cc-ci/machine-docs/JOURNAL-redfix.md

# JOURNAL — phase `redfix`

## 2026-06-17T23:20Z — Bootstrap

Read phase plan + plan.md §6.1/§7/§9 + canon DECISIONS exceptions (lines ~1494–1552). Six
canon-sweep failures to investigate. Confirmed cc-ci access, no run in flight, sweep timer next
fires 2026-06-21 (3-day window), disk 38G free.

Isolation mechanism understood: `runner/nightly_sweep.run_on_tag` = `abra.recipe_checkout(r, tag)` +
`run_recipe_ci.py RECIPE=<r> CCCI_SKIP_FETCH=1` cold/full. I reproduce each failure by running ONE
recipe at a time with no concurrent load.

Starting canonical state notable: **mumble canonical IS present** (`1.0.0+v1.6.870-0`, written
20260617T180501Z — during today's nixenv sweep). The canon DECISIONS recorded mumble RED
(`test_handshake_completes_with_channel_presence`). A canonical only gets written on a GREEN cold run
on latest → mumble flipped green in a recent run. Strong early evidence for the operator's "mumble
passed before" → load flake hypothesis. Must confirm with a clean isolation re-run + check whether the
canon-sweep red was under concurrent load.

Next: start M1 investigation. Plan order (cheap/informative first): triage the existing sweep logs on
cc-ci to pin the EXACT assertion/error for each (mumble, mattermost-lts restore, gitea app.ini,
bluesky routing, discourse compose), then run isolation re-runs. discourse's recorded cause is an
UPSTREAM compose defect (`sidekiq.depends_on: discourse` while service is `app`) that FATAs before any
deploy — that's deterministic, not a load timeout, so it may not even need a long isolation run to
confirm; verify the compose at the latest tag directly first.

## 2026-06-17T23:40Z — M1: discourse isolation run — CANON ROOT-CAUSE WAS WRONG

Ran discourse ALONE on cc-ci (`recipe_checkout discourse 0.8.1+3.5.0` + `RECIPE=discourse
CCCI_SKIP_FETCH=1 cc-ci-run runner/run_recipe_ci.py`, log `/tmp/redfix-discourse.log`).

RESULT: **install PASS, upgrade FAIL, backup PASS, restore PASS, custom PASS** — the recipe deploys,
serves (200 /srv/status), backs up and restores cleanly. NOT a deploy timeout, NOT a 51-min wedge,
NOT a deploy FATA. The canon DECISIONS root-cause ("`abra app deploy` FATAs: service sidekiq depends
on undefined service discourse → invalid compose project") is **misattributed**: that string appears
ONLY from the non-fatal prepull `docker compose config --images` (rc=15, harness logs "skipping
(deploy will pull as usual)"). The real `abra app deploy` is a swarm `docker stack deploy`, which
ignores `depends_on` entirely → the stack converges (`UpdateStatus=completed`).

The ONLY failure is the cc-ci upgrade OVERLAY `tests/discourse/test_upgrade.py`:
- `test_head_runs_official_image_not_bitnamilegacy` — app image is `bitnamilegacy/discourse:3.5.0`;
  test demands `discourse/discourse:3.5.3` (official).
- `test_sidekiq_service_dropped_by_head` — services `['app','db','redis','sidekiq']`; test demands
  sidekiq dropped.

These `prevb`-phase overlay tests are PR-FAITHFULNESS assertions for a specific migration PR
(bitnamilegacy → official `discourse/discourse:3.5.3`, drop sidekiq). Verified that migration exists
in **NO upstream release tag and NOT in main** — `git show main:compose.yml` and every tag
(`0.1.0…0.8.1+3.5.0`) all use `bitnamilegacy/discourse:3.5.0` + sidekiq. So the overlay asserts a
state that doesn't exist anywhere upstream → deterministic RED whenever the sweep tests the latest
release tag. The head DID deploy (chaos-version label = head f87c612d+U, converged) — the test
expectation is simply wrong for the released recipe.

Note (M2 design): migrating discourse from the deprecated `bitnamilegacy` image to official
`discourse/discourse` is a MAJOR recipe rewrite (different fs layout, entrypoint, no `/opt/bitnami`
sidekiq run.sh) — not a 1-line image swap. So the overlay test's `discourse/discourse:3.5.3`
expectation may not be a realistic near-term recipe change. The bitnamilegacy deprecation is real
(bitnami sunset legacy images), so a migration is the right long-term direction, but the test as
written hard-codes a migration target absent upstream. Classification + fix approach to settle in M1
table / M2.

Classification: **stale/PR-specific cc-ci OVERLAY test mismatched to the canonical-sweep context**
(NOT a flake, NOT a load timeout, NOT a recipe-deploy defect, NOT warm-machinery). Teardown clean (no
discourse stack left). Evidence: `/tmp/redfix-discourse.log` on cc-ci; junit under
`/var/lib/cc-ci-runs/manual/junit/upgrade__cc-ci__test_upgrade.xml`.

## 2026-06-18T00:05Z — M1: mattermost-lts isolation run — DETERMINISTIC restore failure (recipe defect)

Ran mattermost-lts ALONE (tag 2.1.9+10.11.15, log /tmp/redfix-mattermost-lts.log).
RESULT: **install/upgrade/backup/custom PASS, restore FAIL** — identical to the canon failure:
`tests/mattermost-lts/test_restore.py::test_restore_returns_state` → `relation "ci_marker" does not
exist` after restore. So it is **deterministic in isolation, NOT a loaded-node race** (canon framing
was wrong). The marker logic is sound (postgres table seeded pre-backup, dropped pre-restore, asserted
post-restore — same pattern immich uses and PASSES).

ROOT CAUSE (recipe backup/restore labels). Compared mattermost-lts vs immich (immich passes the
IDENTICAL test):
- immich `database` svc: `backupbot.backup.pre-hook: /pg_backup.sh backup`,
  `backupbot.backup.volumes.postgres.path: backup.sql` (backs up ONLY the dump file), and
  **`backupbot.restore.post-hook: /pg_backup.sh restore`** (replays the dump on restore). → round-trips.
- mattermost-lts `postgres` svc: `pre-hook: pg_dump > /var/lib/postgresql/data/postgres-backup.sql`,
  `backup.path: /var/lib/postgresql/data/` (backs up the WHOLE live/hot PGDATA dir + the dump),
  `post-hook: rm .../postgres-backup.sql`, and **NO `backupbot.restore.post-hook`**. So on restore,
  abra restores the files but NOTHING replays the dump, and a hot-copied live PGDATA over a running
  postgres does not reload → `ci_marker` lost. Restore log confirms `Restoring Snapshot b0495d36 at /`
  with no post-hook reimport.

Classification: **GENUINE RECIPE DEFECT at latest** (postgres backup/restore does not round-trip —
missing restore post-hook + backs up hot PGDATA instead of dump-only). NOT a flake, NOT cc-ci test
weakening (test is correct & unmodified; immich proves the pattern works). Fix (M2) = recipe PR
adopting the immich-style postgres backup/restore (a `/pg_backup.sh`-style dump + restore post-hook).
Teardown clean (no matt stack). Evidence: /tmp/redfix-mattermost-lts.log; junit
restore__cc-ci__test_restore.xml.

Tooling note: my background "waiter" loop `while pgrep -f run_recipe_ci.py` self-matched (its own
cmdline contains the string) → never exited, falsely showed a run active. Use `pgrep -f
"[r]un_recipe_ci.py"` or match the python invocation. Killed the stuck waiters; node confirmed free.

## 2026-06-18T00:18Z — M1: mumble isolation run — GREEN (flake confirmed)

Ran mumble ALONE (tag 1.0.0+v1.6.870-0, log /tmp/redfix-mumble.log). RESULT: **ALL tiers PASS**
(install/upgrade/backup/restore/custom), including `custom/test_protocol_handshake.py::
test_handshake_completes_with_channel_presence` PASSED. No orphan stacks. The canon sweep recorded
this RED (`test_handshake…` failed under concurrent sweep load); it is GREEN here in isolation, and
its canonical was already written green TODAY (1.0.0+v1.6.870-0 @20260617T180501Z) under the lighter
nixenv sweep. → **load/timing FLAKE** on the control-channel handshake, NOT a recipe defect.

The handshake test already retries (`retry_handshake(attempts=12, interval=5.0)` = 60s). So the flake
is the voice server not completing the TLS+ServerSync handshake within ~60s under heavy concurrent
node load (deploy contention). M2 fix = harness stabilization (stronger readiness gate before the
custom tier / longer-or-smarter retry / serialize), based on the load failure mode. Classification:
**FLAKE (load/concurrency)** → harness stabilization.

Reproducibility: 1 green isolation run here + canonical green today + documented red under canon load.
Will do 1–2 more isolation repeats before the M1 claim to firm "reproducibly green in isolation."

## 2026-06-18T00:45Z — M1: bluesky-pds isolation run — 000 REPRODUCES; root cause = `app` DNS collision on shared proxy

Ran bluesky-pds ALONE (tag 0.3.0+v0.4.219, log /tmp/redfix-bluesky-pds.log). Cold lifecycle GREEN
(install/backup/restore/custom pass; upgrade EXPECTED_NA per recipe_meta — moving pds:0.4 tag). Then
WC5 promote-on-green-cold FAILED exactly as canon: `warm-bluesky-pds.ci.commoninternet.net: not
healthy over HTTPS /xrpc/_health (last status 0)`. So **the 000 reproduces deterministically in
isolation — NOT a sweep-load/ACME-rate-limit flake** (my first hypothesis, refuted).

LIVE DIAGNOSIS (stack left deployed by the failed promote; probed before teardown):
- app service 1/1, healthy: `docker exec app wget localhost:3000/xrpc/_health` → `{"version":"0.4.219"}`;
  app listens on `:::3000`; no restarts. So the PDS itself is fine.
- HTTPS to warm domain → 000. caddy logs flood:
  `tls "failed to get permission for on-demand certificate" domain=warm-bluesky-pds…
   error=… Get "http://app:3000/tls-check?domain=…": dial tcp 10.10.0.X:3000: connect: connection refused`
  (X varies: .2 .4 .5 .6 .8 .9 .10 .12).
- bluesky uses caddy **on-demand TLS** (Caddyfile: `on_demand_tls { ask http://app:3000/tls-check }`,
  `tls { on_demand }`, `reverse_proxy app:3000`). caddy must reach app:3000/tls-check to be GRANTED a
  cert before serving TLS. It can't → no cert → TLS handshake fails → 000.
- WHY can't caddy reach app: **service-name `app` collision on the shared `proxy` overlay.**
  - app is on `warm-bluesky-pds…_internal` ONLY (IP 10.0.3.3). caddy is on `proxy` (10.10.50.223) +
    `…_internal` (10.0.3.6).
  - `docker exec caddy getent hosts app` → returns ONLY proxy IPs (8/8 tries: 10.10.0.4/.5/.6/.10/.12),
    **NEVER the internal 10.0.3.3.** The proxy-net `app` alias shadows bluesky's own internal app.
  - `docker network inspect proxy` shows EVERY stack aliases its main service `app`:
    `drone…_app=10.10.0.2`, `traefik…_app=10.10.0.5`, `warm-keycloak…_app=10.10.0.9`,
    `ccci-reports/bridge/dashboard_app`, … — exactly the IPs caddy hits. None listens a PDS on 3000 →
    connection refused.
  So caddy resolves bare `app` to OTHER stacks' app endpoints on the shared proxy, never its own PDS.

WHY cold passes / warm fails: cold's health window is long (HTTP_TIMEOUT=600) and on first success
caddy CACHES the issued cert; the promote's shorter health window doesn't give caddy a chance to ever
resolve correctly (and here it provably never resolves to 10.0.3.3 at all). The collision is the root
cause; the promote machinery is CORRECT (it refused to write a canonical for an unhealthy 000 — no
canonical.json written, verified).

Classification: **genuine ROUTING/recipe defect — caddy↔app cross-stack `app`-alias collision on the
shared proxy net**, deterministic, reproducible in isolation. NOT a flake; NOT a promote-machinery bug.
Fix approach (M2): recipe PR giving the PDS service a UNIQUE name/alias (e.g. rename `app`→`pds`) so
caddy's `reverse_proxy`/`tls-check` resolve only bluesky's own internal service (no shared-proxy `app`
collision). (Alternatively a caddy-side internal-only resolution; renaming is cleanest.) Will confirm
the exact fix in M2 + verify the warm domain then serves 200.

Cleanup: removed orphaned warm-bluesky-pds stack + its volumes/secrets (promote had left it deployed;
no canonical written). Node clean.

## 2026-06-18T01:05Z — M1: keycloak — warm-domain namespace collision (harness), classification complete

keycloak was de-enrolled (WARM_CANONICAL=False) because its data-warm canonical domain would collide
with the LIVE-warm OIDC provider. Verified the collision STRUCTURALLY (code, no run needed):
- `canonical.canonical_domain(r)` → `warm.stable_domain(r)` → `f"warm-{r}.ci.commoninternet.net"`
  (runner/harness/canonical.py:42-44, warm.py:44-48).
- `warm.WARM_DOMAINS["keycloak"] = "warm-keycloak.ci.commoninternet.net"` (warm.py:27-29) — the
  always-on shared OIDC provider lasuite-*/drone consume for SSO; kept current by roll_warm_infra.
- So `canonical_domain("keycloak") == WARM_DOMAINS["keycloak"]` EXACTLY. Enrolling keycloak as a
  data-warm canonical → the sweep's promote deploy/teardown at warm-keycloak collides with the live
  provider. Confirmed live keycloak healthy (200 /realms/master) — I did not disturb it.

The collision is unique to keycloak: it is the ONLY recipe that is both a live-warm provider (in
WARM_DOMAINS) AND would want a canonical. No collision-free canonical namespace exists today.

Classification: **HARNESS defect — warm canonical domain namespace can collide with a live-warm
provider.** NOT a recipe/flake. Fix approach (M2): make `canonical_domain(r)` collision-free when `r`
is a live-warm provider — e.g. `warm-canon-<r>` (or unconditionally) so the canonical deploy gets a
distinct domain → distinct stack → cannot touch the live `warm-keycloak`. Then set keycloak
WARM_CANONICAL=True and verify it promotes at the collision-free domain WITHOUT disrupting live
keycloak. Minimal blast radius: special-case only providers in WARM_DOMAINS (the 15 other canonicals
keep `warm-<r>`); confirm in M2.

## 2026-06-18T01:05Z — M1: gitea first advance attempt hit a LEFTOVER confound (not the real crash)

First gitea cold@3.6.0 run: cold lifecycle (install/upgrade/backup/restore/custom) ALL PASS; promote
advance FAILED with `FATA warm-gitea.ci.commoninternet.net is already deployed` — NOT the app.ini
crash. Cause: warm-gitea was left DEPLOYED at 3.5.3 by the nixenv-phase sweep (registry said
status=idle but the stack was actually running — a state inconsistency). The advance does `abra app
deploy warm-gitea` assuming the canonical is idle/undeployed; finding it deployed, abra FATAs. This is
the same GREEN-BUT-PROMOTE-FAILED the nixenv phase saw. To reproduce the REAL app.ini issue I undeployed
warm-gitea (docker stack rm; retained data+config volumes → proper idle state) and re-ran gitea
cold@3.6.0 (gitea2). Result pending. NOTE: the "already deployed" promote-failure-when-left-deployed
may be a secondary promote-machinery robustness gap (advance should undeploy-or-chaos an
already-deployed canonical) — will assess after confirming the primary app.ini crash.

## 2026-06-18T00:14Z — M1: gitea warm advance — app.ini read-only JWT crash CONFIRMED (recipe defect)

After restoring warm-gitea to proper idle state (undeployed, 3.5.3 data+config volumes retained),
re-ran gitea cold@3.6.0 (gitea2, log /tmp/redfix-gitea2.log). Cold lifecycle ALL PASS
(install/upgrade/backup/restore/custom — incl. the cold FRESH 3.5.3→3.6.0 upgrade tier). WC5 promote
advance then crash-loops. Live container logs (warm-gitea_..._app, repeated Failed/exit 1):

  modules/setting/setting.go:105:LoadCommonSettings() [F] Unable to load settings from config:
  error saving JWT Secret for custom config: failed to save "/etc/gitea/app.ini":
  open /etc/gitea/app.ini: read-only file system

EXACTLY the canon-documented crash. Mechanism: the recipe mounts app.ini as a docker `config`
(read-only by design) at /etc/gitea/app.ini (compose `configs: - source: app_ini target:
/etc/gitea/app.ini`, app.ini.tmpl). gitea 1.24.2 (3.6.0), on the warm REATTACH of the retained
3.5.3 config volume, decides to (re)generate+SAVE a JWT secret to app.ini → read-only fs → FATA at
config-load, BEFORE any DB migration (so the 3.5.3 data volume stays intact — confirmed canon).

Why cold passes but warm crashes: the cold fresh deploy + cold chaos-upgrade use freshly-generated
secrets consistent with a freshly-initialized config, so gitea never needs to rewrite app.ini. The
warm advance reattaches an OLDER retained config-volume state (seeded under 3.5.3) against the new
run's secrets/3.6.0 binary → gitea reconciles by trying to persist a JWT secret → read-only crash.

Classification: **genuine RECIPE defect** (gitea 3.6.0/1.24.2 + read-only app.ini docker-config mount
on the warm-reattach advance), deterministic, reproduced first-hand. NOT a flake, NOT promote
machinery. Fix approach (M2): recipe PR making app.ini writable on the advance path — e.g. render the
config into the WRITABLE `config:/etc/gitea` volume via an entrypoint (not a read-only docker config),
OR ensure the persisted secrets are accepted without rewrite. (Secondary harness option: canonical
advance falls back to clean re-deploy when in-place config rewrite is impossible — but that loses the
reattach data-warm property; recipe fix preferred.) Ties to LFS PR #1 (app.ini secret handling).

ACTION NEEDED after run exits: warm-gitea is left crash-looping at 3.6.0 → restore it to 3.5.3
(redeploy the known-good canonical version) so the canonical is healthy again. Data volume intact.

## 2026-06-18T00:25Z — M1 CLAIMED (6/6 investigated, isolated, classified)

mumble repeat #2 (mumble2): ALL tiers green again incl. handshake; canonical re-promoted green
(ts 20260618T001730Z). So mumble = 2× reproducibly green in isolation → load/timing FLAKE confirmed.

All six classified with first-hand isolation evidence (or code proof for keycloak). Two canon
root-causes were CORRECTED by isolation: discourse (not a timeout/deploy-FATA — it's a stale cc-ci
overlay test asserting an unreleased migration) and mattermost-lts (not a loaded-node race — a
deterministic recipe restore defect: missing `backupbot.restore.post-hook`). bluesky's 000 is NOT a
load/rate-limit flake (my initial hypothesis) but a deterministic caddy↔app `app`-alias DNS collision
on the shared proxy. gitea app.ini read-only JWT crash reproduced first-hand. keycloak collision proven
structurally in code.

Node clean: warm-gitea idle@3.5.3 (volumes retained), orphaned warm-bluesky removed, only live
warm-keycloak up (healthy 200). Claiming M1; will start M2 fix design while awaiting the Adversary
verdict (keep an unblocked item in hand).

## 2026-06-18T00:25Z — M2 prep (gated on M1 PASS): bluesky fix refinement

While parked at the M1 gate (no node deploys — Adversary cold-verifying), refined the bluesky fix:
cc-ci's bluesky tests probe via HTTP (/xrpc/_health), but the GENERIC harness defaults to
`service="app"` (deployed_identity/_app_container). So RENAMING the recipe's `app` service → `pds`
could break generic harness assumptions. Cleaner fix: keep the service named `app` but give it a
UNIQUE network ALIAS on the internal net (e.g. `aliases: [pds-internal]`) and point caddy at
`pds-internal:3000` (reverse_proxy + on_demand_tls ask). A unique alias has no collision on the shared
proxy (only the bare `app` alias collides), and the service name stays `app` → zero cc-ci-side
breakage. Will validate this exact approach in M2 after M1 PASS.

## 2026-06-18T01:21Z — M1 PASS; starting M2

Adversary M1 verdict: **PASS** @01:18Z — all 6 classifications cold-verified CORRECT by its OWN
isolation re-runs (discourse/mattermost/mumble/bluesky/gitea) + code-verify (keycloak). No VETO.
"Builder cleared to proceed to M2." Two canon root-causes corrected and confirmed (discourse: not a
timeout, stale overlay; mattermost: not a load race, recipe defect). bluesky reclassification (recipe,
not warm-machinery) confirmed against the plan's prior.

Starting M2. Plan: recipe PRs (mattermost-lts, bluesky-pds, gitea) via the recipe mirror+PR flow
(`!testme`-verified, never merge); harness fixes (keycloak collision-free canonical_domain + enroll;
mumble handshake stabilization) on a cc-ci branch; discourse overlay-scope decision. Node now mine
(Adversary done). Will examine the recipe-create-pr flow first, then execute one fix at a time.