notes

I need to setup a password manager to keep track of my projects' credentials. Right now they are spread across so many places. Some in notes, some as files, some in .env, and some actually in a password manager. No idea what's where anymore. Also the links and details of all projects are getting out of hand. Which VPS, which domain, what renews when, where's the login. Need to plan something for these as well.

Testing Dockhand on my local, looks very interesting. Good UI, does almost everything a docker management tool needs to do, also feels simpler to use than Portainer CE. Will share more when I find any more interesting stuff about Dockhand.

Exploring a lightweight way to monitor Docker containers and if any errors, triage them, identify issues and help resolve. I use Beszel for monitoring but Docker logs are not there. Have used Dozzle for another project which handles Docker logs and basic management. Still exploring options, haven't finalized on the final setup.

Can feel the exhaustion. Did so many things last week. This week is going to be slow and I need to stick to the working hours I set for myself. AI doesn't replace the brain thinking, if anything it increases thinking and reading and reviewing. More to think than to write code.

DeepSeek Harness launched — everything is a plugin, ~139k GitHub stars in 4 days. Very curious to try it out, see what it offers and if it's useful for me. Link: https://deepseek.com/harness GitHub: https://github.com/deepseek-ai/deepseek-harness

GLM-5.3 launched this week. Same base model as 5.2 — no new pretraining, everything from post-training. Z.ai's own line: "Scaling post-training is all we did." Benchmarks look promising, but need to test this to see if it makes any major improvements in real life.

Configured CamoFox browser with DataImpulse proxy. Picked DataImpulse because they have a pay-as-you-go plan and wanted to test it anyway. Costs $1/1GB for regular residential proxy, $5/1GB for premium residential. Will see where this takes us in research and browsing.

Refactored the large Rust files in LiteBin and set up CI with fmt, clippy, and tests. Release now runs these before releasing. Need more tests, but for now focusing on the next core features as documented in planning.

Started refactoring LiteBin. Over the period some files have been growing in size and becoming god files — difficult to maintain in the future. Starting those now. Will review and see what all remains after the refactor. LiteBin is written in Rust and I've been postponing CI. Maybe after this I'll add clippy and a test run as part of the version release.

Exploring HTTP/3 and QUIC — an interesting read. HTTP/2 introduced multiplexing of HTTP streams over a single TCP connection. HTTP/3 gave multiplexing over QUIC on top of UDP.

Tried to summarize a Reddit post from the Oracle Hermes setup. Even with Jina and Camofox, the setup struggled — Reddit and similar sites aggressively block datacenter IPs. So the last resort: setting up a residential proxy strictly for reading websites when everything else fails. Not the first tool to reach for — SearXNG, Jina, and Camofox handle most cases. But for sites that hard-block datacenter ranges, a residential proxy is the only thing that reliably gets through.

Enabled Syncthing in the Docker Hermes setup on Oracle so vault files sync back to the laptop. Had to open the Syncthing port via Oracle Cloud's network security rules (their VCN ingress rules) — not just on the VPS itself. Oracle Cloud has that extra network layer on top of the OS firewall. If the provider didn't have this extra network rules layer, you'd just open the port directly on the VPS (iptables/ufw). Good to know which providers add an external firewall — Oracle does, Hetzner doesn't (by default), etc.

Hit blocks doing research from the VPS-hosted Hermes agent — web search and extraction kept getting blocked. The unblocking stack: SearXNG for self-hosted meta-search, Jina AI for extracting pages as markdown, and Camofox as a stealth browser with VNC attached for visual inspection. The goal is simple: enough tooling for a smooth research-supported assistant without getting IP-blocked or bot-detected on every call. Each piece solves a specific failure mode — SearXNG aggregates search engines (no API keys, no shared rate limits), Jina handles the "give me this page as readable text" job, and Camofox handles the "site refuses non-browser clients" job.

Having that feeling that agent upkeep is a hidden layer of work nobody accounts for when setting these up on a VPS. It's easy to say "just deploy it" — the complexity that comes after is the real cost. Every tool bolted on to unblock a problem becomes its own thing to maintain. SearXNG, Jina, Camofox, VNC, Directus, the VPS itself — suddenly the agent isn't an agent, it's a small platform with a platform's maintenance load. The honest question: is all this worth it for the benefits, apart from the learning (which is genuinely good)? The trap is that infra work feels productive but doesn't produce content. If upkeep starts eating the time it was supposed to free up, that's the signal to simplify.

Hermes is live on Oracle Cloud. Created profiles for my brother. Now using DeepSeek V4 as default model via Neural Watt — will monitor usage and results, then decide if we need to switch to another model or provider.

Spent the last day getting LiteBin to work with remote agent node deployment. Deploying Hermes on Docker helped me see the gaps in LiteBin's agent node deployment — now I'm fixing them.

Retried uploading the 900MB image today. Still failed. Need to enhance LiteBin — two things on the list: Resumability — so failed uploads don't start from scratch Direct-to-agent upload — skip the master node entirely Planning to tackle both today. The flow I'm thinking for direct upload: when deploying to an agent node, give the user an option to skip the master. The master requests the agent to generate a token, takes the URL with that token, and passes it back to the client. The client then uploads directly to the agent, which verifies the token and accepts the file. That's the MVP plan for now.

LiteBin has multi-node support but a simpler deployment flow. From my local machine, I build a custom Docker Compose image with profile customizations for Hermes, compress it, and send it directly to the VPS for deployment — no Docker image registry needed. This simplifies things, but it has its own set of problems. Our custom image was a 900MB tar file that needs to go from my machine to the VPS. The current image upload doesn't support resumability. If it stops mid-way, I have to start over from scratch. I added a retry to the LiteBin CLI today — earlier, after 3 failed uploads it would exit and I'd have to recreate the image too. Now at least it retries the upload without rebuilding. These experiments are helping me find the gaps in LiteBin and see the real use cases as well.

Spun up an Oracle Ampere VPS — 2 OCPU, 12GB RAM, Mumbai region — for the Hermes Docker setup. Decided to use LiteBin (l8bin.com) as the deployment tool. Instead of running it standalone, I made this new VPS an agent node to my existing LiteBin master in Europe (Netcup VPS). Didn't work. Jio (my ISP) has connectivity issues with European servers and the connection keeps timing out. Couldn't deploy the Hermes setup. Will retry tomorrow. There's also a design issue with LiteBin I hit today. Right now, the agent node can only communicate through the master. So when I deploy an image, the tar file has to travel from my machine to Europe first, then from Europe to Mumbai. That's a lot of unnecessary latency. I could fix this by making LiteBin talk to the agent directly during deployments. But that needs more planning. For now, I'll retry the deployment tomorrow and deal with the architecture later.

Spent the last day setting up a Docker Compose for Hermes Agent with: Custom onboarding flow Skills pre-configured to set Neural Watt as default LLM provider Now I can configure this for my brother — he gets his own profile. Onboarding is triggered from Telegram, asks questions, and helps set up an Obsidian vault based on his needs. Next plan: deploy this on free Oracle tier and see how it goes.

Had to fight with flows in Directus. The agent was struggling and failed then I helped it, but finally we succeeded.

Codinglog is live end-to-end. I capture notes via Telegram; Hermes writes each one to the Obsidian vault as markdown and publishes it to Directus through a scoped author role. The full pipeline works.

Backup setup testing done on test server. Next: deploy the restic backup setup on the real VPS where Hermes runs. On it today.

Not worried about SQLite page-torn for now. The backup covers MD files fine, and SQLite consistency will be handled when DB backup gets its proper phase later. Current phase: file backup only, ship it.

Kept DB backup as a separate task — DBs need more thought around locks and corruption during backup. For now: file backup only, which is all that matters for the current usecase. Current backup targets: Hermes + Obsidian vault. Hermes has a SQLite db — concern: hope it doesn't break with a plain file-level restic snapshot.

Setting up external backup on an old VPS. Backup VPS: keep ~1 month of backups here. Tight on resources → tooling choice matters. Home NAS: weekly copy of those backups (offsite tier). Tool picked: restic — chose it for the resource/efficiency fit on the constrained box.

Picked Astro + Directus as the stack for the Codinglog blog.