context window optimizer

tooling for agent context management — the problem of deciding what goes into a limited context window when agents work with large, complex knowledge bases. most agents either stuff everything in (hitting the limit, degrading quality) or use naive retrieval (top-k semantic search, which misses structural dependencies). neither approach is good. the optimizer sits between the knowledge base and the agent, managing what gets surfaced, in what order, at what point in the task.

the technical research here is genuinely interesting: context window management involves tradeoffs between recency, relevance, diversity, and dependency structure. a code editing agent needs the current file, related interfaces, recent change history, and relevant documentation — and those come from different retrieval strategies. the 40% context utilization target in spec-driven dev kit is a practical heuristic: leave room for the agent's own output and reasoning, otherwise you get truncated context that causes errors mid-task. good tooling would make this principled: profile the task type, estimate context needs, and dynamically manage the budget across a multi-step workflow.

related: axon, AI tooling, hard docs writer