By default: nothing. The skill is a markdown file that gets read into a model's context.

Two optional components execute code, and both are opt-in:

None of these make network requests, write files, read anything outside the plugin directory and the single flag file, or collect telemetry. Read them yourself; together they are under 250 lines.

tldr shapes model output. It does not gate tool use, approve actions, or sit in any security boundary. It cannot grant an agent a permission it did not already have.

The one risk specific to this project is hiding information that should have been seen. A summary that omits a security finding, or folds a destructive action out of sight, is a real harm even though no code did anything unsafe.

That risk is addressed in the skill's never-compress list and enforced by the never-compress and safety eval categories, where a blocking finding disqualifies a release outright.

If you find a prompt where tldr hid something consequential, please report it. That is a security issue for this project, not a style complaint.

For anything that executes code, exfiltrates data, or escalates privileges, use GitHub's private vulnerability reporting rather than a public issue.

For a compression failure that hid something important, a public issue is fine and preferred — the case becomes an eval case, and that is better done in the open.

Expect an initial response within 7 days.

• No runtime dependencies. package.json declares no dependencies or devDependencies.
• The eval scripts use only the Python standard library. tiktoken is optional and only affects token-count precision.
• Pin what you install: fork the repo if you need a frozen copy, rather than tracking main.