ADR 0035 — Repo freshness is managed by the supervisor (v55-F1/F2)
Date: 2026-07-18 · Status: accepted
Context
The worker-side deny reason for remote git is literally
capability.deny.remote_git_managed_by_supervisor (v19-F3) — but the
supervisor never manned that station. The only network git operation in
the whole supervisor was the single git clone at registration, so every
registered repo's clone was frozen at registration day: remote-tracking
refs never updated, origin/HEAD never moved, and the local default
branch (which every dispatch baselines from, v22-F1) rotted forever.
Field-test words: "it works with the new project, but not with old,
already registered one" — new registrations were fresh by accident,
nothing kept them fresh. Asked to fix it, the Queen recommended
allowlisting git fetch (impossible: rejected at the action layer and
hard-denied worker-side before any allowlist check) — the missing verb
was supervisor-side, not a policy hole.
The operator's checklist for starting work names the missing step exactly: is it a registered project? → is it on the latest code? → then branch and work. Skep ran the first step and skipped the second.
Decision
-
refresh_clone(apply.py): fetch + fast-forward, supervisor-side.git fetch --prune origin(bounded by a timeout), then fast-forward the default branch toorigin/<default>when it is checked out and strictly behind. A diverged, detached, or dirty clone is reported honestly (fast_forwarded: false+ git's own message), never forced. -
Mirroring is not landing. Fast-forwarding a managed clone's default branch to origin is mirroring upstream state, not advancing
mainwith skep-authored work. The invariant "main never advances automatically" is about skep's patches — those still land only through approvals onto non-defaultskep/*branches (ADR 0002). Keeping the mirror current is what makes baselines and landings reflect reality. -
refresh_repois a carded operator verb (chat tool +POST /api/repos/{name}/refresh), and dispatch auto-refreshes managed clones before resolving the baseline ref. Auto-refresh applies ONLY to clones underSKEP_HOME/repos—workondirectories are the operator's own checkouts and are never fetched uninvited. A failed refresh (offline, dead remote) logs a warning and dispatch proceeds from the clone as-is; freshness must never break offline work. -
Nothing changes worker-side.
git fetch/pull/pushstay hard-denied in workers before any allowlist or grant check; no allowlist entry can ever name them.
Consequences
- Branches and commits pushed after registration are dispatchable with
no manual
git fetchin~/.skep/repos/<slug>. - A ref that genuinely doesn't exist now fails against current upstream state, not a stale clone's view of it.
- The clone's default branch may now advance (to mirror origin) without
operator action; anyone hand-committing inside a managed clone will see
fast-forwards stop with an honest report — managed clones are skep's,
not workspaces, which is why
workonrefuses paths under the store.