Blog
Engineering notes on code agents, context, and the systems around them.
Latest
Codebase analytics through MCP: what query_artifact_metadata can answer
How an agent selects code by cognitive complexity and call relationships, summarizes modules, and finds type declarations to prepare QA test cases.
Read articleMore articles
10CodeAlive vs open-source code graph indexers, checked against the compiler
On Playwright, Zod and FluentValidation, 98.5–98.9% of CodeAlive's call edges matched the compiler and it found 82–91% of all calls. The strongest open-source tool we could score reached 41–65% and 40–51%.
CodeAlive harness vs OpenCode: 18.7 points on the same model
Same model, same 40 RepoContextBench v2 questions. The CodeAlive research harness scored 65.3 and OpenCode 46.6. On the 1.5M-line repository CodeAlive also used 3.6 times fewer tokens and spent a third of the time per task.
GraphRAG for code: when the call graph actually helps an agent
On a 0.5M-line repository the call graph barely moved RepoContextBench scores; on a 1.5M-line one it helped the agent find noticeably more of the right code. The harder problem is getting the agent to use the graph at all.
Improving instruction following with next-step hints in tool call results
Ending a tool result with a local instruction helps smaller models interpret the data and choose the next useful action.
How we automated MCP OAuth 2.1 browser testing
Testing the MCP browser authorization flow with a real loopback callback, PKCE, and Playwright.
Giving Kimi K3 a faster path through code
Preliminary RepoContextBench results show CodeAlive cutting Kimi Code task time by 31% and answerer token use by 28% with Kimi K3.
CodeAlive 3.0: From Context Engine to Code Research Agent
A look at CodeAlive 3.0: our production context engine, code research agent, RepoContextBench, and the new Tool API and MCP surface.
How CodeAlive MCP evolved: from three tools to a research interface
How CodeAlive MCP changed its tools, retrieval flow, response formats, and error handling on the way to v3.
RepoContextBench: How Qwen 3.6-35B beats Sonnet 4.6
Qwen 3.6-35B A3B scored 77.6 with the CodeAlive agent harness, ahead of Opus 4.8 at 77.45 and Sonnet 4.6 at 71.95.
Canonicalization for small agent LLMs: a practical AX pattern
A scalar string broke a string-array tool call. We fixed it once at the Microsoft Agent Framework invocation boundary.