10 KiB
Step 15: Hybrid prose plus precise Counts/Events/Chords/Notes lanes
Objective and accepted prerequisite
Step14 is reviewed and independently verified (158 new pitch/plan/native tests). Read report14/current pitch.py/strict plans/composer, description/service/tool feedback and native codecs. Implement the actual PUBLIC description path for the agreed hybrid notation. No new public tool, new dependency, commit, delegation, later prompts, model service, audio/host or user-app changes.
Prose supplies context/defaults; notation fixes events where supplied. Explicit pitches/timing/articulation/ties cannot be overwritten by improvisation. Spacing width is NEVER timing. Contradictory prose is an actionable error, not silent precedence. Unknown material is not ignored into an approximate success.
Owned surfaces
- One focused
notation.py/hybrid.pyparser as appropriate, reused from description.py and compose.py. Keep module/dependency direction acyclic. - Minimal reusable unresolved/typed hybrid plan in schemas.py if needed. Reuse Step14 pitch resolution and ExplicitAttackPlan, never duplicate voicing.
- service.py/result models only to expose accurate interpreted events/assumptions.
- errors.py/server.py/schema descriptions and current docs for actual support.
- Unit/integration/real stdio/installed-package hybrid tests, original fixtures.
- prompts/report_15.md.
Canonical public example that MUST work directly
Lead guitar, 4/4, 120 BPM, 2 bars. Bar 1: Counts: 1 & 2 & 3 & [4 trip let] Events: x x x x x x [x x ^] Chords: F#2 F#2 D2 D2 Bb2 Bb2 - - - Notes: - - - - - - C4 E4 G4 Hold the final G4 through the end of bar 2 without retriggering.
This is six PM chord attacks followed by PM C4, PM E4, PH G4 sustained across the bar boundary. Exact native/MIDI pitches must be F#2 major42/46/49, D2 major38/42/45, Bb2 major46/50/53, C4=60 E4=64 G4=67; explicit chords are not retuned into a key. Exactly21 musical note-ons, not counting switches. Onsets MIDI960: chord groups0,480,960,1440,1920,2400; C4=2880/E4=3200/ G4=3520; first durations480/two triplets320; final G4 end7680.
Also support original single-note two-measure request using lanes: Lead guitar, E minor, 4/4, 120 BPM, 2 bars. Bar 1: Counts: 1& 2& 3& [4 trip let] Events: xx xx xx xx^ Notes: E2E2 E2E2 E2E2 E2E2E2 Hold the final E2 through the end of bar 2 without retriggering.
That MUST preserve the exact previous original phrase's nine musical events, PM26/PH31 switches, durations, pitches and seed0 velocities/bytes where feasible. Accept normal whitespace and tabs, not column-position heuristics.
Count/grid semantics
- Parse labeled Counts:/Events:/Chords:/Notes: rows case-insensitively, preserve pitch-token capitalization semantics from pitch.py. Rows describe one selected pitched part; lead/guitar or bass inferred from prose/part override. No simultaneous drum-lane language is invented in this implementation.
- Numeric labels identify counted meter-denominator beats (quarter in4/4).
Complete
1 &/1&= two eighth positions. Complete1 e & a= four sixteenth positions, uppercase E/A equivalent (not a different rhythm). A bare single beat label = one beat. Explicit[4 trip let]is a3-way eighth-triplet beat; bracket syntax declares the subdivision, not visual width. - Support an explicit Grid:/Subdivision: eighths/sixteenths/eighth-triplets
directive for disambiguation, with a clear tested count-label interpretation.
Bare
1&awithout explicit sixteenth/triplet grid is AMBIGUOUS and errors. Never infer a triplet just from three attacks or lowercase letters. - Counts ordered within bar, range checked against meter, no duplicate/regressed position. Use exact Fraction math and proper denominator units, reject unsupported grids rather than floating rounding. At least different 3/4 simple-meter case plus4/4; do not claim idiomatic compound grouping.
- Multiple
Bar N:blocks and/or matching|bar boundaries may be supported if kept simple; at least a second explicit continuation bar must work. Validate all labeled rows' flattened slots/bar spans agree. No silently truncated/zip-shortened mismatched rows. - Brackets on Events/Notes may group matching triplet tokens; they are not
simultaneous chord syntax there. Chords:
[C3 E3 G3]is one explicit simultaneous voicing token. Document and test this distinction. Avoid nested bracket heuristics silently swallowing a voiced chord as three attacks.
Event/pitch/hold semantics
o= normal/open attack (not open-string fingering).x= palm-muted attack.^= pinch-harmonic attack (evidenced AME only).-in Events = rest (ends active prior sound unless it ended already).~= extend PREVIOUS sounding attack/chord without retrigger; cannot follow silence, conflict with new pitch/articulation or silently create a new note.|= bar boundary, when used consistently./and\in Events are RESERVED slide directions. Without an implemented evidence-backed origin/destination/timing representation, return a specific unsupported-slide error with guidance; do NOT silently emit a normal attack or confuse Chords:Em4/B2 with an Events:/ symbol.
- Each slot has one event symbol and exactly one sounding pitch source or none.
Chords/Notes rows are optional individually; absent row is all
-. Both sounding at one slot is a conflict, not an arbitrary merge. A pitch-lane-means empty lane, not necessarily global rest. Active event with no pitch may use the explicitly disclosed low-tonic context default; never invent pitches when another precise pitch source exists. - Whitespace separates tokens but runs may be compact: Events:
xxxxxx^; Chords:F#F# DD BbBb ---; Notes:-- -- -- C4E4G4. Use Step14 deterministic tokenizer to resolve atoms; values never depend on how many spaces were used. Slash/Over bass, octave-qualified chords and [explicit voicing] must remain one slot, not be split by naive whitespace. - Explicit notes/chords override key-based generation by design, not by silently altering prose: a stated key is context; raw pitches stay exact. If prose directly specifies a different pitch for the SAME event, flag the conflict.
- Resolve octave defaults through Step14 profile/routing policy, propagating all
chosen roots/bass octaves and voicings as visible assumptions/interpretation.
Em4OverB2=[47,64,67,71], Em4/B=[59,64,67,71]; reject explicit bass not lowest. No dropped chord tone, automatic key retune or hidden revoicing. - Hold directive supports natural
Hold/Tie the final [note|chord|G4|E2] through/to the end of bar/measure 2 without retriggering.An optional pitch must match that final note. Extend to exact bar end, no second attack.Bar 2: hold previous note to bar endshould also work and not be consumed by the unsupported globalto/fromrule. Restrict recognized clauses first. - A second
Countsbar with all Events:~and empty pitch lanes must yield one uninterrupted prior note/chord. Sustaining PH preserves its original articulation. Explicit rests/new attacks cannot be overridden by a prose hold. Reject a hold spanning specified conflicting notation. - Structured scalar overrides retain existing documented precedence ONLY when compatible with explicit timing/pitches; a shorter bars override or incompatible meter must error instead of dropping/regenerating exact material.
Prose/context and observed interpretation
- Support documented common context around the rows (guitar/bass, key incl.Em, numeric BPM/bars/meter, "two bars", low-register chord voicings, aggressive/ tight metal style and bounded velocity variation). State actual policy: exact notation cannot be altered by style; no unrequested CC/effects.
- Prose-only inputs continue on old paths byte-identically. Do not remove old unsupported-tempo-change safeguards or accept arbitrary instructions.
- When prose specifies a global technique and lanes explicitly differ for the same events, raise an actionable conflict. Do not mistake explicit intentional phrase overrides (first PM then PH) for contradictory accidental global input.
- The tool should show its interpretation. Add a backward-compatible optional/ default result interpretation surface if necessary: exact resolved pitch sets, start/duration as rational beats, technique and holds. Include canonical voicings/register assumptions in complete JSON text and structuredContent. Do not dump private raw text or paths in errors. Existing simple callers should not need to provide new required fields.
- Help text must clearly advertise tested lane syntax/count grid/pitch convention, not overclaim arbitrary natural-language rhythm comprehension.
Required tests
Unit: canonical/compact examples, width/case invariance, complete eighth/ sixteenth/explicit triplet grids, ambiguous1&a rejection/explicit-grid recovery, row slot mismatches, brackets/voicing/slash bass, rests/lane-empty vs rest, continuation/hold no retrigger, contradictory pitch/articulation/prose, bad bars/meters/labels/keys, per-part capability errors (PH on bass), slide reserved, malformed/huge inputs/budget, raw explicit pitches preserved outside key.
Integration:
- Real source stdio AND installed-wheel native/MIDI with text-only JSON parity.
- Independently assert21 tones for the exact chord example and9 for the original E2 example, exact onsets/durations/PH and no bar2 retrigger, every chord tone, legal native fingering and actual controller absence.
- Em4OverB2/Em4/B/explicit voicing survive the public description to real MIDI/ griff without retune/drop, default bass/root assumptions visible.
- All-~ bar matches prose-tie bytes/semantics; dash/rest truly separates attacks.
- Genuine invalid/mismatched/ambiguous notation errors publish no artifacts.
- All current template/native/CC/atomic/privacy/old-phrase/pitch regressions.
Run focused new tests, full current suite, Ruff lint/format, strict mypy and pip. No version/release metadata changes yet (Step16 owns final0.2.0).
When complete, write your end-of-step report to ./prompts/report_15.md including: what was implemented, test results (pass/fail with details), any issues encountered, and whether you're ready for the next step
Entire report in exactly one plain triple-backtick fence. Include supported grammar/policies/limits, actual native/MIDI event proof, compact and nearby examples, exact counts/checks, result interpretation and Step16 readiness.