| 1 | //! Shared frontmatter reader for Skills, agent profiles, and installation. |
| 2 | //! This leaf is also included by the installation acceptance harness. |
| 3 | |
| 4 | use std::collections::HashMap; |
| 5 | use std::path::{Component, Path}; |
| 6 | |
| 7 | /// Parsed frontmatter: lowercased metadata keys and the body after the fence. |
| 8 | pub(crate) type Frontmatter<'a> = (HashMap<String, String>, &'a str); |
| 9 | |
| 10 | #[derive(Clone, Copy, PartialEq, Eq)] |
| 11 | pub(crate) enum SkillValidationMode { |
| 12 | Lenient, |
| 13 | Strict, |
| 14 | } |
| 15 | |
| 16 | /// Runtime accepts incomplete authoring metadata with visible warnings; |
| 17 | /// installation requires frontmatter, a path-safe name, and a description. |
| 18 | /// Neither mode grants tools, model selection, execution, or fork authority. |
| 19 | pub(crate) fn validate_skill_frontmatter( |
| 20 | metadata: Option<&HashMap<String, String>>, |
| 21 | path: Option<&Path>, |
| 22 | mode: SkillValidationMode, |
| 23 | ) -> Result<Vec<String>, String> { |
| 24 | let Some(metadata) = metadata else { |
| 25 | return match mode { |
| 26 | SkillValidationMode::Strict => { |
| 27 | Err("SKILL.md is missing the leading '---' frontmatter fence".into()) |
| 28 | } |
| 29 | SkillValidationMode::Lenient => Ok(vec![ |
| 30 | "missing frontmatter; using the Markdown heading as the skill name".into(), |
| 31 | "missing description".into(), |
| 32 | ]), |
| 33 | }; |
| 34 | }; |
| 35 | let name = metadata |
| 36 | .get("name") |
| 37 | .filter(|name| !name.trim().is_empty()) |
| 38 | .ok_or_else(|| "missing required frontmatter field: name".to_string())?; |
| 39 | if mode == SkillValidationMode::Strict && !is_path_safe_skill_name(name) { |
| 40 | return Err(format!( |
| 41 | "SKILL.md `name` must be a single path-safe segment (got '{name}')" |
| 42 | )); |
| 43 | } |
| 44 | let mut warnings = Vec::new(); |
| 45 | if !metadata |
| 46 | .get("description") |
| 47 | .is_some_and(|value| !value.trim().is_empty()) |
| 48 | { |
| 49 | if mode == SkillValidationMode::Strict { |
| 50 | return Err("missing required frontmatter field: description".into()); |
| 51 | } |
| 52 | warnings.push("missing description".into()); |
| 53 | } |
| 54 | if let Some(directory) = path |
| 55 | .and_then(Path::parent) |
| 56 | .and_then(Path::file_name) |
| 57 | .and_then(|name| name.to_str()) |
| 58 | && directory != name |
| 59 | { |
| 60 | warnings.push(format!( |
| 61 | "frontmatter name `{name}` differs from directory `{directory}`" |
| 62 | )); |
| 63 | } |
| 64 | let mut keys: Vec<_> = metadata.keys().collect(); |
| 65 | keys.sort(); |
| 66 | for key in keys { |
| 67 | if key.starts_with("metadata.") || key.starts_with("x-") || key.starts_with("description_") |
| 68 | { |
| 69 | continue; |
| 70 | } |
| 71 | let warning = match key.as_str() { |
| 72 | "name" | "description" | "invocation" | "aliases-for" | "license" | "compatibility" |
| 73 | | "metadata" | "when_to_use" | "argument-hint" => None, |
| 74 | "disable-model-invocation" | "user-invocable" => { |
| 75 | if parse_frontmatter_bool(&metadata[key]).is_none() { |
| 76 | let warning = format!("invalid boolean `{key}`; invocation fails closed"); |
| 77 | if mode == SkillValidationMode::Strict { |
| 78 | return Err(warning); |
| 79 | } |
| 80 | Some(warning) |
| 81 | } else { |
| 82 | None |
| 83 | } |
| 84 | } |
| 85 | "allowed-tools" | "disallowed-tools" => Some(format!( |
| 86 | "`{key}` ignored: skills grant no tool or approval authority" |
| 87 | )), |
| 88 | "model" => Some("`model` ignored: skills do not select providers or models".into()), |
| 89 | "context" | "agent" => Some(format!( |
| 90 | "`{key}` unsupported: skills do not create a separate execution context" |
| 91 | )), |
| 92 | _ => Some(format!("unknown frontmatter key `{key}` ignored")), |
| 93 | }; |
| 94 | if let Some(warning) = warning { |
| 95 | warnings.push(warning); |
| 96 | } |
| 97 | } |
| 98 | Ok(warnings) |
| 99 | } |
| 100 | |
| 101 | pub(crate) fn parse_frontmatter_bool(value: &str) -> Option<bool> { |
| 102 | match value.trim().to_ascii_lowercase().as_str() { |
| 103 | "true" | "yes" | "on" | "1" => Some(true), |
| 104 | "false" | "no" | "off" | "0" => Some(false), |
| 105 | _ => None, |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | pub(crate) fn is_path_safe_skill_name(name: &str) -> bool { |
| 110 | if name.is_empty() |
| 111 | || name.trim() != name |
| 112 | || name.chars().any(char::is_whitespace) |
| 113 | || name.contains(['/', '\\']) |
| 114 | { |
| 115 | return false; |
| 116 | } |
| 117 | let mut components = Path::new(name).components(); |
| 118 | matches!(components.next(), Some(Component::Normal(_))) && components.next().is_none() |
| 119 | } |
| 120 | |
| 121 | /// Split a Markdown file into its `---` frontmatter metadata and body. |
| 122 | /// |
| 123 | /// Returns `Ok(None)` when the file does not open with a `---` fence. Keys are |
| 124 | /// lowercased; values are unquoted, and YAML block scalars (`>`, `|`, with |
| 125 | /// chomping) are folded the way `SKILL.md` has always read them. This is the |
| 126 | /// one frontmatter reader: skills and Claude Code agent files both use it. |
| 127 | pub(crate) fn parse_frontmatter( |
| 128 | content: &str, |
| 129 | ) -> std::result::Result<Option<Frontmatter<'_>>, String> { |
| 130 | let content = content |
| 131 | .strip_prefix('\u{feff}') |
| 132 | .unwrap_or(content) |
| 133 | .trim_start(); |
| 134 | let opening = content.split_inclusive('\n').next().unwrap_or_default(); |
| 135 | if opening.trim_end() != "---" { |
| 136 | return Ok(None); |
| 137 | } |
| 138 | let rest = &content[opening.len()..]; |
| 139 | let mut offset = 0; |
| 140 | let end = rest |
| 141 | .split_inclusive('\n') |
| 142 | .find_map(|line| { |
| 143 | let start = offset; |
| 144 | offset += line.len(); |
| 145 | (line.trim_end() == "---").then_some(start) |
| 146 | }) |
| 147 | .ok_or_else(|| "missing frontmatter closing delimiter".to_string())?; |
| 148 | let frontmatter = &rest[..end]; |
| 149 | let body = &rest[end + 3..]; |
| 150 | |
| 151 | let mut metadata = HashMap::new(); |
| 152 | let indentation = |line: &str| line.chars().take_while(|ch| ch.is_whitespace()).count(); |
| 153 | let lines: Vec<&str> = frontmatter.lines().collect(); |
| 154 | let mut i = 0; |
| 155 | let mut maps: Vec<(usize, String)> = Vec::new(); |
| 156 | while i < lines.len() { |
| 157 | let raw = lines[i]; |
| 158 | let line = raw.trim(); |
| 159 | if line.is_empty() || line.starts_with('#') { |
| 160 | i += 1; |
| 161 | continue; |
| 162 | } |
| 163 | if let Some((key, value)) = line.split_once(':') { |
| 164 | let indent = indentation(raw); |
| 165 | while maps |
| 166 | .last() |
| 167 | .is_some_and(|(parent_indent, _)| *parent_indent >= indent) |
| 168 | { |
| 169 | maps.pop(); |
| 170 | } |
| 171 | let key = key.trim().to_ascii_lowercase(); |
| 172 | let key = maps |
| 173 | .last() |
| 174 | .map_or_else(|| key.clone(), |(_, parent)| format!("{parent}.{key}")); |
| 175 | let value = value.trim(); |
| 176 | // Check for YAML block scalar indicators: > (folded), | (literal), |
| 177 | // optionally with chomping: >-, >+, |-, |+ |
| 178 | let is_block_scalar = matches!(value, ">" | "|" | ">-" | ">+" | "|-" | "|+"); |
| 179 | if is_block_scalar { |
| 180 | let is_folded = value.starts_with('>'); |
| 181 | let chomp = if value.ends_with('-') { |
| 182 | "strip" |
| 183 | } else if value.ends_with('+') { |
| 184 | "keep" |
| 185 | } else { |
| 186 | "clip" |
| 187 | }; |
| 188 | // Determine the base indentation from the key line |
| 189 | let base_indent = indentation(raw); |
| 190 | let mut block_lines: Vec<&str> = Vec::new(); |
| 191 | let mut content_indent: Option<usize> = None; |
| 192 | i += 1; |
| 193 | while i < lines.len() { |
| 194 | let raw_line = lines[i]; |
| 195 | if raw_line.trim().is_empty() { |
| 196 | // Empty lines are part of the block |
| 197 | block_lines.push(""); |
| 198 | i += 1; |
| 199 | continue; |
| 200 | } |
| 201 | let line_indent = indentation(raw_line); |
| 202 | if line_indent > base_indent { |
| 203 | // Track content indent from the first non-empty |
| 204 | // line so we strip only that one level of |
| 205 | // leading whitespace, preserving any deeper |
| 206 | // relative indentation (YAML §8.1.2). |
| 207 | if content_indent.is_none() { |
| 208 | content_indent = Some(line_indent); |
| 209 | } |
| 210 | block_lines.push(raw_line); |
| 211 | i += 1; |
| 212 | } else { |
| 213 | break; |
| 214 | } |
| 215 | } |
| 216 | let content_indent = content_indent.unwrap_or(base_indent); |
| 217 | // Strip only the content indent from each non-empty |
| 218 | // line so nested indentation survives. |
| 219 | let block_lines: Vec<&str> = block_lines |
| 220 | .iter() |
| 221 | .map(|raw| { |
| 222 | if raw.is_empty() { |
| 223 | "" |
| 224 | } else { |
| 225 | let indent = indentation(raw); |
| 226 | let strip = std::cmp::min(indent, content_indent); |
| 227 | let byte = raw.char_indices().nth(strip).map_or(raw.len(), |(i, _)| i); |
| 228 | &raw[byte..] |
| 229 | } |
| 230 | }) |
| 231 | .collect(); |
| 232 | // Apply chomping to trailing empty lines before folding. |
| 233 | // Chomping operates on the raw block_lines (before join), so |
| 234 | // strip / keep / clip behave per the YAML spec. |
| 235 | let block_lines = if matches!(chomp, "strip") { |
| 236 | // strip: remove all trailing empty lines |
| 237 | let mut lines = block_lines; |
| 238 | while lines.last().is_some_and(|s| s.is_empty()) { |
| 239 | lines.pop(); |
| 240 | } |
| 241 | lines |
| 242 | } else if matches!(chomp, "keep") { |
| 243 | // keep: no modification |
| 244 | block_lines |
| 245 | } else { |
| 246 | // clip: keep at most one trailing empty line |
| 247 | let mut lines = block_lines; |
| 248 | while lines.len() >= 2 |
| 249 | && lines[lines.len() - 1].is_empty() |
| 250 | && lines[lines.len() - 2].is_empty() |
| 251 | { |
| 252 | lines.pop(); |
| 253 | } |
| 254 | lines |
| 255 | }; |
| 256 | let description = if is_folded { |
| 257 | // Folded: join non-empty lines with spaces; empty |
| 258 | // lines become paragraph breaks. |
| 259 | let mut result = String::new(); |
| 260 | let mut pending_space = false; |
| 261 | for line in &block_lines { |
| 262 | if line.is_empty() { |
| 263 | result.push('\n'); |
| 264 | pending_space = false; |
| 265 | } else { |
| 266 | if pending_space { |
| 267 | result.push(' '); |
| 268 | } |
| 269 | result.push_str(line); |
| 270 | pending_space = true; |
| 271 | } |
| 272 | } |
| 273 | result |
| 274 | } else { |
| 275 | // Literal: join with newlines. |
| 276 | block_lines.join("\n") |
| 277 | }; |
| 278 | metadata.insert(key, description); |
| 279 | } else if value.is_empty() |
| 280 | && lines |
| 281 | .get(i + 1) |
| 282 | .is_some_and(|next| is_block_sequence_item(next)) |
| 283 | { |
| 284 | // A block sequence (`tools:` then ` - Read` lines) becomes |
| 285 | // one comma-separated value, the same as the flow form |
| 286 | // `tools: Read, Grep`. Dropping it would read as "no list". |
| 287 | let mut items = Vec::new(); |
| 288 | i += 1; |
| 289 | while let Some(next) = lines.get(i).filter(|next| is_block_sequence_item(next)) { |
| 290 | let item = next.trim()[1..].trim(); |
| 291 | let item = item |
| 292 | .strip_prefix('"') |
| 293 | .and_then(|v| v.strip_suffix('"')) |
| 294 | .or_else(|| item.strip_prefix('\'').and_then(|v| v.strip_suffix('\''))) |
| 295 | .unwrap_or(item); |
| 296 | if !item.is_empty() { |
| 297 | items.push(item); |
| 298 | } |
| 299 | i += 1; |
| 300 | } |
| 301 | metadata.insert(key, items.join(", ")); |
| 302 | } else if value.is_empty() { |
| 303 | // Child fields retain their map path. In particular, |
| 304 | // metadata.name must never replace the skill's own name. |
| 305 | metadata.insert(key.clone(), String::new()); |
| 306 | maps.push((indent, key)); |
| 307 | i += 1; |
| 308 | } else if value.starts_with('[') { |
| 309 | // Reuse the installed YAML reader for quoted flow items rather |
| 310 | // than splitting commas inside quoted tool names or aliases. |
| 311 | let documents = yaml_rust2::YamlLoader::load_from_str(value) |
| 312 | .map_err(|err| format!("invalid frontmatter sequence `{key}`: {err}"))?; |
| 313 | let items = documents |
| 314 | .first() |
| 315 | .and_then(yaml_rust2::Yaml::as_vec) |
| 316 | .ok_or_else(|| format!("frontmatter `{key}` must be a flow sequence"))?; |
| 317 | let values: Result<Vec<_>, _> = items |
| 318 | .iter() |
| 319 | .map(|item| { |
| 320 | item.as_str().ok_or_else(|| { |
| 321 | format!("frontmatter `{key}` sequence items must be strings") |
| 322 | }) |
| 323 | }) |
| 324 | .collect(); |
| 325 | metadata.insert(key, values?.join(", ")); |
| 326 | i += 1; |
| 327 | } else { |
| 328 | let unquoted = match value { |
| 329 | v if (v.starts_with('"') && v.ends_with('"') && v.len() >= 2) |
| 330 | || (v.starts_with('\'') && v.ends_with('\'') && v.len() >= 2) => |
| 331 | { |
| 332 | &v[1..v.len() - 1] |
| 333 | } |
| 334 | _ => value, |
| 335 | }; |
| 336 | i += 1; |
| 337 | let mut text = unquoted.to_string(); |
| 338 | // Wrapped plain scalars continue at a deeper indentation. |
| 339 | // A colon in that continuation belongs to the value, not a |
| 340 | // new metadata key. Quoted/flow values retain their grammar. |
| 341 | if !value.is_empty() && !value.starts_with(['"', '\'', '[', '{']) { |
| 342 | while let Some(next) = lines.get(i) { |
| 343 | if next.trim().is_empty() || indentation(next) <= indentation(raw) { |
| 344 | break; |
| 345 | } |
| 346 | if !next.trim_start().starts_with('#') { |
| 347 | text.push(' '); |
| 348 | text.push_str(next.trim()); |
| 349 | } |
| 350 | i += 1; |
| 351 | } |
| 352 | } |
| 353 | metadata.insert(key, text); |
| 354 | } |
| 355 | } else { |
| 356 | i += 1; |
| 357 | } |
| 358 | } |
| 359 | |
| 360 | Ok(Some((metadata, body))) |
| 361 | } |
| 362 | |
| 363 | /// A YAML block-sequence entry: `- item` (or a bare `-`) on its own line. |
| 364 | fn is_block_sequence_item(line: &str) -> bool { |
| 365 | let line = line.trim(); |
| 366 | line == "-" || line.starts_with("- ") |
| 367 | } |
| 368 | |
| 369 | #[cfg(test)] |
| 370 | mod tests { |
| 371 | use super::*; |
| 372 | |
| 373 | #[test] |
| 374 | fn nested_metadata_does_not_override_name_or_description() { |
| 375 | let (metadata, _) = parse_frontmatter("---\nname: outer\ndescription: real routing\nmetadata:\n name: impostor\n description: hidden\n nested:\n name: deeper\nlicense: MIT\n---\nbody").unwrap().unwrap(); |
| 376 | assert_eq!(metadata["name"], "outer"); |
| 377 | assert_eq!(metadata["description"], "real routing"); |
| 378 | assert_eq!(metadata["metadata.name"], "impostor"); |
| 379 | assert_eq!(metadata["metadata.nested.name"], "deeper"); |
| 380 | assert_eq!(metadata["license"], "MIT"); |
| 381 | } |
| 382 | |
| 383 | #[test] |
| 384 | fn flow_list_parsed() { |
| 385 | let (metadata, _) = parse_frontmatter("---\nname: demo\nallowed-tools: [Read, 'Bash(ls *)', \"Grep, Glob\"]\naliases-for: [other, another]\n---\nbody").unwrap().unwrap(); |
| 386 | assert_eq!(metadata["allowed-tools"], "Read, Bash(ls *), Grep, Glob"); |
| 387 | assert_eq!(metadata["aliases-for"], "other, another"); |
| 388 | assert!(parse_frontmatter("---\nname: demo\nallowed-tools: [Read\n---\nbody").is_err()); |
| 389 | } |
| 390 | |
| 391 | #[test] |
| 392 | fn metadata_short_description_captured() { |
| 393 | // Actual Codex sample frontmatter, retaining its nested metadata map. |
| 394 | let (metadata, _) = parse_frontmatter(include_str!( |
| 395 | "../../tests/fixtures/skills/codex-skill-creator.md" |
| 396 | )) |
| 397 | .unwrap() |
| 398 | .unwrap(); |
| 399 | assert_eq!( |
| 400 | metadata["metadata.short-description"], |
| 401 | "Create or update a skill" |
| 402 | ); |
| 403 | assert!(!metadata.contains_key("short-description")); |
| 404 | assert!( |
| 405 | validate_skill_frontmatter(Some(&metadata), None, SkillValidationMode::Strict) |
| 406 | .unwrap() |
| 407 | .is_empty() |
| 408 | ); |
| 409 | } |
| 410 | |
| 411 | #[test] |
| 412 | fn real_claude_fixture_preserves_block_sequence_and_authority_warning() { |
| 413 | let (metadata, _) = |
| 414 | parse_frontmatter(include_str!("../../tests/fixtures/skills/claude-access.md")) |
| 415 | .unwrap() |
| 416 | .unwrap(); |
| 417 | assert_eq!(metadata["name"], "access"); |
| 418 | assert_eq!( |
| 419 | metadata["allowed-tools"], |
| 420 | "Read, Write, Bash(ls *), Bash(mkdir *)" |
| 421 | ); |
| 422 | let warnings = |
| 423 | validate_skill_frontmatter(Some(&metadata), None, SkillValidationMode::Strict).unwrap(); |
| 424 | assert_eq!(warnings.len(), 1); |
| 425 | assert!(warnings[0].contains("grant no tool or approval authority")); |
| 426 | } |
| 427 | |
| 428 | #[test] |
| 429 | fn runtime_and_install_validation_agree() { |
| 430 | let (metadata, _) = parse_frontmatter("---\nname: valid\ndescription: routing\n---\nbody") |
| 431 | .unwrap() |
| 432 | .unwrap(); |
| 433 | assert_eq!( |
| 434 | validate_skill_frontmatter(Some(&metadata), None, SkillValidationMode::Lenient), |
| 435 | validate_skill_frontmatter(Some(&metadata), None, SkillValidationMode::Strict) |
| 436 | ); |
| 437 | let mut missing = metadata.clone(); |
| 438 | missing.remove("description"); |
| 439 | assert_eq!( |
| 440 | validate_skill_frontmatter(Some(&missing), None, SkillValidationMode::Lenient).unwrap(), |
| 441 | vec!["missing description"] |
| 442 | ); |
| 443 | assert!( |
| 444 | validate_skill_frontmatter(Some(&missing), None, SkillValidationMode::Strict) |
| 445 | .unwrap_err() |
| 446 | .contains("description") |
| 447 | ); |
| 448 | missing.remove("name"); |
| 449 | assert!( |
| 450 | validate_skill_frontmatter(Some(&missing), None, SkillValidationMode::Lenient).is_err() |
| 451 | ); |
| 452 | let mut unsafe_name = metadata; |
| 453 | unsafe_name.insert("name".into(), "../escape".into()); |
| 454 | assert!( |
| 455 | validate_skill_frontmatter(Some(&unsafe_name), None, SkillValidationMode::Strict) |
| 456 | .is_err() |
| 457 | ); |
| 458 | } |
| 459 | |
| 460 | #[test] |
| 461 | fn unknown_keys_warn_once_spec_keys_silent() { |
| 462 | let (metadata, _) = parse_frontmatter("---\nname: demo\ndescription: routing\nlicense: MIT\ncompatibility: Codewhale\nmetadata:\n vendor-field: fine\nx-custom: fine\nallowed-tools: [Read]\nmodel: example\ncontext: fork\nmystery: ignored\n---\nbody").unwrap().unwrap(); |
| 463 | let warnings = |
| 464 | validate_skill_frontmatter(Some(&metadata), None, SkillValidationMode::Lenient) |
| 465 | .unwrap(); |
| 466 | assert_eq!(warnings.len(), 4, "{warnings:?}"); |
| 467 | for key in ["allowed-tools", "model", "context", "mystery"] { |
| 468 | assert_eq!( |
| 469 | warnings |
| 470 | .iter() |
| 471 | .filter(|warning| warning.contains(key)) |
| 472 | .count(), |
| 473 | 1 |
| 474 | ); |
| 475 | } |
| 476 | } |
| 477 | |
| 478 | #[test] |
| 479 | fn invalid_invocation_booleans_fail_closed() { |
| 480 | let (metadata, _) = parse_frontmatter( |
| 481 | "---\nname: demo\ndescription: routing\nuser-invocable: maybe\n---\nbody", |
| 482 | ) |
| 483 | .unwrap() |
| 484 | .unwrap(); |
| 485 | assert!( |
| 486 | validate_skill_frontmatter(Some(&metadata), None, SkillValidationMode::Lenient) |
| 487 | .unwrap()[0] |
| 488 | .contains("fails closed") |
| 489 | ); |
| 490 | assert!( |
| 491 | validate_skill_frontmatter(Some(&metadata), None, SkillValidationMode::Strict).is_err() |
| 492 | ); |
| 493 | for value in ["yes", "on", "1", "true"] { |
| 494 | assert_eq!(parse_frontmatter_bool(value), Some(true)); |
| 495 | } |
| 496 | for value in ["no", "off", "0", "false"] { |
| 497 | assert_eq!(parse_frontmatter_bool(value), Some(false)); |
| 498 | } |
| 499 | } |
| 500 | } |
| 501 |