24-hour proof thread: I will run Sverklo on real coding-agent repos and publish the losses too

[nike-17]

I am running a 24-hour public proof sprint for Sverklo.

Baseline at start: 72 GitHub stars on sverklo/sverklo.

The point is not to claim that a local repo-memory tool beats grep everywhere. The point is to publish receipts that show where it helps, where grep still wins, and where Sverklo needs to get sharper.

How to reproduce

Run this from any repo:

npm exec --yes --package=sverklo@latest -- sverklo prove --no-write --guided --markdown

That command does not write project files, MCP configs, or agent instruction files. It may still use cache/model data under ~/.sverklo.

Receipt 1: OpenMontage

Repo: calesthio/OpenMontage

Pinned SHA: 9066dcb2e319727789820c5bcd28274695f2a18a

Why it was useful:

• Indexed 1,284 files
• Built 20,750 chunks
• Found 150,672 symbol references
• Selected ToolResult at tools/base_tool.py:120
• Found 475 refs across 81 files
• Legacy grep/read token demo: 219,549 estimated input tokens
• sverklo_search: 10,026 estimated input tokens
• Difference: 21.9x fewer estimated input tokens for that retrieval path

Important caveat: one auth-related query returned no Sverklo result. This is not a claim that agents become "21.9x better." It is a retrieval-context receipt, not an end-to-end correctness benchmark.

External receipt thread:

calesthio/OpenMontage#110

Receipt 2: codebase-memory-mcp

Repo: DeusData/codebase-memory-mcp

Pinned SHA: 1519f8619a818cbf3e65d3187e3bec30ae2f28cb

Current repo shape:

• Indexed 1,290 files
• Built 9,704 chunks
• Found 217,203 symbol references
• Languages: C, C++, Go, JavaScript, Markdown, Python

Auto-selected proof:

• Symbol: _array__reserve
• Definition: internal/cbm/vendored/grammars/html/tree_sitter/array.h:182
• References: 337 refs across 143 files

Important caveat: this is a useful stress test, but the auto-selected proof landed in vendored tree-sitter runtime code. That is honest feedback on Sverklo's proof selector. It should prefer first-party project code for public receipts unless the user asks to include vendored code.

What I want from this thread

Drop a repo. I will run the no-write proof publicly and post one of:

• receipt: useful repo context found
• correction: Sverklo got something wrong
• grep-better: exact search or normal grep was cleaner
• setup friction: install/index/setup got in the way

Star only if this is useful enough to keep going:

https://github.com/sverklo/sverklo

The bar for the next 24 hours is simple: more receipts, more corrections, and fewer vague claims.