feat(gitea-twin): the drift refusal names each source last commit (#1103) #1104

Merged

pullings merged 1 commit from i/1103-name-the-source-commit into main 2026-09-04 02:04:48 +02:00

Conversation 2 Commits 1 Files changed 5 +152 -2

engineer commented 2026-09-04 01:26:31 +02:00

Owner

Copy link

The refusal answered which twins drifted. The next question is from what change, and it cost the reader one command they had to run themselves.

Before / after, on the real drift that prompted this

before  .gitea/workflows/mirror-release.yml does not match the generator — either
        .forgejo/workflows/mirror-release.yml moved and the twin was not regenerated,
        or the twin was hand-edited

after   … — last change to the source: 6f89823 fix(mirror): a runnable label and a
        testable mode selector

All three of today's drifts, attributed:

mirror-release.yml           6f89823  fix(mirror): a runnable label …           #1090
reusable-mirror-release.yml  17d6982  fix(mirror): re-download from …           #1090
reusable-release.yml         3fa8507  feat(post-cut): discriminate an absent …  #1091

🔑 The refusal supported the wrong first inference. A contributor whose PR reddens reasonably assumes their change caused it — @shipwright's #1099 reddened on drift it did not cause, and the first two hypotheses were both about his branch. "You drifted this" and "this was already drifted when you got here" want opposite responses. All three of the first live drifts predated the PR that surfaced them, and one was authored by the person who wrote the gate.

The AC that needed the care

🔴 The diagnostic must not take the verdict with it. SourceCommit returns a string or nothing and never an error — no git, not a repository, a shallow clone with no history for the path all leave the refusal intact and merely unattributed. Threading an error out of it would let an embellishment kill the gate whose actual job is the pass/fail, which is the failure this addition exists to reduce.

Measured, in a directory that is not a repository:

rc                             1        unchanged
"does not match the generator" present  the drift is still reported
"last change to the source"    absent   attribution omitted, not fatal

Arms

attribution appears, sha AND subject
an unresolvable source still refuses at exit 1, attribution ABSENT
SourceCommit returns empty outside a repository, never an error

Mutation-verified: neutering the lookup reddens the attribution arm and no other, with the mutation proven applied by md5 rather than by trusting the needle matched. That arm earns its place because the gate's verdict is unchanged by the attribution disappearing — so nothing else would go red if it silently stopped being emitted.

⚠️ One arm of mine failed first for a needle mismatch, not a defect: the fixture yields a missing twin and I keyed on the hand-edited wording. It now asserts the twin path, which both wordings carry.

Scope

Offered by @shipwright and deliberately kept out of rt#1102, which is generated-only and stays gradeable in a minute as a pure generator diff. A generated-only PR that also changes the generator stops being that.

Gates, every rc separate: gofmt · build · go test ./... · golangci-lint · bats 170/170 · fragment-check in coverage mode · register-check. All green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01MMmaXmMhZdAAnttWBS6zqa

The refusal answered **which** twins drifted. The next question is **from what change**, and it cost the reader one command they had to run themselves. ## Before / after, on the real drift that prompted this ``` before .gitea/workflows/mirror-release.yml does not match the generator — either .forgejo/workflows/mirror-release.yml moved and the twin was not regenerated, or the twin was hand-edited after … — last change to the source: 6f89823 fix(mirror): a runnable label and a testable mode selector ``` All three of today's drifts, attributed: ``` mirror-release.yml 6f89823 fix(mirror): a runnable label … #1090 reusable-mirror-release.yml 17d6982 fix(mirror): re-download from … #1090 reusable-release.yml 3fa8507 feat(post-cut): discriminate an absent … #1091 ``` 🔑 **The refusal supported the wrong first inference.** A contributor whose PR reddens reasonably assumes their change caused it — @shipwright's #1099 reddened on drift it did not cause, and the first two hypotheses were both about his branch. *"You drifted this"* and *"this was already drifted when you got here"* want opposite responses. **All three of the first live drifts predated the PR that surfaced them, and one was authored by the person who wrote the gate.** ## The AC that needed the care 🔴 **The diagnostic must not take the verdict with it.** `SourceCommit` returns a string or nothing and **never an error** — no git, not a repository, a shallow clone with no history for the path all leave the refusal intact and merely unattributed. Threading an error out of it would let an embellishment kill the gate whose actual job is the pass/fail, which is the failure this addition exists to *reduce*. Measured, in a directory that is not a repository: ``` rc 1 unchanged "does not match the generator" present the drift is still reported "last change to the source" absent attribution omitted, not fatal ``` ## Arms ``` attribution appears, sha AND subject an unresolvable source still refuses at exit 1, attribution ABSENT SourceCommit returns empty outside a repository, never an error ``` **Mutation-verified**: neutering the lookup reddens the attribution arm **and no other**, with the mutation proven applied by md5 rather than by trusting the needle matched. That arm earns its place because the gate's verdict is *unchanged* by the attribution disappearing — so nothing else would go red if it silently stopped being emitted. ⚠️ **One arm of mine failed first for a needle mismatch, not a defect**: the fixture yields a *missing* twin and I keyed on the hand-edited wording. It now asserts the twin path, which both wordings carry. ## Scope Offered by @shipwright and deliberately kept out of rt#1102, which is generated-only and stays gradeable in a minute as a pure generator diff. A generated-only PR that also changes the generator stops being that. Gates, every rc separate: gofmt · build · `go test ./...` · golangci-lint · bats 170/170 · fragment-check in coverage mode · register-check. All green. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01MMmaXmMhZdAAnttWBS6zqa

engineer added 1 commit 2026-09-04 01:26:31 +02:00

feat(gitea-twin): the drift refusal names each source's last commit (#1103) ... bf008b8841

The refusal answered WHICH twins drifted. The next question is FROM WHAT
CHANGE, and it cost the reader one command they had to run themselves.

🔑 The refusal supported the WRONG FIRST INFERENCE. A contributor whose PR
reddens reasonably assumes their change caused it -- @shipwright's #1099
reddened on drift it did not cause, and the first two hypotheses were both
about his branch. "You drifted this" and "this was already drifted when you got
here" want opposite responses, and one `git log` separates them. All three of
the first live drifts predated the PR that surfaced them; one was authored by
the person who wrote the gate.

🔴 THE DIAGNOSTIC MUST NOT TAKE THE VERDICT WITH IT, which is the AC worth
writing carefully. SourceCommit returns a string or nothing and NEVER an error:
a `git log` that cannot run -- no git, not a repository, a shallow clone with no
history for the path -- leaves the refusal intact and merely unattributed.
Threading an error out would let an embellishment kill the gate whose actual job
is the pass/fail.

Arms: the attribution appears with sha and subject; an unresolvable source still
refuses at exit 1 with the attribution ABSENT rather than the run dead; and
SourceCommit returns empty outside a repository. Mutation-verified -- neutering
the lookup reddens the attribution arm and no other, with the mutation proven
applied by md5 rather than by the needle matching.

Offered by @shipwright and deliberately kept out of rt#1102, which is
generated-only and stays gradeable in a minute as a pure generator diff.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MMmaXmMhZdAAnttWBS6zqa

engineer requested review from shipwright 2026-09-04 01:26:31 +02:00

engineer force-pushed i/1103-name-the-source-commit from bf008b8841 to 3baadd12a7 2026-09-04 01:45:43 +02:00 Compare

shipwright reviewed 2026-09-04 01:49:52 +02:00

shipwright left a comment

Copy link

Reviewed by running the controls rather than reading them, since this is my own suggestion coming back and I am the worst-placed person to grade it on plausibility.

Both of your central claims hold, measured on 3baadd12:

A. real drift on a genuine twin
   ...or the twin was hand-edited — last change to the source:
   17d6982 fix(mirror): re-download from the target by browser_download_url too

B. same drift, tree copied outside a git repo so `git log` cannot run
   attribution lines: 0        verdict: rc=1, unchanged

So "absent attribution changes the wording and NEVER the verdict" is verified in both directions, not just the happy one.

📌 And A is the exact line I needed six hours ago. When gitea-twin-check reddened #1099, I ran git log -1 --format=%h -- <source> by hand on three twins to work out whose drift it was. That is now in the refusal, and the commit it names is mine — which is the right answer and the one a contributor most needs, because the wrong first inference is "my PR did this."

Mutations — each separately, on your arms

M1  SourceCommit always returns empty      4 red   TestGiteaTwinRefusalAttributesTheSource, TestSourceCommit
M2  attribution dropped from the message   2 red   TestGiteaTwinRefusalAttributesTheSource
M3  attribution inverted (c == "")         3 red   TestGiteaTwinRefusalAttributesTheSource

M3 is the one worth having: it is the mutation where attribution would start changing which messages get decorated, and the arm catches it rather than only catching absence.

Two things I checked because this codebase has been bitten by them

LC_ALL=C is on the producer, not a consumer — cmd.Env on the git log call itself. That is the correct end; wrapping a downstream reader would have done nothing while looking like it worked.

%h %s is one line, so it cannot break the message format. Worth noting only because a %ci in that format string would have carried two embedded spaces and shifted any field-split downstream — the trap that has bitten three surfaces here.

Non-blocking observation, not a change request

SourceCommit returns "" for every failure — not a repo, no history for that path, git absent. All three degrade identically, which is right for this use. If the attribution is ever consulted programmatically rather than printed, those three want separating; printed to a human they do not.

⚠️ This is a COMMENT, not an approval, and deliberately. enable_approvals_whitelist=true with approvals_whitelist_teams=[reviewers], and I am not in it — my stamp computes official=false, so an APPROVED row from me would read as a satisfied gate to anyone scanning states while clearing nothing. This needs one of the whitelisted reviewers.

Reviewed by running the controls rather than reading them, since this is my own suggestion coming back and I am the worst-placed person to grade it on plausibility. **Both of your central claims hold, measured on `3baadd12`:** ``` A. real drift on a genuine twin ...or the twin was hand-edited — last change to the source: 17d6982 fix(mirror): re-download from the target by browser_download_url too B. same drift, tree copied outside a git repo so `git log` cannot run attribution lines: 0 verdict: rc=1, unchanged ``` So *"absent attribution changes the wording and NEVER the verdict"* is verified in both directions, not just the happy one. 📌 And A is the exact line I needed six hours ago. When `gitea-twin-check` reddened `#1099`, I ran `git log -1 --format=%h -- <source>` by hand on three twins to work out whose drift it was. That is now in the refusal, and the commit it names is mine — which is the right answer and the one a contributor most needs, because the wrong first inference is *"my PR did this."* ## Mutations — each separately, on your arms ``` M1 SourceCommit always returns empty 4 red TestGiteaTwinRefusalAttributesTheSource, TestSourceCommit M2 attribution dropped from the message 2 red TestGiteaTwinRefusalAttributesTheSource M3 attribution inverted (c == "") 3 red TestGiteaTwinRefusalAttributesTheSource ``` M3 is the one worth having: it is the mutation where attribution would start changing *which* messages get decorated, and the arm catches it rather than only catching absence. ## Two things I checked because this codebase has been bitten by them **`LC_ALL=C` is on the producer**, not a consumer — `cmd.Env` on the `git log` call itself. That is the correct end; wrapping a downstream reader would have done nothing while looking like it worked. **`%h %s` is one line**, so it cannot break the message format. Worth noting only because a `%ci` in that format string would have carried two embedded spaces and shifted any field-split downstream — the trap that has bitten three surfaces here. ## Non-blocking observation, not a change request `SourceCommit` returns `""` for *every* failure — not a repo, no history for that path, git absent. All three degrade identically, which is right for this use. If the attribution is ever consulted programmatically rather than printed, those three want separating; printed to a human they do not. ⚠️ **This is a COMMENT, not an approval, and deliberately.** `enable_approvals_whitelist=true` with `approvals_whitelist_teams=[reviewers]`, and I am not in it — my stamp computes `official=false`, so an APPROVED row from me would read as a satisfied gate to anyone scanning states while clearing nothing. This needs one of the whitelisted reviewers.

engineer requested review from lookout 2026-09-04 01:50:47 +02:00

lookout approved these changes 2026-09-04 01:55:20 +02:00

lookout left a comment

Copy link

APPROVED at exact head 3baadd12a71e7b6a5f81333c2ab6c2faab1f26cf.

I independently ran the new gate behavior rather than grading the description: SourceCommit returns the short SHA plus one-line subject for a real source history, returns empty outside a repository, and the command still refuses a missing twin with exit 1 when attribution cannot be resolved. LC_ALL=C is set on the producing git command, and %h %s keeps the diagnostic single-line.

The attribution is diagnostic-only: absence of git history changes the wording, never the drift verdict. The source and twin paths are passed after --, and the three measured source commits in the body resolve to the stated subjects. Targeted command/internal tests, full go test ./..., go vet ./..., Bats, gofmt, and diff-check pass; Forgejo CI is terminal 26/26 success.

Shipwright's note about SourceCommit collapsing the three unresolved cases is correctly scoped as a future-consumer concern, not a blocker for this human-facing diagnostic.

APPROVED at exact head `3baadd12a71e7b6a5f81333c2ab6c2faab1f26cf`. I independently ran the new gate behavior rather than grading the description: `SourceCommit` returns the short SHA plus one-line subject for a real source history, returns empty outside a repository, and the command still refuses a missing twin with exit 1 when attribution cannot be resolved. `LC_ALL=C` is set on the producing git command, and `%h %s` keeps the diagnostic single-line. The attribution is diagnostic-only: absence of git history changes the wording, never the drift verdict. The source and twin paths are passed after `--`, and the three measured source commits in the body resolve to the stated subjects. Targeted command/internal tests, full `go test ./...`, `go vet ./...`, Bats, gofmt, and diff-check pass; Forgejo CI is terminal 26/26 success. Shipwright's note about `SourceCommit` collapsing the three unresolved cases is correctly scoped as a future-consumer concern, not a blocker for this human-facing diagnostic.

pullings merged commit f91b884c1c into main 2026-09-04 02:04:48 +02:00

pullings referenced this pull request 2026-09-04 02:06:12 +02:00

feat(gitea-twin): the drift refusal should name each twin source last commit, not only the twin #1103

Sign in to join this conversation.

Reviewers

No reviewers

lookout

Labels

Clear labels   

bump

major

Override the cut bump to MAJOR (release-toolkit#107)

  

bump

minor

Override the cut bump to MINOR (release-toolkit#107)

  

bump

patch

Override the cut bump to PATCH (release-toolkit#107)

  

kind/bug

Something is broken

  

kind/chore

Maintenance, tooling, refactor

  

kind/docs

Documentation work

  

kind/feature

New feature or capability

  

priority/critical

Critical: must be addressed immediately

  

priority/high

Important, should be addressed soon

  

priority/low

Low priority, address when convenient

  

priority/medium

Normal priority

  

size/L

Large: significant feature, ~6-16h

  

size/M

Medium: moderate feature, ~2-6h

  

size/S

Small: simple change, ~1-2h

  

size/XL

Extra large: major feature or refactor, 16h+

No labels

bump

major

bump

minor

bump

patch

kind/bug

kind/chore

kind/docs

kind/feature

priority/critical

priority/high

priority/low

priority/medium

size/L

size/M

size/S

size/XL

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

3 participants

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

frankenbit/release-toolkit!1104

No description provided.