DeepSeek Harness has released v0.1.3-alpha.1, adding arbitrary-file uploads to the Web interface while reshaping the session machinery underneath. The practical headline is simple: files and images can now share a preview area, uploads expose progress and cancellation, and models can read saved files through existing tools when needed.
DeepSeek Watch analysis
What happened
The 4 September release also adds a macOS x64 runtime wheel to the Python SDK, renders images directly in Web tool cards and improves model discovery for custom providers and Anthropic-native model lists.
The less glamorous work may prove just as important. Session persistence now uses lifecycle-scoped SessionHandle objects, creating an agent loop is asynchronous and a new lock ensures that one session is held by at most one process. The session format also moves to v2, with migration support for older logs.
Our top picks
- Files join the chat
Upload arbitrary file types, track progress, cancel transfers and keep them visible while switching conversations. - Intel Macs get a wheel
The Python SDK can install and run its bundled runtime on macOS x64 machines. - Images render properly
readimage results now appear directly in Web tool cards instead of as raw attachment objects. - Model discovery gets broader
Custom providers and native Anthropic model lists can supply model names, context windows and output limits. - Sessions get a bouncer
Per-session locking prevents multiple processes from taking control of the same session. - Old logs get a route forward
v0 and v1 session logs can migrate through immutable adjacent generations into the v2 format.
Why it matters
This is a sizeable quality-of-life release disguised as an alpha patch. Bringing files into the Web workflow and making tool output more legible removes friction at exactly the points where agent sessions tend to become fiddly.
The architectural changes point to a more dependable runtime too. Session ownership, resumable streams and durable settlements are not flashy features, but they are the difference between an agent system that demos well and one people can trust with longer-running work.
Our read
Worth trying for developers already living in DeepSeek Harness, particularly if file-heavy workflows or Intel Macs are part of the setup. Treat it as an alpha, though: the project itself flags a performance regression affecting some historical session loads.
What to watch
- Whether the known session-loading slowdown is fixed in the next release.
- How reliably uploads resume, cancel and remain available across conversation switches.
- Compatibility for integrations moving to asynchronous agent-loop creation and SessionHandle-based persistence.
- Whether model metadata backfilling works consistently across custom providers.
Discussion spark: Which change would make DeepSeek Harness more useful in your workflow: file uploads, stronger session persistence or broader model discovery?
Sources and evidence
- v0.1.3-alpha.1 (4 September 2026, 11:34 UTC)
not affiliated with or endorsed by DeepSeek