Prisma Compute boot-failure repro

Paired working/failing reproductions of three Compute issues, found while deploying a small Hono + Prisma app on 2026-07-05.

Environment: @prisma/cli 3.0.0-beta.22 · workspace "Gregory's Space" · project findcrewai-ops · region fra · authed as boch@prisma.io.

The three issues

1. --env prevents the process from ever starting

Same artifact, same app, same command — only difference is --env <file>:

Deploy	Result
app deploy --entry server.js	boots, serves
app deploy --entry server.js --env env.sample	never starts, platform 404

Base64-encoding values changed nothing → the delivery mechanism, not escaping. Reproduced ~8 times, both on tiny and large artifacts.

2. Large bundles boot nondeterministically (without env)

• control/server.mjs (300 B, no deps): boots 100% (10+ deploys).
• A 5.3 MB comment-padded copy of the same hello: boots → not size alone.
• failing/server.js (5.4 MB esbuild bundle: hono + @hono/node-server + generated Prisma client + @prisma/adapter-pg + pg): booted once across ~10 deploys of byte-identical artifacts. Boots first-try under local node server.js and bun server.js every time.
• Intermediate bundles (hono-only 28 KB; hono+prisma 5.3 MB probe) booted when tried once each — consistent with per-deploy nondeterminism rather than a specific dependency.

3. Failed boots are invisible

Every failing deployment still reports status: "running", live: true, promoted: true, and app logs streams nothing (not even a first line). There is no way to self-diagnose from the CLI.

Secondary CLI quirks (same version)

• prisma.compute.ts is never loaded — deploySettings.config.path is null whether the file sits in the app dir or the repo root, with @prisma/compute-sdk installed and the documented defineComputeConfig default export.
• --framework / --entry are silently ignored when a framework is detected from package.json dependencies (result says "status": "inferred").
• app deploy [dir] reads like a directory argument but is a config target name; the error only surfaces when no config file exists either.

Reference deployment IDs (our workspace, safe to inspect)

What	ID / URL
Main app (many failed boots)	cps_cmr7msfsz0i5f2hdu88hsc1h9 → cmr7msfsz…fra.prisma.build
Working hello deploy	app hello-repro, e.g. cpv_cmr7o00io…
Failing large-bundle deploys	cpv_cmr7n8lbh0imozofbddye55dq, cpv_cmr7ncjk20mfwysfas7apmsci, cpv_cmr7nff0c0i8v2hduza6amz2a, more on request
Fresh app, same failure	app ops-live → cmr7quucj…fra.prisma.build (its hello baseline boots)

Repro steps

See deploy.sh. The failing/server.js artifact contains no secrets — env is read at runtime (env.sample shows the shape with dummy values).

Contact: Gregory (boch@prisma.io). Full deploy-by-deploy timeline available.