spring-boot-coding-agent

spring-boot-coding-agent

Clones a Git repository into a sandbox, reads files, and proposes a patch. Exercises the primitives the personal-assistant sample does not touch:

• Sandbox — every file and command goes through the SPI. Docker is the production default; the in-process provider is the dev fallback.
• AgentResumeHandle — a long-running clone + patch flow registers with the RunRegistry so a disconnected client can reattach by runId and replay missed events (wire-in lands in Phase 1.5).

Running

# Build
./mvnw compile -pl samples/spring-boot-coding-agent

# Run with an OpenAI key (optional — sample works without)
export OPENAI_API_KEY=sk-your-key

# Start the sample
./mvnw spring-boot:run -pl samples/spring-boot-coding-agent

# Open http://localhost:8081 and ask:
#   "clone https://github.com/atmosphere/atmosphere.git and read README.md"

Sandbox requirements

The sample prefers the Docker sandbox backend. If docker is on PATH and the daemon is reachable, operations run inside a fresh alpine:3.20 container with 1 CPU / 512 MB / 5 min / no network beyond the initial clone.

When Docker is absent, the sample falls back to the in-process provider for demonstration purposes only. The in-process provider is NOT a security boundary — never expose the in-process backend to untrusted input in production.

What the agent does

1. Extracts the repo URL from the user message.
2. Provisions a sandbox with default limits.
3. Clones the repository at depth 1.
4. Reads README.md (or README) and returns the first 800 characters.

Extending to real patch proposals is left to the reader — the @SandboxTool annotation wires a method to a sandbox backend, and Sandbox.writeFile + Sandbox.exec(List.of("git", "diff", ...)) produce the diff.

Notes

• The sample does not git commit the proposed patch. A production coding agent pairs this flow with PermissionMode.PLAN from the AgentIdentity primitive so the user approves before the commit lands.
• Egress policy extension (per v0.6 follow-up): swap SandboxLimits.network from boolean to NetworkPolicy (NONE / GIT_ONLY / ALLOWLIST) so the clone step can reach GitHub while downstream tool calls cannot reach the wider internet.