Repository navigation
perf(sdk): start a stdio server with the protocol module it already loaded - #923
Merged
frontegg-david merged 1 commit intoOct 21, 2026
Conversation
Contributor
|
Important Review skippedReview was skipped as selected files did not have any reviewable changes. ⚙️ Run configuration
You can disable this status message by setting the Use the checkbox below for a quick retry:
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info
📝 Walkthrough
Merge Risk: ⚪ Minimal · up to
|
| Check name | Status | Explanation |
|---|---|---|
| Title check |  |
The title clearly and concisely describes the main change: using the already loaded protocol module to start a stdio server. |
| Description check | |
The description explains the motivation, implementation, performance impact, behavior across environments, benchmark results, and extensive test coverage. The optional user-facing change section is no… |
| Docstring Coverage | |
No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 1… |
| Linked Issues check | |
Check skipped because no linked issues were found for this pull request. |
| Out of Scope Changes check | |
Check skipped because no linked issues were found for this pull request. |
Full details: Description check
-
Explanation
The description explains the motivation, implementation, performance impact, behavior across environments, benchmark results, and extensive test coverage. The optional user-facing change section is not included, but the description is otherwise complete and directly relevant.
Full details: Docstring Coverage
-
Explanation
No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 1 files.
✨ Finishing Touches
-
📝 Generate docstrings
-
- Commit to this branch
- Create a new PR
🧪 Generate unit tests (beta)
-
- Commit to this branch
- Create a new PR
-
- Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts
Comment @coderabbitai help to get the list of available commands.
Contributor
Author
|
@coderabbitai review |
Contributor
✅ Action performedReview finished.
|
Contributor
Performance Test ResultsStatus: ✅ All tests passed Summary
Total: 101 tests across 21 projects 📊 View full report in workflow run Generated at: 2026-10-21T12:20:40.233Z |
alexmercerpo
approved these changes
Oct 21, 2026
frontegg-david
force-pushed
the
perf/1.10.x-stdio-single-protocol
branch
from
October 21, 2026 11:06
179183d to
eb4b803
Compare
Contributor
Author
|
@coderabbitai full review |
Contributor
✅ Action performedFull review finished. Your included review limit is currently reached under our Fair Usage Limits Policy. Your current included review allowance is based on your included PR review attempts over the past 7 days. This review may still proceed through usage-based billing if eligible. Your next included review will be available in 19 minutes. |
…oaded
runStdio() took McpServer and StdioServerTransport from
`await import('@frontmcp/protocol')`. In a CommonJS process that import
resolves the package's `import` condition, so a stdio server loaded a second,
ESM copy of @frontmcp/protocol and, through it, of the MCP SDK (types.js and
its zod schemas, the client, shared/protocol.js, zod-to-json-schema), on top of
the CommonJS copy the SDK had already required.
runStdio() now uses the statically imported @frontmcp/protocol, as the HTTP
transport already does, so a CommonJS stdio server uses one copy of the MCP
SDK. In an ESM process the import resolves the same ESM module as before.
frontegg-david
force-pushed
the
perf/1.10.x-stdio-single-protocol
branch
from
October 21, 2026 12:04
eb4b803 to
5b60ca5
Compare
frontegg-david
added a commit
that referenced
this pull request
Oct 21, 2026
…oaded (#923) runStdio() took McpServer and StdioServerTransport from `await import('@frontmcp/protocol')`. In a CommonJS process that import resolves the package's `import` condition, so a stdio server loaded a second, ESM copy of @frontmcp/protocol and, through it, of the MCP SDK (types.js and its zod schemas, the client, shared/protocol.js, zod-to-json-schema), on top of the CommonJS copy the SDK had already required. runStdio() now uses the statically imported @frontmcp/protocol, as the HTTP transport already does, so a CommonJS stdio server uses one copy of the MCP SDK. In an ESM process the import resolves the same ESM module as before. (cherry picked from commit 9201f8a)
frontegg-david
added a commit
that referenced
this pull request
Oct 21, 2026
to #925, #932) (#931) * perf(build): publish ASCII-only JavaScript, which V8 keeps at one byte per character (#919) V8 keeps the source of every loaded script for the life of the process, and stores it with two bytes per character when one character is above U+00FF. esbuild escapes non-ASCII characters in strings and identifiers but copies the comments it keeps as they are, so an em dash or a box-drawing rule in a comment doubled the memory the SDK bundle's source takes (6 MB instead of 3 MB); tsc output (the CLI, @frontmcp/nx) keeps them in strings too. Each publishable package's build now ends with scripts/escape-non-ascii.mjs, which rewrites every non-ASCII character in its .js, .mjs and .cjs output as a \uXXXX escape. A file is rewritten only when esbuild compiles the escaped text to the same program as the original, so a raw template string or a character after a backslash stops the build instead. CI checks the built output with --check. (cherry picked from commit 12cfece) * perf(protocol): load the MCP client and its transports on first use under require() (#920) * perf(protocol): load the MCP client and its transports on first use under require() @frontmcp/protocol's root re-exported the MCP SDK's Client, StreamableHTTPClientTransport, SSEClientTransport and StdioClientTransport, so every CommonJS server loaded the SDK's client, its OAuth client, the SSE and streamable HTTP client transports, eventsource and cross-spawn (31 files) at startup, though a server only uses them for remote apps. The client exports now come from a #client subpath import. Under the require condition it resolves to lazy-client.js, whose getters load each MCP SDK module the first time the export is read; worker, browser and ESM importers get the same static re-exports as before (an ESM importer gets esm/client.mjs, which strip-dist now maps for an `import` target, so the client still comes from the MCP SDK's ESM build). The public exports don't change. * test(protocol): cover the package's exports, now that it has tests and so a coverage threshold lazy-client.spec.ts is @frontmcp/protocol's first test file, so CI's protocol:test --coverage now applies the package's 95% threshold, and the re-export modules nobody had tested left it at 21%. exports.spec.ts checks that the Node modules hand back the MCP SDK's classes, that the browser stand-ins throw their errors, and the 2026-07-28 version helpers. index.ts is left out of coverage, as in most of the libraries here. (cherry picked from commit 0f7c402) * perf(utils): take the ReDoS analyzer from @enclave-vm/ast/redos, which loads none of the parser (#921) @frontmcp/utils imported analyzeForReDoS and REDOS_THRESHOLDS from the @enclave-vm/ast root, which loads the whole AST validator and its parser (acorn, acorn-walk, astring), so every FrontMCP server loaded them at startup. @enclave-vm/ast 2.17.0 exports the analyzer alone as @enclave-vm/ast/redos (agentfront/enclave#104). @enclave-vm/core moves to 2.17.0 with it in the root and plugin-codecall, since core pins ast exactly; the tree keeps one @enclave-vm/ast. (cherry picked from commit 0ac0461) * perf(build): publish whitespace-minified CommonJS and ESM builds and no source maps (#922) * perf(build): publish whitespace-minified CommonJS and ESM builds and no source maps V8 keeps the source of every loaded script for the life of the process, and parses all of it at load. The SDK's CommonJS bundle was 3.07 million characters, of which comments and indentation are about a third. Every package's build-cjs and build-esm now set esbuild's minifyWhitespace: comments and whitespace go, identifiers and syntax stay. The SDK's index.js drops from 3.07 MB to 2.0 MB. The tsc-built packages (frontmcp, @frontmcp/nx, @frontmcp/skills) stop emitting .js.map files, so no package ships source maps. minifyWhitespace also drops the comments bundlers read. The one that changes behavior is the webpackIgnore / @vite-ignore pair on the ESM loader's import() of a computed specifier: without it, rspack (the CLI's lambda and vercel bundler) replaces the import with a context that throws "Cannot find module". That function now lives in its own module, reached through the #import-computed subpath import and built unminified, so the comments stay. The browser builds (dist/browser) are not minified, so the @__PURE__ annotations their per-module tree-shaking relies on stay too. * perf(build): keep the browser-facing packages unminified, so bundlers keep their pure annotations esbuild's minifyWhitespace drops /* @__PURE__ */ annotations, which bundlers use to tree-shake @frontmcp/react, @frontmcp/ui and @frontmcp/plugin-webmcp out of browser apps. Servers don't load those packages for their own startup, so minifying them saved nothing this stack measures. (cherry picked from commit 86bccf8) * perf(sdk): start a stdio server with the protocol module it already loaded (#923) runStdio() took McpServer and StdioServerTransport from `await import('@frontmcp/protocol')`. In a CommonJS process that import resolves the package's `import` condition, so a stdio server loaded a second, ESM copy of @frontmcp/protocol and, through it, of the MCP SDK (types.js and its zod schemas, the client, shared/protocol.js, zod-to-json-schema), on top of the CommonJS copy the SDK had already required. runStdio() now uses the statically imported @frontmcp/protocol, as the HTTP transport already does, so a CommonJS stdio server uses one copy of the MCP SDK. In an ESM process the import resolves the same ESM module as before. (cherry picked from commit 9201f8a) * perf(sdk,protocol): load the Streamable HTTP server transport with the first HTTP session (#924) * perf(sdk,protocol): load the Streamable HTTP server transport with the first HTTP session A stdio server loaded the MCP SDK's Node Streamable HTTP transport at startup, and with it @hono/node-server, which reads the global Request and Response when it loads and so makes Node load its bundled undici. Two things did it: @frontmcp/protocol's CommonJS root required #mcp-streamable-http, and the SDK's RecreateableStreamableHTTPServerTransport, which extends the MCP SDK class, was defined when the SDK's root initialized the transport barrel. - #mcp-streamable-http gets a require variant whose getters load each transport the first time it is read, and a types entry ahead of it. - TransportStreamableHttpAdapter imports the recreateable transport as a type and loads it with loadStreamableHttpTransport(); TransportService awaits it where it creates a LocalTransporter, which it already did asynchronously. - The SDK root imports the in-memory server, web-fetch handler, web-standard helpers and transport types from their modules instead of the barrel, which initialized every module transport/index.ts re-exports. An HTTP server now loads the transport with its first session; a stdio server never does. * perf(sdk): load the Streamable HTTP transport only for Streamable HTTP and stateless HTTP sessions An SSE session uses TransportSSEAdapter, which never touches the Streamable HTTP class, so an SSE-only server no longer loads it either. (cherry picked from commit 4129311) * perf(sdk): import only the semver functions the ESM loader uses (#925) semver.utils.ts imported the semver package root, which loads all 46 of its files, at startup. It uses seven functions; importing them from their own modules (semver/functions/*.js, semver/ranges/*.js) loads 23 files, with the same results. (cherry picked from commit f9bf70c) * fix(sdk): create one stateless transport when the first requests for its key arrive together (#932) * fix(sdk): create one stateless transport when the first requests for its key arrive together getOrCreateAnonymousStatelessTransport and getOrCreateAuthenticatedStatelessTransport checked lookupLocal(key), then awaited the transport's creation and ready() before insertLocal(key). Concurrent first requests could each create one; the later insert replaced the earlier, and the earlier one's disposal then evicted the replacement by key. Both now go through the creationMutex createTransporter already uses. (cherry picked from commit 21c4da8) * build(scripts): say why escape-non-ascii stops on a regular-expression literal or tagged template, and how to fix it esbuild keeps non-ASCII characters in regular-expression literals and tagged templates as written, so it can't confirm that an escape there leaves the program the same, and after a backslash an escape doesn't. The script still stops the build, now naming both cases and the fix: write the character as a \\uXXXX escape in the source. (cherry picked from commit 998f5d5) * build(scripts): tell how to write a tagged template without non-ASCII, since an escape changes its raw text A tag that reads raw text, like String.raw, gets the escape's backslash and letters, so the advice to write \uXXXX escapes only holds for regular-expression literals. For a tagged template the error now says to write it without them, for example with an interpolation under String.raw. (cherry picked from commit 26a3aab)
alexmercerpo
pushed a commit
that referenced
this pull request
Oct 21, 2026
…he loaded protocol module (#934) createInMemoryServer(), the direct client's notification and elicitation handlers, and runWebStandardMcp() took their classes and schemas from `await import('@frontmcp/protocol')`. In a CommonJS process that import resolves the package's `import` condition, so connect(), create() clients and createFetchHandler() loaded a second, ESM copy of @frontmcp/protocol and, through it, of the MCP SDK, on top of the CommonJS copy the SDK had already required. They now use the statically imported @frontmcp/protocol, as runStdio() does since #923. In an ESM process the import resolves the same ESM module as before.
alexmercerpo
pushed a commit
that referenced
this pull request
Oct 21, 2026
* perf(sdk): serve in-memory, direct-client and web-standard MCP with the loaded protocol module (#934) createInMemoryServer(), the direct client's notification and elicitation handlers, and runWebStandardMcp() took their classes and schemas from `await import('@frontmcp/protocol')`. In a CommonJS process that import resolves the package's `import` condition, so connect(), create() clients and createFetchHandler() loaded a second, ESM copy of @frontmcp/protocol and, through it, of the MCP SDK, on top of the CommonJS copy the SDK had already required. They now use the statically imported @frontmcp/protocol, as runStdio() does since #923. In an ESM process the import resolves the same ESM module as before. (cherry picked from commit 5b54191) * perf(sdk): load js-yaml, semver and the uipack renderer on first use under require() (#935) A server loaded js-yaml (25 files), semver (23) and the @frontmcp/uipack root at startup, for code most servers never run: SKILL.md and YAML config parsing, the ESM loader's version ranges, and the auth-UI pages. They now come through three subpath imports, #js-yaml, #semver and #uipack-render, whose `require` variant reads each function from its package the first time it is used, like @frontmcp/protocol's #client. The CommonJS build keeps them external and ships both variants; the ESM and browser builds resolve them at build time to the static variant, so those bundles import the packages as before. (cherry picked from commit b7ee831) * perf(sdk): load the SSE transport with the first SSE session (#937) A stdio server initializes the scope's TransportService, and with it TransportSSEAdapter, which defined RecreateableSSEServerTransport at startup. Its base class reads POST bodies with raw-body and content-type, so a stdio server loaded raw-body, content-type, iconv-lite, http-errors and their dependencies (16 files) that it never uses. TransportSSEAdapter now imports the transport as a type and loads it with loadSseTransport(), which TransportService awaits where it creates the LocalTransporter of an SSE session, as it does for the Streamable HTTP transport since #924. (cherry picked from commit f13b8af)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack 5 of 7, based on #922 (
perf/1.10.x-minify-whitespace). Merge in order.What and why
FrontMcpInstance.runStdio()tookMcpServerandStdioServerTransportfromawait importWithRequireFallback(() => import('@frontmcp/protocol'), ...). The SDK's CommonJS build keeps thatimport()as a native dynamic import, and in a CommonJS process it resolves@frontmcp/protocol'simportcondition. So a stdio server loaded the package a second time, as ESM, and through it a second copy of the MCP SDK: its ESMtypes.jswith every protocol schema,shared/protocol.js, the client and its transports,zod-compat, andzod-to-json-schema's 39 ESM files. That is on top of the CommonJS copies the SDK had already required, and on the path to the first answer.runStdio()now uses the statically imported@frontmcp/protocol, as the HTTP transport (LocalTransportAdapter) already does. A CommonJS stdio server uses one copy of@frontmcp/protocoland of the MCP SDK, the CommonJS one, the same copy its HTTP transport uses; it loads 357 modules instead of 549. An ESM process resolves the same ESM module as before. The browser and Worker builds keep@frontmcp/protocolexternal, as before, and resolve its stubs for the stdio transport.What changes besides startup:
McpServerandStdioServerTransporton the stdio path now come from the MCP SDK's CommonJS build instead of its ESM build (same package version), so errors they raise are instances of the sameMcpErrorclass the rest of a CommonJS FrontMCP server uses.Results
Apple M4 Pro (14 cores, 48 GiB), macOS 15.6.1, Node.js 24.18.1, on AC power, with the machine's normal background load (an IDE and other jobs). The server is the benchmark's FrontMCP implementation (
frontmcp build --target node), run withnodeand no flags, againstnpm packed builds of each branch of this stack and ofrelease/1.10.x, installed into scratch copies of the implementation and built withnpm run build, all measured in one session in randomized, interleaved rounds with the benchmark's own harness functions. Cold start is the time from spawningnodeto the firsttools/callresponse,initializeincluded: one discarded start per server, then 8 rounds of 4 starts per transport in random order. The 100-tool startup is spawn to the first fulltools/list, 5 rounds of 2 starts. Throughput is the benchmark's closed-looptools/callat concurrency 1, 16 and 64 (3 s warmup, 10 s measured, 5 rounds); RSS after startup is read 1 s after the first call, RSS after load after the last level, median of the 5 rounds. Heap used isprocess.memoryUsage().heapUsed700 ms after start (after the first call, no forced GC), median of 10 interleaved starts in a separate run right after. MB is 10^6 bytes. The 1-minute load average was 3.7 to 15.9 during the session (median 5.3); the stdio heap run, 10 interleaved stdio starts with the same preload, ran at 5.0 to 8.6.perf/1.10.x-minify-whitespace)tools/list, p50Per round, the median stdio cold start moved by -56.1 ms (from -73.1 to -25.0 ms). The HTTP path doesn't run this code; its rows are within their run-to-run spread.
Tests
The
runStdiotests (run-stdio-guard,run-stdio-input,run-stdio-primary-scope,run-stdio-revision-router,front-mcp-stdio.decorator; 137 tests in 17 suites with the rest ofsrc/front-mcp) pass; they mock@frontmcp/protocol, which the static import picks up the same way.Loaded modules of the benchmark's stdio server, recorded with a
module.registerHookspreload: 549 before, 357 after, and no.mjsordist/esmfile of@frontmcp/protocolor the MCP SDK after.Run on the top of this stack (
perf/1.10.x-sdk-semver-subpaths), which contains this change:nx run-many -t test,lint,typecheck --projects=tag:scope:publishable(27 projects): all pass, no lint errors. SDK 7,131 tests, auth 2,810, cli 2,105, utils 1,909, plugin-codecall 1,208, adapters 1,011, react 758, uipack 537, plugin-remember 515, observability 496, ui 488, nx-plugin 424, plugin-skilled-openapi 407, testing 377 (8 skipped), and the rest.node --test scripts/*.test.mjs: 83 pass.e2e: bundle-size 8, stdio-transport 26, transport-recreation 40, direct 77, esm 26 (and
test:e2e26), cloudflare 16 (5 skipped), serverless 17, cli-exec 198, dts-consumer 7, remote 31, remote-proxy 3, testing 21, local-auth 73, standalone 49, mcpb 19, unix-socket 20, codecall 37, skills 188 (2 skipped), browser-bundle (Playwright) 11 pass. demo-e2e-public passes 11 of 12, in two runs: the other test could not bind 127.0.0.1:50000 to 50002 (listen EADDRINUSE: address already in use 127.0.0.1:50000), which WebStorm on this machine listens on, inside the suite's port range. The suite takes no port or base-port override (the port registry's free-port check binds the wildcard address, so it doesn't see those three), so that test is left to CI, which has nothing on those ports.The benchmark's correctness gate (HTTP, stdio and the 100-tool server) passes against the build of every branch in the stack.
Summary by CodeRabbit