Memory footprint reduced in instance-ai module

OlegIvaniv

·Apr 7, 2026·#27967fix(instance-ai): reduce memory footprint (no-changelog)

Memory optimizations address heap profiling findings in instance-ai: storage-only Mastra singletons prevent agent closure leakage, Brotli compression is disabled to free ~8.6 MB per connection, node types now read from static JSON instead of a 31 MB registry rebuild, and a regex-based JSON highlighter is replaced with a structural walker.

Heap profiling revealed multiple memory leaks and inefficient patterns in the instance-ai module. The most severe: Mastra singletons were storing agent objects with their closures (tools, prompts, thread context) in the internal #agents dict — in multi-user environments this meant User A's agent got pinned in shared memory for all subsequent users. The fix creates Mastra with only storage, while agents get the back-reference needed for suspend/resume via __registerMastra() without being added to the agents dictionary.

A separate bottleneck: collectTypes() triggers a full rebuild of the node type registry, costing ~31 MB each time. The adapter service now reads from the static JSON file that FrontendService writes at startup instead.

Brotli compressors from Express response compression retained ~8.6 MB of native memory each via keep-alive connections. A middleware now strips br from Accept-Encoding to downgrade to gzip, and SSE streams have compression disabled entirely.

The Mastra upgrade brings lazy tiktoken loading, avoiding ~15 MB of WASM allocation on import. A global stub in prevents Mastra's workspace tools from eagerly loading the 66 MB BPE table — a character-based estimate is sufficient for truncation.

On the tracing side, LangSmith requests now use a gzipFetch wrapper that forces gzip encoding and suppresses noisy 409 "payloads already received" errors on retries.

A bug fix: DaytonaFilesystem.stat() threw DaytonaNotFoundError which Mastra didn't recognize, causing write_file to fail on new files. The error is now translated to FileNotFoundError.

In the frontend, a regex-based JSON syntax highlighter was vulnerable to catastrophic backtracking on large tool outputs. It has been replaced with a structural character walker.

View Original GitHub Description

Summary

Fixes several memory and stability issues in instance-ai discovered during heap profiling.

Memory fixes

The cached Mastra singleton in registerWithMastra() and ensureMastraRegistered() was passing agents to the constructor, which stored the first agent's closures (tools, prompts, thread context) permanently in Mastra's internal #agents dict. In a multi-user environment this means User A's agent gets pinned in the singleton shared by all subsequent users. Fixed by creating Mastra with only storage. Agents get the back-reference for suspend/resume via __registerMastra() without being added to #agents.
InstanceAiAdapterService was calling collectTypes() which triggers postProcessLoaders(), an expensive rebuild of the full node type registry (~31 MB). Switched to reading from the static JSON file that FrontendService writes at startup.
Brotli compressors from Express response compression were costing ~8.6 MB of native memory each, retained via keep-alive connections. Added a controller middleware that strips br from Accept-Encoding to downgrade to gzip, and disabled compression entirely on instance-ai SSE streams.
Upgraded Mastra packages to get lazy tiktoken loading (avoids ~15 MB WASM allocation on import) and a token growth fix. Also patched the module loader to prevent Mastra from eagerly importing tiktoken and tokenizer (~60 MB combined).
LangSmith tracing now uses a gzipFetch wrapper that forces gzip encoding for both proxy and direct client paths, avoiding the brotli decompressor overhead on the tracing side.

Other fixes

LangSmith patchRun was logging noisy 409 "payloads already received" errors on retries. The gzipFetch wrapper now returns a synthetic 200 for 409 responses since the data already landed.
Mastra's wrapWithReadTracker calls stat() before file writes and catches FileNotFoundError for new files. But DaytonaFilesystem.stat() was throwing Daytona SDK's DaytonaNotFoundError which Mastra didn't recognize, causing write_file to fail on new files that don't exist yet. Now translates Daytona 404s to Mastra's FileNotFoundError.
The regex-based JSON syntax highlighter in ToolResultJson.vue was vulnerable to catastrophic backtracking on large tool outputs. Replaced with a structural character walker.

Review / Merge checklist

PR title and summary are descriptive. (conventions)
Docs updated or follow-up ticket created.
Tests included.
PR Labeled with Backport to Beta, Backport to Stable, or Backport to v1 (if the PR is an urgent fix that needs to be backported)