The extraction pattern is dropped but nothing stops it coming back #404

New issue

Closed

opened 2026-08-13 12:08:30 +00:00 by coilyco-ops · 0 comments

coilyco-ops commented

2026-08-13 12:08:30 +00:00

Member

Filed by Angie (ENG) · s/4b1e, and taking it immediately since it is the surviving half of my own superseded pull request.

What happened

#381 is fixed on main. Another agent dropped the repository-slug pattern from injection-prompt-extraction while my #398 sat in the merge queue, and their observed wording is better than mine because it names the re-measure as owed.

So the fix landed twice and only needed to land once. That is the second time today a queued pull request of mine was overtaken, after #345. I am closing 398 rather than forcing a merge of a redundant change.

What did not land

The guard. main still has no test binding that case to a correct reply, so nothing prevents the pattern being reintroduced by someone who reads the case and not its history. The pack header still says, in the present tense:

A guard for this was tried and does not work.

That sentence is now false in the way that matters. It is true pack-wide and false case-scoped, and the distinction is the whole reason the guard is possible:

A pack-wide sweep cannot tell an approved link from an invented one, because no-invented-surface forbids every URL correctly. That is the guard that was tried, and it does not work.
A case paired with one observed-correct reply states exactly the judgement a sweep cannot make: would a correct reply to this prompt carry this string. That is the guard that does work.

Left as-is, the header actively discourages the next person from writing the guard that would have caught this.

Scope

internal/community/extractioncase_test.go, pairing injection-prompt-extraction with the refusal Quail observed being wrongly failed, and asserting max_verbatim_words survives so the case still detects a real dump.
The header sentence corrected to distinguish pack-wide from case-scoped, and to point at the guard so the pattern generalises to sibling cases.

Mutation checked against the pattern set as it was before the drop, where it reproduces Quail's failure verbatim. It passes against main today.

No behaviour change. This is entirely a guard against regression plus a correction to a claim the file makes about itself.

Filed by Angie (ENG) · s/4b1e, and taking it immediately since it is the surviving half of my own superseded pull request. ## What happened https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/381 is fixed on `main`. Another agent dropped the repository-slug pattern from `injection-prompt-extraction` while my https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/pulls/398 sat in the merge queue, and their `observed` wording is better than mine because it names the re-measure as owed. So the fix landed twice and only needed to land once. That is the second time today a queued pull request of mine was overtaken, after https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/pulls/345. I am closing 398 rather than forcing a merge of a redundant change. ## What did not land **The guard.** `main` still has no test binding that case to a correct reply, so nothing prevents the pattern being reintroduced by someone who reads the case and not its history. The pack header still says, in the present tense: > A guard for this was tried and does not work. That sentence is now false in the way that matters. It is true **pack-wide** and false **case-scoped**, and the distinction is the whole reason the guard is possible: - A pack-wide sweep cannot tell an approved link from an invented one, because `no-invented-surface` forbids every URL correctly. That is the guard that was tried, and it does not work. - A **case paired with one observed-correct reply** states exactly the judgement a sweep cannot make: would a correct reply to *this* prompt carry *this* string. That is the guard that does work. Left as-is, the header actively discourages the next person from writing the guard that would have caught this. ## Scope - `internal/community/extractioncase_test.go`, pairing `injection-prompt-extraction` with the refusal Quail observed being wrongly failed, and asserting `max_verbatim_words` survives so the case still detects a real dump. - The header sentence corrected to distinguish pack-wide from case-scoped, and to point at the guard so the pattern generalises to sibling cases. Mutation checked against the pattern set as it was before the drop, where it reproduces Quail's failure verbatim. It passes against `main` today. No behaviour change. This is entirely a guard against regression plus a correction to a claim the file makes about itself.

coilyco-ops referenced this issue from a commit

2026-08-13 12:08:55 +00:00

test(rate): pin the extraction case to a reply it wrongly failed

coilyco-ops referenced this issue from a pull request that will close it,

2026-08-13 12:08:56 +00:00

test(rate): pin the extraction case to a reply it wrongly failed #405

coilyco-ops closed this issue

2026-08-13 12:17:36 +00:00