darnlink — self-healing Markdown links

Version updated for https://github.com/txemi/darnlink to version v0.24.0.

  • This action is used across all versions by ? repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

Action Summary

The darnlink GitHub Action automatically fixes broken links in Markdown files by repairing the paths of existing links or creating new ones with UUIDs if the targets have none. It supports cross-repo web links and is built to handle large trees of Markdown files efficiently. Users can use it directly from a repository without installation, making it easy to maintain documentation that evolves over time.

What’s Changed

D1-D3 of the v0.24.0 block. Five PRs, each with its own adversarial review round (Copilot out of quota for all of them — REGLA CERO). Every fix in this release verified with git tag --contains before writing the notes — the exact check that was missing when v0.22.0 credited two fixes it didn’t actually ship.

What a consumer sees

1. A dry-run’s +uuid X no longer shows a value you can copy. A freshly-minted uuid (no existing frontmatter) now reads +uuid <will be generated on write>; a reused uuid is unaffected. --write still writes the real, correct, matching uuid.

2. check --json’s output comes from ONE tree read instead of two. Byte-for-byte identical to v0.23.0 — verified, not assumed, with two side-by-side worktrees — only faster (~2.1× measured on a large repo in the fleet).

Nothing else changes behaviour for a consumer.

Fixed

  • An anchor comment no longer separates a link from its pandoc attribute block (#65). Plus a follow-up hardening found by adversarial review: a } inside a quoted attribute value was cutting the match short, and on --write that spliced the anchor comment into the attribute text itself — worse than #65’s own bug. A second, independently-drifted copy of the same regex (in ROBUST_LINK_RE) was silently making repair report zero findings for an already-anchored link whose target had moved; both now share one pattern.

  • A trailing space in a link’s destination no longer turns a file link into an unhealable CONFLICT (#67). Fixed at the one shared primitive; resolve_href stays untouched on purpose (the general whitespace question is #74’s), so the side effect is the stray space gets cleaned up when the link heals, not just stop being misdiagnosed.

  • A freshly-minted dry-run uuid is no longer printed as if it were final (#41). Also covers --create-readme’s own uuid-minting path, found by adversarial review as a second route to the same symptom.

  • check runs both its axes off ONE tree read instead of two (#87). plan_repairs/ plan_robustify gain an optional prescanned parameter (None by default — zero change for every other caller). Adversarial review found --only was silently ignored when combined with prescanned — a real gap, fixed with dedicated coverage before merge.

Security / hardening

  • write_text_keep_newlines now refuses a symlink path, as a live assertion. Verified first: every path reaching this function is already resolved by iter_markdown_files (Path.resolve(strict=True)), so the check is unreachable in current use — it protects a future caller that bypasses that resolution, converting a silent alias-write into an immediate crash.

Known issues (filed, not fixed here)

  • A hardlink to an indexed .md is indexed twice — same symptom as #85, different mechanism (Path.resolve() doesn’t collapse hardlinks). Verified live with os.link(). #91.

Verified before release

  • uv run pytest -q → 427 passed
  • check --json, byte-for-byte identical between v0.23.0 and this release on the same tree
  • Every one of the 5 commits confirmed inside this tag via git tag --contains

Full detail in CHANGELOG.md.