Benchmarks, retrieval architecture and the engineering behind indexing a codebase once and serving it to every tool that asks.
11–20 of 89 · page 2 of 9
A pull request changes 11 lines, the tests pass, every automated reviewer approves, and three weeks later a reporting service is quietly returning wrong revenue. Nothing about the diff was incorrect. What was wrong lived in another repository nobody in the pipeline had read. This is where diff scoped review has nothing to say, and it is now the majority of what breaks in a large organisation.
Spec Kit spent a year and gathered well over 100,000 stars teaching several million developers that the specification, not the code, should be the source of truth. That is category creation at a scale no startup could have afforded, and it lands directly on our thesis. It also assumes a blank page, which is true exactly once in the life of a system.
The deepest MCP integration anybody has shipped, with 16 tools, 7 resources, skills and Claude Code hooks, running with zero server on a local graph. The developer experience is the best in the category and that is why it grew. Two things are worth thinking through before you build anything organisational on top, and only one of them is technical.
Local deterministic parsing over 36 tree sitter grammars, typed nodes, provenance tagged edges, Apache 2.0, no vector store and no telemetry. That is a clean design and the privacy position is better than most of the category. The problem starts the moment your organisation has more than one repository, and it is not a problem more parsing solves.
Of every commercial tool in automated code review, Greptile made the closest architectural call to ours: index the whole repository first, then review the change against it. That decision is why it catches 82% of bugs against CodeRabbit's 44%. It also arrives with about 11 false positives per run, and the shape of those false positives tells you exactly what an index of symbols can and cannot know about intent.
Repomix is the leader in context packing, with 26,000 stars and 255,000 npm downloads a month, and its tree sitter compression genuinely cuts token count by about 70%. Apply that to 10 million files and you go from three orders of magnitude short of the context window to slightly less than three orders of magnitude short. Every tool in this category publishes a token reduction number. None of them publishes a correctness number.
Serena is the best open source symbol level tool for agents, with 25,200 stars, 170 contributors and over 40 languages behind a clean LSP to MCP design. A language server gives you exact call edges and zero intent. In the MSR 2026 dataset of rejected agentic pull requests, only 36% failed on the code itself and 31% failed on a decision the project had already taken, which is the third no symbol table can reach.
Sourcegraph built the most complete cross repository code search anyone has shipped, and it did so on an assumption that held for fifteen years: a human is asking. Humans supply the meaning the index leaves out. An agent cannot, and at 10 million files a 1 million token window holds 0.01% of the estate, so the gap is not one model generation away.
Compliance with an agent's own explicit constraints falls from 73% at turn 5 to 33% at turn 16 across 4,416 measured trials, and Claude Code auto compaction reduces 132,000 tokens of session state to roughly 2,300. Shorter sessions and cleaner instruction files treat the symptom. The window fills with low value tokens because the agent rebuilds an understanding of your codebase from raw source on every task.
DeepSeek published its agent runtime on 13 August 2026 and it passed a hundred thousand stars in two days, but almost everything written about it was the install command. This is a plain English read of the source. What an agent runtime actually is, what everything is a plugin means in practice, the three layers inside every capability, the four modes, the security defaults, what the benchmark numbers really measure, and the rough edges nobody is listing.
