Around ninety percent of what Tsunami is tuned for is engineering. Not snippets in a sandbox — reading your logs, editing your files, deploying your app, and checking the result before it says it is done.
It knows the difference between your production server and your laptop, and it checks which machine it is on before it runs anything — so it never fires a Linux command at a Mac, or scans a whole disk when Spotlight would answer in a second.
→ Checking disk and memory df -h --total && free -h ✓ 936G free · 201G RAM available → Listing processes pm2 list ✓ 13 online → Scanning logs for errors ✓ found a stuck job at 103% CPU → Checking open ports ✓ 30 listening · flagged Telnet on 23 → Writing this up for the team ✓ logged to updates
Three separate surfaces, each with its own rules, so it never confuses one for another.
Each room holds its own SSH credentials, encrypted at rest. Tsunami reads logs, edits code, runs builds and restarts services on that room's machine.
Through the local agent you install yourself. It checks which OS you are on and adapts — Spotlight on a Mac, PowerShell on Windows.
Tsunami can inspect and maintain its own infrastructure, which is how it keeps the model pool healthy without anyone watching it.
The failure modes matter as much as the wins. Here is how it behaves when things go wrong.
It does not blindly retry the same thing. It is told to narrow the approach — scan one folder rather than a whole disk — and to say plainly if it still cannot measure something.
Rather than discarding the work and showing an error, it stops and writes up what it learned, what it did not reach, and the single best next move.
On a Mac it recognises the privacy prompt for what it is and tells you which setting to grant, instead of insisting the file does not exist.
Esc stops it immediately. Finished steps stay finished; nothing is left half-applied.
Every room, every advisor, every project. No token meter, no per-seat maths.