Cubase-MCP-GAJ3T.MusicGen.R.../prompts/step_11.md

5.7 KiB

Step 11: Fix LM Studio-visible results and actionable request recovery

Objective and verified reproduction

Fix the actual client-facing root cause, not speculative key capitalization. Read current server/schemas/errors/profile/service tests and current architecture. Keep the strict Python/Mido/native implementation. No commits or delegation.

Verified before editing:

  • "E minor" and "E Minor" both validate and normalize to E minor.
  • "Em" currently fails the structured key regex.
  • The active LM Studio chat's FIRST actual generate_music arguments had parts=["lead"], formats=["sd3p","mid"], key="E minor", profiles={"lead":"aggressive_rock"}, and unrequested CC91.
  • Lead+sd3p produces INVALID_REQUEST with issues field="$request", reason="value_error". The text content is only "Correct the invalid request fields identified in details and retry."
  • The stored LM Studio tool result given to the model contains ONLY that generic text. Structured details are not visible in that client message.
  • Profile text currently says only "6 profiles; inspect mapping_kind and limitations", so that text-only client cannot see the real IDs either.
  • Exact user description is also currently unsupported; that musical feature is Step 12 after this fix is reviewed. Do not pretend argument changes alone make it supported.

Do not read/edit unrelated LM Studio chats/configuration. Do not upload local conversation data. Do not modify the user's active chat or DAW.

Owned implementation

  • server.py: all client-visible result/error content and useful tool descriptions.
  • schemas.py / errors.py: strict public validation context and field guidance, canonical key shorthand if appropriate.
  • Directly related unit/integration/stdio/schema/installed tests and docs.
  • prompts/report_11.md.

Required behavior

  1. For every ProfileListResult, GenerationResult and GenerationError, include the COMPLETE safe public JSON in TextContent as well as structuredContent. JSON text must decode to the same public object; do not only print a vague summary and expect every client to consume structuredContent. This follows MCP structured-output backward-compatibility guidance.
  2. Preserve isError, schemas, stdout framing, SDK lifecycle, and all native/MIDI functionality. Text-only clients must see exact profile IDs, artifacts, warnings and useful error details without private exception inputs.
  3. Cross-field errors identify the actual field and correction, e.g. formats: sd3p/sdp require drums; lead uses griff/agp or mid, bass briff/agp. Use typed/custom error reasons or an equally safe precise approach. Never expose raw Pydantic input/context/exception strings or unknown keys.
  4. Include safe actionable expected grammar/bounds for key, meter, profile IDs, parts/formats and CC shape/grid. Do not infer that the key is wrong for every generic root error. Backend strict types, limits and privacy remain enforced.
  5. Accept conventional structured key shorthand such as Em/F#m/Bb major and canonicalize correctly without breaking existing major/minor full forms. Make advertised schema/examples match actual accepted input. Preserve malformed/unknown-mode rejection; do not weaken into arbitrary defaults.
  6. Add concise input/tool descriptions explaining:
    • omit formats for automatic correct per-part output;
    • guitar lead -> griff, bass -> briff, drums -> sd3p, mid works for all;
    • omit profiles for defaults, or use exact IDs from list_instrument_profiles;
    • never invent profile IDs or add CC events without user intent;
    • supported text vocabulary/known musical limits, and pass user intent faithfully rather than embellishing unsupported instructions.
  7. Do not strip JSON Schema constraints or change transport spec on an unproven theory that llama.cpp cannot parse them. The verified bug is hidden text feedback. Any schema simplification needs concrete engine evidence.
  8. Known profile-selection/unsupported-description domain errors must also expose an actionable public message/details in text. No success-shaped approximations.

Tests

Unit:

  • json.loads(TextContent.text) equals structuredContent for both tools, success and validation/domain errors; exact IDs/corrections are visible through text.
  • Reproduce lead+sd3p with valid E minor: error points to formats and lists valid lead choices, not generic root value_error.
  • Em, E minor, E Minor and other conventional aliases normalize identically; invalid modes/values remain errors and privacy sentinels remain absent.
  • Wrong/invented profile has an exact list-profiles/default recovery action.
  • CC, meter, missing/unknown fields retain strict validation and useful context.

Integration:

  • Real stdio text-only consumer can list IDs, receive wrong-format correction, repair the call and generate actual guitar native/MIDI without guessing.
  • Real installed-wheel client sees equivalent complete text/structured payloads.
  • Existing native/CC/atomic/schema/stdio/source-template regressions still pass.

Run focused tests, full current suite, Ruff lint/format, strict mypy and pip check. Fix actual regressions without weakening assertions. No model change, network service, extra features, triplet composer implementation or later prompts here.

When complete, write your end-of-step report to ./prompts/report_11.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 inside exactly one plain triple-backtick fence. Include actual before/after text-only recovery evidence, exact tests/counts, privacy checks, schema/key decisions and Step 12 readiness. Require LM Studio MCP restart/fresh chat to refresh its cached tool schema; do not claim its live UI was restarted.