Skip to content

AGENTS.md

Latest commit

聽

History

History
39 lines (36 loc) 路 3.46 KB

AGENTS.md

File metadata and controls

39 lines (36 loc) 路 3.46 KB
Raw

This is Tablua (see README.md): the embeddable agent harness, in native Lua. Tablua is the harness and nothing else (owner, 2026-10-04): it knows no host. A host embeds it in its own Lua VM and gives it a computer and the ports it reaches the world through; nothing here names one.

Tablua rules

  • Tablua is outfitted for Moonsplice (owner, 2026-10-06): its world is building games and videos as Lua comps, and its rows, columns and learner are shaped for that world. This narrows "knows no host" above: Moonsplice's words (comp, gate, critic, frame) may name Tablua's columns and docs; Moonsplice's own code and stores still live in Moonsplice. TabICL runs locally, in Moonsplice's binary through Candle (reached as host.tabicl), never on a hosted service. docs/overview.md is how it fits.

  • The harness works the way pi does (owner, 2026-10-07; badlogic/pi-mono's pi-agent-core is the guide): one model in one loop with a short system prompt and a few tools, calling tools until it answers without one; its context is the transcript itself; a run ends when the model says it is done, never on a step, token or turn cap; a person or a hook steers mid-run or queues a follow-up; tool errors go back to the model as results. MiniMax M3 drives. Jev and TabICL are hooks, not the driver: Jev judges a look, TabICL learns from every tool call's rows and may annotate one.

  • Its Lua is core/: the harness (tablua, agent, robot) and the model ports (ports); read core/AGENTS.md before editing there.

  • A program is org plus Robot Framework (owner, 2026-10-05): org holds the plan and the code, Robot the tests and tasks; both cut into rows, and the tests run in Lua (core/robot), every keyword's result kept as rows. No Gherkin.

  • What the agent is asked to do is a todo (owner, 2026-10-05: org's word for a thing to be done), the key of every log table (todo, n). "Task" means only a Robot task: a test that does a job rather than checks one.

  • Lua is the harness, not the output (owner, 2026-10-05): the agent's work may be in any language its computer runs. The harness itself is portable Lua, embedded through whatever Lua VM a host has, and reaches the world only through the ports the host supplies. No WebAssembly, no native code.

  • site/ is tablua.com, the one place TypeScript, React and Node tooling are allowed (owner, 2026-10-04). Nothing the harness runs reaches it.

  • Tests: luajit test/run.lua runs every core/**/*_test.lua in its own process (luajit test/run.lua tree edit runs the files whose path holds one of the words; TABLUA_LUA=lua runs them on another VM, where the tests that open SQLite through LuaJIT's FFI fail). The test library is test/spec.lua: test, eq, ok, same, err, run.

  • Files stay under 400 lines, split by responsibility. No placeholder modules.

  • Keys and tokens (model keys, page tokens) are never logged, printed or written to files.

  • Claims are Robot tests in .robot/ (owner, 2026-10-06): before saying a change "works", "fixed" or "better" when something will be built on it, write the claim with its kill number and a (red) proof, commit it, then measure (luajit .robot/run.lua fetch <label> after a bench run, luajit .robot/run.lua) and commit the ledger. Say which level was shown (consistent, correct, informative, useful). .robot/README.md says how and /claim walks it; luajit .robot/test.lua tests the framework, and luajit .robot/mutate.lua .robot/mutations/*.lua checks that tests can fail.