ProjectHelm — making multi-agent coding observable

The question

When two or three agents touch the same repo in one evening, how do I answer on Monday: what was decided, what actually ran, and what still needs me?

Cursor is strong at local edits. It is weak as a system of record. Plans live in chat. Child agents disappear when the tab closes. Failed checks become another prompt instead of a named Debug phase. I kept losing the thread between "we agreed to fix the tests" and "which agent touched writes.ts?"

ProjectHelm (the JARVIS workspace) is my answer: a thin control plane on my machine. Not another chat UI. State, phases, and a ledger around the Cursor SDK.

The control plane

The FastAPI app owns projects and runs. The web UI is for me to start work, watch status, and kill a loop that went sideways. Agents call into the Cursor SDK with bounded context. Outcomes append to an event ledger. If it isn't in the ledger, I treat it as rumor.

Phases I refuse to leave implicit

Plan forces a scoped breakdown: owners, done criteria, what "good" looks like before anyone edits. Vague goals stay stuck here on purpose.

Agent dispatches work. Child agents get a job, not the whole chat history. That boundary is the whole point; context rot is how you get confident wrong patches.

Debug starts when verification fails. Evidence in, guesswork out. I still improvise sometimes, but the phase name makes the failure visible instead of burying it in a longer prompt.

Ledger is the audit trail. Monday-me can reconstruct Tuesday-night-me without scrolling Discord.

Design choices I'm keeping

Local-first. No per-agent cloud bill for a personal workshop. Explicit phase names beat prompt tone. Kill switch in the UI so I don't hunt PIDs. The IDE still edits files; ProjectHelm coordinates.

I'm not building multi-user permissions or auto-merge. Merge stays human. Fully autonomous shipping is a different product and I'm not pretending this is that.

Where it still breaks

Ledger richness depends on how honest the agent wrappers are about what they did. If a child agent skips writing an event, the dashboard looks calm and lies. Plan quality is still on me; a sloppy plan just becomes a faster mess. Desktop + API + IDE is more moving parts than "open Cursor and vibe," and some weeks I skip the control plane when the task is tiny. That's fine. Tools you only use for multi-agent nights still earn their keep.

Why I published this

Delivery leadership and agent systems share a habit: name the loop, keep evidence, don't confuse activity with outcomes. ProjectHelm is the workshop version of that habit. Full diagrams: JARVIS (ProjectHelm) — architecture.

Takeaway

Treat agents like workers with a foreman. Build the foreman layer first: phases, state, and a log you can read when you're tired. The model is replaceable. The contract isn't.