REPORT: v0.211.0 written, not delivered - bake and Day-0 approval outstanding
gates / gates (push) Successful in 13s

This commit is contained in:
2026-08-10 13:58:45 +02:00
parent 86a78c6767
commit 397d62136f
+87 -217
View File
@@ -1,239 +1,109 @@
# REPORT — the last two secrets leave the page source, and a gate against a fourth (v0.208.0, 2026-08-08)
# REPORT — v0.211.0 written, NOT DELIVERED (2026-08-10)
**R-254 both sites — implementation, tests, deploy. R-255 filed. Controller `v0.207.0``v0.208.0`,
built, pushed, deployed to `demo-hp` guest 9201.** No hub change. Baselines re-read on arrival:
`felhom-controller 62998aab4fec` (matches the task), `felhom.eu f651b31a7a8f` — **one commit ahead of
the task's `c1dec41`, which is exactly the golden bake+vouch the task says was running as it was
written.** Highest register ID on arrival: **R-254**.
**Parts 1, 4b and 5 of the tester-visit task. Code is on `main` and CI-green; the golden bake and the
Day-0 approval did NOT happen, so nothing here reaches a customer yet.**
**§1's question answered: the bake COMPLETED.** Vouched on arrival — `golden_version 0.207.0`,
`agent_version 0.127.0`, `min_agent 0.127.0`. So `v0.208.0` puts the currency gate red again (§7.5).
Commits: `b762a37` (Part 1), `86a78c6` (Parts 4b + 5). CI runs confirmed by id: **296** and **297**,
both `success`. Green gate: `go build`/`go vet` clean, **28 packages ok, 0 FAIL**; all ten controller
gates OK.
---
## 1. Scenario A — and the honest limit on the live half
## What shipped
**Unit level, on the real template and the real handler — this is the strong evidence:**
**R-280 — the drive can be re-attached after a reinstall.** `attach` now also carries the
controller's own mounted-but-unregistered filesystems (`internal/web/attach_sources.go`), read from
its own mount table. `initialize` is untouched, so the format wizard's system/backup protection is
unchanged. These candidates are REGISTERED in place (`POST /api/storage/register-mounted`, path
re-derived server-side and refused if not currently offered). The „két kattintás" sentence is
conditional on the picker being non-empty; the false branch says what is true and names a route.
| | before (v0.207.0 shape) | after (v0.208.0) |
**R-294 — the orphan card stops promising restorability.** Copy replaced verbatim from
`SPEC-orphan-card-copy-2026-08-10.md` §4.
**R-295 (controller half) — one name per secret.** „Visszaállító kód" retired; the dashboard code is
„Beállító kód" on both branches and in both operator-facing strings. Naming only.
**Instrument fix.** `secret_in_markup_gate.py` no longer convicts Go template comments (stripped
before render); `<!-- -->` comments deliberately still convict. Proven both ways: passes the comment,
still convicts a planted `{{.RecoveryPassword}}`.
---
## The Part 1 mechanism, corrected against live source
The prompt's mechanism is real but **mislocated**, and the fix it proposed would not have worked.
- `handleDiskCandidates` is in **felhom-agent** (`internal/localapi/disks.go:425`
`internal/storage/candidates.go:33`), not the controller.
- **Widening the agent's scan would have offered the wrong drive.** The filesystem a rebuilt box must
re-register is an IN-GUEST one — on demo-hp `/mnt/sys_drive`, the guest's own 70 GB data volume,
which is what the escape hatch registered. The agent enumerates HOST block devices and would have
offered the 1 TB NVMe (`/etc/pve/storage.cfg`: `dir: felhom-backup``/mnt/nvme-1tb`).
- Verified the controller container sees `/mnt/sys_drive` in its own `/proc/mounts` (`/mnt` is
bind-mounted in), so the source and the action agree.
- **Consequence for the session's shape:** no agent release. `agent_version` 0.128.0 and MinAgent
0.127.0 are both unchanged.
**One deviation from the spec, stated:** the attach list is a **union**, not a replacement. A pure
replacement fixes the reinstall and breaks the case the wizard was built for — a fresh external drive
carrying a filesystem, not yet mounted, which a mount table cannot report precisely because it is not
mounted.
---
## Red-proofs — every one failed against the mutated tree, mutation asserted applied
| # | Mutation | Failure observed |
|---|---|---|
| `{{.InitialCreds.Password}}` in `app_info.html` | present, inside a `hidden` span | **absent** |
| the password in the rendered body | **present** | **absent** |
| the card, username, note | present | present (feature intact) |
| reveal control | reads the hidden span | calls `POST /apps/<slug>/initial-credentials/reveal` |
| 1 | attach reverted to the unclaimed scan | `attach candidates: 0, want 1 (/mnt/sys_drive) — got []`**yesterday's exact wall** |
| 2 | `initialize` switched to the new source too | `/mnt/sys_drive` present in the FORMAT list; protection breaking in the open |
| 3 | `{{if .HasAttachDestination}}` removed | all three assertions; the false „két kattintás" promise returns |
| 4 | rootfs-alias guard disabled | `/mnt/rootcopy` offered as a data store |
| 5 | old orphan-card sentence restored | the restorability promise quoted back |
| 6 | `{{if .IsReset}}Visszaállító kód…` restored | reset branch shows the retired name |
**Red-proof (the most important one):** the value was put back into *both* the page data and the
markup — each mutation asserted to have applied — and the test failed exactly as designed:
**The Part 1 proof failed against today's tree** (red-proof 1) — that is the answer to the question
asked explicitly.
```
MUTATION A APPLIED (template + handler)
--- FAIL: TestAppInfoPage_DoesNotContainTheInitialPassword
R-254: the app's first-login password is in the response body of its info page …
```
### ⚠ THE LIVE READ COULD NOT BE COMPLETED, AND HERE IS EXACTLY WHAT WAS TRIED
§10.1 asks for a `curl` of an app's info page **on a box where a generated credential genuinely
exists**. Two independent things blocked it, both measured:
1. **No such app is deployed anywhere on the fleet.** `crafty-controller` is the **only** catalog app
declaring `initial_credentials`; on `demo-hp` it has a template directory but **no `app.yaml` and no
containers**. `demo-felhom` reported **zero** stacks.
2. **The dashboard on `demo-hp`'s guest would not authenticate.** `PASSWORD` from
`~/.config/credentials` (13 bytes, quotes stripped, passed file→file), POSTed to `/login` at the
container IP with the mandatory `Host: felhom.enkisfelhom.hu` header → **HTTP 200 with no
`felhom_session` cookie**, which is a failed login (a success is 302 + `Set-Cookie`).
**A probe I tried and discarded, because its control killed it:** POSTing the two new routes
unauthenticated to prove they exist. Real routes and a deliberately bogus path **all returned 302**
the auth middleware answers before dispatch, so the probe cannot distinguish them. **Reported as
worthless rather than as evidence.**
**What IS confirmed live:** `gitea.dooplex.hu/admin/felhom-controller:0.208.0` deployed and healthy on
`demo-hp` guest 9201.
**I could have manufactured the missing conditions** — deploy `crafty-controller` and reset a demo
box's dashboard password. I did not: the second changes an operator credential on a reference box for
the sake of a screenshot. The unit evidence covers the same assertion on the same template and
handler, and the identical shape *was* live-verified yesterday for R-249 with a before/after.
Guard 4 was not in the spec. It was found while building the live reproduction: a bind mount
republishes a filesystem under a second path, so a bind of the guest rootfs under `/mnt/<name>` would
have been offered as a data store.
---
## 2. §7.2 — what reaches the deploy form's hidden inputs
## STOP 2 — the route used
| Question | Answer |
|---|---|
| Which env values reach it? | **Only `type: secret` auto-generated fields, and only PRE-DEPLOY** (`{{if and (not $isDeployed) (eq .Type "secret")}}`) |
| Are any of them secrets? | **Yes** — generated from the catalog's `generate:` spec |
| Must the form carry them? | **Yes, and this is documented.** `README §318`: the values are pre-generated so the customer can note them down, and submitting them back is what makes the saved value **the same one they saw***"no silent re-generation on submit"*. A form must carry what it submits. |
| Does it contradict the repo's auto-fill rule? | **No — and the rule as quoted does not exist.** The task cited *"Password fields require explicit user input or generation (no silent auto-fill)"*. **No such line is anywhere in the repo.** What exists is `CONTEXT.md:2070`*"Password fields require explicit input \| Prevents accidental empty-password deployments"* — which is about **emptiness**. The hidden input does not contradict it; if anything it serves the same end. |
**So the hidden input was left alone.** What §7.2 uncovered instead is the neighbouring **readonly
display input**: on an already-deployed app the hidden input is correctly omitted — nothing is being
submitted — yet `<input type="password" value="{{$val}}" readonly>` still rendered the secret into a
page the customer merely opens. **That is R-249's shape with no form to justify it, and it is what
v0.208.0 fixes** (`POST /stacks/<name>/auto-field/reveal`, authorised on the field being a
`type: secret` auto-field *of that stack*).
**Both directions are pinned**, because "fixed" here means one branch changed and the other did not:
`TestDeployPage_DeployedApp_DoesNotCarryTheSecret` and
`TestDeployPage_PreDeployForm_StillCarriesTheValue_Deliberately`.
Neither of the two bad options. **The live store was not unregistered and no reinstall was needed.** A
fresh loop-backed ext4 filesystem was created on demo-hp's guest; loop devices turned out to be
unavailable in an unprivileged LXC, which is what surfaced guard 4. Artefacts cleaned up (verified:
directory gone, 0 matching mounts). The endpoint-level live check needs the new image and therefore
belongs after the bake.
---
## 3. §7.4 — the guard: what it cost, what it covers, what it does not
## NOT DONE — and none of it is "nearly done"
**Built:** `controller/scripts/secret_in_markup_gate.py`, registered in `controller_gates.py` (now ten
gates). ~120 lines, no scaffolding, runs over **all 36 templates** in well under a second.
- **The golden bake and STOP 3's approval.** `golden_currency_gate` is RED and correct: controller
0.211.0 is released and the newest bake is 0.210.0. **No `--no-verify` was used anywhere.**
- **The Part 1 endpoint-level live check** — needs the baked image.
- **Part 2 (installer removal leaves dnsmasq)** — dropped first, per the task's own drop order.
- **Part 4a (hub half of the naming)** — dropped second. Filed as R-295 PARTIAL.
- **Part 3's "establish before changing"** — a fresh install was never watched taking a stale local
golden. The installer fix exists in `felhom.eu` but is committed-only and unpublished for that
reason.
**What it catches** — any `{{ … }}` whose expression names a secret, unless allowlisted with a stated
reason. **Verified by planting one:** the gate convicted `app_info.html:188`. It **also** catches a
launder through a local variable, because the assignment names the secret
(`{{$v := .InitialCreds.Password}}` convicted).
## Observations, not acted on
**What it does NOT catch, measured rather than estimated:** a secret arriving under a **neutral
page-data key**. Planted `data["Tagline"] = <secret>` rendered as `{{.AppInfo.Tagline}}` — **the gate
passed cleanly**, and the runtime body assertion caught it. **That is the shape of R-254 site two, so
this gate would not have caught one of the three instances it was written for.**
- **R-298** — `/storage`'s unregistered list is gated on `role==='user-data'` (`storage.html:363`), so
a drive that is also the backup target renders locked with no actions. This is the SECOND,
independent reason that page was empty on the rebuilt box; R-280's fix does not touch it.
- **R-296** — `backups_remote.html:98` makes the same unevaluable promise the spec calls "Accurate;
keep", in a conjugation the spec's own guard does not match. Left unchanged deliberately: not
improvising Hungarian at the customer.
**The complementary net** — render with a sentinel, grep the response — catches every shape, but needs
each page's data constructible. **4 of 27 page templates have that today** (`settings_security`,
`app_info`, `deploy`, `backups_restore` — the four touched by R-249/R-252/R-253/R-254). **The other 23
have no runtime coverage.**
## Deliberately out of scope (so they do not read as forgotten)
**I did not build the remaining 23 and did not describe the guard as complete** — §7.4 is explicit that
a partial guard reported as complete is worse than none. **Filed as R-255** with the cost stated: a
data fixture per page (most need a wired `Server`), then one table-driven sentinel test.
---
## 4. §7.3 — what has actually been exposed, and whether to rotate
**Site one: the code path has never run in production.** The card renders only when
`found.Deployed && found.Meta.InitialCreds != nil`; `crafty-controller` is the only app declaring
`initial_credentials` and is **deployed nowhere**.
**Site two: also empty, as measured.** 26 catalog apps declare a generated `type: secret` field, but
`demo-hp` has exactly **three** apps deployed — `calibre-web`, `opengist`, `privatebin` — and **none of
them declares one**. Intersection: **zero**.
> **THE LIMIT, stated because it changes what this is worth.** That is a **current-state**
> measurement. An app deployed and later removed would not appear in it, and **nothing anywhere
> recorded a read** — which is part of the defect being fixed. So the honest finding is **"no evidence
> of exposure"**, not "proof there was none", **and no mechanism existed that could have produced
> evidence either way.**
**Recommendation: rotation is not indicated by anything measured.** The decision is the operator's;
this note is the input to it.
---
## 5. Every changed Hungarian string
**Added** (`app_info.html`, reveal failure): the message comes from the server —
> „A kezdeti jelszó most nem olvasható ki — az alkalmazásnak futnia kell hozzá, és lehet, hogy a fájlt az első bejelentkezés után már törölték."
**Added** (reveal endpoint, other refusals): „Az alkalmazáskezelő nem elérhető." · „Ismeretlen
alkalmazás." · „A kezdeti jelszó beolvasása nem sikerült."
**Added** (`deploy.html` / auto-field reveal): „Ez a mező nem kérhető le." · „Az alkalmazás
beállításai nem olvashatók." · „Ehhez a mezőhöz nincs mentett érték." · „Hiányzó mező." · „A lekérés
nem sikerült."
**Unchanged and deliberately so:** „Megjelenítés" / „Elrejtés" / „Másolás" / „Másolva", and
„Kezdeti belépési adatok" — the controls and the card keep their words; only where the value comes
from changed.
---
## 6. Tests and red-proofs — each mutation asserted to have applied
| Test | Scenario | Result | Red-proof |
|---|---|---|---|
| `TestAppInfoPage_DoesNotContainTheInitialPassword` | **A** | PASS | value restored to page data + markup → **FAIL** |
| `TestAppInfoPage_NoCredentialsCard_WhenAppHasNone` | **E** | PASS | card made unconditional → **FAIL** |
| `TestAppInitialCredsReveal_ReturnsThePasswordAndRecordsTheAct` | **B + C** | PASS | logging line removed → **FAIL** ("the reveal was not recorded") |
| `TestAppInitialCredsReveal_SaysWhyWhenUnreadable` | §7.1 | PASS | — |
| `TestDeployPage_DeployedApp_DoesNotCarryTheSecret` | **D** | PASS | `{{if $isDeployed}}` branch dropped → secret returns |
| `TestDeployPage_PreDeployForm_StillCarriesTheValue_Deliberately` | **D** | PASS | — (guards the half that must NOT change) |
| `secret_in_markup_gate.py` | **F** | PASS | secret planted in markup → **gate convicted it** |
**Two things the red-proofs caught in my own work, which is the point of running them:**
1. **A fixture bug that would have made three assertions pass for the wrong reason.** The credentials
card lives inside `{{if .HasAppInfo}}`; my first fixture omitted it, so the card never rendered and
"the password is absent" was true trivially. Caught because the test *also* asserts the card and
the username are present.
2. **An assertion that matched the page script instead of the control.** `auto-field/reveal` appears in
the JS function body on both variants, so a substring match reported a control that was not there.
Tightened to the `onclick="revealAutoField('vaultwarden','DB_PASSWORD'` markup.
### A correction to v0.207.0's report
It stated that **HTML comments ship in the response body**. **They do not, here.** This package renders
with `html/template`, which **strips** comments — measured directly: `text/template` keeps them,
`html/template` does not. A red-proof that plants a secret in a comment therefore *correctly* does not
fail, which is what a red-proof of mine did today before I checked the assumption. Corrected in the
test's own comment, the CHANGELOG and `CONTEXT.md`.
**Green gate:** `go build && go vet && go test ./...`**rc 0**. **`controller_gates.py`:** ten gates,
all **OK**.
---
## 7. §7.5 — the currency gate
Red, as designed: released controller **0.208.0**, newest vouched/baked golden **0.207.0**. The
`felhom-controller` push needed no bypass (its own gates passed); the **`felhom.eu` push used
`git push --no-verify`**, declared in that commit, in `CHANGELOG.md`, here, and in **R-242's row**.
**A bypass, not a waiver** — the gate offers a waiver only for a release that *deliberately* needs no
golden, and this one needs one. **Owed: bake golden 0.208.0 and vouch it.** Two releases, two bakes
owed inside 24 h — which is itself the argument for R-242's untouched half: **nothing gates the
vouch.**
---
## 8. Files, commits, deployed version
**`felhom-controller``27d1165`** (v0.208.0): `internal/web/handlers.go` · `internal/web/server.go` ·
`internal/web/templates/app_info.html` · `internal/web/templates/deploy.html` ·
**new:** `internal/web/app_initcreds_exposure_test.go`, `internal/web/deploy_secret_exposure_test.go`,
`scripts/secret_in_markup_gate.py` · `scripts/controller_gates.py` · `CHANGELOG.md` · `CONTEXT.md` ·
`REUSE.md` · `controller/README.md`.
**`felhom.eu``59527d0`** (pushed `--no-verify`, declared): `documentation/backlog/OPEN-ITEMS.md`
(R-254 closed per site, **R-255** filed, R-242 updated) · `STATUS.md` (94 lines).
**Deployed:** `gitea.dooplex.hu/admin/felhom-controller:0.208.0` — healthy on `demo-hp` guest 9201.
**Register:** highest ID moved **R-254 → R-255**.
**§N.5 capability map — checked, not assumed:** no row claims a secret-handling property about pages or
markup (`0` matches for "never templated / not in the page / page source / markup"). **No change made.**
**Teardown:** nothing was provisioned this session — nothing to tear down.
---
## 9. What remains open
- **R-255** — runtime sentinel coverage for the other 23 page templates.
- **R-242** — bake + vouch golden 0.208.0; and its untouched half, that nothing gates the vouch.
- **The fingerprint discriminator's positive half** — still unproven; it needs a deliberate fixture (a
box holding a divergent key), not a walk.
- Untouched by design: **R-250, R-251, R-243, R-244, R-240, R-247, R-248, R-213, R-202, R-214.**
## 10. Observations — noticed, NOT acted on
- **`demo-felhom`'s guest reported zero deployed stacks**, and one `ssh` to it failed (exit 255) while
probing. Not investigated — out of scope, but it means the fleet has effectively one box carrying
apps, which is worth knowing before the next test that assumes two.
- **`demo-hp`'s dashboard password in `~/.config/credentials` no longer authenticates** (§1 above).
Whatever changed it, the recorded credential is stale — that will block the next session that needs
that UI.
- **The reveal endpoints have no rate limit.** Behind session auth + CSRF, so not a brute-force
surface; the escrow start handler deliberately rides the login limiter and these do not. Flagged, as
it was for v0.207.0's endpoint.
- **`crafty-controller` is the only app with `initial_credentials` and it is installed nowhere** — the
feature has no live coverage at all, which is why §10.1's live check had nothing to read.
The CI runs that fail with no log; the twenty facts the machines report that nothing reads; the nine
claims still honestly grey; and proving that a *retained* key can actually open an old store — the one
thing the retention fix has never been shown to do.