Cubase-MCP-GAJ3T.MusicGen.R.../prompts/step_6.md

129 lines
7.5 KiB
Markdown

# Step 6: Text interpretation and coordinated composition
## Objective and prerequisites
Implement Step 6 of the amended approved plan. Read the continuation decision
and accepted `report_5_fix.md`: standard MIDI fallback profiles now exist and
were independently verified. Their normal/sustain-only limitation is explicit;
real vendor maps and native/application acceptance are not claimed.
Build genuine deterministic text -> intent -> score -> mapped MIDI behavior,
not a generic loop unrelated to the text. Do not implement MCP service wiring,
native serialization, later prompts, or extra scope. No commits or delegation.
## Owned files
- Create `src/riff_groove_builder/description.py`.
- Create `src/riff_groove_builder/compose.py`.
- Create `tests/unit/test_description.py`.
- Create `tests/unit/test_compose.py`.
- Create `tests/integration/test_text_to_midi.py`.
- Add small text/JSON cases under `tests/fixtures/descriptions/` if useful.
- Update `README.md` and `docs/architecture.md` with the actual supported grammar,
template/default policies, limitations, and focused validation commands.
- Write `prompts/report_6.md`.
Reuse the authoritative request/intent/score models, profile registry, compiler,
exact tick helpers, writer and named errors. Explain/test any necessary tightly
coupled shared change rather than replacing existing working contracts.
## Text-resolution requirements
1. Provide a small typed API resolving GenerateMusicRequest into MusicalIntent
plus explicit assumptions/override records for future GenerationResult.
2. Extract numeric tempo (`120 BPM`), bar count (`4 bars`), meter (`in 7/8`),
key (`E minor`), and explicit selected part words (`lead`/`guitar`, `bass`,
`drums`). No external model API, network, audio, or new parser framework.
3. Structured override values win over text and are disclosed. Contradictory
text without an applicable override is an actionable named error.
Never silently choose one of two distinct numeric tempos/keys/meters.
4. Omitted tempo/bars/meter/key resolve to 120 BPM, four bars, 4/4, E minor, with
assumptions. If no explicit part-selection language exists, default to all
three and disclose that assumption; structured parts win over text.
5. Support a documented bounded vocabulary for riff/lead/bass groove/drum groove,
straight quarter/eighth/sixteenth-note subdivisions, rests, and optional
eighth-note triplets if represented exactly. Document deterministic rules
for qualitative slow/fast and simple rock/metal/groove pattern terms if used.
Do not add a general genre engine, song sections, or extra modes.
6. Meaningful unsupported instructions (unrecognized instrument/technique,
unsupported mode, tempo changes, polyrhythm, audio requests, etc.) must fail
explicitly rather than disappear into a default loop. Ignore only documented
filler. Match whole words, including negation/part-only selection, not
accidental substrings. Do not treat "no drums" as a request for drums.
7. Recognize canonical technique phrases such as palm mute without pretending the
standard profile can realize them. Validate techniques against selected
profiles and return UnsupportedArticulationError for unavailable techniques.
Normal/sustain must work. Do not map palm mute to softer generic MIDI.
8. Keep parsing and composition separate. Do not select production instruments
by guessing vendor names from text; registry selection remains authoritative.
## Composition requirements
1. Compose selected tracks over exactly bars * numerator * 4/denominator
quarter-note beats on the common rational grid. Each track shares start/end.
2. Generate original simple musical content from the resolved intent:
scale-based lead/riff patterns, root/fifth-oriented bass, and kick/snare/hat
drum grooves with documented meter/subdivision behavior.
Use seeded local variation, not copied factory/demo sequences.
3. Respect major/minor key pitch classes, selected profile ranges/voices and
supported articulations. Define the MIDI octave-number convention explicitly.
Prefer sensible lead/bass registers, but choose notes within the actual map.
4. Use exact fractions and supported tick divisions. Gate notes/rests to avoid
ambiguous same-pitch overlap; truncate only as an explicit composition policy
at the shared bar/end boundary, never alter requested musical duration.
5. Drum voice names resolve through the selected profile's aliases. Do not
assume every profile contains every GM sound or use guessed physical notes.
6. No global random state, unsafe pitch/velocity clamping, silent out-of-range
notes, per-track drift, or dropped requested techniques. Invalid/narrow maps
that cannot support the pattern must produce useful named errors.
7. Enforce the event budget before constructing excessive notes; account for
at least note-on/off pairs and leave final control-event validation to the
existing compiler. Keep resource use bounded for extreme valid inputs.
8. Do not change the current Step 3 server handler; the pipeline is exercised
directly until Step 9 connects it.
## Required unit tests
- Valid structured/text values; defaults and assumption records; override
precedence/disclosure; contradictory numeric/key/meter instructions.
- Part inference/ordering, exclusions, only-selection, explicit parts override,
and unsupported instrument/mode/technique/material instruction failures.
- Whole-word extraction, filler, case/spacing, malformed/empty descriptions,
fraction/tempo/bar boundaries, and unsupported tempo changes.
- Different meters, exact bar/end boundaries, seed reproducibility/variation,
key/register/range compliance, no same-pitch overlaps, and subdivision/rest
effects on actual generated content.
- Standard normal/sustain succeeds; requested palm mute fails honestly under
standard-only maps. Synthetic switch profiles may test real compiler coupling
but must never be advertised as vendor compatibility.
- Missing voices, unrepresentable timing, and event-budget rejection.
## Required integration tests
- Lead-only, bass-only, drums-only and all-parts descriptions resolve/compose,
select real standard profiles, compile, write actual MIDI and read it back.
- Independently assert tempo/meter, musical duration, parts/channels, key/range,
rhythm changes from text, and emitted note pairing, not merely file existence.
- Include 3/4, 6/8, 7/8 and nondefault tempo/key/length cases.
- Same input/seed yields deterministic output; changed supported intent changes
measurable output. Compare precise expected events for simple cases.
- Unsupported vendor techniques fail without artifacts.
- Preserve all existing contracts/compiler/MIDI/protocol/package regressions.
## Validation and acceptance
Run focused new tests, relevant prior checks, Ruff lint/format, strict mypy, and
the full current suite after shared changes. The known working venv pytest CLI
fallback is acceptable. Clean up owned test files/children.
Accept only tested text-driven, bounded, aligned standard-MIDI composition.
Real vendor articulation/native/application gates remain open and documented.
## Required report
When complete, write your end-of-step report to ./prompts/report_6.md including: what was implemented, test results (pass/fail with details), any issues encountered, and whether you're ready for the next step
Put the complete report inside exactly one plain triple-backtick fence.
Include APIs, exact supported grammar/default policies, measurable text-to-MIDI
evidence, exact commands/results/skips, limitations and next-step readiness.