specflow

name	specflow
description	Use when the user asks Claude Code to create, inspect, validate, run, dynamically review, patch, pause, continue, or resume Specflow/Aflow agent workflows in the current repository.

Specflow

Use the plugin-provided Specflow MCP tools for real workflow operations. Do not shell out to specflow run when the user wants Aflow Agent parity; that command is the standalone CLI runner and intentionally does not cover dynamic runtime patching or pauseAfterRun ACP conversations.

Workspace

• Work in the user's target repository, not in the plugin directory.
• Pass the repository path as cwd to MCP tools when it is known.
• If the user gives a serverUrl, pass it through. Otherwise let MCP locate or start the persistent workspace Specflow server.
• Treat returned serverUrl and runId as durable handles for follow-up tool calls.
• Once a Specflow MCP tool first returns a serverUrl in the current chat, proactively surface the Specflow UI once for that serverUrl: try to open the URL in the available browser surface if Claude Code exposes one. If no browser action is available, do not block. Do not open duplicate tabs for the same serverUrl in the same chat. Always tell the user both the UI URL and the /design URL.

Workflow Authoring

• Before creating a new workflow, read references/create-workflow.md and references/workflow-authoring.md.
• Before adapting an existing workflow, read references/fork-adapt-workflow.md and references/workflow-authoring.md.
• For small edits to an existing workflow, read references/workflow-authoring.md only when schema details are needed.
• List workflows with specflow_list_workflows.
• Read YAML with specflow_read_workflow.
• Before adapting an existing workflow, fork it with specflow_fork_workflow_to_local.
• Write complete workflow YAML with specflow_write_workflow; local drafts should stay in agentflows-local unless the user explicitly wants a shared workflow.
• Usually reference files already in the repository by relative path instead of importing them.
• Use specflow_import_assets rarely: only when the user explicitly wants to copy files from outside the repository into .aflow/.specflow/agentflow/assets/... so the workflow is durable, shareable, commit-friendly, or reproducible.
• Validate with specflow_validate_workflow before running.

Agent Servers

• Use specflow_list_agent_servers before writing workflow session agentServerId values.
• Use specflow_list_agent_registry when the user wants to inspect available registry agents.
• Use specflow_install_registry_agent or specflow_update_registry_agent only when the user explicitly asks to add or update a registry-backed agent server.
• Use specflow_remove_agent_server only when the user explicitly asks to remove an agent server.
• Do not treat workflow authoring as permission to change agent server config.
• For custom/headless or non-registry agents, ask the user to configure the agent in Specflow UI; do not write raw agent server JSON.
• Agent authentication is handled in Specflow UI or Aflow. If prepare/start reports auth is required, ask the user to authenticate there and retry.
• Use specflow_get_agent_capabilities or specflow_refresh_agent_capabilities to inspect supported ACP modes, permissions, and config options before hard-coding them in workflow YAML.

Running

• Prepare a run first with specflow_prepare_run when variables or auth may be missing.
• Ask the user for missing variables in chat, then call prepare/start again with variableValues.
• If prepare/start reports missing agent authentication, tell the user to open Specflow UI or Aflow and authenticate there, then retry. Claude Code should not drive terminal/TUI auth through this plugin.
• Before starting a run, ask the user to choose Normal run or Dynamic review unless they already specified the mode.
• Start normal or dynamic runs with specflow_start_run.
• For long-running workflows, use bounded waits. If a tool returns before the run finishes, keep the runId and poll with specflow_get_run or continue with the relevant run tool.
• Use specflow_pause_run, specflow_play_run, specflow_interrupt_run, and specflow_stop_run only for existing active runs. pause and interrupt can be resumed with play when the server reports the run can continue. stop is terminal for that run; use continuation, rerun, or a new run instead of trying to play it afterward.
• Use specflow_rerun, specflow_delete_run, and specflow_save_run_best_practice only when the user explicitly asks to rerun, delete, or preserve a successful runtime snapshot as a workflow.

Dynamic Review

Dynamic review is Claude-driven checkpoint review, not user-driven stepping. It pauses after activations so Claude can inspect the completed node text, optionally patch only the current run snapshot, and continue without asking the user at every checkpoint.

• Start with dynamicReview: true.
• Inspect with specflow_get_run_checkpoint.
• Patch only when there is clear evidence that future workflow behavior should change.
• Patch with structured specflow_patch_run_graph operations, not full YAML.
• Continue with specflow_run_to_next_checkpoint.
• Ask the user during dynamic review only for explicit pauseAfterRun, permission/elicitation, auth, missing variables, ambiguity, or high-risk actions.

pauseAfterRun ACP Conversations

pauseAfterRun is a server-owned ACP paused-node session, not MCP elicitation.

When a run reports paused nodes:

1. Call specflow_list_paused_nodes if needed.
2. Use specflow_prompt_paused_node for one or more messages to the paused ACP node.
3. Use specflow_continue_paused_node to close the authored pause.

In dynamic mode, call specflow_continue_paused_node with play: false, then call specflow_run_to_next_checkpoint so the next activation is armed with dynamic pause behavior.

Permission, Elicitation, Sessions

• Use specflow_list_pending_interactions and specflow_respond_interaction for ACP permission or elicitation requests.
• Ask the user before selecting permission options unless their instruction is already explicit.
• Use session tools to list, restore, prompt, cancel, or close ACP sessions.
• Use specflow_get_native_resume_commands or specflow_get_agent_session_native_resume_command when the user asks how to resume a recorded session in a native agent CLI. Trust only returned verified commands; if unavailable, use ACP restore/inspect or ask the user.
• Do not look for MCP auth tools. Agent authentication is handled in Specflow UI or Aflow, then the Claude Code run can be retried.