@fitchmultz/pi-workflows0.2.8
Pi-native recipe workflows and scripted agent()/pipeline() runtime
Untested on omp
This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.
Reports 0
No reports yet.
Readme
pi-workflows
pi-workflows adds development recipes and scripted workflows to Pi. Recipes use specialist agents; scripted workflows coordinate parallel workers.
Install and start
pi-workflows requires Pi; see the compatibility requirements. The scripted workflow / agent() / pipeline() runtime works independently, without other extensions.
Run these commands from your project directory:
pi install npm:@fitchmultz/pi-workflows
pi
Then run this command in Pi:
/workflows
/workflows lists 17 recipes, 31 skills and 25 agents. It reports unavailable recipe integrations without disabling scripted workflows.
Optional recipe integrations
Recipes inherently use isolated specialists, parallel reviews, progress tools and explicit approval gates. They require a supported Pi Subagents version (or a compatible subagent executor with agent/task/label/tasks), todo_list, ask_question, and agent_browser_web_search. All four are hard runtime prerequisites for every recipe, even recipes without external research. These are extension integrations, not native Pi tools.
pi install npm:@fitchmultz/pi-subagents
pi install git:github.com/fitchmultz/pi-todo-list
pi install npm:@fitchmultz/pi-ask-question
pi install npm:pi-agent-browser-native
Configure EXA_API_KEY for Exa or BRAVE_API_KEY for Brave when using recipe research. Provider charges apply.
Owned /recipe-* and /skill:recipe-* invocations stop before model work when required tools are missing, disabled or incompatible. Cold Subagents is supported through its advertised load_subagent loader; native discovery can restore selected, instruction-group-deferred tools. A denied recipe reports the exact tools and does not steer an unrelated running turn. Text/JSON print mode also writes the diagnostic to stderr and exits nonzero.
Recipe instructions then verify specialist profiles and provider/dialog readiness before use; a missing capability is reported, not replaced with inline specialists, serialized reviews or assumed approval. The same instructions apply when a model reads a recipe skill directly. That model-read path is guidance, not a deterministic runtime guard.
Choose a recipe
Recipes use named agents and include approval points. Type /recipe to see recipe commands.
| Task | Command |
|---|---|
| Plan a feature | /recipe-design Add account recovery |
| Build a backend or general change | /recipe-implement Add rate limiting to the public API |
| Investigate a bug | /recipe-diagnose Login returns 500 after deploy |
| Design frontend screens | /recipe-front-design Add account recovery screens |
| Build a full-stack feature | /recipe-fullstack-implement Add account recovery |
| Review a design | /recipe-review docs/design/rate-limit.md |
For frontend work, run /recipe-front-design. Next, run /recipe-front-plan. Then run /recipe-front-build.
Run a scripted workflow
Use a scripted workflow for repeated tasks across many files. Ask Pi to write one:
use a workflow to audit every route handler under src/routes/ for missing auth
The editor highlights the trigger phrase. Press Alt+W / Option+W to dismiss the highlight.
The terminal UI shows these choices before the script runs:
| Choice | Action |
|---|---|
| Once | Run the script once. |
| Always | Skip future approval for this workflow name in this project. |
| View | Read or edit the script. Then return to the approval card. |
| Deny | Cancel the script before it runs. |
Run /deep-research <question> to use the bundled research script.
Scripted workers use native file and shell tools. They do not inherit recipe agents or browser tools.
Workers can change files and run shell commands. Approve only scripts you trust. The runtime has no security sandbox or CPU deadline.
RPC requests confirmation for names without Always approval. Non-UI mode runs scripts without confirmation.
Each worker makes model calls. Large runs use more of your model allowance.
Control a run
Use /workflows to view, pause, resume, stop or save a run. Runs and the resume cache last for the current extension session.
Save writes the script to .pi/workflows/<name>.js. Pi adds the /<name> command in the next session.
/workflow-size shows the agent-count guideline. Set it with /workflow-size small, medium, large or unrestricted. The default is medium: fewer than 15 agents.
The runtime permits at most 16 workers at once and 1,000 calls per run. Size guidelines do not change these limits.
More detail
Read Scripted workflows for script examples, worker results and resume behavior.
Read Development for compatibility requirements, Git installs, the repository layout, build commands and host qualification.
Read the changelog for changes. Report problems or suggestions in issues.
License
MIT.