Upstream: requirements, design, and architecture phase (including ADRs) where early review prevents costly rework.
Midstream: implementation and pull request phase focused on code quality, security, and developer experience.
Downstream: test/QA/release-prep phase. Verify coverage and resilience, and prevent regressions. Ensure AI review aligns with release readiness.
Skill: a YAML frontmatter + Markdown unit of review guidance executed by River Review.
Stream Router: logic that selects and runs skills based on the requested phase and change context.
Riverbed Memory (Future): persistent context layer for previous findings, ADR references, and WontFix decisions to keep reviews consistent over time.
Context Engineering: the discipline of systematically designing, selecting, and controlling the context passed to an LLM. The focus is on what to load, when, and at what fidelity—not on prompt wording.
Context Budget: the upper bound on tokens available for a single review run. Skills, diffs, and memory are selected to maximize review quality within this budget.
Attention Budget: a context allocation strategy that concentrates LLM attention on high-signal information by excluding low-signal noise so that important findings are not buried.
Progressive Disclosure: a staged loading strategy that provides context at increasing levels of detail only when needed. Metadata is loaded first; full skill instructions are loaded only for selected skills.
Review Artifact: a structured output record of a review run. It includes the plan (selected/skipped skills), findings, decisions, and debug information, and serves as input for auditing, memory ingestion, and evaluation.