fflow

name	fflow
description	Start a workflow run. Default uses fflow start with CLI tracking. Use --lite for lite mode, --markdown for prompt-only mode.

Start Workflow Run

Initialize a new workflow run from a workflow YAML file.

Usage

• /fflow PATH — normal mode (default). Uses fflow start with full CLI tracking.
• /fflow --lite PATH — lite mode. Uses fflow start --lite (simplified cards on revisits).
• /fflow --markdown PATH — markdown-only mode. Uses fflow render to inject the full workflow as a prompt. No CLI state tracking or hooks.

Mode Detection

1. If args contain --markdown: use markdown mode (see below).
2. If args contain --lite: use lite mode (see below).
3. Otherwise: use normal mode (see below).

Strip the mode flag from args before extracting PATH.

---

Common Steps (all modes)

1. Check for active run

If there is a remembered run_id from a previous /fflow in this conversation, run fflow current --run-id <run_id>. If the current state is not done, prompt the user:

• "You have an active workflow <run_id> in state <state>. Abort it and start a new one?"
• Options: "Abort and start new" / "Keep it, start new anyway"
• If the user chooses to abort, run fflow abort --run-id <run_id> first.
• If the state is done or the run doesn't exist, skip this step silently.

2. Generate a run ID

Use the format <workflow-name>-$(date '+%Y%m%d%H%M%S') where workflow-name is derived from the workflow filename. Use lowercase letters, numbers, and hyphens.

---

Normal Mode (/fflow PATH)

Full CLI-tracked mode with event sourcing and hooks.

Process

1. Complete the Common Steps above.

2. Run the CLI command:

fflow start <PATH> --run-id <run-id>

Never omit --run-id. PATH can be a workflow name (e.g. spec-gen) or a full path. The CLI resolves it automatically.

3. Remember the run ID — Store the run_id value for use in subsequent fflow current --run-id <run-id> and fflow goto <state> --run-id <run-id> --on <transition-label> calls within this conversation.

4. Flow CLI output

WARNING — DO NOT TRUNCATE CLI OUTPUT

fflow start, fflow goto, and fflow current output state cards that contain critical workflow instructions. You MUST read the complete, untruncated output of these commands. NEVER pipe them through tail, head, or any other tool that truncates output. NEVER use a line-limit or byte-limit on their output. Truncating the output will cause you to miss guide rules, state instructions, and transition definitions, which will lead to incorrect workflow execution.

fflow start will output the initial state card. fflow goto will output the new state card. If the target state is done, the workflow is completed.

The state card consists of instructions, todos and valid state transitions. Follow the instructions and transition to the correct state based on the output of your actions.

Execution model: After every state transition, immediately execute the new state's instructions. You may summarize progress or report status, but do NOT stop between states. Keep driving the workflow forward until you reach a terminal state (a state with no transitions). Only a terminal state ends the workflow.

Before ending a turn: You MUST run fflow current --run-id <run_id> before ending your turn to check if there is remaining work in the current state. Only end your turn if the current state has no actionable work left or requires user input. This prevents accidentally dropping tasks mid-state.

If the exit code of any CLI is not 0, the CLI will output the error message. Follow the error message on right actions to take.

---

Lite Mode (/fflow --lite PATH)

Same as normal mode but with --lite flag. Revisited states show simplified cards.

Process

1. Complete the Common Steps above.

2. Run the CLI command:

fflow start <PATH> --run-id <run-id> --lite

3. Follow the same Flow CLI output instructions as normal mode above.

---

Markdown Mode (/fflow --markdown PATH)

The agent reads the full workflow once and self-manages state transitions — no CLI calls after the initial render. Uses fflow start --markdown to create a minimal run directory and output the full workflow as markdown.

Process

1. Complete the Common Steps above.
2. Run:

fflow start <PATH> --run-id <run-id> --markdown

This creates a run directory with only fsm.meta.json (no event sourcing) and outputs the full resolved workflow as markdown to stdout with ${workflow_dir} and ${run_dir} substituted.

3. Remember the run ID — you can use fflow current --run-id <run-id> to retrieve metadata (run_id, mode, workflow_dir, run_dir), but goto and finish are blocked in markdown mode.

4. Read the complete, untruncated output. This is the entire workflow specification.

5. You are now running a workflow. The full workflow is rendered above. Follow these rules:

   • Start at the initial state (the first state in the document).
   • Read the state's instructions and execute them fully.
   • Check transitions: after completing a state's work, evaluate which transition condition is met and move to that target state.
   • Track your current state yourself — there are no CLI commands to call (goto and finish will error with MARKDOWN_MODE).
   • Keep driving forward: after every state transition, immediately execute the new state's instructions. Do NOT stop between states.
   • Terminal state: a state with no transitions ends the workflow. Only stop when you reach a terminal state.
   • Guide rules: if the workflow has a guide section, follow those rules throughout all states.

---

workflow_dir — Workflow Directory

The CLI commands (fflow start, fflow current, fflow goto) include a workflow_dir field in their output. In markdown mode (fflow render), it appears as an HTML comment: <!-- workflow_dir: /path -->. This is the directory containing the workflow YAML file.

Use workflow_dir from CLI output whenever workflow prompts reference files relative to the workflow directory (e.g., {workflow_dir}/../references/github-cli.md).

Error Handling

• RUN_EXISTS — The generated run_id is already taken. Generate a different slug and retry.
• SCHEMA_INVALID — The YAML file has validation errors. Show the error message and suggest using /fflow-author to build a valid workflow.