Adds /api/v2/narrative, serving a one-sentence summary of how unusual a given
day was at a location. Generation is a batch job (scripts/generate_narratives.py)
that talks to a local Ollama; the endpoint only ever reads the cache, so no
request path depends on the inference host being reachable or powered on. A miss
answers 200 with narrative:null and status "pending" rather than 404.
The model is handed tier labels and percentile directions only, never a
temperature or rainfall figure, so it has nothing to misreport; the sanitizer
then rejects any output that carries units, a forecast or advice. Fact lines
state direction ("warmer than 82% of nights at this time of year here") because
the earlier "82nd percentile (High)" form was read backwards and phrased a warm
night as a cool one. Narratives are unit-free as a result, which also keeps one
cached row per cell/date instead of a per-country degree variant.
Storage is a new `narrative` table in the derived store, LOGGED on Postgres
unlike its two UNLOGGED siblings: losing it costs an LLM pass over the whole
grid, not a recompute from parquet.
Twelve documents under docs/onboarding/ covering orientation, local setup,
the repo map, per-domain deep dives, the cross-service contracts, CI and the
release flow, infra and secrets, observability, task recipes, and a list of
which docs in this tree are currently stale.
Every command in the setup guide was run against this checkout: the backend
suite (429 passed, 8 skipped), the frontend Go suite, a venv boot of the
backend, and a build+boot of the Go frontend.
Two findings recorded along the way:
- static/units.js's F_REGIONS is guarded by no test, despite three source
comments claiming "a test asserts all three stay identical". The Go test
only cross-checks the Go copy against the backend's Python. All four copies
are currently identical.
- backend/ and frontend/docker-compose.test.yml still default to the retired
emi/thermograph-backend/app image path, and the frontend harness pins the
split-era v0.0.2-split-ci tag.
Claude-Session: https://claude.ai/code/session_01AfXqHrxCJLs2D7hpQkiUiJ
The dev → main promotion could not merge: `main` sat 22 commits ahead of `dev`
with content dev had never seen, and git found genuine conflicts in CLAUDE.md
and README.md plus modify/delete on the eight workflow files #87 removed.
The cause is that recent promotions were merged with squash rather than "Create
merge commit". A squash replays dev's commits onto main under new SHAs, so the
two branches share content but not history — main appears to have independently
edited the same files, and every subsequent promotion inherits the divergence.
`promote` warns about exactly this.
Resolution: both conflicts take dev's side, not as a default but because main's
side describes workflows that no longer exist (`*-deploy-dev.yml`,
`backend-build-push` and friends). Keeping it would reintroduce the class of
staleness the root CLAUDE.md now forbids — statements naming files that are not
there. The eight deleted workflows stay deleted; the merge result carries the
nine that #87 left behind.
Verified after resolution: no conflict markers anywhere, no unmerged paths, no
document naming a deleted workflow, all nine workflow files parse, and
shellcheck is clean across the tree.
Promotions from here should use "Create merge commit" so this does not recur.
shell-lint failed on two single-quoted strings. Both are deliberate and
double-quoting either one would break it silently rather than loudly.
In render_centralis_secrets, the verification step runs `bash -c 'set -a; . "$1"'`
with the rendered file passed as an argument. The "$1" must be expanded by that
inner shell. Double-quoted, it would interpolate the OUTER $1 — the function's
own first argument, the repo root — so the check would source the wrong path and
still report success. That check exists precisely to catch a bad render, so a
false pass there is worse than no check.
In verify-centralis-render.sh, the string is executed by the REMOTE shell over
ssh. Double-quoted, mktemp would run locally and the script would send the host
a path that does not exist there.
Suppressed with the reason at each site rather than loosened globally.
Claude-Session: https://claude.ai/code/session_0182KTMrsTHJc3TcewCatJFY
Brings the last two credential stores under SOPS, so all three environments and
the control plane are managed the same way.
dev.yaml — dev was not "intentionally not wired in", it was broken
The vault README claimed dev had no real secrets. Both halves were false.
Dev needed secrets it did not have: thermograph-dev-daemon-1 has been in a crash
loop, restarting every ~60s with "neither THERMOGRAPH_INTERNAL_TOKEN nor
THERMOGRAPH_AUTH_SECRET is set; refusing to start" — so dev has had no Discord
gateway and no job timers. With THERMOGRAPH_BASE_URL unset the app also falls
back to https://thermograph.org, so dev's IndexNow pings and verification links
claimed to be prod. dev.yaml gives it its own generated AUTH_SECRET and a real
base URL, and 12 values total.
And dev already held production credentials: backend-deploy-dev.yml injects
S3_ACCESS_KEY/S3_SECRET_KEY into every dev deploy, and the only provisioned
keypair for that bucket is read-write — on the bucket holding prod's backups.
Dev does not need them; without bucket creds the lake service 503s and history
falls through to the archive. Those workflow lines still need removing.
dev renders dev.yaml ALONE, via THERMOGRAPH_SECRETS_SKIP_COMMON=1. Eleven of
common.yaml's sixteen values are live production credentials, the render is a
plaintext concatenation consumed through env_file:, and dev is the operator's
desktop AND the Forgejo runner executing unreviewed dev-branch code with the
docker socket mounted. An override in dev.yaml would not help — last-wins
governs consumers, but prod's value is still physically a line in the file.
Verified: current renderer leaks 28 lines onto dev including both S3 keypairs
and the VAPID private key; patched gives exactly 12 keys and no prod credential.
Beta's render is byte-identical either way.
centralis.prod.yaml — its own file, its own renderer
/etc/centralis.env was hand-edited, which is how a JSON registry got written
unquoted into a shell-sourced file tonight and silently collapsed three
identities to one. The renderer now owns the file.
It is service- AND environment-scoped, not folded into prod.yaml, because
/etc/thermograph.env is loaded into every container in the app stack. Folding
these in would put CENTRALIS_AUTH_TOKEN and CENTRALIS_TOKENS — which
authenticate an endpoint carrying run_on_host and sql_query(write) — into the
web backend's environment, turning a read-anything bug in the app into a
foothold on the control plane.
Quoting is guaranteed three ways rather than assumed. Critically, `sops -d
--output-type dotenv` emits values verbatim, so reusing the existing render path
would have reproduced tonight's bug exactly. The new function decrypts to JSON
and emits POSIX single-quoted assignments; it then sources its own output in a
clean shell and compares every value before touching /etc; and it refuses to
write unless CENTRALIS_TOKENS parses as a non-empty subject->token object after
sourcing. Tested against 16 hostile values including embedded quotes, newlines
and ;rm -rf /.
Value-identity proven, not asserted: rendered vs live compared as *effective*
values on prod (a textual diff would report a false difference, and would report
a false match if the vault had captured quote characters as part of the value),
plus both env files producing a byte-identical `docker compose config` digest.
9 keys, both token subjects preserved. Nothing deployed.
Note for later: /etc/thermograph.env is also shell-sourced, and survives on raw
dotenv only because none of its 32 current values contains a shell-special
character. It is one quoted secret away from the same bug.
Claude-Session: https://claude.ai/code/session_0182KTMrsTHJc3TcewCatJFY
Two changes to how credentials are stored and distributed.
1. common.yaml now exists.
The renderer has always concatenated common.yaml then <env>.yaml, host winning,
and the README has always documented common.yaml with a checkmark. The file was
never created. So every value shared by prod and beta was duplicated in both
vaults, free to drift apart with nothing to detect it.
16 values move to common.yaml: VAPID keypair, metrics token, IndexNow key,
REGISTRY_TOKEN, the S3 endpoint/bucket and both S3 keypairs, plus shared config.
5 stay per-host because their values genuinely differ (APP_CPUS, DB_CPUS,
DB_MEMORY, WORKERS, THERMOGRAPH_BASE_URL). 8 exist only on prod — the Discord
and mail credentials, which beta does not have at all.
Three more are held back deliberately despite being identical today:
POSTGRES_PASSWORD, THERMOGRAPH_AUTH_SECRET, THERMOGRAPH_DATABASE_URL. These are
the credentials that let one environment act as another, and beta is the more
exposed box — it serves public Forgejo and Grafana. They match only because beta
was seeded from prod. Keeping them per-host costs one line each and preserves
the ability to diverge; putting them in common.yaml would encode the equivalence
as intentional and make breaking it a migration rather than an edit. Reasoning
recorded in the vault README, since a future reader will otherwise "fix" it.
Done without any plaintext leaving prod: ciphertext was shipped up, decrypted
against the host's age key, recombined, re-encrypted, and shipped back. The
consolidation refuses to write unless it has proved merged(common + env) equals
the original env vault exactly — same keys, same values — and re-verifies from
the written files afterwards. Both checks passed for prod and beta.
2. render_thermograph_secrets no longer fails open.
One `return 0` covered two different situations: "this host is not configured
for SOPS" (true of the LAN dev box, and correct to succeed) and "this host IS
configured but its vault is not where we looked". The second is a failure, and
returning 0 made it a silent no-op wearing a success code — a deploy would
report success having rendered nothing, and the host would keep serving whatever
/etc/thermograph.env already held, including after a rotation.
The likely cause is passing the wrong root: the function wants the directory
containing deploy/secrets, which on the hosts is /opt/thermograph/infra, not
/opt/thermograph. That mistake looked exactly like "not configured here", which
is why it went unnoticed. It now exits 1 and names the probable cause.
Claude-Session: https://claude.ai/code/session_0182KTMrsTHJc3TcewCatJFY
The previous commit swept in .claude/worktrees/city-resolver and
.claude/worktrees/thermograph-mentions as embedded git repositories. Those are
other Claude sessions' live checkouts and have no business in this tree.
Ignoring the directory so `git add -A` cannot do it again — the repo already
has a documented history of parallel sessions colliding through shared
checkouts, and this is the same hazard wearing a different hat.
Claude-Session: https://claude.ai/code/session_0182KTMrsTHJc3TcewCatJFY
`grep -oE "^[A-Z_]+="` cannot match a key name containing a digit. This estate
has exactly six such names — all the *_S3_* keys — so a prod audit returned 26
keys against 32 real ones, and reported the S3 and lake credentials as missing
from both live hosts.
They were present the whole time. The phantom was independently reproduced
twice by re-running the same pattern, which is what made it convincing, and it
was briefly recorded as one of two root causes of the lake being unqueryable
(#56). That issue has one cause: the missing duckdb-lake image.
An audit that under-reports is worse than no audit. A missing-secret finding
sends someone to provision a credential that already exists, and in a rotation
tool it would justify writing over one.
Fixed to `^[A-Z][A-Z0-9_]*=` in both the skill and key_gaps.py's docstring. The
match still stops at the `=`, so no value is read — that property is the reason
this grep exists rather than a parser.
Centralis's secrets_gaps carried the same bug and is fixed separately, with a
regression test.
Claude-Session: https://claude.ai/code/session_0182KTMrsTHJc3TcewCatJFY
The estate had zero alert rules. The only contact point was Grafana's factory
default pointing at the literal string <example@email.com>, and beta's
untracked compose override routed Grafana's SMTP at prod's Postfix — so
alerts, had any existed, would have been delivered by the box most likely to
be on fire. Prod served 86 5xx in 24h including five /healthz failures and
nobody was told.
Alerting (deploys to beta, which is where Grafana runs):
- 12 Loki-based rules. There is no Prometheus in this estate and Loki is the
only datasource, so every rule is log-derived.
- Thresholds come from a 24h backtest that happens to contain a real ~20min
prod outage at 04:20Z. Absolute counts, not ratios: prod runs ~7 req/min, so
one bad request is 1.4% and a ratio alert would scream all night. The 5xx
burst rule fires on the outage's 45 and 22 buckets and on nothing else in
the day; the largest benign bucket all day was 5.
- Routed to a new private #ops-alerts channel, not to any existing channel —
#weather-events, #announcements and #prod are product surfaces that notify
real subscribers.
- AlertingWatchdog is a dead-man's switch; its value is its absence.
- Delivery is proven, not assumed: the rules were provisioned into a throwaway
Grafana against beta's live Loki and a real alert arrived in Discord. This
matters because GET /api/v1/provisioning/contact-points returns [REDACTED]
for the URL — a contact point holding an uninterpolated env var looks
perfectly healthy and pages nobody. The only proof is a message arriving.
CI gains a structural check, because the existing one only proves YAML parses:
an alert rule whose condition names a missing refId is valid YAML, provisions
cleanly, and never fires. It also hard-fails on a literal Discord webhook in
the repo. Verified against all three breakages deliberately introduced.
Postfix supervision:
- The 13h outage was a boot-ordering race, not a Docker renumbering: postfix
started at 08:09:49, wg0 came up at :51, postfix fataled at :52 on a missing
docker_gwbridge address, and dockerd did not finish starting until 08:10:16.
Stock postfix@.service is ordered only After=network-online.target and ships
no Restart=, so one lost race became a permanent outage.
- An ExecStartPre gate now blocks up to 60s until every inet_interfaces
address actually exists, which absorbs the transient case inside a single
start attempt. That makes bounded retry correct: 5 attempts in 600s, then
failed — a genuinely broken config reaches a visible failed state in ~100s
instead of re-fataling every 15s forever.
- A 5-minute watchdog timer retries indefinitely and runs reset-failed, so
"failed" still self-heals. Worst case is ~5 minutes, not 13 hours.
- Wants=, not Requires=: a dockerd failure must not take down the loopback and
mesh listeners that do not depend on Docker at all.
Health checks must read config with `postconf -c`, never postmulti/postqueue/
postfix — those three RESOLVE inet_interfaces and so fatal precisely when an
address is missing, which made the first version of this check report
status=ok bound=0/0. A health check that fails open is worse than none.
DEPLOY.md carries the monitoring contract, including that systemctl is-active
postfix is a known-false signal: postfix.service is a wrapper whose
ExecStart=/bin/true, so it reports active forever while the real postfix@-
instance is failed with zero listeners. Reproduced live.
Known gap, documented not fixed: Alloy ships Docker stdout, Caddy files and
app JSONL, not journald — so no Postfix line reaches Loki and the mail rules
cannot fire until loki.source.journal is added.
Claude-Session: https://claude.ai/code/session_0182KTMrsTHJc3TcewCatJFY
After a cell's archive advances, the first content request recomputes a 45yr payload cold. Add warm_cities.warm_content — idempotent, cache-only, bounded — that pre-populates the content derived-store under the same (kind,key,content_token) the routes read, wired into the leader-gated notifier loop (50 cities/30-min tick) so the ~1000-city set stays warm within hours of each daily advance.
The three content handlers loaded the full ~45yr archive before checking the cache, so even hits paid the load. Restructure to resolve cell + compute content_token cheaply first, check the store, and load history only inside the miss-only build() closure. Cache hits and 304s now return without touching the archive. Keys/kinds/ETag unchanged.
Content pages keyed cache validity on the full-history hist_end, invalidating ~1-2x/day and paying a 45yr load even to compute the token. Add content_token(cell_id) = PAYLOAD_VER:CONTENT_VER:max_date, backed by an indexed MAX(date) (climate_store.history_max_date / climate.history_max_date), so it survives intra-day top-ups and needs no full-history load.
