eddy-v2

name	eddy-v2
description	Run Eddy V2 proof-gated YouTube edits from raw footage folders. Use for raw creator footage to long video, Shorts, launch kit, scorecard, and receipts.

Eddy V2 Skill

This is the shared Codex and Claude skill for Eddy V2. Use MCP tools first when available. Fall back to the CLI.

Default workflow

1. Run eddy doctor.
2. Run eddy edit <folder> for autonomous editing.
3. Read <folder>/eddy-runs/<run>/scorecard.md.
4. If status is blocked, report the exact blocker and receipts path.
5. If Studio Sound credentials become available after a run, use eddy audio-proof <run> to retry cloud audio proof from the existing extracted WAV instead of rerunning the full edit.
6. If Lennox gives 8/10 quality scores, record them with eddy review <run> instead of editing the scorecard by hand.

For MCP host-agent repair loops, call eddy_v2_scorecard with json: true or format: "json" when you need the same structured payload as eddy scorecard <run> --json; omit those options for the human-facing Markdown scorecard.

Motion proof lives under each run's motion/ projects. Expect frame.md, storyboard.md, storyboard.html, motion-plan.json, motion-collision-proof.json, motion-lint.json, and motion-inspect.json before treating overlays or Shorts cards as compositor-safe.

Host-agent intent path

If you have enough context to provide or repair the edit intent yourself, pass a JSON object with target_duration_s, identity, shorts_target, hook, and title:

• MCP: call eddy_v2_edit_start with intent or intent_json.
• CLI fallback: write the object to a local file and run eddy edit <folder> --intent <intent.json>.

Eddy validates the host intent against frozen identities, source duration, and the default YouTube floor, records host_intent receipts, and skips OpenRouter for that run.

Review path

eddy review <run> --long-edit <score> --motion <score> --audio <score> --shorts <score> records the human bakeoff verdict. The command keeps publishable_8_of_10 false when any score is below 8 or when machine/audio quality blockers such as strong_studio_sound_not_proven remain.

Audio proof retry

eddy audio-proof <run> verifies the run's source hashes, reuses audio/source-audio.wav, retries the configured cloud audio providers under the current cost cap, and remuxes final/video.mp4 only when cloud audio passes parity. Use --local-only when cloud egress is forbidden; it will refuse provider branches and keep the existing audio blockers.

When credentials live in 1Password, do not print or paste token values. First make sure op whoami is signed in, then run the retry through 1Password references:

op run --env-file .env.audio -- eddy audio-proof <run> --cloud-budget 25 --json

Prefer DESCRIPT_API_KEY for Strong Studio Sound. AUPHONIC_API_KEY plus AUPHONIC_PRESET / AUPHONIC_PRESET_UUID, or ELEVENLABS_API_KEY, can prove an approved cloud fallback but not Descript Studio Sound.

Hard boundaries

• Never edit, move, delete, or overwrite raw source files.
• No platform publish, full-video upload, or scheduling actions exist in Eddy V2.
• The only permitted egress is receipted, cost-capped model/audio quality work when cloud mode is allowed and credentials are configured.
• --local-only must be used whenever the user forbids cloud/model/audio/image egress.
• Do not call a run publishable until Lennox scores long edit, motion, audio, and Shorts at 8/10+.