🚧 THIS PAGE IS UNDER CONSTRUCTION 🚧

~ CODE GRAPH RESEARCH ~

😺 😻 😼 😽 🙀 😿 🐱

*** WELCOME 2 MY RESEARCH DOSSIER!!! ***

★ NEW ★ Last updated: 21 AUG 2026  |  4 AGENTS · 15 THREADS

🐱 HOT NEWS 🐱 SERENA TOOK 11 HOURS 2 INDEX LLVM 🐱 CODEGRAPH DIED AT 44% 🐱 AIDER WENT 16X OVER BUDGET 🐱 NOBODY HAS EVER EVALUATED ANY OF THIS 🐱

😻 THE VERDICT 😻

BUILD IT. ADOPT NOTHING.

No tool off the shelf does build a graph → rank it → summarize it. Every single one has a documented scale failure, zero independent evidence, or a star count u cannot verify!!!

Meanwhile structural retrieval beats dense embeddings in EVERY published head-to-head we found, and 120 lines of stdlib python scouted repos nobody had ever opened in under a second. 😺

🐱😺😻😼😽😾😿🙀🐱😺😻😼😽😾😿🙀

🙀 THE SCALE WALL 🙀

every number here is from a GITHUB ISSUE not a blog!!! 🙀

TOOLWHAT HAPPENEDCAT VERDICT
Serena11 HOURS to index LLVM (7,749 files)🙀
Serena~30GB RAM, froze the machine, 3x🙀
CodeGraphOOM crash at 44% on 137,699 files😿
CodeGraph1.3% of edges WRONG after 80 commits🙀
aiderasked for 1,024 tokens, got 16,419😿
grepaiZERO independent reports. ever.🙈
OUR THING1,780 files in 9.4 SECONDS, 0 deps😻

(not apples 2 apples — serena builds a full semantic index. but we didnt ask for one!!)

😼 4 THINGS THAT CHANGED THE PLAN 😼

😺 #1 — REPO OVERVIEWS MIGHT NOT EVEN WORK

N=438 tasks. Static prose repo overviews did NOT reduce steps-to-first-file and did NOT improve resolution. LLM-generated ones MADE IT WORSE by 0.5–2%!!!

⇒ so: cache the COMMAND, not the MAP. it runs in 0.34s. why store it.

😻 #2 — STRUCTURE BEATS EMBEDDINGS

WHORESULT
DyCoder (ASE 2026)graph 43.06% vs dense 23.60% Pass@1
LocAgent (ACL 2025)+10pp file · +19pp function
4 hybrid studieshybrid wins by 9–30% relative

😼 #3 — BARREL FILES: NOBODY HAS SOLVED THIS

when index.ts re-exports everything, EVERY import points at the barrel and ur ranking dies. dependency-cruiser? no. skott? no. grimp? pydeps? NO!!! 🙀

the algorithm exists inside NEXT.JS as a bundler transform. nobody ever applied it 2 ranking.

😽 #4 — WE ACCIDENTALLY REBUILT AIDER EXACTLY

every constant matches repomap.py VERBATIM. and NOBODY HAS EVER EVALUATED IT.

“No other agent in the corpus uses graph-theoretic relevance ranking.”
— Inside the Scaffold, arXiv:2604.03515

so theres no state of the art 2 be behind!!! 😺

🙀😿😾😽😼😻😺🐱🙀😿😾😽😼😻😺🐱

🙈 THE NUMBER THAT HURTS 🙈

🙀
p = 0.087

adding an index gave +7.9pp resolve rate over NO index... but measured against a strong grep agent, the gain was NOT STATISTICALLY SIGNIFICANT.

translation: the case that ANY of this beats a competent agent with grep is not yet made. 🙀

😺 THE DECISION 😺

  1. Generate on demand. persist the recipe not the orientation
  2. Rank over IMPORTS (names give u get() and close(), useless!!) + linguist vendor exclusions
  3. Add the top rung — what is this project FOR
  4. ts.resolveModuleName + resolution cache. the cache is MANDATORY

NOT BUILDING: barrel collapse (checked ur repos — zero TS barrels!) and monorepo machinery (microservices!!) 😻

🏆 BEST VIEWED IN NETSCAPE NAVIGATOR 4.0
🏆 800x600 OR BETTER
🏆 CAT LOVERS WEBRING MEMBER
🏆 100% HAND-CODED CATS

YOU ARE VISITOR NUMBER
0000437

😻 SIGN MY GUESTBOOK! · EMAIL ME · ◀ WEBRING ▶ 😻

🎵 NOW PLAYING: canon_in_d.mid 🎵 🐱 thx 4 visiting my page!!! 🐱 dont forget 2 bookmark 🐱 🎵
🚧 MORE CATS COMING SOON 🚧

every finding on this page is real and sourced.
the sober version lives at plans/code-graph-report.html
full citations: held outside this repository.

🐱 made with NOTEPAD.EXE 🐱