renato97/ableton-mcp-ai
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 1 Wiki Activity
Files
main
ableton-mcp-ai/docs/SPRINT_v0.1.38_NEXT_GLM_PROJECT_EDITING_AND_COHERENCE.md

11 KiB
Raw Permalink Blame History

SPRINT v0.1.38 - NEXT FOR GLM

Expand Open-Project Editing Tools And Add Real Coherence Auditing

Owner: GLM via OpenCode
Reviewer: Codex
Fecha: 2026-04-02
Mode: Edit existing project, not generate new song

Primary benchmark project:

  • C:\Users\ren\Desktop\song Project\song.als

Reference docs:

  • C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\docs\PROJECT_AUDIT_song_2026-04-03.md
  • C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\docs\SPRINT_v0.1.37_VALIDATION_REPORT.md
  • C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\docs\SPRINT_v0.1.37_NEXT_GLM_PROJECT_EDITING.md

1. Reviewer Summary

GLM did advance the editing runtime path, but the report in SPRINT_v0.1.37_VALIDATION_REPORT.md still overclaimed closure.

Codex reviewed:

  • the report
  • the active code
  • the live MCP state against the currently open benchmark set
  • the editing runtime and tests

The real outcome of v0.1.37

  • good progress
  • not complete

The main reason is simple:

  • the report claimed HARMONY_PIANO_MIDI had an arrangement clip
  • the live MCP state now shows HARMONY_PIANO_MIDI with:
    • session_clip_count = 0
    • arrangement_clip_count = 0

So the report cannot be treated as final truth for the actual open benchmark project.

2. What Codex Fixed After The Report

These fixes are already in the tree and are the new baseline.

2.1 MCP/runtime fixes

Codex fixed:

  • get_clips routing in both runtimes
  • track_type propagation for new editing commands
  • arrangement clip fallback via session-record-to-arrangement when direct track.create_clip(...) is unavailable
  • set_device_parameter(...) public API now supports parameter_name
  • public get_device_parameters(...) was added
  • audit_current_project() now reports repeated_clip_overuse

2.2 Product-level cleanup

Codex also fixed:

  • duplicate MCP tool definitions in server.py
  • broken create_clip(...) validation path that used nonexistent _validate_float
  • fragile side effect where arrangement fallback could leave tracks armed; it now preserves/restores arm state more safely
  • get_track_info(...) now carries more useful nested clip/device details for editing workflows

2.3 Current validated tests

Passing locally:

  • test_runtime_truth.py
  • test_selection_coherence.py
  • test_piano_forward.py

3. Live Truth You Must Respect

Codex re-checked the current project through MCP.

Current live facts:

  • Ableton is reachable
  • session is at 95 BPM
  • current project has 16 tracks and 6 scenes
  • HARMONY_PIANO_MIDI is still present as track 15
  • but it is still empty in Arrangement in the current live set

So for this sprint:

  • do not use the v0.1.37 report as your source of truth
  • use live MCP responses as the source of truth

4. Code Review Findings

4.1 Main report failure

The report treated a transient or unpersisted edit as if it were benchmark truth.

That is not acceptable for project-editing work.

If a report claims a live edit succeeded, you must prove:

  1. the edit exists in the open set
  2. MCP sees it
  3. it still exists after re-inspection

If not:

  • status is not COMPLETED

4.2 Editing tools are improving, but still not complete enough

The MCP now has useful editing surface:

  • get_clips
  • get_devices
  • get_device_parameters
  • create_clip
  • add_notes_to_clip
  • create_arrangement_clip
  • add_notes_to_arrangement_clip
  • duplicate_clip_to_arrangement
  • track mute/pan/send/device control
  • audit_current_project

But serious open-project editing still needs more public tools.

4.3 Coherence is still under-tooled for open projects

Right now the project audit sees structural issues, but not enough musical ones.

The system still lacks a first-class project coherence audit for an already-open song.

That is the next real product gap.

5. Product Direction For v0.1.38

We now have two parallel goals:

  1. give MCP more real edit tools for already-open projects
  2. make coherence measurable on the open project itself

This sprint is not about a new generation path. It is about:

  • editing a real .als
  • auditing coherence in that real .als

6. Important Semantic Rule

The user comes from FL and may say piano roll.

For this product, that must be interpreted as:

  • harmonic MIDI editing
  • note editing
  • arrangement MIDI support

It does not mean:

  • force piano timbre
  • add more piano sounds
  • make the song piano-forward

If the track is named HARMONY_PIANO_MIDI, treat it as:

  • a harmonic MIDI support lane

Do not turn this sprint into piano sound design.

7. P0 Tasks

P0.1 Re-validate the live benchmark truth

Before any new work, restart:

  1. Ableton
  2. OpenCode

Then reconnect MCP and inspect the live benchmark again.

You must verify:

  • get_session_info()
  • get_tracks()
  • get_track_info(track_index=15)

The goal is to start from real live state, not from old screenshots or stale outputs.

Required evidence:

  • exact responses

P0.2 Add missing public edit tools for real project work

The MCP still needs more low-level editing primitives exposed publicly.

Add public MCP tools in server.py for:

Session / clip editing

  • delete_clip(track_index, clip_index)
  • set_clip_name(track_index, clip_index, name)
  • set_clip_loop(track_index, clip_index, loop_start, loop_end, loop_length=0.0, looping=True)
  • fire_clip(track_index, clip_index)
  • stop_clip(track_index, clip_index)

These already exist or are very close in runtime support. Expose them cleanly.

Transport / editing support

  • jump_to(time)
  • set_loop(enabled)
  • set_loop_region(start, length)
  • show_arrangement_view()

Do not rely on shell or hidden helpers when MCP can own these editing actions.

Track inspection quality

Extend get_track_info(...) so it can optionally accept:

  • track_type="track" | "return" | "master"

and do not leave return/master inspection as second-class behavior.

P0.3 Add a first-class project coherence audit

Add a new public MCP tool:

  • audit_project_coherence()

This is not a generation manifest audit. It is an audit of the currently open live project.

Minimum required outputs:

  • longest_drum_gap
  • longest_harmonic_gap
  • tracks_with_zero_arrangement_clips
  • harmonic_midi_tracks_without_arrangement_clips
  • dominant_repeated_audio_sources
  • repetition_by_track
  • harmonic_coverage_ratio
  • drum_coverage_ratio
  • same_sample_overuse_flags
  • coherence_summary

The audit must use:

  • current track names
  • arrangement clips
  • device presence
  • real repeated source names

Do not fake this from manifests alone.

P0.4 Make the coherence audit useful for this actual benchmark

The benchmark project currently shows exactly the kind of issues we care about:

  • repetitive one-shots
  • arrangement islands
  • empty harmonic MIDI support
  • continuity depending too much on audio bass only

audit_project_coherence() must be able to make those issues visible.

At minimum it must highlight:

  • HARMONY_PIANO_MIDI exists but has no arrangement clips
  • AUDIO KICK, AUDIO CLAP, AUDIO HAT, AUDIO PERC MAIN reuse the same small set of sources heavily
  • AUDIO SYNTH LOOP and AUDIO SYNTH PEAK are sparse
  • the song still has continuity/coherence gaps despite looking more structured

P0.5 Validate real editing persistence

Pick one real edit on song.als and prove it persists across re-inspection.

Examples:

  • create a session clip on HARMONY_PIANO_MIDI
  • add notes to it
  • duplicate it to arrangement
  • rename the created session clip
  • inspect again

Required proof:

  1. action call
  2. immediate success result
  3. re-inspection result showing it exists

If re-inspection does not confirm it:

  • the edit does not count as validated

8. P1 Tasks

P1.1 Add arrangement-specific clip inspection

Current get_clip_info(...) remains session-slot-centric.

That is still too weak for serious arrangement editing.

Implement one of:

  • get_arrangement_clip_info(track_index, clip_start_time, track_type="track")
  • or extend get_clip_info(...) with a view argument

This tool should report:

  • clip name
  • start time
  • length
  • audio vs MIDI
  • loop state when available

Do not leave arrangement inspection ambiguous.

P1.2 Add a safe orchestration helper for project refinement

Only after P0 works.

Add a tool like:

  • refine_existing_project_section(...)

This must not be a fake black box. It must return:

  • actions taken
  • tracks touched
  • clips created/duplicated/edited

Good first scope:

  • fill one harmonic gap
  • extend one rhythmic section
  • reduce one repeated-source hotspot

Keep it bounded.

9. Non-Goals

Do not spend this sprint on:

  • new genre generation logic
  • reference remake quality
  • new piano timbre logic
  • vocal generation
  • cosmetic md writing

The focus is:

  • project editing tools
  • live project coherence visibility

10. Required Files

Primary:

  • C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\AbletonMCP_AI\AbletonMCP_AI\MCP_Server\server.py
  • C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\abletonmcp_init.py
  • C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\AbletonMCP_AI\abletonmcp_runtime.py
  • C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\AbletonMCP_AI\AbletonMCP_AI\MCP_Server\tests\test_runtime_truth.py

Benchmark:

  • C:\Users\ren\Desktop\song Project\song.als

11. Required Validation

Minimum local:

python -m py_compile "C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\AbletonMCP_AI\AbletonMCP_AI\MCP_Server\server.py" "C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\abletonmcp_init.py" "C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\AbletonMCP_AI\abletonmcp_runtime.py"
python "C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\AbletonMCP_AI\AbletonMCP_AI\MCP_Server\tests\test_runtime_truth.py"

Required live validation:

  • inspect benchmark project with MCP
  • run at least one real edit
  • re-inspect to prove persistence
  • run audit_current_project()
  • run audit_project_coherence()

12. Deliverable

Write:

  • C:\ProgramData\Ableton\Live 12 Suite\Resources\MIDI Remote Scripts\docs\SPRINT_v0.1.38_VALIDATION_REPORT.md

Must include:

A. Runtime truth

  • restart evidence
  • MCP connection evidence
  • benchmark loaded evidence

B. Tool results

  • exact call
  • exact output
  • pass/fail

C. Coherence audit results

  • exact output of audit_project_coherence()
  • interpretation against the benchmark project

D. Real edit proof

  • action
  • success result
  • re-inspection result

E. Reviewer conclusion

Only one of:

  • COMPLETED
  • PARTIAL
  • BLOCKED

If the live benchmark still contradicts the report, it is not COMPLETED.

13. Final Rule

Do not close this sprint because the MCP can call editing commands.

Close it only if:

  • the editing commands work on the open benchmark project
  • the effects are visible on re-inspection
  • the new coherence audit makes the benchmark's musical problems legible
Reference in New Issue View Git Blame Copy Permalink