Audit: files too large for agents, tooling, and GitHub UI

[tannergooding]

Summary

Three distinct problems — agent/LLM token cost, GitHub tool interaction, and GitHub UI/search — all stem from the same root cause: files that have grown without a size budget. This issue tracks candidates for splitting and flags files that break specific documented thresholds.

Audit methodology: git ls-tree -r -l HEAD for byte sizes, git show | split for per-file line and max-line-length, git log --oneline --since=2.years for 2-year commit counts. No large files were opened whole; the audit itself was non-destructive.

---

Confirmed thresholds

Limit	Value	Source
view tool truncation	20 KB	Tool documentation
Agent token cost	morph.cpp (~585 KB) ≈ 137k tokens ≈ 200 AI credits per full read	Measured in this repo (tiktoken/o200k_base)
Code search file exclusion	> 350 KiB	docs.github.com
Code search line truncation	Lines > 1,024 chars	Same
Code search full exclusion	Files with > 1 line over 4,096 bytes	Same
Per-PR diff	> 20,000 lines or > 1 MB	docs.github.com
Per-file diff	> 20,000 lines or > 500 KB not shown; 400 lines / 20 KB auto-loaded	Same
Blob view syntax highlighting	> ~1 MB (unverified — no official docs page found)	Community reports

160 files in the repo currently exceed the 350 KiB code search exclusion.

---

Native (JIT core)

Files are hand-maintained, in active development, and exceed the code search exclusion and/or the per-file diff limit.

Path	Size (KB)	Lines	2y commits	Exceeds
src/coreclr/jit/gentree.cpp	1,210	36,981	289	All limits; ~282k tokens, context-window overrun
src/coreclr/jit/compiler.h	524	13,724	392	Code search, diff; highest-churn file in repo
src/coreclr/jit/emitxarch.cpp	637	20,974	82	Code search, diff; at the 20k-line diff boundary
src/coreclr/jit/emitarm64sve.cpp	811	20,030	23	Code search, diff; 20k-line boundary
src/coreclr/jit/morph.cpp	572	16,231	190	Code search, diff
src/coreclr/jit/valuenum.cpp	548	16,149	115	Code search, diff
src/coreclr/jit/lsra.cpp	557	14,024	46	Code search, diff
src/coreclr/jit/importer.cpp	563	14,788	129	Code search, diff
src/coreclr/jit/importercalls.cpp	502	12,725	158	Code search, diff
src/coreclr/jit/lower.cpp	479	13,249	177	Code search
src/coreclr/jit/compiler.cpp	362	10,851	169	Code search
src/coreclr/jit/codegenxarch.cpp	428	11,645	104	Code search
src/coreclr/jit/lowerxarch.cpp	412	10,919	103	Code search
src/coreclr/jit/instrsxarch.h	439	1,411	42	Code search; 1,411 lines in 439 KB (avg 318 bytes/line)
src/coreclr/jit/emit.cpp	359	10,922	65	Code search
src/coreclr/jit/gentree.h	306	10,700	122	Code search

Native (VM / Debug / Mono)

Path	Size (KB)	Lines	2y commits	Exceeds
src/coreclr/vm/jitinterface.cpp	521	15,982	188	Code search, diff
src/coreclr/vm/methodtablebuilder.cpp	520	13,172	53	Code search, diff
src/coreclr/vm/excep.cpp	376	10,760	82	Code search
src/coreclr/vm/proftoeeinterfaceimpl.cpp	324	11,093	19	Code search
src/coreclr/debug/ee/debugger.cpp	529	16,018	65	Code search, diff
src/coreclr/debug/di/process.cpp	529	14,793	50	Code search, diff
src/coreclr/debug/di/rspriv.h	421	11,634	41	Code search
src/coreclr/tools/superpmi/superpmi-shared/methodcontext.cpp	320	7,837	59	Code search
src/mono/mono/mini/mini-llvm.c	510	15,532	16	Code search, diff
src/mono/mono/mini/aot-compiler.c	498	15,976	12	Code search
src/mono/mono/mini/method-to-ir.c	452	14,048	12	Code search

Managed

Path	Size (KB)	Lines	Max line (chars)	2y commits	Exceeds
src/libraries/System.Runtime.Intrinsics/ref/System.Runtime.Intrinsics.cs	2,059	11,987	781	84	Code search; ~490k tokens; unverified blob-view limit
src/libraries/System.Runtime/ref/System.Runtime.cs	1,571	18,273	2,652	111	Code search; max line trips 1,024-char search truncation; ~374k tokens
src/libraries/System.Private.CoreLib/src/System/Runtime/Intrinsics/Arm/AdvSimd.cs	975	16,318	355	4	Code search; ~233k tokens
src/libraries/System.Private.CoreLib/src/System/Threading/Tasks/Task.cs	387	7,759	233	30	Code search
src/libraries/System.Private.Xml/src/System/Xml/Core/XmlTextReaderImpl.cs	353	9,751	185	3	Code search
src/libraries/System.Text.RegularExpressions/gen/RegexGenerator.Emitter.cs	327	5,902	259	24	Code search
src/libraries/System.Linq.Parallel/src/System/Linq/ParallelEnumerable.cs	322	5,967	215	5	Code search
src/libraries/System.Private.CoreLib/src/System/MemoryExtensions.cs	338	6,473	262	19	Code search

.github instruction / skill files

A separate class of problem: these files are loaded as context by LLM agents. The 20 KB view tool truncation applies to files an agent reads at runtime (build config, KB docs, referenced instruction files). Empirical testing confirmed that SKILL.md files are injected in full by the skill loader — not via view — so their content is not silently truncated on invocation. Large skill files still consume more context budget per use, and any secondary files those skills reference at runtime (e.g. pr-assessment.md, .github/instructions/*.md) remain subject to the 20 KB truncation.

File	Size	Lines	Problem
.github/workflows/ci-failure-scan-feedback.lock.yml	126 KB	—	Auto-generated gh-aw lockfile; looks like a real workflow to glob, contains only SHAs and hashes
.github/workflows/holistic-review.lock.yml	121 KB	—	Same
.github/workflows/ci-failure-fix.lock.yml	119 KB	—	Same
.github/workflows/closed-issue-reference-check.lock.yml	118 KB	—	Same
.github/workflows/ci-failure-scan.lock.yml	114 KB	—	Same
.github/workflows/breaking-change-doc.lock.yml	112 KB	—	Same
.github/workflows/holistic-review-orchestrator.yml	34.9 KB	736	Real workflow; 1.75x view truncation limit; truncated when read at runtime
.github/workflows/ci-failure-scan.md	32.4 KB	409	Knowledge-base doc; truncated when read at runtime
Directory.Build.props	32 KB	405	Read via view when investigating any build issue; always truncated
.github/workflows/ci-failure-fix.md	29.1 KB	415	Truncated when read at runtime
.github/skills/issue-triage/SKILL.md	28.2 KB	487	Injected in full on invocation (empirically verified — all 487 lines present); 28 KB context cost per use
.github/agents/system-net-review.agent.md	26 KB	427	Agent definition; loading mechanism not empirically tested
.github/workflows/ci-failure-scan-feedback.md	27.2 KB	349	Truncated when read at runtime
.github/skills/code-review/SKILL.md	22.6 KB	203	Injected in full on invocation (empirically verified — all 203 lines present); 22 KB context cost per use
.github/skills/api-proposal/SKILL.md	22.4 KB	389	Expected to inject in full (same mechanism); context cost per use
.github/workflows/shared/create-kbe.instructions.md	22.1 KB	522	Truncated when read at runtime
.github/agents/extensions-reviewer.agent.md	20.2 KB	245	Agent definition; loading mechanism not empirically tested

The six .lock.yml files are the most acute: each is ~120 KB of binary-encoded metadata that an agent cannot reason about, and all six appear in any *.yml glob of .github/workflows/.

Generated / by-design (not action items)

• src/coreclr/pal/prebuilt/inc/corprof.h / cordebug.h / xclrdata.h — MIDL-generated ("this ALWAYS GENERATED file")
• src/coreclr/ilasm/prebuilt/asmparse.cpp — Bison-generated
• src/coreclr/debug/ee/amd64/amd64InstrDecode.h — machine-generated, see gen_amd64InstrDecode/README.md
• src/tools/ilasm/src/ILAssembler/gen/CILParser.cs — <auto-generated>
• src/tests/JIT/jit64/opt/cse/hugeexpr1.cs and siblings — intentionally enormous CSE stress tests
• src/native/external/** — vendored external code (brotli, zlib-ng, zstd, valgrind)
• *.PlatformNotSupported.cs intrinsics stubs — maintained alongside their real counterparts by design

---

Candidate work items (priority order)

1. Split src/coreclr/jit/gentree.cpp — 1.2 MB / 37k lines / 289 commits/2y. One full read exceeds the 128k-token context of most models. Natural seams: node construction, printing, fgMorphXxx helpers, flag management, assertion logic.

2. Decompose src/coreclr/jit/compiler.h — 524 KB / 13.7k lines / 392 commits/2y (highest churn in repo). Included by every JIT translation unit. Split by phase state, liveness helpers, assertion structs, opt flags.

3. Split importer pipeline — importer.cpp (563 KB / 129 commits) + importercalls.cpp (502 KB / 158 commits). Both exceed code search and diff limits. Further split by opcode category (arithmetic/compare, memory, control flow, calls).

4. Address ref assembly size/line-length — System.Runtime.cs (1.57 MB, max line 2,652 chars trips code search 1,024-char truncation). System.Runtime.Intrinsics.cs (2.05 MB, ~490k tokens). Whether split by logical area or otherwise is an api-review-process question.

5. Split src/coreclr/vm/jitinterface.cpp — 521 KB / 16k lines / 188 commits/2y. Second-highest-churn VM file; every meaningful change produces a partially-hidden diff. Natural split: helper implementations, EE callbacks for type/method queries, JIT code manager helpers.

6. Split src/coreclr/jit/morph.cpp + lower.cpp — morph.cpp (572 KB / 190 commits) is the calibration file at ~200 AI credits per read. lower.cpp (479 KB / 177 commits) already has a precedent split with lowerxarch.cpp. Both exceed code search and diff limits.

Note

This issue was researched and drafted with GitHub Copilot (Claude Sonnet 4.6). The analysis methodology, threshold verification, and prioritization were performed autonomously from git metadata without opening large files. Thresholds were confirmed against official documentation where available; the blob-view rendering limit (~1 MB) is noted as unverified.

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-meta
See info in area-owners.md if you want to be subscribed.