Skip to content

v0.8.0 — streaming uploads, durable-capture consolidation, Harbor trial capture, 3-tool MCP

Choose a tag to compare

@richardwei6 richardwei6 released this 24 Jul 01:59

CLI/SDK probe-research 0.8.0 — streaming uploads, durable-capture consolidation, Harbor trial capture, and the 3-tool MCP surface.

Highlights

  • Streaming artifact uploads (#48): large artifacts (model weights, eval sets) stream to storage instead of being read whole into memory, so probe artifact add / trial * no longer OOM on big files. Content-Length is measured from the file so header and body can't disagree; retries re-open and resend from byte 0.
  • Large artifacts by path reference (#49): file:// reference registration for artifacts that live on a shared volume, instead of uploading them.
  • Durable-capture consolidation (#48): one stdlib-only probe.sdk.durable (atomic write / dir-fsync / flock / JSON read) shared by spool, capture, harbor, harbor_export, and Miles; PEP 562 lazy package init so a distributed actor imports no httpx to write metric batches to disk.

Harbor trial capture (SDK-owned)

  • Durable capture ledger + write-ahead spool (#37)
  • Own the producer staging contract in the SDK; ingest staged trial-capture descriptors; bind captures after offline run resolution
  • Durable Miles metric adapter
  • docs(harbor): SDK-owned producer handoff

MCP

  • Consolidated to three tools, rewritten instructions, added the eval (#35)
  • research_browse — enumerate structure without a query (#34)

Client & surface

  • Run-scoped artifact inheritance + vertical move (#47)
  • Report CLI and plugin versions (#45); SessionStart hook nudging on a stale CLI/plugin (#40)
  • Tag /v1 requests with X-Probe-Surface (cli/sdk/mcp) + X-Probe-Tool (#44)
  • Restore device pairing (tap pair/revoke) (#33)

Release / plugin plumbing

  • One-command chat-triggered releases (#43); also release the plugin (plugin.json + manifest) (#46)
  • Plugin bumps 0.5.0 → 0.6.0 (#36) → 0.7.0 (#41)

Note: 0.8.0 was cut from the accumulated main and includes work from several contributors (Harbor trial capture, the MCP surface rewrite, and the streaming/consolidation).

Full changelog: v0.7.1...v0.8.0