actions/checkout leaves an Authorization header in .git/config unless persist-credentials: false is set — write-capable while this instance's Actions default token permission is permissive, where it stands after a Restricted trial was reverted on 2026-08-05 (#748). Every workflow except review-verdict.yml omitted it, so a job handed that credential to every script it ran, not merely to the checkout.
15 of the repo's 16 checkouts now set it: docker-build.yml 8, pr-checks.yml 5, dependency-scan.yml 1, plus review-verdict.yml's pre-existing 1.
The ordering, which is the point
#743 deferred this rather than doing it blind. Five steps ran git fetch ... origin "$base_ref" || true and fed the result into changed-file skip logic, so a credential regression would not have reddened anything — it yields an empty changed set the skip logic reads as "nothing changed". The masking comes out first, then the flag goes on.
--depth came off those five fetches in the same step. All five run in fetch-depth: 0 jobs, and git fetch --depth=N grafts a complete clone shallow — measured on a 150-commit repo branched 130 back: .git/shallow appears and git diff origin/main...HEAD fails fatal: no merge base; the same fetch without --depth returns the correct file list.
Loud is not the same as blocking.api-docs, format and decisions lifecycle gate on this diff and redden on a failed fetch. docs-reminder is a non-blocking nudge whose advisory red would join the combined status the merge gate reads, so its parity step joins its two siblings in carrying continue-on-error: true.
Population, swept by shape not by the issue's file list
The issue names six docker-build.yml checkouts and two fetches. Sweeping git ls-files for the shape found eight checkouts there, five more in pr-checks.yml, one in dependency-scan.yml, and five masked fetches plus three masked git diffs. Recorded on the issue so acceptance is judged against what shipped.
ci-image.yml's one checkout is excluded — not because it is safe (its unfiltered push trigger is head-resolved in the sense that matters) but because any commit touching that file re-points ci-image-pin's expected at itself and reddens a blocking job. It rides with #744.
Filed, not fixed here
#835 — the machine-checked guard for this convention. Blocked by #744 so it can ship with no exemption list.
#836 — :latest images ship InformationalVersion 0.0.0-<sha> (measured live: 0.0.0-5a7a4ed7). Pre-existing, found while verifying the --depth behaviour: ci-detect-docs-only.sh's push-path --depth=2 grafts the build job's fetch-depth: 0 checkout, so git describe returns nothing.
Verification
scripts/tests — 956 passed, 2 skipped (re-run after every review round and after each rebase).
decisions_validate.py, build_decisions_catalog.py --check, check-kickoff-guard.sh, check-doc-narrative.py — all clean.
Shallow-graft mechanism reproduced synthetically; 0.0.0-5a7a4ed7 measured on the live ersatztv-test container.
Nine cold review rounds, alternating two independent reviewers (one a different model family). Rounds 1-8 blocked; every blocking finding was documentation accuracy, never the workflow code, which was independently re-derived nine times. Round 9: clean, no findings.
Done-when box 3 ("demonstrated non-empty changed-file set, not merely a green run") is satisfied by this PR's own docs-reminder / api-docs logs, which run the new unmasked code against real changed files.
Closes #746.
`actions/checkout` leaves an `Authorization` header in `.git/config` unless `persist-credentials: false` is set — write-capable while this instance's Actions default token permission is `permissive`, where it stands after a Restricted trial was reverted on 2026-08-05 (#748). Every workflow except `review-verdict.yml` omitted it, so a job handed that credential to every script it ran, not merely to the checkout.
**15 of the repo's 16 checkouts now set it**: `docker-build.yml` 8, `pr-checks.yml` 5, `dependency-scan.yml` 1, plus `review-verdict.yml`'s pre-existing 1.
## The ordering, which is the point
#743 deferred this rather than doing it blind. Five steps ran `git fetch ... origin "$base_ref" || true` and fed the result into changed-file skip logic, so a credential regression would not have reddened anything — it yields an empty changed set the skip logic reads as "nothing changed". **The masking comes out first, then the flag goes on.**
`--depth` came off those five fetches in the same step. All five run in `fetch-depth: 0` jobs, and `git fetch --depth=N` grafts a *complete* clone shallow — measured on a 150-commit repo branched 130 back: `.git/shallow` appears and `git diff origin/main...HEAD` fails `fatal: no merge base`; the same fetch without `--depth` returns the correct file list.
**Loud is not the same as blocking.** `api-docs`, `format` and `decisions lifecycle` gate on this diff and redden on a failed fetch. `docs-reminder` is a non-blocking nudge whose advisory red would join the combined status the merge gate reads, so its parity step joins its two siblings in carrying `continue-on-error: true`.
## Population, swept by shape not by the issue's file list
The issue names six `docker-build.yml` checkouts and two fetches. Sweeping `git ls-files` for the shape found **eight** checkouts there, five more in `pr-checks.yml`, one in `dependency-scan.yml`, and **five** masked fetches plus **three** masked `git diff`s. Recorded on the issue so acceptance is judged against what shipped.
`ci-image.yml`'s one checkout is excluded — not because it is safe (its unfiltered `push` trigger is head-resolved in the sense that matters) but because any commit touching that file re-points `ci-image-pin`'s `expected` at itself and reddens a blocking job. It rides with #744.
## Filed, not fixed here
- **#835** — the machine-checked guard for this convention. Blocked by #744 so it can ship with no exemption list.
- **#836** — `:latest` images ship `InformationalVersion 0.0.0-<sha>` (measured live: `0.0.0-5a7a4ed7`). Pre-existing, found while verifying the `--depth` behaviour: `ci-detect-docs-only.sh`'s push-path `--depth=2` grafts the `build` job's `fetch-depth: 0` checkout, so `git describe` returns nothing.
## Verification
- `scripts/tests` — 956 passed, 2 skipped (re-run after every review round and after each rebase).
- `decisions_validate.py`, `build_decisions_catalog.py --check`, `check-kickoff-guard.sh`, `check-doc-narrative.py` — all clean.
- Shallow-graft mechanism reproduced synthetically; `0.0.0-5a7a4ed7` measured on the live `ersatztv-test` container.
- **Nine cold review rounds**, alternating two independent reviewers (one a different model family). Rounds 1-8 blocked; every blocking finding was documentation accuracy, never the workflow code, which was independently re-derived nine times. Round 9: clean, no findings.
- Done-when box 3 ("demonstrated non-empty changed-file set, not merely a green run") is satisfied by this PR's own `docs-reminder` / `api-docs` logs, which run the new unmasked code against real changed files.
`actions/checkout` leaves an `Authorization` header in `.git/config` unless
`persist-credentials: false` is set — write-capable while this instance's Actions
default token permission is `permissive`, where it stands after a Restricted trial
was reverted on 2026-08-05 (#748). Every workflow here except `review-verdict.yml`
omitted it, so a job handed that credential to every script it ran, not merely to the
checkout. 15 of the repo's 16 checkouts now set it.
Setting the flag first would have been unsafe, which is why #743 deferred it rather
than doing it blind: five steps ran `git fetch ... origin "$base_ref" || true` and
fed the result into changed-file skip logic, so a credential regression would have
presented as an empty changed set on the three steps that also masked `git diff`, and
a possibly-absent `origin/<base>` handed to a fail-open Python validator on the other
two — a silently self-skipping check on a green job either way.
So the masking comes out first, and only then does the flag go on. `git diff`'s own
`2>/dev/null || true` goes with it on the three sites that had one.
`--depth` came off those five fetches in the same step, and that is the part the cold
reviews converged on. All five run in jobs that check out `fetch-depth: 0`, and
`git fetch --depth=N` grafts a *complete* clone shallow — measured on a 150-commit
repo branched 130 back: `.git/shallow` appears and `git diff origin/main...HEAD` fails
`fatal: no merge base`, while the same fetch without `--depth` returns the correct
file list. Under `|| true` that was a silent skip; unmasked it would have reddened
`api-docs` and `format` on any branch not rebased inside the depth window, and left
`decisions lifecycle` green with its diff-dependent checks silently skipped.
Loud is not the same as blocking. `api-docs`, `format` and `decisions lifecycle` gate
on this diff and now redden on a failed fetch. `docs-reminder` does not — it is a
non-blocking nudge whose advisory red would still join the combined status the merge
gate reads, which is why its `Set up Python` and doc-narrative steps already carried
`continue-on-error: true`. The parity nudge now carries it for the same reason, so all
three of its non-checkout steps have it and the `::error::` reaches the log without
becoming a merge block. `Checkout` stays blocking.
The population is wider than #746's file list. Sweeping the shape rather than the
named file found 8 checkouts in `docker-build.yml` (not the six #743 recorded), 5 in
`pr-checks.yml` and 1 in `dependency-scan.yml`, and three masked fetches in
`pr-checks.yml` beside the two in `docker-build.yml`. That widening is recorded on the
issue so acceptance is judged against what shipped. No job that checks out pushes with
git, and the repo is public with anonymous `git-upload-pack` answering, so nothing
loses a capability it was using. Renovate does push branches, but `renovate.yml` has
no checkout step and carries its own `RENOVATE_TOKEN`.
Two further masked fetches, in `scripts/ci-detect-docs-only.sh` and
`scripts/ci-detect-already-validated.sh`, keep their masking on purpose: in both, a
failed fetch can only leave the base revision unresolvable, and an unresolvable base
emits `false` — run the full matrix. Neither can turn a failed fetch into a skip.
Their `--depth` is a different matter and is NOT cleared here: the push-path
`--depth=2` grafts the `build` job's `fetch-depth: 0` checkout, which is why `:latest`
images ship `InformationalVersion 0.0.0-<sha>` — measured on the live test container
as `0.0.0-5a7a4ed7` and filed as #836. Four stale statements of that script's
shallow-consumer set (two in docs/ci-cd.md, one in the script itself, one in
`ci.docs-only-detect-shallow-safe`) said `fetch-depth: 1` and named two of the three
consumers; they are corrected here, because this change states the same fact
again and would otherwise leave the corpus disagreeing with itself.
`ci-image.yml`'s one checkout is the exception, and not because it is safe: its
unfiltered `push` trigger is head-resolved in the sense that matters. It is excluded
because any commit touching that file re-points `ci-image-pin`'s `expected` at itself
and reddens a blocking job, and because the same trigger would publish a toolchain
image from the editing branch. It rides with #744.
No machine-checked guard holds the convention yet: writing one now means a hardcoded
`ci-image.yml` exemption whose only justification is a temporary condition in another
issue, and that entry would outlive its reason silently. Filed as #835, blocked by
#744, to be written with no exemption list at all.
fixes#746
Decisions-Edit: yes
Nine cold rounds, two independent reviewers alternating (one a different model family). Rounds 1-8 blocked; every blocking finding was documentation accuracy, never the workflow code, which was independently re-derived nine times. Round 9 clean, no findings. 956 pytest passed; decisions validator, catalog check and kickoff guard clean; shallow-graft mechanism reproduced synthetically; Done-when box 3 satisfied by this PR's own docs-reminder log showing a non-empty 9-file changed set from the unmasked credential-free fetch on a real runner.
Review-verdict: MERGEABLE @ 4196a05
Nine cold rounds, two independent reviewers alternating (one a different model family). Rounds 1-8 blocked; every blocking finding was documentation accuracy, never the workflow code, which was independently re-derived nine times. Round 9 clean, no findings. 956 pytest passed; decisions validator, catalog check and kickoff guard clean; shallow-graft mechanism reproduced synthetically; Done-when box 3 satisfied by this PR's own docs-reminder log showing a non-empty 9-file changed set from the unmasked credential-free fetch on a real runner.
timothy
merged commit f2551b778e into main2026-08-26 20:50:57 +02:00
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Closes #746.
actions/checkoutleaves anAuthorizationheader in.git/configunlesspersist-credentials: falseis set — write-capable while this instance's Actions default token permission ispermissive, where it stands after a Restricted trial was reverted on 2026-08-05 (#748). Every workflow exceptreview-verdict.ymlomitted it, so a job handed that credential to every script it ran, not merely to the checkout.15 of the repo's 16 checkouts now set it:
docker-build.yml8,pr-checks.yml5,dependency-scan.yml1, plusreview-verdict.yml's pre-existing 1.The ordering, which is the point
#743 deferred this rather than doing it blind. Five steps ran
git fetch ... origin "$base_ref" || trueand fed the result into changed-file skip logic, so a credential regression would not have reddened anything — it yields an empty changed set the skip logic reads as "nothing changed". The masking comes out first, then the flag goes on.--depthcame off those five fetches in the same step. All five run infetch-depth: 0jobs, andgit fetch --depth=Ngrafts a complete clone shallow — measured on a 150-commit repo branched 130 back:.git/shallowappears andgit diff origin/main...HEADfailsfatal: no merge base; the same fetch without--depthreturns the correct file list.Loud is not the same as blocking.
api-docs,formatanddecisions lifecyclegate on this diff and redden on a failed fetch.docs-reminderis a non-blocking nudge whose advisory red would join the combined status the merge gate reads, so its parity step joins its two siblings in carryingcontinue-on-error: true.Population, swept by shape not by the issue's file list
The issue names six
docker-build.ymlcheckouts and two fetches. Sweepinggit ls-filesfor the shape found eight checkouts there, five more inpr-checks.yml, one independency-scan.yml, and five masked fetches plus three maskedgit diffs. Recorded on the issue so acceptance is judged against what shipped.ci-image.yml's one checkout is excluded — not because it is safe (its unfilteredpushtrigger is head-resolved in the sense that matters) but because any commit touching that file re-pointsci-image-pin'sexpectedat itself and reddens a blocking job. It rides with #744.Filed, not fixed here
:latestimages shipInformationalVersion 0.0.0-<sha>(measured live:0.0.0-5a7a4ed7). Pre-existing, found while verifying the--depthbehaviour:ci-detect-docs-only.sh's push-path--depth=2grafts thebuildjob'sfetch-depth: 0checkout, sogit describereturns nothing.Verification
scripts/tests— 956 passed, 2 skipped (re-run after every review round and after each rebase).decisions_validate.py,build_decisions_catalog.py --check,check-kickoff-guard.sh,check-doc-narrative.py— all clean.0.0.0-5a7a4ed7measured on the liveersatztv-testcontainer.docs-reminder/api-docslogs, which run the new unmasked code against real changed files.Review-verdict: MERGEABLE @
4196a05Nine cold rounds, two independent reviewers alternating (one a different model family). Rounds 1-8 blocked; every blocking finding was documentation accuracy, never the workflow code, which was independently re-derived nine times. Round 9 clean, no findings. 956 pytest passed; decisions validator, catalog check and kickoff guard clean; shallow-graft mechanism reproduced synthetically; Done-when box 3 satisfied by this PR's own docs-reminder log showing a non-empty 9-file changed set from the unmasked credential-free fetch on a real runner.