Skip to content

Changelog

0.3.1

  • geyser login works again with browser sign-in. The server no longer sends a pre-filled approval link (so a device code can't be used to trick someone into approving it), and the CLI crashed without one. It now opens the approval page, where you type the code it shows.

0.3.0

  • Call your workspace’s own models, such as one you taught, from your code with the official OpenAI or Anthropic SDK. New models:infer credential scope, client.list_models(), openai_base_url() / anthropic_base_url() helpers, geyser models list, OpenAPI routes under /api/v1/openai and /api/v1/anthropic, and a Use your own models guide. Requires the matching Cell release.

0.2.0

  • Current membership, Agent assignment, session audience, project and Cell generation are rechecked for developer access.
  • CLI profiles bind credentials to the issued API URL. Exact loopback checks reject deceptive HTTP hosts.
  • Publisher trust is administrator-owned; Cell and Agent verify exact Sigstore archive bytes and identity. Installation becomes active only after real sandboxed tests and a current acknowledgement.
  • Owned JSON inputs, typed results and an Agent task consumer replace metadata-only task acceptance. Admission preserves capabilities, budgets and outcome contracts; expiring claims fence stale workers.
  • Sync/async evaluation, cancellation and fork contracts match the server. Event watching streams immediately and drains terminal events; retry delays honor server rate limits.
  • Unknown emulator effects require reconciliation. Scaffolds execute real JSON handlers. New issue-normalization, source-review and Agent-review applications show useful workflows.
  • Console project/credential/publisher setup and explicit revocation are available with the matching server rollout. OpenAPI is exported from actual server routes.
  • Signed skills, model-selection assertions and Agent Bundles apply to individual qualified Open tasks, with explicit reference omissions and no shared-settings changes.
  • Remote replay creates an atomic project-owned task with model-only, exact-stub, workspace-read-only or fresh-authority full execution.
  • Approval-required tasks pause before actual writes; rejection resumes without execution. Approval responses expose action digests and the SDK distinguishes run and approval versions.
  • Agent HTTP runtime dependencies move to corrected httpx2/httpcore2 releases. Undispatched inspection forks and unavailable native import formats are labeled explicitly.

Read migration details. Publishing SDK artifacts and rolling out Cell/Agent software are separate release actions; source availability does not assert either has occurred.

0.1.0 — published

Initial public Python SDK/CLI, standalone CLI, Homebrew distribution, emulator, schema and documentation release. It does not include the 0.2.0 corrections above. Existing release bytes remain immutable.