Beanflow is a skill. Install the Beans CLI and link skills/beanflow into
~/.agents/skills/beanflow; the dotfiles installer already creates this link.
There is no package build, MCP server, or Stop hook to install.
Ask to plan, start, check, resume, pause, or land a run. The skill owns the workflow and its run-record reference defines how Beans and Git preserve progress. Planning and status requests do not authorize start.
Start each new Beanflow run in its own chat/task with the current implementer profile's
model and reasoning. The implementer writes the code, repairs findings,
verifies checkpoints and communicates with you. It is not a subagent of the launching
chat. The planner defines bounded review slices and release triggers alongside each Milestone.
The implementer dispatches independent reviewer subagents at those Task checkpoints
and continues permitted work. The final Milestone review covers gaps and integration,
without repeating prior coverage. The implementer answers questions,
dispositions findings and repairs confirmed defects at a safe Task boundary. There is
no separate validator or automatic second review; the skill defines when a repair
warrants another review.
When launching from another task, transfer the accepted scope and durable checkpoint, record the destination chat ID, and let that chat own the run. Resume an existing run in its recorded chat instead of creating a duplicate. Planning and status requests alone do not launch implementation.
Read ~/.codex/agents/ROLES.md before dispatch. Model choices live in those host
profiles; the skill's role-routing rules
define how they apply to existing chats, new chats and review agents.
Continuation depends on the agent following the skill. If the host ends or interrupts the chat, explicitly resume from the recorded checkpoint. There is no background process that restarts the agent.