The nightly refresh.yml and deploy.yml each substituted asset/nginx/
site.conf.tmpl with their own inline python. When bb2f5b1 templated the
listen line as {{WEB_LISTEN}} (moving the vhost behind oolon's stream SNI
router), it added the substitution to the template and deploy.yml but not
to refresh.yml. The daily refresh then rsynced a literal
`listen {{WEB_LISTEN}};` into /etc/nginx/conf.d/rob.tn.conf, `nginx -t`
failed for the whole edge, and — because the file is written into the live
conf.d before it is tested — every vhost's reload (including the step@
cert renewals) stayed frozen. Internal vhosts, cichlid.internal among
them, served certs that had expired days earlier while the renewed certs
sat unused on disk.
- Replace both inline renderers with script/render-site-conf.py, shared by
deploy.yml and refresh.yml so they cannot drift on what they substitute.
- Guard rails: the renderer fails if any {{PLACEHOLDER}} lacks an env value
or survives substitution, so a forgotten/misnamed variable is a red build
on the runner instead of a broken vhost on the edge.
- Add the missing WEB_LISTEN to refresh.yml's env (the immediate drift).
- Rename the template's {{DOCROOT}} to {{WEB_ROOT}} so every placeholder
maps to the env var of the same name.
- Remove script/deploy.sh: the third, unused renderer of the same template
(superseded by the Actions workflows) and a standing source of drift.
- Docs (readme, CLAUDE.md) updated to the Actions-only deploy path.
Known follow-up (needs a sudoers change + infra-setup re-run on oolon, so
out of scope here): the rendered vhost is still rsynced straight into the
live conf.d and only then `nginx -t`'d, so a valid-but-wrong config could
still wedge nginx. Stage-validate-swap with rollback would close that.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QsH1rcWQYtRVhvaftiKm22
oolon's TCP 443 belongs to the stream SNI router, which ssl_prereads the
handshake and forwards to the local https tier on 127.0.0.1:14443 with
PROXY protocol. site.conf.tmpl predates that and still bound 443 itself,
so every deploy and every daily refresh rsynced a vhost that collides
with the router.
Nothing in the pipeline caught it. `nginx -t` only detects duplicate
listeners within a context, not across http{} and stream{}, and
`systemctl reload` merely sends SIGHUP, so it exits 0 while nginx logs
"bind() to 0.0.0.0:443 failed (98: Address already in use) ... still
could not bind()", aborts the reconfiguration and keeps its old cycle.
The deploy went green while oolon's running config was frozen. It stayed
frozen for a day, stranding every cert the step@ timers renewed on disk
until eleven internal vhosts were serving expired certs. A cold start
would have failed outright, taking the whole public edge down.
Template the listen line from WEB_LISTEN (manifest web.config.listen for
script/deploy.sh, which renders the same template), and assert that the
reload landed by requiring a fresh worker generation, dumping the nginx
error log when it did not.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0182wzZE8DguMPWhxD21gfP2
The fedora-44 runner reaches the public internet (the CV gist on
api.github.com prerendered fine) but not the public rob.tn — split-horizon
DNS — so every moments-API-sourced route baked empty loading states while the
build still "succeeded". prefetchQuery swallows fetch errors, so the empty
snapshot shipped silently.
- Point VITE_API_BASE at the internal API the deploy already reaches
(http://nikola.kosherinata.internal:42424/v1), verified reachable over the
mesh. Only affects the SSR/prerender build; the browser bundle still uses the
relative /api/v1.
- run-prerender.mjs now aborts (exit 1) if zero dynamic routes are enumerated,
i.e. the API was unreachable — so an empty snapshot can never deploy again.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7zF7Kf4JqDwa6M8Qgge9M
CI's pnpm did not honor the build-script allowlist from package.json (removed
in pnpm 10) nor from pnpm-workspace.yaml under `pnpm --dir ui`, so build-web
kept failing with ERR_PNPM_IGNORED_BUILDS. Make it version- and discovery-
independent: run in ui/ via working-directory, install with --ignore-scripts
(no approval gate), then `pnpm rebuild @swc/core esbuild` to place the native
binaries vite needs. Verified locally: cold install + rebuild + vite build all
succeed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7zF7Kf4JqDwa6M8Qgge9M
The `rust` runner image has cargo + musl but no node/pnpm, so the web build
(and the previous npm-install workaround) can't run there. The fedora runner
images bake in node + pnpm + rsync. Split the build:
- build-binaries on `rust` (cargo musl + lint/test gate)
- build-web on `fedora-44` (pnpm install + prerender, no install step)
Deploy jobs move to `fedora-44` (has rsync/ssh/pnpm/ca-trust) and depend on
the relevant build job.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7zF7Kf4JqDwa6M8Qgge9M
The build job failed with `corepack: command not found`: Fedora's nodejs
package (gongfoo runner-rust image) ships node + npm but not corepack. The
job runs as root, so install pnpm globally via npm instead. Longer term,
bake pnpm into the runner image to drop this step.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7zF7Kf4JqDwa6M8Qgge9M
Make the site fully prerendered so a plain curl returns complete content
for every route (crawlers / AI screening tools see real text, not an empty
#root), while humans keep full client interactivity.
Prerender:
- Build-time per-route render: prefetch data, renderToString, inline the
dehydrated react-query cache as window.__RQ_STATE__; client hydrateRoots
and refetches live (activity stays fresh; crawlers get the baked snapshot).
- New entry-server.tsx + prerender/{prefetch,routes,meta}.ts + run-prerender.mjs;
shared lib/ranges.ts keeps SSR and client query keys identical.
- pnpm build now: tsc -b -> vite client build -> ssr build -> prerender.
- API base absolute at build (VITE_API_BASE), relative /api/v1 in the browser.
- CSS imports moved to the client entry so the tree imports under Node.
- schema.org Person + Occupation JSON-LD and per-route title/description/og.
- UTC + explicit field widths on shared date formatting so SSR and client
hydration match byte-for-byte (fixes hydration mismatch on /activity).
- Strip non-text gist content from the CV fetch (1MB -> 25KB gzipped page).
Deploy (Gitea Actions, replaces script/deploy.sh):
- deploy.yml: on push to main, lint/test gate, build api+worker as static
musl binaries (pure-rustls, no glibc skew) + prerendered web, deploy each
over SSH as gitea_ci with scoped sudo.
- refresh.yml: daily cron re-bakes only the web snapshot so gist/activity
edits propagate without a push or bouncing the api/worker.
- script/infra-setup.sh + asset/sudoers.d/{api,worker,web}-host.conf for
one-time per-host provisioning. Secrets: RSYNC_SSH_KEY, QUERY_GITHUB_TOKEN,
QUERY_GITEA_TOKEN.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X7zF7Kf4JqDwa6M8Qgge9M
