Session 0047.0 — Transcript

Date: 2026-05-30 Started: 2026-05-30T11-02 · Ended: 2026-05-30T12-47 (PST implied) Goal: Advance ROADMAP #38 (Track Δ) — turn the session-0046 e2e scaffold (rfc-app PR #4) into a working, green Tier-1 local-Docker suite. No VM work (Track Δ's VM pieces stay gated/deferred to their own session).

Outcome: #38 Tier 1 is live and green. docker compose up --build stands up the full stack (frontend+backend+SQLite+Mailpit+real disposable Gitea) and the Playwright suite passes 4/4, deterministically across repeated runs: health, landing render, and the OTC sign-in vertical (API round-trip + full browser UI). Pushed to rfc-app PR #4 (session-0046/e2e-tier1-scaffold, fast-forward a3f48ba..15af891). One default-secure framework change rode along (ratelimit.py test bypass). Not a release; no deploy; PR #4 left for the operator to merge.

Pre-session state

flotilla: main @ 5f1689f (v1.4.0). v1 has shipped — relevant below.
rfc-app: main @ 0c972c8 (v0.31.0). PR #4 open from session 0046: branch session-0046/e2e-tier1-scaffold @ a3f48ba, an e2e/ scaffold (compose topology + helpers + test.fixme specs) — not yet brought up or validated; Dockerfiles + the Gitea bootstrap were the explicit TODO.
OHM live: v0.31.0 (deploys.id=46). Untouched this session.
Other sessions in flight at claim time: 0045 (--INPROGRESS, UX pass on the admin Users page — rfc-app frontend, disjoint from e2e/), plus stale --INPROGRESS placeholders 0026/0027.

Turn-by-turn arc

Arc 1 — Orient + the v1-gate question

Read SESSION-PROTOCOL.md + the ROADMAP Track-Δ section (#37–#42). Claimed session 0047 via claim-session-id.sh (placeholder pushed at e76b1d3).

The launch prompt insisted on confirming flotilla v1 timing before any VM-side work. Asked the operator (Tier-1-only vs PPE-in-scope). The operator turned it back to me: "can we do #2 this session or stick to #1?"

My call: stick to Tier-1 (#1) this session. Surfaced a factual correction — flotilla v1 has already shipped (v1.4.0), so the "gated on v1" condition on Track Δ is already satisfied; #37 (PPE) isn't blocked by v1 timing, it's just a large, separate live-infra effort (new GCP project + micro VM + colocated ppe.git Gitea + DNS + a real flotilla multi-deployment code change). #38 Tier-1 is prerequisite-free, operator- gesture-free, and de-risks PPE (Tier 2 reuses the same suite). House discipline is one feature per session.

Arc 2 — Map the backend (Explore subagent)

Dispatched an Explore agent to produce a concrete reference: config.py required vars, boot/migration path (app.main:app, db.run_migrations), OTC endpoints (POST /auth/otc/request → 202; POST /auth/otc/verify → session cookie rfc_session), the Turnstile knob (TURNSTILE_REQUIRED), the gitea.py client surface, the seed scripts, the frontend build (VITE_APP_NAME required), /api/health, and the deploy/nginx proxy shape. Verified the load-bearing details directly afterward.

Key catch: the rfc_session cookie is Secure by default → browsers drop it on http://localhost → sign-in wouldn't stick in Tier 1. Confirmed the opt-out is SESSION_COOKIE_SECURE=false (main.py:152).

Arc 3 — Build the Tier-1 stack (in an isolated worktree)

Per the parallel-session rule (0045 in flight, same repo), worked in a dedicated git worktree on session-0047/e2e-tier1 (off the scaffold tip), never the shared checkout. Wrote:

e2e/docker/backend.Dockerfile (+ backend-entrypoint.sh) — repo-mirror layout so parents[2]/VERSION and the seed scripts' parent.parent resolve as in prod; entrypoint reads the minted bot token into GITEA_BOT_TOKEN. Same image runs the gitea-seed one-shot.
e2e/docker/frontend.Dockerfile + nginx.conf — Vite build → nginx serving the SPA + proxying /api,/auth to backend:8000.
scripts/bootstrap_e2e_gitea.py — create the org, then reuse seed_meta_repo.py to create + seed META_REPO (ohm-content).
Rewrote docker-compose.yml: gitea (real gitea/gitea:1.22), gitea-init (mint bot user+token via CLI → shared volume), gitea-seed (org + seed), backend, frontend, mailpit, chained by depends_on.
.env.example: added SESSION_COOKIE_SECURE=false, GITEA_ADMIN_PASSWORD, annotated the runtime-overridden GITEA_BOT_TOKEN; root .dockerignore.

Arc 4 — Bring it up; three real obstacles, fixed

Iterated against a live docker compose up:

Gitea refuses to run as root → every gitea admin user create FATAL'd. Fix: container is root only to chown the fresh token volume; the gitea commands drop to the git user via su-exec.
gitea-init not idempotent on an up over a persisted volume (user + token-name e2e already existed → exit 1). Fix: short-circuit if the token file exists; tolerate an existing user; unique token name.
Per-IP OTC rate limiter (5 sends / 5 min, ratelimit.py) — the suite drives all auth from one IP, so a normal run + CI retries trip it. This is a real determinism issue, not just test churn. Fix: a default-secure, env-gated RFC_APP_RATELIMIT_DISABLED bypass in allow() (mirrors RFC_APP_INSECURE_WEBHOOKS; durable per-email lockouts unaffected) + tests/test_ratelimit_bypass.py (incl. the default-still-enforces case, run green via a throwaway venv).

After fixes: clean down -v && up --build → all services healthy, /api/health → {"version":"0.31.0","status":"ok"}, gitea seeded.

Arc 5 — Green the suite

Finalized the test.fixme specs against the live DOM:

smoke.spec.js: health shape + landing <h1>/title.
otc-login.spec.js: API round-trip (request→Mailpit→verify→/me, asserts pending/needs_profile) and a browser sign-in.

Two real DOM obstacles in the browser test: the v0.13.0 cookie-consent banner intercepts clicks (added a reusable helpers/app.js dismissCookieConsent), and a fresh OTC user redirects to the pending "request is in review" page rather than staying on the capture-profile copy (asserted the "Sign out" link + the review heading, dodging a strict-mode double-match). 4/4 green, twice, ~1.7s.

Arc 6 — Ship + clean up

Committed 15af891, verified fast-forward safety against the unchanged remote tip (a3f48ba), pushed to PR #4. Tore down Docker (down -v), removed the worktree + throwaway venv. Shared rfc-app checkout never touched.

Cut state (end of session)


rfc-app PR #4	`session-0046/e2e-tier1-scaffold` @ `15af891` (fast-forward over scaffold `a3f48ba`). Open, not merged. Tier-1 suite green.
rfc-app `main`	unchanged @ `0c972c8` (v0.31.0)
OHM live	unchanged — deploys.id=46, v0.31.0, healthy
ohm-rfc pin	unchanged (v0.31.0)
flotilla	unchanged @ `5f1689f` (v1.4.0)

Track Δ ledger	Status
#37 PPE pre-prod env	⏸ deferred — own session (live infra; v1 gate already satisfied)
#38 e2e suite + gate	🟡 Tier 1 shipped to PR #4 (green); Tier 2 gated on #37
#39 blue/green slots	⏸ depends on #37
#40 compat gate	⏸ Tier-1 harness is the next #38-adjacent piece
#41 canary routing	⏸ depends on #39/#40
#42 auto-rollback	⏸ depends on #41

§19.2 / process candidates surfaced

The e2e suite needs a stable test-posture profile. Three env knobs (SESSION_COOKIE_SECURE=false, RFC_APP_RATELIMIT_DISABLED=1, RFC_APP_INSECURE_WEBHOOKS=1) + TURNSTILE_REQUIRED=0 are all "deterministic-test" toggles. PPE (#37) will want the same posture except SESSION_COOKIE_SECURE (https). Worth codifying as a named profile rather than scattered env when Tier 2 lands.
The (B) Gitea stub vs (A) real-Gitea split is now half-built. (A) is the default everywhere today; (B) (the fast-inner-loop optimization) remains a TODO behind gitea.py. Decide whether (B) earns its keep once the suite has enough non-Gitea scenarios to make inner-loop speed bite.
flotilla multi-deployment audit (#37 dependency). The §8 deploy phases assume one host/project; driving a second deployment (PPE) on its own VM/gcloud-config is the real flotilla code change behind #37 — audit before standing up PPE.

What lands on the operator's plate

Merge rfc-app PR #4 — the Tier-1 suite is green and one fast-forward commit (15af891) over the scaffold. Left unmerged because it's your PR from 0046 and the 0046 heads-up flagged a wedged Gitea merge-queue; merging is a clean operator gesture (additive e2e/ + a default-secure ratelimit.py gate; no version, no deploy). When it merges, the ohm-rfc ROADMAP #38 row can be updated to "Tier 1 shipped."
Decide #37 PPE timing — it's unblocked (v1 shipped); it's a full live-infra session (new GCP project + micro VM + ppe.git Gitea + DNS + flotilla multi-deployment support). Confirm before any VM stand-up.

Prompt the operator can paste into the next Claude Code session

OHM driver session. Read ~/git/ohm-infra/SESSION-PROTOCOL.md and
~/projects/wiggleverse/ohm-rfc/ROADMAP.md (Phase F / Track Δ, #37–#42).
Claim your session id via claim-session-id.sh --start <now>.

State: flotilla v1 has shipped (v1.4.0) — the "gated on v1" condition on
Track Δ is ALREADY SATISFIED. Session 0047 shipped #38 **Tier 1**: a
working, green local-Docker e2e suite on rfc-app PR #4
(branch session-0046/e2e-tier1-scaffold, tip 15af891). `cd e2e && cp
.env.example .env && docker compose up -d --build && npx playwright test`
→ 4/4 green (health, landing, OTC API round-trip, OTC browser sign-in)
against a real disposable Gitea + Mailpit + fresh SQLite. A default-secure
RFC_APP_RATELIMIT_DISABLED gate was added to backend/app/ratelimit.py
(test-posture only; unit-tested). PR #4 is OPEN, not merged.

Pick one:
  (a) Merge rfc-app PR #4 (operator gesture; additive, no deploy), then
      update the ohm-rfc ROADMAP #38 row to "Tier 1 shipped."
  (b) #38 follow-ons that need no VM: the (B) in-memory gitea.py stub
      (fast inner loop), broader scenarios (propose-RFC / contribute-PR /
      invites / notifications — reuse helpers/mailpit.js + helpers/app.js),
      or the #40 Tier-1 compat-gate harness (N-1 image → migrate → N).
  (c) #37 PPE — the big one. Unblocked but live-infra: a SEPARATE GCP
      project + micro VM + colocated ppe.git Gitea + DNS + a real flotilla
      multi-deployment code change (audit the §8 deploy phases first).
      CONFIRM with the operator before standing up anything on a VM.

Heads-up (still relevant): a wedged Gitea merge-queue was seen in 0046
(persistent 405); verify git state against origin (git ls-remote) before
assuming a push/merge landed; put --max-time on curls; do hands-on git in
an isolated worktree (other sessions share the checkout).

11 KiB Raw Blame History Unescape Escape