[derencius]

Problem

The generated workflow entrypoint statically re-exports every workflow (export { default as X } from 'P';). Temporal's reusable VM caches modules loaded during an activation in that workflow's own moduleCache, so every cached workflow evaluates the whole catalog.

Mechanism

@temporalio/worker/lib/workflow/reusable-vm.js:138-156: __webpack_module_cache__ is a Proxy. A set before any activator exists goes into the shared sharedModules map; a set during an activation goes into that workflow's moduleCache (lines 150-156). A static entrypoint is evaluated inside each activation, so each cached workflow holds its own copy of every workflow module.

Change

Each user workflow and alias export in the entrypoint becomes an async wrapper that does await import( /* webpackMode: "eager" */ 'P' ) and calls the default export. Eager mode keeps everything in one bundle. $catalog stays a static export. Workflow type resolution is by export name, which is unchanged. The framework reads workflow metadata (METADATA_ACCESS_SYMBOL) only in Node (importComponents in the loader and the catalog), never from the bundle, so the wrapper loses nothing. --check bundles the same entrypoint through bundleWorkflowCode; a one-workflow lazy entrypoint bundled successfully with that call.

Measurements

Measured 2026-10-19 on a GrowthX os-workflows worker (@outputai/core 0.13.0, Temporal TS SDK 1.20.3, reuseV8Context default true, macOS footprint, workflows parked in sleep). The lazy rows come from patching the same shape by hand in node_modules.

Setup	Per cached workflow	Idle worker
228-workflow catalog, static entrypoint	~114 MB	1.9 GB
1-workflow catalog, static entrypoint	~4.6 MB	0.76 GB
228-workflow catalog, lazy entrypoint	~6.5 MB	1.54 GB

Prod pods reached 7-8 GB at 25 cached workflows and the fleet autoscaled to its 20-pod max. bundlerOptions.preloadModules did not help; preloading @outputai/core breaks activations ("Cannot assign to read only property 'lastIndex' of object '[object RegExp]'"). With the hand patch, a parked workflow and a real workflow with 5 activities and 3 nested child workflows completed across forced worker restarts with zero nondeterminism or module errors.

Follow-up local A/B benchmark

Repeated 2026-10-30 after rebasing against main, using Node 24.15.0, Temporal TS SDK 1.24.0, reuseV8Context: true, and the 72-workflow test_workflows catalog. Each variant started from a stopped Compose environment. Memory values are medians of three docker stats samples before and after starting 100 unique sleep workflows parked on their first timer. Worker telemetry confirmed 101 cached workflows, including the catalog, with no in-flight work.

Metric	Static entrypoint	Lazy entrypoint	Impact
Idle worker	397.6 MiB	400.2 MiB	+2.6 MiB
Loaded worker, 100 cached executions	764.9 MiB	563.0 MiB	-201.9 MiB (-26.4%)
Incremental memory, 100 executions	367.3 MiB	162.8 MiB	-204.5 MiB (-55.7%)
Incremental memory per execution	3.673 MiB	1.628 MiB	-2.045 MiB (-55.7%)
Workflow bundle size	4.51 MB	4.53 MB	+0.02 MB
Workflow bundle compilation	992 ms	986-999 ms	No meaningful change
Compose start to API ready	25.823 s	26.425-27.795 s	No meaningful change

This smaller catalog produces smaller absolute savings than the production catalog, but independently confirms the same effect on the current Temporal SDK and rebased implementation.

Not tested: restart during an in-flight activity; a prod-sized mixed cache.

🤖 Generated with Claude Code

[github-actions]

PR review

Verdict

✅ PASS

Findings

None.

Categories

• Design: ✅ PASS
• Quality: ✅ PASS
• Correctness: ✅ PASS
• Documentation: ✅ PASS
• Changeset: ✅ PASS
• Tests: ✅ PASS
• Security: ✅ PASS
• Compatibility: ✅ PASS