fixtureblame FIELD NOTES / 001

FOR AGENT-WRITTEN, DATABASE-BACKED TEST SUITES

The build is red.
Find the test
that changed the data.

A spec passes alone and breaks the next one. fixtureblame finds a deterministic polluter with controlled reruns, then shows the PostgreSQL rows it changed — inside your own CI.

Start a two-week pilot Inspect the evidence

Runs after a failing test step. Hard rerun and time budgets. Unknown when evidence runs out.

THE METHOD

A finding you can reproduce.

  1. 01 / ISOLATE

    Repeat the victim.

    Reset and run it alone twice. Alternating outcomes are nondeterministic, not proof of pollution.

  2. 02 / NARROW

    Find the polluter.

    Try changed specs first, then bisect known earlier specs in the same worker. Verify the singleton and reset.

  3. 03 / DIFF

    Name the rows.

    Snapshot, run the polluter alone, snapshot again. Label inserted, deleted and changed rows with your fixtures.

ILLUSTRATIVE REPORT / NOT CUSTOMER DATA

Evidence before blame.

[COUPLED]PR attribution: newly_observed6 experiments · 1.2 CI-min

recap.spec.ts fails after rsvp.spec.ts.

The named polluter inserted a shared attendee. The saved base report passed; that supports newly observed, not confirmed PR causality. A PR can introduce the polluter itself.

Polluter-only database mutation
OperationTable / keyFixture / owner
+ insertpublic.attendees
id = 2		shared e2etester attendee
specs/recap
  1. Victim alone ×2: passed
  2. After named polluter: failed
  3. After reset: passed

Mechanism and PR attribution are separate. Missing, skipped or stale baseline observations stay unknown.

THE FAILURE HAS A PATTERN

Shared state hides the real input.

“The production behavior was right; the test boundary was wrong.”
Bob / TimeToBuildBob · 27 July 2026
“The tests pass when run in isolation and fail when run together — or worse, pass in one order and fail in another.”
Peyton Green · 7 April 2026

These are independent operator accounts of the problem, not customer endorsements. Isolation and explicit fixture ownership remain the long-term fix.

TWO WEEKS / YOUR TEAM / YOUR CI

Measure whether the evidence earns its cost.

The pilot is for an agent-heavy engineering team with a real CI budget. Your team independently confirms the polluter and at least one mutation. These are proposed pilot thresholds, not results or guarantees.

  • ≥ 5coupled or pre-existing failures observed
  • ≥ 70%coupled findings confirmed by your team
  • ≤ 1confirmed-wrong blame
  • ≤ 15 minmedian CI cost per failure

Customer-local by design. No telemetry, hosted ingest or model API. Data stays in your CI; optional GitHub output uses your token.

Start a two-week pilot