Wasted tokens: stop AI coding agents re-reading files — persistent repo context

As teams adopt AI coding agents, those agents frequently re-send large swaths of repository files to LLMs on every invocation, wasting tokens, increasing latency and producing inconsistent answers as unchanged context is reparsed. This problem hits engineering teams building in-IDE assistants, CI/automation bots, and code review tools across the 28M professional developers market where organizations already spend roughly $1,500 per developer annually on tooling and AI assistants. You could build a persistent repo-context layer: a lightweight daemon or managed service plus IDE SDK that stores content-addressed snapshots, maintains embeddings and small local vector indexes, and transmits only diffs or semantic pointers to the model instead of full files. Key features would be deterministic cache invalidation, low-latency retrieval, privacy/compliance hooks, and instrumentation that reports token savings and correctness metrics; delivered as an SDK and optional hosted control plane, pricing could be per-repo or per-seat, with teams plausibly reducing repeated context by 30–70% depending on workflows. This is timely because token bills and latency are becoming visible line items, retrieval-augmented designs are standard, and developers expect editors to host fast AI assistants. To stand out you need rigorous correctness guarantees (avoid stale-code answers), tight editor integrations, measurable ROI dashboards, and clear security controls; competition is medium and fragmented so execution and partnerships matter. The main challenges are cross-platform integration, cache-consistency in high-velocity repos, and proving savings on real workloads, but a solution that demonstrably lowers token spend while preserving correctness addresses a specific, growing pain point teams are already motivated to solve.