Skip to main content

Server 0.1.9

2026-09-01 · Tag server/v0.1.9 · 当前设备

New Features​

  • Standard Workflow MCP endpoint: Streamable HTTP exposes entrypoints as tools and reuses the existing run queue, persistence, and audit stack.
  • Conversation OpenAI gateway: a bearer-only /v1 serving Responses, Conversations, and Models subsets with per-project keys.
  • Related project KV grants: mutual or one-sided confirmation with read/write or prefix-scoped grants that take effect as soon as they are revoked.

Breaking Changes​

  • Global KV, the old workflow scope, kv.workflow, persistentValue.workflow, and standalone data grants are removed. Migration: KV and PersistentValue keep only project and conversation scopes, isolated per storage type.

Added​

  • Added a standard Streamable HTTP MCP endpoint for Workflow projects, authenticated with account or admin keys.
  • Added a bearer-only OpenAI Responses gateway and project key management for Conversation, isolating resources per user and project.
  • Added wfpk_* full keys shown only once on creation or rotation, with old keys invalidated immediately.
  • Added run records written when an OpenAI Responses call starts and updated in place on completion, failure, cancellation, or restart.
  • Added KV read, write, and compare-and-set APIs for the current and related projects, authorized against both latest declarations.
  • Added dedicated project storage, revision, and related-project dependency tables, leaving legacy tables, functions, and views untouched.

Changed​

  • Changed Kanban bridges, project group locks, downloads, and releases to resolve only confirmed related projects.
  • Changed the OpenAI Responses gateway to report output limits and content filter end states as incomplete while keeping partial output.
  • Changed Conversation deletion to keep its Items, with transactional locks around image reuse, deletion, and expiry cleanup.
  • Changed persisted incomplete responses to be resumable through previous_response_id.

Fixed​

  • Fixed production builds inlining the OpenAI Responses Shared runtime instead of leaving TypeScript-only packages in deploy dependencies.
  • Fixed MCP elicitation returning invalid content or failing transport or recovery no longer leaving waiting_for_input runs behind.
  • Fixed OpenAI call metadata trusting only STEAMHARNESS_API_PUBLIC_BASE_URL or forwarded headers from a trusted proxy.
  • Fixed partial image upload failures now returning explicit validation errors and reclaiming the partial files.
  • Fixed the gateway accepting system and developer messages, text.verbosity, and the compatibility include parameter.