Skip to content

TODO.md

Latest commit

 

History

History
56 lines (49 loc) · 2.83 KB

TODO.md

File metadata and controls

56 lines (49 loc) · 2.83 KB
Raw

Temporary Workarounds

Track repository workarounds that must be removed after an upstream dependency ships. Each entry must identify an upstream reference, removal condition, and cleanup validation.

Installed Adapter Descriptor Metadata

  • Status: Waiting for adapter wheel metadata support
  • Added: July 19, 2026
  • Affected implementation: crates/fabric-core/src/config.rs, crates/fabric-cli/src/presets.rs, crates/fabric-cli/assets/adapters/, and the Python binding
  • Reason: The published CLI crate currently carries package-local copies of adapter descriptors because Cargo cannot package descriptors from the repository-level adapter directories. A drift test keeps those copies aligned with the canonical descriptors, but the duplication is a packaging workaround. CLI presets stage the embedded descriptors under a temporary adapters/ directory, and core also probes a compile-time repository path.
  • Upstream resolution: NeMo Fabric adapter distributions advertise their descriptors through installed wheel metadata (internal packaging milestone; no external dependency).
  • Removal condition: Python can discover installed adapter descriptors from wheel metadata and pass them to core through an explicit, typed adapter registry.
  • Cleanup: Add adapter registrations to ResolveContext; remove implicit repository and base_dir/adapters discovery from core; have the Python binding register wheel-owned descriptors; revisit and remove the duplicated CLI descriptor assets once wheel metadata is consumable; retain embedded or staged files only for standalone CLI assets such as the scripted runner; and validate both installed-wheel and standalone Rust CLI behavior before removing this entry.

Planned Work

Example Authoring Skill

  • Status: Not started
  • Purpose: Add a guided workflow for users who need help designing a new example, selecting the closest preset or maintained example, customizing the generated code, and adding documentation and tests.
  • Boundary: Keep nemo-fabric example init as the deterministic operation for users who already know which example and variant they want to customize. The skill provides authoring judgment and validation around that operation.
  • Implementation: Have the skill invoke or reuse example init as its starting point, then modify and validate the generated application. Do not create a separate template collection or scaffolding implementation in the skill.
  • Completion condition: The skill can guide a user from an example idea to a runnable, documented, and tested Python or Rust example while keeping the CLI scaffold as the single source for generated starter code.