Tok
diff --git a/‎demo/scenarios/bach-italian-3rd.sh‎
Lines changed: 152 additions & 98 deletions b/‎demo/scenarios/bach-italian-3rd.sh‎
Lines changed: 152 additions & 98 deletions
@@ -6,20 +6,28 @@
 # Deliberately simple: score playback + filter sweeps + FX, not a jam.
 #
 #   - Bass score + sequencer stay LOCKED; nobody writes to them.
-#   - Pan is handled by a native LFO (state.lfo[0] targeting BassPan),
-#     configured once via the API — no scripted Python loop, no agent
-#     touching pan, no HTTP hammering during playback.
-#   - One PAD agent drifts the 303 filter (cutoff + resonance) across
-#     the movement.  That's its ONLY remit.
-#   - 808 arrives mid-piece; a KIT agent writes a half-time pulse
-#     (kicks on 0/8, closed hat on 4/12 of a 16-step loop).  Agent-
-#     driven so the drums read as a live decision, not a pre-baked
-#     pattern; scope=kit_a keeps it away from the bass sequencer.
-#   - One FX agent escalates reverb + bitcrush into the outro.
+#   - Bass voices start on ENGINE DEFAULTS (known-good sound from
+#     `./start.sh` → import MIDI).  During the 30 s intro, three
+#     direct api_params nudges move cutoff / resonance a little so
+#     there's audible filter motion before the drums arrive — no
+#     LFOs (every LFO variant on the bass filter clicked on note
+#     retriggers).
+#   - 808 arrives at 30 s; a KIT agent writes a 32-step half-time
+#     pulse and is re-asked twice during the outro to evolve the
+#     groove.  Scope=kit_a keeps it away from the bass sequencer.
+#   - FX outro is SCRIPTED, not agent-driven: three api_params nudges
+#     ramp reverb + bitcrush in three stages.  Gemma's FX lane kept
+#     emitting only reverb / ring_mod and never bitcrush keys, so
+#     the bitcrush stayed at 0 in every prior take — scripted keeps
+#     it predictable.  The whole bus passes through FX, so drums get
+#     bitcrushed too.
+#   - All 4 LFO slots are locked so no agent's mod lane can sneak a
+#     modulation onto bass params; every non-reverb / non-bitcrush
+#     FX channel is locked off so nothing else audible creeps in.
 #   - Camera tracks each addition: focus_on the new module, linger a
 #     few seconds, then return to the score.
 
-scene_count 6
+scene_count 5
 
 BACH_MIDI="${BACH_MIDI:-demo/scenarios/bach-italian-3rd.mid}"
 
@@ -45,13 +53,13 @@ set_style ""
 # Bach-specific Huth-per-component pinning: everything grayscale
 # except the sequencer step dots + event-stream history, which are
 # Huth-coloured unconditionally.  `show_thinking_in_log` is on so
-# the PAD / KIT / FX agents' reasoning traces are visible.
+# the KIT / FX agents' reasoning traces are visible.
 set_ui_prefs '{"huth_piano": false, "huth_bar_osc": false, "huth_ring_osc": false, "huth_spectrum": false, "show_thinking_in_log": true, "enable_thinking": true}'
 
-# Master volume tuned down from the default 0.85 — the resonance
-# sweeps + bitcrush outro push peak loudness into clip territory at
-# the default gain.  0.60 leaves headroom for the FX escalation.
-api_params '{"fx":{"master_volume":0.60}}'
+# Master volume stays at its default — the 0.60 override we had here
+# was adding clicky transients on the bass with no audible gain in
+# headroom (the FX agent ramps its own wet/depth, which bounds the
+# outro loudness itself).
 
 add_instrument bass
 look_at ai
@@ -68,145 +76,191 @@ PLAY_DURATION="$MIDI_DURATION"
 PLAY_DURATION_INT=${PLAY_DURATION%.*}
 echo "  [scenario] will play for ${PLAY_DURATION_INT}s at ${MIDI_BPM} BPM"
 
-# Starting timbre.  Voice 0 (right hand) brighter and L-panned; voice
-# 1 (left hand) darker and R-panned.  These are the static defaults —
-# the LFO below modulates the global bass pan on top of voice 0's
-# starting position (DSP applies one pan to the summed bass bus, so
-# the LFO moves both voices together in the current engine; real
-# anti-phase per-voice pan is a planned DSP refactor — see PLAN.md).
-api_params '{
-  "bass_voices": [
-    { "enabled": true, "volume": 1.0, "cutoff": 0.55, "resonance": 0.35, "env_mod": 0.45, "decay": 0.40, "accent_level": 0.5, "distortion": 0.10, "pan": -0.25 },
-    { "enabled": true, "volume": 1.0, "cutoff": 0.32, "resonance": 0.55, "env_mod": 0.35, "decay": 0.65, "accent_level": 0.4, "distortion": 0.05, "pan":  0.25 }
-  ]
-}'
-
-# Native pan LFO: state.lfo[0] targets BassPan with a slow sine.  The
-# audio thread runs it every block — zero HTTP traffic during
-# playback.  Rate 0.03 ≈ ~0.2 Hz (5-second cycle at the LFO's
-# internal scaling); depth 0.7 so pan swings ~±0.7 around the
-# starting position.
+# Starting timbre: DEFAULTS on the bass filter / env.  Earlier takes
+# that tweaked env_mod / decay or enabled BassCutoff / BassResonance
+# LFOs came back clicky.  `./start.sh` → import MIDI → default 303
+# is the known-good sound; this scenario reproduces it.  Cutoff and
+# resonance move during Scene 3 via `sweep_pad` (scripted api_params,
+# no LFO — the helper has been proven clean by the intro demo).
+
+# Pan LFO only: slow stereo drift on the bass bus.  Pan-only LFOs
+# don't retrigger with the filter envelope and have never caused the
+# clicks we saw with BassCutoff / BassResonance LFOs.  Locked below
+# so nothing rewires it.
 api_params '{
   "lfo": [
-    { "enabled": true, "target": "BassPan", "waveform": "Sine", "rate": 0.28, "depth": 0.7, "phase_offset": 0.0 }
+    { "enabled": true, "target": "BassPan", "waveform": "Sine", "rate": 0.28, "depth": 0.70, "phase_offset": 0.00 }
   ]
 }'
 
-# Lock the score + the entire sequencer so nothing (agent or jam
-# loop) can overwrite the imported notes.  The PAD agent is scoped
-# to "bass" synth params only; these locks are belt-and-braces.
+# Lock the imported score and every FX bus channel except the two we
+# want audible (reverb + bitcrush).  The fx state is a flat struct —
+# delay / chorus / phaser / ring_mod / compressor / distortion /
+# tape / waveshaper / eq / autotune all run in the DSP regardless of
+# which modules are in the rack.  Last run the FX agent kept writing
+# ring_mod_mix despite the prompt saying reverb+bitcrush only, and
+# since no ring-mod module was visible, it sounded like "invisible
+# FX".  Locking their mix/gain paths here means the agent can't turn
+# any of them on, so what's audible matches what's in the rack.
+# All four lfo slots are locked too — the planner keeps picking the
+# mod lane for any scoped agent, and an unlocked slot lets an LFO
+# onto something like `bass_voices.1.env_mod` that then reads as
+# "invisible LFO making the bass bleeP".
 lock \
     "sequencer.bass_pattern" \
     "sequencer.bass_patterns" \
     "sequencer.bass_steps" \
+    "sequencer.bass2_steps" \
+    "sequencer.bass3_steps" \
+    "sequencer.bass4_steps" \
     "sequencer.bass_voice_steps" \
+    "sequencer.bass_notes" \
+    "sequencer.bass_accents" \
+    "sequencer.bass_slides" \
+    "sequencer.bass_pans" \
+    "sequencer.bass2_notes" \
+    "sequencer.bass2_accents" \
+    "sequencer.bass2_slides" \
+    "sequencer.bass2_pans" \
+    "sequencer.bass3_notes" \
+    "sequencer.bass3_accents" \
+    "sequencer.bass3_slides" \
+    "sequencer.bass3_pans" \
+    "sequencer.bass4_notes" \
+    "sequencer.bass4_accents" \
+    "sequencer.bass4_slides" \
+    "sequencer.bass4_pans" \
     "sequencer.bpm" \
     "sequencer.steps" \
-    "bass_voices.0.note" \
-    "bass_voices.1.note" \
-    "bass_voices.0.pan" \
-    "bass_voices.1.pan" \
-    "bass.pan" \
-    "lfo[0]"
+    "lfo[0]" "lfo[1]" "lfo[2]" "lfo[3]" \
+    "fx.delay_mix" "fx.chorus_mix" "fx.phaser_mix" "fx.ring_mod_mix" \
+    "fx.waveshaper_mix" "fx.distortion_mix" "fx.compressor_mix" \
+    "fx.tape_mix" \
+    "fx.eq_low_gain" "fx.eq_mid_gain" "fx.eq_hi_gain"
 
 show_then_return bass 4
 
-# ── Scene 2: Play + PAD agent ───────────────────────────────────────────────
+# ── Scene 2: Play ──────────────────────────────────────────────────────────
+# No PAD agent.  The three LFOs configured in Setup sweep pan /
+# cutoff / resonance natively — that's the filter performance.
+# AI narrative for this scenario lives on the KIT + FX agents added
+# below.
 
 scene "Play"
 
 look_at sequencer
 play
 
-add_agent PAD gemma bass
-wait_for_model
-
-# PAD's ONLY remit: squelch the filter (cutoff + resonance).  No pan
-# (LFO owns it), no notes, no patterns, no anything else.  Bar-based
-# ramps so the motion develops across pattern cycles, not bumps.
-ask "PAD: squelch the 303 filter across the Bach outro.  Move ONLY cutoff and resonance — everything else is locked or off-limits.  Cutoff in [0.05, 0.55], resonance in [0.70, 1.0].  Keep voice 0 and voice 1 within 0.12 of each other so the two hands read as one squelching instrument.
-
-Use \`ramps\` with \`bars\`: 4 or 8.  Never step-jump.  Example per turn:
-
-  {\"ramps\":[
-    {\"param\":\"bass.cutoff\",\"to\":0.12,\"bars\":8},
-    {\"param\":\"bass.resonance\",\"to\":0.95,\"bars\":8},
-    {\"param\":\"bass_voices.1.cutoff\",\"to\":0.14,\"bars\":8},
-    {\"param\":\"bass_voices.1.resonance\",\"to\":0.92,\"bars\":8}
-  ]}
-
-Each next cycle, re-target to new corners of the cutoff/resonance pad so the motion is a continuous drift.  DO NOT touch: notes, steps, patterns, gate, accent, slide, volume, pan, distortion, env_mod, decay.  DO NOT add ramps for any of those.  Cutoff + resonance on both voices — that is your entire scope." PAD 0
-
 # ── Scene 3: Drums arrive (+60s) ────────────────────────────────────────────
-# Deterministic half-time kick + hat pattern via direct api_params.
-# No KIT agent — the pattern is fixed, the scene is just showing the
-# 808 layering in.  chain_advance_preserve_non_bass carries the
-# drums across bank swaps so they play for the rest of the piece.
+# KIT writes a 32-step half-time pulse under the 240 BPM bass, and
+# gets re-asked twice during the outro so the groove evolves rather
+# than loops unchanged.  Scope=kit_a keeps it away from the bass
+# sequencer.  chain_advance_preserve_non_bass carries the drums
+# across bank swaps so they play for the rest of the piece.
 
 scene "Drums arrive"
 
-# 60 s of Bach + PAD-driven filter drift before the 808 layers in.
-# Scroll around instead of staring at one place: filter pad → agent
-# console → back-panel cable tour → front → sequencer, then the 808
-# arrival reads as a deliberate reveal after the camera tour.
-fill_wait 60
+# 30 s of Bach before the 808 layers in.  Smoothly sweep the filter
+# pad via the shared `sweep_pad` helper — same motion the intro /
+# default acid demo uses, lerped keyframes at ~7.5 fps so cutoff and
+# resonance move together in a visible arc on the XY pad.  No LFO
+# (LFOs on BassCutoff kept clicking on note retriggers); this is
+# pure scripted api_params inside the helper, which we've proven
+# doesn't click.  `sweep_pad` writes to voice 0 only (matches the
+# acid demo) — voice 1 stays at defaults so the left-hand line keeps
+# a stable timbre under the sweeping right hand.
+look_at bass
+wait_seconds 4
+sweep_pad 22
+look_at sequencer
+wait_seconds 4
 
 add_instrument 808
 show_then_return 808 5
 
-# Bass 2× drums — Bach stays centre-stage.  Tuned well below the
-# intro mix because the filter sweep can get loud on resonance peaks.
-api_params '{"kit_a": {"kick": {"volume": 0.45}, "hihat_closed": {"volume": 0.30}}}'
+# Drums below Bach + stretch kit_a voices to 32 steps so the kick /
+# hat / snare rows span two bars of the 64-step bass pattern instead
+# of looping inside a 16-step window (which read as "only half a
+# bar" under the moving score).
+api_params '{
+  "kit_a": {"kick": {"volume": 0.45}, "hihat_closed": {"volume": 0.30}},
+  "sequencer": {"drum_lengths": {"kick_a": 32, "snare_a": 32, "hihat_a": 32, "hihat_a_open": 32}}
+}'
 
 add_agent KIT gemma "kit_a"
 wait_for_model
 
-# KIT writes a half-time pulse that feels like 120 under the running
-# 240 BPM bass.  Scope=kit_a keeps it away from the bass sequencer.
-# Pattern is deliberately specific — kick on 0/8, closed hat on 4/12
-# of a 16-step loop — so the first response has concrete targets
-# instead of drifting to a full 4-on-the-floor that'd overwhelm Bach.
-ask "KIT: half-time drums under a 240 BPM Bach outro.  Kit_a only, 16-step loop.  Kick on steps 0 and 8 ONLY (not every 4 — that's too dense at this tempo).  Closed hi-hat on steps 4 and 12.  NO snare on 2 and 4 — this is not pop.  Add one or two quiet ghost hits somewhere unusual for texture (a soft hat on a weird step, or a single tom ping).  Everything quiet, velocity ~0.4 max.  The drums should feel half the speed of the bass — a slow pulse under a running melodic line, not a dance beat.  DO NOT touch kit_b / hoover / an1x / bass — kit_a only." KIT 0
+# First ask: sparse 32-step half-time pulse.  Explicit step targets so
+# the model has concrete positions instead of drifting to a dense
+# 4-on-the-floor that'd bury Bach.
+ask "KIT: half-time drums under a 240 BPM Bach score.  Kit_a only, 32-step loop (two bars).  Kick on steps 0, 8, 16, 24.  Closed hi-hat on steps 4, 12, 20, 28.  NO snare on 2 and 4 — this is not pop.  Add two or three quiet ghost hits somewhere unusual for texture across the 32 steps (a soft hat on a weird step, a single tom ping, a muted kick).  Everything quiet, velocity ~0.4 max.  Half the speed of the bass — a slow pulse under a running melodic line, not a dance beat.  DO NOT touch kit_b / hoover / an1x / bass — kit_a only." KIT 0
 
 # ── Scene 4: FX outro ───────────────────────────────────────────────────────
-# One FX agent, tight scope: reverb + bitcrush only.  Three asks
-# ~30 s apart so the escalation reads as a progression, and the
-# camera stays on the FX module after each ask so you can see the
-# bitcrush wet knob swing.
+# FX is now DRIVEN DIRECTLY from bash via api_params, not by an FX
+# agent.  Reason: the FX lane of Gemma 4 reliably emits `reverb_*`
+# and `ring_mod_*` but never `bitcrush_*` in its raw output, no
+# matter how explicitly the prompt spells the field names.  Three
+# consecutive prior recordings all had bitcrush stuck at its default
+# (rate = 0, mix = 0) because the agent never wrote the keys.
+# Scripted api_params gives a predictable escalation — three nudges,
+# roughly 30 s apart, with each one bumping the three bitcrush
+# channels together so the crunch builds audibly into the ending.
 
 scene "FX outro"
 
-FX_START_OFFSET=$(echo "$PLAY_DURATION - 60 - 90" | bc -l)
+# Align FX outro so stage 3 ends ~21 s before the MIDI's natural stop
+# (previous cut used 15 s of headroom; TTS was still landing 6 s
+# after music, so another 6 s earlier).  Stages are 30 + 30 + 32 =
+# 92 s; drums phase was 30 s; everything else comes out of this
+# pre-FX fill.
+FX_START_OFFSET=$(echo "$PLAY_DURATION - 30 - 92 - 21" | bc -l)
 # Long stretch between drums-arrive and FX-outro — fill it with
 # camera motion instead of staring at the sequencer.
 fill_wait "$FX_START_OFFSET"
 
+# Add the visible FX modules.  Both reverb and bitcrush are drawn to
+# the rack so what the viewer sees matches what they hear — the bus
+# DSP runs reverb unconditionally whenever fx.reverb_mix > 0, so
+# without a reverb module visible the previous take had "invisible
+# reverb".  Other FX channels are locked above so nothing else
+# audible can sneak in.
+add_effect reverb
+show_then_return reverb 4
+enable_fx reverb
 add_effect bitcrush
 show_then_return bitcrush 4
-# Flip bitcrush on — it was added disabled by default so the module
-# could sit in the rack briefly without clicking the signal.  The FX
-# agent's ramps below then have something to affect.
 enable_fx bitcrush
 
-add_agent FX gemma fx
+# FX field-name note: state keys are `fx.bitcrush_mix`,
+# `fx.bitcrush_bits`, `fx.bitcrush_rate`, `fx.reverb_mix`,
+# `fx.reverb_size`.  `bitcrush_bits` is INVERTED: 1.0 = full quality
+# (bypass), 0.0 = 1-bit.  So "more crunch" means bits DOWN.
 
-# Stage 1: gentle.
-ask "FX: outro starts now.  Nudge reverb_mix up to about 0.30 and lengthen reverb_size a touch.  Bring bitcrush in light: wet around 0.20, bit_depth_reduction around 0.30, sample_rate_reduction around 0.25.  Leave delay / chorus / phaser / eq / compressor ALONE — this scenario only uses reverb and bitcrush.  Use ramps with bars: 4 for every move." FX 0
+# Stage 1: clearly audible bitcrush + medium reverb.  Previous values
+# were too subtle — bumped bitcrush_mix higher and bits lower so the
+# crunch is unmistakable from the first stage.  KIT gets nudged to
+# evolve so the drums don't loop a single two-bar pattern through
+# the whole outro.
+api_params '{"fx":{"reverb_mix":0.40,"reverb_size":0.60,"bitcrush_mix":0.45,"bitcrush_bits":0.55,"bitcrush_rate":0.40}}'
+ask "KIT: evolve the 32-step groove.  Keep it sparse and half-time (kick on 0/8/16/24 still anchors it), but add a couple of extra ghost kicks or off-grid hats — maybe an anticipation before the downbeat (step 7 or 23), or a pair of 16th hats somewhere in the second bar.  Still no backbeat snare.  32-step loop, velocities ~0.4 max, kit_a only." KIT 0
 
 look_at fx
 wait_seconds 30
 
-# Stage 2: dirtier.
-ask "FX: stage two.  Push bitcrush harder — wet 0.55, bit_depth_reduction 0.60, sample_rate_reduction 0.55.  Raise reverb_mix another step to about 0.45.  Still reverb + bitcrush only — nothing else.  Ramps with bars: 4." FX 0
+# Stage 2: heavier crunch + fuller reverb.
+api_params '{"fx":{"reverb_mix":0.60,"reverb_size":0.80,"bitcrush_mix":0.75,"bitcrush_bits":0.25,"bitcrush_rate":0.70}}'
+ask "KIT: push the groove further.  Stay 32 steps, half-time, kit_a only.  Vary the kick placement (can drop one of 0/8/16/24 for a breath, add one at 11 or 27 for tension), let the hats get a bit more active in the second bar.  One sparse snare ghost is allowed now — somewhere unexpected like 14 or 22, never on 4 or 12.  Velocities still quiet." KIT 0
 
 look_at fx
 wait_seconds 30
 
-# Stage 3: full chaos.
-ask "FX: final stage.  Bitcrush cranked: wet 0.85+, bit_depth_reduction 0.90, sample_rate_reduction 0.85.  Reverb_mix to 0.70 with reverb_size close to 1.0 for a cavernous tail.  Reverb + bitcrush ONLY.  Ramps with bars: 2 or 4 for a faster collapse into the final bars." FX 0
+# Stage 3: full chaos — hard digital crunch + cavernous reverb.
+# Stage 3 is 32 s so the final bitcrush values sit long enough to
+# actually register as the "collapse" moment of the piece.
+api_params '{"fx":{"reverb_mix":0.85,"reverb_size":1.00,"bitcrush_mix":1.00,"bitcrush_bits":0.05,"bitcrush_rate":0.92}}'
 
 look_at fx
-wait_seconds 30
+wait_seconds 32
 
 # ── Scene 5: End ────────────────────────────────────────────────────────────
 # chain_loop=false stops the transport on its own at the end of the