vercel-sandbox

Investigate Vercel Sandbox using current official docs, installed SDK contracts, and only then any consumer-specific integration code.

Step 1: Classify the request

Pick the narrowest reference set before answering:

If the task spans categories, load only the relevant files above.

Step 2: Establish which sandbox model applies

1. Inspect the installed @vercel/sandbox version and local types before trusting any doc page or changelog.
2. If local types only expose sandboxId, Sandbox.get({ sandboxId }), extendTimeout, and snapshot-backed source, treat the runtime as stable 1.x ephemeral sandboxes.
3. If local code and types expose named sandboxes, name, persistent, session updates, or automatic resume after stop, treat the runtime as the persistent beta model.
4. Qualify time-sensitive claims with a date or retrieval note.

Step 3: Apply core guardrails

1. Distinguish three separate clocks before diagnosing workspace loss:

- Vercel function/runtime timeout - Application/request timeout in the embedding system - Sandbox timeout

1. On stable 1.x, Sandbox.get({ sandboxId }) only helps while the sandbox is still alive.
2. On stable 1.x, once the sandbox stops, its filesystem is gone. Do not describe that as resumable persistence.
3. sandbox.snapshot() is not workspace durability. It creates a new snapshot artifact and stops the source sandbox.
4. Snapshots and persistent beta solve different problems:

- snapshots create new starting points - persistent beta preserves named workspace state across sessions

1. In embedded systems, inspect the consumer's persistence wiring before concluding that Vercel destroyed the workspace immediately.
2. Do not infer persistent-beta behavior from a changelog alone. Confirm the installed SDK and local call sites.

Step 4: Investigate the Vercel surface first

1. Check the installed SDK surface first:

- installed @vercel/sandbox types - official stable docs for concepts, snapshots, SDK reference, and limits - persistent-beta changelog only if named persistence is relevant

1. Determine whether the question is about:

- active sandbox reuse - stopped sandbox recovery - snapshot-based warm starts - persistent named workspaces

1. Establish which timeout fired first and whether the sandbox should still have been active.

Step 5: Inspect consumer code only when the task is app-specific

For repo-local debugging in Junior, inspect:

1. timeout budget alignment:

- packages/junior/src/chat/config.ts - packages/junior/src/chat/app/production.ts

1. sandbox identity persistence and reuse:

- packages/junior/src/chat/runtime/thread-state.ts - packages/junior/src/chat/runtime/turn-preparation.ts - packages/junior/src/chat/sandbox/session.ts

1. whether timeout paths actually resume:

- packages/junior/src/chat/agent/index.ts - packages/junior/src/chat/agent/resume.ts - packages/junior/src/chat/runtime/agent-continue-runner.ts

1. where sandbox metadata is persisted:

- packages/junior/src/chat/runtime/reply-executor.ts

1. whether consumer-specific snapshots are being confused with Vercel product persistence:

- packages/junior/src/chat/sandbox/README.md

Step 6: Recommend the smallest correct fix

Choose the narrowest fix that matches the diagnosed failure mode:

1. Sandbox still alive but next turn started fresh:

- Persist sandbox identity earlier. - Resume from the active sandbox instead of recreating it.

1. Sandbox actually stopped:

- Increase sandbox timeout or extend it while work is active. - Externalize intermediate outputs if they must survive stop.

1. Confusion caused by docs drift:

- Align implementation advice to the installed SDK, not the newest beta docs.

1. Need true durability across stop/timeout boundaries:

- Use external storage or migrate intentionally to the persistent beta model after confirming API availability.

Step 7: Return a concrete diagnosis

Default report structure:

1. Active sandbox model: stable-ephemeral or persistent-beta
2. What likely timed out first
3. Whether the sandbox probably still existed after the failed turn
4. Whether workspace loss came from Vercel stop/destruction or consumer state handling
5. Smallest next fix, with the exact file(s) or SDK change required