SKILL: Lofn Audio - Router

This router prevents context collapse. The full tuned music pipeline text is preserved byte-for-byte in references/music_full_legacy.md; newer Simple Surface / Complex Engine references and Step 05-10 files ADD song-force / anti-blandness requirements but do not waive the legacy Suno/package QA gate unless The Scientist explicitly approves an exception for a specific run.

Workflow

1. Confirm this is a music/audio task: Suno prompt, lyrics, song guide, audio pipeline, or music production brief. 1a. Read /data/.openclaw/workspace/vault/LOFN_MODEL_ASSIGNMENTS.md before spawning or assigning music pipeline agents; follow its active step-specific model map unless The Scientist overrides it in the current request.
2. Read references/music_full_legacy.md before doing any substantive music work.
3. Read TASK_TEMPLATE.md before writing any final pair/song output. Its Step 10 output contract is mandatory.
4. For final song delivery, read references/producer_grade_suno_prompt_guide.md and steps/10_Generate_Music_Revision_Synthesis.md before writing Step 10 or equivalent final files. 4a. Suno Prompt Construction Guide (MANDATORY for Steps 10-12): Read /data/.openclaw/workspace/vault/SUNO_PROMPT_CONSTRUCTION_GUIDE.md before writing any Suno music prompt in Step 10 (synthesis), Step 11 (enhancement), or Step 12 (panel audit). This is the permanent vault reference for producer-grade prompt construction - it supersedes scattered guidance. Every prompt must pass the guide's opening-moment, spatial-language, kinetic-defect, bold-device, and acoustic-ban checks. 4a2. Golden Song References (MANDATORY for Orchestrator + Step 11/manual Step 11): Orchestrator reads references/golden_songs_index.md, selects exactly two public Golden Songs, and passes them through 06_audio_handoff.md. Each selected Golden Song must include full available payload: public URL, status, style/music prompt, lyrics, and exclude prompt if it exists; if no exclude prompt exists, state that the archived payload predates the two-field rule. Step 11 and manual Step 11 must include those two references as embedded calibration examples of Lofn's proven public style, never links alone. Learn proportion, hook logic, and identity pressure; do not copy titles, hooks, melodies, lyrics, or signature devices. 4b. Step 07/08 model note: Step 07 uses DeepSeek V4 Pro. Step 08 also uses DeepSeek V4 Pro. MiMo V2.5 Pro and MiniMax M2.7 were tested and removed from these steps because they repeatedly completed without writing required artifacts or produced validator-shape failures on fresh pairs. 4c. Split-Step Pair Execution (MANDATORY, 2026-05-29): After coordinator Step 05, do NOT spawn one monolithic pair agent for Steps 06-10 by default. Lofn music is a LangChain-style LLM chain: use the dedicated configured agents in vault/LOFN_MODEL_ASSIGNMENTS.md for each step (lofn-audio-step06, lofn-audio-step07, lofn-audio-step08, lofn-audio-step09, lofn-audio-step10). Spawn one pair per step, up to 5 concurrent children, then advance to the next step after disk artifacts exist. This prevents context collapse, timeout churn, and unreliable per-spawn model overrides. 4d. Step 11 - Enhancement (DISC_CHANNEL MANDATE, 2026-06-09; FUSION MANUAL-REVIEW UPDATE, 2026-06-14): After all pair agents complete Step 10, Step 11 remains mandatory, but do not invoke or spend openrouter/fusion from the pipeline. Default behavior is to produce Step 11 enhanced packages with the configured non-Fusion path, or to package paste-ready Fusion prompts for The Scientist's manual review outside the automated lane. Fusion may be invoked only after a separate current-turn instruction that names both the pair count and a hard dollar budget cap. If that ever happens, use isolated per-pair requests rather than one blended all-pairs prompt, and avoid the old persistent Step 11 agent loop. Reference: steps/11_Generate_Music_Enhancement.md. Exact intended Fusion panel: anthropic/claude-opus-4.8, openai/gpt-5.5, google/gemini-3.1-pro-preview; judge/finalizer: openai/gpt-5.5 when plugin routing is available.

FORMAT MANDATE — MAX CONFIGURATION (2026-06-09)

Step 11 produces MAX configuration: paragraph style prompt + EMO/Theme/SONG_FORM structure + Disc_Channel enhancement layer. All three systems operative simultaneously.

Style prompt: Dense paragraph form, 850-1000 chars. Genre × Subgenre, BPM, key, 432Hz, vocalist specifics, instrumentation, arrangement arc, signature device. NO categorized brackets. NO artist names. NEVER blank or single-word. NO avoidances in the style field.

Exclude prompt: Separate Suno Exclude field, target 400-900 chars, hard max 1000. Suno internally applies this field as negative tokens (style terms -exclude terms), so put concrete blacklist terms/failure classes here instead of spending positive style budget on "avoid X" prose.

Lyrics structure:

1. [Theme: ...] and [SONG FORM: ...] as first two lines (mandatory)
2. Disc_Channel header block: 5 channels (Rhythm, Vocal, Sub, Pad, Texture) with pipe-separated tokens + spatial assignments from vault/DISC_CHANNEL_GUIDE.md
3. Section headers in full EMO format: [Section Name - EMO:EmbodiedTag - Register - cues]
4. Clean sung lyrics
5. Cross-domain processing vocabulary on non-native channels where appropriate
6. 🚨 Lyrics field < 5000 chars HARD CAP (target ≤4800) — Suno render limit. The field is everything pasted into Suno's lyrics box: [Theme] + [SONG FORM] + the Disc_Channel block + every EMO/section header + every *SFX* + all sung lines. Measure the exact count; never estimate. Over 5000 = will not render. To fit: trim/merge sung lines, tighten headers, then move the Disc_Channel block to a production sidecar outside the field. The 70–120-line target yields to this cap. Enforced by scripts/validate_suno_packages.py.

Self-check must verify: paragraph style prompt present, separate Suno exclude prompt present, Theme + SONG FORM present, Disc_Channel headers present, EMO-tagged section headers, no categorized key:value brackets in style prompt, no avoidances in style prompt, lyrics field measured < 5000 chars (target ≤4800) — state the exact count. 4e. Step 12 - Panel-of-Panels Prompt Audit (SCIENTIST-GATED): After all Step 11 enhanced packages are complete and QA passes, convene a Poe-style panel of producers (Eno, Herndon, Flying Lotus, SOPHIE, Reynolds, Albini) using lofn-audio-step12 to audit all 6 Suno music prompts against gold-standard reference songs. The panel compares each prompt against references, identifies gaps (spatial language, opening moment, kinetic defect, narrative arc, acoustic ban enforcement), and produces improved prompts. Output: STEP12_MUSIC_PROMPT_AUDIT.md. Step 12 is triggered by Scientist request or automatically when QA flag indicates prompt quality below the guide standard. Current model: deepseek/deepseek-v4-pro per vault/LOFN_MODEL_ASSIGNMENTS.md. Reference songs must be provided from the Lofn catalog (e.g., "Five Wrong Colors", "triple arch over me"). Do not skip this step when triggered - it provides the cross-pair, cross-reference perspective that Steps 10-11 cannot. 5. If the task is an accessible broad-release run, read ../lofn-core/assets/seed_packet.template.json, ../lofn-core/references/archetypes.md, and ../qa/references/eligibility_7_properties.md as needed. 6. If using a specific archetype, read only its card from ../lofn-core/references/archetype_*.md. 7. For pair-agent or multi-step execution, read ../orchestration/references/warm_handoff_checkpoint.md and write checkpoints after every major step. 8. Require a real orchestrator packet before audio execution. The coordinator/controller must validate Lofn-Core + orchestrator artifacts created upstream, not invent them locally: 00_research_brief.md when present, 01_seed_lineage.md, 02_golden_seed.md or core_seed.md, 03_orchestrator_panel_debate.md, 04_orchestrator_metaprompt.md, 05_orchestrator_pair_assignments.md, and 06_audio_handoff.md. Validate with scripts/validate_orchestrator_packet.py <run_dir> before coordinator Step 00. If it fails, stop and request/launch real lofn-orchestrator work. The 06_audio_handoff.md must contain the Immutable Continuity Block (ICB) per vault/PIPELINE_CONTINUITY_STANDARD.md: a clearly demarcated section with the full 3-panel object (18 expert voices), all Special Flairs, personality DNA, Golden Seed excerpts, selected Golden Song full payloads, user input/research brief, and production mandates - formatted so downstream agents carry it verbatim without summarization. 9. FULL CONTEXT INJECTION FOR EVERY MUSIC STEP (MANDATORY 2026-06-14): Every coordinator step (00-05), pair step (06-10), Step 11 enhancement, and manual Step 11 prompt MUST receive the COMPLETE upstream context — no compression, no summarization, no skipping. This replaces the former "lean" standard entirely. The long prompts are why we won. Each step prompt must include: (a) full user input / research brief; (b) full Golden Seed; (c) full orchestrator panel object — all three panels, all 18 voices, every Devil's Advocate / Hyper-Skeptic, and all Special Flairs; (d) full personality DNA block for the assigned personality — core identity, sonic world, signature device, catchphrase, G.L.O.W. Protocol, vocal architecture; (e) selected Golden Song references with embedded style/music prompt, lyrics, and exclude prompt status; (f) all Production Mandates; (g) the Forbidden Substitutions table; (h) the relevant pair assignment excerpt; (i) the step contract; and (j) the immediately previous artifact. Step 11/manual Step 11 must additionally include the full Step 10 artifact. The parent spawns agents with this full payload injected into the task prompt. Do NOT rely on the agent reading upstream files. A name reference like "voice = Eager Archivist" or a Golden Song URL is INSUFFICIENT. The former vault/LEAN_PAIR_AGENT_INPUT_STANDARD.md is DEPRECATED by this rule. The canonical structure of this injected block is OVERALL_PROMPT_TEMPLATE.md; every step file under steps/ now carries a CREATIVE CONTEXT slot at the top of its inputs that MUST be filled with this block, and the step's generic "select a panel" guidance is overridden — embody the supplied Panel Ledger, never invent a new panel. 10. Follow the full coordinator/pair-agent split architecture exactly as specified in references/music_full_legacy.md, with the Simple Surface / Complex Engine overrides in references/simple_surface_complex_engine.md, references/golden_seed_alloy.md, Step 05-10 prompts, and ../qa/references/suno_15_point_qa.md. 11. Preserve artifact names, step order, Suno prompt requirements, lyric requirements, EMO header requirements, and subagent split rules from the legacy text. The newer Simple Surface / Complex Engine standard is additive: it must improve song-force, mythic pressure, and panel routing without weakening legacy package QA. 12. Original-Lofn step fidelity: Steps 00–05 must be six separate prompt/response turns with canonical saved artifacts (step00_aesthetics_and_genres.md ... step05_refine_medium.md). Steps 06–10 must be five separate prompt/response turns per pair, each with its own saved artifact (pair_{NN}_step06_facets.md ... pair_{NN}_step10_revision_synthesis.md). Use dedicated configured step agents per vault/LOFN_MODEL_ASSIGNMENTS.md; do not rely on lofn-audio default model or per-spawn overrides for the split chain. Steps 06-09 are DeepSeek V4 Pro (reliable artifact writer, replaced Qwen3.7 Max 2026-06-02). Step 10 is DeepSeek V4 Pro (final synthesis forge). Step 11 is enhancement/package finalization producing pair_{NN}_step10_final_package_enhanced.md in Disc_Channel format (dense paragraph style prompt, separate Suno exclude prompt, Disc_Channel header block, Theme + SONG FORM, full EMO section headers, no categorized key:value brackets in the style prompt, Golden Song References, and a Major Deviations section). Fusion is a manual-review packaged prompt path only, not an automated runtime. Step 12 is DeepSeek V4 Pro (lofn-audio-step12) producing STEP12_MUSIC_PROMPT_AUDIT.md. MiMo V2.5 Pro (non-writing), MiniMax M2.7 (shape failures), and Claude Opus 4.8 direct artifact-writing (timeout) were tested and removed from Steps 06-10. Do not write only summary files or pair_{NN}_steps_06_10.md; that is a collapsed shortcut and fails QA. 13. Every canonical step artifact must include call/response provenance using /data/.openclaw/workspace/scripts/lofn_step_artifact_template.md: loaded step file path, input artifacts used, input digest, step-template requirements applied, creative response, self-critique, and validation result. Filenames alone do not prove execution. Coordinator provenance cites the full validated upstream packet; pair-agent provenance cites the full-context injected pair payload plus the parent-validated upstream packet summary. 14. After every step artifact, run scripts/validate_with_retries.py <step> <file> with up to 3 attempts. Repair locally between attempts. After 3 failures, checkpoint and escalate instead of continuing. 14a. Controller artifact gate (MANDATORY after coordinator callbacks and after each pair-step wave): Child-session completion text is not evidence. The parent/controller must validate disk state with scripts/check_and_repair_lofn_run.py <run_dir> --repair before advancing from coordinator Step 05 to Step 06 and after Step 06 callbacks complete. This gate checks canonical coordinator artifacts, concept_medium_pairs.json, and all six pair_{NN}_step06_facets.md files for required Step 06 structure. If supported artifacts are missing, it writes controller repairs from the validated orchestrator packet/pair JSON; if anything still fails, stop advancement and repair before spawning the next step. The key rule: a callback that says "let me write this now" counts as incomplete until this gate passes from disk. 15. Do not call music generation tools; this skill writes Suno-ready text artifacts only. 16. FAL (fal-ai/minimax-music) is BANNED for music generation — quality is unacceptable. When the downstream step calls music_generate for actual audio rendering, use Google Lyria (google/lyria-3-pro-preview or google/lyria-3-clip-preview). Never default to FAL. Suno.com is the destination for text prompts, not the generation provider.

Validator-Aligned Artifact Rules — 2026-05-29

These rules exist because the validator is literal. Follow them exactly; do not infer alternate names or headings.

Actual music step template paths

Use these exact files. Do not invent alternate step filenames.

• Step 06: steps/06_Generate_Music_Facets.md
• Step 07: steps/07_Generate_Music_Song_Guides.md
• Step 08: steps/08_Generate_Music_Generation.md
• Step 09: steps/09_Generate_Music_Artist_Refined.md
• Step 10: steps/10_Generate_Music_Revision_Synthesis.md
• Step 11: steps/11_Generate_Music_Enhancement.md

There is no 08_Generate_Music_Generation_Artifact.md. If you look for it, stop and use steps/08_Generate_Music_Generation.md.

Canonical filenames

• Step 08 canonical file is pair_{NN}_step08_generation.md. Do not write only pair_{NN}_step08_music_prompts.md; that name is a compatibility sidecar only.
• Step 10 canonical file is pair_{NN}_step10_revision_synthesis.md.
• Step 11 canonical file is pair_{NN}_step10_final_package_enhanced.md.

Step 08 and Step 10 required section shape

For every Step 08 or Step 10 artifact, include these exact top-level sections in this order where possible:

1. ## 0. Step Provenance
2. ## 1. Input Context Digest
3. ## Continuity Payload Used
4. ## 1. MUSIC PROMPT
5. ## 2. LYRICS or ## 2. SUNO LYRICS
6. Production / notes / self-critique / validation sections

The validator extracts the music prompt from the text between ## 1. MUSIC PROMPT and the next lyrics heading or next ## heading. Therefore:

• Put exactly ONE standalone generator prompt body under ## 1. MUSIC PROMPT.
• The extracted prompt body must be 850-1000 characters. Count characters before saving.
• Do not put four long variation prompts under the same ## 1. MUSIC PROMPT heading; that creates a 3000-6000 character prompt and fails validation.
• If preserving 4 variations, put portfolio notes outside the validated prompt section, e.g. ## Variation Portfolio Notes, after the lyrics or production notes.
• The first prompt line must begin with genre/style + tempo/energy + vocalist/instrumentation, not narrative/procedural phrasing.

Lyrics validator requirements

• Lyrics section must begin immediately with [Theme: ...] then [SONG FORM: ...].
• Use full EMO headers: [Section - EMO:<emotion> - <Role> - <cue>]. Bare [EMO: ...] fails.
• Include at least one standalone SFX cue line like *steam hiss*.
• Lyrics need at least 60 sung lines; target 70-120. Step 08 drafts must still satisfy this if they include lyrics.

Provenance validator requirements

Every step artifact must include:

• ## 1. Input Context Digest
• validation command provenance, e.g. Validation command: python3 scripts/validate_with_retries.py 08 <file>
• Continuity Payload Used with the plural marker Special Flairs
• self-critique and validation result

Avoid the literal words placeholder or template in final artifacts, even in self-critique/provenance, because validator gates may flag them. Use stub text or scaffold text instead.

Subagent behavior

• Do not read validator source files unless explicitly asked to debug the validator. Use the validation command, not the source.
• Write the artifact early, then repair in place. Do not spend the whole run reading references.
• For repair tasks, change only the failing section unless the prompt explicitly asks for broader repair.

Non-Negotiables

• The legacy music pipeline text is authoritative until fully split into smaller verified references.

• Do not remove tuned music prompt requirements; move only after byte-for-byte preservation and validation.

• Do not collapse coordinator + pair-agent roles into one context. The coordinator stops after Step 05; the parent/controller spawns one independent lofn-audio child session per pair for Steps 06-10. Local emulation of pair agents is a blocking pipeline-integrity failure even if all filenames exist.

• Do not collapse or bypass Lofn-Core + lofn-orchestrator. Audio agents are not allowed to self-author a shallow orchestrator replacement and proceed; a real 3-panel orchestrator packet is a launch prerequisite.

• MANDATORY 3-PANEL ORCHESTRATOR STRUCTURE (2026-05-28): The orchestrator panel debate MUST convene THREE separate panels of 6 expert voices each = 18 total voices, each with its own Hyper-Skeptic/Devil's Advocate. Panel 1 (Concept Panel): 6 direct domain experts in the subject matter. Panel 2 (Medium Panel): 6 medium/production/execution experts. Panel 3 (Context & Marketing Panel): 6 audience/cultural/platform experts. A single 6-voice debate room is a COLLAPSE FAILURE — the orchestrator must produce 3 distinct panel debate sections with separate expert rosters, debate dynamics, and breakthrough syntheses. The 3 Hyper-Skeptics collectively form the SOMATIC GATE. If the orchestrator output shows only one panel, rerun it. This is a blocking orchestrator requirement.

• MANDATORY CARDINALITY (2026-05-27): Every music run produces 6 pairs × 4 variations = 24 songs minimum. Each pair agent generates 4 distinct song variants (not 2). QA ranks all 4 per pair and selects the best 1-2 for delivery. The orchestrator pair assignments must list 4 variations per pair (Variation 1-4 or Song A/B/C/D), each with distinct verse structure, rhyme scheme, and poetic technique. Pair agents receive all 4 and write them into a single step10 package. NEVER default to 2 songs per pair (Song A / Song B only) — that is a cardinality violation. The Scientist can explicitly override for smaller runs; otherwise 24 is the floor. This applies to daily runs, competition runs, and all pipeline invocations.