Files
session-history/ohm/0066/SESSION-0066.0-TRANSCRIPT-2026-06-03T17-00--2026-06-03T19-20.md

131 lines
8.0 KiB
Markdown

# Session 0066.0 — Transcript
> App: ohm
> Start: 2026-06-03T17-00 (PST) · End: 2026-06-03T19-20 (PST)
> Driver: Claude (Opus) + operator (ben.stull@wiggleverse.org)
> Status: **FINALIZED**
## Launch prompt
> "Let's migrate all of the app repos to match the current pattern of one repo per app."
(Session ran from the `~/git` parent, not a single repo — the work spanned many
Wiggleverse repos. Filed under **ohm** because OHM was the heaviest leg and the
only live production deploy this session.)
## Pre-state
- The recorded repo-topology decision (engineering README §5.3 + flotilla-core
onboarding memory): **repo-per-deployment default; kill the per-app
`<app>-flotilla` operator-CLI shim** → deploy config becomes a `deployment.toml`
read by shared `flotilla-core`. `wiggle-snip` already conformed.
- Three apps still carried a `-flotilla` shim: **ecomm** (empty placeholder),
**bdd-release-planner** (real bespoke `planner_flotilla`), **OHM**
(`ohm-rfc-app-flotilla`, a thin shim over flotilla-core + the Patchwatch radar
infra). OHM's 0.31.5 deploy was "pending".
## Plan (as executed)
Sequenced easiest→riskiest, each leg its own PR; archive (not delete) retired
shims; framework/content/meta/prototype repos stay split.
1. **ecomm** — drop the empty shim from app.json + README note; archive shim.
2. **bdd-release-planner** — author flotilla-core-schema `deployment.toml` + VERSION;
archive shim.
3. **OHM** — fold `deployment.toml` + infra (runner/scripts/docs) into `ben.stull/ohm-rfc`;
repoint Patchwatch runner at flotilla-core; archive shim.
4. **handbook** — rewrite engineering §5.3/§8.2 to the new pattern; update memory.
## Turn-by-turn arc
1. **Explore + plan.** Two read-only Explore agents mapped flotilla-core's operator
interface and inventoried the OHM shim. Key finds: OHM's shim was *already* a thin
wrapper (`context.py` + `build_cli`), and the **Patchwatch `audit` logic already
lives in flotilla-core**. Entered plan mode; resolved scope via AskUserQuestion
(shims-only; include OHM; archive). Plan approved.
2. **Leg 1 — ecomm** (PR #1): removed the placeholder shim entry, README deploy note,
archived `wiggleverse-ecomm-flotilla`. Clean.
3. **Leg 2 — bdd** (PR #2): discovered bdd uses a *different* deployment.toml schema
(bespoke `planner_flotilla`). Translated to flotilla-core schema. **Wrong-turn /
friction:** flotilla-core's `deployment validate` flagged `db-under-install`
(planner uses `PLANNER_DB`, not `DATABASE_PATH` — mirrored it) and a
`name-derivation` FAIL that **even the reference `wiggle-snip` toml trips** → a
flotilla-core validator over-strictness (`<name>-ppe.<domain>` is the real
convention, not `ppe.<name>.<domain>`). Logged as out-of-scope; spawned a task.
Asked operator about bdd's live cutover → "full cutover + archive now."
4. **Leg 3a — flotilla-core v0.8.0** (PR #2): hit a hard blocker — `deployment.toml`'s
single `[app].repo` fanned out to all three registry repo roles, so OHM's
**framework-reuse** topology (code = `ben.stull/rfc-app`, pin+corpus = `ben/ohm-rfc`)
was inexpressible. Operator chose "enhance flotilla-core now." Added optional
`[app].app_source_repo` + `[app].pin_source_repo` (default to repo); 448 tests green.
5. **Leg 3 — OHM** (ohm-rfc PR #11): derived `deployment.toml` from the **live operator
registry** (`~/.ohm-rfc-app-flotilla/…db`) — caught drift vs the Slice-1
`register-ohm.sh` (live had `ohm-bot`, `ohm-content`, rfc.wiggleverse.org, the
session-history overlay vars). Moved runner/scripts/docs into `infra/`; repointed
the Patchwatch image to install flotilla-core + import the toml. Dropped the shim's
`pin_floor` (operator-approved). **Wrong-turn:** opened the PR against `ben/ohm-rfc`
→ 307 redirect; canonical owner is `ben.stull`. Force-push to fix a URL was
policy-denied → recovered via a non-force follow-up commit.
6. **Leg 4 — handbook** (engineering PR #6): rewrote §5.3 (one-repo-per-app + when to
split) and §8.2 ("operator engine", deployment.toml). Done in a throwaway worktree
off `origin/main` so the operator's in-flight `slice3` WIP was untouched.
7. **"do 1-3 if you can" → merges.** Self-merge was first policy-denied ("ambiguous");
operator then explicitly authorized. Merged all PRs in order (flotilla-core first),
tagged `flotilla-core v0.8.0` + bdd `v0.1.0`, upgraded global flotilla-core to 0.8.0.
8. **OHM deploy — failed, then fixed.** Registered OHM into `~/.flotilla`; dry-run
GREEN. **First real deploy FAILED at phase 3:** `pathspec 'v0.31.5' did not match`.
Prod-SSH diagnosis was twice policy-gated; once explicitly permitted, the SSH
revealed the **root cause**: the VM clones rfc-app from
**`git.benstull.org/benstull/rfc-app`** (a *different host* than canonical
`git.wiggleverse.org`), and that mirror lacked `v0.31.5`. Verified both hosts share
identical history at v0.31.4, pushed the `v0.31.5` tag to git.benstull.org,
re-deployed → **9/9 green, OHM live on v0.31.5** (`{"version":"0.31.5","status":"ok"}`).
9. **bdd cutover — found unrunnable, fixed the app half.** flotilla-core verify needs
`body.version == tag`, but planner `/api/health` returned no `version` and its
`health_url` (`rfc.wiggleverse.org`) is **OHM's** domain (`bdd.ecomm…` 301-redirects
there). Shipped `/api/health` version reporting (PR #3, VERSION→0.1.1, tag v0.1.1).
Operator decided planner's future host = `rfc.wiggleverse.org/bdd/{TOKEN}`; the live
deploy (needs nginx routing + prod SSH) was deferred.
## Cut state (where things stand at session end)
- **Merged + tagged:** ecomm #1, bdd #2/#3, flotilla-core #2 (v0.8.0), ohm-rfc #11,
engineering #6. Tags: flotilla-core `v0.8.0`, bdd `v0.1.0`/`v0.1.1`, rfc-app
`v0.31.5` mirrored to git.benstull.org.
- **3 shims archived** read-only; none deleted.
- **OHM deployed live on v0.31.5** via the new flotilla-core 0.8.0 path (verified).
- Global `flotilla-core` = 0.8.0; merged feature branches pruned (local + origin).
- Memory updated in the **flotilla-core per-repo bucket** (the central per-project
files had just been reorganized into per-repo auto-memory buckets mid-session).
## Deferred / follow-ups (in memory)
- **bdd live cutover** — host decided (`rfc.wiggleverse.org/bdd/{TOKEN}`, deprovision
`bdd.ecomm…` after); needs nginx routing on the rfc edge + `edge.domain`/`health_url`
+ `flotilla-core deploy planner`. The `planner-gitea-read-token` secret exists.
- **rfc-app mirror divergence** — the OHM VM pulls from git.benstull.org, not canonical
git.wiggleverse.org; mirror tags forward each release or repoint the VM origin.
- **Patchwatch runner rebuild** — non-urgent (radar runs its existing image); needs
Docker + a flotilla-core-readable Gitea token.
- **flotilla-core `name-derivation` validator** over-strictness — spawned task.
## Operator plate (what needed a human / a permission this session)
- Created `wiggleverse-ecomm-bdd-app/planner-gitea-read-token` (copied bytes from an
existing read-token secret, piped — never displayed).
- Authorized self-merge of the PRs; authorized prod SSH for the OHM diagnosis/deploy.
- Decided: include OHM, enhance flotilla-core now, planner's future host.
## Next-session prompt
> Continue the Wiggleverse one-repo-per-app migration — the only piece left is the
> bdd-release-planner LIVE cutover (every other app migrated; OHM deployed on v0.31.5).
> Read the flotilla-core memory bucket (project_flotilla_core_onboarding.md) first.
> Repo-side is done: deployment.toml merged, planner /api/health reports its version
> (v0.1.1). Decided host: planner serves at rfc.wiggleverse.org/bdd/{PLANNER_TOKEN};
> bdd.ecomm.wiggleverse.org is deprovisioned afterward. First move: design the nginx
> routing on the rfc.wiggleverse.org edge so /bdd/ proxies to the planner (alongside
> OHM at root) with a health path flotilla-core can verify, then set
> edge.domain/health_url in wiggleverse-ecomm-bdd-release-planner-app/deployment.toml
> and run `flotilla-core deploy planner`. Deploys SSH into prod VMs (permission-gated).