Skip to content

feat: warm-boot snapshot for the Vercel Sandbox eval runner - #201

Draft
Rodriguespn wants to merge 3 commits into
mattrossman/ai-912-spike-vercel-sandbox-runner-for-evalsfrom
ai-912-vercel-sandbox-snapshot
Draft

feat: warm-boot snapshot for the Vercel Sandbox eval runner#201
Rodriguespn wants to merge 3 commits into
mattrossman/ai-912-spike-vercel-sandbox-runner-for-evalsfrom
ai-912-vercel-sandbox-snapshot

Conversation

@Rodriguespn

@Rodriguespn Rodriguespn commented Aug 14, 2026

Copy link
Copy Markdown
Contributor

Follow-up to #192.
Adds an opt-in Vercel Sandbox warm-boot snapshot to the run-vercel-evals controller, measures the net gain under the same conditions as #192, and importantly, keeps the default path cold so it cannot regress the scheduled refresh.

Full refresh at scale: warm vs cold, both concurrency=250 / runs=2:

pairs run-evals boots failures
Warm (this PR)run 31802584549 237 28m10s 237 warm / 0 cold 0
Cold (#192) — run 31735507954 220 30m03s

Every one of the 237 pairs booted warm from the prebuilt snapshot (vars.EVAL_SNAPSHOT_ID): 0 cold, 0 sandbox failures.
The warm run did more pairs (237 vs 220) in less wall-clock, with zero build cost this run (the snapshot was built out-of-band and reused) - ≈13% faster per-pair throughput. The remaining ~28 min is dominated by agent execution + the sequential 2-attempt retry (the separate parallel-attempts lever, not touched here).

The warm run's benchmark pass rate is 194/220 = 88.2%, identical to main's cold history (194/220 = 88.2%) — the environment change didn't move outcomes.

Recommended CI wiring (follow-up, not in this draft)

Keep the scheduled refresh on the default cold path, and add a small "build snapshot" job keyed on pnpm-lock.yaml + packages/sandbox/Dockerfile + the pinned CLI/skills versions that builds once and publishes the id (repo/environment variable or artifact); eval-refresh consumes it via --snapshot-id.

Pre-bakes docker + node_modules + the agent base image + the Supabase
stack images into a snapshot so per-pair VMs boot warm (dockerd restart +
git fetch/checkout + cache-hit pnpm install) instead of cold-provisioning
and pulling images on every pair.

- scripts/vercel-sandbox.ts: shared VM primitives + coldProvision, extracted
  so the runner and the builder no longer import each other (breaks a cycle).
- scripts/vercel-snapshot.ts: one-shot builder; runs coldProvision + bakes,
  then tears the prewarmed stack fully down so only images survive into the
  snapshot (a warm agent must see the same clean docker state as cold).
- run-vercel-evals.ts: warm/cold boot branch; --snapshot builds one for this
  run, --snapshot-id reuses a prebuilt one; default is unchanged (cold).

Measured on build-cli-001-bootstrap-app (codex-gpt-5.4-mini, runs=1):
cold eval 205s -> warm 131s, snapshot build ~180s, same 7/7 checks cold and
warm. Net win requires reuse via --snapshot-id (a per-run build regresses at
concurrency >= pairs), so the default stays cold to avoid any CI regression.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@vercel

vercel Bot commented Aug 14, 2026

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

Project Deployment Actions Updated (UTC)
evals Ready Ready Preview Aug 14, 2026 1:29pm

Request Review

Passes --snapshot-id to eval:vercel when the repo variable is set, so the
full refresh boots pairs warm from a snapshot built out-of-band. Unset ->
cold boots, so this can't regress the default path.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
@Rodriguespn Rodriguespn added the run-evals Add to a PR to refresh benchmark evals label Aug 14, 2026
@mattrossman
mattrossman force-pushed the mattrossman/ai-912-spike-vercel-sandbox-runner-for-evals branch from 117ec3a to 489b380 Compare August 14, 2026 15:18
@Rodriguespn Rodriguespn changed the title feat: warm-boot snapshot for the Vercel Sandbox eval runner (C3 follow-up to #192) feat: warm-boot snapshot for the Vercel Sandbox eval runner Aug 14, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

run-evals Add to a PR to refresh benchmark evals

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant