pr 29008 real dev environment verification, 2026-11-03, linux boat vm baseline $ git worktree add ../main github/main HEAD cefb544af3c98224874419198c0359b652ca3c94 $ cd ../main && timeout 600 vp i passed $ T3CODE_DEV_SEED=0 setsid nohup vp run dev > /tmp/main-dev.log 2>&1 & captured pid 20392; serverPort=16733 webPort=8693; isolated /home/user/main/.t3 repository playwright and /usr/bin/google-chrome drove the web client. added /home/user/t3code as a real project through add project > local folder. selected codex gpt-6.1-sol and submitted the read-only verification prompt. real response: 40 repetitions of $file-upload $unknown-inline-check in one paragraph. actual provider catalog: 29 skills. cdp conditional logpoint at the loaded matcher's skills.find expression: one complete scan of that 1,399-character response paragraph recorded 80 catalog searches. no mocked provider, fixture messages, synthetic skill catalog, or source instrumentation. $ kill -TERM -- -20392 main server stopped; both main ports absent before starting pr. pr $ git fetch https://github.com/luke2x/t3code.git perf/cache-repeated-inline-skill-lookups $ git switch -c perf/cache-repeated-inline-skill-lookups ca27c136898bcc2180f3630a55e45ea82d28162e exact HEAD ca27c136898bcc2180f3630a55e45ea82d28162e $ timeout 600 vp i passed $ T3CODE_HOME=/home/user/t3code/.t3 T3CODE_DEV_SEED=0 setsid nohup vp run dev > /tmp/dev.log 2>&1 & captured pid 27392; serverPort=13773 webPort=5733; isolated /home/user/t3code/.t3 created project and thread through the web app; same codex gpt-6.1-sol prompt. real response: same 1,399-character repeated paragraph and actual 29-skill catalog. equivalent cdp logpoint: 2 catalog searches, one known name and one unknown name. this measures catalog searches in the real renderer, not end-to-end latency or frame time. nearby behavior through actual client interactions 40 assistant skill chips rendered; all 40 unknown tokens stayed plain text. inline code: $file-upload; link: $file-upload -> https://example.com/ chip > view instructions opened /home/user/.agents/skills/file-upload/SKILL.md. closed the file panel, reloaded the conversation, and asserted again: {"chips":40,"unknown":40,"code":["$file-upload"],"links":[{"text":"$file-upload","href":"https://example.com/"}]} captured and inspected before/after screenshots and a 5.8-second interaction gif. checks, all through blacksmith-check $ vp i && vp test run packages/shared/src/inlineSkills.test.ts 1 file passed, 6 tests passed. remote snapshot initially lacked history, so fetched the contributor branch with depth 16. $ vp run check:changed 30084f0018c5751a07291222e8d04e16d513700a 2 changed files; format passed; lint passed; knip 1/2 passed; knip 2/2 passed; 33 affected package typechecks passed; 2 selected files / 7 tests passed. unverified native desktop, ios/android, other providers, catalog changes between renders, production-build/high-volume performance, and end-to-end frame-time/latency improvements. no code changes, commits, or pushes were needed.