agent-testing

Deterministic fault injection and autonomous resilience tests for DeepSeek Harness tools

Nine deterministic hooks that block a coding agent from weakening tests, swallowing errors, or stubbing type checks — each backed by a planted-failure test proving the guard can actually fire. Runs on Claude Code and DeepSeek Harness.

Record normalized DeepSeek Harness tool results and replay them offline with strict matching and zero tool-body execution.