Measure CodeGuard context cost across delivery paths

[thomas-bartlett]

Summary

Define a simple way to compare how much context CodeGuard uses across its main delivery paths.

Scope

Compare the current options:

• Direct rule files
• Agent Skills
• MCP rule access
• codeguard-reviewer / subagent review
• Flattened prompt or bundled guidance

Questions

• What should we measure: tokens, rules loaded, latency, setup complexity, or output quality?
• When is direct rule loading good enough?
• When do skills, MCP, subagents, or progressive loading provide enough benefit to justify added complexity?

Acceptance Criteria

• Define a small set of context-cost measurements.
• Add a simple comparison template or checklist.
• Recommend the first context-reduction approach worth testing.
• Clarify how this relates to Evaluate CodeGuard with local models and constrained context windows #73 and Add evaluation harness for Project CodeGuard #86.

Related: discussion #41, #73, #86.