Add service job provider runtime

[ai-virtual-b]

Summary

• add acp serve commands for service-job provider runtimes: init, start, endpoints, deploy bundle, stop, logs
• add handler/budget/offering scaffolds and a runtime that connects outbound to the BE service-job namespace
• keep x402/MPP public endpoints on BE while the provider self-hosts or deploys the handler runtime
• keep --settle-8183 reserved but disabled for x402/MPP until ERC-8183 service-job support exists

Notes

• deploy currently creates a Docker-ready bundle and next steps; it does not execute a hosting-provider deployment
• provider runtime needs ACP_AGENT_TOKEN and outbound access to BE

Tests

• npm run build
• tsx bin/acp.ts serve endpoints --dir /tmp/acp-cli-smoke --json

---

Note

High Risk
Adds a new provider runtime that verifies and settles x402/MPP payments and signs on-chain transactions via a configured private key, which is security- and funds-critical. Also introduces new token-auth flows and deployment bundling that may affect how credentials and config are handled in production.

Overview
Adds a new acp serve feature set to scaffold, run, and (optionally) deploy a provider-side “service-job” runtime that connects outbound to the BE /service-jobs Socket.IO namespace.

The runtime loads a developer handler.ts (plus optional budget.ts), serves a /health endpoint, can run a native ACP listener for registry jobs, and for BE-mediated jobs builds payment challenges and verifies/settles x402 (EIP-3009 USDC) and MPP (tempo via mppx) credentials before invoking the handler; --settle-8183 is explicitly reserved but blocked for x402/MPP.

CLI/config/auth updates support the new flow: registers serve commands (init/start/endpoints/deploy/stop/logs), adds agent-token issuance via /auth/agent (message signing) and stores it in config/env, centralizes API URL selection via getApiUrl(), and adds new dependencies (@x402/*, mppx) plus viem/lockfile bumps.

^{Reviewed by Cursor Bugbot for commit 6eccf1e. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 6eccf1e. Configure here.}

[cursor]

Unconditional break prevents fallback to legacy PID files

Medium Severity

The break statement in the stop command's PID file loop is unconditional — it sits outside the try/catch, so it always executes after the first existing PID file is found, even when process.kill() throws (e.g., stale PID). This prevents the loop from falling through to check the legacy PID file path, which is the entire reason the loop and pidFiles array exist.

 

^{Reviewed by Cursor Bugbot for commit 6eccf1e. Configure here.}

[cursor]

Deploy bundle copies from cwd instead of CLI source

Medium Severity

copyRuntimeBundle resolves CLI source directories (bin, src, serve, etc.) relative to process.cwd(), but process.cwd() is the user's working directory, not the CLI package's installation directory. The init command correctly uses import.meta.url to locate scaffold files, but deploy does not, so the bundle will silently miss CLI source files if the command is run from a directory other than the repo root.

 

^{Reviewed by Cursor Bugbot for commit 6eccf1e. Configure here.}