Repository navigation
Releases
v0.4.0: the live badge
The badge went live, and the registry enrolls itself. A static plaque goes stale the moment a project changes its skills, so listed projects no longer wear a file committed to their own repo: their badge and plaque are served from the registry site on GitHub Pages (dongdarealest.github.io/skilldoctor), generated from the latest verification. Enrollment is one paste of a personal address that carries the wearer's own GitHub owner and repo (the registry page opens with a small generator that builds the snippet), and a scheduled workflow does the rest each day: it searches GitHub code search for the badge address, enrolls the repos it finds, re-audits every listed repo from a fresh clone, and republishes the registry. Keep passing and the plaque shows the current score; stop passing and it flips to a grey recheck design with no score, so an old result is never silently kept; take the badge down and the listing is dropped. One scoring rule now applies to everyone, the health score the audit reports for the project's audit path, which moves SkillPort's plaque from the weakest pack figure of 92 to the whole collection figure of 98. The registry data lives in registry/projects.json, the generator in tools/registry_build.py (its audit, parsing and bookkeeping steps run on local folders and fixtures, which is what its tests use), and the plain audited mark is served alongside, carrying no score and enrolling nothing. The static --write-badge files remain as a dated snapshot for announcements and offline use, and the audit report's closing lines now present the live route first. Seventy four tests pass.
v0.3.9
The audit now ends with something worth showing. A passing audit earns the SkillDoctor badge, and this release gives the badge a real design (a gold pulse on ink, the project mark), a large plaque format for a place of honor at the top of a README, and a registry to make both mean something: the picture alone proves nothing, so listings in REGISTRY.md are added only after the maintainers re-run the audit and the results, file hashes included, match the submitted report. Run with --write-badge and a passing audit writes the personalized badge (score and month on its face), the plaque (health score, skill count and the zero failures line) and the machine report into the audited folder; a failing audit writes nothing and says why. The JSON report now carries the SHA-256 of every audited SKILL.md and the audit date. The conflict check also learned the shape of a skill collection: a skills/ source and its for- distribution copies inside one tree are one skill, not a conflict, while genuine duplicates still flag, and template folders holding scaffolds are no longer audited as if they were installed skills. The installers now download skilldoctor.py and its checksum as release assets instead of raw files, so installs are countable, with the same one line command and the same checksum verification. Fifty three tests pass.
SkillDoctor v0.3.8
Six fixes from a careful external review (thank you, jcc2010; issue #18). The security scan now reads the whole SKILL.md, frontmatter included, and scans the text files bundled in a skill folder, listing anything it cannot scan as not audited. Invisible Unicode characters are a security error. Token cost splits into always-on (name and description) and body (per use), and the quarantine line frees the always-on figure. Spec length limits are checked, and plugin caches no longer produce false duplicate conflicts. The installers pin to this release tag and verify the published SHA-256 before installing. See RELEASE-NOTES.md for the full entry.
SkillDoctor v0.3.7
SkillDoctor now recognizes Muse skills. A Muse skill keeps its metadata in a skill.yaml file beside the SKILL.md, with no frontmatter by design, and the auditor used to fail every one of them for that missing block. It now reads the skill.yaml, takes the name and description from there, validates the name as snake_case, and runs the full security checks on the body as before. Claude format skills are audited exactly as they were.
SkillDoctor v0.3.6
Adds a visitor counter badge to the README, so the reach of the repo is visible at a glance. No checker changes.
v0.3.5
The PyPI page shows the pictures now: the README images point at absolute addresses that PyPI can load, instead of relative ones only GitHub could resolve.
v0.3.4
SkillDoctor is on PyPI as skillcheckup: pip install skillcheckup, then run the skilldoctor command. Publishing now runs through PyPI trusted publishing from this repos release workflow, with no stored token anywhere.
v0.3.3
Getting in is easier. The Windows installer now creates a real skilldoctor command instead of leaving you to call the file by hand, and every release from now on ships standalone executables for Windows, macOS and Linux, so no Python installation is needed at all. The README action example points at this version.
v0.3.2
SkillDoctor now looks deeper. Real collections nest their skills a few folders down, and the scanner only looked one level deep, so pointing it at a collection root found nothing. It now walks the tree. The Windows install line also gets its own code block with a copy button, like the Mac and Linux one.
v0.3.1
The report now speaks first and guides after. A plain language verdict opens every report, an empty scan explains where to point the tool instead of dead ending, and a failing audit ends with the exact next command and the tokens it frees. The README quick start moved up to where a new visitor looks for it. The CI matrix also drops Python 3.8, which current GitHub runners no longer provide, and runs each version to completion instead of cancelling the rest.