v2.0.0

A complete, breaking rewrite. AudioCLI is repositioned from "ML data-prep tool" to "scriptable, batch-capable, cross-platform audio power-tool that runs DAW-quality effects from the command line." The engine is now Spotify's pedalboard; the v1 ML-flavored stack is gone.

Migration from v1

• Package rename: AudioCLI (capital A) → audiocli (lowercase). Update imports: from AudioCLI.modules.core.process import ... → from audiocli import run_per_file, AudioBuffer, op, ....
• Command flattening: the process / target / download grouping is gone. Every command is now top-level.
  • process resample 44100 → audiocli resample --target … --sr 44100
  • process mono -o → audiocli mono --target … -o (per-op flag)
  • target set <path> → in REPL: set targets <path>; in one-shot: --target <path>
  • target output <path> → in REPL: set output <path>; in one-shot: --output <path>
  • process file script.acli → audiocli run-script script.acli
  • process hook script.py → audiocli hook script.py --target …
• phaseflip renamed to polarity.
• v1 .acli scripts will not run on v2 — the command grammar changed. Update each line to the new shape.
• pip install audiocli now works on Windows (v1's readline install requirement broke Windows entirely).

Removed

• Torch and torchaudio entirely. The default install no longer pulls torch. The --pt save flag is gone. ML-flavored ops (RandPool, the noise-injection primitives) are removed; the new plugin system is the path forward for users who want them.
• librosa, scipy, aeiou, tqdm, termcolor, readline, icli, einops, pdoc are no longer dependencies.
• download http recursive scrape — punted to a possible future audiocli-plugin-http-scrape plugin.
• process file unreachable command path — its loader-confusion bug is fixed by the new run-script command.
• one_shot_args global state — replaced by explicit per-call run_per_file(...) parameters.

Added — engine

• pedalboard is the engine. All effects (Compressor, Limiter, HighpassFilter, LowpassFilter, Reverb, Delay, Chorus, Phaser, Distortion, Bitcrush, PitchShift, libsamplerate Resample), file I/O (WAV/FLAC/MP3/OGG/AAC via pedalboard.io.AudioFile), and VST3/AU hosting come from pedalboard's pre-built wheels. No system ffmpeg/sox required on any platform.
• AudioBuffer value type: (channels, samples) float32 ndarray + sr + subtype + optional format / quality overrides. Same shape pedalboard expects, so passing buffers in/out of pedalboard chains is a noop.

Added — pipeline

• Real parallelism. Single ThreadPoolExecutor, all files submitted at once, results consumed via as_completed so worker exceptions surface as Result(ok=False, error=…) instead of vanishing into an unconsumed iterator (the v1 batcher's bug).
• Per-file Result + end-of-job JobReport with ok_count, failed_count, failures, duration_s, exit_code (capped at 255).
• Cancellation. run_per_file accepts a cancel_token: threading.Event; flipping it from any thread skips pending submissions, lets in-flight files finish, and returns a partial JobReport.
• Bulletproof: a job over 1000 files where 50 are deliberately broken finishes, reports 950 successes + 50 failures with reasons, exits non-zero, and never hangs.
• Modular execution seams. Event dispatch, output path resolution, worker-count policy, and CLI rendering live in focused modules so the pipeline remains library-only execution code.

Added — CLI

• Flat command namespace. Every op lives at the top level (audiocli normalize, not audiocli process normalize).
• Consistent flags across every op: --target (repeatable), --output, --workers N, --recursive/--no-recursive, -o, --json.
• Lazy imports: audiocli --help returns in ~120 ms. numpy / pedalboard / pyloudnorm / platformdirs / click_repl / prompt_toolkit / rich only load when an op actually runs.
• Cross-platform REPL via click-repl + prompt_toolkit (replaces v1's Unix-only readline-based shell). Same Typer app powers REPL and one-shot; commands are identical. ; chaining preserved. History persists across sessions.
• --json event mode: every op emits newline-delimited {"type": "start" | "progress" | "file_done" | "error" | "done", ...} events on stdout. Same protocol the eventual desktop app uses to subscribe via on_event=.
• Persistent settings at the platformdirs user-config location (XDG on Linux, ~/Library/Application Support/AudioCLI on macOS, %APPDATA%\AudioCLI on Windows), JSON with a "schema": 1 field for future migrations.
• First-party command modules. Special-case commands (info, remove-silent, chunk) are direct Typer commands, with reusable analysis, silence detection, and chunking logic split into focused library modules.

Added — first-party ops

Group	Ops
Levels & channels	gain, normalize (peak / LUFS), polarity, mono, stereo
Time-domain	resample, pitch, trim, fade (linear / exp / cosine)
Format	convert (wav/flac/mp3/ogg), bitdepth (8/16/24/32)
Effects (pedalboard)	compress, limit, highpass, lowpass, reverb, delay, chorus, phaser, distortion, bitcrush
VST/AU	vst --plugin-path … --param key=value
Special-case (analysis / side-effect / multi-output)	info, remove-silent, chunk
Power-user	hook, run-script, shell

Added — plugin system

• Public @op decorator (from audiocli import op, AudioBuffer). First-party ops use the same decorator plugin authors do.
• Discovery via Python entry-points: [project.entry-points."audiocli.ops"] my-op = "my_pkg:my_func". AudioCLI loads them via importlib.metadata.entry_points(group="audiocli.ops") at startup.
• Conflict resolution: first-party wins on name collision; the conflict is logged at startup. Malformed plugin signatures raise PluginError at registration time.
• Plugin contract is filter-shape only in v2.0 ((buf: AudioBuffer, **params) -> AudioBuffer). The contract may broaden in v2.1+.

Added — library API

• Public surface: from audiocli import run_per_file, run_one, list_ops, get_op, AudioBuffer, op, Op, OpInfo, ParamInfo, AudioCLIError, LoadError, SaveError, OpError, PluginError, ConfigError.
• list_ops() returns frozen dataclasses describing every registered op (name, help, kind, params with type/default/help/required) — designed to drive dynamic UI in a future desktop frontend.
• No print() / sys.exit() in any library module. The CLI layer renders; the library raises.

Added — error model

Public exception hierarchy:

AudioCLIError
├── LoadError
├── SaveError
├── OpError
├── PluginError
└── ConfigError

Tooling

• Ruff lints + formats. ruff check . && ruff format --check . is the full lint pass.
• 313 tests covering every op (round-trip + correctness + negative), the pipeline (50-files-with-3-corrupt isolation, cancellation mid-batch, no-thread-leak), the CLI (every command's --help, --json JSON parseability, exit codes), the REPL (;-chaining, history, set persistence), the plugin loader (entry-point discovery, conflict resolution, malformed signature → PluginError), and the library API (list_ops() shape, run_per_file callable surface).
• CI matrix: {ubuntu, macos, windows}-latest × {3.10, 3.11, 3.12} running ruff check, ruff format --check, pytest --cov.