A minimal, typed scaffold for agents: deterministic prompts, strict JSON, and a session that records plans, tool calls, and staged edits—so you can replay runs and ship audit logs.
* Deterministic Markdown prompt trees (dataclass input/output, tool contracts)
* On‑disk overrides with hash checks; Git‑root discovery
* Event bus (ToolInvoked/PromptExecuted), reducers, rollbackable session state
* Built‑ins: planning, sandboxed VFS, Python‑eval (via asteval)
* Optional adapters: OpenAI/LiteLLM conversation loop + JSON‑Schema outputs
Install: uv add weakincentives (extras as needed)
Start with the code‑review example; it ties together prompt construction, overrides, telemetry and using the adapters.
Status: Alpha; APIs may change
Roadmap: parallel sub-agents, built-in GEPA prompt optimizer that uses the overrides store