaves this run's local note, or records the run that holds it, and the
next write of that file is refused from that record without a round-trip (rechecked after
60 s). A second write of the same file while the claim is in flight waits for that claim
instead of racing it, and a claim this run already won is never reported as another
run's. The acquire push is bounded at 60 s. Local leases on a local record plane are
unchanged: they are a file create and are never backgrounded. Same three files on the
same remote with 1.23.0: 4.0, 5.2 and 4.0 s, with the remaining time spent in local file
writes on a machine at load average ~500 (two 50-byte writes profiled at 1.2 s).

The skill description is a router again (AS-27): 445 characters, down from 936; every
trigger the hub routes on (lib/triggers.js, route agent-sync) is still in it, in
English and Russian, with the boundary and the refusal phrase. The "before editing a
shared registry, even unasked" rule moved into the body.

• Tests: test/hooks_session_test.py gains 8 cases (7 watched failing on 5bd8bec);
  new test/guard_budget_test.py, 5 cases over a local bare remote with a git shim that
  slows ls-remote and push by 5 s (3 watched failing on 5bd8bec: guard calls of
  11.06 s, 5.51 s and 11.36 s). Both run in npm test and CI.
• Not covered: the window between a write that proceeds and its claim landing; a run that
  writes a file another run claimed in that window is refused only from its next write.
  Trigger evals were not re-measured for the new description (backlog AS-28).