What problem does it solve? Long-running agent tasks lose state, drift into free-form chat, and improvise branching decisions when a conversation is interrupted. This Skill enforces a deterministic start → yield → observation → resume → done loop where a bundled runtime owns state, retries, and branch selection while the host agent only executes the current step. ## Core Features & Use Cases - Bridge-driven execution loop: scripts/bridge.py start returns a yield, done, or error response; the host executes the yielded prompt envelope and replies with a structured Observation via bridge.py resume until the runtime declares the run terminal. - Published workflow catalog: workflow-binding.json ships workflows such as demo-prompt-loop, academic-research-pipeline, ios-ai-assisted-development-flow, co-storm-autonomous-research, and earnings-reviewer, selectable explicitly with --workflow-id or via the configured default. - Companion subskills: setup, inject, pack, register, delete, and workflow-creator handle installing slash shortcuts, injecting catalog blocks into AGENTS.md/CLAUDE.md, exporting/importing .flow archives, removing workflows, and authoring new workflows from a spec.json blueprint. - Use Case: Run /durable-workflow-runtime academic-research-pipeline to drive a multi-stage research workflow where each stage is verified, retries surface a compact retry_context, and blocked steps pause safely for external input instead of failing silently. ## Quick Start Invoke /durable-workflow-runtime with an optional workflow id and let the agent allocate host I/O paths with scripts/host_io.py before calling bridge.py start.
