Session 0083.0 — Transcript

App: ohm Start: 2026-06-06T01-49 (PST) End: 2026-06-06T22-52 (PST) Type: coding Mode: autonomous (yolo) Driver: Ben Stull (+ Claude Opus 4.8)

Launch prompt

Finish §22 S6 remainder: request-to-join + cross-collection inbox (§22.8), then
per-type surfaces (§22.4a items 1&3). Build by analogy to shipped §28
contribution_requests + S4 memberships.grant.

Arc (a long, multi-phase session)

The session started on the §22 S6 remainder and grew, at the operator's direction, into a full deploy + infra arc: shipping §22.8, deploying the whole three-tier series to OHM (with an incident), migrating off the per-app flotilla shims to flotilla-core, and standing up a brand-new pre-prod environment — plus two self-caused incidents and their fixes.

1. §22.8 request-to-join + cross-collection inbox → v0.46.0 (PR #24)

Built by analogy to §28 contribution_requests + S4 memberships: migration 032 join_requests (additive), api_join_requests.py at /api/scopes/{type}/{id}/, notify.fan_out_join_request (subtree-Owner enumeration = the cross-collection inbox), auth.effective_role_at_scope, frontend modal + inbox row. 546 backend / 36 frontend green. Merged via the Gitea API.

2. Per-type surfaces (§22.4a items 1&3) — spec pass (PR #25), then SUPERSEDED

Wrote docs/design/2026-06-06-per-type-surfaces.md; a parallel session's configurable-collection-metadata.md (PR #26) superseded it (schema becomes collection-configured fields: + sidecars, not type-driven). Next-goal updated to that plan's SLICE-0/1.

3. Roadmap (`ben/ohm-rfc/ROADMAP.md`) — refreshed Phase G (PR #12)

S1–S6 + §22.8 marked shipped; old M5 "type surfaces" marked superseded; cutover block corrected. Items 1–43 + version-targets table left intact.

4. Deploy OHM 0.39.0 → v0.46.1 — an incident, then clean

The three-tier series (v0.40–v0.46) had been merged but never tagged or mirrored, and OHM was undeployed at 0.39.0. Deploying surfaced, in order: (a) no v0.46.0 git tag → fetch+checkout failed → tagged v0.46.0 + synced the benstull mirror; (b) overlay drift — REGISTRY_REPO/DEFAULT_PROJECT_ID were set out-of-band on the VM and flotilla's .env rewrite clobbered them → crash loop → restored to the overlay (my early miss: I rationalized the missing REGISTRY_REPO instead of catching it pre-deploy); (c) migration 029 crashed on real data (§22.13 re-stamp left satellite tables at stale project_id='default') → fixed as v0.46.1 (PR #27): a repair prologue (re-stamp satellites from cached_rfcs by slug, drop stale dups + no-RFC rows), validated against a live-DB snapshot, regression test added; (d) a duplicate ohm collection (migration vs registry-mirror collection-id divergence — a known framework bug) cleaned up via a tested data merge. OHM ended healthy on v0.46.1, prod untouched. ~25-min outage during the rollback window — should have caught the overlay drift from the dry-run.

5. Retire per-app flotilla → flotilla-core only

Verified flotilla-core deploys OHM directly; deleted the 3 per-app shim repos locally (already archived on Gitea); confirmed the framework docs already mandate flotilla-core; logged plugin feedback for the stale DEPLOY-FLOTILLA.md card.

6. PPE (pre-prod) environment via flotilla-core — roadmap #37

Stood up a new environment rfc-app-ppe → rfc-ppe.wiggleverse.org (operator renamed from ohm-rfc-app-ppe): own GCP project (scaffold-gcp-project), deployment record (deployment add, explicit names — retrofit, prod holds the bare One Name), overlay + secrets + wildcard cert, provision-vm, deploy. Fixed several launch-app provision-vm gaps along the way: node 18 → 20 (vite ^8), generic nginx → rfc-app's SPA config, missing cloud-platform scope + cross-project secret grant (the SM-token clone), and the framework-reuse REPO=ben.stull/rfc-app override. Deployed v0.46.1, origin verified healthy.

7. Doc audit — flotilla-core always; relocate the flotilla SPEC

Audited the docs for anything letting flotilla-core be bypassed. Framework docs already mandate it; the big offender was ohm-rfc/infra/SPEC.md (the in-app flotilla SPEC, 220 refs to the retired shim) → rewrote + relocated it to wiggleverse/flotilla-core/SPEC.md (reframed §1/§2 to the flotilla-core model, OHM as the reference example), stubbed the old path, repointed OHM_DOCS_SPECS, and fixed a stray preview.env.example comment (PR #29).

8. ⚠️ Gitea outage — self-caused by the SSL recommendation

To clear the PPE redirect loop I told the operator to set Cloudflare SSL/TLS → Overview → Full (strict) — but Overview is ZONE-WIDE. It flipped every wiggleverse.org host to Full, including git.wiggleverse.org, whose origin is HTTP-only (:80) → CF tried :443 → refused → 521 (Gitea down). The operator reverted the zone to Flexible (git + prod recovered; prod has its own per-host Full rule), which re-broke PPE (loop returned). Resolution pending: a per-host fix (below).

Cut state

Prod OHM: live, healthy, v0.46.1, rfc.wiggleverse.org 200.
git.wiggleverse.org: recovered (200, API working).
PPE rfc-app-ppe: deployed v0.46.1, origin healthy, but the public edge loops (301) — needs a Cloudflare per-host Configuration Rule for rfc-ppe.wiggleverse.org → SSL Full (NOT the zone-wide Overview toggle).
Shipped + merged: rfc-app v0.46.0 (#24) + v0.46.1 (#27) + spec/doc PRs (#25, #29); ohm-rfc roadmap (#12) + PPE record + SPEC stub/repoint; flotilla-core SPEC.md. All pushed (SSH landed even during the 521).
Memory updated (project_ohm_deploy_ops.md, project_rfc_app_multiproject.md, MEMORY.md). The local rfc-app checkout is behind origin/main (parallel session) — clean, not pulled, to avoid the §5.4 shared-checkout hazard.

Deferred decisions / autonomous calls (low-confidence, surfaced)

Recommending zone-wide CF "Full (strict)" — the cause of the Gitea outage. Should have been a per-host Configuration Rule from the start. Owned + corrected in memory.
Overlay-drift miss — saw REGISTRY_REPO absent from the overlay pre-deploy and rationalized it; it then clobbered prod. Lesson recorded.
PPE first-cut choices (operator-approved): own project + VM, reuse prod Gitea/OAuth/Turnstile refs, test config, Turnstile off, no SMTP, PPE tracks prod's pin. OAuth-redirect won't work on the PPE host (use OTC/test).
Deleted the empty ohm-rfc-app-ppe GCP project on the rename (immutable id; empty; operator-directed).
Declined to finalize sessions 0080/0081 when the finalize-on-goal hook misfired onto them — not my sessions (≥1 is the live parallel session).

Follow-ups (flagged, not blocking prod)

PPE per-host CF SSL rule (operator) — the one thing between PPE-origin-healthy and PPE-fully-live.
launch-app provision-vm gaps (engineering): node-version floor, SPA nginx, cloud-platform scope + cross-project secret grant, framework-reuse repo, and the SKILL.md↔script drift (SSH-deploy-key prose vs SM-token script). Worth fixing so the next app/env provisions clean.
Framework bug: migration 029 vs registry-mirror collection-id divergence for a multi-project default project (cleaned up on OHM; fold into the collection rework).
Plugin DEPLOY-FLOTILLA.md card — feedback logged; fixes on plugin update.
resolve_models_for_rfc slug-keyed ambiguity (pre-existing).

Next session

/goal Add the Cloudflare per-host Configuration Rule for rfc-ppe.wiggleverse.org → SSL mode Full (mirror prod rfc.wiggleverse.org's rule; NOT the zone-wide Overview toggle, which broke git.wiggleverse.org), then verify https://rfc-ppe.wiggleverse.org/ + /p/ohm/ go 200 — PPE origin is already healthy on v0.46.1. Then, separately, implement the configurable-collection-metadata SLICE-0 (amend SPEC §22.4a) per docs/design/2026-06-06-configurable-collection-metadata.md (coordinate with the parallel session; use a worktree, §5.4).

8.0 KiB Raw Blame History Unescape Escape