Originally created by: lodar
A new 5dive bug verb that opens a public GitHub issue from a payload built as an allowlist, never a denylist: every field is named once in a single jq -n object literal, so a field that is not written cannot leak.
That direction matters more than it sounds. Stripping risky keys off a richer object fails open — the next field someone adds to the source object ships by default. Naming each field fails closed.
tests/bug_report_pii_allowlist_unit.sh plants a PII-shaped marker into the excluded free-text fields (.reason, .detail) and greps every downstream surface for it. 17/0.
I asked for a key-set arm pinning the payload's top-level keys, believing check [#1] asserted shape rather than an exact key set. It already was the pin — a literal expected-keys list that fails on any new field.
The instructive part is how I convinced myself otherwise. I mutated by adding an unused --arg leaked_home and the suite stayed 17/0, which I read as the arm being blind. dev2 reproduced it, then diffed the emitted payload and found it byte-identical: jq -cn only emits a var the final object literal names, so the --arg never reached the output. I verified that the argument was added and never that the output changed — and an unused --arg is invisible precisely because jq is designed to ignore it. Wiring leaked_home into the object literal for real turned check [#1] red at 16/1, naming the extra key.
So: no new arm, and a 6-line comment on check [#1] recording why the obvious mutation looks vacuous — the next reviewer will have the same doubt, and documenting it beats a second arm testing the same property.
A mutation you did not confirm changed the output is not a mutation, and a green under one is worth nothing.
No logic change, 17/17, pii-scan clean.
Authored by dev2; reviewed, mutation-checked and merged by main (dev2 holds no gh credential).
🤖 Generated with Claude Code
Ticket changed by: lodar