Fault Recovery

Debug by narrowing, not by guessing. Random edits hide bugs more often than they fix them — and leave you unsure which change mattered. Reproduce the failure reliably, read what the system actually said, isolate the smallest trigger, test one hypothesis at a time, fix the cause, and lock it with a regression test.

This skill is for methodical diagnosis in dev, CI, staging, or post-mitigation production investigation. When users are actively impacted and you need to stop the bleeding first, pair with [[incident-response]] — mitigate before deep root-cause work. For preventing failures at boundaries, see [[resilience]]. For telemetry during diagnosis, see [[observability]].

When to Use

• A crash, exception, failing test, or wrong output
• Behavior that contradicts what the code should do
• A regression after a deploy or local change
• Flaky or intermittent failures — "works sometimes"
• A bug report that needs reproduction before fix
• Browser-only or environment-specific failures ([[browser-checks]])
• Post-incident root-cause analysis after service is stable ([[incident-response]])

Skip the full process for a typo with an obvious one-line fix — but still confirm with a test or reproduction if the surface area isn't trivial.

Not this skill alone: active production outage with no mitigation → [[incident-response]] first.

Process

Work in order. Don't edit code until reproduction and the error are understood.

1. Capture what you know — before touching code

Write a failure record — even three lines — so you don't lose facts while guessing:

Expected: checkout total includes tax for EU addresses
Actual: tax line shows €0.00
First seen: after commit abc123 / in CI job X / only on Safari
Repro rate: always / ~30% / only under load
Error text: (paste or link)

If you can't fill expected vs actual, clarify with the reporter ([[spec-first]]) before debugging.

2. Reproduce it reliably

A bug you can't trigger on demand, you can't prove you fixed.

Make reproduction deterministic:

• Pin inputs — exact URL, user role, feature flags, request payload, seed data
• Pin environment — OS, browser, Node/Python version, env vars, local vs CI
• Pin timing — race bugs need a script or stress loop, not "I clicked fast"
• Minimize steps — shortest path from cold start to failure

If intermittent:

• Run repeatedly (--repeat, loop script) to estimate frequency
• Look for shared preconditions — cache state, clock, concurrency, test order
• Check test pollution — does it pass alone but fail in the full suite?

No repro after reasonable effort → gather more telemetry ([[observability]]), add logging at boundaries, or pair with the reporter. Don't "fix" what you can't see fail.

3. Read the actual error — fully

The message, stack trace, and logs usually name the neighborhood. Read them before theorizing.

Stack traces

• Start at the top meaningful frame in your code — not the framework wrapper
• Walk caused by chains — root exception may be deeper
• Note file, line, and function — open that code before editing elsewhere

Logs

• Filter by request/correlation id ([[observability]])
• Read surrounding context — what happened in the 10 lines before the error?
• Check level — is this the first error or a downstream symptom?

Tests

• Read the assertion message — expected vs actual often pin the bug
• Run the single failing test isolated — pytest path::test, npm test -- -t name

Generic errors ("Internal Server Error", "Something went wrong") still point somewhere — trace from the handler to where the real exception was swallowed ([[resilience]] red flag).

4. Narrow the surface

Shrink the problem space until the failure is small and local.

Git bisect — regression after a range of commits:

git bisect start
git bisect bad          # current broken
git bisect good <sha>   # known good
# test each step; git bisect good/bad until found
git bisect reset

Code bisect — disable, stub, or bypass halves:

• Comment out middleware, feature branches, or optional steps
• Replace remote call with fixture — does it still fail?
• Binary search through a pipeline: parse → validate → transform → persist

Input minimization — shrink to smallest failing case:

• Smallest JSON payload, shortest string, single row in DB
• HTML/CSS bisect for layout bugs — remove sections until failure stops

Scope by layer

Document what you ruled out — "not the database" is progress.

5. Form one hypothesis

A hypothesis is a testable statement that explains all observed evidence:

Bad: "Something's wrong with tax"
Good: "Tax is zero because regionCode is null when EU address uses legacy ISO field"

One hypothesis at a time. If you have two equally likely theories, pick the one you can disprove fastest (cheapest experiment).

Confirm assumptions against source ([[source-first]]) — API signature, default value, query result — not memory of how it "should" work.

6. Test with a single change

Change one thing, observe, record result. Repeat.

Keep a debug log for non-obvious sessions:

H1: null region → added log in TaxService → regionCode null on EU path → confirmed
H2: legacy field not mapped → fixed mapper → test green

If a change doesn't help, revert it before the next experiment. Debug commits should not accumulate noise.

Prove the hypothesis:

• Does the failure disappear when you undo the suspected cause?
• Does artificially triggering the cause reproduce the failure again?

7. Fix the cause, not the symptom

Symptom fixes ship silence — the bug persists or corrupts data elsewhere. If a workaround is required temporarily (production hotfix), track a follow-up to fix root cause ([[incident-response]]).

8. Lock it in with a regression test

Before closing the bug ([[test-first]]):

1. Failing test that reproduces the bug (or the minimal case you isolated)
2. Fix the cause
3. Test green — proves fix and guards against return
4. Related cases — edge cases the bug revealed

Name the test after the behavior: applies_eu_tax_when_legacy_region_field_set.

For flaky bugs, the test may need deterministic timing or explicit ordering — document why.

9. If stuck — widen the lens systematically

After one focused hour (or two failed hypotheses with no new data), check common non-code causes:

Environment & versions

• Dependency version diff local vs CI ([[dependency-hygiene]])
• Wrong env var, missing secret, stale .env
• Node/Java/Python version mismatch
• Docker image not rebuilt

Data

• Production data shape vs fixture — nulls, unicode, huge values, time zones
• Stale cache, migration not applied, seed out of date
• Wrong tenant/user context

Concurrency & timing

• Race between async steps — add ordering or await
• Test parallelization — pytest -n0 to compare
• Clock skew, TTL expiry, debounce timing

Assumptions

• List what you "know" — verify each against source ([[source-first]])
• Ask: what would I expect to see in logs if this theory were true? Do I see it?

Get a second pair of eyes with the failure record and debug log — not "it doesn't work."

Time-box — if still stuck, escalate: more telemetry, pair debug, or scope reduction.

10. Scenario playbooks

Failing unit/integration test

1. Run isolated → read assertion → open test + implementation.
2. git diff recent changes if regression.
3. Hypothesis → single fix → green → keep test.

CI-only failure

1. Diff CI vs local: env vars, versions, services, parallelism.
2. Re-run failed job with debug logging enabled.
3. Reproduce in CI-like container locally.

Flaky test

1. Estimate rate — loop 100×.
2. Run alone vs full suite — order dependency?
3. Fix determinism: fixed clock, seeded RNG, await async, isolate shared state.
4. Never "retry 3×" as the only fix without understanding why.

Browser / UI bug ([[browser-checks]])

1. Reproduce in browser with DevTools open — console, network first.
2. Compare network response to what UI shows — backend vs front-end split.
3. Narrow: component state → props → fetch → API.
4. Screenshot + HAR for intermittent UI issues.

Wrong output, no exception

1. Trace data flow from input to output — log at boundaries.
2. Find where value first goes wrong — bisect pipeline.
3. Often config, mapping, or unit mismatch — not algorithm.

Production bug (service stable)

1. Confirm mitigation if impact ongoing ([[incident-response]]).
2. Use traces/logs/metrics to find failing requests ([[observability]]).
3. Reproduce in staging with production-like data snapshot when possible.
4. Fix + test + staged rollout ([[launch-readiness]]).

"Worked yesterday" regression

1. git bisect or scan deploy changelog.
2. First bad commit → read that diff carefully — often one logical change.
3. Revert to confirm before forward fix.

Common Rationalizations

• "I'll just try a few things." — Untracked changes create new bugs and hide the real one.
• "I don't need to reproduce it." — You can't prove the fix worked.
• "The error message is generic." — Trace deeper; generic often means swallowed exception.
• "It's probably a fluke." — Intermittent bugs have a condition you haven't found yet.
• "I'll add a quick fix and clean up later." — Quick symptom fixes become permanent debt.
• "Tests pass now, ship it." — Did you add a regression test? Did you run the full suite?
• "I know this API." — Verify against source; wrong mental models waste hours ([[source-first]]).
• "More logging everywhere." — Log at boundaries with correlation ids; don't flood and bury signal.

Red Flags

• Changing code before reproducing the failure
• Multiple simultaneous edits with no record of which mattered
• Skimming the stack trace and guessing from the file name
• Fixing by masking — catch-ignore, retry-until-pass, hide error in UI
• Fix with no regression test ([[test-first]])
• Debugging production under active user impact without mitigation ([[incident-response]])
• console.log / print spam committed as the "fix"
• Flaky test "fixed" with retries only
• Same bug reopened twice — previous fix was symptom-only
• Hours spent without a written failure record or hypothesis list

Verification

• Failure record written — expected vs actual, environment, repro steps
• Failure reproduced reliably (or repro gap documented with telemetry plan)
• Error, stack trace, and logs read fully — not skimmed
• Surface narrowed — git/code/input bisect; ruled-out areas noted
• One hypothesis tested at a time; debug log or clear commit history
• Suspect behavior confirmed against source ([[source-first]])
• Root cause fixed — not symptom masked without tracked follow-up
• Regression test added and full relevant suite green ([[test-first]])
• Flaky root cause addressed if applicable — not retry-only
• Production fix staged/rolled out with monitoring ([[launch-readiness]], [[observability]])