Skip to content

fix: the skill says personal data for what is sent to another service - #107

Merged
alkhatibdev merged 1 commit into
mainfrom
skill-personal-data
Nov 18, 2026
Merged

alkhatibdev merged 1 commit into
mainfrom
skill-personal-data

Conversation

@alkhatibdev

Copy link
Copy Markdown
Owner

The skill says personal data for what is sent to another service

  • The Claude Code skill's description said to use DPOLens "when sending data to another
    service". It now says "when sending personal data to another service", as the server's
    instructions and the tool descriptions do. In the published trigger test, the plugin searched
    on a page view sent to Segment, which carries data about no one, in all three runs, where the
    MCP server on its own never did.
  • A test reads every text that says when to search (the skill's description, the instructions'
    trigger sentence, and search_policies on a self-hosted instance and on a community version)
    and fails on a bare "data". The lines are joined first, so "personal" at the end of one line
    and "data" at the start of the next stay one phrase. On main it fails on the skill's
    description, and only there.
  • evals/trigger-test.md replaces "the likely reason" with what was found and measured. The
    page-view task had two ways in: the word, and in one run the plugin's /policy-check command,
    which Claude Code ran on its own. With this change alone, the plugin arm was run again: 36 of
    36 (all before the first edit), 1 of 45 and 9 of 9. The page-view task never searched and
    /policy-check never ran. The one unneeded search was session-length, in one run of three.

Checked

  • The rerun: the 30 held-out tasks, three runs each, Claude Code 2.1.288 with
    claude-sonnet-5-5, from 1acbf05 with only the skill's change, on 17 and 18 November. It is
    not written to evals/results/trigger.json: it ran before fix: the trigger test's instance chooses the law its searches cover #104, so its instance had no law
    chosen and every search came back as an error, which --write refuses. The published result
    stays October's. What it counts is the call, which comes before any answer, so the counts hold.
  • ruff, ruff format --check, mypy, the import contracts, scripts/client_text.py and the
    whole test suite with the integration tests (1501 passed) in a clean copy, with caches off and
    GITHUB_ACTIONS=true. The new test fails on main and passes here.

@alkhatibdev
alkhatibdev merged commit bf5b595 into main Nov 18, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant