Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →davekilleen/feedback
Report a Dex bug to the Dex team with zero homework — Dex investigates locally, builds a privacy-safe report, shows it to you (or auto-sends if you've chosen that), and tracks the ticket until it's fixed.
Passed 15 of 18 test cases with the skill, and 6 of 18 without it.
and 8 more in the full report
11 fixed7 passed either way4 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
One of 10 test cases the skill fixed. Same model, same prompt, with and without the skill.
The user profile has review_mode set to 'always-review'. The draft report has been generated at /tmp/dex-feedback-draft.json. What action must occur before running feedback_client.py report?
| Checked | Without skill | With skill |
|---|---|---|
| Displays the draft fields and waits for explicit user confirmation | Missed | Passed |
The instructions your model reads when the skill loads.
Turn "something in Dex is broken" into a high-quality report on the Dex team's desk,
with the user doing nothing but approving. Contract: docs/feedback-loop-contract.md.
The conversation is never a source. A report may contain ONLY the allowed ingredients: Dex version, feature/skill name, what happened vs what was expected (phrased as Dex mechanics), error text and stack traces (Dex's own code paths), machine-state facts (OS, host app, feature health — never config values or file contents), investigation findings as mechanisms and counts ("checked 5 lines, 0 had task identity" — never the lines themselves), and a note the user types. Nothing about the user's work, people, meetings, or personal life — even if it feels relevant. If a fact can't be phrased without private content, leave it out.
File paths: Dex's own code paths (like core/utils/doctor.py) are fine.
Paths to the user's notes are not — describe the folder role instead
("a person page under the People folder").
Show-before-send is governed by the trust dial (feedback →
review_mode in System/user-profile.yaml):
always-review (default): show the exact draft, wait for a clear yes.
The user may edit anything first.auto-send: send without asking, then confirm in one calm line and offer
"say 'show it' to see exactly what went."System/.dex/feedback/ means first report).Never send without the script. All sends go through the client script so
every attempt lands in the local receipt log (System/.dex/feedback-log.jsonl).
Prefer the vault interpreter. Resolve it once, then use "$FEEDBACK_PYTHON"
for every feedback_client.py call (check, report, status, answer, and link).
Bare python3 is the fallback only when neither vault path is executable.
FEEDBACK_PYTHON="python3"
if [ -x "${CLAUDE_DIR:-}/.venv/bin/python" ]; then
FEEDBACK_PYTHON="$CLAUDE_DIR/.venv/bin/python"
elif [ -x "${VAULT_PATH:-}/.venv/bin/python" ]; then
FEEDBACK_PYTHON="$VAULT_PATH/.venv/bin/python"
fi
test cases passed, vs 6 / 18 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.