Does the Sverklo proof command write project files?

No. With --no-write, Sverklo does not write project files, MCP config, or agent instruction files. First run may cache a local model and index under ~/.sverklo.

What should I share after a proof run?

Share one receipt, one correction, one case where grep was better, or one setup-friction report. Negative receipts are useful because they expose product gaps.

No-write proof receipt

Run the proof before setup.

Sverklo can print repo-memory evidence before it writes project files, MCP config, or agent instructions. Use this to decide whether Sverklo understands your repo well enough to install.

Command

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

Requires Node 24 or newer. First run may cache the local ONNX model and index under ~/.sverklo. The command should not write project files or MCP config.

What to check

Central files: do the selected files look like real hubs, not random examples?
Selected symbol: is the symbol meaningful for the repo?
Callers: does the receipt name real usage sites?
Prompt: would the paste-ready prompt help an agent before editing?

Classify the result

Receipt: useful evidence from a real repo.
Correction: the graph picked the wrong center or missed important files.
Grep-better: an exact string search was simpler and more reliable.
Setup friction: Node, npm, model cache, or command output blocked the run.

Why this matters

The current adoption bottleneck is external proof. Maintainer-seeded receipts are useful for showing the shape, but third-party receipts and corrections are what make the product better and make the claim credible.

Share the useful or broken part

Post the receipt, correction, or setup friction in the public proof thread. If the repo is private, redact names and keep the failure mode.

Next: read what repo memory means, then use install or the Codex CLI recipe if the proof looks useful.