[avezou]

setup-ai.sh:

• build whisper-cli with the current Cmake-based layout
• download models/ggml-medium.bin
• prompt for LLM provider settings and write them to .env
• update docs to reflect changes

Summary

Adds a one-shot setup-ai.sh bootstrap for the AI pipeline so users can setup whisper.cpp, download the default mode, and configure LLM provider from one guided flow.

• clone/recover whisper.cpp and build whisper-cli
• download models/ggml-medium.bin
• prompt for LLM provider settins and write them to .env

Changes

• setup-ai.sh  — (NEW) adds the one-shot AI bootstrap flow to clone/build  whisper.cpp , download the model, and write LLM settings into  .env .
•  server/lib/whisper-transcriber.js  — aligns the server’s auto-build path with the current  whisper.cpp  CMake layout so transcription can build the CLI correctly.
•  tests/test-setup-ai.sh  — adds a smoke test that runs the setup script against a fake toolchain and verifies the expected outputs.
•  run-pre-validation.sh  — includes the new setup script test in the repo’s validation flow and runs shell tests with bash.
•  package.json  — adds an  npm run test:setup-ai  shortcut for the new smoke test.
•  README.md  — updates the AI setup docs to point users at  ./setup-ai.sh  and notes the required  cmake  dependency.
•  QUICKSTART.md  — mentions the optional AI bootstrap step in first-time setup.
•  web/js/main.js  — points capability-gating hints at  ./setup-ai.sh  instead of manual whisper.cpp/LLM steps.
•  .gitignore  — ignores the local  whisper.cpp/  checkout and downloaded model files.
•  CHANGELOG.md  — records the new AI bootstrap script in the unreleased notes.

Test Plan

• Server tests pass (npm test)
• Manual test: start session, verify audio
• Tested in Chrome (chromium on Archlinux)
• Tested in Firefox

Screenshots

[msitarzewski]

Merged — thank you, and I'm sorry this sat for four weeks. That's on me, not on the PR.

This is a genuinely useful contribution. setup-ai.sh was sitting near the top of my roadmap notes as the next thing to build, and you showed up and built it. A few things I particularly appreciated reading through it:

• The whisper-transcriber.js fix is the real win. The old auto-build path was still assuming the pre-CMake layout, so it had quietly stopped working — you caught a live bug while adding a feature.
• tests/test-setup-ai.sh running the script against a fake toolchain is exactly the right way to test a bootstrap script. That pattern is going to get reused.
• Wiring the capability-gating hints in main.js to point at ./setup-ai.sh closes the loop on the gating work from v0.3.2 nicely. Good instinct.

Heads-up so it doesn't look like I'm silently rewriting your work: I'm pushing some follow-up hardening on top, mostly around failure modes that only show up on real networks and real .env files:

• Model download goes to a temp path and gets renamed on success — right now an interrupted 1.5 GB download leaves a truncated file that later runs accept as valid
• Reading just the LLM_* keys out of .env instead of sourcing it, since real secrets with $ or spaces in them break under set -u
• Sorting out the .gitignore / gitlink situation, which is a pre-existing mess in this repo that your PR happened to land in the middle of

None of that detracts from the contribution — it's the kind of polish that comes from running a thing against a hostile network a few times. Thanks again for the patience and for the work.