Show the recorded two-chat demo in the README

README.md

@@ -21,7 +21,7 @@
   <a href="https://chatgpt.com/plugins/plugins_6aa482a5d9048191a727260b5f898078">Get the plugin</a> ·
   <a href="#install">Install</a> ·
   <a href="docs/try-it.md">Try it yourself</a> ·
-  <a href="docs/demo.md">30-second demo</a> ·
+  <a href="docs/cli-demo.md">29-second demo</a> ·
   <a href="#how-it-works">How it works</a> ·
   <a href="#evidence">Evidence</a> ·
   <a href="#control-and-inspection">Controls</a> ·
@@ -39,10 +39,15 @@ Memory is built and organized locally, **without LLM calls or additional billed
 tokens for memory management**. These results cover one task with existing
 prior work; see [Evidence](#evidence) for the comparison and its limits.
 
-**[Watch the 30-second walkthrough](docs/demo.md)** or
-**[try it in two chats](docs/try-it.md)**. Ask ordinary questions: Eggshell
-automatically brings relevant findings from the first investigation into the
-follow-up. The walkthrough is an edited English summary of the measured study.
+**Watch Eggshell carry work into a new chat — 29 seconds.**
+
+https://github.com/user-attachments/assets/939eca34-4897-4bcf-9dbc-a6326bc9ba3a
+
+Actual Codex CLI, Luna xhigh, independent ephemeral chats. In this recorded pair,
+Eggshell used **81.8% fewer input + output tokens in the second chat**; both
+answers passed static source review. One pair, with time edited for clarity.
+[Demo and measurement details](docs/cli-demo.md) ·
+[Try it in two chats](docs/try-it.md).
 
 Use the published **Codex plugin**, or the separate, experimental
 [adapters for Claude Code, Gemini CLI, Cursor, and OpenCode](adapters/README.md).

docs/benchmarks/cli-demo.json

@@ -0,0 +1,73 @@
+{
+  "description": "One recorded Clang follow-up comparison in independent ephemeral Codex chats.",
+  "model": "gpt-5.6-luna",
+  "reasoning_effort": "xhigh",
+  "ephemeral": true,
+  "llvm_commit": "6dfe1677ab8dffbc6ec13d53a1e0215d75147689",
+  "source_tree": "1dbf5dbe50ba286d5e9e008bbd230e47d4de0dbd",
+  "initial_prompt_sha256": "ac3c77019e4a2c83c7b0c8253a391fb1275cc7d01b0d063c5d42b3b910c1b31f",
+  "followup_prompt_sha256": "8a6ee07fec059fbea973da71398ffd6ba60c9f78c0765fce6fe5b9d57eae4acf",
+  "prior_egg_sha256": "f087d3548e83de69f8dc9029578203886a51857d304239ed215bb5c83c3b0591",
+  "runtime_sha256": "d9d687ce1cb4cc994ebfa99d9bed7b4a6784841b7a83c53f09ed7592d19813f0",
+  "shared_initial_investigation": true,
+  "token_definition": "Input plus output, including cached input. Reasoning output is already included in output.",
+  "runs": {
+    "initial": {
+      "input_tokens": 828632,
+      "output_tokens": 6673,
+      "cached_input_tokens": 725504,
+      "total_tokens": 835305
+    },
+    "fresh": {
+      "input_tokens": 1066830,
+      "output_tokens": 7667,
+      "cached_input_tokens": 939520,
+      "total_tokens": 1074497
+    },
+    "eggshell": {
+      "input_tokens": 188254,
+      "output_tokens": 7114,
+      "cached_input_tokens": 149760,
+      "total_tokens": 195368
+    }
+  },
+  "totals": {
+    "fresh_plus_fresh": 1909802,
+    "fresh_plus_eggshell": 1030673,
+    "followup_reduction": 0.8181772494478812,
+    "overall_reduction": 0.4603246828728842
+  },
+  "quality": {
+    "method": "Single-reviewer, non-blinded comparison with the fixed source snapshot and native execution records. No runtime tests were requested or run; this is not a general equivalence claim.",
+    "criteria": [
+      "Direct target and architecture options can affect the triple before initial host toolchain lookup.",
+      "Architecture-scoped wrapper expansion happens later and does not retroactively choose the already selected host toolchain.",
+      "The answer covers architecture/offload matching and the parser or NoXarchOption restrictions.",
+      "The answer names a relevant existing test and source locations.",
+      "The answer distinguishes static source evidence from runtime behavior that was not tested."
+    ],
+    "fresh": {
+      "result": "usable",
+      "criteria_satisfied": 5,
+      "answer_sha256": "25bdcdfb449d725d7d0aa7ddd48aa015fe92edf096b781d16a55e759f00820e6"
+    },
+    "eggshell": {
+      "result": "usable",
+      "criteria_satisfied": 5,
+      "answer_sha256": "d894cb324e92f1619c97a12c26f2204fcc489c8790a439b6447511e20ca77954"
+    }
+  },
+  "video": {
+    "url": "https://github.com/user-attachments/assets/939eca34-4897-4bcf-9dbc-a6326bc9ba3a",
+    "duration_seconds": 29,
+    "sha256": "736434e89b1393b2c0949530613bd47295f14481e22e34b77fe01896ee174261",
+    "editing": "Elapsed time compressed, with pauses and editorial annotations at handoff delivery and explicit reuse. Native output text and measurements unchanged."
+  },
+  "limitations": [
+    "One recorded pair; results vary by task.",
+    "Shared initial investigation is charged to both routes, not two independent initial runs.",
+    "Static source review by one non-blinded reviewer; no Clang runtime tests.",
+    "Token reduction is not a dollar-cost estimate.",
+    "Raw private transcripts are not published; hashes identify local evidence and do not replace access to it."
+  ]
+}

docs/cli-demo.md

@@ -0,0 +1,55 @@
+# Two chats. Less repeated work.
+
+[Project home](../README.md) · [Try it yourself](try-it.md) · [Measurement record](benchmarks/cli-demo.json)
+
+https://github.com/user-attachments/assets/939eca34-4897-4bcf-9dbc-a6326bc9ba3a
+
+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.
+
+## The recorded comparison
+
+| 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. The README’s [ten-trial study](../README.md#evidence)
+and its [earlier walkthrough](demo.md) are a separate experiment.
+
+## The follow-up question
+
+```text
+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](https://github.com/llvm/llvm-project/tree/6dfe1677ab8dffbc6ec13d53a1e0215d75147689).
+Source, prompt, prior-work, runtime, answer and video hashes are included in the
+[measurement record](benchmarks/cli-demo.json). 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](https://github.com/llvm/llvm-project/blob/6dfe1677ab8dffbc6ec13d53a1e0215d75147689/clang/lib/Driver/Driver.cpp#L1695-L1722)
+and [per-toolchain argument translation](https://github.com/llvm/llvm-project/blob/6dfe1677ab8dffbc6ec13d53a1e0215d75147689/clang/lib/Driver/Compilation.cpp#L63-L111).