Meow Flow

Coordinate MeowFlow staged agent work through the mfl CLI.

Hard Rules

• Do not use chat history as the source of truth for thread state. Discover and update MeowFlow state with mfl before acting.
• If a slash command arrives in the middle of an existing Paseo agent chat, this chat is coordinator-mode unless mfl agent update-self confirms it is already the requested MeowFlow stage agent for the current thread.
• In coordinator-mode, do not perform plan/code/review/execute/validate work locally. Launch the requested stage with mfl run --stage <stage> ..., report the returned ids, and stop.
• If mfl agent update-self cannot classify this session as a MeowFlow stage agent, stay in coordinator-mode and use mfl run for the requested stage.

Command Contract

• Syntax: /meow-flow [--provider <provider>] [content] or /mfl [--provider <provider>] [content].
• Continuation syntax: /mfl plan [--provider <provider>], /mfl code [--provider <provider>], /mfl review [--provider <provider>], /mfl execute [--provider <provider>], /mfl validate [--provider <provider>], /mfl commit, /mfl archive, and /mfl delete.
• The current thread state, worktree occupation, agents, request body, and handoffs come from mfl, not from chat history alone.
• Thread names set with mfl thread set name must be kebab-case matching ^[a-z0-9]+(-[a-z0-9]+)*$.
• Agent names from mfl run are <Stage>: <request summary> until the thread has a name, then <Stage>: <thread-name> (<agent-sequence>). If a stage agent changes the thread name, update the current idle or running Paseo agent with paseo agent update <agent-id> --name "<Stage>: <thread-name> (<sequence>)".

Startup

1. If running inside a Paseo agent, run mfl agent update-self.

2. Run mfl status.

3. If status is repository-root, tell the user to create a linked worktree with mfl worktree new and stop.

4. If status is idle and the user provided request content, launch the first plan agent. When the entry command includes --provider <provider>, pass it to mfl run; otherwise rely on mfl run provider resolution:

   mfl run --stage plan "<content>"
   mfl run --stage plan --provider <provider> "<content>"
   

5. If status is occupied, report the thread name or id and latest agent id, then:

   • For /mfl plan, /mfl code, /mfl review, /mfl execute, or /mfl validate, launch the matching stage with mfl run.
   • For /meow-flow, /mfl, or /meow-plan with request content and no explicit continuation stage, launch a plan stage in the current thread with mfl run --stage plan ....
   • If there is no request content or the user asks for a different thread, ask how they want to proceed before launching anything.

When mfl run returns agent-id: <id> and next-seq: <seq>, include both in your response. Also include thread-id and worktree when mfl run prints them, then direct the user to continue in the new agent chat.

Session Mode

After startup, decide whether this chat is a coordinator or a stage agent:

• Coordinator-mode: any normal Paseo agent chat, any mfl agent update-self failure, or any chat that is not confirmed as the requested stage. Launch with mfl run --stage <stage> ... and do not perform the stage work locally.

• Stage-agent mode: mfl agent update-self confirms this session belongs to the current thread and requested stage. Before acting, read:

  mfl status
  mfl thread status <id> --no-color
  mfl handoff get -n 5
  

Stage Dispatch

For staged continuation commands, run the matching stage through mfl run in the current thread worktree. Pass --provider <provider> when the user included it; otherwise rely on mfl run provider resolution:

mfl run --stage plan "<optional-content>"
mfl run --stage code "<optional-content>"
mfl run --stage review "<optional-content>"
mfl run --stage execute "<optional-content>"
mfl run --stage validate "<optional-content>"

mfl run --stage code --provider <provider> "<optional-content>"

The launched stage agent reads mfl thread status <id> --no-color and recent handoffs before acting. The coordinator stops after reporting the launched agent details.

Handoffs

Stage agents should read recent handoffs with:

mfl handoff get -n 5
mfl handoff get --since <seq>

Before finishing, agents that produce planning, implementation, review, execution, or validation results append a compact handoff:

mfl handoff append --stage plan "planned X; proposal Y; next code"
mfl handoff append --stage code "implemented X; tests Y passed"
mfl handoff append --stage review "approved; checked A and B"
mfl handoff append --stage execute "generated script X; output at Y"
mfl handoff append --stage validate "validated X with command Y"

Keep handoffs short and concrete. They are coordination state for later agents.

Continuation Actions

• /mfl commit: read new handoffs, inspect the current diff, commit with a suitable repository title, and push when a remote tracking branch is configured.
• /mfl archive: follow the same behavior as /meow-archive.
• /mfl delete: follow the same behavior as /meow-archive delete; do not revert code changes.

Output

For launches, report:

• the stage launched
• thread-id
• worktree
• provider
• agent-id
• next-seq
• the next chat or command the user should use