docs/TASKS.md

TASKS.md — Midnight Static Execution Tracker

Deadline: June 15, 2026. Target submit: June 14 evening.

Status legend:

• [x] done
• [~] partially done / usable but not final
• [ ] not started
• [>] next / current focus
• [D] deferred or cut unless time allows

Current Position

We have a live Hugging Face Space with a working Gradio fallback, one instant showcase rerun, a schema-valid fixture writer, Kokoro TTS, and a crude mixer.

Live Space:

https://huggingface.co/spaces/build-small-hackathon/midnight-static

Current recommended focus:

1. Build the custom radio UI.
2. Post social update #1 after the UI is visually credible.
3. Run real Nemotron smoke on GPU infra.

Phase 1 — Foundation And Repo

1.1. [x] Initialize repo via Codex with project skeleton.

1.2. [x] Add AGENTS.md, docs, uv project metadata, tests.

1.3. [x] Set repo-local Git author to Codex <codex@openai.com>.

1.4. [x] Add .env protection and .env.example.

1.5. [x] Keep uv run pytest green before pushes.

Current evidence:

• First commit: 8c16939 Initialize Midnight Static scaffold
• Latest test count: 14 passing

Phase 2 — Script Schema And Writer Boundary

2.1. [x] Define shared script schema in midnight_static.schema.

2.2. [x] Add deterministic fixture writer.

2.3. [x] Add writer contract and --writer fixture pipeline option.

2.4. [x] Add five-premise writer smoke test.

2.5. [~] Add NemotronWriter implementation boundary.

Notes:

• NemotronWriter targets nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16.
• Local macOS BF16 smoke failed because the model requires mamba-ssm.
• Llama-Nemotron fallback avoided that import path but checkpoint fetch was interrupted after 3m47s.
• Real Nemotron smoke should move to GPU infra.

2.6. [>] Run real Nemotron smoke on deployment GPU path.

2.7. [ ] Add one retry on invalid model JSON, then fixture fallback.

Phase 3 — Voice And Audio Pipeline

3.1. [x] Verify Kokoro voice IDs from hexgrad/Kokoro-82M.

3.2. [x] Lock VoiceID enum.

3.3. [x] Render full fixture scene with multi-voice Kokoro.

3.4. [x] Add crude mixer: dialogue + hardcoded bed + SFX.

3.5. [x] One-command typed premise → mixed WAV:

uv run python -m midnight_static.pipeline --mode crude "a phone booth that rings under the sea"

3.6. [ ] Add production mixer chain:

• AM radio band-pass / saturation
• vinyl crackle
• ducking under dialogue
• LUFS normalization
• MP3 export

3.7. [ ] Add SFX retrieval/generation trace interface.

3.8. [ ] Add music bed/sting library interface.

Phase 4 — Live Space

4.1. [x] Add Gradio app entrypoint.

4.2. [x] Fix Space build packaging.

4.3. [x] Fix Space runtime launch.

4.4. [x] Publish Space under hackathon org.

4.5. [x] Add README Space metadata/frontmatter.

4.6. [x] Add model inventory table.

4.7. [x] Add Codex and Modal sections to README.

4.8. [x] Add three instant showcase reruns.

4.9. [x] Add showcase selector in Gradio UI.

4.10. [ ] Add README links for demo video, social post, blog, dataset, adapter, and traces as they become available.

Phase 5 — Showcase And Demo Quality

5.1. [x] Showcase 1: a phone booth that rings under the sea.

5.2. [x] Showcase 2: two rival lighthouse keepers, one light.

5.3. [x] Showcase 3: a wedding where every guest recognizes the wrong bride.

5.4. [x] Add showcase selector in Gradio UI.

5.5. [x] Manual listen test for all showcase broadcasts.

• First listen found the three reruns too similar.
• Fixed by adding hand-authored showcase scripts with distinct genres, casts, voices, and scene structures.

5.6. [ ] Tune levels if showcases are harsh, too quiet, or unclear.

Never cut:

• showcase broadcasts
• listen test
• README tags/model table
• demo video

Phase 6 — Custom Radio UI

6.1. [x] Build static radio visual shell.

6.2. [x] Genre dial interaction.

6.3. [x] Premise input as telegram slip.

6.4. [x] ON AIR loading/state sequence.

6.5. [>] Playback with script display.

6.6. [ ] Keep Gradio fallback available.

6.7. [D] Call-in microphone / ASR mode.

6.8. [D] FLUX poster generation.

Notes:

• Off-Brand score depends heavily on this phase.
• Do not block the live Space on this phase; the plain app is already live.

Phase 7 — Modal, Dataset, And Fine-Tune

7.1. [ ] Add modal/gen_dataset.py.

7.2. [ ] Generate 400 validated examples.

7.3. [ ] Human skim pass, 45 minute cap.

7.4. [ ] Publish dataset to HF Hub.

7.5. [ ] Add modal/finetune.py.

7.6. [ ] Train LoRA and publish adapter.

7.7. [ ] Evaluate base vs LoRA.

7.8. [ ] Add modal/batch_audio.py.

7.9. [ ] Generate SFX/music library and manifest.

Notes:

• This phase needs Modal credentials.
• This phase is high prize evidence but not required for the app to be live.

Phase 8 — Social, Blog, And Submission

8.1. [>] Social post #1: build-in-public with live Space link.

8.2. [ ] Add social post link to README.

8.3. [ ] Social post #2: radio UI reveal.

8.4. [ ] Demo video, 60-90 seconds.

8.5. [ ] Add demo video link to README.

8.6. [ ] Blog / Field Notes post.

8.7. [ ] Publish traces dataset.

8.8. [ ] Run docs/SUBMISSION.md checklist.

8.9. [ ] Final fresh-browser test on desktop and mobile.

8.10. [ ] Submit.

Deferred / Cut Unless Ahead

9.1. [D] VoxCPM2 G1 gate.

Reason: Kokoro is working, and OpenBMB/MiniCPM evidence is not core to the current path.

9.2. [D] BillShield-lite side project.

Reason: Midnight Static still needs showcase, UI, social, and submission work.

9.3. [D] FLUX posters.

Reason: only after custom radio UI and demo basics are done.

Standing Rules

• Commit small via Codex throughout.
• Keep uv run pytest green before every push.
• If a task takes more than 2x estimate, cut scope or downgrade it.
• Never commit .env or tokens.
• Prefer live, audible, demonstrable progress over deeper model work that is invisible to judges.