River Review: AI code review via Review Judgment as Code

River Review's starting point was an observation about long agent sessions: aggregating 111 Codex sessions and 8,096 requests, the cited TOKIUM article showed that even with a high prompt-cache hit rate, a larger context means more is resent per request. One suggested fix was isolating large outputs, such as tests and log analysis, in sub-agents, which act as a boundary rather than only a way to parallelize.

That led the project away from simply shrinking a context budget. River Review asks instead what to read, which responsibility reads it, what to keep as evidence, and what to hand to the next decision. Its context presets cap review modes at 1,024, 4,000 and 16,000 maxTokens, and Skill loading uses progressive disclosure, so the budget is treated as an attention budget as much as a cost budget. The current implementation actually computes only path proximity between changed files, since symbolUsage, siblingTest and commitRecency are configurable but not calculated in the ranking.

The design also draws a line the project does not cross: context lifecycle stays outside the core. When to end a session, compress context or retry is left to the agent host, and River Review provides the judgment material before and after that boundary. A planned Organizer would project existing findings, coverage and verification results into a human-facing view deterministically, without re-judging severity or deciding GO or NO-GO, but it is still at the design stage.