Backtestshugorun-20260930-140149#15234
templates.Defer: identical block bodies in different output formats share one escaping mode, chosen non-deterministically per build
Our patch was judged better than the merged fix.
In this judgment, Patch B = our agent, Patch A = the merged fix.
Both patches fix the root cause the same way. Each adds the owner's IsPlainText flag to the hashed input for the deferred ID in handleWith. HTML and plain-text owners with byte-identical bodies therefore no longer collide in addDeferredTemplate. Same-mode sharing, including the documented cache-by-key/content behaviour across layouts of one format, is preserved. So correctness is essentially equal.
Patch B has two small advantages on the fix itself:
- It only changes the hash input for plain-text owners, so existing HTML deferred IDs stay byte-identical.
- It avoids relying on fmt being imported.
Patch A's fmt.Sprintf("%t:%s") is fine but changes every ID.
The tests separate the two patches:
- The bug is non-deterministic: map iteration order picks the winner, with roughly a 60/40 split. Patch A's single-build test would pass by chance a sizable fraction of the time on unfixed code, so it is a weak regression test.
- Patch B builds the reproduction 20 times, which makes a pre-fix pass practically impossible.
- Patch B also adds a test that same-format identical bodies with one key still execute once. That directly guards the constraint the issue says any fix must preserve.
Scope is minimal in both. Patch B wins on test quality.
The judge saw only the two patches, labelled A and B in random order, with issue links and version numbers removed. The reasoning above is shown verbatim.
Solve
27m 2s20 turns
Wall-clock time in the container
Full test suite
No new failures
3,736 passed · 99 skipped · 3,837 total
Contamination audit
Clean
No sign the agent saw the real fix
Patch size
63vs 29 merged
Changed lines, ours vs merged
The issue
View on GitHubWhat version of Hugo are you using (hugo version)?
<pre>
$ hugo version
hugo v0.164.0-ce2470e7012b5ab5fc4e10ebe4027e9f8d9e00dc+extended windows/amd64 BuildDate=2026-07-06T16:39:30Z VendorInfo=gohugoio
</pre>
Does this issue reproduce with the latest release?
Yes.
hugo v0.165.0-76a5e1880ab46688155b02e99bab9be2a6134492+extended windows/amd64 BuildDate=2026-08-12T14:26:28Z VendorInfo=gohugoio
Over one run of 40 consecutive builds of the unchanged reproduction site below with that binary, 28 builds published the wrong value in public/index.html and 12 published the wrong value in public/index.txt. All 40 exited 0, and none printed a warning or an error.
Summary
When two templates belonging to different output formats contain byte-identical templates.Defer block bodies, both blocks do execute, each with its own correct output-format context. What they share is the escaping applied inside the block, and which of the two escaping modes survives changes from build to build of an unchanged site. On every build exactly one of the two outputs is therefore wrong: either an HTML page publishes a value raw that it escapes correctly a few characters earlier outside the block, or a plain-text page receives HTML entities.
Reproduction
Three files in an empty directory. No content, no modules, no external tooling.
hugo.toml:
baseURL = 'https://example.org/'
title = 'defer probe'
disableKinds = ['page', 'section', 'taxonomy', 'term', 'RSS', 'sitemap', 'robotsTXT']
[outputFormats.probe]
mediaType = 'text/plain'
baseName = 'index'
isPlainText = true
[outputs]
home = ['html', 'probe']
layouts/home.html:
{{ $v := "<b>&" }}OUT[{{ $v }}]{{ with (templates.Defer (dict "key" "K" "data" (dict "v" $v))) }}IN[{{ .v }}]{{ end }}
layouts/home.probe.txt, byte-identical to layouts/home.html (both md5 c921197a9c8637852e38d94a0352359d):
{{ $v := "<b>&" }}OUT[{{ $v }}]{{ with (templates.Defer (dict "key" "K" "data" (dict "v" $v))) }}IN[{{ .v }}]{{ end }}
OUT[...] interpolates the value outside the deferred block, IN[...] interpolates the same value inside it. Then run hugo several times without changing anything, and look at public/index.html and public/index.txt after each run.
What happens
Two consecutive runs of the same command on the unchanged site, pasted verbatim:
$ hugo --logLevel info
Start building sites …
hugo v0.164.0-ce2470e7012b5ab5fc4e10ebe4027e9f8d9e00dc+extended windows/amd64 BuildDate=2026-07-06T16:39:30Z VendorInfo=gohugoio
INFO build: step process substep collect files 0 files_total 0 pagesources_total 0 resourcesources_total 0 duration 0s
INFO build: step process duration 0s
INFO build: step assemble duration 0s
INFO build: step render substep pages site en outputFormat html duration 1.3703ms
INFO build: step render substep pages site en outputFormat probe duration 509.5µs
INFO build: step render pages 2 content 0 duration 1.8798ms
INFO build: step render deferred count 2 duration 1.7766ms
INFO build: step postProcess duration 0s
INFO build: duration 4.171ms
│ EN
──────────────────┼────
Pages │ 2
Paginator pages │ 0
Non-page files │ 0
Static files │ 0
Processed images │ 0
Aliases │ 0
Cleaned │ 0
Total in 12 ms
--- public/index.html ---
OUT[<b>&]IN[<b>&]
--- public/index.txt ---
OUT[<b>&]IN[<b>&]
$ hugo --logLevel info
Start building sites …
hugo v0.164.0-ce2470e7012b5ab5fc4e10ebe4027e9f8d9e00dc+extended windows/amd64 BuildDate=2026-07-06T16:39:30Z VendorInfo=gohugoio
INFO build: step process substep collect files 0 files_total 0 pagesources_total 0 resourcesources_total 0 duration 513.8µs
INFO build: step process duration 513.8µs
INFO build: step assemble duration 0s
INFO build: step render substep pages site en outputFormat html duration 517.9µs
INFO build: step render substep pages site en outputFormat probe duration 517.2µs
INFO build: step render pages 2 content 0 duration 1.0351ms
INFO build: step render deferred count 2 duration 2.0736ms
INFO build: step postProcess duration 0s
INFO build: duration 3.6225ms
│ EN
──────────────────┼────
Pages │ 2
Paginator pages │ 0
Non-page files │ 0
Static files │ 0
Processed images │ 0
Aliases │ 0
Cleaned │ 0
Total in 12 ms
--- public/index.html ---
OUT[<b>&]IN[<b>&]
--- public/index.txt ---
OUT[<b>&]IN[<b>&]
In the first run the HTML page publishes <b>& raw inside the block, a few characters after escaping the same value correctly outside it. In the second run the plain-text page receives <b>&. Nothing changed between the two runs.
Quantified over builds of the unchanged site, each build a separate one-shot hugo invocation:
| Hugo version | builds | index.html published raw <b>& inside the block | index.txt published <b>& inside the block |
|---|---|---|---|
| v0.164.0, fresh copy of the site per build | 40 | 25 | 15 |
| v0.164.0, fresh copy of the site per build, second run | 40 | 23 | 17 |
| v0.165.0, fresh copy of the site per build | 40 | 26 | 14 |
| v0.165.0, fresh copy of the site per build, second run | 40 | 28 | 12 |
| v0.164.0, repeated builds in one directory | 12 | 8 | 4 |
| v0.165.0, repeated builds in one directory | 8 | 5 | 3 |
In all 180 of those builds the two output formats carried the same string inside the block, so exactly one of the two published files was wrong every time; outside the block the value was escaped correctly for its format in every build.
What I expected
Each format's deferred block rendered under that format's own escaping — IN[<b>&] in public/index.html and IN[<b>&] in public/index.txt — and the same result on every build of an unchanged site.
The documentation says "The output format, site, and language will be the same, even if the execution is deferred." The output-format context does hold up, as the identity-token control below shows; it is the escaping regime that does not follow the output format.
Controls
All of these were run at both v0.164.0 and v0.165.0.
The block really does execute once per output format. Adding {{ warnf "DEFER-EXEC-%d" math.Counter }} to the start of the still byte-identical block body logs two executions in every build:
WARN DEFER-EXEC-1
WARN DEFER-EXEC-2
INFO build: step render deferred count 2 duration 1.7497ms
--- public/index.html ---
OUT[<b>&]IN[<b>&]
--- public/index.txt ---
OUT[<b>&]IN[<b>&]
So this is not one rendering substituted into two files; it is two renderings sharing one escaping mode.
Different key values do not separate them. With "KH" in layouts/home.html and "KT" in layouts/home.probe.txt, the block bodies still byte-identical, the escaping stayed shared and kept flipping: 36 builds, 22 with the HTML page wrong and 14 with the plain-text page wrong.
Different block body text does separate them, stably. Changing the body to INH[{{ .v }}] in the HTML template and INT[{{ .v }}] in the plain-text template gives the correct per-format result in 36 of 36 builds, with no flip: INH[<b>&] in public/index.html and INT[<b>&] in public/index.txt.
The per-format context inside the block is correct. With the body still byte-identical in both templates, but carrying an output-format identity token next to the value:
{{ $v := "<b>&" }}OUT[{{ $v }}]{{ with (templates.Defer (dict "key" "K" "data" (dict "v" $v))) }}IN[{{ (index page.AlternativeOutputFormats 0).Name }}|{{ .v }}]{{ end }}
With exactly two formats configured the single alternative names the other one, so IN[probe|...] marks the execution that produced the HTML page and IN[html|...] the one that produced the plain-text page. That token was correct in 36 of 36 builds, while the escaping of the value beside it in the very same body was shared and flipped across those builds:
public/index.html OUT[<b>&]IN[probe|<b>&]
public/index.txt OUT[<b>&]IN[html|<b>&]
public/index.html OUT[<b>&]IN[probe|<b>&]
public/index.txt OUT[<b>&]IN[html|<b>&]
Where this appears to come from
Offered as a lead from reading the source, not as an instrumented finding.
tpl/tplimpl/templatetransform.go, handleWith (v0.164.0, function at line 409) registers the extracted block under a hash of the block body alone, at lines 450-459:
inner := withNode.List.CopyList()
s := inner.String()
...
innerHash := hashing.XxHashFromStringHexEncoded(s)
deferredID := tpl.HugoDeferredTemplatePrefix + innerHash
...
c.deferNodes[deferredID] = inner
Nothing about the owning template, its output format or its media type enters that identifier, so two templates in different output formats with the same body produce the same deferredID.
tpl/tplimpl/templatestore.go, addDeferredTemplate (v0.164.0, function at line 916) then keeps only the first registration under that name, and the first owner fixes the parse namespace for whoever comes later:
if _, found := t.templatesByPath.Get(name); found {
return nil
}
...
if owner.D.IsPlainText {
prototype := t.tns.parseText
...
} else {
prototype := t.tns.parseHTML
...
}
That is exactly the escaping regime the two output formats disagree about, and the loser silently inherits the winner's.
Which owner gets there first looks order-dependent. transformTemplates (function at line 1916) reaches the registrations by ranging s.templates() at line 1937, and that iterator's inner loop at line 1784 ranges a map[nodeKey]*TemplInfo — the value type of treeMain, declared at line 447 — whose iteration order Go randomizes per run. I found no sort on that path.
One contrast worth noting: the sibling transform in the same file, handleWithPartial (function at line 324), hashes the owning template's name into its identifier at line 335 — hashing.XxHashFromStringHexEncoded(c.t.Name() + withNodeInnerString) — while the older defer path does not.
I checked the implicated files at v0.164.0, v0.165.0 and master 7b5199fdeff7d7682e9f979f468d55bc7b67b934: tpl/tplimpl/templatestore.go, tpl/templates/templates.go and tpl/template.go are byte-identical at all three, and tpl/tplimpl/templatetransform.go differs at master only outside this path — the handleWith lines quoted above are unchanged there, at 405-414.
A constraint on any fix
Hashing the owning template's name into the deferred identifier, the way handleWithPartial does, would break something the documentation promises for key: "This will, combined with a hash of the template content, be used as a cache key." Two different layouts of the same output format with identical bodies and one shared key share a single execution today, and that is worth keeping — at both versions, on a variant of the site above reduced to one output format and with the page kind enabled plus one content file, so that single.html renders at all, I measured a single DEFER-EXEC-1 for a home.html and a single.html written that way, with both pages publishing the shared result.
Separating registrations by the property that actually differs here — the owner's IsPlainText, or its output format — would fix the escaping while leaving that documented same-format sharing untouched.
Related, not duplicates
- #7774, fixed by #8462 ("Fixes #7774", merged 2021-04-23): per-build nondeterminism in output-format template lookup for shortcodes. Same failure shape, different subsystem, and it predates
templates.Defer. - #13236: two empty deferred blocks in the same template collided on the identifier
__hdeferred/cafc7706cee4572band panicked; closed as completed. That is the same body-only identifier, in a different situation — the collision there is inside one template, and the cross-output-format escaping consequence does not come up.
I did not find an existing report covering byte-identical templates.Defer bodies across output formats.
What I did not establish
- Which registration wins in any individual build. I inferred the winner from the escaping in the published output; I did not instrument the running binary to observe the registration.
- Platform generality. Everything above was measured on one machine, Windows 11, amd64 extended builds. Whether the split differs on another platform or core count is unmeasured.
- Two non-plain-text formats sharing a body. The escaping observable needs an
isPlainTextformat paired with an HTML one, so two HTML-ish formats sharing a body were not tested and the reach across format combinations is unknown.
The two patches
tpl/templates/defer_integration_test.goindex 4f67898bc..b5dde868e 100644@@ -367,3 +367,61 @@ Home. b.Assert(err, qt.Not(qt.IsNil)) b.Assert(err.Error(), qt.Contains, "templates.Defer cannot be used inside a partialCached partial") }++// Identical deferred block bodies in templates for different output formats+// (one HTML, one plain text) must not share a registration, as that would+// apply the wrong escaping in one of them.+func TestDeferIdenticalBodiesAcrossOutputFormats(t *testing.T) {+ t.Parallel()++ files := `+-- hugo.toml --+baseURL = 'https://example.org/'+disableKinds = ['page', 'section', 'taxonomy', 'term', 'rss', 'sitemap', 'robotsTXT']+[outputFormats.probe]+mediaType = 'text/plain'+baseName = 'index'+isPlainText = true+[outputs]+home = ['html', 'probe']+-- layouts/home.html --+{{ $v := "<b>&" }}OUT[{{ $v }}]{{ with (templates.Defer (dict "key" "K" "data" (dict "v" $v))) }}IN[{{ .v }}]{{ end }}+-- layouts/home.probe.txt --+{{ $v := "<b>&" }}OUT[{{ $v }}]{{ with (templates.Defer (dict "key" "K" "data" (dict "v" $v))) }}IN[{{ .v }}]{{ end }}+`++ // Which registration won depended on map iteration order, so build several times.+ for range 20 {+ b := hugolib.Test(t, files)++ b.AssertFileContent("public/index.html", "OUT[<b>&]IN[<b>&]")+ b.AssertFileContent("public/index.txt", "OUT[<b>&]IN[<b>&]")+ }+}++// Identical deferred block bodies with the same key in the same output format+// share a registration and execute once.+func TestDeferIdenticalBodiesSameOutputFormatExecuteOnce(t *testing.T) {+ t.Parallel()++ files := `+-- hugo.toml --+baseURL = 'https://example.org/'+disableKinds = ['section', 'taxonomy', 'term', 'rss', 'sitemap', 'robotsTXT', '404']+-- content/p1.md --+---+title: "P1"+---+-- layouts/home.html --+Home|{{ with (templates.Defer (dict "key" "K")) }}EXEC-{{ math.Counter }}{{ end }}|+-- layouts/page.html --+Page|{{ with (templates.Defer (dict "key" "K")) }}EXEC-{{ math.Counter }}{{ end }}|+`++ for range 5 {+ b := hugolib.Test(t, files)++ b.AssertFileContent("public/index.html", "Home|EXEC-1|")+ b.AssertFileContent("public/p1/index.html", "Page|EXEC-1|")+ }+}tpl/tplimpl/templatetransform.goindex ab42412eb..f5fd7a4b2 100644@@ -408,6 +408,11 @@ func (c *templateTransformContext) handleWith(withNode *parse.WithNode) { c.err = errors.New("resources.PostProcess cannot be used in a deferred template") return }+ if c.t.D.IsPlainText {+ // Text and HTML templates are escaped differently,+ // so they cannot share the same deferred template.+ s = "text:" + s+ } innerHash := hashing.XxHashFromStringHexEncoded(s) deferredID := tpl.HugoDeferredTemplatePrefix + innerHash
tpl/templates/defer_integration_test.goindex 4f67898bc..46c126ac1 100644@@ -323,6 +323,30 @@ End. b.AssertFileContent("public/index.html", "Home.", "Defer 1", "Defer 2", "Defer 3", "End.") } +// See issue 15234.+func TestDeferDifferentOutputFormats(t *testing.T) {+ t.Parallel()++ files := `+-- hugo.toml --+[outputFormats.probe]+mediaType = "text/plain"+baseName = "index"+isPlainText = true+[outputs]+home = ["html", "probe"]+-- layouts/home.html --+{{ $v := "<b>&" }}OUT[{{ $v }}]{{ with (templates.Defer (dict "key" "K" "data" (dict "v" $v))) }}IN[{{ .v }}]{{ end }}+-- layouts/home.probe.txt --+{{ $v := "<b>&" }}OUT[{{ $v }}]{{ with (templates.Defer (dict "key" "K" "data" (dict "v" $v))) }}IN[{{ .v }}]{{ end }}+`++ b := hugolib.Test(t, files)++ b.AssertFileContent("public/index.html", "OUT[<b>&]IN[<b>&]")+ b.AssertFileContent("public/index.txt", "OUT[<b>&]IN[<b>&]")+}+ // See issue 13492. func TestDeferInsidePartialCachedShouldFail(t *testing.T) { t.Parallel()tpl/tplimpl/templatetransform.goindex ab42412eb..dcb59abe7 100644@@ -408,7 +408,10 @@ func (c *templateTransformContext) handleWith(withNode *parse.WithNode) { c.err = errors.New("resources.PostProcess cannot be used in a deferred template") return }- innerHash := hashing.XxHashFromStringHexEncoded(s)+ // The deferred template is parsed using the owner's escaping mode. Keep+ // identical bodies shared within that mode, but don't let HTML and plain+ // text templates register the same deferred template.+ innerHash := hashing.XxHashFromStringHexEncoded(fmt.Sprintf("%t:%s", c.t.D.IsPlainText, s)) deferredID := tpl.HugoDeferredTemplatePrefix + innerHash c.deferNodes[deferredID] = inner
What the agent reported
I fixed the bug. HTML and plain-text templates with byte-identical templates.Defer block bodies now each get their own escaping, on every build. I couldn't use the required ./bt-test gate: it failed before running any tests because it tries to download Go modules and the network is blocked. So all test results below come from go test against the pre-filled module cache at /bt/cache/go-mod (with GOPROXY=off).
What was wrong: handleWith in tpl/tplimpl/templatetransform.go named each deferred block from a hash of the block body only. So home.html and home.probe.txt produced the same name. addDeferredTemplate keeps only the first registration under a name and uses that owner's IsPlainText to decide between text and HTML escaping. Which owner came first depended on Go's random map ordering, so the other output format silently got the wrong escaping. Your diagnosis was right.
What I changed: one production edit, in tpl/tplimpl/templatetransform.go (handleWith). When the owning template is plain text, "text:" is mixed into the string before hashing. HTML and plain-text owners now always get separate registrations. Two templates in the same output format with identical bodies and the same key still share one registration and run once, as the docs promise. HTML block names are exactly what they were before.
Tests (added to tpl/templates/defer_integration_test.go):
TestDeferIdenticalBodiesAcrossOutputFormatsis your reproduction site, built 20 times. It checks thatpublic/index.htmlcontainsOUT[<b>&]IN[<b>&]andpublic/index.txtcontainsOUT[<b>&]IN[<b>&]. It failed before the fix (the HTML page got a rawIN[<b>&]) and passes after.TestDeferIdenticalBodiesSameOutputFormatExecuteOncecovershome.htmlandpage.htmlwith identical bodies and the same key. It checks that both pages showEXEC-1, meaning the shared block ran once. It passes both before and after: it protects the documented sharing rather than catching this bug.- All 18
TestDefer*tests pass, and so do all./tpl/...packages excepttpl/tplimpl, where one test fails. - Five tests fail with and without my change, so the change didn't cause them:
TestVimeoShortcodeintpl/tplimpl, andTestModulePrivateRepoplus threeTestGitInfoFromGitModule*tests inhugolib. The rest ofhugolibpasses.
A reviewer agent confirmed three things. The plain-text flag the fix reads is the same one addDeferredTemplate uses to choose escaping. No other code depends on the exact form of the block name. The diff touches nothing else.
AGENTS.md and CLAUDE.md were already deleted in the working tree before I started, and I left them that way. Nothing is committed.
What the report asked for:
- Each format's block uses its own escaping, the same on every build: done, and the first test shows it.
- Keep same-format sharing (identical body and key run once): kept, and the second test checks it.
- Separate by
IsPlainText, not by template name: done as you suggested. I didn't add the owner's name the wayhandleWithPartialdoes, because that would break the sharing in item 2. - Two non-plain-text formats sharing a body: they still share one registration. That is correct, because both use HTML escaping, and the fix only separates by the plain-text flag.
- Which registration won in a given build: no longer relevant, since the two can't collide any more.
The reviewer also flagged a separate, older issue I left alone: in handleWith's sibling, handleWithPartial (templatetransform.go around lines 305–316), deferred blocks inside a partial-decorator body seem never to be registered.