multi-step-workflow
A deterministic N-step research pipeline on pugmark. No tools, no model-driven
branching, just task → plan → step 1 → step 2 → step 3 → answer → pause. Each
call is one event on the append-only log, so the whole pipeline is
crash-resumable from any step and forkable at any step.
It answers the question that comes up after the react-tool-agent hero: “what does pugmark look like when you don’t have a tool-using react loop, just a multi-step workflow?” Full source on GitHub.
Prerequisites
brew install firetiger-oss/tap/pugmark
pip install pugmark anthropic
export ANTHROPIC_API_KEY=sk-ant-...
Run
pugmark switch file:///tmp/pug-workflow # point pugmark at this example's bucket
pugmark run --start -- python workflow.py
pugmark log -A # → task, plan, step(1), step(2), step(3), answer, then pause
The handler seeds a demo task on the first invocation when the log is empty, then runs the full pipeline to completion.
Fork to try a different plan
Fork at the plan snapshot and push a different plan event to steer the workflow
without redoing the prior research:
pugmark fork -s <session>:2
pugmark push <<<'{"kind":"plan","steps":["focus on async support","focus on type safety","focus on community size"]}'
pugmark run -- python workflow.py
pugmark log -A # branching tree
The downstream step(1) / step(2) / step(3) / answer for the forked session use
the new plan. The parent’s history is untouched.
What it shows
| Pattern | Where in the code |
|---|---|
Walk the log via pugmark.read() and decode each entry as JSON | for obj in pugmark.read(): last_event = obj.decode_json() |
fold(state, event) to update a small in-process state one event at a time | accumulates task, plan, completed steps |
Structural pattern matching with guards (if idx < NUM_STEPS) | match last_event: in main() |
| One write per invocation, then return (pugmark’s run loop calls back) | every terminal in the dispatch |
The same shape as the react agent, but with a four-event protocol instead of a
four-event conversation. No tools, no model SDK features beyond plain
messages.create, just the substrate.