@ai-agent-forge/plugin-workflow
v0.88.12
Published
Workflow orchestration capability plugin for Agent Forge (stage DAG state machine over session entries)
Maintainers
Readme
@agent-forge/plugin-workflow
Workflow orchestration capability plugin for Agent Forge (design contract:
docs/design/工作流编排插件设计.md). It registers workflow_set and workflow_advance: a fourth self-contained state
machine on top of the goal/plan/todo layers that orchestrates a stage dependency graph (DAG) — dependency unlocking,
parallel activation guarded by write-scope checks, in-place retry and cross-stage retreat with DAG-propagated resets, and
a per-transition audit stream. Workflows are an upgrade shape for genuinely multi-stage parallel work, not the default:
the goal/plan/todo layers keep serving linear small tasks.
State persists as append-only workflow.state session entries; every conversion (including held or rejected activation
attempts) appends one workflow.transition audit entry with a closed-catalog reason code, and each persisted change is
announced as a workflow.changed event. This plugin is the repository's first event subscriber: plan.changed is used
as a trigger signal only — all truth is re-derived from the current branch entries, and the handler is fully isolated so
it never rethrows at the publisher. Other domains are only read through the shared session entry protocol (soft
references: goalId, stages[].planId, stages[].role).
Install
# from a local checkout
agent-forge plugin install ./plugins/workflow
# from npm
agent-forge plugin install @ai-agent-forge/plugin-workflowThe plugin requires the session, events, and tools host capabilities — the same public surface as the goal/plan
plugins — and also exposes the /workflow text command (panorama, why, log) through the public registerCommand
face so any host shape can inspect "why is it stopped here" without a TUI.
