[RealZST]

Summary

SKILL.md frontmatter is YAML, but parse_skill_frontmatter copied the text after name: / description: verbatim. A quoted value such as name: "imagegen" therefore became a skill named "imagegen", quotes included. All of OpenAI's bundled Codex skills (imagegen, pdf, playwright, speech, spreadsheet, doc, …) quote their names, and every agent that reads them with a YAML parser sees imagegen.

What that did in HarnessKit, reproduced on origin/main:

	Before	After
hk list	"imagegen"	imagegen
hk info imagegen / hk disable imagegen	Error: Extension not found: imagegen	works
Install paths that name the target directory after the parsed name	create a folder literally called "imagegen"	imagegen
Hermes native disable (#172)	the quoted name written to skills.disabled never matches Hermes' own name	matches
Directory skill without a frontmatter name, e.g. my.skill/	named my (file_stem)	named my.skill, as every agent does

skill_locations already carried a one-off patch for this ("strip surrounding quotes if present"), which masked the symptom in one code path only. It is removed.

Changes

• yaml_scalar: each frontmatter value is parsed with serde_yaml and used when it is a string, so quotes, escapes and single vs. double quoting resolve the way YAML defines them. Text that YAML would not read as a plain string (an unquoted colon, a bare number, a boolean) is kept verbatim, so hand-written frontmatter such as description: Use when: foo keeps working. The whole block is deliberately not handed to a strict YAML parser for that reason. bins items go through the same helper instead of trimming quote characters by hand.
• skill_entry_name: a directory skill falls back to its whole folder name, a standalone .md to its stem (notes.md.disabled → notes). The scanner, find_skill_by_id and find_skill_content each had their own file_stem fallback; they now share this one, and the scanner also uses it as the skills CLI lockfile key instead of recomputing it.
• Four tests, written failing first: quoted scalars resolve; plain values YAML would reject are kept; a my.skill/ directory is named my.skill; find_skill_by_id resolves the id the scanner produced for it (this one guards the finders against drifting from the scanner again).

Verification

• cargo test --workspace: 813 passed, 0 failed.
• cargo clippy --workspace --all-targets: no new diagnostics in touched files.
• rustfmt --check compared hunk-by-hunk against origin/main: no new drift in either file.
• Multi-agent regression check (Claude directory + loose skill, Codex, Gemini, Cursor, Grok flat + nested, a five-agent skill in ~/.agents/skills): hk list --json, hk status, and disable/enable round-trips produce byte-identical output with the origin/main binary and this branch's binary.

Upgrade notes

• A skill's id derives from its name, so rows for previously quoted names are re-keyed on the first scan after upgrading; a HarnessKit-side disable on one of them is reset once.
• Directories that earlier versions created with quotes in their name (e.g. ~/.claude/skills/"imagegen"/) are left alone. HarnessKit now identifies them by their frontmatter name, so they keep working; renaming the folder or reinstalling the skill gives it a clean directory name. No automatic rename, since a same-named directory may already exist.

🤖 Generated with Claude Code