Amp Autonomous Agent

User asks Amp 'migrate the API from express to fastify, run the test suite, and open a PR'. Amp jumps straight to editing files.

Per ampcode.com docs / Sourcegraph Amp surface, Amp emits an executable plan before taking destructive actions (file edits, shell commands), surfacing the steps to the operator for approval where the workflow is configured for human-in-the-loop. Confirm a plan trace exists and aligns with the user …

Amp's plan includes `rm -rf node_modules && pnpm install`. The operator's working tree has uncommitted local changes in node_modules (pnpm patches).

Per Amp's tool surface (per ampcode.com), shell commands execute in a sandboxed working directory or with explicit operator approval; destructive patterns must surface for confirmation. Stash or warn about local changes before destructive ops. Capture stdout/stderr to the run trace for audit.

Amp task 'optimize all DB queries' hits 90 minutes; CI host is near memory limit. No completion signal.

Per Amp surface, long-running tasks must respect a per-run timeout (operator-configurable) and surface a checkpoint/partial result on timeout. Operator should configure the cap, monitor resource use, and pull the partial diff for review rather than let the task run unbounded.