osiris · experiments
experiments, evals, and half-finished ideas. everything here is real code that actually runs. some of it works. most of it almost does.
agent loops, memory, evals, gateways. the main lane.
minimal coding agent for any openai-compatible api. stdlib only, zero dependencies. the one that gets improved the most.
full cognitive runtime in stdlib-only python. layered memory with decay, rehearsal, forgetting curves.
multi-agent orchestration for any openai-compatible api. debate, vote, supervise, pipeline, fanout.
openai-compatible llm gateway. one local endpoint fanning out to many model backends.
model eval harness. json in, pass/fail out. for checking if a model actually does what it claims.
run one prompt against several system prompts or models. a/b testing for people who write system prompts.
scriptable fake chat completions server so projects stop hand-rolling their own test doubles.
process memory toolkit for debugging and reverse engineering. attach, scan, pattern match.
small clis that do one thing and do it fast.
bundle a repo into one file for pasting into coding agents. respects gitignore, token estimate included.
load a .env file and run a command with those vars set. the glue between everything.
rerun a shell command whenever files change. zero deps, debounced, ignores the junk by default.
semantic search over any project folder. stdlib only, any embeddings api on the back.
http/1.1 server written from scratch. sockets, routing, keep-alive, chunked encoding. because why not.
built to learn something, kept because it worked.
full chip-8 plus super-chip 1.1. all 35 opcodes, hires mode, disassembler, cli debugger.
record/replay proxy for openai-compatible apis. agent tests that run offline and deterministic.
the next experiment hasn't been built yet. check back, there's usually a new repo every day or two.
everything here started as a real problem. mule exists because typing the same agent loop from scratch got old. ctxpack exists because pasting repos into cursor by hand is miserable. the lab is just where they live before they earn a real docs page.
want the flagship? that's mule.osiris.rocks.