Resonate records each step of a durable execution as a promise. That is a good storage model that requires an understanding of the protocol to create a debugging surface.
In other words, a failed workflow sits in a store as a set of rows, and getting the story out of that means knowing which query to write.

resonate-ui reads those promises back and reassembles them into the call graph they came from, so a run arrives as a picture. The UI is Apache-2.0, it builds to static files, and the browser talks to your server directly.

The Workflows view. Four workflow call graphs sit side by side as small node diagrams. checkout-7f3a and fanout-2b6d are resolved and drawn in teal; import-legacy-1c5f is canceled and drawn in grey; payout-91bc is rejected and drawn in plum, with one child still pending in amber. Each card carries a state badge, a step count, a done/failed/pending tally and a total duration.

Five states, drawn as five

A rejection, a cancellation and a missed deadline call for three different responses respectively, including different colours and labels in every view. Graphs lay out automatically, pan and zoom, and hold their shape while the data refreshes underneath. Parameters and return values arrive base64-encoded; the console decodes them and pretty-prints the JSON, so the reason a step failed reads as BalanceError: insufficient balance: need 4210, have 1980 on the page instead of in your clipboard on the way to a decoder.

The promise detail page for payout-91bc.2. A status panel reads REJECTED, a timing panel gives created, times-out, settled and a 1.8s duration, and a tags panel lists resonate:parent, resonate:origin and resonate:scope. A Parameters section shows decoded JSON naming the function transfer_funds. A Value section shows decoded JSON with type BalanceError and the message "insufficient balance: need 4210, have 1980".

Context for an assistant

Every record-bearing view carries an Ask AI button that turns what you are looking at into a single Markdown document: the schema, the view state, the structure, the raw records and the decoded payloads. It is built in your browser and copied to your clipboard. A human can read the graph, while an AI assistant reads the bundle.

The document closes by listing what it left out: records past the cap, values it could not decode, a view that failed to load. That last item took a fix. The all-clear branch used to fire whenever there were no caps and no redactions, which is exactly the state a failed load leaves behind, so the most confident document the feature could emit came from the least informed capture.

The Ask AI dialog, titled "Context for an assistant". It explains that everything the view is rendering is turned into text you can paste into any assistant, built in this browser and copied to your clipboard. It reports 4 records and 6 KB, carries an amber warning that the bundle includes payload data and should be treated as public the moment it is copied, and previews the generated Markdown. Buttons offer Copy to clipboard, Download .md and Close.

Run it against the fixture

npm install
npm run mock     # a fixture speaking the real protocol, on :8099
npm run dev      # the console, on :5173

Every screenshot above is that fixture. It seeds six named workflows and a batch queue anchored to the moment you start it, so the dashboard has a live hour behind it and the schedules have a next run ahead of them.

The README is the full tour: CORS and auth setup, the API compatibility notes, the known limitations and the roadmap. The concepts are in the docs, and questions go to Discord.

Try it!

Resonate HQ

A console for durable execution