Ops Ar

/ops:ops-ar — A&R Command

Load ops-rules before acting. Public repo (no personal data). Outbound: one draft → one approval → one send. If AskUserQuestion / Workflow are missing, follow Rule 10 in ops-rules (Hermes: numbered options / two-turn Telegram card; delegate_task).

A&R the given record(s) like a pop/dance-hit label owner + master producer. The deliverable is always the full A&R card per track:

VERDICT (hit/10 + sign / develop / pass) → WHAT'S WORKING → WHAT'S HOLDING IT BACK → THE PLAN (producer moves) → REFERENCE & POSITIONING → NEXT.

Configuration (templatable — no hardcoded personal data)

The skill must read these at runtime — never hardcode user names, mailboxes, label names, or absolute /Users/... paths.

Taste profile injection (mandatory when ar.profile exists)

Before spawning any ar-producer agent, read the profile once:

bashCopy code
PREFS="${CLAUDE_PLUGIN_DATA_DIR:-$HOME/.claude/plugins/data/ops-ops-marketplace}/preferences.json"
jq '.ar.profile // empty' "$PREFS"

If non-empty, inject the whole JSON object verbatim into every ar-producer spawn prompt, with the instruction: "Judge as the A&R for THIS owner/label. Calibrate the verdict per verdict_calibration. Weigh hooks and toplines against songwriting_taste. REFERENCE & POSITIONING must position against catalog_recent, signature_classics, and label_roster_third_party — name the closest catalog comparison, not generic genre acts. Check tempo against tempo_sweet_spot and brand fit against lane. If signing_structure is present, the NEXT section of a sign/develop verdict must state which entity signs the master."

Fields the profile may carry (all optional): owner, label, lane, tempo_sweet_spot, ar_team, signing_structure, reference_acts, catalog_recent, signature_classics, label_roster_third_party, songwriting_taste, verdict_calibration. If the profile is absent, fall back to the generic dance-pop / feel-good house default and say so in the card header.

Imprint gate (mandatory when ar.imprints is non-empty)

Imprints are data, not code. Never hardcode an imprint's name, sound, or test — read them:

bashCopy code
jq -r '.ar.imprints // {} | keys[]' "$PREFS"          # imprint keys
jq '.ar.imprints["<key>"]' "$PREFS"                   # one imprint

For each imprint whose lane the track falls in, inject that object verbatim into the ar-producer spawn prompt except brand_book_corpus (the full scraped brand-book text — too large for a spawn prompt; consult it only when a card needs exact brand-book language, via jq '.ar.imprints["<key>"].brand_book_corpus.pages | keys' then the specific page). Always inject sound_description — it is the imprint's sonic north star and the tiebreaker on the subjective points.

Add an IMPRINT section to the A&R card between VERDICT and WHAT'S WORKING: score the track against that imprint's own ten_point_test per its scoring_rule (one line per point, pass/fail/N-A, with the failing evidence named), then one routing line — eligible for <imprint display_name>, or route to parent label. An imprint is a strict subset of the label: a failed gate never downgrades the main verdict, it only routes the release. Owner's own tracks always get the scorecard; third-party demos get it only when they are in that imprint's lane.

Each imprint object may carry: display_name, lane, sound_description, ten_point_test, scoring_rule, brand_book_corpus. If ar.imprints is absent or empty, skip the IMPRINT section entirely — do not invent an imprint.

Modes

1. Single track — /ops:ops-ar <file|url|latest>

Spawn the ar-producer agent (Opus) with the track:

• Local file → pass the path directly.
• URL/Dropbox → agent downloads first (Dropbox: append &dl=1; YouTube: yt-dlp -x --audio-format mp3).
• latest / empty → newest audio file (.mp3, .wav, .m4a, .aiff, .flac, .ogg, .aac) in ~/.claude/jobs/*/tmp/ by mtime — ignore JSON, PNG, and other non-audio artifacts.
• Subagent MCP rule: the spawn prompt MUST name the audio-ar tools and include the literal instruction ToolSearch select:mcp__audio-ar__full_ar_report,mcp__audio-ar__analyze_track,mcp__audio-ar__mood_score,mcp__audio-ar__transcribe_vocals,mcp__audio-ar__separate_stems,mcp__audio-ar__render_visuals,mcp__audio-ar__analyze_stems,mcp__audio-ar__cyanite_analyze,mcp__audio-ar__musicai_analyze,mcp__audio-ar__soundcharts_lookup — subagents don't inherit MCP discovery.
• Relay the agent's A&R card back verbatim.

2. Batch — /ops:ops-ar <file1> <file2> ... (or multiple URLs)

A&R multiple local paths or URLs in one invocation:

1. Collect inputs — every argument after the skill name is a track (local file or URL/Dropbox). Download URLs first (same rules as single-track).
2. Dedupe by md5 (same file copied under different names counts once).
3. Analyze per track — spawn ar-producer (Opus) per deduped track, in waves of ≤2 (stem separation is CPU-heavy; check nproc/uptime first). Same subagent MCP rule as single-track mode (include the full ToolSearch select:mcp__audio-ar__... list). Relay each agent's full A&R card back verbatim.
4. Summarize — compile a ranked verdict table from the per-track cards.

3. Inbox sweep — /ops:ops-ar inbox [from <sender> ...]

Pull every demo/song from the user's Gmail inbox and A&R them all:

1. Find demos: gog gmail search 'has:attachment (filename:mp3 OR filename:wav OR filename:m4a OR filename:aiff OR filename:flac OR filename:ogg OR filename:aac)' (add from: filters if senders given). Confirm scope with the user if the set is large (>10 threads).
2. Download attachments to disk without flooding context: per thread, gog gmail thread get <tid> -j → message ids from thread.messages[].id (envelope {downloaded, thread: {messages: [...]}} — NOT top-level messages) → gog gmail raw <mid> -j piped to jq for audio parts (filename + attachmentId) → gog gmail attachment <mid> <aid> --out <dir>/<label>__<file>. Also grep text parts for external links (postal.music, disco.ac, wetransfer, dropbox) — flag link-only demos that need a login as NOT ANALYZED and tell the user to request a file re-send.
3. Dedupe by md5 (forwarded demos repeat across threads).
4. Analyze per track — spawn ar-producer (Opus) per deduped demo, in waves of ≤2 (stem separation is CPU-heavy; check nproc/uptime first). Same subagent MCP rule as single-track mode (include the full ToolSearch select:mcp__audio-ar__... list). Relay each agent's full A&R card back verbatim.
5. Summarize — compile a ranked verdict table from the per-track cards.

4. Email delivery — "send the verdict to my email"

House rule: every A&R email ALWAYS includes (a) a per-track DIRECT LISTEN LINK and (b) the FULL A&R card per track — never just the ranked summary.

• DIRECT listen link (mandatory): a link that actually PLAYS the audio — any external streaming link found in the thread (postal.music, disco.ac, …), OR upload the demo to Google Drive (gog drive upload <file> → share → direct link). An email-thread link alone is NOT sufficient.
• Also include the Gmail deep-link (gog gmail url <threadId>) — both direct + email link is ideal.
• Rule 6 applies in full: stage the final draft, get explicit per-message approval, then send (one approval = one send).

Pro APIs (Cyanite / Music.ai / Soundcharts) — operational notes

• Cyanite ($AUDIO_AR_HOME/venv/bin/python pro_apis.py cyanite <file>): returns genreTags, subgenreTags, moodTags, instrumentTags, bpmRangeAdjusted, bpmPrediction{value confidence}, keyPrediction{value confidence}, timeSignature, energyLevel, valence, arousal, voicePresenceProfile, predominantVoiceGender, musicalEraTag, transformerCaption and per-mood scores (mood{happy uplifting energetic dark sad calm epic romantic}). Free/trial plans have a LIFETIME library cap — deleting tracks does NOT free quota. On librarySizeLimitExceededError, route through Music.ai instead.
  • Schema drift (2026-10): fileUploadRequest now returns { uploadUrl id } (the field was uploadId), then libraryTrackCreate(input: { uploadId: $id }) with $id: ID!. Older pro_apis.py copies that read uploadId, send a String id, or query removed fields fail validation — fix the stack, not the card.
  • Auth test: query { ping } answers without a valid key. Prove a key with libraryTracks(first: 1) { edges { node { id } } } and check that data.libraryTracks is an object.
  • API access needs a webhook URL on the Cyanite integration, even if you only poll. Point it at an HTTPS endpoint you control, and verify the Signature header (HMAC-SHA512 of the raw body with the webhook secret) before processing any payload; reject mismatches. A tiny Cloudflare Worker is enough.
  • Mood scores beat CLAP on dark/bright. When CLAP says "dark" but Cyanite mood.dark is low (< 0.2) and happy/uplifting are high, report the Cyanite reading.
• Music.ai (pro_apis.py musicai <file>, needs $MUSICAI_WORKFLOW, e.g. a "Metadata Suite" workflow): same Cyanite engine on separate billing + extras — ai_voice (Real vs AI-GENERATED — always flag AI guide vocals: they're placeholders needing a real singer), voice_gender, instruments. Implementation gotchas (already handled in pro_apis.py): requests need a browser User-Agent (Cloudflare 1010 blocks default python-urllib), and the upload-URL request must be a clean GET with no body. Convert .m4a to mp3 before upload.
• Soundcharts (pro_apis.py soundcharts <query>): released-catalogue lookup only — useless for unreleased demos; use it for reference-track benchmarking in the REFERENCE section. Auth is the x-app-id + x-api-key pair, and both must come from the same console app — a key from one app with another app's id fails auth and looks like an expired key. song/{uuid} gives audio.key/audio.mode/audio.tempo, handy for checking that a remix keeps the original's key.

Interpretation guardrails (carry into every card)

• Demo bounces are loud and dull on top — judge song/topline/lane, not the demo master.
• Never infer missing verses / song incompleteness from a sparse Whisper transcript (low vocal in the bounce ≠ unwritten song).
• librosa BPM can read doubled/halved — trust the pro-layer bpmRangeAdjusted (the prediction folded into 60–180 BPM, so it can differ from raw bpmPrediction.value, e.g. 200 vs 100) and use bpmPrediction.confidence to judge how sure it is, when available; otherwise confirm by groove.
• Verify hit-claims with data (CLAP commercial lean, valence/arousal), but the verdict is producer judgment, not a printout.

Fallback

If mcp__audio-ar__* is unavailable, run the stack directly via Bash from $AUDIO_AR_HOME (venv/bin/python analyze.py <file>, clap_score.py, transcribe.py, pro_apis.py). Never fabricate analysis — if nothing ran, say so.

Agent Teams support

If CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 is set, use Agent Teams when dispatching multiple ar-producer agents in batch or inbox-sweep mode. This enables:

• Agents share partial findings in real time (e.g., one agent finds a stem issue → others adjust their verdict framing accordingly)
• You can steer mid-sweep: "prioritize the demo from sender X first"
• Progress is visible per track as agents report back

Team setup (only when flag is enabled, batch/inbox dispatch phase):

Copy code
TeamCreate("ar-batch")
Agent(team_name="ar-batch", name="ar-[track-slug]", ...)

If the flag is NOT set, use standard parallel subagents (fire-and-forget, waves of ≤2).