You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
Feedback loops: input.feedback: { from, when, max } sends a later step's output back to an earlier step and runs the steps in between again while that output still matches when, for at most max extra rounds (default 3, up to 10). A loop also ends when its output repeats, a run fails before it would pass 200 step runs, and the run summary says why each loop ended. The editor draws loop-back edges and edits their limits.
Decision steps (type: decision) answer typed questions (choice, yes_no, score) about the upstream outputs and can route on the answers the way a branch does. With TYPESAFE_API_KEY set, TypeSafe's Jev answers with calibrated probabilities; without it, the step's opencode agent is asked for the same probabilities. for_each asks the same questions about every item of a list. The editor adds and edits decision steps.
Eight workflow templates in workflows/templates (review your changes, fix a bug, add tests, research a question, draft release notes, triage GitHub issues, explain a codebase, turn notes into action items) and a Templates view in the editor that opens one as a new workflow.
An AI chat beside the canvas: describe a change, and a local opencode server (opencode 1 or 2) answers with a complete workflow. The proposal shows the steps and connections it adds, removes or changes, its validation problems and the YAML; Apply puts it on the canvas and Undo puts the previous workflow back. The chat's notes on the file format cover feedback loops, decision steps and the Claude Code, Codex and omo mappings.
agentry-mcp runs workflows on Claude Code or Codex as well as opencode. run_workflow takes runtime (opencode, claude, codex or auto); on Claude Code and Codex each model step is one headless claude -p or codex exec on your own account, and read-only roles run read-only. auto follows AGENTRY_RUNTIME, keeps a project that has the opencode plugin on opencode, and otherwise uses the CLI of the MCP client that connected. agentry-mcp run <workflow> runs one from a terminal, and in the editor a step can name a Claude Code agent and a Codex profile.
runtime: omo runs each model step as one headless omo -p on your own account: read-only roles get only omo's read tools, the others a short list (read, search, web, edit, write, shell), and a step started from an omo session does not touch that session. agent_mapping.omo picks a step's model and AGENTRY_OMO_MODEL the default; auto recognizes omo as the MCP client that connected.
npm packages: agentry-mcp, agentry-watch and the opencode plugin as opencode-agentry. npx -y agentry-mcp and npx -y agentry-watch run the latest release, and each release is published there by CI through npm trusted publishing.
Fixed
The opencode plugin loads and runs on opencode 2 (checked on 2.1.2) and keeps working on opencode 1. On opencode 2, which gives plugins no log, progress goes to .agentry/agentry.log.
agentry-watch and agentry-mcp wake projects on opencode 2 servers, and a server that wants a password says to start it with OPENCODE_SERVER_PASSWORD.
agentry-watch and run_workflow wait up to 5 minutes for a project's first start, while opencode installs the plugin's dependency, and a timeout says what was happening.
A model step that fails writes its log with the reason, as tool steps do, and get_run reports the error for that step.