How do I install Whyts?

Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.

Does Whyts need an API key?

No required environment variable is listed on this record.

When was Whyts last checked?

Last checked 2026-11-09. Find out why your TypeScript project is slow. Evidence, not guesses.

Whyts MCP-Server

whyts

Find out why your TypeScript project is slow.

A small CLI that turns compiler diagnostics, traces, and import relationships into evidence you can act on. It says where the check is slow, and for three known patterns it also says why and what to do. No account, API key, or LLM required.

npx whyts

Run it in the folder of your tsconfig.json. It runs the real compiler once and prints at most three actions. To prove that a type change helped, use npx whyts compare (see below). To give the same evidence to an AI coding agent, use claude mcp add --transport stdio whyts -- npx -y whyts-mcp.

Found with whyts: a 7.4 s type check that went to 0.06 s (TanStack Form) and vue-tsc going from 80 s to 0.56 s (Vue). See Real cases.

whyts terminal report

Try it

Requires Node.js 20+ and a TypeScript project with dependencies installed.

npx whyts --project .

For repeatable runs, pin a version: npx [email protected] --project .

To run from a checkout:

git clone https://github.com/musatoktas/whyts.git
cd whyts
npm ci --ignore-scripts
npm run demo
node src/cli.js --project /path/to/your/tsconfig.json

Diagnoses and remedies

By default, whyts prints at most three actions and one summary line. Each action has a pattern, a location, a short remedy and a link. The image above is a recorded run. Use --verbose for the full report.

whyts 0.8.5 · TypeScript 5.9.3 · repro-project/tsconfig.json

1. [measured] Recursive type instantiation: DeepKeys<Values>
   Where: k.ts:4:17
   TS2589 at this place: Type instantiation is excessively deep and possibly infinite.
   It hit the instantiation limit inside DeepKeysAndValuesImpl (../src/util-types.ts:151).
   DeepKeysAndValuesImpl refers to itself: DeepKeysAndValuesImpl -> DeepKeyAndValueArray -> DeepKeysAndValuesImpl.
   Type argument JsonData (k.ts:2): JsonData refers to itself.
   Fix: Stop the expansion of DeepKeysAndValuesImpl when it meets a type it already visited. Or add a depth limit to it. Check that the fix keeps the same results.
   Cost: 10082.6 ms, at most 98.8% of Check time (upper bound, not a saving)
   Docs: https://github.com/microsoft/TypeScript/wiki/Performance

Check 10.20s · Total 10.80s · 65 files · 1 compiler error · 2 findings in the full report. Use --verbose to see them.

Eight patterns have a remedy. Each one was seen in a real project before it was added.

Pattern	Trace evidence	Remedy
`recursive-type-instantiation`	A TS2589 error and the `instantiateType_DepthLimit` events. whyts finds the type alias that owns the type parameters at the limit, a reference cycle by name between aliases, and a type argument that refers to itself	Stop the expansion on a type that was visited, or add a depth limit
`partial-constraint`	A TS2589 error at a type reference, and `instantiateType_DepthLimit` events inside a generic type. A type parameter of the referenced declaration has a constraint that gives that generic type fewer type arguments than it declares, so the omitted ones take their defaults	Give all type arguments in the constraint, for example `any` for each. If the constraint is in a dependency, tell its maintainers
`instantiation-expression`	A TS2589 error at a type reference to a type alias whose body is a conditional type, such as `ReturnType` or `Parameters`. A type argument is an instantiation expression (`typeof f<T>`), and the `instantiateType_DepthLimit` events are not inside that alias	Write the conditional type in place, with the expression as the checked type. For `ReturnType`, the declared return type of `f` can be used instead
`builder-chain`	A TS2589 error at the start of a chain of method calls, and `instantiateType_DepthLimit` events inside a class. The method called most often in the chain, followed back through variables, takes and returns that class, and is called at least 32 times	Build the calls as separate values of at most 32 calls each, and join them with the same method. Cutting the chain into variables does not help
`nested-declaration`	A TS2589 error at a type or at the name of a declaration, and `instantiateType_DepthLimit` events inside a generic type. Through type aliases, interfaces and `typeof`, the type reaches a value whose written type in a declaration file nests that generic type at least 8 levels deep	Write the data type by hand as named interfaces, and annotate the value in its source with a type that names them. If the declaration is in a dependency, tell its maintainers
`callback-inference`	A check chain whose focus is a call that took 200 ms or more, and recorded type comparisons and nested checks cover at most 25% of it. The trace has no event for type argument inference. The call gives no type arguments, a generic declaration of the callee has a parameter whose type uses its type parameters, and the argument for it is a function with its own type: a name of a function, or a function whose parameters all have types

A program-load action looks like this (vuetify 4.2.4, a project that only has import { createVuetify } from "vuetify"; the lines are shortened):

1. [measured] Program load: 421 dependency files, vuetify has 407; Check time is 0.01s of 4.22s
   Where: ../node_modules/vuetify/lib/framework.d.ts:1:1
   The program has 480 files: 421 in node_modules (13.8 MB of source), 58 compiler library files and 1 of your own. ... Memory used 577 MB in the traced run.
   Largest packages by files: vuetify 407 (12 MB), @vuetify/v0 3 (0 MB), vue 2 (0 MB), ...
   409 of them (11.9 MB) come only through 218 typeof import() type references in ../node_modules/vuetify/lib/framework.d.ts, to ../node_modules/vuetify/lib/components/index.d.ts (192), ...
   Shortest import chain: index.ts -> ../node_modules/vuetify/lib/framework.d.ts -> ../node_modules/vuetify/lib/components/index.d.ts
   Fix: Each typeof import() reference in ... loads its target for every project that imports the package. Ask the maintainers to move them to an opt-in entry point. skipLibCheck skips checking these files, not loading them. ...

Rules:

A diagnosis is measured when the compiler recorded it (the error or the trace events). The name-based walk over type aliases is syntax only: it can miss a cycle that runs through imports it cannot match by name, or through a type that is not a declaration of the program.
The variance pattern needs 100 ms of union time in total, and only events of 50 ms or more count. It lists only types declared in the project. A type from a dependency gets no remedy, because the project cannot change it.
The cost is the largest recorded interval for the place, shown as at most N% of Check time. It is an upper bound under tracing. It is not a predicted saving.
In the compact report, a slow check without a known pattern shows only above 3% of Check time. A check that a diagnosis already explains by type name is not listed twice.
The trace events exist in TypeScript 5.9, 6.0 and 7.0 (instantiateType_DepthLimit, getVariancesWorker). In TypeScript 7 they carry a checkerId, and whyts uses one checker, so the type ids stay unique.
The program-load pattern needs all of these: 200 files or more from node_modules, a Total time of 1 s or more, a Check time of at most 1 s and at most 10% of the Total time, and one file that alone brings 100 files or more (and 25% of the dependency files). The compiler library files and your own files do not count as dependency files. A project with a normal amount of its own code has a larger Check time and does not match. The counters come from the traced run, where Memory used is higher than in a plain run.
program-load is a measurement of the program, not of the check. It does not say that removing the files makes the check faster: the check time of such a program is already small. It reads typeof import('pkg') in declaration files as an import. A file that is also brought by a regular import is not counted as brought only by those references.
callback-inference is a match of the call shape and of the time that the trace does not explain. The trace has no event for inference, so whyts does not prove that inference is the cost. It reads only the syntax of the callee declarations and of the argument. It does not know which overload the compiler chose. A type parameter that is wrapped in NoInfer deep inside the parameter type still matches: in telegraf (#2016), that form did not change the time, and NoInfer around the whole element type did.
Remedies are short instructions. They name a direction. whyts did not measure any remedy on your project.

Real cases

whyts found the place in each case below. The tRPC fix is merged. The other fixes are open pull requests in those projects, so they are proposals until the maintainers merge them. The Vuetify row is a measurement in an issue, with no fix. Each number comes from my own measurement on the stated compiler version.

Project	What whyts found	Result
tRPC (#6997)	TS2589 in `Serialize` (`serialize.ts:35`) when the type contains a recursive JSON type such as `Prisma.JsonValue`	A fix in PR #7660, merged on 2026-10-19. The repro had TS2589 on main and has no error with the fix. Instantiations went from 89,700 to 78,502. The type-check time of `packages/server` did not change (within noise)
TanStack Form (#1474)	TS2589 in `DeepKeys`: the type alias refers to itself through `DeepKeyAndValueArray`	A fix in PR #2422. Check time of the repro went from 7.4 s to 0.06 s (`whyts compare`, 10 runs, separated). `form-core` stayed within noise
TypeORM (#8559)	TS2589 in `QueryDeepPartialEntity` (`QueryPartialEntity.ts:17`) for types that contain themselves	A fix in PR #12943. whyts reports the diagnosis before the change and reports none after it
Vue (#14323)	`defineModel` with a `defineEmits` call signature that ends in `void` makes `OverloadUnionRecursive` (`@vue/shared`) run to the depth limit. It is 100% of the Check time	A fix in PR #15851. `vue-tsc` with 1, 2 and 3 emits went from 31.1, 55.3 and 80.0 s to 0.42, 0.49 and 0.56 s (TypeScript 5.9.3). A component with 40 overloads and a model still hits TS2589, as it did before
Hono (#2399)	TS2589 in `JSONParsed` (`src/utils/types.ts:53`) with a recursive JSON type	A fix that stops at a visited type, in . A depth limit made the output 313,638 bytes; the fix keeps it at 762 bytes

What it finds

Finding	Evidence	What it does not claim
Source check chains	Related expression/variable locations grouped by same-thread temporal nesting, inclusive chain duration, contained type comparisons with names and declaration locations	That nesting or a contained comparison proves the cause of a slowdown
Broad inclusion	Effective `include` and configured root count	That every broadly included file is unnecessary
Roots worth reviewing	Generated and output roots with no observed importers. Test and spec roots are only counted	That unimported files are safe to delete
Duplicate type versions	Multiple versions of an `@types` package loaded into the program	That all duplicate versions can safely be aligned
Barrel reach	Reexport count, importers, transitive file reach, reached files already configured as roots	Exclusive check cost, bundler behavior, or guaranteed savings
Slow file checks	Recorded `checkSourceFile` trace intervals	Complete attribution of total time or the exact expensive type
Compiler errors	The first five errors, the error code counts, and a warning when unresolved modules dominate	That the timings are valid when dependencies are missing
Recursive type instantiation	TS2589 location, the type alias at the instantiation limit, name-based reference cycles, self-referencing type arguments	That the listed cycle is the only cause
Variance computation	Outermost `getVariancesWorker` events for project types, nested types, union time	That a variance annotation removes the cost

Findings are labeled measured, observed, or review. Measured source check chains appear first, with project code prioritized over dependencies. Chain spans below 10 ms and file spans below 100 ms remain in the recorded data without becoming findings. Structural suggestions follow in a separate section. Suggested changes require a fresh measurement and behavior checks. whyts never invents a predicted speedup.

From a file to an expression and its types

npm run demo:types

The 16-line example assigns a mapped API client with 676 routes to a public client shape. The report can show the assignment at client.ts:16:14, the FullClient to PublicClient comparison recorded inside that check, and their declarations at lines 8 and 11. For example, one local TypeScript 5.9.3 run recorded 38.4 ms for the variable check and 31.8 ms for this comparison. These are inclusive samples from that run, not stable benchmarks or speedup claims. A fast run can omit sampled expression/comparison events.

Nested checks in the same file and trace thread are shown as one chain, leaving room for other expensive checks. Each chain includes its entry location and up to five related checks. The focus is the deepest recorded non-identifier check covering at least 80% of the chain's recorded duration, falling back to its largest member. This is a navigation heuristic; deferred checks can visit earlier source lines, and temporal nesting is not AST containment or proof of causality. Inclusive member durations overlap and must not be added.

Type IDs are resolved selectively from the compiler's types.json, including files larger than 128 MiB. whyts streams individual descriptors and retains only IDs needed by the selected report entries. Named types, compiler flags, union/intersection member counts when available, and declaration locations give a concrete place to investigate. The project's compiler scanner skips leading whitespace and comments to show the declaration's first token. Unresolved IDs remain explicit. Comparisons without a containing recorded expression are listed separately without source attribution. Definition locations may point into a dependency, and are context rather than a proposed fix.

Project file intervals have their own list, so large dependency checks cannot hide every project file. If all files reached by a barrel are already configured roots, whyts explains why changing that import alone will not remove them from the program.

Share of Check time

Each measured chain and file interval shows an upper bound. The terminal report says "at most N% of Check time". JSON has checkTimeShareUpperBoundPercent.

The value is the inclusive interval divided by the compiler Check time. It is not a prediction. The compiler recorded the interval during a trace run, and the trace run is slower than a normal run. A change to the code does not remove the whole share.

Explain a file

node src/cli.js explain src/shared/value.ts --project /path/to/project

File paths are relative to the selected tsconfig directory. The result shows a configured-root reason or one shortest observed import/reference chain, plus direct importers. Import resolution respects TypeScript paths, reexports, dynamic imports, and ESM/CJS usage modes.

This is a compact explanation, not a complete implementation of tsc --explainFiles. Library inclusion, ambient type directives, symlinks, and project-reference redirection can require the compiler's own explanation.

JSON and automation

node src/cli.js --project apps/web/tsconfig.json --json > report.json
node src/cli.js --project apps/web --timeout 300 --no-color

Version 0.7 adds the additive field diagnoses and does not change the existing fields. Each entry has pattern, title, confidence (measured), location, milliseconds, checkTimeShareUpperBoundPercent, evidence, remedy and docs. evidence has the event, the typeName, the location and the time, and the pattern fields. compilerErrors has the new list excessiveDepth with the TS2589 sites. The --verbose option changes only the terminal output. The JSON is always complete.

JSON has schemaVersion, compiler version, summary, diagnostics with units, findings with evidence, hotspots, and warnings. Version 0.3 adds sourceGroups and typeDescriptors while preserving schema version 1 and the existing hotspots, projectHotspots, sourceHotspots, and typeHotspots meanings. Measured source findings now use rule source-check-chain and chain evidence. sourceGroups includes root, members, memberCount, inclusive milliseconds, and the chosen expression's focusMilliseconds. Version 0.8.10 adds focusRecordedMilliseconds: the part of focusMilliseconds inside recorded type comparisons and checks nested in the focus. typeDescriptors records file/scanned/retained bytes and requested/resolved ID counts. Source/type records include durations in milliseconds; source and declaration locations use one-based line/character, and source pos/end are compiler UTF-16 offsets. Progress goes to stderr. Exit codes:

Code	Meaning
`0`	Analysis completed and type checking succeeded
`1`	Tool/configuration failure, unsupported compiler, or timeout
`2`	Report available, but the compiler reported errors

Findings alone do not fail CI. This release diagnoses; it does not enforce a performance budget.

For a native TypeScript 7 run, the JSON has four more fields: compiler (native), experimental (true), graphTypescriptVersion and checkers (1). typescriptVersion is the version of the native compiler.

Measure repeatedly

One run is weak evidence. In our tests, Check time changed by up to 10% between runs on the same tree. Use --runs to measure more than once.

npx whyts --project . --runs 5

How it works:

whyts first does the normal traced run. All results about chains and files come from this run.
Then whyts does N more runs without a trace. A trace slows the compiler and adds a large type dump.
Each run uses a new temporary incremental cache. Each run checks the whole program.
The flags are the same as in the traced run, but without --generateTrace. Your own build cache stays unchanged.
The traced run has read every file before. The file cache is warm for all N runs. whyts adds no warm-up run in this mode.
The report shows the median, the smallest value, the largest value and the spread. The spread is (largest minus smallest) divided by the median, in percent.
The default is --runs 1. One run adds no extra runs and no timings field.

The Check time and Total time in the first lines of the report come from the traced run. Use the untraced values for before and after claims.

The JSON gets an additive timings field. Schema version 1 stays the same. The field has mode, runs, checkers, flags, compilerFiles, compilerExitCodes, checkTime and totalTime. Each time object has unit (s), runs, values, median, min, max and spreadPercent. The compiler prints a rounded value. The JavaScript compiler prints 10 ms steps.

TypeScript 7: the untraced runs use --checkers 1, like the traced run. We measured the Check time of the tRPC packages/server project with 15 interleaved runs for each mode. The spread was 18.1% with one checker and 19.2% with the default checkers. We saw no gain in stability from the default mode. One checker keeps the flags equal to the traced run, so a chain time and a Check time describe the same work. Do not compare this time with a default parallel tsc run. Several checkers are not supported in this release.

Compare two projects or two reports

Live comparison runs both projects on this machine:

npx whyts compare --baseline ./before --candidate ./after --runs 10

Each side is a directory or a tsconfig file. --runs is the number of measured runs for each side. The default is 5, and the range is 1 to 100. The options --timeout, --max-old-space-size, --typescript and --json also work. A shared --typescript makes both sides use one compiler.

Offline comparison reads two JSON reports:

npx whyts --project ./before --runs 10 --json > before.json
npx whyts --project ./after --runs 10 --json > after.json
npx whyts compare before.json after.json

Order and warm-up

Each side runs once first, baseline then candidate. These warm-up runs are in the JSON under warmup and not in the numbers.
The first run reads all files from disk. The side that runs first would pay that cost alone. In our test, a first run with a cold file cache had a Total time 10% to 15% above the later runs. The Check time did not change.
The measured runs follow in pairs, and the order inside a pair flips each time: A B, B A, A B, B A. A is the baseline.
With this order, a slow drift of the machine gives no side an advantage. With an odd N, the baseline goes first once more than the candidate.
All runs use the untraced mode of --runs. The JSON shows the order in order.

The noise rule

The rule is simple, and it is not a statistical test.

For each side, take the smallest and the largest measured value.
If the two ranges overlap, the result is within-noise. Ranges that only touch overlap.
If the ranges do not overlap, the result is separated. The candidate is faster or slower.
Each side needs at least 3 measured runs. With fewer, the result is insufficient-runs.

The difference of the medians is in ms and in percent of the baseline median. The rule gives no confidence interval and no effect size.

If both sides have the same distribution and the runs are independent, the chance of ranges that do not overlap is 2 divided by C(2N, N). This is 10% for N = 3, 0.79% for N = 5 and 0.0011% for N = 10. The JSON shows it as a fraction in rule.chanceWithoutDifference. A load spike or a drift breaks the assumption. Use at least 5 runs. Stop other heavy work on the machine during a comparison.

The rule is conservative. A small slowdown can give within-noise. More runs widen the ranges, so more runs do not always help. In our tests, a Check time increase of about 10% was sometimes within-noise, and an increase of about 30% was always separated. See VALIDATION.md for the numbers. An unseparated result does not prove that there is no difference.

Checks before the comparison

whyts stops with exit code 1 when:

one side uses the native compiler and the other side does not,
the checker counts differ,
the flags of the measurement runs differ,
one report has timings and the other report has none,
a report has an unknown schema version, or a file is not a whyts report.

whyts prints a warning and continues when:

the TypeScript versions differ,
the file counts of the two programs differ (the Files line of the compiler),
a live comparison finds different compiler options in the two tsconfig files,
both sides report the same compiler errors (the warning starts with COMPILER ERRORS),
unresolved modules are most of the errors,
the whyts versions of the two reports differ.

Compiler errors and comparable sides

A compiler can stop early on an error, for example TS2589. Then the compiler does less work and its time is short. Two sides that do different work have no valid time comparison.

whyts calls the sides not comparable when:

one side has compiler errors and the other side has none,
both sides have errors, and the error counts or the error codes differ.

If the sides are not comparable, whyts prints NOT COMPARABLE and the reason. It shows the raw times and the error summary of each side. It shows no difference and no verdict.

If both sides have the same errors, whyts keeps the verdict and prints the COMPILER ERRORS warning. A report that has no error data gives no code comparison. A clean side still decides.

Chains in an offline comparison

whyts matches source check chains by file, line and character, and file check intervals by file. An edit above a chain moves its line. The chain then shows as gone at the old line and new at the new line. The result lists each entry as new, gone, grown, shrunk or unchanged. An entry grew or shrank when it changed by at least 10 ms and at least 10%. These limits are display limits, not statistics.

These values come from one traced run for each report. They are samples. A chain that is only in one report can be below the 10 ms limit in the other report, or outside its five largest chains. whyts compares such a chain when the other report records it. It lists the chain as new or gone when the other report does not.

Output and exit codes

--json prints the comparison. It has kind (comparison), mode, baseline, candidate, checkTime, totalTime and warnings. A live comparison adds runs, order, warmup and measuredRuns. An offline comparison adds findings. Progress goes to stderr.

Version 0.7.1 adds three fields. Schema version 1 stays the same. comparable is true or false. reason is null, or the text that says why the sides are not comparable. baseline.compilerErrors and candidate.compilerErrors have count and codes, or null when a report has no error data. When comparable is false, the verdict of checkTime and totalTime is not-comparable. The difference fields are null. The time summaries stay.

Code	Meaning
`0`	The comparison finished and the sides are comparable. Equal compiler errors on both sides give a visible warning, not another code.
`1`	A failure: a bad argument, an unreadable file, a compiler failure or a timeout, or a rejected comparison. Also sides that are not comparable.

The exit code is 1 for sides that are not comparable, because the run has no valid result. A script that checks only the exit code must not read it as a pass. The output is still complete. Use comparable in the JSON to tell this case from a failure. A failure prints to stderr and gives no JSON.

Use with AI coding agents

An AI coding agent cannot find slow types by reading code. The raw traces are too large for its context. whyts has an MCP server (Model Context Protocol, over stdio). It gives the agent a few KB of measured JSON instead.

Start it with npx -y whyts-mcp. The package whyts-mcp depends on whyts and on the MCP packages, so npm install whyts does not install them. These tools are available:

Tool	Use it to	Time
`analyze`	Find where the check time goes. Returns Check time, compiler error summary, `diagnoses` with a remedy, the top 5 measured findings, and warnings.	One traced type check, plus `runs` untraced checks
`compare`	Prove that a change made the check faster or slower. Returns `separated`, `within-noise` or `not-comparable`, with medians and the reason.	About 2 x `runs` + 2 type checks (`runs` is 5 by default)
`explain`	Show why one file is in the program: the import chain and the direct importers.	Seconds. It does not run the type checker.

Each result is one JSON text, at most 12,000 characters. If whyts shortens a result, the field truncated says what it cut. Paths are absolute. Errors return isError with a kind and a hint. whyts runs one call at a time, because two type checks at once distort the timings. When the client sends a progress token, whyts sends a progress notification every 15 seconds and at each timing run.

Example, a compare result (runs 5):

{"tool":"whyts","mode":"compare","decision":"separated","direction":"faster",
 "rationale":"Every run of the candidate was faster than every baseline run (5 runs per side; the min-max ranges do not overlap). ...",
 "checkTime":{"baseline":{"median":0.21,"min":0.2,"max":0.22},"candidate":{"median":0.03,"min":0.03,"max":0.04},"deltaMilliseconds":-180,"deltaPercent":-85.7}}

Set up a client

Claude Code:

claude mcp add --transport stdio whyts -- npx -y whyts-mcp

Add --scope project to write the entry to .mcp.json for your team. Claude Code limits each tool call with the timeout field (milliseconds) of the entry. Progress notifications do not extend that limit.

Cursor, in .cursor/mcp.json (project) or ~/.cursor/mcp.json (all projects):

{
  "mcpServers": {
    "whyts": { "command": "npx", "args": ["-y", "whyts-mcp"] }
  }
}

VS Code with GitHub Copilot, in .vscode/mcp.json:

{
  "servers": {
    "whyts": { "command": "npx", "args": ["-y", "whyts-mcp"] }
  }
}

Codex:

codex mcp add whyts -- npx -y whyts-mcp

Codex stops a tool call after 60 seconds, and it stops a server start after 10 seconds. Raise both in ~/.codex/config.toml:

[mcp_servers.whyts]
command = "npx"
args = ["-y", "whyts-mcp"]
startup_timeout_sec = 60
tool_timeout_sec = 900

Other clients: run the command npx -y whyts-mcp over stdio. The server supports the 2025-11-25 protocol and the 2026-07-28 protocol.

The server is published to the MCP Registry as io.github.musatoktas/whyts.

Give your agent the short rules in docs/agents.md.

The server runs the TypeScript compiler of the project that you name, as tsc does. It does not edit the project. The typescript input names a compiler package to run, so give the agent only paths that you trust.

The MCP packages (@modelcontextprotocol/server and zod) are not dependencies of whyts. npm install whyts installs only typescript, as before. The package whyts-mcp has the same version as whyts and depends on that exact version. The command whyts mcp still works. If the MCP packages are missing, it prints npx -y whyts-mcp.

Alternative, if you want to start the server from the whyts package: npx -y -p [email protected] -p @modelcontextprotocol/server@2 -p zod@4 whyts mcp.

Options for large projects

Option	Meaning
`--verbose`	Print the full report. The default prints at most three actions and one summary line.
`--timeout <seconds>`	Compiler time limit. The default is 900.
`--runs <N>`	Do N untraced runs after the traced run. The report and the JSON show the median, smallest and largest value. The default is 1. See Measure repeatedly.
`--max-old-space-size <MB>`	Set the heap limit of the compiler process. The range is 256 to 1048576. TypeScript 7 ignores it.
`--typescript <path>`	Use this TypeScript package directory, or its `lib/typescript.js`, instead of the project compiler. The version may be 5.x, 6.x or 7.x.

node src/cli.js --project packages/compiler --max-old-space-size 10240
node src/cli.js --project . --typescript node_modules/typescript

The trace writes a type dump after the check. The compiler does not include the dump in Total time. In the drizzle-orm type-tests project, Total time was 13.58 s and Dump types time was 249.3 s. The compiler process took 263.5 s. The report shows the wall time and the dump time separately.

The default timeout is 900 s. That is more than 3 times the longest compiler process in the 0.4 tests (263.5 s). A progress note goes to stderr every 30 s while the compiler runs.

If a signal stops the compiler, the error shows the last lines of compiler stderr. For a heap error, SIGABRT, SIGSEGV or SIGKILL, the error tells you to use --max-old-space-size. The operating system out-of-memory killer or a native crash can also cause SIGKILL or SIGABRT.

The TypeSpec compiler project stopped in the default heap. It finished with --max-old-space-size 10240.

Compiler errors

The report shows the first five errors with file, line and code. It also shows the count of each error code.

Half or more of the errors can be unresolved modules or declarations (TS2307, TS2792, TS2688, TS7016, TS6053). Then the dependencies are probably missing or not built. The timings can be wrong. The report shows a warning.

Example: react-router in the TanStack Router repository gave 1887 errors and a Check time of 4.76 s before the build. After the build it gave 0 errors and 12.43 s.

JSON version 0.4 adds these fields. Schema version 1 and all existing fields stay the same.

compilerErrors: total, first, codes, missingDependencyErrors, measurementMayBeInvalid.
summary.dumpTypesSeconds.
checkTimeShareUpperBoundPercent on file intervals, source checks and chains.
evidence.testRootsExcluded on review-root-files.

summary.errorCount now counts the parsed compiler error lines.

TypeScript 7 (experimental)

whyts can profile a project with the native TypeScript 7 compiler. This support is experimental.

To use it, install typescript@7 in the project, or pass the package directory with --typescript. Then run whyts as usual.

npx whyts --project . --typescript node_modules/typescript

How it works:

whyts starts bin/tsc of the TypeScript 7 package with --checkers 1.
The native compiler numbers the types of each checker separately. With more than one checker, the type IDs in the trace collide. One checker keeps each ID unique.
The native trace stores source positions as UTF-8 byte offsets. whyts converts them to UTF-16 offsets.
The native types_0.json file stores declaration paths in lower case. whyts matches them to the real file names.
TypeScript 7 has no JavaScript API for module resolution. whyts builds the import graph and reads the tsconfig with its own typescript 5.x or 6.x package.

Known limits:

The report and the JSON mark the run as experimental. compiler is native, experimental is true, and checkers is 1.
Check time comes from one checker. A default tsc run uses more checkers, so its Check time can differ. Do not compare the two.
--max-old-space-size has no effect, because the native compiler is not a Node.js process. whyts prints a warning and ignores the option.
The native compiler prints no Dump types time line. The report does not show it, and summary.dumpTypesSeconds is null.
whyts warns when the graph and the native compiler report different file counts. Module resolution can differ between the two compilers. Treat the import graph findings as uncertain then.
whyts stops with an error when its own TypeScript cannot read your tsconfig. A setting that only TypeScript 7 knows can cause this.
We tested TypeScript 7.0.2 on Linux x64 only. We did not test Windows, macOS or other 7.x versions.

How it works

Resolve the project's installed TypeScript compiler, falling back to the bundled 5.9 compiler. A TypeScript 7 package selects the native compiler.
Read the effective config using the compiler API, including JSONC and extends.
Build a source/import graph and inspect the loaded type packages.
Run that compiler with --noEmit, --extendedDiagnostics, and --generateTrace.
Group temporally nested source checks within each file/thread, match contained type comparisons, stream selected type IDs from types.json, and rank measurements before structural findings.
Delete temporary trace/cache files.

The compiler receives a new temporary incremental cache. Your existing build cache is neither read nor overwritten. Source files and normal build outputs are unchanged. No package lifecycle scripts are invoked and the CLI makes no network requests. Installing dependencies or fetching the CLI can, of course, access the network.

Measurement boundaries

Measures a fresh-cache, traced type check. With --runs, it also measures fresh-cache untraced checks. It is not a timer for Next.js builds, bundling, tests, or the language server.
Trace generation and an empty cache affect timings. Compare runs with the same compiler, flags, hardware, and tracing mode. The report's wall timer excludes the earlier graph analysis.
Recorded check intervals are inclusive and expression/type events may be sampled. Overlapping intervals for the same file, source check, chain root, or type pair are merged. The five largest file checks and five largest project file checks are reported separately, plus five source chains (project first), up to five members and three contained comparisons per chain, and five global type comparisons. The five individual source checks remain in JSON for compatibility. These lists overlap and cannot be summed into total check time.
A comparison is associated with the innermost recorded source check only when its whole span fits inside that check on the same process/thread. The chain also collects comparisons contained in its members; they may occur outside the selected focus expression. Unrecorded checks, missing positions, or unavailable type descriptors reduce detail; whyts does not infer a missing causal chain.
A check-hotspot finding is omitted when a single source check chain in the same file covers at least 80% of that file's recorded check time, because the chain finding already reports the same time. The file stays in the "Largest recorded project file-check intervals" list. The threshold is CHAIN_COVERAGE_THRESHOLD (0.8) in src/analyze.js.
Barrel reach counts observed source edges, including type-only edges. Files may already be configured roots, and a direct import may leave the program size unchanged.
To bound graph traversal, at most 40 barrel candidates are checked, ranked first by reexport count; five are reported. Trace files over 128 MiB are not parsed. Type descriptors are streamed in 64 KiB chunks with a 1 GiB scan budget, 4 MiB per-record budget and 16 MiB retained-data budget. Scanning stops when the selected IDs are found, so the remaining file is not validated. Limits or missing IDs produce warnings; file size alone does not disable type resolution. Compiler output is bounded at 16 MiB. Source snippets are limited to 240 characters and type labels to 160 characters.
Select a leaf tsconfig in a monorepo. whyts does not run tsc --build or build referenced projects. Missing/stale referenced declaration outputs can affect results.
Types, Instantiations and Memory used in diagnostics come from the trace run. The trace makes them larger. In the TypeSpec compiler project, the traced run reported 10,480,436 types. An earlier test run without a trace reported 135,624 types. We did not repeat the run without a trace for 0.4. Do not compare them with output of a normal run.

For further trace exploration, use Microsoft's analyze-trace. Compiler documentation: extendedDiagnostics, generateTrace, and performance guidance.

Development

npm ci --ignore-scripts
npm run check
npm test
npm pack --dry-run

Tests cover import chains, path aliases, cycles, dynamic imports, inherited configs, actually loaded duplicate types, nested and out-of-order trace events, chain grouping across source positions, thread/file isolation, comment-aware declarations, selective descriptor streaming with UTF-8 chunk boundaries and budgets, missing descriptors, project prioritization, barrel root overlap, terminal sanitization, real traced checks, cache preservation, compiler errors, JSON output, and timeouts. Further tests cover the run order, the warm-up exclusion, the summary values, the noise rule, the checks before a comparison, and the offline matching of findings. The MCP tests start whyts mcp with a raw stdio client and check the tool list, the compact results, the error paths, progress notifications and the stop of a running compiler. GitHub Actions runs Node 20/22/24 on Linux, Windows, and macOS.

The implementation is plain ESM JavaScript so a checkout runs without a build step. TypeScript is the only required dependency. The MCP server is in the separate package whyts-mcp, which adds two more packages.

Bug reports are most useful with a tiny reproduction, Node/TypeScript versions, and expected versus actual output. Reports and traces may contain private filenames, source snippets, and type names; review them before sharing.

MIT license. Created by Musa Toktas.

Whyts

Diesen Server hinzufügen

Host a copy with our keys and logs.

Server facts

Tools

Questions