Project home · Try it yourself · Measurement record
eggshell-29s-demo-intro.mp4
This 29-second demo uses actual Codex CLI recordings with Luna xhigh and
independent ephemeral chats. One chat investigates Clang’s toolchain selection.
Two new chats receive the same related question: one starts fresh with Eggshell
hooks disabled; the other receives relevant findings from the saved .egg.
The recording highlights the handoff delivery and the agent’s explicit reuse statement. Elapsed time is compressed, with pauses for those annotations. The terminal output and measured token totals are unchanged.
| Work | Fresh + fresh | Fresh + Eggshell |
|---|---|---|
| Shared first investigation | 835,305 | 835,305 |
| Second-chat follow-up | 1,074,497 | 195,368 |
| Both chats | 1,909,802 | 1,030,673 |
81.8% fewer tokens in the second chat; 46.0% fewer across both chats. The same first investigation is charged to both routes. Tokens mean input plus output, including cached input; reasoning is already included in output. These are token counts, not dollar-cost estimates. Memory organization makes no extra LLM calls; the selected handoff still consumes normal model input tokens.
Both final answers passed five static-source review criteria covering selection order, wrapper restrictions, relevant source/tests, and accurate reporting of what was not run. This was one recorded pair and a single non-blinded reviewer, without Clang runtime tests. It does not establish general savings or equivalent answer quality across tasks. This recording uses an earlier build; the README’s current three-topic study measures the indexed source build separately. The earlier walkthrough is also a separate recording.
In this Clang source snapshot, can an architecture-scoped -Xarch_* argument
change the already selected host toolchain in the same way that a direct
-target, -m32, or -m64 option can? Explain the relevant ordering and constraints,
with function names and source-line citations, and identify a relevant existing
driver test. Keep the final answer within 250 words. Do not modify files or
build LLVM. Distinguish static source findings from untested runtime behavior.
Both follow-ups used the same prompt, model, effort, and fixed LLVM source commit. Source, prompt, prior-work, runtime, answer and video hashes are included in the measurement record. Raw private execution logs remain local; hashes identify those records but are not a substitute for access to them.
The key ordering is visible in host toolchain selection and per-toolchain argument translation.