| name | live-evidence-branch-freeze |
| description | Run BEFORE any force-push, rebase, branch restart (`checkout -B`), or squash-merge-and-reuse of a branch in this repo — and BEFORE dispatching any SSH-free/self-reporting RunPod lane (open-judge, virtue-evals, independence-eval-v2, long-horizon, wisdom-pilot) — to check whether live pods or workflows are still PUSHING EVIDENCE to that branch. Those pods clone the branch at pod start and push receipt commits back to it; a mid-run history rewrite makes their pushes non-fast-forward, which either times out the launcher (a wasted paid pod-hour) or silently drops the receipts while the Actions run still reports SUCCESS (green-but- resultless). Fires on: "rebase the PR branch", "restart the branch from main", "force-push", "re-use the merged branch", or dispatching a lane whose workflow says "pod self-reports via git push".
|
| metadata | {"short-description":"Never rewrite a branch that live self-reporting pods are pushing evidence to"} |
Live-evidence branch freeze
The incident this prevents (2026-07-06, PR #655 / runs 28772411875 + 28772413187)
Eleven benchmark lanes were dispatched against a PR branch. While three SSH-free pods were
still running, the branch history was rewritten twice (a lease-caught force-push collision,
then a checkout -B restart from main after the PR merged). Consequences, both real:
- open-judge (run 28772411875): the pod finished but its receipt push could no longer
fast-forward; the launcher's 60-min wait timed out and deleted the pod — one paid pod-hour
spent, zero receipts (
[selfreport] wait timed out after 60 min; deleting pod ... to be safe).
- virtue-evals (run 28772413187): the Actions run concluded success, but no
sophrosyne/dikaiosyne receipts ever landed on any branch — the pod's push was silently
orphaned. Green-but-resultless, exactly the failure mode
wisdom-gpu-prebaked §6 warns
about ("Actions success is not proof of a result").
An earlier, luckier variant in the same session: --force-with-lease on the same branch
rejected a push because independence/long-horizon evidence commits had landed mid-rebase;
the 5 clobbered bot commits had to be recovered by cherry-pick (all were, zero loss — but
only because the lease fired and the objects were still fetchable).
The rules
- Before ANY history rewrite of a branch (force-push, rebase-push,
checkout -B
restart, branch deletion): list in-flight runs pinned to it —
curl -sf "https://api.github.com/repos/tomyimkc/sophia-agi/actions/runs?branch=<branch>&status=in_progress"
(also check status=waiting and status=queued). Any self-reporting lane in that list →
freeze: wait for it to conclude AND for its receipts to appear on the branch, or
cancel it deliberately and note the spend. Do not rewrite under it.
- Prefer dispatching self-reporting lanes on
main (or another branch nobody rewrites),
not on an active PR branch that may be rebased/restarted/squash-merged while pods run.
- After a lease-rejected push, investigate before overriding: the commits that moved the
ref may be a pod's evidence.
git fetch the branch, git log <yours>..origin/<branch>,
and cherry-pick/preserve bot commits (authors like github-actions[bot], subjects like
ckpt ..., pod heartbeat, evidence(...)) before any --force-with-lease retry.
- A green self-reporting run is unverified until its receipt is on the branch. Always
confirm the artifact path the workflow names (e.g.
agi-proof/benchmark-results/{sophrosyne,dikaiosyne}/,
agi-proof/benchmark-results/wisdom-market/open-judge-regrade.json) actually gained a
commit. No receipt → treat as failed, re-dispatch on a stable ref.
- Merged-PR branch restarts (
checkout -B <branch> origin/main) count as rewrites:
do step 1 first — evidence pods dispatched pre-merge may still be running against the
pre-merge history.
