fix(setup-ai): harden AI bootstrap against real-world failure modes

[msitarzewski]

Follow-up to #12. That PR delivered the bootstrap script we needed; this hardens the paths where it breaks on a real network with a real .env.

Fixes

Interrupted model downloads were cached as success. The 1.5 GB model was written straight to its final path, so a dropped transfer left a truncated file that every later run accepted. Because server/lib/capabilities.js:58 detects the model with a bare existsSync, the UI would un-gate Transcribe over a corrupt model — the exact failure the v0.3.2 capability gating exists to prevent. Reproduced with a curl stub that dies mid-transfer:

RUN 1: curl: (18) transfer closed with bytes remaining to read
       -> models/ggml-medium.bin left on disk, 11 bytes
RUN 2 (network healthy): OK: Whisper model already present
                         OK: AI pipeline setup complete   # over an 11-byte model

Downloads now land in models/ggml-medium.bin.part, are size-checked, and only then renamed. Also dropped curl --silent — a multi-minute silent download reads as a hung script.

.env was sourced as shell. Under set -euo pipefail, both of these abort the script:

JWT_SECRET=k7$Rm2pQ9wZx        -> .env: line 1: Rm2pQ9wZx: unbound variable   (exit 1)
STATION_NAME=Pirate Radio One  -> .env: line 1: Radio: command not found      (exit 127)

A value containing ; or $(...) would execute. Our own .env carries JWT_SECRET, ICECAST_*_PASSWORD, and COTURN_PASSWORD, any of which can hold a $. Only the three LLM_* keys are read now, via sed.

whisper.cpp was a tracked gitlink that .gitignore could not touch. #12 added whisper.cpp/ to .gitignore while the path remained a 160000 entry with no .gitmodules backing — and git never ignores a tracked path:

git check-ignore -v whisper.cpp  -> exit 1 (NOT ignored)
git ls-files -s whisper.cpp      -> 160000 1fe009ca...

Meanwhile setup-ai.sh clones --depth 1 of the default branch, which will not equal the pinned SHA — so anyone who ran the script got a permanently dirty tree showing modified: whisper.cpp (new commits). Untracked it, which is what the setup script already assumed and what .gitignore was reaching for.

Missing ffmpeg aborted the whole bootstrap, though neither ffmpeg nor ffprobe is needed to clone whisper.cpp, build it, or fetch the model. Now warns and continues.

Also removed a dead already-built check that printed OK: whisper.cpp already built twice.

Testing

Kept the fake-toolchain harness from #12 — it is a good pattern — and added a regression case per bug. Both new cases were verified to fail against the pre-fix script, so they are testing something real:

vs pre-fix setup-ai.sh:  [2] FAIL: truncated model left on disk
                         [3] FAIL: setup-ai.sh choked on a realistic .env
vs this branch:          4 checks, all PASS

Full suite 18/18 (9 signaling + 9 rooms), server running as CI starts it.

tests/test-setup-ai.sh is now wired into ci.yml. It was only reachable via run-pre-validation.sh and an npm script, so it would have rotted.

Note

npm test alone reports 0/9 on a clean main — the tests need the signaling server already listening on 6736, which ci.yml starts but the npm script does not. Not addressed here, but the npm script is a misleading front door.

Thanks to @avezou for the original script — the CMake layout fix in it caught a live bug in the server auto-build path.

🤖 Generated with Claude Code