Radar, deterministic code search/navigation

Radar, a deterministic code search and navigation tool, builds compact maps and exact source routes that enable coding agents to read fewer files and spend fewer tokens, achieving a 37.1x faster clean build (0.056 s vs. 2.069 s) and a 79.62% faster tie-break-heavy route lookup compared to code-review-graph. In a 7-task, 3-run-per-condition live cost test, Radar reduced agent costs from $0.155542 to $0.056409 per session, a 72.4% drop in uncached input, and projects weekly savings of $2.23-$22.32 for 100-1,000 daily searches. The tool also cut prompt tokens significantly, with the generated skill falling from 368 to 156 tokens and the navigation contract from 835 to 462, while using 93.590x less token volume than Semble in a frozen 40-case ledger.

same host / 300 Python files ↗ /evidence.html local-comparison 37.1x faster clean build 0.056 s radar map versus 2.069 s code-review-graph build. Why: compact maps, not a full review graph. radar builds committed maps and exact source routes. Coding agents read fewer files, spend fewer tokens, and finish navigation sooner. $ radar query "token verification" FINAL SOURCE ANCHOR src/auth.rs verify token measured advantage Radar builds the smallest useful index, returns a source pointer instead of a repository tour, and keeps models out of the lookup path. same host / 300 Python files ↗ /evidence.html local-comparison 0.056 s radar map versus 2.069 s code-review-graph build. Why: compact maps, not a full review graph. 2.825 ms radar route versus 321.052 ms code-review-graph search. Latest query internals: scorer-heavy fresh-query p50 is now 14.65%-25.90% faster; tie-break-heavy route lookup is 79.62% faster. frozen 40-case ledger / native contracts ↗ /evidence.html retrieval Radar used 919 actual cl100k base tokens versus TokenSave 13,947 ; Semble used 86,009 , or 93.590x Radar's volume. Why: exact anchors are smaller than bundled context or semantic snippets. 7 tasks × 3 runs per condition / median sums ↗ /evidence.html live-cost $0.056409 with radar versus $0.155542 with direct source search; uncached input fell 72.4%. Why: the agent invokes one exact query instead of exploring files. agent prompt surfaces / current batch ↗ /evidence.html prompt-surfaces The generated skill fell 368 → 156 tokens, the navigation contract fell 835 → 462 , route-add wrapper keys fell 577 → 152 , and a 100-1,000-search daily session projects 159,400-1,594,000 input tokens saved. Why: Radar budgets the instructions and fallback files agents actually read. 100-1,000 code searches / one session per day ↗ /evidence.html cost-projection Projection: $0.4402-$4.4020 → $0.1214-$1.2140 per 100-1,000-search daily chat session, or $2.23-$22.32/week . Why: fewer exploratory source reads become fewer billable input tokens. Scope matters. The external-tool rows compare different native output contracts on pinned versions; the agent-cost row is a published-rate reference scenario, not a Spark invoice. Broader graph and editing tools can be the better choice when an anchor is not enough. Exact routes replace repository tours. Less context goes in, and the source stays in charge. Repository structure stays beside the code. Exact definition packets reject stale answers and whole-file reads. Callers, tests, dependencies, and co-change history. Abstention looked safe on 34 calibration answers, then returned 4 wrong anchors among 9 holdout directs. Radar published it and rolled it back. Skills, MAPs, route caches, symbol indexes, MCP schemas, and overview rows are measured as agent input, not treated as free text. Different tools, different jobs. This is where radar is the sharper fit. | project | shape | signal | radar boundary | radar is better when | |---|---|---|---|---| | Rival numbers stay labeled. See the full comparison ↗ /evidence.html alternatives .