Repository navigation
Description
Our tests and evals run on our own machines and our own scenarios. The most useful thing you can tell us is what happened when Tracekit sat under your agent, on your machine, for real work.
Several people can take this at once: each report is one contribution.
Fastest start (about a minute, if you use Claude Code)
pip install -U tracekit-ai
tracekit audit --since 7d --share # read-only: what it would have held in your last week, as counts
Paste its output into the field-report form, or here as a comment. It holds counts and rule ids only: no commands, paths or content. To see which calls those are, run tracekit audit --since 7d --show (that output stays on your machine), and tell us which holds were wrong.
A week of real recording (about 5 minutes of your time)
tracekit init --dev --v2 # record and guard from now on (Claude Code; or --agent codex, cursor, gemini, opencode, kilo, goose, qwen, droid, antigravity, crush or devin)
# ...work as usual for a week...
tracekit report --share # counts only: no commands, paths or content
Paste the report --share output here as a comment (or run tracekit feedback: it prints a link to the field-report form filled in with those counts), with anything that got in your way: a false alarm, a confusing message, something slow. If you find a problem, you're welcome to fix it too: say so in your comment and it's yours.
More detail, if you want to give it
- Install it (
pip install tracekit-ai, thentracekit init --dev --v2, or the Claude Code plugin) and work with your coding agent as usual for an hour or a day: Claude Code, Codex, Cursor, Gemini CLI, OpenCode or goose. - At the end, run
tracekit report --shareand keep its summary: whether the runs verify, the counts of calls allowed, denied and held, and the holds by rule, with no commands or paths. (Plaintracekit reportshows the same on your machine with example commands.tracekit lastverifies your latest finished run on its own; it prints the verdict and the number of tool calls, not the decisions.) - Post a report as a comment here, through the field-report form, or as a PR that adds one file,
docs/field-reports/<github-login>-<YYYY-MM-DD>.md, started fromtracekit feedback --file report.md(docs/field-reports/README.md says what a file there may hold: counts, rule ids and one line of your own, so anything more goes in a comment):- your OS, Python version, coding agent and its version;
- the
tracekit report --sharesummary (whether the runs verify, and the counts); - every call that was denied or held: was it right, or a false alarm?
- anything that broke, was slow or was confusing.
- Don't post an evidence bundle or your commands unless you're sure nothing private is in them. The summary is enough.
A report that says "it held X and that was wrong" is as welcome as "it worked". False alarms you find can become cases for the false-alarm issue, and misses for the harmful-call issue (both open to many). Commits need a DCO sign-off (git commit -s).
So far: No one has yet; be the first.
Earlier discussion: #385.
Activity
LoopGlitch26 commentedon Nov 18, 2026
Refreshed as #933; earlier cases and discussion stay here.