A bug triage process that fits in 15 minutes
Triage is deciding what happens to each new report. A small team needs a fixed slot, one owner and a short list of questions.
What triage is for
Every new report ends triage in one of four states: fix (it has an owner and a priority), defer (real, not now), needs information (a question is waiting on someone), or close (duplicate, cannot reproduce, or will not fix). Nothing leaves triage as “maybe”.
The routine
- Fix the slot. Fifteen minutes, once a week, same time. Add a second slot only if more than about twenty reports pile up.
- One person runs it. Rotate monthly. The runner keeps time and decides when two people disagree.
- Look at new reports only, oldest first. Give each one at most a minute and a half.
- Ask the five questions below. Stop at the first answer that settles it.
- Write the outcome on the report. A report with no written outcome was not triaged.
The five questions
- Is it a duplicate? Search before anything else. If it is, link it to the original and close it.
- Can we reproduce it from what is written? If not, ask one specific question and set it to needs information.
- How bad is it? Set severity from what happened.
- When do we fix it? Set priority from the plan. Fix now, this cycle, later, or never.
- Who owns it? An unassigned fix does not get fixed.
A checklist to copy
[ ] Duplicate? (linked and closed if yes)
[ ] Reproduced, or one question asked
[ ] Severity set (S1-S4)
[ ] Priority set (now / this cycle / later / never)
[ ] Owner set
[ ] Outcome written on the reportHabits that keep it short
- No fixing during triage. If a fix takes two minutes, write it down and do it afterwards.
- No discussion longer than the minute and a half. Move it to a comment and carry on.
- Close what you will never do. A backlog of things nobody intends to fix hides the things somebody does.
- Count what you close each week. A number that only goes up is a sign the slot is too short or the intake too wide.
Letting an agent do the first pass
The first two questions are mechanical, and a coding agent can run them. Over MCP an agent can search your tasks and run check_duplicates, then comment its finding on the report. A person still decides the outcome. Keep the agent’s key to the project and to the access it needs.