34 KiB
34 KiB
CURRENT — STEP 10 FINAL REAL NATIVE/MIDI SOFTWARE ACCEPTANCE
Date: 2026-10-20
Executed: prompts\step_10_fix.md
RESULT / CERTIFICATION BOUNDARY
PASS: implemented Eclipse 4.0.1 Riffer XML 2.1 .griff, Ray5 4.1.0 .briff,
SD3 3.4.4/SL-DFH template-song .sd3p and Standard MIDI software subset.
PASS: full suite, lint/format, strict types, pip consistency, final wheel/sdist,
EXACT persisted final wheel real-stdio generation outside the checkout,
independent saved-byte music checks and genuine READ-ONLY local-template run.
Native serialization is implemented, not a renamed-MIDI or unavailable-writer
placeholder. agp/sdp are request aliases; artifact metadata uses real suffixes.
UNRUN: actual Eclipse/Ray5/SD3 open/inspect/play/save/close/reload, audible
articulation/controller response, Cubase MIDI/native import/routing and playback.
The previously observed Cubase15 unnamed user project (PID 41836) was never
manipulated, closed or saved. No global input, audio host/driver launch, audio
analysis/rendering, plugin hosting or DAW automation was performed.
Read-only process inspection is not an application test or permission to host.
No full-library/all-version/arbitrary-technique or acoustic compatibility claim.
Software readiness is COMPLETE; the original full application definition of
done remains uncertified. These limits are not pytest skips or fake passes.
WHAT CHANGED / WHAT WAS PRESERVED
Read the latest plan, CURRENT native/service/CC reports 7_fix, 8_fix, 9_fix,
9_fix_controls, current codecs/service/contracts/composition/tests and docs.
Preserved all pre-existing native implementation and unrelated working changes.
No production-code defect was exposed; no speculative production patch,
dependency-manifest change, package/dependency restore, commit, delegation,
later prompt execution or unrelated feature was introduced.
Directly necessary acceptance changes:
- tests\integration\test_distribution.py now additionally verifies installed
target-default missing-template partial fallback and disabled-fallback error,
actual-name native-only requests, and native collision byte preservation.
Existing musical/control/privacy/atomic assertions remain intact.
- tests\unit\test_readme.py adds eight tests: six executable JSON contracts,
their inventory, and current local documentation file/anchor links.
Requests also resolve, compose and compile against current profiles.
- README/current architecture/profiles/format/compatibility docs and plan now
identify actual native defaults, aliases, CC/technique boundaries, current
software acceptance and separate UNRUN live checks. Exact filler vocabulary
is documented; stale five-profile/current MIDI-only status is removed.
- Found a real coupled documentation defect: README's publication-safety link
targeted #implemented-in-step-9, lost when the native-fix heading changed.
Restored an explicit stable anchor and added the local-link regression.
The compatibility section's existing public anchor is also retained.
- This report is CURRENT-first; the entire prior MIDI-only report remains below
as explicitly retracted historical scope, not approved native completion.
FINAL EXECUTED VALIDATION — ALL PASS
PowerShell, existing repository .venv, Windows 11 build 10.0.26200 AMD64,
CPython 3.13.13 / MSC v.1944 64 bit.
Scratch/build temporary locations were explicitly repository-local:
$scratch = (Resolve-Path '.test-artifacts\scratch').Path
$env:TEMP=$scratch; $env:TMP=$scratch; $env:TMPDIR=$scratch
$env:PIP_NO_INDEX='1'
Full-suite genuine-template activation:
$env:RIFF_SD3_TEMPLATE='C:\Users\g3\Documents\Toontrack\Superior Drummer 3\sd3-demo.sd3p'
.\.venv\Scripts\python.exe -m pytest -q --tb=short
1421 passed in 63.67s; 0 failed, 0 skipped, 0 xfails.
Prior current baseline independently rerun: 1413 passed in 61.51s.
Eight new documentation contract/link cases account for the count change.
Includes native/MIDI/CC/physical-string/profile/stdin/packaging regressions,
read-only genuine-template cases, real Windows junctions, concurrent
publication, all-or-nothing multi-artifact failures and no-overwrite.
Final log: .test-artifacts\step10-full-final.log
.\.venv\Scripts\python.exe -m pytest tests\unit\test_readme.py tests\integration\test_distribution.py -q --tb=short
11 passed in 12.66s; 0 failures/skips.
Log: .test-artifacts\step10-focused-final.log
.\.venv\Scripts\python.exe -m ruff check .
All checks passed.
.\.venv\Scripts\python.exe -m ruff format --check .
85 files already formatted.
.\.venv\Scripts\python.exe -m mypy --strict src tests
Success: no issues found in 47 source files.
.\.venv\Scripts\python.exe -m pip check
No broken requirements found.
git --no-pager diff --check
PASS, no whitespace errors.
Editor Problems on both changed Python tests: no errors.
Editor runTests: "No tests found"; not represented as a pass or skip.
Actual repository-venv pytest above is the authoritative test execution.
Static log: .test-artifacts\step10-static-final.log
Actual versions:
mcp 1.30.0; pydantic 2.14.1; mido 1.3.3; MIDIUtil 1.2.1;
pytest 9.1.1; anyio 4.15.1; ruff 0.17.1; mypy 2.4.0;
build 1.6.1; setuptools 84.2.0; wheel 0.48.0; packaging 26.3;
jsonschema 4.26.0; types-jsonschema 4.26.0.20261006; pip 26.0.1.
No claim of testing Python 3.11/3.12, other OSes or every permitted dependency
version. Existing dev-venv dependencies were reused for installed-wheel runtime;
this is not a fresh complete dependency-environment certification.
FINAL PERSISTED DISTRIBUTIONS / SOURCE IDENTITY
After all source/test/README/current-doc/plan edits:
.\.venv\Scripts\python.exe -I -B -m build --no-isolation --outdir .\dist\step10-native-20261020
PASS: real sdist followed by wheel built from that sdist; offline.
Retained FINAL files, package version 0.1.0:
dist\step10-native-20261020\gaj3t_musicgen_riff_groove_builder-0.1.0-py3-none-any.whl
74822 bytes; SHA256
cb4e6978d7d778d96eda88ce70b5f5c43a17e44aae9fcfdc834e3ea3a0d0502d
dist\step10-native-20261020\gaj3t_musicgen_riff_groove_builder-0.1.0.tar.gz
85030 bytes; SHA256
26671d01c39c9746fd1ea251aca1a2cb20496c8fd3a1d88e09533614b27552de
Final 68-file source/test/fixture/docs/manifest/README/plan SHA256 manifest:
5765b606bf661da59858914f4d146956c647a7beaf88ea2f168a3301d3a07d9e
Stored as source_manifest in the retained final acceptance.json. The digest
covers sorted compact JSON of relative paths -> file-byte SHA256; it excludes
this self-referential report, caches, build metadata and ignored evidence.
Separate delivery checks rehash every file and both final archives.
All 22 actual package Python/JSON files match BOTH archives byte-for-byte;
sdist README matches final README. Wheel has 26 entries including metadata,
16 Python modules/entry locations, SIX real JSON profile resources and actual
formats\agp.py / formats\sdp.py codecs. All six resource hashes also match
the installed wheel resources. Archive inspections exclude private templates,
factory/native musical files, manuals, executables/plugins and scratch trees.
No private template, kit, factory song, licensed manual, binary, credential or
upload was added to source/tests/distributions. Original portable fixtures are
newly authored test content, not stripped private exports.
Prior dist files were preserved; initial/pre-link builds under separately
labeled dist directories are NOT the final accepted wheel above.
EXACT FINAL-WHEEL REAL STDIO / SAVED ORIGINAL EVIDENCE
Retained root (R below):
.test-artifacts\step10-final-native-20261020\final-verified
Driver/checkers retained in its parent:
accept.py; verify_saved.py; verify_delivery.py
Final execution:
.\.venv\Scripts\python.exe -B .\.test-artifacts\step10-final-native-20261020\accept.py
.\.venv\Scripts\python.exe -I -B .\.test-artifacts\step10-final-native-20261020\verify_saved.py
.\.venv\Scripts\python.exe -I -B .\.test-artifacts\step10-final-native-20261020\verify_delivery.py
The driver installed THAT exact persisted wheel with
--no-index --no-deps --no-cache-dir --target R\installed-wheel, noneditably.
Every real server subprocess used -I -B and cwd:
D:\_Repos\Code\GAJ3T\MCPs\MusicGen
That existing directory is OUTSIDE the checkout; no files were created there.
All 16 imported package/submodule/entry locations and distribution metadata
resolve under R\installed-wheel, not editable src. Actual service location:
R\installed-wheel\riff_groove_builder\service.py
Installed importlib.resources reads all six real profile JSONs there.
Probe/resource/module hashes and complete requests/results/live schemas are
saved in R\acceptance.json. No earlier build substitutes for this proof.
Three actual initialized/EOF stdio sessions, stderr EMPTY and exit 0 each:
portable: 11 successful generations, 5 expected errors, 20 stdout lines
missing-template: 2 successful generations, 2 expected errors, 7 stdout lines
genuine-private: 1 successful generation, 1 expected error, 5 stdout lines
All 32 UTF-8 stdout lines validate as MCP JSON-RPC. Missing-template disabled
fallback, invalid associations/endpoints, unsupported native CC disabled
fallback and collisions leave output snapshots unchanged.
Separate corrupt-template startup: exit 2, ZERO stdout bytes, ZERO outputs,
visible stderr error; no successful fallback. Only an owned deliberately bad
portable-template file was removed after that probe.
Portable cases use an unchanged copy of the ORIGINAL fixture
tests\fixtures\sd3\original-template.sd3p, not a genuine/private kit template.
R\portable contains:
default: actual mid + griff + briff + sd3p (four outputs);
aliases: agp+sdp -> griff/briff/sd3p only;
actual: griff+briff+sd3p -> those three native outputs only;
midi: explicit MIDI-only with native fallback disabled;
final-cc-pitched/drums/other/midi/tick/omitted/empty: seven exact CC cases.
R\missing-template contains:
fallback: griff/briff plus ONE MIDI, fallback_for=["drums"] only;
midi-no-template: explicit MIDI plus pitched CC1/reset, no SD3 template needed.
R\genuine-private\genuine-original contains four genuine-template outputs.
Total: 14 successful generation scenarios, 8 intentionally rejected tool calls,
38 actual retained musical artifacts; no .agp/.sdp mislabeled files.
Independent SECOND process imports NO riff_groove_builder modules. It reopens
every saved file using Mido, ElementTree and a separate quoted-brace/SD3 triple
decoder, with fixed numeric music/velocity/fingering expectations, not the
production serializer/compiler as its expected output.
R\independent-saved-checks.json: 38 passed; 0 failed; 0 skipped;
package_imported=false; source_templates_unchanged=true.
It checks exact note on/off timing/velocities, native physical string/fret,
PM/PMP, tempo/meter/key/end, switches/channels/CC raw bytes/native stairs,
artifact size/suffix/profile metadata, subset/fallback warnings, directory
coverage, original no-control hashes and exact non-song template bytes.
R\delivery-checks.json additionally confirms 66 current local documentation
file/anchor links and all retained source/build/artifact hashes.
All six README fenced JSON examples pass current typed contracts; all three
request examples compile, and the real installed server's discovered schemas
also validate the documented requests/response/error.
EXACT ORIGINAL MUSICAL CONTENT / DEFAULT HASHES
Request: seed 42; original "lead bass drums palm mute 137.5 BPM 2 bars 7/8
Bb minor quarter rests"; target defaults; fallback disabled when configured.
Resolved A# minor; MIDI PPQ960, start0/end6720; native PPQ480/end3360;
SD3 interval0..7 quarter beats. Tempo 436364 us/quarter, meter7/8,
Conductor/Lead/Bass/Drums, channels0/1/9, every MIDI track EOT6720.
Pitched onsets: 0,960,1920,3360,4320,5280; duration960; release velocity0.
Lead pitches: 58,61,66,58,61,65; velocities96,64,96,64,64,96.
Bass pitches: 46,53,46,46,53,46; velocities48,51,55,48,51,51.
Drums: simultaneous hat42 and hit36,38,36,36,38,36 at those onsets;
each duration240; paired velocities80,96,96,96,64,64.
Lead selector26/bass24, switch velocity100, interval0..30; no drum switch.
Native lead PM, bass PMP, not separate D0 bass PM.
Physical lead pitch -> (string,fret): 58->(2,1),61->(2,4),
66->(1,4),65->(1,3); bass46->(3,1),53->(2,3).
Notes neither dropped nor shortened; final rests preserve full score duration.
R\portable\default:
default.mid 336 bytes
8fc4cfbe66c0706f223ec33a61692ead046a56a9dba51c6566f967690cee186c
default_lead.griff 1296 bytes
fe7a2a5c98d0ae16f6894e681e7f001a658c90d11b3f1beb313d5c98b6c810c9
default_bass.briff 1282 bytes
5da9b05e92e6989366f41f8abf22a0f8b88ae14591bc8b544e4c7c0cf47827c4
default_drums.sd3p 1491 bytes
8bee8a86a7e60320d02c4404e36d818639af1d8f4ebce27f6c23a25e7ebd0589
Aliases/actual native-only originals match these native bytes. Omitted and
empty CC requests remain pinned to these PRE-controls report_9_fix hashes.
CC1 MIDI (tick,controller,value):
lead channel0: (0,1,11),(480,1,73),(6720,1,0)
bass channel1: (0,1,21),(960,1,127),(6720,1,0)
Native tick/value stairs:
lead: (0,11),(240,73),(3360,0)
bass: (0,21),(480,127),(3360,0)
End zeros are EXPLICIT requests, never invented resets. Same-tick controls
precede note-ons. All-part CC4 drum points (320,4,96),(6720,4,0) yield ONLY
drum fallback, reason sd3_native_controller_not_supported; Riffer files survive.
Lead CC74 (320,74,127),(6720,74,0) yields ONLY lead native fallback, reason
unsupported_native_controller; bass/SD3 natives survive. MIDI-only retains all
ten mixed controls; one-tick lead CC1=99 at tick1 has no invented reset.
No audible modulation/vibrato/effect semantics are inferred from raw CC data.
GENUINE READ-ONLY LOCAL TEMPLATE / PRIVATE STATE
Read in place only:
C:\Users\g3\Documents\Toontrack\Superior Drummer 3\sd3-demo.sd3p
Before/after identity/content verified both by service and separate saved check:
size75393; inode166633186213030518; mtime_ns1788614183091903900;
SHA256 06f54b02079e4b32d532ecc232b813c3fe1865799e6f5b2b21b5ddf398b0533d
All 170 old source-song notes are replaced by twelve ORIGINAL generated notes,
one track/clip, exact137.5BPM/7/8/0..7beats/channel9/velocities64..96.
R\genuine-private\genuine-original\genuine-original_drums.sd3p:
66325 bytes; SHA256
265535e1b885ff7765becce8281873ec627e28b35b5404a0f0cd9b533c424680
Non-song bytes, including kit/sampler/mixer/tracker/host-sync, are identical:
SHA256 10340dc5c057f2580c474ebc56277b2deae3263cee49e4674e07b8c9cc421d32
Its MIDI/griff/briff hashes equal the portable originals above.
Generated genuine evidence RETAINS PRIVATE KIT STATE; it stays local/ignored
under .test-artifacts, never source/tests/distributions/uploads.
Portable original source1110bytes has SHA256
ef72e05498d2809ac47432e925d6072258cbe3c918de13b9079c2353af1759a4
and also remains unchanged in bytes/identity/size/mtime.
FAULTS / CLEANUP / HONEST LIMITS
Full-suite tests prove partial last-native writes, corrupted saved bytes,
same-size/mtime-restored earlier-artifact corruption, ownership replacement,
unknown stage entries, denied cleanup, publish/model/serialization/IO faults
do not publish partial successes or overwrite existing output. Concurrent
four-file publication has one winner and one OUTPUT_EXISTS. Unexpected faults
never become MIDI successes. Explicit corrupt templates are errors.
Only issues encountered here: documented editor test discovery absence;
the real stale documentation anchor fixed/tested above; an initial independent
evidence script assumed LF rather than preserving genuine CRLF and omitted
null profile-selection fields in its comparison. Those checker assumptions
were corrected without changing musical expectations or production behavior.
Final independent checks pass on actual original native bytes.
Every owned stdio child was EOF-closed and awaited; no acceptance host/server
is left running. UUID distribution/native-service/stdio/control test workspaces
were cleaned by fixtures; no incomplete stage remains in retained final trees.
Retained original outputs, installed-wheel provenance, all JSON checks and final
builds are intentionally preserved and ignored. Earlier evidence/builds were
not deleted or represented as this exact-wheel acceptance.
Trusted local ACL/no-hostile-race/power-loss limits remain documented.
READY: all actionable final native/MIDI SOFTWARE acceptance is verified and
persistent. No implementation/package blocker remains for the declared subset.
NOT CERTIFIED: live receiver/plugin/Cubase behavior, audible performance, broader
libraries/versions/techniques/controllers, other runtimes/OSes or fresh complete
dependency installation. No later prompt was executed or new host launched.
======================================================================
HISTORICAL — PRIOR MIDI-ONLY STEP 10 REPORT (RETRACTED INTERIM SCOPE)
The following original report is retained verbatim as history. Its MIDI-only
closure and BLOCKED native implementation findings are superseded above;
they were not user approval to reduce the native-output requirement.
======================================================================
Step 10 — installed software acceptance and truthful delivery documentation
Date: 2026-10-19
SEPARATE OUTCOMES
Independent Windows Standard MIDI software/package/docs acceptance: PASS.
Complete current pytest suite: 1119 passed; 0 failed; 0 skipped.
Actual final persisted wheel server outside checkout and disk read-back: PASS.
Final real sdist/wheel build, lint, format, strict types and pip check: PASS.
Verified Eclips/Ray/SD3 vendor mappings: BLOCKED; not supplied or certified.
Real AGP/SDP serialization/native acceptance: BLOCKED.
Cubase/plugin/receiver import and native load/save/reload: UNRUN.
No full native-project completion or original full release acceptance is claimed.
IMPLEMENTED AND PRESERVED
Executed prompts\step_10.md under the amended approved functional-MIDI plan.
Reviewed the amended plan, report_9.md/prior report evidence, native fix
handoffs, and current implementation, documentation and tests.
Updated:
- tests\integration\test_distribution.py: one new installed-wheel server
acceptance test, extending the existing offline sdist/wheel build fixture.
Reuses RawStdio/TIMEOUT from test_stdio.py instead of duplicating transport
client logic. Existing separate-venv import/distribution tests remain.
- README.md: operational standard-only identity, tested configuration,
setup/existing-root requirements, absolute Windows MCP configuration, profile
discovery, accepted requests, schema-valid concrete responses/errors,
exact limits/defaults/overrides, routing and filesystem limitations,
installed-wheel acceptance and explicit remaining native handoffs.
- docs\architecture.md: removes obsolete current empty-registry/unfinished
generation statements, documents actual service layers and installed-wheel
acceptance, preserves native gates and trusted-ACL/race/crash limits.
- docs\compatibility.md: actual current Python/OS/dependency versions and
software results, separately BLOCKED native/vendor and UNRUN application rows.
- docs\articulation-profiles.md: installed standard-only resources/routing,
external receiver setup and unchanged vendor limitations.
- docs\formats\agp.md and sdp.md: current operational service/fallback status,
explicit BLOCKED step_7_fix.md/step_8_fix.md handoffs and missing evidence.
- prompts\plan.md: corrects obsolete proposed/awaiting-review status to the
approved partial-native status; labels the old review instruction historical.
The pre-existing continuation decision and original full definition of done
remain intact.
- prompts\report_10.md: this persistent report.
No production source, profile JSON, dependency manifest, prior report or native
fix prompt was changed. No actual production defect was exposed by acceptance,
so no speculative source fix or artificial regression was added. The new test
closes the real installed-server acceptance gap. No commits, delegation, later
prompts, extra formats, vendor profiles, native fixtures or application claims.
Pre-existing unrelated/untracked implementation work was preserved.
TESTED ENVIRONMENT
Windows 11 / 10.0.26200 / AMD64.
CPython 3.13.13, MSC v.1944 64 bit (AMD64).
Existing interpreter:
D:\_Repos\Code\GAJ3T\MCPs\MusicGen\GAJ3T.MusicGen.Riff-Groove-Builder\.venv\Scripts\python.exe
Actual versions:
mcp 1.30.0; pydantic 2.14.1; mido 1.3.3; MIDIUtil 1.2.1;
pytest 9.1.1; anyio 4.15.1; ruff 0.17.1; mypy 2.4.0;
build 1.6.1; setuptools 84.2.0; wheel 0.48.0; packaging 26.3;
jsonschema 4.26.0; types-jsonschema 4.26.0.20261006; pip 26.0.1.
No project dependency installation/restoration/reinstallation was performed.
Tests installed only their built application wheels offline into owned targets
and the existing isolated import-test venv. Runtime server dependencies were
reused from the development venv. This is NOT a fresh full dependency-install
certification, a test of all allowed dependency versions, Python 3.11/3.12,
another OS, or actual receiver/plugin behavior.
EXACT VALIDATION COMMANDS AND RESULTS
Commands ran in PowerShell from the repository root. Repository-local
test/build scratch setup (repeated for each relevant shell):
$scratch = New-Item -ItemType Directory -Force '.test-artifacts\scratch'
$env:TEMP = $scratch.FullName
$env:TMP = $scratch.FullName
$env:TMPDIR = $scratch.FullName
$env:PIP_NO_INDEX = '1'
Editor runTests was attempted for the absolute distribution test path:
"No tests found in the files." This is the known discovery limitation,
not a pass or pytest skip. The working repository-venv CLI was used.
Final focused acceptance:
& '.\.venv\Scripts\python.exe' -m pytest tests\integration\test_distribution.py -q -s --tb=short
PASS: 3 passed in 10.12s; 0 failed; 0 skipped.
Includes both original build/import tests and the new real installed-wheel
server generation/error/EOF acceptance. An earlier focused run before the last
assertion additions also passed 3 tests in 9.81s.
Final complete current suite:
& '.\.venv\Scripts\python.exe' -m pytest -q --tb=short
PASS: 1119 passed in 38.13s; 0 failed; 0 skipped.
Preserves all source/distribution/profile/compiler/description/composition/
MIDI/MIDIUtil/native-unavailability/service/stdio tests, including genuine
Windows junction and competing publication regressions. Native/application
evidence missing outside this executable software suite is not a pytest skip.
& '.\.venv\Scripts\python.exe' -m ruff check .
PASS: All checks passed!
& '.\.venv\Scripts\python.exe' -m ruff format --check .
PASS: 69 files already formatted on the final rerun after persisting this report
(68 on the earlier pre-report run).
& '.\.venv\Scripts\python.exe' -m mypy --strict src tests
PASS: Success: no issues found in 38 source files.
& '.\.venv\Scripts\python.exe' -m pip check
PASS: No broken requirements found.
git --no-pager diff --check
PASS: no tracked-diff whitespace errors. Existing implementation is largely
untracked; Ruff separately checked the actual Python files.
Editor problems: no errors in the changed Python test.
Final distributions, after the final test/source/README edits:
& '.\.venv\Scripts\python.exe' -m build --no-isolation
PASS: Successfully built gaj3t_musicgen_riff_groove_builder-0.1.0.tar.gz
and gaj3t_musicgen_riff_groove_builder-0.1.0-py3-none-any.whl.
The default build built the wheel from the real sdist. Index access disabled.
Persisted under dist:
- gaj3t_musicgen_riff_groove_builder-0.1.0-py3-none-any.whl
51,462 bytes; SHA256:
841DF0BF582257EEE3FA5A5F9165A41776937CC74D1726D342827FC897E9C0AE
- gaj3t_musicgen_riff_groove_builder-0.1.0.tar.gz
52,548 bytes; SHA256:
8AAE99D375BF34F62FCD4B98C91EBEA7C376A5DF8DCA873087FEE197F057B274
An additional archive inspection compared all 19 current Python/JSON source
files byte-for-byte to BOTH final distributions; all matched. The sdist README
matched the current README byte-for-byte. Wheel contents included no synthetic
fixtures, native files or manuals. Hashes identify these builds; byte-for-byte
reproducibility across build times is not asserted.
README validation additionally parsed all five fenced JSON examples: one MCP
configuration, two accepted requests, one success response, one typed error.
Requests passed strict models, profile selection, composition and compilation;
both ended at 6720 ticks. Success passed the published output schema and model,
including exact assumptions; error passed its model. Local documentation file
links were checked (the report link was completed when this file was written).
Current docs contain no active empty-profile/unfinished-generation claims.
Final persistence check confirmed exactly one plain report fence, every local
documentation file link resolving, unchanged final archive hashes and cleanup;
lint/format/strict types/pip check all passed again after report creation.
INSTALLED-WHEEL SERVER PROOF
Normal test procedure:
- Build real sdist and wheel offline from current source.
- pip install --no-index --no-deps --no-cache-dir --target <owned-target> <wheel>
- Launch the repository-venv Python with -I -B and an explicit bootstrap that
prepends the wheel target to sys.path, asserts all application import/spec
locations and distribution metadata there, then executes installed __main__
with --output-dir <existing-owned-output-root>.
- Working directory:
D:\_Repos\Code\GAJ3T\MCPs\MusicGen
This is the existing parent OUTSIDE the checkout. No file was created there.
- Dependency MCP import:
<repository>\.venv\Lib\site-packages\mcp\__init__.py
- Application/version metadata: non-editable wheel 0.1.0, not src.
Final focused test workspace (subsequently removed):
<repository>\.test-artifacts\distribution-fc21b3bebca9492c9b263bd561028085
Its installed server PID 33388 exited 0; 13 protocol stdout lines were checked.
To prove the EXACT persisted final wheel, not just a separately built wheel,
the same acceptance function was also replayed against a hash-identical copy
of dist's final wheel. Exact replay command:
@'
import asyncio
import hashlib
import os
from pathlib import Path
import shutil
import sys
import uuid
root = Path.cwd().resolve()
sys.path.insert(0, str(root.joinpath('tests', 'integration')))
from test_distribution import test_built_wheel_real_stdio_generation_outside_checkout
wheel = root.joinpath('dist', 'gaj3t_musicgen_riff_groove_builder-0.1.0-py3-none-any.whl')
workspace = root.joinpath('.test-artifacts', 'step10-final-wheel-' + uuid.uuid4().hex)
workspace.joinpath('dist').mkdir(parents=True)
scratch = workspace.joinpath('scratch')
scratch.mkdir()
env = os.environ.copy()
for name in ('PYTHONPATH', 'PYTHONHOME'):
env.pop(name, None)
env.update(TEMP=str(scratch), TMP=str(scratch), TMPDIR=str(scratch), PIP_NO_INDEX='1', PIP_NO_CACHE_DIR='1', PIP_DISABLE_PIP_VERSION_CHECK='1', PYTHONDONTWRITEBYTECODE='1')
try:
copied = workspace.joinpath('dist', wheel.name)
shutil.copy2(wheel, copied)
expected = hashlib.sha256(wheel.read_bytes()).hexdigest()
assert hashlib.sha256(copied.read_bytes()).hexdigest() == expected
asyncio.run(test_built_wheel_real_stdio_generation_outside_checkout((workspace, env)))
print('FINAL persisted wheel acceptance replay: PASS; SHA256=' + expected)
finally:
shutil.rmtree(workspace)
assert not workspace.exists()
print('Final-wheel replay workspace removed: ' + str(workspace))
'@ | & '.\.venv\Scripts\python.exe' -
PASS: exact persisted wheel SHA256 matches 841DF0BF...E9C0AE above.
Actual replay target:
<repository>\.test-artifacts\step10-final-wheel-d903ac35181845ceb73e70d8e2c5e04c\installed-server
Actual replay package path:
<target>\riff_groove_builder\__init__.py
Server/service paths:
<target>\riff_groove_builder\server.py
<target>\riff_groove_builder\service.py
All 16 asserted package/submodule entry locations:
__init__, __main__, server, service, schemas, errors, description, compose,
profiles, articulations, midi, formats.__init__, formats.agp, formats.sdp,
data.__init__, data.profiles.__init__.
None resolved into editable src. Both probe and actual server bootstrap
enforce these checks; this is not merely a separate import smoke test.
Actual wheel JSON members compared to installed bytes and loaded as resources:
general-midi-percussion-v1.json; standard-midi-bass-v1.json;
standard-midi-lead-v1.json. Exactly three resources, no synthetic/vendor profile.
Replay PID 46304 initialized through the real SDK, listed exactly
list_instrument_profiles and generate_music, listed exactly standard.bass.v1,
standard.gm-drums.v1 and standard.lead.v1, all standard_midi/native-unavailable/
normal-only. Schemas equal the authoritative contracts. Requests and all
profile/success/error structured outputs were independently schema-validated.
Actual request description, seed 42:
lead bass drums 137.5 BPM 2 bars 7/8 Bb minor quarter rests
Actual published files relative to that replay's server-outputs root:
- explicit\explicit.mid — ["mid"], complete, fallback_for=[], 3 mapping warnings.
- default\default.mid — formats omitted, midi_fallback, fallback_for all three
parts, 3 mapping plus 3 native-unavailable warnings.
- native\native.mid — ["agp","sdp"], the same explicit fallback/warning coverage.
Each actual disk file was 324 bytes, began MThd and had SHA256:
7d61b8d4a8c42a19091126ce7301b159744bbae4e78456b0cd1894f6269aea4b
Independent saved-file Mido parsing verified SMF type 1, PPQ 960,
137.5 quarter-note BPM encoded as 436364 microseconds/quarter at tick zero,
7/8 at zero, tracks Conductor/Lead/Bass/Drums, and one EOT per track at 6720.
Resolved tonality A# minor, exact profile IDs, parts, measured sizes and paths
matched actual artifacts. Standard channels were zero-based 0/1/9.
Note-on counts were 0/6/6/12; pitched onsets 0,960,1920,3360,4320,5280;
each drum onset had two hits. Pitched gates 960 ticks, drum gates 240;
every note was legal, positive-onset-velocity, non-overlapping and pair-complete.
Drum pitches were 36/38/42, with no program/bank/CC/keyswitch payloads.
Identical content across format policies is MIDI substitution, not native data.
The same process verified typed failures and unchanged output directory lists:
- Native-disabled drums -> NATIVE_FORMAT_UNAVAILABLE, sdp_layout_not_established.
- Native-disabled lead -> NATIVE_FORMAT_UNAVAILABLE, agp_layout_not_established.
- Lead palm mute under standard map -> UNSUPPORTED_ARTICULATION.
- Unicode unsupported description -> UNSUPPORTED_DESCRIPTION, no text leakage.
- Existing explicit basename -> OUTPUT_EXISTS; no new directory appeared.
No disabled/unsupported/Unicode directory or stage/native artifact was created.
Every captured server stdout line was newline-delimited UTF-8, parsed as JSON
and validated as MCP JSON-RPC; exactly 13 lines, no extra nonprotocol output.
Deliberate malformed Unicode input yielded only the SDK generic protocol
notification on stdout, separate sanitized diagnostics on stderr, and successful
ping recovery. Closing stdin reaped the initialized server with exit code 0.
A second installed-wheel child also exited 0 with zero stdout/stderr on EOF
before initialization. Reads/writes/shutdown have bounded waits; cleanup can
kill only the owned child and reap it if normal EOF fails.
ISSUES AND CLEANUP
Initial static checks found two wrapping findings, a loop-closure lint finding
and a test-local dictionary-type inference conflict. Refactored the schema-call
helper, explicitly typed heterogeneous failure options and formatted the changed
test. No production bug, weakened check, broad ignore or dependency change.
All final checks pass. A build-directory inspection found it already absent;
subsequent explicit cleanup checks confirmed that absence, not a build failure.
All owned UUID distribution/install/server/output/junction/test workspaces were
removed, including the final-wheel replay outputs after actual disk inspection.
The evidence paths above identify observed files, not retained MIDI downloads.
Only pre-existing scratch and step9-verified remain under .test-artifacts.
No .riff-stage-* directories or Python module-server processes remain; process
enumeration verified this after the focused test and replay. build is absent.
The repository venv/caches, prior step9 evidence and unrelated work were preserved.
Final dist wheel/sdist remain as intentional delivery artifacts.
No scratch scripts, background helpers, native binaries, secrets, manuals or
factory songs were added. No external/vendor/application work was fabricated.
REMAINING ORIGINAL NATIVE/VENDOR/APPLICATION BLOCKERS
1. Vendor MIDI maps: exact Eclips/Ray identity/build and SD3 build/library/preset/
selected map, real numeric ranges/voices/channels/keyswitch/CC semantics,
velocity layers and reset/timing evidence remain unverified. Standard maps
and synthetic compiler tests cannot close this gate.
2. AGP: actual content-bearing native layout/version/encoding, lawful controlled
empty/one-note/multi-note exports and isolated tempo/timing/velocity/technique
changes, source/fixture permission and true native round trips are missing.
Guitar and bass must be established independently; no shared layout inferred.
prompts\step_7_fix.md remains BLOCKED and was not executed or rewritten.
3. SDP: real song/groove-versus-preset content category, container/version/schema,
mapping/library dependencies, lawful controlled hit/variable-change exports,
permission and actual native round trips remain missing.
prompts\step_8_fix.md remains BLOCKED and was not executed or rewritten.
4. Applications: actual Eclips/Ray/SD3 native import/inspect/save/close/reload,
receiver routing, Cubase MIDI import/tempo/meter/alignment/dynamics/control
behavior and native/MIDI content equivalence are UNRUN. Exact versions,
workflows and evidence are unavailable. Mido opening a MIDI file is not
any of these application acceptances.
5. If a requested native extension cannot hold the musical content, that is
an unresolved scope contradiction requiring real evidence/decision, not
permission to rename MIDI, emit preset-only files or substitute formats.
SOFTWARE READINESS / STOP
Ready for use/review as the documented Windows Standard MIDI MCP generator on
the tested configuration, including real built-wheel execution. All independent
Step 10 software acceptance is complete and persistent in tests/docs/artifacts.
Not ready to claim the original native/application definition of done.
No next worker prompt, commit or delegated task was created. Stopped after
verified software/report completion, not merely because native gates exist.