You click through a real debugger in your browser. Your AI agent drives the same session from the command line. Both of you see every stop.
An agent sets a breakpoint, runs and steps. The tab follows along. Swipe the recording to see the locals.
Whatever your agent does shows up in your tab as it happens. Whatever you do, your agent can read back.
Record a Milo program once, then start at the failure and go back: Reverse Continue lands on the last time a breakpoint was hit, Step Back on the stop before. Every run of the recording computes the same thing, so the stop you go back to has the same locals it had the first time.
Five reverse continues from the failure find the order that was parsed wrong. Swipe the recording to see the locals.
dapweb speaks the Debug Adapter Protocol, the same one VS Code uses. Nearly every debugger speaks it too.
Plus anything else with a DAP adapter: --dapPath "gdb -i dap".
Everything you expect, plus views of the machine most debuggers hide.
Every action is a dapweb api command with a JSON reply. No plugin, no protocol to learn.
Conditions, hit counts and log points, kept across restarts.
Hex dump colored by region (stack, heap, code), with your locals labeled where they live.
What changed since the last stop, and instruction-level stepping.
Reuse the VS Code launch configs you already have.
Every event is journaled, so you or your agent can ask what happened last time.
curl -fsSL https://milo-language.github.io/dapweb/install.sh | sh
macOS and Linux. Read the script (it downloads the release into ./dapweb-*).
Already installed? dapweb update swaps in the latest build.