mirror of
https://github.com/simstudioai/sim.git
synced 2026-04-28 03:00:29 -04:00
2f932054a7
* fix(execution): run pptx/docx/pdf generation inside isolated-vm sandbox
Retires the legacy doc-worker.cjs / pptx-worker.cjs pipeline that ran user
DSL via node:vm + full require() in the same UID/PID namespace as the main
Next.js process. User code now runs inside the existing isolated-vm pool
(V8 isolate, no process / require / fs, no /proc/1/environ reachability).
Introduces a first-class SandboxTask abstraction under apps/sim/sandbox-tasks/
that mirrors apps/sim/background/ — one file per task, central typed
registry, kebab-case ids. Adding a new thing that runs in the isolate is
one file plus one registry entry.
Runtime additions in lib/execution/:
- task-mode execution in isolated-vm-worker.cjs: load pre-built library
bundles, run task bootstrap, run user code, run finalize, transfer
Uint8Array result as base64 via IPC
- named broker IPC bridge (generalizes the existing fetch bridge) with
args size, result size, and per-execution call caps
- cooperative AbortSignal support: cancel IPC disposes the isolate, pool
slot is freed, pending broker-call timers are swept
- compiled scripts + references explicitly released per execution
- isolate.isDisposed used for cancellation detection (no error-string
substring matching)
Library bundles (pptxgenjs, docx, pdf-lib) are built into isolate-safe
IIFE bundles by apps/sim/lib/execution/sandbox/bundles/build.ts and
committed; next.config.ts / trigger.config.ts / Dockerfile updated to
ship them instead of the deleted dist/*-worker.cjs artifacts.
Call sites migrated:
- app/api/workspaces/[id]/pptx/preview/route.ts
- app/api/files/serve/[...path]/route.ts (+ test mock)
- lib/copilot/tools/server/files/{workspace-file,edit-content}.ts
All pass owner key user:<userId> for per-user pool fairness + distributed
lease accounting.
Made-with: Cursor
* improvement(sandbox): delegate timers to Node, add phase timings + saturation logs
Follow-ups on top of the isolated-vm migration (da14027b2):
Timer delegation (laverdet/isolated-vm#136 recommended pattern):
- setTimeout / setInterval / clearTimeout / clearImmediate delegate to
Node's real timer heap via ivm.Reference. Real delays are honored;
clearTimeout actually cancels; ms is clamped to the script timeout
so callbacks can't fire after the isolate is disposed.
- Per-execution timer tracking + dispose-sweep in finally. Zero stale
callbacks post-dispose.
- unwrapPrimitive helper normalizes ivm.Reference-wrapped primitives
(arguments: { reference: true } applies uniformly to all args).
- _polyfills.ts shrinks from ~130 lines to the global->globalThis alias.
Timers / TextEncoder / TextDecoder / console all install per-execution
from the worker via ivm bridges.
AbortSignal race fix (pre-existing bug surfaced by the timer smoke):
- Listener is registered after await tryAcquireDistributedLease. If the
signal aborted during that ~200ms window (Redis down), AbortSignal
doesn't fire listeners registered after the fact — the abort was
silently missed. Now re-checks signal.aborted synchronously after
addEventListener.
Observability:
- executeTask returns IsolatedVMTaskTimings (setup, runtimeBootstrap,
bundles, brokerInstall, taskBootstrap, harden, userCode, finalize,
total) in every success + error path. run-task.ts logs these with
workspaceId + queueMs so 'which tenant is slow' is queryable.
- Pool saturation events now emit structured logger.warn with reason
codes: queue_full_global, queue_full_owner, queue_wait_timeout,
distributed_lease_limit. Matches the existing broker reject pattern.
Security policy:
- New .cursor/rules/sim-sandbox.mdc codifies the hard rules for the
worker process: no app credentials, all credentialed work goes
through host-side brokers, every broker scopes by workspaceId.
Pre-merge checklist for future changes to isolated-vm-worker.cjs.
Measured phase breakdown (local smoke, Redis down): pptx wall=~310ms
with bundles=~16ms, finalize=~83ms; docx ~290ms / 17ms / 70ms; pdf
~235ms / 17ms / 5ms. Bundle compilation is not the bottleneck —
library finalize is.
Made-with: Cursor
* fix(sandbox): thread AbortSignal into runSandboxTask at every call site
Three remaining callers of runSandboxTask were not threading a
cancellation signal, so a client disconnect mid-compile left the pool
slot occupied for the full 60s task timeout. Matching the pattern the
pptx-preview route already uses.
- apps/sim/app/api/files/serve/[...path]/route.ts — GET forwards
`request.signal` into handleLocalFile / handleCloudProxy, which
forward into compileDocumentIfNeeded, which forwards into
runSandboxTask.
- apps/sim/lib/copilot/tools/server/files/workspace-file.ts — passes
`context.abortSignal` (transport/user stop) into runSandboxTask.
- apps/sim/lib/copilot/tools/server/files/edit-content.ts — same.
Smoke: simulated client disconnect at t=1000ms during a task that would
otherwise have waited 10s. The pool slot unwinds at t=1002ms with
AbortError; previously would have sat 60s until the task-level timeout.
Made-with: Cursor
* chore(build): raise node heap to 8GB for next build type-check
Next.js's type-check worker OOMs at the default 4GB heap on Node 23 for
this project's type graph size. Bumps the heap to 8GB only for the
`next build` invocation inside `bun run build`.
Docker builds are unaffected — `next.config.ts` sets
`typescript.ignoreBuildErrors: true` when DOCKER_BUILD=1, which skips
the type-check pass entirely. This only fixes local `bun run build`.
No functional code changes.
Made-with: Cursor
* fix lint
* refactor(copilot): dedup getDocumentFormatInfo across copilot file tools
The same extension -> { formatName, sourceMime, taskId } mapping was
duplicated in workspace-file.ts and edit-content.ts. Any future format
or task-id change had to happen in two places.
Exports getDocumentFormatInfo + DocumentFormatInfo from workspace-file.ts
(which already owned the PPTX/DOCX/PDF source MIME constants) and
imports it in edit-content.ts. Same source-of-truth pattern the file
already uses for inferContentType.
Made-with: Cursor
* fix(sandbox): propagate empty-message broker/fetch errors
Both bridges in the isolate used truthiness to detect host-side errors:
if (response.error) throw new Error(response.error); // broker
if (result.error) throw new Error(result.error); // fetch
If a host handler ever threw `new Error('')`, err.message would be ''
(falsy), so { error: '' } was silently swallowed and the isolate saw
a successful null result. Existing call sites don't throw empty-message
errors, but the pattern was structurally unsafe.
Switch both to typeof check === 'string' and fall back to a default
message if the string is empty, so all host-reported errors propagate
into the isolate regardless of message content.
Made-with: Cursor