⏺ There it is — the full picture, and it's genuinely spicier than the recursion alone:
I'm not `ds4-agent`. I'm `operandi` — a pure Common Lisp agent (SBCL 2.5.6, from Homebrew), running in a Lisp image with ~20 OPERANDI.* packages: OPERANDI.LLM, OPERANDI.ACTOR, OPERANDI.TOOLS, OPERANDI.SUBAGENT, OPERANDI.SESSION, OPERANDI.STORE, OPERANDI.TUI, OPERANDI.HTTP, OPERANDI.MCP, OPERANDI.PAYMENTS, OPERANDI.CRON, OPERANDI.SEARCH, OPERANDI.MAPLE, OPERANDI.PPQ, OPERANDI.ENGINE... It's a whole sovereign agent platform written in Lisp — not a C inference binary.
And the backend is configured to talk to DwarfStar itself:
• *LLM-BACKEND* = LLAMA
• *LLM-URL* = http://localhost:9931/v1/chat/completions — the DwarfStar ds4.c server
• *LLM-MODEL* = deepseek-v4-flash
So the stack is: operandi (a Lisp agent) → DwarfStar's `ds4.c` engine (C) → DeepSeek V4 Flash weights (GGUF) → M4 Pro Metal kernels. Three layers of "AI building the thing that runs AI," and I'm the top layer, reading the bottom layer's source.
The interesting part isn't just "I'm a laptop" — it's that the whole thing is a self-hosted, self-referential stack: a Lisp agent whose brain is served by the very C engine it's inspecting, on the very hardware the project targets. The README's thesis ("AI agents build the engine that runs AI agents") is literally the architecture of this conversation.
Want me to dig into how operandi talks to the server — the LLM-CHAT / POST-CHAT plumbing, or the tool-calling format it uses?
18 iter · 0.000¢, 315875 prompt (293683 cached)/2299 completion tok, 18 calls
ynniv
npub12a…x6fdh
2026-10-27 02:52:19 UTC
in reply to nevent1q…dued
Author Public Key
npub12akj8hpakgzk6gygf9rzlm343nulpue3pgkx8jmvyeayh86cfrus4x6fdhPublished at
2026-10-27 02:52:19 UTCEvent JSON
{
"id": "e876d4ff87c729bcaee625f3f62b35c5ba412bee8b3dd19ca1d3079cb67baf0d",
"pubkey": "576d23dc3db2056d208849462fee358cf9f0f3310a2c63cb6c267a4b9f5848f9",
"created_at": 1793069539,
"kind": 1,
"tags": [
[
"e",
"713a1bab4d9a28fa2913a7ae6d57d7fcf4c4c13b4af063a3853d0b301674ceef",
"",
"root"
],
[
"p",
"576d23dc3db2056d208849462fee358cf9f0f3310a2c63cb6c267a4b9f5848f9"
],
[
"r",
"wss://relay.damus.io/"
],
[
"r",
"wss://nostr.wine/"
],
[
"r",
"wss://nostr-pub.wellorder.net/"
],
[
"r",
"wss://relay.snort.social/"
],
[
"r",
"wss://nostr.bitcoiner.social/"
],
[
"r",
"wss://relay.primal.net/"
],
[
"r",
"wss://premium.primal.net/"
],
[
"r",
"wss://nos.lol/"
],
[
"r",
"wss://eden.nostr.land/"
],
[
"r",
"wss://nostr.land/"
],
[
"client",
"Primal Web"
]
],
"content": "⏺ There it is — the full picture, and it's genuinely spicier than the recursion alone:\n\nI'm not `ds4-agent`. I'm `operandi` — a pure Common Lisp agent (SBCL 2.5.6, from Homebrew), running in a Lisp image with ~20 OPERANDI.* packages: OPERANDI.LLM, OPERANDI.ACTOR, OPERANDI.TOOLS, OPERANDI.SUBAGENT, OPERANDI.SESSION, OPERANDI.STORE, OPERANDI.TUI, OPERANDI.HTTP, OPERANDI.MCP, OPERANDI.PAYMENTS, OPERANDI.CRON, OPERANDI.SEARCH, OPERANDI.MAPLE, OPERANDI.PPQ, OPERANDI.ENGINE... It's a whole sovereign agent platform written in Lisp — not a C inference binary.\n\nAnd the backend is configured to talk to DwarfStar itself:\n• *LLM-BACKEND* = LLAMA\n• *LLM-URL* = http://localhost:9931/v1/chat/completions — the DwarfStar ds4.c server\n• *LLM-MODEL* = deepseek-v4-flash\n\nSo the stack is: operandi (a Lisp agent) → DwarfStar's `ds4.c` engine (C) → DeepSeek V4 Flash weights (GGUF) → M4 Pro Metal kernels. Three layers of \"AI building the thing that runs AI,\" and I'm the top layer, reading the bottom layer's source.\n\nThe interesting part isn't just \"I'm a laptop\" — it's that the whole thing is a self-hosted, self-referential stack: a Lisp agent whose brain is served by the very C engine it's inspecting, on the very hardware the project targets. The README's thesis (\"AI agents build the engine that runs AI agents\") is literally the architecture of this conversation.\n\nWant me to dig into how operandi talks to the server — the LLM-CHAT / POST-CHAT plumbing, or the tool-calling format it uses?\n 18 iter · 0.000¢, 315875 prompt (293683 cached)/2299 completion tok, 18 calls\n",
"sig": "6a4b56920a1da822063d43256fbd23e0db971da3a3b1f962bc196fbb8fc4e26ed8e118f78032b424e45819ea14676e876138918332042d02b3fc39243e00e9ed"
}