README.md

6.9 Build a Minimal Coding Harness in 100 Lines and Equip Its First Skill

This experiment implements a complete, minimal coding harness in under 170 lines of pure Python standard library code (zero third-party dependencies):

1. Three orthogonal atomic tools: read_file, write_file, and bash (with safe_path workspace boundary checks and a 15-second command timeout).
2. Progressive disclosure for Skills: On startup, the harness scans the YAML frontmatter of workspace/.agents/skills/*/SKILL.md and injects only a one-line summary (< 50 tokens) into the system prompt. When the model sees a task related to a Skill, it calls read_file on demand to read the full SKILL.md checklist.
3. Agent main loop: A simple loop that appends model tool calls and environment outputs to messages until all unit tests and ledger reconciliation checks pass.

Requirements

Only the Python 3 standard library and an AI Builder Space API key. It runs on any machine.

How to Run

cd 6_9_mini_harness
export AI_BUILDER_TOKEN=sk_your_key
python3 mini_harness.py

You can also pass a custom instruction:

python3 mini_harness.py --task "Add a max_discount=50.0 cap parameter to settle_order in settlement.py, and add a unit test that passes"

What You Should See

By default, the task asks the agent to fix the order settlement function in workspace/settlement.py. When you run it, you will see a complete 6-turn closed-loop trajectory (a reference log is saved in sample_output/deepseek_run.txt):

• Turn 1: After seeing the Skill summary in the system prompt, the model proactively calls read_file to read .agents/skills/settlement-reconcile/SKILL.md.
• Turns 2-3: It reads settlement.py, test_settlement.py, and verify_ledger.py, then runs bash to reproduce the two bugs (premature per-item rounding and pre-discount coupon threshold checking).
• Turns 4-5: It calls write_file to fix the code, then runs both unittest and verify_ledger.py in bash as required by the SKILL.md checklist.
• Turn 6: Once both commands return exit_code=0, it outputs the final summary and exits the loop.