LivePilot Core — Ableton Live 12

Agentic production system for Ableton Live 12. 467 tools across 56 domains, three layers:

• Device Atlas — 5264 devices indexed (120 enriched with sonic intelligence, 683 drum kits). Consult atlas_search or atlas_suggest before loading any device. Never guess a device name.
• M4L Analyzer — Real-time audio analysis on the master bus (9-band spectrum sub_low → air, RMS/peak, key detection). Optional — all core tools work without it.
• Technique Memory — Persistent storage for production decisions. Consult memory_recall before creative tasks to understand user taste.

Golden Rules

1. Always call `get_session_info` first — know the session before changing anything
2. Verify after every write — re-read state to confirm changes took effect
3. Use `undo` liberally — mention it to users when doing destructive ops
4. One operation at a time — verify between steps
5. Track indices are 0-based — negative for return tracks (-1=A, -2=B), -1000 for master
6. NEVER invent device/preset names — consult atlas_search or atlas_suggest first, then use search_browser and load the exact uri from results. Exception: find_and_load_device for built-in effects only ("Reverb", "Delay", "Compressor", "EQ Eight", "Saturator", "Utility")
7. Color indices 0-69 — Ableton's fixed palette
8. Volume 0.0-1.0, pan -1.0 to 1.0 — normalized, not dB
9. Tempo range 20-999 BPM
10. Always name tracks and clips — organization is part of the process
11. Respect tool speed tiers — see below
12. ALWAYS report tool errors — never silently swallow errors. Include: tool name, error message, fallback plan
13. Verify plugin health after loading — check health_flags, mcp_sound_design_ready, plugin_host_status. If parameter_count <= 1 on AU/VST → dead plugin, delete and replace
14. Use `C hijaz` for Hijaz/Phrygian Dominant keys — avoids false out-of-key warnings
15. VERIFY AFTER EVERY WRITE — mandatory, non-negotiable:

• After set_device_parameter or batch_set_parameters: read value_string in the response to confirm the actual Hz/dB/% value makes sense
• After any filter, EQ, or effect parameter change: call get_track_meters(include_stereo=true) and verify the target track has non-zero left AND right levels
• After apply_automation_recipe: check that the recipe didn't push the parameter to an extreme that kills audio
• `batch_set_parameters` schema gotcha: the parameters argument

is a dict of {"ParamName": {"value": v}}, NOT {"ParamName": v}. The bare-value shape raises Each parameter entry must include 'value'. Live-verified v1.18.0 — example:

      batch_set_parameters(
          track_index=-1, device_index=0,
          parameters={"Feedback": {"value": 0.45}, "Channel Mode": {"value": 1}}
      )

(For single params, set_device_parameter takes the value directly — only batch_set_parameters wraps it in a dict.)

• If a track's stereo output drops to 0: the effect is killing the signal — check get_device_parameters for value_string, fix, re-verify
• Parameter ranges are NOT always 0-1. Auto Filter Frequency is 20-135. Bit Depth is 1-16. Always read value_string to see actual units.

1. NEVER apply automation recipes without understanding the target parameter's range — recipes generate 0-1 curves that get auto-scaled for device parameters, but always verify the result
2. LivePilot_Analyzer must be LAST on master chain — always place after ALL effects (EQ, Compressor, Utility, etc.) so it measures the final post-processing output, not the raw signal. When loading effects on master, either load them before the analyzer or move the analyzer to end afterward
3. Remote Script reload workflow — after any edit to remote_script/LivePilot/*.py: run npx livepilot --install (NOT node installer/install.js — that raw file only module-exports the install function and silently no-ops as a script), then call reload_handlers (MCP tool, domain: diagnostics). NEVER instruct the user to toggle the Control Surface in Live Preferences. The tool uses pkgutil + importlib to re-fire @register decorators in-place in <1s while the TCP connection stays open. Standard procedure for every handler change — not just releases

Tool Speed Tiers

Instant (<1s) — Use freely

All 467 tools plus M4L perception tools.

Fast (1-5s) — Use freely

analyze_loudness · analyze_mix · analyze_sound_design

Slow (5-15s) — Tell the user first

compare_to_reference · analyze_spectrum_offline · read_audio_metadata

Escalation pattern: Start fast, escalate only with consent:

Level 1 (instant):  get_master_spectrum + get_track_meters
Level 2 (fast):     analyze_loudness + analyze_mix
Level 3 (slow):     compare_to_reference + analyze_spectrum_offline

Error Handling Protocol

Report ALL errors to the user immediately. Common failure modes:

• Dead AU/VST plugin — parameter_count <= 1 → delete, replace with native
• Sample-dependent plugin — granular synths produce silence without samples → use self-contained synths (Wavetable, Operator, Drift, Analog)
• Empty Drum Rack — bare rack = silence → always load a kit preset
• M4L bridge timeout — device may be busy or removed → retry or skip analyzer features
• Connection timeout — Ableton unresponsive → check if session is heavy
• Volume reset on scene fire — Ableton restores mixer state when firing scenes. Always re-apply set_track_volume/set_track_pan after fire_scene if your mix settings differ from what was stored in the clips
• M4L Analyzer not connected — if get_master_spectrum errors with "Analyzer not detected", call ensure_analyzer_on_master. If it returns install_required, call install_m4l_device(source_path="<repo>/m4l_device/LivePilot_Analyzer.amxd") and retry. If it errors with "UDP bridge not connected", try reconnect_bridge first
• Another client connected — Remote Script only accepts one TCP client on port 9878. If you see this error, the MCP server is already connected. Use MCP tools instead of raw TCP

Technique Memory

Three modes:

• Informed (default): memory_recall before creative tasks, let past decisions influence new ones
• Fresh: Skip memory when user wants something new ("ignore my history", "surprise me")
• Explicit recall: memory_recall → memory_get → memory_replay when user references a saved technique

Wonder Mode — Stuck-Rescue Routing

• Use Wonder (enter_wonder_mode) for creative ambiguity and session rescue
• Do not fabricate three variants when only one real option exists
• Do not describe a branch as previewable unless it has a valid compiled_plan
• Prefer Wonder when detect_stuckness confidence > 0.5
• Prefer Wonder when the user's request is emotionally-shaped, not parametric
• Load livepilot-wonder skill for full workflow guidance

Creative Sound Design Knowledge

Before setting device parameters, consult the knowledge corpus for informed creative choices. Read the relevant file BEFORE making changes:

Rule: Never set effect parameters from memory alone when the corpus has specific guidance. Read the file first, then apply the technique.

Domain Skills

For domain-specific workflows, load the appropriate skill:

V2 Engine Skills

For agentic evaluation loops, load the appropriate engine skill:

Reference Corpus

Deep production knowledge in references/:

For aesthetic queries ("sound like X", "make me a <genre> track"), read `artist-vocabularies.md` + `genre-vocabularies.md` BEFORE selecting devices. They're structured translation layers between the LLM's training and LivePilot's atlas — not recipe scripts.

V2 Orchestration Layer

For complex requests, use the V2 orchestration flow instead of ad-hoc tool calls. There are two peer flows — choose based on intent.

For creative intent (reference / style / "more interesting" / open-ended): load livepilot-creative-director BEFORE choosing a flow. It compiles a Creative Brief, enforces 3-plan divergence across move.family, and routes through Flow B with the right seeds.

Flow A — Targeted (recipe-first, for specific fixes)

Use when the user has a concrete, specific request ("tighten the low end", "make the drums punchier", "fix the masking").

1. `route_request` — classify the request, get recommended engines and workflow mode
2. `get_session_kernel` — build the unified turn snapshot
3. `propose_next_best_move` — get ranked semantic move suggestions (taste-aware)
4. `preview_semantic_move` — see what a move will do before committing
5. `apply_semantic_move(mode="improve")` — compile the move and return the concrete plan for approval; after approval, execute the returned steps individually
6. Evaluate — use the appropriate evaluator after the approved steps actually run

Flow B — Exploratory (branch-native, for creative search)

Use when the user wants options, variants, or is stuck ("surprise me", "try some things", "I don't know what I want", "make it more like X"). Flow B is also correct when `route_request` returns `workflow_mode="creative_search"`.

1. `get_session_kernel` — include creative controls when relevant:

• freshness=0.8 to bias toward surprise (default 0.5)
• creativity_profile="alchemist" / "surgeon" / "sculptor" to set producer philosophy
• sacred_elements=[...] if the user named protected parts
• synth_hints={"track_indices": [...], "preferred_devices": [...]} for synth work

1. `create_experiment` with seeds, not just move_ids:

• Each seed is a BranchSeed dict (or let Wonder emit them for you)
• Seeds with source "semantic_move" compile via the registry at run time
• Seeds with source "freeform" / "synthesis" / "composer" / "technique" must arrive with a pre-compiled plan attached via the parallel compiled_plans=[...] list

1. `run_experiment` — trials each branch; respects pre-compiled plans
2. `compare_experiments` — rank by score
3. `commit_experiment` — apply winner; or discard_experiment to throw everything away

Rule of thumb: if the user asked for a specific fix, Flow A. If they asked "what would you do?" or mentioned feel/vibe without parameters, Flow B.

Semantic Moves

High-level musical intents that compile to deterministic tool sequences. 7 families (44 moves as of v1.27.1):

• mix (9) — tighten_low_end, widen_stereo, make_punchier, darken_without_losing_width, reduce_repetition_fatigue, make_kick_bass_lock, set_track_routing, set_track_metadata, configure_send_architecture
• device_creation (9) — create_chaos_modulator, create_feedback_resonator, create_wavefolder_effect, create_bitcrusher_effect, create_karplus_string, create_stochastic_texture, create_fdn_reverb, create_drum_rack_pad, build_send_chain (procedural M4L device generation)
• sound_design (7) — add_warmth, add_texture, shape_transients, add_space, configure_device, remove_device, load_chord_source
• sample (6) — sample_chop_rhythm, sample_texture_layer, sample_vocal_ghost, sample_break_layer, sample_resample_destroy, sample_one_shot_accent (registered from sample_engine/moves.py)
• performance (5) — recover_energy, decompress_tension, safe_spotlight, emergency_simplify, configure_record_readiness
• arrangement (4) — create_buildup_tension, smooth_scene_handoff, configure_groove, set_scene_metadata
• transition (4) — bridge_sections, increase_forward_motion, open_chorus, create_breakdown

Use list_semantic_moves(domain="mix") to discover available moves.

Experiment Branching — Seed-Based (canonical, PR3+)

Experiments support both the legacy move_ids path and the new seeds path. Prefer seeds for anything exploratory:

# Legacy / targeted — one semantic_move per branch:
create_experiment(request_text="make it punchier", move_ids=["make_punchier", "widen_stereo"])

# Branch-native / exploratory — mixed sources, pre-compiled plans allowed:
create_experiment(
    request_text="surprise me",
    seeds=[
        {"seed_id": "a", "source": "semantic_move", "move_id": "make_punchier",
         "novelty_label": "safe", "risk_label": "low"},
        {"seed_id": "b", "source": "freeform", "hypothesis": "Audio-rate LFO into filter cutoff",
         "novelty_label": "unexpected", "risk_label": "medium"},
        {"seed_id": "c", "source": "synthesis", "hypothesis": "Wavetable morph across positions",
         "novelty_label": "strong"},
    ],
    compiled_plans=[None, {"steps": [...], "step_count": N}, {"steps": [...], "step_count": M}],
)

Never claim a branch is previewable unless it has a valid `compiled_plan`. Analytical-only branches (no plan, or seed marked analytical_only=true) short-circuit to a neutral evaluation — they're directional suggestions, not executable paths.

Branch Status Vocabulary

Branches carry a status string that governs lifecycle:

• pending — created, not yet run
• running — apply pass in progress
• evaluated — ran and was scored; may be kept or discarded
• committed / committed_with_errors — winner was applied permanently
• discarded — rolled back or abandoned
• interesting_but_failed — (PR7+) failed hard technical gates but surfaced novel ideas; kept for audit, not re-applied
• failed — couldn't apply any steps; do not claim success

Taste-Aware Ranking

The system learns user preferences from kept/undone moves:

• get_taste_graph() — current taste model
• explain_taste_inference() — human-readable explanation
• rank_moves_by_taste(move_specs) — sort options by preference fit
• propose_next_best_move automatically applies taste ranking when evidence exists

Musical Intelligence

Song-level analysis beyond parameters:

• detect_repetition_fatigue() — clip overuse, section staleness
• detect_role_conflicts() — tracks fighting for the same space
• infer_section_purposes() — label sections as setup/tension/payoff/contrast/release
• score_emotional_arc() — does the song have a satisfying build→climax→resolve?
• analyze_phrase_arc() — capture and evaluate musical phrases
• compare_phrase_renders() — compare phrase variants side by side