No-write code-intelligence proof run on OpenMontage

[nike-17]

I ran a no-write Sverklo proof pass against OpenMontage because it is trending today and it is an agent-heavy repo, so it is a better sanity check than a toy project.

Pinned repo:

• calesthio/OpenMontage@9066dcb2e319727789820c5bcd28274695f2a18a
• 397 code files in the quick benchmark file set
• sverklo@0.29.2

Proof run:

• Command shape: sverklo prove --no-write --guided --markdown <repo>
• Indexed 1,284 files, 20,750 chunks, and 150,672 symbol references
• Selected ToolResult at tools/base_tool.py:120
• Found 475 references across 81 files
• Generated next agent prompt: Use sverklo impact on ToolResult and tell me what would break if I changed its signature.

Legacy token-efficiency demo:

• grep + reading the top 10 matched files over 5 exploratory questions: 219,549 estimated input tokens
• sverklo_search: 10,026 estimated input tokens
• Aggregate: 21.9x fewer estimated tokens

Important caveat: this is the legacy token-efficiency benchmark, not a correctness-gated benchmark. One query, auth_middleware, returned no Sverklo result, so I would not treat this as a quality win. The honest claim is only that repo-memory search can cut token input sharply on exploratory codebase lookup, while still needing human/project feedback on whether the selected symbol is meaningful.

What I would like from OpenMontage maintainers/users:

1. Is ToolResult actually a useful high-blast-radius symbol to inspect?
2. Did this miss a more important abstraction in the repo?
3. Is this kind of no-write proof useful enough to run on agent-heavy repos, or is it mostly noise?

Sverklo repo: https://github.com/sverklo/sverklo