markup/asciidocext: Improve Asciidoctor integration

[jmooring]

Fixes an issue where improper attribute derivation from the page's relative permalink caused failures with outdir, imagesoutdir, and imagesdir when markup.asciidocext.workingFolderCurrent is enabled. The updated logic now correctly handles:

• Multi-byte characters
• Multilingual multi-host sites
• Site builds from a subdirectory
• Pages using ugly URLs

Supports diagram caching as implemented in v3.1.0 of the asciidoctor-diagram extension:

• Enables caching by default
• Sets default cache location to the compiled value of caches.misc.dir

Reduces duration of integration tests by:

• Generating GoAT diagrams instead of Ditaa diagrams
• Taking advantage of asciidoctor-diagram caching

Closes #9202
Closes #10183
Closes #10473
Closes #14160

[jmooring]

I'm not thrilled about the duration of the integration test (it's the second largest contributor to overall test time), but I don't want to pare it down either.

[Copilot]

Pull Request Overview

Copilot reviewed 12 out of 12 changed files in this pull request and generated no new comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jmooring]

most important test cases

Rebased on master. Integration test will only run when IsRealCI is true, and then only for three of the sub-tests.

if !slices.Contains([]string{"Test 4", "Test 8", "Test 16"}, tc.name) {
    continue
}

Our CI testing... 15s on Linux and 30s on Windows. Caching, plus GoAT instead of Ditaa diagrams, should make this much faster. Spinning up JRE for each diagram isn't cheap. GoAT support will be in the next asciidoctor-diagram release (PR merged yesterday).

[jmooring]

@bep Let's hold off on this for a day or two. There's a new upstream release that implements caching and GoAT diagrams.

[github-actions]

AI Detection Analysis 🔍

Confidence Score: 15%

Reasoning: The content in this pull request shows clear signs of human authorship, particularly due to the nuanced understanding of the Hugo framework, the AsciiDoctor ecosystem, and handling of multilingual site constructs and rendering edge cases. The implementation includes highly context-specific logic, consistent variable naming, detailed test cases, and explicit intention (e.g., skipping tests under CI or specific OS conditions). Most importantly, the test coverage illustrates deep knowledge of Hugo’s configuration and its interaction with advanced AsciiDoc features (e.g., diagram caching, multilingual URLs, diagram output directories), which currently would be difficult to fully automate using generative AI tools.

The writing in the description is polished yet technical, aligning with an experienced open source contributor familiar with Hugo and AsciiDoc internals. The highly structured test matrix with twenty complex test cases involving multilingual, uglyURLs, and baseURL edge cases, and the creation of content using idiomatic Go templating, are particularly difficult to produce via AI without overgeneralization or errors.

Key Indicators:

• Deep Hugo internals knowledge (e.g., .Conf.CacheDirMisc(), markup config nesting, language-specific baseURL handling, etc.)
• Manual, structured test scenarios across 20 permutations of multilingual and URL representations
• Precision use of Go standard libraries, TOML config placeholders, and integration with Ruby-based AsciiDoctor tooling (including use of gem CLI)
• Human-like formatting and in-line comments tailored to future maintainers and contributors
• Strategic skips based on platform and CI environment detection (in test logic)
• The PR exhibits design coherence and maintains Hugo's code quality and documentation standards, which is more typical of experienced maintainers than code-generating AI.

This body of work suggests deliberate engineering and domain-specific expertise far beyond current AI-generated code capabilities.

Key Indicators:

• Complex test scenarios tailored to niche requirements (multilingual, subdirs, URI encoding)
• Context-aware OS and environment handling with skip logic
• Advanced error messages and inline usage documentation
• High integration between multiple Hugo components (markup, config, testing, i18n)
• Lack of generic or boilerplate phrasing common in LLM output
• No signs of typical LLM limitations (hallucinated APIs, improper tagging, misunderstanding of Go syntax or Hugo config semantic model)

Conclusion: Human-authored with strong coherence and domain-specific reasoning.

✅ No strong indicators of AI generation detected

[jmooring]

Perhaps this is a hangover from the outage(s) a few hours ago, but integration tests fail every time:

failed to execute binary "npm" with args [install]: npm warn tar TAR_ENTRY_ERROR ENOSPC: no space left on device

[github-actions]

AI Detection Analysis 🔍

Confidence Score: 15%

Reasoning:
This pull request is highly detailed and contextually deep in domain-specific integration for Hugo's AsciiDoc support. It spans tests, conditional logic, diagrams, multilingual content, and caching—an extremely thorough and nuanced development effort. The content includes highly specific technical details such as which diagrams to use (e.g., GoAT vs. Ditaa), low-level file layout decisions, multilingual multilingualism with subdirectories and hosts, and deep knowledge of the AsciiDoctor diagram plugin implementation, including caching behavior and filesystem interactions. Automated tools may aid in generating some scaffolding, but the comprehensive nature of the implementation, test coverage, and interpretive decisions strongly suggests an experienced human developer.

Key Indicators:

• Thoughtful treatment of multiple edge cases and site configurations, such as multilingual multi-host deployments and URL structures ("uglyURLs").
• Deep integration across Hugo internal logic: cache directories, diagrams rendering, content pipeline, TOC extraction.
• Custom fallback logic, error handling, and nuanced CLI checks for external binaries and gems.
• Extensive and varied integration test scenarios, including validations for multilingual SVG outputs and exceptional path handling (e.g., UTF-8 paths like "Straßen/Frühling").
• Rich and coherent documentation in comment blocks and structuring, showing awareness of long-term maintainability.
• Custom test skipping logic depending on CI vs local, and OS-specific exclusions for Windows.

These reflect a level of contextual and experiential understanding not easily achievable with current AI content generators.

✅ No strong indicators of AI generation detected

[bep]

Confidence Score: 15%

There's only a 15% chance you're not human, according to OpenAI. Running on every PR force push doesn't make sense, though ... I'll see if I can fix that.

[jmooring]

There's only a 15% chance you're not human

Yeah, well, I'd argue the number is higher than that, but I appreciate its charitable assessment.

[jmooring]

@bep I could use another set of eyes on this. I cannot figure out why the CI runs continue to fail. The new integration test passes:

ok  	github.com/gohugoio/hugo/markup/asciidocext	53.224s

But the job dies later in the process.

[bep]

@jmooring I will have a look. We have had "space issues" before, which is the reason we currently do not run on Darwin. But the "free-disk-space" should work, and it's strange that it works in other PRs.

[bep]

I have tried to narrow it down in #14190 ... This I don't understand, it must be some issue on the GitHub side of the fence. I need to look at this with a fresh head tomorrow.

[bep]

OK, #14196 should fix it, will merge once it goes green.

[jmooring]

It's looks like there are still some space issues or something on Linux.

[bep]

OK, now this issue has spread to #14197 -- which I guess also gives me a hint about what the real issue is: Test binary sizes now crossing a breaking threshold. I will see if I can find a fix.

[jmooring]

This is ready for final review.

Integration test details:

• Generates GoAT diagrams instead of Ditaa diagrams
• Enables the caching mechanism introduced in v3.0.1 of the asciidoctor-diagram extension
• Runs only when IsRealCI is true
• Runs all subtests on Linux
• Runs 3 of the subtests on Windows

When testing locally on Linux with Ditaa diagrams, the benefits of diagram caching are obvious. When using GoAT diagrams, the benefit is less noticeable because generation is very fast. On Windows, the benefits of diagram caching for simple diagrams are negligible due to the overall slower performance of the Asciidoctor calls.

Current CI test duration:

• Linux: ~50s (runs all 20 subtests)
• Windows: ~30s (runs 3 of 20 subtests)