Public: anyone can read this; nobody can change it.
GATE: assert exactly ONE system webhook after a dispatcher restart + reviewer engages without the 10-min fallback #67
Closed without the crew
Second half shipped: nightly app-E2E in the cold-start gate (https://git.open-platform.sh/plat/mitosis/pulls/75)
Covered as of that PR (all in scripts/coldstart-assert.sh, the ONE gate both CI and humans run):
- (1) webhook singleton: already covered — A2 (exactly one admin hook, probed BY ID) + A3 (restart
deploy/agents, STILL exactly one; version-aware log check). - (2) reviewer engages without the ~10-min fallback: now asserted nightly by
--e2eon the CI daughter, with a REAL canary app generated from_app-templateand shipped unattended:- E4 prime PR auto-merges (builder → preview → reviewer verdict → merge)
- E5 prod serves 200 in-cluster
- E6 dispatcher log has zero
preview liveness unconfirmedlines - E7 builder-finish → reviewer-queued < 120 s (the fallback path takes ~10 min, so it cannot pass this)
- E8 the app's prod namespace carries
plat.sh/app+ generated ResourceQuota + NetworkPolicies
Runs on the nightly cron and on any workflow_dispatch with e2e=true, AFTER the standard gate, hard-capped at 35 min. Humans: scripts/coldstart-assert.sh --e2e <ctx> <domain>.
Default-safe: the phase is SKIP until you seal a credential. No secret exists today (the staged host token is flagged compromised-pending-rotation, so it was deliberately NOT wired in). e2e-flagged runs currently print one loud E2E SKIP line and exit 0 — the standard gate still gates.
To turn it on (one command, with a FRESH Claude Code OAuth token):
curl -X PUT -H "Authorization: token " -H 'Content-Type: application/json' \
-d '{"data":"<fresh-claude-code-oauth-token>"}' \
https://git.open-platform.sh/api/v1/repos/plat/mitosis/actions/secrets/E2E_CLAUDE_TOKEN
What happens once sealed: every nightly run (and every e2e=true dispatch) germinates the CI daughter WITH the credential sealed into plat-agents-secrets, then runs E1–E8 above. Rotation = same PUT with a new token; DELETE .../actions/secrets/E2E_CLAUDE_TOKEN returns to loud-SKIP.
The full --e2e path was validated manually on a scratch germination (evidence in the PR body) — marked manual-validation there since CI itself can't run it until the secret is sealed.
Both gate assertions this issue asked for are now LIVE in CI (PR #75 merged, its own gate run 14 green): (1) exactly-one-webhook-after-dispatcher-restart = A3, running on every gate since #71; (2) reviewer-engages-without-the-10-min-fallback = E6/E7 in the --e2e mode, proven manually end-to-end (unattended canary shipped in 793s, builder→reviewer gap 34s, zero fallback lines). The nightly E2E activates the moment a FRESH Claude credential is sealed: curl -X PUT -H "Authorization: token <PAT>" -d '{"data":"<token>"}' https://git.open-platform.sh/api/v1/repos/plat/mitosis/actions/secrets/E2E_CLAUDE_TOKEN — until then it SKIPs loudly and exits 0 (the compromised staged token was deliberately NOT wired). Closing: the assertions exist, run, and are proven; activation is a one-command operator step.
Two failure modes found while validating today's wb-gate germination (fresh plat.local daughter) both SELF-HEAL: the app still ships and serves 200, so a naive E2E passes while the platform is degraded. Exactly the class of regression the acceptance gate (#46) / CI-1 harness (#21) must assert explicitly, or agents 0.7.6 (#66) fixes them once and nothing keeps them fixed.
(1) exactly ONE system webhook survives a dispatcher restart (#63). Proven live: one
kubectl rollout restart deploy/dispatcher→ admin hooks id 1 AND id 3, same dispatcher URL, both active → every Forgejo event delivers twice (doubled runs/noise; the RBAC-5 HMAC re-seal can never patch either). Fix in flight: plat/agents#10 (persist the created hook id; GET-by-id on boot). Gate assertion: restart the dispatcher twice, then assert exactly one admin hook targets the dispatcher URL — probeGET /api/v1/admin/hooks/{id}by id, since the LIST endpoint demonstrably returns[]while hooks exist — AND one PR event produces exactly one dispatched run.(2) reviewer engages without the ~10-min fallback (#57, root-caused today: the liveness probe resolved the preview via loopback from inside the dispatcher pod, so it never confirmed a preview that was already serving 200 with the correct /version sha). Fix in flight: plat/agents#11 (in-cluster DNS first). Gate assertion: on a PR build, the reviewer engages within 3 min of the preview serving 200 with
/version== PR head sha, and dispatcher logs contain NO "preview liveness unconfirmed" fallback line.Fix direction: add both as assertions in the CI-1
validate-coldstartset (#21) and the #46 unattended gate run; validate first on the #66 release-train germination. Feeds #46 — does not block the MVP path.