harn_skills/
lib.rs

1//! Embedded Harn skill corpus.
2//!
3//! This crate exposes the bundled corpus as metadata plus `SKILL.md`
4//! bodies. CLI commands that enumerate, dump, or install these skills
5//! are layered above this foundation.
6
7use std::collections::BTreeMap;
8use std::env;
9use std::fmt;
10use std::fs;
11use std::io;
12use std::path::{Path, PathBuf};
13use std::sync::OnceLock;
14
15/// Environment override for canonical Harn skill discovery.
16pub const HARN_SKILLS_DIR_ENV: &str = "HARN_SKILLS_DIR";
17
18/// Frontmatter fields embedded with each bundled skill.
19#[derive(Debug, Clone, Copy, PartialEq, Eq)]
20pub struct SkillFrontmatter {
21    pub name: &'static str,
22    pub short: &'static str,
23    pub description: &'static str,
24    pub when_to_use: Option<&'static str>,
25}
26
27/// A single skill embedded into the Harn build.
28#[derive(Debug, Clone, Copy, PartialEq, Eq)]
29pub struct EmbeddedSkill {
30    pub name: &'static str,
31    pub frontmatter: SkillFrontmatter,
32    pub body: &'static str,
33    /// The full original SKILL.md source — frontmatter delimiter,
34    /// frontmatter block, blank line, and body — exactly as embedded.
35    /// Use this when round-tripping a skill back to disk so the dumped
36    /// copy is byte-identical to the binary's canonical record.
37    pub source: &'static str,
38}
39
40/// Owned frontmatter fields loaded from a `SKILL.md` on disk.
41#[derive(Debug, Clone, PartialEq, Eq)]
42pub struct DiskSkillFrontmatter {
43    pub name: String,
44    pub short: String,
45    pub description: String,
46    pub when_to_use: Option<String>,
47}
48
49/// A single skill discovered recursively from `HARN_SKILLS_DIR`.
50#[derive(Debug, Clone, PartialEq, Eq)]
51pub struct DiskSkill {
52    pub name: String,
53    pub frontmatter: DiskSkillFrontmatter,
54    pub body: String,
55    pub source: String,
56    pub path: PathBuf,
57}
58
59/// The active canonical corpus used by `harn skill list/get`.
60#[derive(Debug, Clone, PartialEq, Eq)]
61pub enum SkillCorpus {
62    Embedded(&'static [EmbeddedSkill]),
63    Disk(Vec<DiskSkill>),
64}
65
66impl SkillCorpus {
67    pub fn is_disk(&self) -> bool {
68        matches!(self, Self::Disk(_))
69    }
70
71    pub fn len(&self) -> usize {
72        match self {
73            Self::Embedded(skills) => skills.len(),
74            Self::Disk(skills) => skills.len(),
75        }
76    }
77
78    pub fn is_empty(&self) -> bool {
79        self.len() == 0
80    }
81}
82
83/// Error returned when disk skill discovery finds malformed files.
84#[derive(Debug)]
85pub enum SkillDiscoveryError {
86    Io {
87        path: PathBuf,
88        source: io::Error,
89    },
90    MissingFrontmatter {
91        path: PathBuf,
92    },
93    MissingField {
94        path: PathBuf,
95        field: &'static str,
96    },
97    DuplicateName {
98        name: String,
99        first: PathBuf,
100        second: PathBuf,
101    },
102}
103
104impl fmt::Display for SkillDiscoveryError {
105    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
106        match self {
107            Self::Io { path, source } => write!(f, "{}: {source}", path.display()),
108            Self::MissingFrontmatter { path } => {
109                write!(f, "{}: missing SKILL.md frontmatter", path.display())
110            }
111            Self::MissingField { path, field } => {
112                write!(f, "{}: missing `{field}` frontmatter field", path.display())
113            }
114            Self::DuplicateName {
115                name,
116                first,
117                second,
118            } => write!(
119                f,
120                "duplicate skill `{name}` in {} and {}",
121                first.display(),
122                second.display()
123            ),
124        }
125    }
126}
127
128impl std::error::Error for SkillDiscoveryError {}
129
130const SOURCES: &[&str] = &[
131    include_str!("corpus/harn-agent/SKILL.md"),
132    include_str!("corpus/harn-de-slop/SKILL.md"),
133    include_str!("corpus/harn-diagnostics/SKILL.md"),
134    include_str!("corpus/harn-language/SKILL.md"),
135    include_str!("corpus/harn-orchestration/SKILL.md"),
136    include_str!("corpus/harn-probe/SKILL.md"),
137    include_str!("corpus/harn-providers/SKILL.md"),
138    include_str!("corpus/harn-rules/SKILL.md"),
139    include_str!("corpus/harn-testing/SKILL.md"),
140    include_str!("corpus/harn-tracing/SKILL.md"),
141    include_str!("corpus/release-harn/SKILL.md"),
142];
143
144static EMBEDDED_SKILLS: OnceLock<Box<[EmbeddedSkill]>> = OnceLock::new();
145
146/// Return every skill bundled into this build.
147pub fn list_embedded_skills() -> &'static [EmbeddedSkill] {
148    EMBEDDED_SKILLS
149        .get_or_init(|| SOURCES.iter().map(|source| parse_skill(source)).collect())
150        .as_ref()
151}
152
153/// Return one bundled skill by canonical skill name.
154pub fn get_embedded_skill(name: &str) -> Option<&'static EmbeddedSkill> {
155    list_embedded_skills()
156        .iter()
157        .find(|skill| skill.name == name)
158}
159
160/// Return the active canonical corpus. `HARN_SKILLS_DIR` wins only
161/// when it contains at least one recursively discovered `SKILL.md`;
162/// otherwise callers fall back to the embedded corpus.
163pub fn resolve_skill_corpus_from_env() -> Result<SkillCorpus, SkillDiscoveryError> {
164    let Ok(dir) = env::var(HARN_SKILLS_DIR_ENV) else {
165        return Ok(SkillCorpus::Embedded(list_embedded_skills()));
166    };
167    if dir.trim().is_empty() {
168        return Ok(SkillCorpus::Embedded(list_embedded_skills()));
169    }
170
171    let skills = list_disk_skills(dir)?;
172    if skills.is_empty() {
173        Ok(SkillCorpus::Embedded(list_embedded_skills()))
174    } else {
175        Ok(SkillCorpus::Disk(skills))
176    }
177}
178
179/// Recursively discover `SKILL.md` files under `root`.
180///
181/// A missing root is treated as an empty disk corpus so
182/// `HARN_SKILLS_DIR` can fall back cleanly to embedded skills.
183pub fn list_disk_skills(root: impl AsRef<Path>) -> Result<Vec<DiskSkill>, SkillDiscoveryError> {
184    let root = root.as_ref();
185    if !root.exists() {
186        return Ok(Vec::new());
187    }
188
189    let mut paths = Vec::new();
190    collect_skill_paths(root, &mut paths)?;
191    paths.sort();
192
193    let mut by_name: BTreeMap<String, DiskSkill> = BTreeMap::new();
194    for path in paths {
195        let skill = parse_disk_skill(&path)?;
196        if let Some(first) = by_name.get(&skill.name) {
197            return Err(SkillDiscoveryError::DuplicateName {
198                name: skill.name,
199                first: first.path.clone(),
200                second: path,
201            });
202        }
203        by_name.insert(skill.name.clone(), skill);
204    }
205
206    Ok(by_name.into_values().collect())
207}
208
209fn parse_skill(source: &'static str) -> EmbeddedSkill {
210    let (frontmatter, body) = split_frontmatter(source);
211    let frontmatter = parse_frontmatter(frontmatter);
212    EmbeddedSkill {
213        name: frontmatter.name,
214        frontmatter,
215        body,
216        source,
217    }
218}
219
220fn collect_skill_paths(dir: &Path, out: &mut Vec<PathBuf>) -> Result<(), SkillDiscoveryError> {
221    let entries = fs::read_dir(dir).map_err(|source| SkillDiscoveryError::Io {
222        path: dir.to_path_buf(),
223        source,
224    })?;
225    for entry in entries {
226        let entry = entry.map_err(|source| SkillDiscoveryError::Io {
227            path: dir.to_path_buf(),
228            source,
229        })?;
230        let path = entry.path();
231        let file_type = entry
232            .file_type()
233            .map_err(|source| SkillDiscoveryError::Io {
234                path: path.clone(),
235                source,
236            })?;
237        if file_type.is_dir() {
238            collect_skill_paths(&path, out)?;
239        } else if file_type.is_file() && entry.file_name() == "SKILL.md" {
240            out.push(path);
241        }
242    }
243    Ok(())
244}
245
246fn parse_disk_skill(path: &Path) -> Result<DiskSkill, SkillDiscoveryError> {
247    let source = fs::read_to_string(path).map_err(|source| SkillDiscoveryError::Io {
248        path: path.to_path_buf(),
249        source,
250    })?;
251    let (frontmatter, body) =
252        split_disk_frontmatter(&source).ok_or_else(|| SkillDiscoveryError::MissingFrontmatter {
253            path: path.to_path_buf(),
254        })?;
255    let frontmatter = parse_disk_frontmatter(path, frontmatter)?;
256    Ok(DiskSkill {
257        name: frontmatter.name.clone(),
258        frontmatter,
259        body: body.to_string(),
260        source,
261        path: path.to_path_buf(),
262    })
263}
264
265fn split_disk_frontmatter(source: &str) -> Option<(&str, &str)> {
266    split_frontmatter_parts(source)
267}
268
269fn parse_disk_frontmatter(
270    path: &Path,
271    frontmatter: &str,
272) -> Result<DiskSkillFrontmatter, SkillDiscoveryError> {
273    let mut name = None;
274    let mut short = None;
275    let mut description = None;
276    let mut when_to_use = None;
277
278    for line in frontmatter.lines() {
279        let Some((key, value)) = line.split_once(':') else {
280            continue;
281        };
282        let value = value.trim().to_string();
283        match key {
284            "name" => name = Some(value),
285            "short" => short = Some(value),
286            "description" => description = Some(value),
287            "when_to_use" => when_to_use = Some(value),
288            _ => {}
289        }
290    }
291
292    Ok(DiskSkillFrontmatter {
293        name: require_disk_field(path, name, "name")?,
294        short: short.unwrap_or_default(),
295        description: require_disk_field(path, description, "description")?,
296        when_to_use,
297    })
298}
299
300fn require_disk_field(
301    path: &Path,
302    value: Option<String>,
303    field: &'static str,
304) -> Result<String, SkillDiscoveryError> {
305    value.ok_or_else(|| SkillDiscoveryError::MissingField {
306        path: path.to_path_buf(),
307        field,
308    })
309}
310
311fn split_frontmatter(source: &'static str) -> (&'static str, &'static str) {
312    let Some((after_open, line_ending)) = split_opening_frontmatter(source) else {
313        panic!("embedded skill source is missing opening frontmatter delimiter");
314    };
315    let Some((frontmatter, body)) = split_closing_frontmatter(after_open, line_ending) else {
316        panic!("embedded skill source is missing closing frontmatter delimiter");
317    };
318    (frontmatter, body)
319}
320
321fn split_frontmatter_parts(source: &str) -> Option<(&str, &str)> {
322    let (after_open, line_ending) = split_opening_frontmatter(source)?;
323    split_closing_frontmatter(after_open, line_ending)
324}
325
326fn split_opening_frontmatter(source: &str) -> Option<(&str, &str)> {
327    if let Some(after_open) = source.strip_prefix("---\n") {
328        Some((after_open, "\n"))
329    } else if let Some(after_open) = source.strip_prefix("---\r\n") {
330        Some((after_open, "\r\n"))
331    } else {
332        None
333    }
334}
335
336fn split_closing_frontmatter<'a>(
337    after_open: &'a str,
338    line_ending: &str,
339) -> Option<(&'a str, &'a str)> {
340    let close = format!("{line_ending}---{line_ending}");
341    let close_offset = after_open.find(&close)?;
342    Some((
343        &after_open[..close_offset],
344        &after_open[close_offset + close.len()..],
345    ))
346}
347
348fn parse_frontmatter(frontmatter: &'static str) -> SkillFrontmatter {
349    let mut name = None;
350    let mut short = None;
351    let mut description = None;
352    let mut when_to_use = None;
353
354    for line in frontmatter.lines() {
355        let Some((key, value)) = line.split_once(':') else {
356            continue;
357        };
358        let value = value.trim();
359        match key {
360            "name" => name = Some(value),
361            "short" => short = Some(value),
362            "description" => description = Some(value),
363            "when_to_use" => when_to_use = Some(value),
364            _ => {}
365        }
366    }
367
368    SkillFrontmatter {
369        name: name.expect("embedded skill frontmatter is missing `name`"),
370        short: short.expect("embedded skill frontmatter is missing `short`"),
371        description: description.expect("embedded skill frontmatter is missing `description`"),
372        when_to_use,
373    }
374}
375
376#[cfg(test)]
377mod tests {
378    use super::*;
379    use std::collections::BTreeSet;
380    use tempfile::TempDir;
381
382    #[test]
383    fn lists_expected_initial_corpus() {
384        let skills = list_embedded_skills();
385        let names: Vec<&str> = skills.iter().map(|skill| skill.name).collect();
386        assert_eq!(
387            names,
388            [
389                "harn-agent",
390                "harn-de-slop",
391                "harn-diagnostics",
392                "harn-language",
393                "harn-orchestration",
394                "harn-probe",
395                "harn-providers",
396                "harn-rules",
397                "harn-testing",
398                "harn-tracing",
399                "release-harn",
400            ]
401        );
402        assert_eq!(skills.len(), SOURCES.len());
403    }
404
405    #[test]
406    fn can_fetch_harn_language_skill() {
407        let skill = get_embedded_skill("harn-language").expect("harn-language skill is embedded");
408        assert_eq!(skill.frontmatter.name, "harn-language");
409        assert!(skill.body.contains("Harn language"));
410    }
411
412    #[test]
413    fn skills_have_unique_names_and_body_only_content() {
414        let mut names = BTreeSet::new();
415        for skill in list_embedded_skills() {
416            assert_eq!(skill.name, skill.frontmatter.name);
417            assert!(names.insert(skill.name), "duplicate skill {}", skill.name);
418            assert!(
419                !skill.body.trim().is_empty(),
420                "{} body is empty",
421                skill.name
422            );
423            assert!(
424                !skill.body.trim_start().starts_with("---"),
425                "{} body includes frontmatter",
426                skill.name
427            );
428        }
429    }
430
431    #[test]
432    fn skills_are_sorted_by_name() {
433        let names: Vec<&str> = list_embedded_skills()
434            .iter()
435            .map(|skill| skill.name)
436            .collect();
437        let mut sorted = names.clone();
438        sorted.sort_unstable();
439        assert_eq!(names, sorted);
440    }
441
442    #[test]
443    fn source_round_trips_to_frontmatter_and_body() {
444        for skill in list_embedded_skills() {
445            assert!(
446                split_frontmatter_parts(skill.source).is_some(),
447                "{} source missing opening fence",
448                skill.name
449            );
450            assert!(
451                skill.source.ends_with(skill.body),
452                "{} source must end with the body so dump output is byte-stable",
453                skill.name
454            );
455            assert!(
456                skill.source.contains(&format!("name: {}\n", skill.name))
457                    || skill.source.contains(&format!("name: {}\r\n", skill.name)),
458                "{} source missing canonical name field",
459                skill.name
460            );
461        }
462    }
463
464    #[test]
465    fn frontmatter_split_accepts_crlf_sources() {
466        let source = "---\r\nname: crlf\r\n---\r\n# Body\r\n";
467        let (frontmatter, body) = split_frontmatter_parts(source).expect("CRLF frontmatter");
468        assert_eq!(frontmatter, "name: crlf");
469        assert_eq!(body, "# Body\r\n");
470    }
471
472    #[test]
473    fn frontmatter_split_rejects_missing_closing_fence() {
474        assert!(split_frontmatter_parts("---\nname: missing\n# Body\n").is_none());
475    }
476
477    #[test]
478    fn embedded_corpus_stays_within_binary_budget() {
479        let bytes: usize = SOURCES.iter().map(|source| source.len()).sum();
480        assert!(
481            bytes <= 200 * 1024,
482            "embedded corpus is {bytes} bytes, expected <= 200 KiB"
483        );
484    }
485
486    #[test]
487    fn skill_bodies_are_focused_and_not_placeholders() {
488        let expectations = [
489            ("harn-agent", ["agent_loop", "session id", "approval"]),
490            ("harn-de-slop", ["typed", "structural guard", "rebase"]),
491            ("harn-diagnostics", ["diagnostic", "repair", "conformance"]),
492            ("harn-language", ["quickref", "type", "conformance"]),
493            ("harn-orchestration", ["agent_loop", "workflow", "host"]),
494            ("harn-probe", ["probe", "fact", "evidence"]),
495            ("harn-providers", ["llm_call", "provider", "schema"]),
496            (
497                "harn-testing",
498                ["conformance", "deterministic", "mock_time"],
499            ),
500            ("harn-tracing", ["replay", "receipts", "transcript"]),
501            ("release-harn", ["release_ship", "merge queue", "tag"]),
502        ];
503
504        for (name, terms) in expectations {
505            let skill = get_embedded_skill(name).expect("expected embedded skill");
506            let body = skill.body.to_ascii_lowercase();
507            assert!(
508                !body.contains("embedded stub") && !body.contains("placeholder"),
509                "{name} should contain real guidance, not stub wording"
510            );
511            for term in terms {
512                assert!(
513                    body.contains(term),
514                    "{name} body should mention focused term `{term}`"
515                );
516            }
517        }
518    }
519
520    #[test]
521    fn skill_bodies_match_split_skill_contract() {
522        for skill in list_embedded_skills() {
523            let lines = skill.body.lines().count();
524            assert!(
525                lines >= 80,
526                "{} body is {lines} lines, expected at least 80",
527                skill.name
528            );
529            assert!(
530                lines <= 300,
531                "{} body is {lines} lines, expected at most 300",
532                skill.name
533            );
534        }
535    }
536
537    #[test]
538    fn skill_cross_links_resolve_to_embedded_skills() {
539        let names: BTreeSet<&str> = list_embedded_skills()
540            .iter()
541            .map(|skill| skill.name)
542            .collect();
543        for skill in list_embedded_skills() {
544            for reference in bracketed_skill_references(skill.body) {
545                assert!(
546                    names.contains(reference),
547                    "{} links to unknown embedded skill [[{}]]",
548                    skill.name,
549                    reference
550                );
551            }
552        }
553    }
554
555    #[test]
556    fn diagnostics_skill_mentions_all_code_categories() {
557        let skill = get_embedded_skill("harn-diagnostics").expect("diagnostics skill");
558        for category in [
559            "TYP", "PAR", "NAM", "CAP", "LLM", "ORC", "STD", "PRM", "MOD", "LNT", "FMT", "IMP",
560            "OWN", "RCV", "MAT",
561        ] {
562            assert!(
563                skill.body.contains(&format!("`{category}`")),
564                "harn-diagnostics should mention diagnostic category `{category}`"
565            );
566        }
567    }
568
569    #[test]
570    fn disk_discovery_finds_recursive_skill_files_sorted_by_name() {
571        let temp = TempDir::new().expect("temp dir");
572        write_skill(
573            &temp.path().join("zeta").join("SKILL.md"),
574            "zeta-skill",
575            "Zeta",
576        );
577        write_skill(
578            &temp.path().join("nested").join("alpha").join("SKILL.md"),
579            "alpha-skill",
580            "Alpha",
581        );
582
583        let skills = list_disk_skills(temp.path()).expect("discover disk skills");
584        let names: Vec<&str> = skills.iter().map(|skill| skill.name.as_str()).collect();
585        assert_eq!(names, ["alpha-skill", "zeta-skill"]);
586        assert_eq!(skills[0].frontmatter.description, "Alpha description");
587        assert!(skills[0].body.contains("Alpha body"));
588    }
589
590    #[test]
591    fn disk_discovery_treats_missing_root_as_empty() {
592        let temp = TempDir::new().expect("temp dir");
593        let skills = list_disk_skills(temp.path().join("missing")).expect("discover disk skills");
594        assert!(skills.is_empty());
595    }
596
597    #[test]
598    fn disk_discovery_rejects_duplicate_skill_names() {
599        let temp = TempDir::new().expect("temp dir");
600        write_skill(
601            &temp.path().join("one").join("SKILL.md"),
602            "same-skill",
603            "One",
604        );
605        write_skill(
606            &temp.path().join("two").join("SKILL.md"),
607            "same-skill",
608            "Two",
609        );
610
611        let error = list_disk_skills(temp.path()).expect_err("duplicate name should fail");
612        assert!(
613            error.to_string().contains("duplicate skill `same-skill`"),
614            "unexpected error: {error}"
615        );
616    }
617
618    fn write_skill(path: &Path, name: &str, label: &str) {
619        fs::create_dir_all(path.parent().expect("skill parent")).expect("create skill parent");
620        fs::write(
621            path,
622            format!(
623                "---\nname: {name}\nshort: {label} short\ndescription: {label} description\n---\n# {label}\n\n{label} body\n"
624            ),
625        )
626        .expect("write SKILL.md");
627    }
628
629    fn bracketed_skill_references(body: &str) -> Vec<&str> {
630        let mut references = Vec::new();
631        let mut rest = body;
632        while let Some(start) = rest.find("[[") {
633            rest = &rest[start + 2..];
634            let Some(end) = rest.find("]]") else {
635                break;
636            };
637            references.push(&rest[..end]);
638            rest = &rest[end + 2..];
639        }
640        references
641    }
642}
harn_skills/lib.rs

harn_skills/
lib.rs
