Deploys the generation-flow refactor (unified generate, streaming fixes,
tool-calling, strict model validation, AddBos::Never).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
GitHub path filters do not match submodule gitlink changes, so the
`paths: apps/**` filter meant a submodule pointer update never triggered
the deploy. Drop the filter so any push to main deploys (matches the
documented "Push to main -> nix build -> systemctl restart").
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Docker was decommissioned 2026-08-02 but docker-build-push.yml and
deploy-docker.yml were left behind. They still listened to
repository_dispatch: [submodule-updated], so every app push queued a
redundant Docker build alongside the intended Nix deploy.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
next start re-reads next.config.ts at runtime; the OLD store had only
.next/public/package.json/node_modules, so images.maximumDiskCacheSize
fell back to defaults (undefined) and the image-optimizer disk LRU did
mkdir(<store>/.next/cache/images) -> EACCES on every optimized image.
Also now that next.config.ts sets maximumDiskCacheSize:0 +
isrFlushToDisk:false, shipping the file makes runtime match the baked
build config, eliminating the EACCES spike.
Neighbor multicast to ff02::1 (MLDv2 reports from other hosts) was
hitting the LOG+DROP tail — ~1800 FW6-DROP lines/6h in journald,
i.e. ~5/min of pure log spam from ICMPv6 layer-2 discovery. Accept
link-local multicast ranges before the logging rule.
Restore upload.asepharyana.my.id/.web.id routing lost during Nix migration
(docker label router removed with container). Upload chain uses 2GB body
buffer + dedicated 300/100 rate limit, service targets bun on host:3000.
DeterminateSystems/nix-installer-action defaults to determinate:true
which tries to auth with FlakeHub via GitHub JWT. We don't use any
FlakeHub flakes, so disable it — skipping the auth entirely.
Separate deploy job failed because it used ls to find store paths
by name, finding OLD local paths instead of the freshly copied CI
paths. Now each service builds, copies, and updates its profile
in a single job using the exact store path from the build output.
No more guessing which path is the right one.
- Fix nix copy URL: ***@ → $VPS_USER@
- Store path from build output, passed across jobs
- Deploy job waits for all builds via needs: build
- SSH key setup in its own step, guarded by main branch
- Only deploy on main branch pushes
- Remote deploy script fetches from VPS nix store
- Add flake.nix with derivations for hub (Next.js), scraper, tools, llm-api
- Create hub systemd unit (port 3099)
- Update Traefik dynamic config to point to host hub service
- Add iptables rule for Docker-to-host communication
- Use fetchGit for submodule source resolution
- Replaced llama-cpp-2 apply_chat_template with minijinja rendering
- reasoning_content separate from content in API response
- Works with MiniCPM5 thinking model natively
