eve-on-Vercel: source deploy 500s (ERR_MODULE_NOT_FOUND for internal #channel/*), and --prebuilt skips sandbox prewarm — no clean deploy for sandbox apps

[davidstrouk]

Link to a minimal reproduction

I don't have a standalone minimal repro repo yet — this was found on a real (private) production deployment, and I can't easily reproduce a fresh Vercel build outside that project right now. I'm filing anyway because the evidence points directly at eve's own source (see below), and I'm very happy to build and link a minimal repro (a tiny eve app with one sandbox-using subagent) if that helps you triage.

Steps to reproduce

1. An eve app with at least one node/subagent that uses a sandbox (so eve provisions a Vercel Sandbox template).
2. Deploy it to Vercel with a source vercel deploy (Vercel builds from source — the path eve's own warning recommends).
3. Open any route of the resulting deployment (/, /eve/v1/slack, …).

Current vs. expected behavior

Current: every route returns HTTP 500 FUNCTION_INVOCATION_FAILED. The runtime logs show ERR_MODULE_NOT_FOUND for an eve-internal #channel/* import (e.g. #channel/compiled-channel). The generated Vercel function's dependency trace appears to drop eve's internal #… subpath files, so they're missing at runtime.

Expected: a source vercel deploy serves the routes normally (the internal #channel/* modules are included in the traced function bundle).

Why this matters (the bind): the two deploy paths are mutually exclusive and both break sandbox apps:

• Source deploy (vercel deploy): eve does provision the Vercel Sandbox templates (the on-Vercel build runs prewarm, gated on VERCEL_DEPLOYMENT_ID) — build log shows eve: initialized N sandbox templates — but every route 500s from the trace gap above.
• Prebuilt deploy (local vercel build + vercel deploy --prebuilt, with the traced output hand-patched to re-add the dropped #channel/* files): routes work, but the sandbox templates were never provisioned (a local build has no VERCEL_DEPLOYMENT_ID, so prewarm is skipped) → any agent that touches a sandbox throws SandboxTemplateNotProvisionedError at runtime.

So there's no single-command deploy for an eve app that uses sandboxes on Vercel. eve's own build even warns about the sandbox half:

Skipped Vercel sandbox template prewarm because VERCEL_DEPLOYMENT_ID is missing. The generated .vercel/output may reference sandbox templates that were not provisioned. Do not deploy it with vercel deploy --prebuilt; use vercel deploy so Vercel builds from source.

…but "use vercel deploy" is exactly the path that 500s from the trace gap.

eve version

0.19.0

Environment

• Node.js 24 (both the local build and the on-Vercel build).
• Vercel (Fluid Compute), deployed via the Vercel CLI.
• Build via vercel build / vercel deploy with VERCEL_USE_EXPERIMENTAL_FRAMEWORKS=1.

Where does the bug occur?

At deploy/runtime on Vercel (a source vercel deploy's deployment). The build succeeds; the failure is at request time.

Deployment

A private production deployment (can share the deployment ID privately with the team to correlate platform-side logs).

Build and runtime logs

• On-Vercel build (provisioning works): eve: initializing 6 sandbox templates … / eve: sandbox template "subagents/<id>" (vercel): preparing base runtime inside sandbox / eve: initialized 6 sandbox templates (0 reused, 6 built).
• Source-deploy runtime: HTTP 500 FUNCTION_INVOCATION_FAILED on / and /eve/v1/slack, with ERR_MODULE_NOT_FOUND for an eve #channel/compiled-channel import.
• Relevant eve source: the prewarm gate is dist/src/internal/nitro/host/vercel-build-prewarm.js → shouldPrewarmVercelBuild() (requires both VERCEL and VERCEL_DEPLOYMENT_ID).

Suggested direction

If eve's package exports (and/or the nitro tracing it emits) followed the internal #channel/* subpaths, a plain source vercel deploy would serve correctly and provision sandboxes in one step — removing the community --prebuilt + trace-patch workaround entirely. Happy to help test a fix or provide a minimal repro.