interview_helper_mux

Turn a long-form interview recording into a mastered podcast (master/master.wav), with an optional The War Room RSS publish path (private S3 + CloudFront).

Details: NORTH_STAR.md · SETUP.md · AGENTS.md · docs/cross-cutting/podcast-rss-hosting.md · terraform/README.md

Pipeline: 72 stages (35 analysis + 37 delivery) in src/interview_mux/v2/config.py.

---

Fresh environment (once per machine)

Requires: macOS Apple Silicon recommended (local STT / image), Python 3.12, ffmpeg, Node (for GUI), Terraform ~> 1.14.7 if you will publish RSS.

Windows / Linux + NVIDIA: MLX is Apple-only. Use ./scripts/bootstrap_venv_windows.sh, which builds the CUDA equivalents (faster-whisper, transformers+bitsandbytes, DeepFilterNet wheel) behind the same pipeline contracts — see docs/cross-cutting/windows-cuda-setup.md.

# 1) Clone / enter repo
cd interview_helper_mux

# 2) Bootstrap core .venv + local MLX stacks (speech, LLM, image) + GUI
./scripts/bootstrap_venv.sh

# 3) Secrets
mkdir -p config/secrets ASSETS/input
cp config/templates/secrets.env.example config/secrets/secrets.env
# Edit config/secrets/secrets.env — set at least OPENAI_API_KEY

# 4) Put source interview audio in ASSETS/input/
#    e.g. ASSETS/input/interview.wav

Optional bootstrap flags: BOOTSTRAP_SKIP_GUI=1, BOOTSTRAP_SKIP_VERIFY=1

Episode covers use OpenAI Images (podcast.cover_image) — see docs/cross-cutting/podcast-cover-theme.md.

---

Daily launch

./scripts/run.sh
# open http://127.0.0.1:8765

Runs and artifacts: ASSETS/executions/exec_*

Other run flags: MUX_PRESERVE_SESSION=1, MUX_REFRESH_DEPS=1, MUX_SKIP_ASSETS_CLEANUP=1, MUX_NO_BROWSER=1, MUX_RUN_ID=…, MUX_FRESH=0

---

The War Room RSS (optional — Terraform)

Infra lives under terraform/ with committed local state (terraform/README.md). Bucket name and show/layout settings live in config/app.defaults.json → podcast. Credentials + CloudFront/feed URLs live in config/secrets/secrets.env. Default resource base: the_war_room_001.

A — Create podcast hosting (S3 + CloudFront OAC)

# secrets.env must include AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY (or AWS_PROFILE)
cp config/terraform.tfvars.example config/terraform.tfvars   # optional overrides
./scripts/tf-init.sh
./scripts/tf-plan.sh
./scripts/tf-apply.sh   # also upserts PODCAST_* into secrets.env

Session restore (if needed): ./scripts/tf-plan.sh --use-session or USE_LATEST_SESSION=1 ./scripts/tf-plan.sh.

B — Seed empty feed + show artwork

python scripts/seed_podcast_origin.py

C — Run app and publish

./scripts/run.sh

Finish a pipeline → G-Publish → Prepare package for this run (local meta/cover/mp3 only). When ready, Upload this run to S3 (or python scripts/sync_podcast_episodes.py --execution-id …). Sync uploads only that execution, is additive, and never deletes remote objects.

Manual CloudFront recovery (boto3, no AWS CLI): ./scripts/invalidate_podcast_cf.sh (uses the live PODCAST_FEED_BASE_URL after apply or ./scripts/tf-rotate-cloudfront-url.sh).

D — Submit the feed

Open PODCAST_FEED_BASE_URL/feed.xml and submit via Apple’s pass-through:

https://podcastsconnect.apple.com/my-podcasts/new-feed?submitfeed=<that feed URL>

Seed / rotate / apply scripts print the encoded pass-through whenever they display a new RSS URL. Notice: Apple rejects an empty seed feed — publish ≥1 episode or a trailer first. Also add the same feed URL in Spotify for Podcasters.

Full ops: docs/cross-cutting/podcast-rss-hosting.md

AWS note: Terraform (scripts/tf-*.sh) owns the stack and updates terraform/state/. App uploads use boto3 + secrets.env. Operators never need aws login or AWS CLI.

---

Layout

scripts/bootstrap_venv.sh          # fresh env
scripts/tf-*.sh                    # Terraform wrappers (podcast stack; updates terraform/state/)
scripts/sync_podcast_tf_secrets.sh # outputs → PODCAST_* in secrets.env
scripts/seed_podcast_origin.py     # seed feed.xml + show art (boto3)
scripts/sync_podcast_episodes.py   # upload one execution (or --all for explicit bulk)
scripts/tf-rotate-cloudfront-url.sh # same S3 bucket, new CloudFront RSS URL
scripts/invalidate_podcast_cf.sh   # CloudFront invalidation of the live feed URL
scripts/run.sh                     # launch (MUX_FULL_AUTO=1 for detached Full-auto)
terraform/                         # S3 + CloudFront OAC (README + committed state/)
src/interview_mux/                 # pipeline + API
frontend/                          # React GUI
config/                            # defaults + secrets + podcast cover + tfvars.example
docs/                              # specs and prompts
tools/                             # CLI helpers + Full-auto launchers