bouch-audio

A portable, evidence-tiered audio domain capability for agents: reasoning for electronic production, mixing and mastering, and audio verification. It includes two tested deterministic tools. It does not depend on any DAW.

Registry identity: dev.bouch/audio. Status: experimental (v0.2.0).

• Three Skills that carry reusable reasoning and route into references:
  • electronic-production: building material that doesn't exist yet (sound design, bass/drums, arrangement, movement, resampling).
  • mixing-and-mastering: balancing and diagnosing material that exists, and what mastering can and cannot fix.
  • audio-verification: proving technical properties by measurement, and routing perceptual and musical judgement to a human listener.
• Shared references (references/): detailed knowledge in which every claim is tagged PROJECT-PROVEN, SOURCE-BACKED, BOTH, PARTIAL, OPEN, HYPOTHESIS, VERDICT PENDING or NOT YET COVERED. Start with references/evidence-status.md.
• Tools (tools/): analyze.py (BS.1770 loudness, true peak, clipping, stereo, spectrum, windowed analysis) and sample_op.py (manifest-verified slice/arrange/reverse/repitch). See tools/README.md.
• Evidence (evidence/): source pins, the quote and licence audit, evidence-tier changes, exclusions and the qualification record.

• Not a DAW adapter. No REAPER, MCP or plugin operation, project state or render scripts. A future adapter should route into these references, not copy them.
• Not a sample library, render corpus or plugin inventory.
• Not a survey of audio production. It covers what Bouch experiments actually exercised or sourced; the gaps are listed, not filled.
• Not an authority on taste. Measurement proves technical properties; listening decides musical ones.

bouch-audio/
├── plugin.json                 # Agent Plugins 1.0.0 manifest (canonical identity)
├── .claude-plugin/plugin.json  # Claude adapter manifest (same identity fields)
├── skills/
│   ├── electronic-production/SKILL.md
│   ├── mixing-and-mastering/SKILL.md
│   └── audio-verification/SKILL.md
├── references/                 # shared, linked from Skills by relative path
├── tools/                      # analyze.py, sample_op.py (+ README)
├── tests/                      # pytest suites for the tools
├── evals/                      # claude plugin eval: activation / non-activation / traversal
├── evidence/                   # PROVENANCE.md, qualification.md
└── scripts/validate-package.sh

There is no root SKILL.md. Skills are discovered at skills/<name>/SKILL.md. The root references/ directory is a working pattern shared with bouch-agent-core; it is not a spec-defined location, and each Skill reaches it through explicit ../../references/… links.

Claude Code, for one session:

claude --plugin-dir /path/to/bouch-audio

Other Agent Skills clients: point the client at skills/, keeping the repository tree intact so the ../../references and ../../tools links resolve. The tools need Python ≥ 3.10 with numpy and scipy, or uv run.

scripts/validate-package.sh                                  # manifests, skills, links, paths, tags, tests
claude plugin eval . --ablation none --runs 3 --trust-plugin # activation evals (model cost)

The current qualification record is evidence/qualification.md.

MIT (see LICENSE). Third-party techniques are paraphrased and cited by title and publisher; no third-party text, audio or renders are included. See evidence/PROVENANCE.md.