SonicCraft v0.1.0

First release: timeline DSL v1.0 and engine v0.1.

Added

• Timeline DSL v1.0 (YAML 1.2 or JSON): output and mastering settings, subtitle layout, timing policy,
  mix reference level, voice defaults, a speakers registry, provider instances, and speech, music and sfx
  tracks with cues, pauses and clips. Values accept unit strings ("350ms", "1.5s", "1m30s",
  "00:01:30.500", "320k").
• Validation with agent-first diagnostics: stable codes, field.sub[index] paths, line/column for YAML,
  did-you-mean hints for unknown keys, enum values, references and providers; every problem reported at
  once; secrets in timelines rejected.
• Timing: sequential placement with pauses, absolute starts, and relative anchors (after_cue,
  with_cue, before_cue, signed offsets) on cues, clips and tracks; dependency-cycle detection; a
  sample-accurate layout engine with the speech-overlap policy (push by default, or warn, error,
  allow) and per-cue allow_overlap.
• TTS providers: ElevenLabs (word timestamps), OpenAI-compatible /audio/speech servers, Cartesia
  (word timestamps), Kokoro via Kokoro-FastAPI (word timestamps), Piper, macOS say, custom commands and a
  deterministic mock voice; content-addressed cache; concurrency limits; retries with backoff; credential
  checks; voice listing; dry-run duration estimates in characters per second.
• Native DSP: decoding, resampling, silence trimming, per-source loudness normalization, loops with
  optional seam crossfades, fades (equal-power, linear, exponential), clip crossfades, constant-power
  panning, sidechain and cue-driven ducking with attack, release, hold and lookahead.
• Mastering: EBU R128 / BS.1770 loudness normalization with presets (podcast -16, streaming -14,
  broadcast -23, atsc -24 LUFS) and an iterative, 4× oversampled true-peak limiter.
• Encoders: WAV (16/24-bit integer, 32-bit float), MP3 (LAME, CBR; cargo feature mp3, on by default),
  FLAC (16/24-bit); ogg, opus and m4a through ffmpeg; fixed-seed TPDF dither.
• Subtitles: WebVTT, SRT, JSON with per-word timestamps, and a HyperFrames-compatible transcript;
  speaker labels; caption overrides.
• Side files: per-track stems, a deterministic manifest, and a self-contained HTML preview (waveform,
  ducking envelopes, captions).
• CLI soniccraft: validate, plan (dry-run), render, schema, init, providers, voices,
  cache, mcp; --json output and exit codes 0/1/2/3.
• MCP server (soniccraft mcp, soniccraft-mcp) serving both protocol eras: tools
  soniccraft_validate_timeline, soniccraft_dry_run, soniccraft_render, soniccraft_get_schema,
  soniccraft_list_providers, soniccraft_list_voices.
• Agent Skill skills/soniccraft with references and five templates; Claude Code plugin and marketplace
  manifests; Codex skill metadata.
• Examples (tech-briefing, two-person-podcast, offline-demo) with procedurally generated,
  royalty-free assets and the script that generates them.
• Dual MIT OR Apache-2.0 licensing; CI on Linux and macOS.
• A container image, ghcr.io/user-github-me/soniccraft (linux/amd64 and linux/arm64, with ffmpeg), published
  from each release's Linux binaries; docker build . builds it from source.

Added in pre-release hardening

• Release workflow: prebuilt soniccraft and soniccraft-mcp for Linux (x86_64, arm64), macOS (Apple
  silicon, Intel) and Windows (x86_64), each with and without the LGPL MP3 encoder, with SHA256SUMS and a
  THIRD-PARTY-LICENSES.txt in every archive.
• The soniccraft-mcp npm launcher (npx -y soniccraft-mcp): runs soniccraft mcp, downloading the
  matching release on first use and verifying it against checksums pinned in the package.
• licenses/ with the LAME and LGPL/GPL texts and the steps to relink a modified LAME; a cargo-deny license,
  ban and source policy (deny.toml).
• CI: the minimum Rust version, Windows (build and smoke render), cargo-deny, rustdoc, and the npm packages
  on Node 20 and 24.
• New diagnostics: comment_in_text (an unquoted # cut a cue's text short), suspicious_value (seconds
  written as milliseconds; {text}/{voice_id} as a command argument of its own), speakers_share_voice,
  unknown_voice (macOS voice not installed, piper model not found, already in validate),
  track_anchor_out_of_order, timeline_too_long (over 24 h, or over the render memory budget),
  loudness_target_missed, true_peak_exceeded, cache_write_failed, audio_repaired,
  channels_downmixed.
• Render memory budget: SONICCRAFT_MAX_RENDER_GB (default 8); a render that would need more stops before
  any TTS request, and validate/plan warn.
• SONICCRAFT_ALLOWED_HOSTS: hosts besides each provider's own API host and servers on this machine that may
  receive API keys.
• MCP: detail: auto|summary|full on validate, dry run, render and get_schema (compact reports above about
  12,000 characters, whole results under about 40,000; full never turns into -32603); a format argument on
  render; at most two concurrent renders (SONICCRAFT_MCP_MAX_RENDERS), queued renders report progress and
  can be cancelled; 2026-07-28 results carry _meta serverInfo.
• CLI: soniccraft mcp --allow-command / SONICCRAFT_ALLOW_COMMAND=1.
• RenderReport.loudness.pre_encode: the master that went into a lossy encoder.
• Sources with more than two channels (5.1, 7.1) are mixed down with ITU-R BS.775 instead of being
  unusable.
• Docs: a Security and a Resources section in the README, a Setup section in the skill (install, verify,
  failed MCP server), Codex approval settings for codex exec.
• Library APIs: soniccraft_core::parse::secret_spans, parse::MAX_NODES, values::parse_duration
  (DurationError), values::MAX_BITRATE_BPS, spelling_of on every name enum, Compiled.assets
  (AssetRef), ir::checked_ms_to_frames, diagnostics::display_path and human_duration; layout
  Layout::estimate_dependent, depends_on_estimates, diagnostics_with_estimates, has_certain_errors;
  dsp downmix, master::master_with_cancel, limiter::limit_in_place, loudness::measure_file,
  measure_with_cancel, duck::cue_regions, wav::read_repaired, new DecodeInfo fields; tts
  render_mock, resolve_macos_voice, resolve_piper_model, probe_reachable, key_host_allowed,
  settings_keys, command_argument_warnings, error_codes, TtsOptions.child_env_remove,
  SynthesisResult.cache_write_error; engine MAX_TIMELINE_BYTES, TimelineSource::from_reader,
  LoudnessReport.pre_encode, working EngineOptions.allow_command and confine_outputs_to; mcp
  ServerOptions::max_concurrent_renders, ALLOW_COMMAND_ENV, MAX_RENDERS_ENV.

Behaviour settled in pre-release hardening

• enum spellings are exact (Speech, MP3, equal-power are invalid_value with a
  did-you-mean); duration units are lowercase and clock seconds must be below 60; bitrate strings take no
  sign or exponent; version: "01.0" is invalid; padded ids (" v ") are invalid_id.
• every time is at most 24 h (larger values, usually epoch milliseconds, are invalid_value),
  and a timeline longer than 24 h is timeline_too_long; target_lufs must be in [-60, -5] (was
  [-70, -5]); a total length of 0 is invalid_value.
• a misspelled speaker (a close match of a defined speaker) is an error; voice keys inside
  settings (settings.voice) are a field_not_applicable warning.
• MCP servers refuse command providers (and piper with its own command) unless started
  with --allow-command; a timeline's own output.path must stay inside the MCP base directory.
• API keys go only to the provider's own API host, to servers on this machine (only the
  provider's own key variable), or to SONICCRAFT_ALLOWED_HOSTS; regional, remote and other local-network
  endpoints must be listed, and so must a local server that should get a variable named by api_key_env.
• command and piper cache keys include the program, its content hash, script files and the
  piper model, so those cues re-synthesize once.
• soniccraft_core::Compiled has a new field assets (build it with
  ..Compiled::default()); DspError has a new variant Cancelled; the name enums' parse is exact
  (case-sensitive, no trimming), so builtin_type("Eleven-Labs") is None; Layout.total_frames is capped
  at 24 h.
• JSON input has the YAML limits (250000 values, 64 levels) and diagnostics for comments and trailing
  commas; stdin/MCP auto-detection no longer reads such JSON as YAML.
• Hints convert values (44.1k → 44100, 50% → 0.5, target_lufs: 16 → -16, youtube →
  streaming, start_at_ms: a2 → start_relative: {after_cue: a2}) and know common key synonyms (file,
  voice_name, start_after, pause_ms), with unit notes for bare numbers.
• secret_in_timeline also covers annotation keys (x-api-key), keys inside provider strings (Bearer,
  --api-key=, URLs) and URL passwords; diagnostics never repeat a secret, and CLI source snippets show it as
  <redacted>.
• The JSON Schema encodes conflicts, applicability by track type, time sign and length rules, and
  null-as-absent.
• Renders commit their outputs atomically: a cancelled or failed render never deletes or half-writes
  earlier outputs, and files that are not SonicCraft outputs are never overwritten.
• Lossy outputs (mp3, ogg, opus, m4a) are decoded and measured as delivered, with codec true-peak and
  loudness-drift compensation (at most 3 passes); the report's loudness describes the delivered file.
• Mastering brackets the gain and keeps the pass closest to the target; an unreachable target gives the
  loudest master within 6 dB of limiting, the same for every target out of reach.
• Dry runs render mock voices exactly; ducking regions are computed from intervals; layout errors that depend
  on estimated durations are warnings in validate and plan; a render...