Skills · Code

CodeSkillDugout-created

/analyse-parallel-bugs

Finds how many real causes sit behind a pile of bug reports, checks each one, and orders them by what they cost.

Created by Piyush Baheti · Source: piyush-tkd/skillsdugout/tree/main/plugins/analyse-parallel-bugs

Who it is for

Anyone handed several problems at once — after a deploy, from support, or from a busy floor — who needs to know what to fix first.

Started as a checklist after too many "six bugs" turned out to be two.

When to use it

  • Several problems are reported together and may share a cause.
  • One report might have more than one cause behind it.
  • You need an ordered list someone can act on, with what is proven and what is not.

Example

You ask

Five things broke after Tuesday's 14:00 deploy: emails with no prices, $0 revenue in the sales report, the mobile discount button does nothing, invoice totals are blank, and product pages show UTC. The deploy renamed price to unit_price. The mobile problem was first reported Monday. The spec says product pages use UTC.

It writes a triage report

Triage: 5 reports after Tuesday 14:00 deploy

Conclusion: the 5 reports map to 3 causes, and only 2 are defects. One cause is the field rename and is likely to be the deploy's fault. One is a mobile discount bug that predates the deploy, so the deploy did not cause it. One is not a bug.

Symptoms (as reported)

  1. Confirmation emails arrive with no prices.
  2. Admin sales report shows $0 revenue for Tuesday afternoon.
  3. On mobile, the discount code button does nothing (first reported Monday).
  4. Invoice PDF shows prices, but the total is blank.
  5. Product page "Last updated" shows UTC, not local time.

Grouping (UNPROVEN)

Cause Reports Status
A. price renamed to unit_price 1, 2, 4 UNPROVEN
B. Mobile discount button failure 3 UNPROVEN, cause unknown
C. UTC timestamp 5 Not a defect, per the spec you gave

Cause A: consumers still read the old price field

  • Evidence:
    • Report 1: the email template is missing prices.
    • Report 2: the revenue report is $0 for Tuesday afternoon.
    • Report 4: the invoice total is blank.
    • All three began at the 14:00 Tuesday deploy, which is the same moment price was renamed to unit_price.
  • Falsifier: if the email, report and invoice code paths do not read order.price, or if the failures predate 14:00, this cause is wrong.
  • Explains: reports 1, 2 and 4.
  • Unexplained wrinkle: report 4 shows prices but a blank total. If the rename were the whole story, the invoice's price column should be blank too. One reading is that line items already use unit_price while the total calculation still reads price. That is untested. Invoices may therefore have a second cause, and fixing the email and report alone may not clear report 4.
  • Classification: broken.
  • Cost to leave:
    • Customers receive wrong emails and invoices, and the cause is still active.
    • Finance sees $0 revenue for the afternoon, with no error shown.
    • The damage grows with every order placed.
  • Cost to fix:
    • Small if it is a field-read change in three consumers, but this is a rough estimate with no code seen.
    • Search for every reader of price, not just these three.
    • Resending the emails and reissuing the invoices is a separate decision.
  • Not verified:
    • Whether stored order data is correct (are the prices saved, or only displayed wrongly?).
    • Whether customers were charged correctly.
    • Whether the report is computed live (fixes itself) or stored (needs a rebuild).
    • Which orders since 14:00 were affected.

Cause B: mobile discount button does nothing

  • Evidence: first reported Monday, before the deploy, so it cannot come from the deploy or the rename. There is nothing else to go on.
  • Falsifier: if the failure reproduces only on builds from after Tuesday 14:00, the Monday report is a different issue.
  • Explains: report 3 only.
  • Classification: broken, with the cause undetermined.
  • Cost to leave: customers on mobile are blocked from applying discounts and may abandon the purchase. Desktop is a possible workaround, but that is unverified.
  • Cost to fix: unknown.
  • Next step to settle it: reproduce on a mobile device, then check the browser console and network call on button press, and the device, browser and app versions in the affected reports.

Cause C: "Last updated" shows UTC

  • Evidence: the product spec says product-page timestamps are shown in UTC, so the behavior matches the spec.
  • Classification: correct and surprising. If the label does not say "UTC", it is also badly explained.
  • Action: do not change the logic. Optionally add a "UTC" suffix to the label. If the reporter expects local time, that is a product decision to change the spec, not a bug fix.
  • Not verified: the spec text itself. I took it from your note.

Priority order

  1. A (wrong data leaving the system). Customer-facing emails and invoices are wrong, and the revenue report is silently wrong.
  2. B (blocked work). Customers cannot apply discounts on mobile.
  3. C (cosmetic). At most a label tweak.

Recommendation

  • Now:
    • Fix A across every reader of price.
    • Re-test the invoice total separately, since it behaves differently from the other two.
  • Ticket:
    • Investigate B as its own bug, and tell the reporter it is not deploy-related.
    • Add "UTC" to the C label, as low priority.
  • Leave: the C logic.

Output is Markdown (.md), shown formatted. Sample run on 2026-10-03 with an illustrative shop scenario. One closing section of follow-up questions trimmed.

Evidence

Our test

No baseline test yet. We will publish one, with and without the skill, when testing is done.

Reported by the community, by app

Claude Code: Tested
No outcomes reported yet
Claude apps: Not tested
No outcomes reported yet
Cursor: Not tested
No outcomes reported yet
Codex: Not tested
No outcomes reported yet
GitHub Copilot: Not tested
No outcomes reported yet

Compatibility labels come from our tests only. Community reports add detail but cannot mark an app as tested.

Editorial assessment · not a community rating

Editorial review in progress

We have not finished testing this listing against doing the task without it. The verdict appears here when we have. Our team made this skill; it gets the same test as external work.

How we review

Setup and customization

  • Works with no setup after install.
  • Works best with access to the code, logs or data it needs to check each cause.

What it can access

Reads
The reports you give it; Code, logs or data you point it to
Changes
Nothing; it writes a report in the conversation

Install

Claude Code: Tested
  1. Add the Skills Dugout marketplace once.
    /plugin marketplace add piyush-tkd/skillsdugout
  2. Install /analyse-parallel-bugs.
    /plugin install analyse-parallel-bugs@skillsdugout
  3. Start a new session and ask for it, for example "here are five things that broke after the deploy".

Community reviews

No reviews yet. Be the first to rate it after you try it.

Each review records the app, version, task and outcome; those outcomes feed the Evidence section. One rating per person; creators reply but cannot rate their own work.

No reviews yet. Be the first to rate it after you try it.

Source, license and versions

Source
github.com/piyush-tkd/skillsdugout/tree/main/plugins/analyse-parallel-bugs
License
MIT, as stated in the repository
  • 0.1.0 · 2026-10-03 · First release.
Show the full skill instructions

Analysing several bugs at once

What this is for

A list of problems arrives together. The job is to find out how many causes there actually are, verify each one, and hand back an ordered list somebody can act on.

The whole reason this needs a method is that the obvious approach — take them in order, diagnose each — gets two things wrong. It misses that six reports are often three causes, and it spends the same effort on the cosmetic one as on the one moving money.

The one rule that matters most

A count of reports is not a count of bugs. Sketch the mapping before going deep on any one report, and revise it as evidence comes in — diagnosis will often change it.

Three shapes, and they need different work:

  • Many reports, one cause. The common case after a deploy. Find it once and six tickets close.
  • One report, many causes. The dangerous case. Fix the first, the symptom persists, and the reporter concludes the fix did not work.
  • Genuinely independent. Then order them and go, but prove it rather than assuming it.

The pass

1. List the symptoms, not the diagnoses

Write down what was observed, in the reporter's words. Not what you think causes it.

A symptom recorded as "the retry logic is broken" has already picked a suspect, and every later step will look for evidence that confirms it. "The row still shows as failed after pressing retry" keeps the field open.

2. Group by suspected shared cause, and mark each group UNPROVEN

Cluster them. Then write the word UNPROVEN next to each cluster, because a grouping is a hypothesis and it is the one most likely to be wrong — related symptoms feel like one bug long before they are shown to be one.

The cheapest test of a grouping is usually a timestamp or a version: did they all start at the same moment? Did they all appear in the same release?

3. Verify each cause separately, and name the falsifier first

For each cause, before you go looking: write down what would make this claim false, then go and look for that.

Confirming evidence is easy to find for a wrong answer. An honest attempt to falsify is the only thing that separates a verified claim from a plausible one.

Specific traps, each of which has produced a confident wrong answer:

  • A sample proves what it contains, never what it lacks. Searching a sample, not finding the thing, and concluding it does not exist is only valid if the sample was guaranteed to contain it. Check that the population you sampled is the population that would hold the answer.
  • Do not generalise from the part you happened to read. Reading a fraction of a diff, a log or a file and concluding about the whole is a guess. Find the check that covers all of it — usually cheaper than reading the rest.
  • Agreement today is not agreement by design. Two numbers that match on young data may be coincidence. Ask what would separate them and whether that has happened yet.
  • Check whether the defect is yours. Before reporting something as pre-existing, confirm it is not from work done in this session. Attributing your own mistake to the codebase wastes everybody's time twice.

4. Separate what is broken from what is merely surprising

Not everything reported is a defect. Before calling any behaviour correct, establish what was expected — from a requirement, a spec, or the person who owns the behaviour. Existing code alone does not prove intent.

Four outcomes, and they read very differently to whoever raised it:

  • Broken — behaves wrongly. Fix it.
  • Correct but badly explained — the behaviour is right and the label, message or name lies about it. Fix the words, and say plainly that the behaviour was right, because "we changed the label" reads as a brush-off unless the reason is given.
  • Correct and surprising — right, well described, and still unexpected. Explain it. Changing it would be the bug.
  • Undetermined — the expected behaviour or the evidence is not settled yet. Say what would settle it, and do not force one of the other three.

Getting this wrong in either direction is expensive. Treating a wording problem as a logic bug produces a migration nobody needed; treating a logic bug as a wording problem ships the wrong numbers with nicer labels on them.

5. Order by consequence, not by effort

Sort by what it costs to leave alone. The default order:

  1. Wrong data leaving the system — anything reaching a customer, a payer, a partner or a ledger
  2. Blocked work — somebody cannot do their job
  3. Invisible failure — it is wrong and nothing says so
  4. Visible annoyance — wrong, and everybody can see it is wrong
  5. Cosmetic

Adjust it with judgment: how many people it reaches, whether the harm is still happening, and whether it can be undone. An invisible failure that corrupts records daily outranks blocked work with a workaround.

A cheap fix low on that list can be done first for momentum; say that you are doing it for that reason rather than letting the order imply it matters more.

6. Report

Per cause, tight:

  • What is wrong, in plain words, and the evidence that settles it — a number, a query result, a code path
  • Which reported symptoms it explains. If a symptom is left over, mark it unexplained — it may be another cause, or an incomplete explanation of this one
  • Broken / badly explained / correct-and-surprising / undetermined
  • What it costs to leave, and roughly what it costs to fix
  • Anything you could not verify, and what would settle it

Then one recommendation for the whole set: what to do now, what to ticket, what to leave.

Say what you did not check

An unverified claim presented alongside verified ones inherits their credibility. Mark it.

"I think X, and the way to know is Y" is worth more than a confident wrong answer, and it costs nothing to write. An analysis with three proven findings and one honest "unproven" is far more useful than four findings of unmarked mixed quality.

Red flags — stop and re-check

  • You have as many causes as there were reports. Suspiciously tidy; look for the shared one
  • A cause was found without anything being run or read
  • Evidence was gathered only to confirm an answer already chosen
  • You are about to report a defect in code you touched this session
  • A symptom is unexplained but the analysis is being presented as complete
  • Everything on the list is high priority