Skip to content

v0.26.0

Choose a tag to compare

@tintinweb tintinweb released this 16 Nov 17:27
· 15 commits to master since this release

What changes for existing users

  • A workflow starts only when you ask for one (#451). "Use 10 agents to …" now runs parallel Agent calls; say "use a workflow" or name a saved workflow to run one.
  • agent({ schema }) is validated strictly against draft-07 (#455). Schemas with unknown keywords (prefixItems, x-*) or a non-draft-07 $schema are refused before the agent starts.

New

  • Workflow progress in FleetView and the inspector: a progress bar, active phases, each agent's tool calls, and full prompt/outcome (#455).
  • Workflows report why an agent() call failed via agentError(label), the journal, and <failures> in the completion notification (#439, #444 — thanks @woksin, @Swarneil1905).
  • workflowToolExposure: "skill" moves the workflow script reference into a workflow-authoring skill, saving about 4k tokens per session.
  • workflowToolExposure: "deferred" keeps SubagentWorkflow out of the tool spec until pi's tool_search loads it (#434, #437 — thanks @M1racleShih).
image image

Changed

  • Workflow scripts behave as Claude Code's do: strict mode, copied return values, pipeline() stops an item on null, plus setTimeout, disallowedTools and case-insensitive effort. See Coming from Claude Code.

Fixed

  • Running workflow agents show live tool-call and token counts (#298, #453 — thanks @cgngtr).

Thanks @woksin, @Swarneil1905, @M1racleShih, @cgngtr.

Full details in CHANGELOG.md.