v0.5.5

Changes

• Fix Docker startup without internet access. The image now installs the complete application
  during the build and starts with uv run --no-sync, so startup does not try to fetch build
  dependencies from PyPI. Container tests cover memory creation and search, restart, and
  recreation with networking disabled and prepared model caches.
• Add an optional external_ref to plans. Use it to identify a plan by an external issue or
  job reference through MCP, REST, or the CLI. References are unique across each user's plans;
  duplicate writes return a conflict, and list_plans can find the existing plan by reference.
• Accept file_ids in MCP create_memory and update_memory. Updates replace the linked
  files, and an empty list clears them. Fixes #76.
• Add get_memory_entities to retrieve all entities directly linked to a memory, including
  their names and types. Fixes #78.
• Add EMBEDDING_ONNX_PROVIDERS and RERANKING_ONNX_PROVIDERS for local FastEmbed models.
  These let you select execution providers, including hardware acceleration when a suitable
  ONNX Runtime package is installed. Empty values keep the existing provider selection.
• Add Nix support for a development shell, application package, and container image, with
  x86_64-linux, aarch64-linux, and aarch64-darwin configurations. See the Nix guide.
• Fix Windows test-server shutdown, correct plan-status documentation to use archived,
  and add an architecture overview and regression coverage for the new behaviour.

Upgrade notes

• The normal startup migration adds a nullable external_ref column and a unique index to
  plans on SQLite and PostgreSQL. Existing plans keep an unset reference.
• Plan references are case-sensitive, trimmed, and limited to 255 characters. Omitting the
  field or passing null during an update leaves the existing reference unchanged.
• Offline Docker use still requires cached embedding/reranking models and tokenizer data.
  Keep those caches in persistent storage when recreating containers. See the
  offline setup guide.
• Accelerated ONNX runtimes must be installed separately. Every configured provider must be
  available in that runtime. DirectML reranking requires RERANKING_WORKERS=1.
  See the configuration guide.
• No embedding rebuild is required. Python dependency versions are unchanged.

Contributors

Thanks to @bertheto for ONNX provider configuration (#79), memory file attachments
(#80), reverse entity lookup (#81), and the Windows test fix (#87).
Thanks to @twstagg for Nix packaging and platform support (#84, #85).

Full changelog