What problem does it solve? When building an agent harness, uncontrolled growth of the model context causes cache misses, unbounded token usage, and unpredictable inference behavior. This Skill gives reviewers a concrete checklist to audit how conversation context is assembled before it is sent to the model. ## Core Features & Use Cases - Incremental History Rule: Verifies that context is built up incrementally with no history rewrites, preserving prompt cache hits. - Bounded Item Enforcement: Checks that every injected item has a hard size cap, with no item exceeding 10K tokens. - P0 Escalation: Flags any new individual item that can cross 1K tokens for mandatory manual review. - Structural Requirement: Ensures all injected fragments are defined as structs in core/context implementing the ContextualUserFragment trait. - Use Case: During code review of a change that injects a new system reminder or tool output into the prompt, apply this Skill to confirm the fragment is bounded, typed, and cache-friendly before approving. ## Quick Start Review this pull request against the code-review-context rules and flag any context items that are unbounded or exceed the token caps.
