Capstone: micro-harness

Starter code for the Module 10 capstone. Each file is a runnable skeleton; the TODOs mark where you extend it as you progress through the milestones.

Layout

agent.py       # the loop (Milestone 1)
tools.py       # tool registry and built-in tools (Milestone 2)
context.py     # compaction and caching (Milestone 3)
subagents.py   # router and specialists (Milestone 4)
permissions.py # permission engine (Milestone 5)
prompts/       # layered system prompts (Milestone 6)
evals/         # eval runner and cases (Milestone 7)
trace.py       # logging and replay (Milestone 8)
main.py        # CLI entry point

Running

pip install anthropic jsonschema
export ANTHROPIC_API_KEY=...
python main.py "your task here"

What is here

Everything you need to implement Milestone 1 exists as stubs with clear TODOs. Later milestones will require you to add files or modules; the starter is intentionally small.

What is not here

A specific model provider. The code uses the Anthropic SDK as a reference; swap for any provider.
A web UI. The harness is CLI-first.
Tests or evals. You write those in Milestone 7.

Hints

Keep agent.py small. If it balloons past 200 lines, you are probably missing an abstraction.
Write tests (or at least a main-block smoke test) as you go, not at the end.
When you get stuck, go back to the module that covers the subsystem you are working on. The answers are there.