Files
session-history/ohm/0055/SESSION-0055.0-TRANSCRIPT-2026-06-01T03-00--2026-06-01T03-35.md
T

128 lines
6.9 KiB
Markdown

# Session 0055.0 — Transcript
> Date: 2026-06-01
> Start: 2026-06-01T03-00 (PST implied)
> Goal (as it emerged): not launched as an OHM driver session — opened as a
> conceptual Q&A comparing the OHM/flotilla self-hosted stack against a PaaS
> like Vercel. It became a **flotilla SPEC design session**: the "which Vercel
> wins could we adopt?" thread converged on per-PR preview environments, and
> the operator asked for that to be specified.
>
> Outcome: **Added a fully-fleshed §19.2 candidate — "Per-PR preview
> environments" — to `ohm-rfc-app-flotilla/SPEC.md`, committed on a branch,
> merged `--no-ff` to `main`, and pushed to `origin`.** Two design decisions
> were settled into the entry by the operator: (1) the §10 deploy audit trail
> is **production-only** — neither per-PR previews nor PPE write to it; and (2)
> previews use **SQLite only**, the same engine as production, which is exactly
> what makes them cheap — no dual-engine/Postgres path. Along the way we
> assessed (and rejected, for now) the idea of dual SQLite+Postgres support in
> rfc-app.
>
> **Scope note (transcript-honest):** this is **pure-flotilla doc work with no
> OHM actuation** — no deploy, no §13 record edit, no `ohm-rfc` operation. Per
> SESSION-PROTOCOL the transcript expectation *softens* for exactly this case;
> this transcript exists because the operator asked for one at session end. The
> ID was therefore claimed at publish time (the protocol's late-claim path),
> not at true session start.
>
> **Timestamp caveat:** the `03-00` start is anchored to the first commit of
> the session (`912e828` at 03:03) and the preceding Q&A; the work actually
> opened a few minutes earlier. A concurrent session published `0054`
> (03-24/03-25) while this one was already in flight — expected under advisory
> numbering; the `claim-session-id.sh` push allocated `0055` race-free.
---
## Pre-session state
- `ohm-rfc-app-flotilla` at `main` = `70ba3b0` (flotilla v1.6.0 merged), clean
tree. SPEC.md at 1875 lines; §19.2 held the post-v1 candidate queue.
- rfc-app checkout present at `~/projects/wiggleverse/rfc-app`
(HEAD `a3f48ba`), used read-only this session to assess the DB layer.
- No preview-environment concept anywhere in flotilla; v1 has two long-lived
environments only (production VM + PPE, a single serialized lane).
## Arc
### Arc 1 — Q&A: the OHM/flotilla stack vs. Vercel
Walked the advantages (ownership/no-lock-in, real persistent backend, flat
predictable cost, secret discipline under operator control, mission-fit
sovereignty, arbitrary architecture) against where Vercel wins (zero ops,
push-deploy + preview envs + instant rollback, edge CDN, autoscaling, frontend
DX). Bottom line: the app's *shape* (stateful backend, modest predictable
traffic) is the opposite of Vercel's sweet spot; flagged Cloud Run as the
"Vercel-like ergonomics without leaving the sovereignty model" middle path.
### Arc 2 — Which Vercel wins to adopt
Ranked the adoptable ones: per-PR previews (highest value), one-command
rollback (flotilla half-has it via pinning), edge CDN via the existing
Cloudflare relationship (cheapest win), atomic/zero-downtime deploys. Operator
zeroed in on **per-PR deployments with synthetic data in a DB** — PPE's single
lane can't do parallel/isolated review.
### Arc 3 — Draft the spec
Read SPEC §3/§5/§8/§10/§12/§13/§19.2 to match conventions, then added the
**"Per-PR preview environments"** §19.2 candidate (placed after its cousin
"Private-branch deploys"): problem statement, proposed `preview up/down/
list/show` verb surface, runtime options (Cloud Run recommended over
VM-containers), seeded-DB model, test-secret posture (zero real secret bytes —
*strengthens* §3 invariant 1), the invariant-5 trigger tension, and the
cross-section touchpoints.
### Arc 4 — Decision: audit trail is production-only
Explained the §10 deploy-history audit trail's purpose (append-only provenance:
who/what/when/which-config for every prod deploy, secret-redacted; combined with
`git log .rfc-app-version` + `gcloud secrets versions list` for full
reconstruction) — and disambiguated it from the unrelated `flotilla audit`
Patchwatch radar (§12.4). Operator decided: **previews and PPE never write the
§10 trail — production only**, so `deploy log` stays high-signal. Recorded as
**settled** in the entry.
### Arc 5 — Dual SQLite+Postgres? Investigated, rejected for now
Operator asked how hard it'd be to support both engines. Inspected rfc-app's DB
layer: raw stdlib `sqlite3` (no ORM) in `backend/app/db.py` with a module-level
singleton connection; **23 raw `.sql` migrations** (~1,300 lines) and **28 of
43 app modules** issuing raw SQL. Quantified the SQLite-isms at the query sites
(`datetime()`/`strftime` in 21 files, `lastrowid` in 12, `ON CONFLICT` in 8,
`INSERT OR REPLACE/IGNORE` in 3, case-sensitive-`LIKE` hazard in 3), plus the
structural mismatches (strict vs. dynamic typing, placeholder style, the
singleton-conn-vs-pool concurrency model). Conclusion: dual-support is
**medium-high effort + a permanent maintenance tax**, and rfc-app §4.2 already
frames Postgres as a *future one-way migration*, not a dual-support goal.
### Arc 6 — Decision: stay SQLite; previews are cheap because of it
Operator: remove the Postgres framing entirely. Rewrote the candidate's DB
paragraph — previews bake a seeded `.sqlite` (synthetic data) into the image
and die with the container; prod and previews share **one engine**, which is
precisely what makes per-PR previews cheap. No dual-engine/Postgres path to
build or maintain; Postgres remains a hypothetical future one-way migration,
never a prerequisite for previews.
### Arc 7 — Ship
Confirmed the entry internally consistent. Committed `912e828` on branch
`spec/preview-environments-candidate`, merged `--no-ff` (`4c6781f`) to `main`,
deleted the branch, pushed `5cf5ecd..4c6781f` to `origin/main`.
## Artifacts
- `ohm-rfc-app-flotilla/SPEC.md` §19.2 — new "Per-PR preview environments"
candidate (+85 lines). Merge commit `4c6781f` on `main`, pushed to origin.
## What this session did NOT do
- No OHM deploy, no §13 deployment-record edit, no `ohm-rfc` operation, no
rfc-app release. rfc-app was read-only (DB-layer assessment).
- Did not start the preview feature's build — it remains a §19.2 candidate
that "earns its session" when concurrent PR review becomes PPE's bottleneck.
- Did not act on the two open forks (Cloud Run vs VM-container runtime; the
invariant-5 trigger question) — left for the build session by design.
## Handoff
The preview-environments candidate is specced and on `main`. When it earns a
build session, the two open decisions are runtime (Cloud Run strongly favored —
also the on-ramp for a broader Cloud Run migration, gated on containerizing
rfc-app) and the invariant-5 trigger (operator-typed `preview up --pr=N` first
cut, deferring auto-on-PR to the hosted-flotilla shift). The DB and
audit-trail questions are already settled (SQLite-only; production-only trail).