Skip to content

ROADMAP.md

Latest commit

 

History

History
65 lines (48 loc) · 4.37 KB

ROADMAP.md

File metadata and controls

65 lines (48 loc) · 4.37 KB
Raw

BC ModelVault roadmap

This roadmap separates currently available capabilities from the work required to turn BC ModelVault into a complete AI asset manager.

Available today

  • Downloads from Hugging Face URLs or repository identifiers, including models, datasets, and Spaces.
  • Tokens for private/gated repositories, multiple profiles, and active-token selection.
  • Pause/resume during the current session and reuse of partial Hugging Face cache files.
  • Pre-download free-space checks.
  • Post-download file presence and size validation.
  • Downloads pinned to the verified commit.
  • Progress, speed, ETA, themes, Italian/English localization, and notification area integration.

P0 — Reliable foundations (implemented)

  1. Persistent download queue: queued/running/paused/failed/completed states, configurable worker count, and restart recovery. SQLite persistence with atomic writes.
  2. Reliable cross-restart resume: persist repository, commit, destination, and current files; reconcile cache and disk state at startup.
  3. Cryptographic integrity: obtain LFS OIDs/ETags or verifiable manifests, stream SHA-256 validation, and expose verified/unverified/mismatch states.
  4. Error log and history: rotating token-free logs, filterable history, and redacted diagnostic exports.

P0 is available in the application. Queue state is stored transactionally in ~/.bc-modelvault/modelvault.sqlite3; interrupted jobs are recovered as paused and reconciled with files on disk. SHA-256 is verified where Hugging Face publishes an LFS digest, while files without a trusted digest remain explicitly unverified.

P1 — AI asset catalog (implemented)

  1. Model metadata: size, extension/format, license, author, revision, and date. Display not declared instead of inventing missing data.
  2. Classification: LLM, checkpoint, LoRA, VAE, and ControlNet using Hugging Face tags, model_index.json, filenames, and manual overrides. Automatic classification must expose its confidence level.
  3. Local library: searchable index, filters, hash-based duplicate detection, integrity status, and Open folder through the operating system API.

P1 is available in the Tools → Local AI asset library window. Metadata uses declared Hugging Face values and displays not declared when absent. Automatic classification stores its confidence independently from manual overrides. Duplicate detection compares cryptographic file hashes rather than filenames.

P2 — ComfyUI integration (implemented)

  1. Multiple ComfyUI profiles: name, installation root, and folder mappings with structure validation and a default profile.
  2. Automatic installation: plan preview, routing to models/checkpoints, loras, vae, controlnet, and similar locations; skip/rename/replace conflict handling and atomic copies.
  3. Optional links: hard links or symbolic links to avoid duplicates after checking volume and permissions, with copying as the fallback.

P2 is available in the ComfyUI sidebar page. Profiles and custom mappings are stored in SQLite; the first profile becomes the default automatically. Every installation requires a visible plan. Replacement requires explicit confirmation, copies use same-directory temporary files plus atomic replacement, and unavailable hard/symbolic links fall back to copying while reporting the fallback count.

P3 — Product quality (implemented)

  1. Clean cancellation, retries with backoff, and bandwidth/concurrency limits.
  2. Catalog updates and revision comparisons without implicit overwrites.
  3. Integration tests, SQLite migrations, executable signing, and a checksum-enabled release pipeline.

P3 adds cooperative cancellation, bounded retry with exponential backoff, the existing aggregate bandwidth/concurrency controls, and non-destructive revision comparison from the local library. SQLite exposes a current schema version and performs compatible migrations at startup. The Windows release workflow runs tests, builds the executable, optionally signs it when certificate secrets are configured, generates SHA-256 checksums, uploads artifacts, and publishes tagged releases.

Minimum completion criteria

Every feature must provide persistent state where applicable, localized messages, automated tests, documented failure behavior, and no token exposure in logs. Destructive or overwriting operations must always provide a preview and require explicit confirmation.