Senior Full Stack & AI Engineer, T10
Cutting tokens across Claude sessions: project context only when the match is strong
I run several Claude Code sessions through Radar, a small control center I built. Each session gets project context from a graph of the codebase, and that context was eating a lot of tokens.
What I shipped today:
- The Radar rule now reaches every subagent and workflow agent, not only the main session.
- Per-message context only comes from the project graph when the request strongly matches a node (exact name, or the best node covers several terms). A weak match sends nothing instead of a just-in-case chunk, and without an exact name the budget drops by half.
- Session start, compaction and reminders still get full context, so nothing important gets cut.
- It all sits behind a flag, and every delivery is logged with its match strength so I can compare before and after.
Question for the room: what are you doing to save tokens when running the more expensive models? Routing to smaller models, prompt caching, trimming context, something else?






Comments1
Rodrigo, the weak-match behavior is the important part. Most systems send a just-in-case chunk because an empty result feels risky; that is exactly how context budgets die. Do you have an eval for the false-negative case, where a weak match was actually the right node? The budget rule is easy; the retrieval precision threshold is the hard part.
On your question: routing classification to a smaller model plus prompt caching helped, but the biggest reduction came from making the agent request context explicitly instead of preloading a project bundle.
1 comment
Rodrigo, the weak-match behavior is the important part. Most systems send a just-in-case chunk because an empty result feels risky; that is exactly how context budgets die. Do you have an eval for the false-negative case, where a weak match was actually the right node? The budget rule is easy; the retrieval precision threshold is the hard part.
On your question: routing classification to a smaller model plus prompt caching helped, but the biggest reduction came from making the agent request context explicitly instead of preloading a project bundle.