I’ve wanted a specific thing for years: AI embedded in the runtime. Catch an exception → run a prompt. API failure → scan docs for schema changes → patch the response.

I discovered the Claude Agent SDK is effectively “Claude Code as a library” (same agent loop + tools), so I asked: what if the agent doesn’t just edit code — what if it lives inside the running process?

Reflexive runs your app and opens a local web UI where the agent can: - watch stdout/stderr and trigger investigations from log patterns - read/edit project files (explicit flags) - start/stop/restart the process - attach to the Node inspector and set real V8 breakpoints - inspect locals/scopes while paused, optionally run prompt-attached breakpoint analysis, then resume

Two moments convinced me this is a different primitive: 1) I started from an empty file. The app exited(1). The agent asked if I wanted it fixed. Minutes later it had converted it into a working webserver, then kept working after I stopped typing (it curl-tested its own endpoints and narrated why it kept the process alive). 2) I set a breakpoint before an API response, mutated one value to “Customer.Hacked”, resumed, and the agent treated it like an incident: it isolated the relevant code path, disabled the risky behavior, restarted the process, and wrote a postmortem.

This is intentionally a dev tool first: the power is behind explicit flags.

If you try it and it breaks, I’d love the failure reports — the whole point is runtime-level feedback loops.