Implement HTTP Range requests for resumable model downloads

[Copilot]

Interrupted model pulls now resume from the last downloaded byte using HTTP Range requests. When ./model-cli pull ai/smollm2:135M-Q4_0 is interrupted and re-run, it continues from where it left off instead of restarting.

Implementation

Resumable layer wrapper (pkg/distribution/distribution/resumable.go)

• Wraps v1.Image and v1.Layer to intercept layer downloads
• Detects existing .incomplete files and their byte offsets
• Issues HTTP requests with Range: bytes=offset- header
• Falls back to full download on HTTP 416/200 responses

Distribution client (pkg/distribution/distribution/client.go)

• Stores HTTP transport for Range request authentication
• Wraps remote models with resumableImage before store write
• Parses references to construct registry blob URLs

Store layer (pkg/distribution/internal/store/blobs.go)

• Opens .incomplete files in append mode when detected
• Writes resumed data directly to existing partial files

Behavior

First attempt (interrupted):
  GET /v2/repo/blobs/sha256:abc... → 50MB downloaded → ^C

Second attempt (resumed):
  GET /v2/repo/blobs/sha256:abc...
  Range: bytes=52428800-         → remaining 450MB downloaded
  HTTP 206 Partial Content

Gracefully degrades to full download when Range requests unsupported.

Original prompt

Integrate HTTP Range requests for resumable downloads

There's part of the reference implementation in:

the reference implementation in @distribution/pull_v2.go

When someone executes:

./model-cli pull ai/smollm2:135M-Q4_0

And that pull gets interrupted. If we pull again, it should resume where we left off.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.