What problem does it solve? AI agents working on the CG-agent-harness Rust codebase tend to confabulate API flags, drift from constraints in long sessions, weaken security posture, and produce fluent but wrong answers. This Skill encodes explicit epistemic, reasoning, and security disciplines so any model executes them consistently on substantive engineering and security work in this repo. ## Core Features & Use Cases - Reasoning and calibration protocol: Decompose questions, test load-bearing premises, mark speculation explicitly, and treat "I don't know" as a valid output. - Harness security lens: Trust-boundary rules for the I6 process isolation, guard chain, write gates, clone jail, and approval binding, with findings-before-writes on core security files. - Cold repository facts: Architecture map, invariants outline, known traps (quoted YAML flags, intentional cross-boundary duplication, CSRF placeholder contract), quality bar commands, and settled decisions that must not be re-litigated. - Use Case: An agent asked to modify src/shim or src/server/guards.rs first reads INVARIANTS.md, reports a findings summary, and includes an explicit invariant statement in the draft PR instead of editing blindly. ## Quick Start Ask the agent to apply the fable-protocol discipline while reviewing or modifying a core security file in the CG-agent-harness repository.
