Skip to content

v0.7.1

Choose a tag to compare

@github-actions github-actions released this 07 Aug 13:12
· 116 commits to main since this release

First public release, shipping the a2wave CLI to npm and multi-arch container images to GHCR.

  • Agent orchestration platform: create, configure and orchestrate Agents in natural language. Execution capability comes from the underlying Agent CLIs (Cursor Agent / Claude Code / Codex); the platform does not intervene in an Agent's runtime decisions.
  • Publish Agents over many channels: API, Feishu, Slack, Discord, A2A, schedules, a built-in chat page, and GitLab / GitHub repository triggers — repository triggers start a run only when a watched merge/pull request actually moves, instead of polling unconditionally.
  • Extend through Skills and MCP Servers: compose Agent capability from Skills and MCP Servers (stdio / sse / http / group). A Skill is creator-private by default; an administrator may publish it to all users.
  • Providers and model discovery: models are probed per bound credential rather than kept as a static catalog that drifts from what the account can actually run.
  • Agent CLIs installed at runtime: the image preinstalls no Agent CLI. They are installed on demand from provider-cli-lock.json with pinned versions and SHA-256 verification, cutting over 1GB from the image; installation state is always probed from PATH, never trusted from the database.
  • Evaluation: replay Case sets against an Agent's current configuration, freezing a provider / model / prompt snapshot for comparison across versions. The evaluation queue is isolated from the run queue, so a large evaluation cannot starve interactive chat.
  • Knowledge bases and workspaces: connect Git and Perforce sources; runs and evaluations execute in isolated workspaces.
  • Enterprise capabilities: password and SSO login, per-Agent owner / editor / viewer permissions, audit logging, rate limiting, and health/readiness probes.
  • Database backends: SQLite single-file deployment by default. PostgreSQL is experimental, aimed at multi-instance deployments, not yet recommended for production, and has no SQLite → PostgreSQL data migration path.
  • In-app user manual: a manual at /wiki, with a bilingual (Chinese / English) interface.
  • Documentation and CI: streamlined README with configuration extracted into its own reference; the dependency license inventory check is now host-independent, and secret scanning uses the repository's own ruleset.