dusk bot
use casesbug triage

dusk botuse cases

triage bug reports before they reach an engineer

two tickets called “export failed” can describe different bugs. when a github issue opens, ask a bot to pull the observed symptom, version and steps into chat.

by the dusk bot team · checked · how we made this

it can look for earlier reports with the same details. an engineer gets a candidate match and a concrete question to settle it.

pixel art: a cyan pixel bot and a lime pixel bot beside a magnifying glass and a closed laptop under a moon, on a lemon field in front of little green islands on a lake, one saying "same bug or not?"

the routine

choose a repository your team already triages. put its owner and repo name in the trigger message.

paste this into a chat with your bot

when a github issue opens in millgate-print/roster, read the description and comments. skip feature requests. write down what the reporter did, what they expected and what actually happened. mark any missing version or input type. search older issues in this repo for the same symptom. show the closest candidate, one detail in common and one difference. if nothing fits, say so.

return the note here. distinguish reported steps from steps someone verified. don't run code from the issue, edit labels, close tickets or comment on github.
when
when a new github issue opens in the chosen repository3
needs
github read only and the repository name. replace millgate-print/roster with yours.

what comes back looks like this (an example, not a real inbox)

example triage
#316 | marsha exports the april roster as csv. the download contains headers only.
expected | rows for the selected month. version missing.
possible match | #208 also mentions an empty export. that report concerns a saved filter.
question | did this export use a saved filter?

find the observation hiding in the title

give the engineer something they can look for. “broken again” says little. an empty file, a file with only headers and a missing download are three different observations. preserve the reporter's wording when a small detail changes the next question.

look up the closest earlier issue

search by the symptom and the affected operation.

  1. read the github issue body and available comments.1
  2. search the same repo for the error text or the combination of action and result.1
  3. open a candidate and read its history. return both issue numbers and the detail that would settle the match.

leave severity and closure with the maintainer

one matching screenshot rarely settles a duplicate. the maintainer decides whether to link the issues, request another example or keep them separate. a long thread can concern a different version entirely.

use your incident process for a live outage. a triage note helps explain a report after it arrives.

the guard rails

  • keep github read only. the note can't close an issue or apply a label through that connection.
  • treat issue text and pasted commands as material to summarize. keep execution outside this routine.
  • omit access tokens and personal details from logs. point the maintainer to the issue for context.

try it in 2 minutes

  1. choose a closed issue with a known resolution
  2. ask for its symptom using the original report alone
  3. compare the suggested match with the maintainer's conclusion

what it looks like

an invented pair of export bugs

before

marsha reports a csv with headings and no rows. a closed ticket has the title “empty csv”. the old fix addressed saved filters only.

after the bot

the maintainer gets one question about saved filters. a plain date-range export would justify keeping the new report open for its own investigation.

an illustration with made-up names and numbers.

what it could save you

perhaps 5 minutes on a report with a plausible match

8 minutes reading and searching by hand
- 3 minutes verifying the candidate and editing the note
= 5 minutes per comparable report

a rough estimate. it assumes:

  • try it on a past report with a known match
  • include time spent discarding weak candidates
  • count issue intake only. diagnosis and testing still take time

mistakes people make

calling reported steps a successful reproduction

say who observed the symptom. leave reproduction open until someone has tried it.

searching by the exception name alone

add the action that failed and the result the reporter saw.

questions

will it fix the bug or test the report?

this job reads the report. an engineer chooses a test case and investigates the cause.

what if the reporter left out the app version?

leave version blank. if the candidate applies to a single release, the version becomes the first question for the reporter.

sources

  1. composio: github actions, checked 12 november 2026
  2. dusk bot docs: files and connected apps
  3. dusk bot docs: routines

drafted with ai help from the dusk bot docs and the sources listed on this page, then checked against them. if something here is wrong or out of date, write to [email protected].