bmad-module-skill-forge 2.2.0 → 4.0.0

package/src/shared/scripts/skf-scan-skill-md-structure.py

@@ -1,5 +1,5 @@
 # /// script
-# requires-python = ">=3.10"
+# requires-python = ">=3.11"
 # dependencies = []
 # ///
 """SKF Scan Skill.md Structure — deterministic structural checks for SKILL.md.
@@ -8,23 +8,34 @@ Replaces the synonym-grep loops, bash fence recipes, and inline Python
 table-parser in `references/coherence-check.md` §§2.1 / 2.2 / 2.3 / 2.6 with
 a single subprocess invocation that emits JSON.
 
+The usage-scope, cross-reference and reference-check subcommands emit the
+facts that test-skill's coherence check (§2.4 zero-usage exports, §§3-4
+file, script and asset references) and verify-stack's Check 4 (literal
+citations between skills) would otherwise count, grep and compare by hand,
+so two runs over the same files always agree.
+
 Subcommands:
   scan <skill-md>
       Emit JSON describing fence balance, bare opening fences (no language
-      tag), and table column-count drift:
+      tag), table column-count drift and the Scripts & Assets section (see
+      "Scripts & Assets section" below):
         {
           "unbalanced_fences": <bool>,
           "fence_count": N,
           "bare_opening_fences": [{"line": N, "text": "..."}, ...],
           "table_drift": [{"line": N, "section": "<heading>",
                            "expected_cols": N, "actual_cols": N,
-                           "row": "..."}]
+                           "row": "..."}],
+          "scripts_assets": {"folders": ["assets", "scripts"],
+                             "section": {"heading": "...", "line": N} | null,
+                             "missing": <bool>}
         }
 
   scan <skill-md> --required-sections
       Emit JSON describing which of the three required section families
       (description / usage / api_surface) are present, and which synonym
-      satisfied the requirement (case-insensitive, `##`/`###` tolerated):
+      satisfied the requirement (see "Heading match rule" below), with the
+      outline a caller reads when a family is not satisfied:
         {
           "description":  {"satisfied": <bool>,
                            "matched_synonym": "<heading>" | null,
@@ -34,14 +45,129 @@ Subcommands:
                            "tried": [...]},
           "api_surface":  {"satisfied": <bool>,
                            "matched_synonym": "..." | null,
-                           "tried": [...]}
+                           "tried": [...]},
+          "frontmatter_description": <bool>,
+          "outline": [{"line": N, "level": N, "text": "<heading>"}, ...]
+        }
+      `outline` lists every heading of the body in document order (the
+      Line model's headings: outside the frontmatter and fenced code).
+      `frontmatter_description` is true when the frontmatter has a
+      `description` key with a non-empty value: inline, or on the
+      indented lines under it (a `>` or `|` block, say).
+
+  usage-scope <skill-md> --exports <json-file-or-'-'> [--kinds <k1,k2,...>]
+              [--body auto|single|split]
+      Emit JSON counting, for each export, the lines that use its name in
+      the usage scope (see "Usage scope" and "Export match" below):
+        {
+          "scope": {
+            "body": "single" | "split",
+            "sections": [{"heading": "<text>", "line": N,
+                          "end_line": N}, ...],
+            "reference_files": ["references/<file>.md", ...]
+          },
+          "exports": [{"name": "...", "kind": "..." | null, "count": N,
+                       "first_hit": {"file": "...", "line": N} | null},
+                      ...],
+          "zero_usage": [{"name": "...", "kind": "..." | null}, ...],
+          "warnings": ["references/<file>.md: cannot read: ...", ...]
+        }
+      --exports holds a JSON array of names or {"name", "kind"} objects,
+      an object with an `exports` array (test-skill's step 3 inventory),
+      or the validate-inventory.py result (its `inventory.exports`).
+      --kinds counts only the exports whose kind it lists; an export
+      given without a kind is always counted. --body picks the scope
+      (default auto; see "Usage scope"). `exports[]` is sorted by name,
+      then kind, with duplicates dropped; `zero_usage[]` lists the
+      exports whose count is 0.
+
+  cross-reference --skills <json-file-or-'-'> [--skills-root <dir>]
+                  [--pairs <json-file-or-'-'>]
+      Emit JSON listing where one skill's SKILL.md literally cites another
+      skill (see "Citation match" below):
+        {
+          "citations": [{"from": "<a>", "to": "<b>",
+                         "substring": "<as written>", "line": N,
+                         "count": N, "excerpt": "...",
+                         "hits": [{"line": N, "substring": "...",
+                                   "excerpt": "..."}, ...]}, ...],
+          "edges": [["<a>", "<b>"], ...],
+          "pairs": [{"library_a": "<a>", "library_b": "<b>",
+                     "a_cites_b": <bool>, "b_cites_a": <bool>,
+                     "cited": <bool>}, ...],
+          "warnings": ["<name>: ...", ...]
+        }
+      --skills holds a JSON array of {"name", "skill_md" | "path",
+      "aliases"} entries (`aliases` optional), or an object with a
+      `skills` array: the skf-enumerate-stack-skills.py inventory, whose
+      `path` is relative to the skills folder, so pass that folder as
+      --skills-root. An entry's `source_repo_basename` and
+      `source_root_basename` (the inventory gives both) are aliases too,
+      unless another entry has the same term as its name, an alias or a
+      source basename, since one word would then cite two skills.
+      `path` names the skill folder, `skill_md` the file;
+      a relative one resolves against --skills-root, else the working
+      directory, and a --skills-root that is not a directory is an
+      error. `citations[]` holds one entry per citing direction, sorted
+      by (from, to): `hits[]` lists every citing line in line order,
+      `count` is their number, and `substring`, `line` and `excerpt`
+      repeat the first hit. `edges[]` is the same set in the shape
+      skf-find-cycles.py reads ({"edges": [[from, to], ...]}), so this
+      output pipes into `find --edges -` as is. --pairs (a JSON array of
+      [a, b] or {"library_a", "library_b"} entries, or an object with a
+      `pairs` array) keeps only the citations between those pairs and
+      adds `pairs[]`, sorted, where `cited` means either direction
+      cites. A SKILL.md that cannot be read is a warning (its skill can
+      still be cited), and so is a pair that names a skill --skills
+      lacks.
+
+  reference-check <skill-md> [--source-root <dir>] [--skills-root <dir>]
+      Emit JSON checking every file, script and asset reference in the
+      SKILL.md body (see "Reference extraction" below):
+        {
+          "roots": {"skill": "<dir>", "source": "<dir>" | null,
+                    "skills": "<dir>" | null},
+          "references": [{"line": N,
+                          "type": "file-path" | "script-asset",
+                          "form": "link" | "mention",
+                          "target": "<as written>",
+                          "canonical": "<realpath>",
+                          "status": "ok" | "missing" | "escapes",
+                          "root": "skill" | "source" | "skills" | null},
+                         ...],
+          "counts": {"total": N, "ok": N, "missing": N, "escapes": N},
+          "scripts_assets": {...},
+          "warnings": ["--source-root is not a directory: <dir>", ...]
         }
+      A target resolves against the SKILL.md folder and through every
+      symlink (os.path.realpath) before the root check. The roots are
+      the SKILL.md folder, --source-root (the extraction tree,
+      metadata.json `source_root`) and --skills-root (the skills output
+      folder, for a stack skill whose references reach its constituent
+      skills); `root` names the first one that holds the target.
+      `escapes`: outside every root (whether it exists is not checked,
+      the escape is the finding); `missing`: inside a root, nothing
+      there; `ok`: inside a root and present. `references[]` is in
+      document order; the same canonical target twice on one line is one
+      reference. A --source-root or --skills-root that is not a
+      directory (a stale metadata.json `source_root`, say) is a warning,
+      not an error.
+
+Scripts & Assets section:
+  `folders` lists the `scripts` and `assets` folders beside SKILL.md,
+  sorted; `section` is the first heading outside the frontmatter and
+  fenced code whose text starts with `Scripts` (case-insensitive, any
+  level: the section the reference extraction below also reads), with its
+  text and line, or null; `missing` is true when a folder exists and no
+  such section does, the gap test-skill's coherence check records.
 
 Heading match rule:
-  Match the first `^#+\\s+<heading>$` (any number of `#`, case-insensitive,
-  surrounding whitespace trimmed) that matches any synonym in a family.
-  The reported `matched_synonym` is the canonical synonym from the list,
-  not the heading text as it appears in the file.
+  Match the first heading of the body (the Line model's headings: any
+  level, a closing `#` sequence dropped, never a `#` line inside fenced or
+  indented code or the frontmatter) whose full text, case-insensitive,
+  is a synonym of a family. The reported `matched_synonym` is the
+  canonical synonym from the list, not the heading text as it appears in
+  the file.
 
 Fence balance:
   Count triple-backtick (```) fence lines (`^```` at start of line, ignoring
@@ -74,20 +200,107 @@ Empty SKILL.md:
   An empty file yields `fence_count: 0`, `unbalanced_fences: false`,
   empty `bare_opening_fences`, empty `table_drift`; for
   `--required-sections`, all three families have `satisfied: false`,
-  `matched_synonym: null`.
+  `matched_synonym: null`, with `frontmatter_description: false` and an
+  empty `outline`.
+
+Line model (scan --required-sections, usage-scope, cross-reference,
+reference-check):
+  A file is read as UTF-8, a leading byte order mark dropped, and split
+  on `\\n` alone (a `\\r\\n` pair loses its `\\r`, a lone `\\r` stays in
+  its line), so every `line` matches `grep -n`. The frontmatter is a
+  leading `---` ... `---` block. A fence opens on a line of three or more
+  backticks or tildes (any indent) and closes on the same character
+  repeated at least as often with nothing after it; an unclosed fence
+  runs to the end of the file. A heading is an ATX heading outside the
+  frontmatter and fenced code: at most three spaces of indent (four make
+  indented code), one to six `#`, then a space, a tab or the end of the
+  line, with a closing `#` sequence dropped from its text. Its section
+  runs to the line before the next heading of the same or a higher level.
+
+Usage scope:
+  Single-body (no `*.md` file under `references/`): the section of the
+  first heading whose text is a usage synonym (REQUIRED_SYNONYMS["usage"],
+  the anchor coherence-check.md §2.1 matches), fenced code inside it
+  included. The Key Exports list is outside it, so listing an export
+  there does not count as using it. With no such heading the scope is
+  empty and every export is zero-usage.
+  Split-body (a `*.md` file under `references/`): every usage-family
+  section of SKILL.md (the headings in USAGE_FAMILY_HEADINGS: every usage
+  synonym plus `Key API Summary`, `Pattern Surface` and `Key Exports`,
+  full text, case-insensitive, any level), fenced code inside them
+  included, plus every `*.md` file under `references/` at any depth, in
+  path order. Overlapping sections count each line once.
+  --body auto (the default) decides by the `references/` test alone.
+  coherence-check.md §2.4 also calls a package with `references/`
+  single-body when its `## Full*` sections carry real content rather than
+  stubs: that is the caller's judgment, passed as --body single.
+
+Export match:
+  The name as a fixed string, case-sensitive, never inside a longer
+  identifier: on a side where the name ends in a letter, digit, `_` or
+  `$`, the next character may not be one of those. So `$state` and `a.b`
+  match only themselves (no shell expansion, no regex wildcard), `get`
+  does not match `getAll`, and `.then` still matches `promise.then`.
+  `count` is the number of matching lines (the `grep -c` count).
+
+Citation match:
+  A skill's name and aliases, case-insensitive, never inside a longer
+  name: on a side where the term ends in a letter, digit, `_` or `-`, the
+  next character may not be one of those. So `React` and `react/client`
+  cite `react`, while `react-dom`, `preact` and `reactive` do not. The
+  whole SKILL.md is searched, frontmatter and code included, and a skill
+  never cites itself. A hit's `substring` is the match as written (the
+  longest term at the leftmost match of its line) and its `excerpt` that
+  line around all of its matches. A match is literal, not a judgment: a
+  skill named after a common word (`next`, `requests`) also matches
+  prose such as `the next step` or a `sending requests` description, so
+  the caller picks from `hits[]` the line it quotes as the citation.
+
+Reference extraction:
+  Outside the frontmatter and fenced code, two forms count:
+  - link: a markdown link, image or reference definition (`[id]: target`,
+    footnotes aside) outside inline code, unless its target is a URL
+    (`scheme:` or `//`), a bare `#anchor`, or holds a placeholder or glob
+    character (`{}<>*?$`). The target may hold one level of balanced
+    parentheses; its backslash escapes and entity references are
+    resolved, the `#fragment` and `?query` dropped and `%XX` escapes
+    decoded.
+  - mention: a `references/`, `scripts/` or `assets/` path, optionally
+    `./`-prefixed, in prose or inline code, not preceded by a letter,
+    digit, `_`, `-`, `.`, `/`, `@` or `:` (so `src/assets/x`, URLs and
+    `[SRC:scripts/x.sh:L1]` provenance citations are not references).
+    Trailing punctuation, a `#fragment` and a `:L12` line suffix are
+    dropped; a bare folder or a placeholder is not a reference. A
+    `scripts/` or `assets/` mention counts only when the package has that
+    folder or it sits in a section whose heading starts with `Scripts`
+    (the Scripts & Assets section): elsewhere it can name a path in the
+    reader's own project, as framework docs do.
+  A target under `scripts/` or `assets/` is `script-asset`; any other
+  target is `file-path`.
+  Known limits: a mention is taken as written (no escape is resolved) and
+  ends at whitespace, so `references/my file.md` in inline code reads as
+  `references/my`; parentheses nest one level deep in a link target; and
+  a one-word line such as `[Note]: see-below` is a reference definition,
+  as CommonMark reads it.
 
 Exit codes:
-  0 — operation succeeded
-  1 — user error (file not found, can't read)
+  0: operation succeeded
+  1: user error (file not found or unreadable, malformed JSON input, both
+     --skills and --pairs on stdin, a cross-reference --skills-root that
+     is not a directory)
+  2: usage error (argparse: missing or unknown argument or subcommand)
 """
 
 from __future__ import annotations
 
 import argparse
+import html
 import json
+import os
 import re
 import sys
 from pathlib import Path
+from urllib.parse import unquote
 
 
 # --------------------------------------------------------------------------
@@ -95,17 +308,19 @@ from pathlib import Path
 # --------------------------------------------------------------------------
 
 
-# These mirror the canonical synonyms documented in
-# `src/skf-test-skill/references/coherence-check.md` §2.1, with the
-# SKF-template-specific headings folded in so they are first-class
-# matches rather than literal-name misses (per the §2.1 "Note"
-# paragraph). The set covers the Deep/create-skill template
-# (`Quick Start`, `Common Workflows`, `Key API Summary`, `Key Types`),
-# the quick-skill template (`Usage Patterns`, `Key Exports`), and the
-# reference-app assembly overrides (`Adoption Steps` replaces Common
-# Workflows for usage; `Pattern Surface` replaces Key API Summary for
-# api_surface), since headings are matched on the full heading text,
-# not a substring.
+# The synonyms test-skill's coherence check (coherence-check.md §2.1)
+# matches, with the SKF-template headings folded in so they are
+# first-class matches rather than literal-name misses: the Deep/create-skill
+# template (`Overview`, `Quick Start`, `Common Workflows`, `Key API
+# Summary`), the quick-skill template (`Usage Patterns`, `Key Exports`),
+# the component-library override (`Component Catalog` replaces Key API
+# Summary for api_surface) and the reference-app overrides (`Adoption
+# Steps` replaces Common Workflows for usage; `Pattern Surface` replaces
+# Key API Summary for api_surface), since headings are matched on the full
+# heading text, not a substring. A template heading is never renamed: a
+# new one widens this list (test-skf-scan-skill-md-structure.py checks the
+# templates against it). A heading outside the list leaves its family
+# unsatisfied, and §2.1 then judges the `outline`.
 REQUIRED_SYNONYMS: dict[str, list[str]] = {
     "description": ["Description", "Overview", "Purpose", "Summary"],
     "usage": [
@@ -129,6 +344,7 @@ REQUIRED_SYNONYMS: dict[str, list[str]] = {
         "Reference",
         "Key API Summary",
         "Pattern Surface",
+        "Component Catalog",
     ],
 }
 
@@ -138,13 +354,39 @@ REQUIRED_SYNONYMS: dict[str, list[str]] = {
 # --------------------------------------------------------------------------
 
 
+# `scan` (table drift) and skf-shard-body.py read headings with this one.
 _HEADING_RE = re.compile(r"^\s*(#+)\s+(.*?)\s*$")
+_FRONTMATTER_DESCRIPTION_RE = re.compile(r"^description[ \t]*:(.*)$")
+# A YAML block scalar header (`>`, `|-`, `>2` ...) holds no text itself.
+_BLOCK_SCALAR_RE = re.compile(r"^[|>](?:[1-9][+-]?|[+-][1-9]?)?$")
+_EMPTY_SCALARS = frozenset({"", '""', "''", "~", "null"})
+
+
+def frontmatter_description(lines: list[str], start: int) -> bool:
+    """True when the frontmatter (`lines[:start]`) has a `description` key
+    with a non-empty value: inline, or on the indented lines under it."""
+    for i in range(1, max(start - 1, 1)):
+        m = _FRONTMATTER_DESCRIPTION_RE.match(lines[i])
+        if not m:
+            continue
+        value = m.group(1).split(" #", 1)[0].strip()
+        if value.lower() not in _EMPTY_SCALARS and not _BLOCK_SCALAR_RE.match(value):
+            return True
+        for nxt in lines[i + 1:start - 1]:
+            if not nxt.strip():
+                continue
+            if nxt[:1] not in (" ", "\t"):
+                break
+            return True
+        return False
+    return False
 
 
-def find_required_sections(text: str) -> dict[str, dict]:
+def find_required_sections(text: str) -> dict[str, object]:
     """For each family, find the first matching heading.
 
-    Walks every line once, lower-cases the heading text, and looks it up
+    Reads the body's headings once (the Line model: outside the frontmatter
+    and fenced code), lower-cases each heading text, and looks it up
     against pre-lowered synonym sets. Returns the structure described in
     the module docstring.
     """
@@ -155,13 +397,13 @@ def find_required_sections(text: str) -> dict[str, dict]:
         for syn in synonyms:
             lookup[syn.lower()] = (family, syn)
 
+    lines = _split_lines(text)
+    start = _frontmatter_end(lines)
+    code = _code_lines(lines, start)
+    headings = _body_headings(lines, start, code)
     matched: dict[str, str] = {}
-    for line in text.splitlines():
-        m = _HEADING_RE.match(line)
-        if not m:
-            continue
-        heading_text = m.group(2).strip().lower()
-        hit = lookup.get(heading_text)
+    for _, _, heading in headings:
+        hit = lookup.get(heading.lower())
         if hit is None:
             continue
         family, canonical = hit
@@ -170,7 +412,7 @@ def find_required_sections(text: str) -> dict[str, dict]:
             continue
         matched[family] = canonical
 
-    result: dict[str, dict] = {}
+    result: dict[str, object] = {}
     for family, synonyms in REQUIRED_SYNONYMS.items():
         if family in matched:
             result[family] = {
@@ -184,6 +426,11 @@ def find_required_sections(text: str) -> dict[str, dict]:
                 "matched_synonym": None,
                 "tried": list(synonyms),
             }
+    result["frontmatter_description"] = frontmatter_description(lines, start)
+    result["outline"] = [
+        {"line": idx + 1, "level": level, "text": heading}
+        for idx, level, heading in headings
+    ]
     return result
 
 
@@ -316,6 +563,636 @@ def find_table_drift(text: str) -> list[dict]:
     return findings
 
 
+# --------------------------------------------------------------------------
+# Line model shared by usage-scope, cross-reference, reference-check and
+# scan --required-sections
+# --------------------------------------------------------------------------
+
+
+_FENCE_RE = re.compile(r"^\s*(`{3,}|~{3,})(.*)$")
+# A CommonMark ATX heading. `_HEADING_RE` stays as it is for `scan` and
+# skf-shard-body.py, which import it.
+_BODY_HEADING_RE = re.compile(r"^ {0,3}(#{1,6})(?:[ \t]+(.*?))?[ \t]*$")
+_CLOSING_HASHES_RE = re.compile(r"(?:^|[ \t]+)#+$")
+
+
+def _split_lines(text: str) -> list[str]:
+    """Split on newlines only, so line numbers match `grep -n`.
+
+    `str.splitlines()` also breaks on form feeds and U+2028, which would
+    shift every later line number away from what an editor or grep shows.
+    """
+    lines = text.replace("\r\n", "\n").split("\n")
+    if lines and lines[-1] == "":
+        lines.pop()
+    return lines
+
+
+def _frontmatter_end(lines: list[str]) -> int:
+    """Index of the first body line, past a leading `---` ... `---` block."""
+    if lines and lines[0].strip() == "---":
+        for i in range(1, len(lines)):
+            if lines[i].strip() == "---":
+                return i + 1
+    return 0
+
+
+def _code_lines(lines: list[str], start: int = 0) -> list[bool]:
+    """Flag every fence line and every line between a fence and its close.
+
+    A backtick line whose info string holds a backtick is inline code, not
+    a fence (CommonMark), and an unclosed fence runs to the end of the file.
+    """
+    code = [False] * len(lines)
+    fence: tuple[str, int] | None = None
+    for i in range(start, len(lines)):
+        m = _FENCE_RE.match(lines[i])
+        if fence is None:
+            if m and not (m.group(1)[0] == "`" and "`" in m.group(2)):
+                fence = (m.group(1)[0], len(m.group(1)))
+                code[i] = True
+            continue
+        code[i] = True
+        if (
+            m
+            and m.group(1)[0] == fence[0]
+            and len(m.group(1)) >= fence[1]
+            and not m.group(2).strip()
+        ):
+            fence = None
+    return code
+
+
+def _body_headings(
+    lines: list[str], start: int, code: list[bool]
+) -> list[tuple[int, int, str]]:
+    """(index, level, text) of each heading outside frontmatter and code.
+
+    A `#` line indented four or more spaces is indented code, and a
+    closing sequence (`## Usage ##`) is not part of the text.
+    """
+    headings: list[tuple[int, int, str]] = []
+    for i in range(start, len(lines)):
+        if code[i]:
+            continue
+        m = _BODY_HEADING_RE.match(lines[i])
+        if m:
+            text = _CLOSING_HASHES_RE.sub("", m.group(2) or "").strip()
+            headings.append((i, len(m.group(1)), text))
+    return headings
+
+
+def _section_spans(
+    headings: list[tuple[int, int, str]], total: int, wanted
+) -> list[tuple[int, int, str]]:
+    """(start, end, heading) of each section whose heading `wanted` accepts.
+
+    A section runs from its heading line to the line before the next
+    heading of the same or a higher level, or to the end of the file;
+    `end` is exclusive.
+    """
+    spans: list[tuple[int, int, str]] = []
+    for pos, (idx, level, text) in enumerate(headings):
+        if not wanted(text):
+            continue
+        end = total
+        for next_idx, next_level, _ in headings[pos + 1:]:
+            if next_level <= level:
+                end = next_idx
+                break
+        spans.append((idx, end, text))
+    return spans
+
+
+def _bounded(term: str, word_chars: str) -> str:
+    """Regex for `term` as a fixed string that is never part of a longer word.
+
+    `word_chars` is a character-class body such as `\\w$`. The guard only
+    applies on a side where the term itself ends in such a character, so
+    `.then` still matches inside `promise.then`.
+    """
+    word = re.compile(f"[{word_chars}]")
+    body = re.escape(term)
+    if word.match(term[0]):
+        body = f"(?<![{word_chars}])" + body
+    if word.match(term[-1]):
+        body += f"(?![{word_chars}])"
+    return body
+
+
+# --------------------------------------------------------------------------
+# Usage scope: per-export usage counts
+# --------------------------------------------------------------------------
+
+
+# coherence-check.md §2.4 counts a split-body skill over every usage-family
+# heading: the usage synonyms above plus the api-surface headings that carry
+# call examples in the SKF templates. A single-body skill counts over its
+# first usage-synonym section alone, so an export does not count as used
+# because Key Exports lists it. Derived from REQUIRED_SYNONYMS so a usage
+# synonym added there joins both scopes.
+USAGE_FAMILY_HEADINGS: list[str] = [
+    *REQUIRED_SYNONYMS["usage"],
+    "Key API Summary",
+    "Pattern Surface",
+    "Key Exports",
+]
+
+# The --body choices after `auto` is resolved.
+_BODIES = ("single", "split")
+
+# Characters that continue an identifier in the languages SKF documents
+# (`$` for JS/TS names such as `$state`).
+_IDENT_CHARS = r"\w$"
+
+
+def usage_scope(
+    skill_text: str,
+    reference_texts: list[tuple[str, str]],
+    exports: list[dict],
+    kinds: set[str] | None = None,
+    skill_label: str = "SKILL.md",
+    body: str | None = None,
+) -> dict:
+    """Count the lines that use each export over the usage scope.
+
+    `reference_texts` holds (relative path, text) for each reference file,
+    in path order. `exports` holds {"name", "kind"} dicts (kind may be
+    None); `kinds`, when given, drops the exports whose kind is known and
+    not listed. `body` is "single" or "split"; None picks "split" when
+    there is a reference text and "single" otherwise, and a single-body
+    scope reads no reference text. Returns the structure described in the
+    module docstring, without `warnings` (the caller owns file reading).
+    """
+    if body is None:
+        body = "split" if reference_texts else "single"
+    if body not in _BODIES:
+        raise ValueError(f"body must be one of {_BODIES}, not {body!r}")
+    lines = _split_lines(skill_text)
+    start = _frontmatter_end(lines)
+    code = _code_lines(lines, start)
+    headings = _body_headings(lines, start, code)
+    if body == "single":
+        usage = {h.lower() for h in REQUIRED_SYNONYMS["usage"]}
+        spans = _section_spans(
+            headings, len(lines), lambda text: text.lower() in usage
+        )[:1]
+        reference_texts = []
+    else:
+        family = {h.lower() for h in USAGE_FAMILY_HEADINGS}
+        spans = _section_spans(
+            headings, len(lines), lambda text: text.lower() in family
+        )
+    in_scope = sorted({i for s, e, _ in spans for i in range(s, e)})
+
+    # (file, 1-based line, text) in scope order: SKILL.md first, then the
+    # reference files, so `first_hit` is deterministic.
+    corpus: list[tuple[str, int, str]] = [
+        (skill_label, i + 1, lines[i]) for i in in_scope
+    ]
+    for rel, text in reference_texts:
+        corpus.extend(
+            (rel, n, line) for n, line in enumerate(_split_lines(text), start=1)
+        )
+
+    selected = {
+        (e["name"], e["kind"])
+        for e in exports
+        if kinds is None or e["kind"] is None or e["kind"] in kinds
+    }
+    results: list[dict] = []
+    for name, kind in sorted(selected, key=lambda k: (k[0], k[1] or "")):
+        pattern = re.compile(_bounded(name, _IDENT_CHARS))
+        count = 0
+        first_hit = None
+        for label, lineno, line in corpus:
+            # the substring test is a cheap pre-filter; the regex decides
+            if name in line and pattern.search(line):
+                count += 1
+                if first_hit is None:
+                    first_hit = {"file": label, "line": lineno}
+        results.append({
+            "name": name,
+            "kind": kind,
+            "count": count,
+            "first_hit": first_hit,
+        })
+
+    return {
+        "scope": {
+            "body": body,
+            "sections": [
+                {"heading": text, "line": s + 1, "end_line": e}
+                for s, e, text in spans
+            ],
+            "reference_files": [rel for rel, _ in reference_texts],
+        },
+        "exports": results,
+        "zero_usage": [
+            {"name": r["name"], "kind": r["kind"]}
+            for r in results
+            if r["count"] == 0
+        ],
+    }
+
+
+# --------------------------------------------------------------------------
+# Cross-reference: literal citations between skills
+# --------------------------------------------------------------------------
+
+
+# Characters that continue a package or skill name (`react-dom` is not
+# `react`).
+_NAME_CHARS = r"\w-"
+_EXCERPT_CONTEXT = 60
+
+
+def _citation_pattern(terms: list[str]) -> re.Pattern[str]:
+    """One case-insensitive regex for a skill's terms, longest term first."""
+    ordered = sorted(terms, key=lambda t: (-len(t), t.lower()))
+    return re.compile(
+        "|".join(_bounded(t, _NAME_CHARS) for t in ordered), re.IGNORECASE
+    )
+
+
+def _excerpt(line: str, start: int, end: int) -> str:
+    """The line around [start, end), whitespace collapsed, `...` at a cut.
+
+    [start, end) runs from a line's first match to its last, so every
+    match on the line shows.
+    """
+    lo = max(0, start - _EXCERPT_CONTEXT)
+    hi = min(len(line), end + _EXCERPT_CONTEXT)
+    text = " ".join(line[lo:hi].split())
+    if lo > 0:
+        text = "..." + text
+    if hi < len(line):
+        text += "..."
+    return text
+
+
+def cross_references(
+    skills: list[dict], pairs: list[tuple[str, str]] | None = None
+) -> dict:
+    """Find where each skill's SKILL.md literally cites another skill.
+
+    `skills` holds {"name", "terms", "text"} dicts, `text` None when the
+    SKILL.md could not be read (the skill can still be cited). `pairs`,
+    when given, keeps only the citations between those unordered pairs and
+    adds `pairs[]`. Each citation lists every citing line in `hits`, since
+    the first one can be a common word rather than the citation. Returns
+    citations, edges, pairs and pair warnings.
+    """
+    patterns = {s["name"]: _citation_pattern(s["terms"]) for s in skills}
+    names = sorted(patterns)
+    wanted: set[tuple[str, str]] | None = None
+    if pairs is not None:
+        wanted = set()
+        for a, b in pairs:
+            wanted.update({(a, b), (b, a)})
+
+    citations: list[dict] = []
+    for source in sorted(skills, key=lambda s: s["name"]):
+        if source["text"] is None:
+            continue
+        lines = _split_lines(source["text"])
+        for target in names:
+            if target == source["name"]:
+                continue
+            if wanted is not None and (source["name"], target) not in wanted:
+                continue
+            pattern = patterns[target]
+            hits: list[dict] = []
+            for lineno, line in enumerate(lines, start=1):
+                matches = list(pattern.finditer(line))
+                if not matches:
+                    continue
+                hits.append({
+                    "line": lineno,
+                    "substring": matches[0].group(0),
+                    "excerpt": _excerpt(
+                        line, matches[0].start(), matches[-1].end()
+                    ),
+                })
+            if not hits:
+                continue
+            citations.append({
+                "from": source["name"],
+                "to": target,
+                "substring": hits[0]["substring"],
+                "line": hits[0]["line"],
+                "count": len(hits),
+                "excerpt": hits[0]["excerpt"],
+                "hits": hits,
+            })
+
+    result: dict = {
+        "citations": citations,
+        "edges": [[c["from"], c["to"]] for c in citations],
+    }
+    warnings: list[str] = []
+    if pairs is not None:
+        cited = {(c["from"], c["to"]) for c in citations}
+        rows: list[dict] = []
+        for a, b in sorted(pairs):
+            for name in (a, b):
+                if name not in patterns:
+                    warnings.append(
+                        f"{name}: named in --pairs but not in --skills"
+                    )
+            rows.append({
+                "library_a": a,
+                "library_b": b,
+                "a_cites_b": (a, b) in cited,
+                "b_cites_a": (b, a) in cited,
+                "cited": (a, b) in cited or (b, a) in cited,
+            })
+        result["pairs"] = rows
+    # dict.fromkeys keeps the first of repeated warnings, in order
+    result["warnings"] = list(dict.fromkeys(warnings))
+    return result
+
+
+# --------------------------------------------------------------------------
+# Reference check: file, script and asset references
+# --------------------------------------------------------------------------
+
+
+# A link title is "...", '...' or (...); a reference definition must end
+# after its title, so a prose line such as `[Note]: see below` is not one.
+_LINK_TITLE = r"(?:\s+(?:\"[^\"\n]*\"|'[^'\n]*'|\([^()\n]*\)))?"
+# A link destination is `<...>`, or a run with no whitespace whose
+# parentheses are backslash-escaped or balanced one level deep
+# (`references/a_(1).md`). A backslash always takes the next character.
+_DEST_CHAR = r"(?:\\\S|[^\s()\\])"
+_LINK_RE = re.compile(
+    rf"!?\[[^\]\n]*\]\(\s*(<[^>\n]*>|(?:{_DEST_CHAR}|\({_DEST_CHAR}*\))+)"
+    + _LINK_TITLE
+    + r"\s*\)"
+)
+_REF_DEFINITION_RE = re.compile(
+    r"^\s{0,3}\[([^\]\n]+)\]:\s*(<[^>\n]*>|\S+)" + _LINK_TITLE + r"\s*$"
+)
+_MENTION_RE = re.compile(
+    r"(?<![\w.:/@-])(?:\./)?(references|scripts|assets)/[^\s`'\"()|,;]+"
+)
+# Backslash-escaped ASCII punctuation, or an entity reference: CommonMark
+# resolves both inside a link destination.
+_ESCAPE_RE = re.compile(
+    r"\\([!-/:-@\[-`{-~])"
+    r"|&(?:#[0-9]{1,7}|#[xX][0-9a-fA-F]{1,6}|[A-Za-z][A-Za-z0-9]*);"
+)
+_URL_SCHEME_RE = re.compile(r"^[A-Za-z][A-Za-z0-9+.-]+:")
+_LINE_SUFFIX_RE = re.compile(r":L?\d+(?:-L?\d+)?$")
+_PLACEHOLDER_CHARS = frozenset("{}<>*?$")
+_MENTION_TRAILING = ".,:;!?*_~>]}"
+_PACKAGE_FOLDERS = ("scripts", "assets")
+
+
+def _inline_code_spans(line: str) -> list[tuple[int, int]]:
+    """[start, end) of each inline code span: a backtick run closed by an
+    equal run. An unmatched run is literal text."""
+    spans: list[tuple[int, int]] = []
+    i, n = 0, len(line)
+    while i < n:
+        if line[i] != "`":
+            i += 1
+            continue
+        j = i
+        while j < n and line[j] == "`":
+            j += 1
+        run = j - i
+        k = j
+        close = -1
+        while k < n:
+            if line[k] != "`":
+                k += 1
+                continue
+            m = k
+            while m < n and line[m] == "`":
+                m += 1
+            if m - k == run:
+                close = k
+                break
+            k = m
+        if close == -1:
+            i = j
+            continue
+        spans.append((i, close + run))
+        i = close + run
+    return spans
+
+
+def _mask(line: str, spans: list[tuple[int, int]]) -> str:
+    """Blank out each span, keeping every offset in place."""
+    chars = list(line)
+    for start, end in spans:
+        chars[start:end] = " " * (end - start)
+    return "".join(chars)
+
+
+def _unescape(text: str) -> str:
+    """Resolve backslash escapes and entity references in one pass, so an
+    escaped `\\&amp;` stays `&amp;`."""
+    return _ESCAPE_RE.sub(
+        lambda m: m.group(1) or html.unescape(m.group(0)), text
+    )
+
+
+def _clean_link(dest: str) -> str | None:
+    """The local path a link destination names, or None when it names none."""
+    if dest.startswith("<") and dest.endswith(">"):
+        dest = dest[1:-1].strip()
+    dest = _unescape(dest)
+    if (
+        not dest
+        or dest.startswith(("#", "//"))
+        or _URL_SCHEME_RE.match(dest)
+    ):
+        return None
+    dest = unquote(dest.split("#", 1)[0].split("?", 1)[0])
+    if not dest or any(c in _PLACEHOLDER_CHARS for c in dest):
+        return None
+    return dest
+
+
+def _clean_mention(token: str) -> str | None:
+    """The path a bare mention names, or None for a folder or placeholder."""
+    token = _LINE_SUFFIX_RE.sub("", token.rstrip(_MENTION_TRAILING))
+    token = token.split("#", 1)[0].rstrip(_MENTION_TRAILING)
+    if token.endswith("/") or any(c in _PLACEHOLDER_CHARS for c in token):
+        return None
+    return token
+
+
+def _reference_type(target: str) -> str:
+    rel = target
+    while rel.startswith("./"):
+        rel = rel[2:]
+    if rel.startswith(tuple(f"{folder}/" for folder in _PACKAGE_FOLDERS)):
+        return "script-asset"
+    return "file-path"
+
+
+def extract_references(
+    text: str, package_folders: frozenset[str] = frozenset()
+) -> list[dict]:
+    """List the file, script and asset references in a SKILL.md body.
+
+    `package_folders` names the folders among `scripts` and `assets` that
+    the skill package has; a bare mention of another one counts only
+    inside the Scripts & Assets section. Returns [{"line", "form",
+    "target"}] in document order (before any existence check).
+    """
+    lines = _split_lines(text)
+    start = _frontmatter_end(lines)
+    code = _code_lines(lines, start)
+    scripts_section = {
+        i
+        for s, e, _ in _section_spans(
+            _body_headings(lines, start, code),
+            len(lines),
+            lambda heading: heading.lower().startswith("scripts"),
+        )
+        for i in range(s, e)
+    }
+
+    refs: list[dict] = []
+    for idx in range(start, len(lines)):
+        if code[idx]:
+            continue
+        line = lines[idx]
+        # links never sit inside inline code; mentions may
+        plain = _mask(line, _inline_code_spans(line))
+        found: list[tuple[int, str, str | None]] = []
+        link_spans: list[tuple[int, int]] = []
+
+        definition = _REF_DEFINITION_RE.match(plain)
+        if definition and not definition.group(1).startswith("^"):
+            target = _clean_link(definition.group(2))
+            found.append((definition.start(2), "link", target))
+            link_spans.append((0, definition.end()))
+        for m in _LINK_RE.finditer(plain):
+            found.append((m.start(1), "link", _clean_link(m.group(1))))
+            link_spans.append(m.span())
+
+        for m in _MENTION_RE.finditer(_mask(line, link_spans)):
+            folder = m.group(1)
+            if (
+                folder in _PACKAGE_FOLDERS
+                and folder not in package_folders
+                and idx not in scripts_section
+            ):
+                continue
+            found.append((m.start(), "mention", _clean_mention(m.group(0))))
+
+        for _, form, target in sorted(found, key=lambda f: f[0]):
+            if target is not None:
+                refs.append({"line": idx + 1, "form": form, "target": target})
+    return refs
+
+
+def scripts_assets_section(text: str, skill_dir: Path) -> dict:
+    """The package's scripts/ and assets/ folders and its Scripts & Assets heading."""
+    lines = _split_lines(text)
+    start = _frontmatter_end(lines)
+    heading = next(
+        (
+            {"heading": t, "line": idx + 1}
+            for idx, _, t in _body_headings(lines, start, _code_lines(lines, start))
+            if t.lower().startswith("scripts")
+        ),
+        None,
+    )
+    folders = sorted(f for f in _PACKAGE_FOLDERS if (skill_dir / f).is_dir())
+    return {
+        "folders": folders,
+        "section": heading,
+        "missing": bool(folders) and heading is None,
+    }
+
+
+def _realpath(path: str) -> str:
+    try:
+        return os.path.realpath(path)
+    except (OSError, ValueError):
+        # a name the OS rejects (an embedded NUL, say) resolves nowhere
+        return os.path.abspath(path)
+
+
+def _within(path: str, root: str) -> bool:
+    path, root = os.path.normcase(path), os.path.normcase(root)
+    return path == root or path.startswith(root.rstrip(os.sep) + os.sep)
+
+
+def check_references(
+    refs: list[dict],
+    skill_dir: str,
+    source_root: str | None = None,
+    skills_root: str | None = None,
+) -> dict:
+    """Resolve each reference and check it exists inside an allowed root.
+
+    A source or skills root that is not a directory becomes a warning.
+    """
+    warnings = [
+        f"--{flag} is not a directory: {path}"
+        for flag, path in (
+            ("source-root", source_root),
+            ("skills-root", skills_root),
+        )
+        if path and not os.path.isdir(path)
+    ]
+    roots = [("skill", _realpath(skill_dir))]
+    if source_root:
+        roots.append(("source", _realpath(source_root)))
+    if skills_root:
+        roots.append(("skills", _realpath(skills_root)))
+    base = roots[0][1]
+
+    references: list[dict] = []
+    counts = {"total": 0, "ok": 0, "missing": 0, "escapes": 0}
+    seen: set[tuple[int, str]] = set()
+    for ref in refs:
+        canonical = _realpath(os.path.join(base, ref["target"]))
+        key = (ref["line"], os.path.normcase(canonical))
+        if key in seen:
+            continue
+        seen.add(key)
+        root = next(
+            (label for label, path in roots if _within(canonical, path)), None
+        )
+        if root is None:
+            status = "escapes"
+        elif os.path.exists(canonical):
+            status = "ok"
+        else:
+            status = "missing"
+        counts["total"] += 1
+        counts[status] += 1
+        references.append({
+            "line": ref["line"],
+            "type": _reference_type(ref["target"]),
+            "form": ref["form"],
+            "target": ref["target"],
+            "canonical": canonical,
+            "status": status,
+            "root": root,
+        })
+
+    labels = dict(roots)
+    return {
+        "roots": {
+            "skill": labels["skill"],
+            "source": labels.get("source"),
+            "skills": labels.get("skills"),
+        },
+        "references": references,
+        "counts": counts,
+        "warnings": warnings,
+    }
+
+
 # --------------------------------------------------------------------------
 # CLI
 # --------------------------------------------------------------------------
@@ -325,14 +1202,22 @@ def _read_text(path: Path) -> str:
     return path.read_text(encoding="utf-8")
 
 
+def _read_markdown(path: Path, errors: str = "strict") -> str:
+    """The file as UTF-8 for the line model: a leading byte order mark is
+    dropped, and every `\\r` is kept, since `read_text` would turn a lone
+    `\\r` into a line break that `grep -n` does not count."""
+    return path.read_bytes().decode("utf-8-sig", errors=errors)
+
+
 def _cmd_scan(args: argparse.Namespace) -> int:
     skill_md = Path(args.skill_md)
     if not skill_md.is_file():
         print(f"error: file not found: {skill_md}", file=sys.stderr)
         return 1
     try:
-        text = _read_text(skill_md)
-    except OSError as exc:
+        # the Line model's read, so an `outline` line matches `grep -n`
+        text = _read_markdown(skill_md)
+    except (OSError, UnicodeDecodeError) as exc:
         print(f"error: cannot read {skill_md}: {exc}", file=sys.stderr)
         return 1
 
@@ -345,6 +1230,7 @@ def _cmd_scan(args: argparse.Namespace) -> int:
             "fence_count": fence_count,
             "bare_opening_fences": bare,
             "table_drift": find_table_drift(text),
+            "scripts_assets": scripts_assets_section(text, skill_md.parent),
         }
 
     json.dump(payload, sys.stdout, indent=2)
@@ -352,13 +1238,332 @@ def _cmd_scan(args: argparse.Namespace) -> int:
     return 0
 
 
+class UserError(Exception):
+    """A caller-facing input error: exit code 1."""
+
+
+def _load_skill_md(raw: str) -> tuple[Path, str]:
+    skill_md = Path(raw)
+    if not skill_md.is_file():
+        raise UserError(f"file not found: {skill_md}")
+    try:
+        return skill_md, _read_markdown(skill_md)
+    except (OSError, UnicodeDecodeError) as exc:
+        raise UserError(f"cannot read {skill_md}: {exc}") from exc
+
+
+def _read_json(source: str, label: str):
+    """Parse JSON from a file path, or from stdin when `source` is '-'."""
+    if source == "-":
+        try:
+            text = sys.stdin.read()
+        except OSError as exc:
+            raise UserError(
+                f"failed to read {label} from stdin: {exc}"
+            ) from exc
+    else:
+        path = Path(source)
+        if not path.is_file():
+            raise UserError(f"{label} file not found: {path}")
+        try:
+            text = _read_text(path)
+        except (OSError, UnicodeDecodeError) as exc:
+            raise UserError(
+                f"failed to read {label} file {path}: {exc}"
+            ) from exc
+    try:
+        return json.loads(text)
+    except json.JSONDecodeError as exc:
+        raise UserError(f"malformed JSON in {label} input: {exc}") from exc
+
+
+def parse_exports(payload) -> list[dict]:
+    """Normalize the --exports JSON into [{"name", "kind"}] dicts.
+
+    Accepts a list of names or {"name", "kind"} objects, an object with an
+    `exports` list, or the validate-inventory.py result (`inventory`).
+    """
+    items = payload
+    if isinstance(payload, dict):
+        if "exports" in payload:
+            items = payload["exports"]
+        elif isinstance(payload.get("inventory"), dict):
+            items = payload["inventory"].get("exports")
+    if not isinstance(items, list):
+        raise UserError(
+            "--exports must be a JSON array, an object with an `exports` "
+            "array, or a validate-inventory.py result"
+        )
+    exports: list[dict] = []
+    for idx, item in enumerate(items):
+        if isinstance(item, str):
+            name, kind = item, None
+        elif isinstance(item, dict):
+            name, kind = item.get("name"), item.get("kind")
+        else:
+            raise UserError(
+                f"--exports[{idx}] must be a name or an object; got {item!r}"
+            )
+        # surrounding whitespace is never part of a name, and a stray space
+        # would make every line miss
+        if not isinstance(name, str) or not name.strip():
+            raise UserError(f"--exports[{idx}] needs a non-empty `name`")
+        if kind is not None and not isinstance(kind, str):
+            raise UserError(f"--exports[{idx}] `kind` must be a string")
+        kind = (kind or "").strip() or None
+        exports.append({"name": name.strip(), "kind": kind})
+    return exports
+
+
+def parse_kinds(raw: str) -> set[str]:
+    kinds = {k.strip() for k in raw.split(",") if k.strip()}
+    if not kinds:
+        raise UserError("--kinds needs at least one kind, e.g. function,method")
+    return kinds
+
+
+# The names an skf-enumerate-stack-skills.py inventory entry gives the
+# repository and the folder its skill was built from.
+SOURCE_ALIAS_KEYS = ("source_repo_basename", "source_root_basename")
+
+
+def _source_aliases(item: dict) -> list[str]:
+    """The entry's SOURCE_ALIAS_KEYS values that are non-empty strings."""
+    found: list[str] = []
+    for key in SOURCE_ALIAS_KEYS:
+        value = item.get(key)
+        if isinstance(value, str) and value.strip():
+            found.append(value.strip())
+    return found
+
+
+def parse_skill_entries(payload) -> list[dict]:
+    """Normalize the --skills JSON into [{"name", "file", "terms"}] dicts.
+
+    `file` is the SKILL.md path as given (`skill_md`, else `path` joined
+    with SKILL.md); `terms` is the name plus its aliases, case-insensitive
+    duplicates dropped. An entry's source basenames (SOURCE_ALIAS_KEYS)
+    are aliases too, unless another entry has the same term as its name,
+    an alias or a source basename: one word would then cite two skills.
+    """
+    items = payload.get("skills") if isinstance(payload, dict) else payload
+    if not isinstance(items, list):
+        raise UserError(
+            "--skills must be a JSON array of skill entries or an object "
+            "with a `skills` array"
+        )
+    parsed: list[tuple[str, str, list[str], list[str]]] = []
+    seen: set[str] = set()
+    for idx, item in enumerate(items):
+        if not isinstance(item, dict):
+            raise UserError(f"--skills[{idx}] must be an object; got {item!r}")
+        name = item.get("name")
+        if not isinstance(name, str) or not name.strip():
+            raise UserError(f"--skills[{idx}] needs a non-empty `name`")
+        name = name.strip()
+        if name in seen:
+            raise UserError(f"--skills names `{name}` twice")
+        seen.add(name)
+        skill_md, folder = item.get("skill_md"), item.get("path")
+        if isinstance(skill_md, str) and skill_md:
+            file = skill_md
+        elif isinstance(folder, str) and folder:
+            file = os.path.join(folder, "SKILL.md")
+        else:
+            raise UserError(
+                f"--skills[{idx}] ({name}) needs `skill_md` or `path`"
+            )
+        aliases = item.get("aliases") or []
+        if not isinstance(aliases, list) or not all(
+            isinstance(a, str) for a in aliases
+        ):
+            raise UserError(
+                f"--skills[{idx}] ({name}) `aliases` must be strings"
+            )
+        parsed.append((name, file, aliases, _source_aliases(item)))
+
+    # How many entries hold each term, case-insensitively.
+    owners: dict[str, int] = {}
+    for name, _file, aliases, source in parsed:
+        held = {t.strip().lower() for t in (name, *aliases, *source) if t.strip()}
+        for term in held:
+            owners[term] = owners.get(term, 0) + 1
+
+    entries: list[dict] = []
+    for name, file, aliases, source in parsed:
+        unique = [term for term in source if owners[term.lower()] == 1]
+        terms: dict[str, str] = {}
+        for term in (name, *aliases, *unique):
+            term = term.strip()
+            if term:
+                terms.setdefault(term.lower(), term)
+        entries.append(
+            {"name": name, "file": file, "terms": list(terms.values())}
+        )
+    return entries
+
+
+def parse_pairs(payload) -> list[tuple[str, str]]:
+    """Normalize the --pairs JSON into unique (a, b) tuples."""
+    items = payload.get("pairs") if isinstance(payload, dict) else payload
+    if not isinstance(items, list):
+        raise UserError(
+            "--pairs must be a JSON array of pairs or an object with a "
+            "`pairs` array"
+        )
+    pairs: list[tuple[str, str]] = []
+    seen: set[frozenset[str]] = set()
+    for idx, item in enumerate(items):
+        if isinstance(item, list) and len(item) == 2:
+            a, b = item
+        elif isinstance(item, dict):
+            a, b = item.get("library_a"), item.get("library_b")
+        else:
+            raise UserError(
+                f"--pairs[{idx}] must be [a, b] or "
+                f"{{\"library_a\", \"library_b\"}}; got {item!r}"
+            )
+        if not (isinstance(a, str) and a and isinstance(b, str) and b):
+            raise UserError(f"--pairs[{idx}] needs two non-empty skill names")
+        if a == b:
+            raise UserError(f"--pairs[{idx}] pairs `{a}` with itself")
+        key = frozenset((a, b))
+        if key not in seen:
+            seen.add(key)
+            pairs.append((a, b))
+    return pairs
+
+
+def _reference_paths(skill_dir: Path) -> list[Path]:
+    """Every `references/**/*.md` under the skill folder, in path order."""
+    ref_dir = skill_dir / "references"
+    if not ref_dir.is_dir():
+        return []
+    paths = [
+        Path(root) / name
+        for root, _dirs, files in os.walk(ref_dir)
+        for name in files
+        if name.lower().endswith(".md")
+    ]
+    return sorted(paths, key=lambda p: p.relative_to(skill_dir).as_posix())
+
+
+def _read_reference_files(
+    skill_dir: Path, paths: list[Path]
+) -> tuple[list[tuple[str, str]], list[str]]:
+    """(relative path, text) of each reference file, in the order given.
+
+    A file that cannot be read becomes a warning; undecodable bytes are
+    replaced, so one bad byte never hides a whole file from the count.
+    """
+    texts: list[tuple[str, str]] = []
+    warnings: list[str] = []
+    for path in paths:
+        rel = path.relative_to(skill_dir).as_posix()
+        try:
+            texts.append((rel, _read_markdown(path, errors="replace")))
+        except OSError as exc:
+            warnings.append(f"{rel}: cannot read: {exc}")
+    return texts, warnings
+
+
+def _emit(payload: dict) -> int:
+    json.dump(payload, sys.stdout, indent=2)
+    sys.stdout.write("\n")
+    return 0
+
+
+def _cmd_usage_scope(args: argparse.Namespace) -> int:
+    skill_md, text = _load_skill_md(args.skill_md)
+    exports = parse_exports(_read_json(args.exports, "--exports"))
+    kinds = parse_kinds(args.kinds) if args.kinds is not None else None
+    paths = _reference_paths(skill_md.parent)
+    body = args.body
+    if body == "auto":
+        # a reference file that cannot be read still makes the body split
+        body = "split" if paths else "single"
+    references: list[tuple[str, str]] = []
+    warnings: list[str] = []
+    if body == "split":
+        references, warnings = _read_reference_files(skill_md.parent, paths)
+    payload = usage_scope(
+        text,
+        references,
+        exports,
+        kinds=kinds,
+        skill_label=skill_md.name,
+        body=body,
+    )
+    payload["warnings"] = warnings
+    return _emit(payload)
+
+
+def _cmd_cross_reference(args: argparse.Namespace) -> int:
+    if args.skills == "-" and args.pairs == "-":
+        raise UserError(
+            "--skills and --pairs cannot both read from stdin ('-')"
+        )
+    # a wrong root would turn every skill into a not-found warning and
+    # leave nothing to cite, which reads like a clean result
+    if args.skills_root is not None and not Path(args.skills_root).is_dir():
+        raise UserError(
+            f"--skills-root is not a directory: {args.skills_root}"
+        )
+    entries = parse_skill_entries(_read_json(args.skills, "--skills"))
+    pairs = None
+    if args.pairs is not None:
+        pairs = parse_pairs(_read_json(args.pairs, "--pairs"))
+
+    skills: list[dict] = []
+    warnings: list[str] = []
+    for entry in entries:
+        skill_md = Path(entry["file"])
+        if args.skills_root and not skill_md.is_absolute():
+            skill_md = Path(args.skills_root) / skill_md
+        text = None
+        if not skill_md.is_file():
+            warnings.append(f"{entry['name']}: SKILL.md not found: {skill_md}")
+        else:
+            try:
+                text = _read_markdown(skill_md)
+            except (OSError, UnicodeDecodeError) as exc:
+                warnings.append(
+                    f"{entry['name']}: cannot read {skill_md}: {exc}"
+                )
+        skills.append(
+            {"name": entry["name"], "terms": entry["terms"], "text": text}
+        )
+
+    payload = cross_references(skills, pairs)
+    payload["warnings"] = warnings + payload["warnings"]
+    return _emit(payload)
+
+
+def _cmd_reference_check(args: argparse.Namespace) -> int:
+    skill_md, text = _load_skill_md(args.skill_md)
+    skill_dir = skill_md.parent
+    package_folders = frozenset(
+        folder for folder in _PACKAGE_FOLDERS if (skill_dir / folder).is_dir()
+    )
+    refs = extract_references(text, package_folders)
+    payload = check_references(
+        refs, str(skill_dir), args.source_root, args.skills_root
+    )
+    warnings = payload.pop("warnings")
+    payload["scripts_assets"] = scripts_assets_section(text, skill_dir)
+    payload["warnings"] = warnings
+    return _emit(payload)
+
+
 def _build_parser() -> argparse.ArgumentParser:
     parser = argparse.ArgumentParser(
         prog="skf-scan-skill-md-structure",
         description=(
             "Deterministic structural scans for SKILL.md: fence balance, "
-            "bare opening fences, table column drift, and required-section "
-            "presence (case-insensitive synonym match)."
+            "bare opening fences, table column drift, required-section "
+            "presence (case-insensitive synonym match), export usage counts "
+            "over the usage sections, literal citations between skills, "
+            "and file, script and asset reference checks."
         ),
     )
     sub = parser.add_subparsers(dest="cmd", required=True)
@@ -372,13 +1577,112 @@ def _build_parser() -> argparse.ArgumentParser:
     )
     p_scan.set_defaults(func=_cmd_scan)
 
+    p_usage = sub.add_parser(
+        "usage-scope",
+        help=(
+            "count the lines that use each export in the usage sections "
+            "(and references/**/*.md for a split-body skill)"
+        ),
+    )
+    p_usage.add_argument("skill_md", help="path to a SKILL.md file")
+    p_usage.add_argument(
+        "--exports",
+        required=True,
+        help=(
+            "path to the exports JSON, or '-' for stdin. Shape: "
+            '["<name>", ...], [{"name": "...", "kind": "..."}, ...], '
+            '{"exports": [...]} or {"inventory": {"exports": [...]}}'
+        ),
+    )
+    p_usage.add_argument(
+        "--kinds",
+        help=(
+            "comma-separated kinds to count, e.g. function,method; an "
+            "export given without a kind is always counted"
+        ),
+    )
+    p_usage.add_argument(
+        "--body",
+        choices=("auto", *_BODIES),
+        default="auto",
+        help=(
+            "single: the first usage-synonym section; split: every "
+            "usage-family section plus references/**/*.md; auto (the "
+            "default): split when the package has a references/**/*.md "
+            "file, else single"
+        ),
+    )
+    p_usage.set_defaults(func=_cmd_usage_scope)
+
+    p_xref = sub.add_parser(
+        "cross-reference",
+        help=(
+            "list the literal citations of each skill's name or aliases "
+            "in the other skills' SKILL.md"
+        ),
+    )
+    p_xref.add_argument(
+        "--skills",
+        required=True,
+        help=(
+            "path to the skills JSON, or '-' for stdin. Shape: "
+            '[{"name": "...", "skill_md" | "path": "...", "aliases": [...]}] '
+            'or {"skills": [...]} (the skf-enumerate-stack-skills.py output, '
+            "whose source basenames a single skill holds count as aliases)"
+        ),
+    )
+    p_xref.add_argument(
+        "--skills-root",
+        help=(
+            "folder that relative `path` and `skill_md` entries resolve "
+            "against"
+        ),
+    )
+    p_xref.add_argument(
+        "--pairs",
+        help=(
+            "path to a pairs JSON, or '-' for stdin, that limits the "
+            'citations to those pairs. Shape: [["<a>", "<b>"], ...], '
+            '[{"library_a": "...", "library_b": "..."}] or {"pairs": [...]}'
+        ),
+    )
+    p_xref.set_defaults(func=_cmd_cross_reference)
+
+    p_refs = sub.add_parser(
+        "reference-check",
+        help=(
+            "check that each file, script and asset reference exists and "
+            "resolves inside the allowed roots"
+        ),
+    )
+    p_refs.add_argument("skill_md", help="path to a SKILL.md file")
+    p_refs.add_argument(
+        "--source-root",
+        help=(
+            "the skill's extraction tree (metadata.json source_root), a "
+            "second allowed root"
+        ),
+    )
+    p_refs.add_argument(
+        "--skills-root",
+        help=(
+            "the skills output folder, an allowed root for a stack skill "
+            "whose references reach its constituent skills"
+        ),
+    )
+    p_refs.set_defaults(func=_cmd_reference_check)
+
     return parser
 
 
 def main(argv: list[str] | None = None) -> int:
     parser = _build_parser()
     args = parser.parse_args(argv)
-    return args.func(args)
+    try:
+        return args.func(args)
+    except UserError as exc:
+        print(f"error: {exc}", file=sys.stderr)
+        return 1
 
 
 if __name__ == "__main__":