Defame1297/holocron
0
0
Fork 0
Code Issues 14 Pull Requests Actions Packages Projects Releases 3 Wiki Activity
Files
807caf22eec39c2b254171444f9a606cf7c11dc8
holocron/scripts/skill-size-check.sh
Defame1297 1d40544075 fix(gates): hold skill versions above main's tip as well as the merge-base 
Why: two branches that both bump a skill 1.0.0 -> 1.0.1 with different
content merge without a conflict, and each passed the gate against its own
merge-base, so main could ship two changes under one version.

Implementation Notes:
- check-skill-version-bump requires the pushed version to exceed both the
  merge-base and the main tip; failures name the baseline they missed.
- Presence is read from the tree, so a blob missing from a partial clone is
  a read failure instead of a silently exempt "new" skill.
- A leading UTF-8 BOM no longer reads as a missing version.
- Version parts reject leading zeros in all three validators
  (check-skill-version-bump, skill-size-check, factory-audit).
- New tests cover equal bumps, moved files, major/minor ordering, bad refs,
  unreadable blobs, mode-only changes, symlinks and tag peeling.

Impact: ADR-0022 amended (reverses "not main's current tip"); gates.md
updated to match, including pre-commit 4.6.1's exact ref selection.

ADR: 0022
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-16 11:24:08 +00:00

1526 lines
79 KiB
Bash
Executable File
Raw Blame History

This file contains invisible Unicode characters
This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
#!/usr/bin/env bash
set -euo pipefail
# Two independent gate families live in this script. Do not conflate them.
#
# 1. SPEC CONFORMANCE (MAX_LINES / MAX_WORDS, unchanged). Enforces
# agentskills.io's skill-authoring.md guidance: keep SKILL.md within 500
# lines and roughly 5,000 tokens, so the full body doesn't crowd out
# conversation history and other active skills once loaded into context.
# MAX_WORDS counts the WHOLE FILE, frontmatter included.
#
# 2. CONTEXT BUDGET (ADR-0020: description chars, body-only words, resolvable
# boundary targets). A skill's name + description is preloaded into every
# agent's context every session whether or not the skill is ever invoked;
# the body is loaded only on invocation, and then competes with the caller's
# live conversation. Those are different costs with different ceilings, so
# they get their own numbers and their own measurements.
#
# The two families measure different things on purpose and neither replaces the
# other: 2,770 whole-file words is a conformance backstop, 900 body-only words
# is a quality gate, and a file can sit well inside one while failing the other.
#
# Vale can't express any of this (its checks operate on text patterns, not raw
# file size), so this is a plain script instead of a Vale rule.
#
# Both spec ceilings are inclusive: a file at exactly MAX_LINES or MAX_WORDS
# passes, and only one past it fails. That matches
# factory-audit/scripts/lib-checks-skill.sh, which has always used `line_count <= 500` as
# its pass condition — the two previously disagreed at exactly 500 lines, so a
# SKILL.md could pass its own audit and still be blocked by the commit hook.
# The ADR-0020 ceilings are inclusive the same way.
#
# Token counts aren't computed exactly here — a whitespace word count is used
# as a proxy (Python's str.split(), the same primitive
# factory-audit/scripts/lib-checks-skill.sh applies to these two constants; `wc -w`
# disagrees with it on Unicode separators, which is why the awk pass that used
# to live in the loop below is gone).
#
# THE MEASUREMENT BASIS, stated because the previous re-measure drifted onto a
# different one and the numbers moved without the prose noticing: characters
# per word is len(text) / len(text.split()) over the WHOLE FILE, whitespace
# included, on plugins/*/.apm/skills/*/SKILL.md. Counting only non-whitespace
# characters gives a materially lower figure (4.90 / 5.52 / 5.54 / 6.19 today)
# and is not the basis MAX_WORDS is calibrated against.
#
# Measured over this repo's 39 in-scope SKILL.md files (2026-08-31, after the
# ADR-0020 retrofit), characters per word runs min 5.93 / median 6.67 / mean
# 6.63 / max 7.34. At the standard ~4-characters-per-token English
# approximation that is 1.48 / 1.67 / 1.66 / 1.84 tokens per word.
#
# MAX_WORDS=2770 is therefore calibrated to the corpus WORST case rather than
# its median: 2770 words at the densest observed 7.34 chars/word is ~20,300
# characters, or ~5,090 tokens at the 4-characters-per-token approximation. So
# what this gate guarantees is "about 5,000 tokens even for the densest prose
# the corpus has produced" — the earlier median-calibrated MAX_WORDS=2900 let
# such a file sit at exactly the ceiling and spend ~5,320 tokens. A
# median-density file at 2770 words spends ~4,620 tokens, so typical prose
# gives up ~140 words of headroom to close that gap. Densest file today:
# git-commits at 7.34 chars/word.
#
# THE CORPUS IS NOWHERE NEAR THIS CEILING ANY MORE, and the note that used to
# stand here — "a gate two files have already grown into" — described the
# pre-retrofit corpus and is now wrong by a factor of three. The largest
# SKILL.md is write-docs at 914 whole-file words, then vale-run at 874;
# skill-author, the old high-water mark at 2,760, is down to 661. MAX_WORDS is
# a spec-conformance backstop with roughly 1,850 words of slack, and the gate
# that actually bites is ADR-0020's 900-word body budget below it. Do not read
# the two as redundant: they measure different spans, and a file can sit well
# inside one while failing the other.
#
# It is a one-sided proxy in the useful direction — nothing under the word
# ceiling is wildly over the token ceiling — but it is not exact BPE
# tokenization and does not replace one. Re-measure the corpus, on the basis
# stated above, before treating any of these numbers as still current.
#
# python3 AND PyYAML are required for the ADR-0020 half, and both are hard
# dependencies rather than best-effort: python3 because pre-commit (which is how
# this script runs) is itself a Python application, and PyYAML because the
# hand-rolled folding reader that used to cover its absence disagreed with a
# real parser across the FAIL boundary. Two readers that measure the same
# description differently is worse than one reader that refuses to start.
# These constants are intentionally duplicated in
# factory-audit/scripts/lib-checks-skill.sh (Python) rather than shared from one
# file: this script is a standalone bash pre-commit hook, that one is an
# in-skill Python check library invoked in a different context.
# tests/test-skill-size-check.sh asserts both files agree on these values, so
# drift between them fails CI rather than silently diverging.
#
# The ADR-0020 constants below are duplicated the same way and carry the same
# warning: factory-audit/scripts/lib-checks-skill.sh holds a second copy of
# DESC_SUGGEST_CHARS / DESC_MAX_CHARS / BODY_SUGGEST_WORDS / BODY_MAX_WORDS,
# and factory-audit/scripts/lib-checks-agent.sh holds a third copy of the two
# description constants (agents take the description gates and, per ADR-0020,
# deliberately take NO body word gate). ADR-0025's merge put those two libraries
# in one directory but did NOT collapse the copies — the two flows gate
# different spans — so the drift risk is unchanged: if they diverge, the audit
# reports a skill ready to ship that the commit hook then rejects.
MAX_LINES=500
MAX_WORDS=2770
# ADR-0020 context-budget gates. SUGGESTION does not fail; FAIL does.
DESC_SUGGEST_CHARS=250
DESC_MAX_CHARS=400
BODY_SUGGEST_WORDS=600
BODY_MAX_WORDS=900
FAIL=0
# ZERO ARGUMENTS IS A USAGE ERROR, exit 2 — not a clean run.
#
# This hook is `pass_filenames: true` in both .pre-commit-config.yaml and
# .pre-commit-hooks.yaml, and pre-commit skips a filename-passing hook entirely
# when nothing matches its `files:` pattern, so it never invokes this script
# with an empty argument list. Every no-argument invocation therefore comes from
# somewhere else — a hand-run command, a wrapper, or a `files:` pattern edited
# into matching nothing — and printing nothing and exiting 0 made all three
# indistinguishable from a clean corpus. A mis-scoped pattern would have
# silently disabled the whole ADR-0020 gate family while every hook reported
# green.
#
# Exit 2, not 1, for the same reason a8cd5e8 split validate-adapter.sh's usage
# exits out: {0,1} are this script's verdict codes (clean / findings), and a
# caller that reads a non-zero exit as "the SKILL.md needs editing" must be able
# to tell a broken invocation from a real finding.
if [[ $# -eq 0 ]]; then
echo "usage: skill-size-check.sh <SKILL.md> [SKILL.md ...]" >&2
echo " Measures the agentskills.io spec ceilings and the ADR-0020 context" >&2
echo " budget for each SKILL.md named on the command line." >&2
echo " No paths were given. This is a usage error, not a clean run: a hook" >&2
echo " whose files: pattern matches nothing would otherwise be" >&2
echo " indistinguishable from a corpus with no findings." >&2
exit 2
fi
# An unreadable path — a broken symlink named SKILL.md is storable in git, and a
# directory named SKILL.md reaches this hook the same way — is diagnosed ONCE,
# in the Python per-file loop below. There used to be a bash pre-loop here doing
# exactly the same stat dance and printing exactly the same sentence, so every
# such path was reported twice with two ERROR lines for one broken file. It is
# still NOT a silent skip; the diagnosis simply lives where the file is read.
#
# The MAX_LINES / MAX_WORDS ceilings are not measured in bash either. They used
# to be, in a single awk pass, and that pass was wrong twice over:
# * `read -r lines words <<< "$(awk ...)"` discarded awk's exit status, so a
# file awk could not read yielded empty variables, bash arithmetic read
# them as 0, and both ceilings passed in total silence — the one outcome
# this script forbids itself.
# * awk's NR/NF do not agree with the Python splitlines()/split() that
# factory-audit/scripts/lib-checks-skill.sh uses for the SAME two constants.
# splitlines() also breaks on \x0b \x0c \x1c \x1d \x1e \x85 U+2028 U+2029
# and split() on every Unicode space, so a body padded with U+2028 read as
# 6 lines here and 606 lines there — hook green, audit FAIL.
# One implementation now owns both: the Python block below already reads every
# file (with a real diagnostic on failure), so it counts there.
if ! command -v python3 > /dev/null 2>&1; then
echo "ERROR: python3 is required for the ADR-0020 description/body/boundary-target gates but was not found on PATH." >&2
echo " Why: skipping them would be a vacuous pass — the hook would go green having checked only the spec ceilings." >&2
echo " Fix: install python3 (pre-commit itself is a Python application, so it is almost certainly already present)." >&2
exit 1
fi
if ! python3 -c 'import yaml' > /dev/null 2>&1; then
echo "ERROR: PyYAML is required for the ADR-0020 description/body/boundary-target gates but is not importable by python3." >&2
echo " Why: the description VALUE has to be measured after YAML folding is resolved, and the hand-rolled reader that used to stand in for PyYAML disagreed with it across the 400-character FAIL boundary. Falling back would make the verdict depend on which reader ran." >&2
echo " Fix: python3 -m pip install PyYAML (or your distro's python3-yaml package)." >&2
exit 1
fi
if ! python3 -u - \
"$DESC_SUGGEST_CHARS" "$DESC_MAX_CHARS" \
"$BODY_SUGGEST_WORDS" "$BODY_MAX_WORDS" "$MAX_WORDS" "$MAX_LINES" "$@" <<'PYTHON'
import glob
import os
import re
import sys
import yaml
# Output is UTF-8 for the same reason input is: under LC_ALL=C the streams
# default to ASCII, and this script's own message text carries em dashes (the
# ADR-0020 boundary SUGGESTION is one). Pinning only the reads moved the crash
# from the read to the write — a UnicodeEncodeError raised while PRINTING, after
# every check has already run, which loses the whole report. read_text() in the
# shared resolver block below pins the reads; this pins the writes.
#
# Deliberately OUTSIDE the ADR-0020 shared boundary resolver block: all three
# scripts in the family need this, but tests/test-adr0020-contract.sh hashes
# that block for byte-identity, so shared-looking edits belong beside it, not
# inside it.
for _stream in (sys.stdout, sys.stderr):
try:
_stream.reconfigure(encoding='utf-8')
except AttributeError: # pragma: no cover — Python < 3.7
pass
DESC_SUGGEST_CHARS = int(sys.argv[1])
DESC_MAX_CHARS = int(sys.argv[2])
BODY_SUGGEST_WORDS = int(sys.argv[3])
BODY_MAX_WORDS = int(sys.argv[4])
MAX_WORDS = int(sys.argv[5])
MAX_LINES = int(sys.argv[6])
files = sys.argv[7:]
failed = False
def error(msg):
global failed
failed = True
print("ERROR: %s" % msg, file=sys.stderr)
def suggest(msg):
# stdout, not stderr, and never touches the exit code. The hook is declared
# `verbose: true` in .pre-commit-config.yaml so this actually reaches a
# human — pre-commit prints nothing at all for a passing hook otherwise,
# which is the exact way ADR-0013 records Vale warnings going invisible.
print("SUGGESTION: %s" % msg)
def info(msg):
# A check that DECLINED to run says so out loud. The one thing this script
# must never do is stay quiet about a measurement it did not take.
print("INFO: %s" % msg)
# Anything inside the BEGIN/END markers below is hashed byte-for-byte against
# the plugin copy by tests/test-adr0020-contract.sh. Never edit the marked span
# in one file alone -- including its comments -- or that test fails. Change both
# copies in one commit, keeping the span's line count intact.
# ===== BEGIN ADR-0020 SHARED BOUNDARY RESOLVER =====
# ONE resolver, embedded VERBATIM in two scripts (ADR-0025 retired the third):
# scripts/skill-size-check.sh
# plugins/kyberforge/.apm/skills/factory-audit/scripts/lib-boundary-resolver.sh
# The block between these markers must stay byte-identical in both. It is copied
# rather than imported because a cache-installed plugin's scripts cannot read
# files outside their own plugin directory, and this repo-root hook resolves via
# .pre-commit-hooks.yaml, where entry[0] is the only token pre-commit rewrites --
# so no single file is reachable by both (the same constraint that duplicates the
# ADR-0020 constants). Edit one copy, then paste it over the other.
#
# Requires: glob, os, re, yaml (imported by the host script; PyYAML is a hard
# dependency, preflighted in bash before the interpreter starts).
# --- Input ----------------------------------------------------------------
# Every file this resolver's callers read goes through read_text(), which pins
# UTF-8 explicitly instead of inheriting locale.getpreferredencoding(). Under
# LC_ALL=C that inherited encoding is ASCII, so a perfectly ordinary em dash in
# a SKILL.md aborted the run with a bare UnicodeDecodeError traceback — loud,
# but pointing at the interpreter rather than at the file or the fix. A file
# that genuinely is not UTF-8 still fails; it just says so.
class EncodingError(Exception):
pass
def read_text(path):
"""File contents as text, UTF-8, with a diagnostic instead of a traceback."""
try:
with open(path, encoding='utf-8') as fh:
return fh.read()
except UnicodeDecodeError as exc:
raise EncodingError(
"not valid UTF-8 (%s at byte %d) — re-save the file as UTF-8; "
"this gate does not guess at other encodings"
% (exc.reason, exc.start))
# --- Universe ------------------------------------------------------------
# The set of names a boundary clause may resolve against is derived from an
# AUTHORING ROOT found by walking up FROM THE TARGET FILE. It is NEVER derived
# from this script's own location: deriving it from ${BASH_SOURCE} leaked
# holocron's whole skill universe into every consumer repo that ran this hook
# through pre-commit, so a consumer skill routing to a holocron skill such as
# `factory-audit` resolved against a plugin it had never installed.
#
# An authoring root is the nearest ancestor holding plugins/*/.apm/skills/ or
# plugins/*/.apm/agents/ (a plugin monorepo), falling back to the nearest
# ancestor holding .git. When one is found the universe is:
# 1. every skill and agent under <root>/plugins/*/ — sibling plugins resolve,
# which is what a monorepo means,
# 2. the target's own apm package,
# 3. the packages that package DECLARES in apm.yml dependencies.apm.
# Deployed .claude/ and .agents/ trees are deliberately NOT consulted when the
# root came from the plugins/ probe. They are `apm install` output, gitignored,
# and present only on a machine that has run it: four cross-plugin targets in
# this repo (gitea-branches -> git-branches, gitea-branches -> git-history,
# gitea-issues -> git-branches, gitea-workflow -> git-workflow) resolved through
# .claude/skills/ alone, so the same commit measured 2 dangling targets on a
# developer machine and 6 on a fresh clone. A gate shipping hot with no baseline
# cannot give two answers.
#
# Deployed trees ARE used when no plugin monorepo was found — whether the walk
# landed on a bare .git ancestor or on nothing at all. That is the consumer
# case: the file being checked lives in or beside a deployed tree, inside an
# ordinary git repo, with no monorepo to read. The two cases are told apart by
# which probe matched, never by how many names a root contributed; see
# known_targets().
def _is_fs_root(path):
return os.path.dirname(path) == path
def _collect_package(pkg_dir, names):
"""Add every skill/agent name a package directory exposes, any layout."""
# glob.escape() the DIRECTORY only. A checkout path containing `[`, `]`,
# `*` or `?` — a worktree named `feature[2]`, say — otherwise turns the
# whole pattern into a character class that matches nothing, and the
# resolver degrades to the "DID NOT RUN" INFO with rc=0 across every file
# in the tree. The wildcards in `sub` are the intended ones and stay raw.
safe_dir = glob.escape(pkg_dir)
for sub in ('.apm/skills/*/', 'skills/*/'):
for path in glob.glob(os.path.join(safe_dir, sub)):
# A directory is a skill only if it HOLDS a SKILL.md. An empty
# leftover — a deleted skill whose directory survived, a scaffolding
# stub, an editor's stray mkdir — is untracked by git, so it exists
# on the machine that made it and nowhere else. Counting it made a
# boundary target resolve locally and dangle in a fresh clone: the
# same install-dependence the deployed-tree rule above exists to
# remove, arriving through a different door.
if os.path.isfile(os.path.join(path, 'SKILL.md')):
names.add(os.path.basename(path.rstrip('/')).lower())
for sub in ('.apm/agents/*.md', 'agents/*.md'):
for path in glob.glob(os.path.join(safe_dir, sub)):
# The same rule one directory over, which until now had no
# counterpart here at all: the skills branch above tests for a
# SKILL.md, the agents branch took every glob hit on trust. A
# DIRECTORY named `ghost-agent.md` matches `*.md` and glob does not
# tell the two apart, so a leftover of that shape resolved a routing
# target on the machine holding it and dangled everywhere else —
# identical install-dependence, arriving through the one door
# nobody guarded.
if not os.path.isfile(path):
continue
base = os.path.basename(path)
if base.endswith('.agent.md'):
base = base[:-len('.agent.md')]
else:
base = base[:-len('.md')]
names.add(base.lower())
def _apm_package_root(start_dir):
"""Nearest ancestor that is an apm package root (apm.yml or .apm/).
The filesystem root is never a candidate: a stray /.apm/skills/ — a
scaffolding test's leftover, say, and one really does exist on at least one
machine here — would otherwise become the package root of every path on it.
Capped at ten levels so a pathological path can't become a filesystem
crawl; that covers every real layout by a wide margin.
"""
current = os.path.abspath(start_dir)
for _ in range(10):
if _is_fs_root(current):
return None
if (os.path.isfile(os.path.join(current, 'apm.yml'))
or os.path.isdir(os.path.join(current, '.apm'))):
return current
current = os.path.dirname(current)
return None
def _authoring_root(start_dir):
"""Nearest ancestor that is a plugin monorepo, else the nearest .git tree.
Returns (root, matched_plugins_probe). The flag reports WHICH probe
matched: True for the plugins/*/.apm/{skills,agents} glob, False for the
.git fallback and for no match at all. known_targets() needs that
distinction — only a real plugins/ root makes the deployed trees
redundant, and a name-count delta cannot tell the two apart.
Two passes, not one interleaved walk: a nested .git (a submodule, a
worktree of a sub-package) must not win over a real plugins/ root further
up. Both passes stop before the filesystem root for the same reason
_apm_package_root does.
"""
probes = (
lambda d: bool(glob.glob(os.path.join(glob.escape(d), 'plugins', '*', '.apm', 'skills'))
or glob.glob(os.path.join(glob.escape(d), 'plugins', '*', '.apm', 'agents'))),
lambda d: os.path.exists(os.path.join(d, '.git')))
for index, probe in enumerate(probes):
current = os.path.abspath(start_dir)
for _ in range(12):
if _is_fs_root(current):
break
if probe(current):
return current, index == 0
current = os.path.dirname(current)
return None, False
def _collect_authoring_root(root, names):
"""Every plugin in the monorepo contributes its names."""
for pkg in glob.glob(os.path.join(glob.escape(root), 'plugins', '*')):
if os.path.isdir(pkg):
_collect_package(pkg, names)
def _declared_dependency_dirs(pkg_dir):
"""Directories of the apm packages pkg_dir's manifest DECLARES.
Reads dependencies.apm and resolves each entry to a directory on disk:
a monorepo-relative `path:` (against the package root and the nearest
ancestor manifest, which is the monorepo root) or an installed
apm_modules/<name>/. Entries that resolve to nothing are skipped — an
undeployed dependency contributes no names rather than an error.
"""
manifest = os.path.join(pkg_dir, 'apm.yml')
if not os.path.isfile(manifest):
return []
try:
data = yaml.safe_load(read_text(manifest)) or {}
except Exception:
return []
if not isinstance(data, dict):
return []
deps = data.get('dependencies')
deps = deps.get('apm') if isinstance(deps, dict) else None
if not isinstance(deps, list):
return []
roots = [pkg_dir]
ancestor = os.path.dirname(os.path.abspath(pkg_dir))
for _ in range(10):
if _is_fs_root(ancestor):
break
if os.path.isfile(os.path.join(ancestor, 'apm.yml')):
roots.append(ancestor)
break
ancestor = os.path.dirname(ancestor)
found = []
for entry in deps:
candidates = []
if isinstance(entry, dict):
rel = entry.get('path')
name = entry.get('name')
if not name and rel:
name = os.path.basename(str(rel).rstrip('/'))
if rel:
candidates.extend(os.path.join(r, str(rel)) for r in roots)
if name:
candidates.append(os.path.join(pkg_dir, 'apm_modules', str(name)))
elif isinstance(entry, str):
name = re.split(r'[#@]', entry)[0].strip().rstrip('/').split('/')[-1]
if name:
candidates.append(os.path.join(pkg_dir, 'apm_modules', name))
candidates.extend(os.path.join(r, 'plugins', name) for r in roots)
for candidate in candidates:
if os.path.isdir(candidate):
found.append(candidate)
return found
def _deployed_roots(start_dir):
""".claude/ and .agents/ trees above start_dir — what a host really sees.
Consulted ONLY when no plugin monorepo root was found; see the header. The
filesystem root is skipped for the same reason _apm_package_root skips it:
a stray /.claude/skills/ must not join every path's universe.
"""
found = []
current = os.path.abspath(start_dir)
for _ in range(10):
if _is_fs_root(current):
break
for name in ('.claude', '.agents'):
base = os.path.join(current, name)
if os.path.isdir(base):
found.append(base)
current = os.path.dirname(current)
return found
def known_targets(start_dir):
"""Every skill/agent name a boundary clause in start_dir may name."""
names = set()
start = os.path.abspath(start_dir)
# Siblings: a cache-installed plugin and a deployed .claude/skills/ tree
# both put peers one level up, with no plugins/ directory above them. The
# grandparent is guarded against the filesystem root exactly like the two
# walk-up loops above — for a start dir of /skills/<x> the grandparent is
# `/`, and collecting there picks up this machine's stray /.apm/skills/.
parent = os.path.dirname(start)
grandparent = os.path.dirname(parent)
if (os.path.basename(parent) in ('skills', 'agents')
and os.path.isdir(parent) and not _is_fs_root(grandparent)):
_collect_package(grandparent, names)
package = _apm_package_root(start)
if package:
_collect_package(package, names)
for dep_dir in _declared_dependency_dirs(package):
_collect_package(dep_dir, names)
# A .git ancestor is an authoring root only if it actually holds plugins.
# _authoring_root() falls back to the nearest .git, so it is truthy in ANY
# git repo; without the distinction that fallback wins in every consumer
# checkout, _collect_authoring_root() contributes nothing, and the deployed
# branch below is dead code in the exact case it exists for. So condition
# on WHICH probe matched, which _authoring_root() reports directly. A
# name-count delta looks equivalent and is not: _collect_authoring_root()
# re-collects the checked file's own plugin, whose names the blocks above
# already added, so a one-plugin monorepo shows a delta of zero and would
# wrongly reach for the deployed trees — including the user's global
# ~/.claude/skills, making the verdict depend on what happens to be
# installed (ADR-0020 lines 118-127).
root, root_has_plugins = _authoring_root(start)
if root:
_collect_authoring_root(root, names)
if not root_has_plugins:
for base in _deployed_roots(start):
_collect_package(base, names)
return names
# --- Extraction -----------------------------------------------------------
# False positives are the design constraint here, not recall. The rules:
# * A BARE target must be hyphenated AND sit in a boundary sentence (one
# carrying "do not"/"instead"/"rather than"/"not for"). Without the second
# condition, pc-run's "run pre-commit hooks" reads as a route to a
# non-existent `pre-commit` skill.
# * A BARE arrow target counts only in ADR-0020's compressed boundary form,
# `Not <thing> -> <skill-name>`. The example that motivated it is gone:
# diagnose's process chain "fix -> regression-test", which without the
# gate read as a route to a non-existent `regression-test` skill, was cut
# when issue #99 retrofitted that description. So the gate is currently
# UNEXERCISED — gating and not gating produce the same verdict corpus-wide.
# Keep it anyway. It is a false-positive guard against prose no one has
# written yet, and any new process chain re-arms it. Unexercised is not the
# same as unnecessary, and the branch it guards is still load-bearing: the
# bare-arrow rule is the sole extractor for three real targets in
# kyberforge (factory-audit -> skill-author, factory-audit ->
# agent-author, apm-orchestrate -> apm-install), all written unbackticked.
# * A backticked hyphenated token counts only inside a boundary sentence.
# Unconditionally, `pre-push` or `commit-msg` in a TRIGGER clause is a hard
# FAIL with no escape hatch. Gating it costs nothing (measured over this
# corpus: 54 targets before and after); DELETING it costs 7 real targets
# across three gitea skills, so it is gated, not removed.
# * SINGLE-WORD targets are deliberately NOT matchable bare — `research`,
# `triage`, `forge`, `prototype` and `tdd` are all real skill names and all
# ordinary English, so a bare-word rule would flag most of the corpus. A
# single-word target must be written `` `forge` `` or /forge to be seen.
# That is a known recall limitation, accepted over the false positives.
# Tool names (Read/Write/Edit) are excluded by the lowercase-only pattern; MCP
# tool names (issue_write) by its rejection of underscores; file names by its
# rejection of dots and slashes.
#
# ATTRIBUTIVE USE. The boundary-sentence gate above does NOT solve the
# `pre-commit` false positive, and the comment that claimed it did was wrong:
# "instead", "rather than", "do not" and "not for" are exactly the words a
# boundary clause uses, so the gate is open precisely where the risk is. All of
# these were hard dangling FAILs with no suppression:
# Use pre-commit hooks instead of ad-hoc scripts.
# Invoke the pull-request template instead of writing one by hand.
# Use conventional-commits formatting rather than free-form messages.
# Composes label-resolution logic instead of duplicating it.
# Do not use for X — run the `pre-push` hooks instead.
# What separates every one of them from a real route is grammar, not marking:
# the hyphenated token is a compound MODIFIER of the noun that follows it
# ("pre-commit hooks", "pull-request template"), where a route target is
# terminal — followed by punctuation, a conjunction, or a boundary word. So a
# target whose next token is an ordinary lowercase noun is CONFIRM-ONLY: it
# still resolves and still counts as a route when the name exists, but it can
# never raise a dangling error on its own.
#
# This is deliberately NOT the simpler "only marked targets may dangle" rule,
# which would have been wrong here: BOTH live true positives in this corpus are
# BARE — research's "(use neuledge-context)" and gitea-issues' "Composes
# gitea-labels-\n milestones", where the `>` fold yields "gitea-labels-
# milestones" and the trailing hyphen is what keeps it terminal. Marking is a
# poor proxy, so the follower token is the signal, and it is applied to
# backticked targets too.
#
# TERMINAL IS NOT ENOUGH — IN-SENTENCE CORROBORATION. The follower test clears
# `pre-push` in the example above only because that example happens to be
# followed by the noun "hooks". Move the same token into terminal position and
# it was a hard FAIL again, with no suppression mechanism anywhere in this gate:
# Do not use for running hooks — run `pre-commit` instead.
# Do not use for the commit message — see `commit-msg`.
# Do not use for type errors — run `type-check` first.
# Instead, use `semantic-release`.
# Do not use for the old flow — use the clean-up instead.
# Do not run end-to-end, run unit-tests.
# Every one of those is grammatically identical to a genuinely broken route:
# "route verb + hyphenated name + terminal" is also exactly how prose cites a
# tool, a hook, a file format or an English compound. Nothing local separates
# them, and the skills most exposed are the ones this contract sends authors
# back to rewrite first — pc-run, pc-author, vale-run, vale-config and the apm-*
# family are all ABOUT hyphenated tools.
#
# So the confidence to BLOCK a commit comes from the sentence, not the token: a
# prose-form target may raise a hard error only when its own sentence names at
# least one OTHER target that RESOLVES. A routing sentence proves itself by
# routing somewhere real; a lone unresolvable name proves nothing. That is not a
# rule fitted to the fixtures — it is the shape of both live true positives,
# which sit beside `write-docs` and `gitea-labels-milestones` respectively, and
# it changes this corpus's verdict by exactly nothing.
#
# An uncorroborated unresolvable target is NOT discarded: every caller reports
# it at its SUGGESTION tier, naming the target. The finding stays visible on
# every run; only the power to block a commit is withdrawn, which is the part
# that had no escape hatch.
#
# EXPLICIT ROUTE NOTATION is exempt from corroboration and always blocks:
# ADR-0020's compressed arrow (`Not <thing> -> <name>`) and Claude Code's
# invocation form (`/<name>`). Neither is ever how English cites a tool — nobody
# writes `-> pre-commit` or `/pre-commit` to mean the hook — so there is no
# ambiguity to resolve, and an author who wants a route checked unconditionally
# has two ways to say so.
#
# BOTH FORMS ARE SWEPT FOR ON THEIR OWN, and that is a repair of the promise
# above rather than a widening of it. Until the sweeps existed, notation was
# only ever seen as the OBJECT OF A ROUTE VERB (`use
# /name`) or as the tail of a `not ... ->` clause with no `;` or sentence end in
# between. Every one of these therefore exited 0 in total silence — no ERROR, no
# SUGGESTION, not even the target's name:
# Do not use for Y — /no-such-skill instead.
# Do not use for Y; /no-such-skill handles that.
# Do not use for Y (/no-such-skill covers it).
# Do not use for Y — that is /no-such-skill's job.
# Do not use for Y — defer to /no-such-skill.
# Do not use for Y — /no-such-skill.
# Do not use for Y; -> no-such-skill covers it.
# For W, /no-such-skill is the right entry point.
# The target was never EXTRACTED, so the notation-first rule in _add() had
# nothing to apply itself to and the "always blocks" promise was false for the
# ordinary way an author writes the thing. The SUGGESTION tier made it worse
# than a gap: its printed remedy tells the author to "write it as `/name` or
# `-> name` and it will be checked properly", and taking that advice turned a
# visible SUGGESTION into silence — the gate teaching the one edit that blinds
# it.
#
# THE TWO SWEEPS ARE GATED DIFFERENTLY, and the asymmetry is the whole point.
# `/name` is Claude Code's invocation syntax and nothing else — no English
# sentence contains one by accident — so the ADR-0020 amendment and
# docs/spec/gates.md both promise it blocks UNCONDITIONALLY, for any name. So
# NOTATION_SLASH is swept over every sentence, boundary marker or not. Gating it
# on BOUNDARY_MARKER made that promise false for the last sentence of
# Do not use for Z — use /real-skill instead.
# For W, /no-such-skill is the right entry point.
# which exited 0 in total silence: the boundary clause is one sentence up, so
# the sweep never looked at the sentence carrying the broken route. Extraction is
# per-sentence by design (corroboration is scoped to one sentence), which is
# exactly what made the gap invisible.
#
# NOTATION_ARROW stays gated on BOUNDARY_MARKER, and so does the backtick sweep.
# Neither form is unambiguous: `-> name` is also how a process chain is written
# ("reproduce -> minimise -> regression-test") and a code span is how a tool, a
# file and a skill are all cited. Ungating either would fire on prose that
# carries no routing intent at all — the false-positive class this whole
# extractor is tuned against.
#
# BOTH `/name` PATTERNS REFUSE A TOKEN THAT IS PART OF A PATH: a following `/`,
# or a `.` followed by a non-space, means `references/foo.md`, `docs/a/b.md` or
# `https://x/y`, not a route. A sentence's closing `.` is not followed by a
# non-space, so `— /no-such-skill.` still counts.
#
# THAT GUARD IS WRITTEN `(?![\w-])` AND NOT `\b`, because `\b` is not a guard at
# all here: it holds after a hyphen, so when the trailing lookahead rejected the
# full segment the engine simply backtracked to a shorter hyphen-terminated
# prefix and reported THAT as a route. Every one of these was a hard blocking
# ERROR naming a skill nobody had written:
# the config lives at /opt-tools/bin/thing. -> 'opt'
# see /api-docs/v2.md for the schema. -> 'api' AND 'api-docs'
# the file /no-such-skill.md documents it. -> 'no-such'
# `(?![\w-])` forbids the shortened prefix outright, so the whole segment is
# rejected as the path it is. MARKED_TARGET carries the same guard: it had no
# trailing lookahead whatsoever, so `see /api-docs/v2.md` raised the second of
# the two errors above through the route-verb path rather than the sweep.
#
# NAMESPACE: `plugin:skill` is live in this repo (native user-scope installs
# still resolve `gitea:gitea-prs`), so the patterns admit an optional
# `<plugin>:` prefix and normalize_target() strips it before resolution.
NS = r"(?:[a-z0-9]+(?:-[a-z0-9]+)*:)?"
NAME_ANY = NS + r"[a-z0-9]+(?:-[a-z0-9]+)*"
NAME_HYPH = NS + r"[a-z0-9]+(?:-[a-z0-9]+)+"
ROUTE_VERB = (r"(?:use|uses|using|run|runs|invoke|invokes|invoking|try|see"
r"|that'?s|compose|composes|call|calls"
r"|routes?\s+to|delegates?\s+to|prefers?|switch(?:es)?\s+to"
r"|hands?\s+off\s+to)")
MARKED_TARGET = (r"(?:`/?(%s)`|(?<![\w./*-])/(%s)(?![\w-])(?!/|\.\S))"
% (NAME_ANY, NAME_ANY))
ANY_TARGET = r"(?:%s|(%s)\b)" % (MARKED_TARGET, NAME_HYPH)
ROUTE_MARKED = re.compile(r"\b%s\s+(?:the\s+|an?\s+)?%s" % (ROUTE_VERB, MARKED_TARGET), re.I)
ROUTE_ANY = re.compile(r"\b%s\s+(?:the\s+|an?\s+)?%s" % (ROUTE_VERB, ANY_TARGET), re.I)
# re.I on ALL of them, uniformly. The patterns are built from the same
# lowercase NAME_* fragments, so half of them carrying the flag and half not
# meant `Skill-Audit` at the start of a boundary sentence was extracted by
# ROUTE_ANY but invisible to BACKTICK — contradicting normalize_target()'s own
# docstring, which exists precisely because extraction is case-insensitive and
# the universe is not.
CONT_MARKED = re.compile(r"\s*(?:or|and|/|,)\s*%s" % MARKED_TARGET, re.I)
CONT_ANY = re.compile(r"\s*(?:or|and|/|,)\s*%s" % ANY_TARGET, re.I)
ARROW_MARKED = re.compile(r"(?:->|→)\s*%s" % MARKED_TARGET, re.I)
# The two EXPLICIT ROUTE NOTATION sweeps. NOTATION_SLASH runs over EVERY
# sentence; NOTATION_ARROW is scoped to a boundary sentence by its caller (see
# the asymmetry note in the header). NOTATION_SLASH is deliberately not a reuse
# of MARKED_TARGET's `/name` alternative: that one only ever runs behind a route
# verb or an arrow, and it may match a namespaced or path-adjacent token in
# positions this free-standing sweep must refuse.
# NOTATION_ARROW is ARROW_BOUNDARY minus its leading `\bnot\b%s*?`, which is
# what made `Do not use for Y; -> no-such-skill covers it.` invisible:
# CLAUSE_BODY cannot cross the `;`, so the clause's own punctuation disarmed the
# check. Dropping that prefix costs the one false positive the bare-arrow bullet
# above names — a process chain ending in a hyphenated word, `Instead, reproduce
# -> minimise -> regression-test.` — and costs it only in a sentence that already
# carries a BOUNDARY_MARKER. That exposure is neither new nor larger: the same
# chain written `Do not use for X — reproduce -> regression-test.` was already a
# hard ERROR under ARROW_BOUNDARY, so this changes which boundary words reach the
# arrow, not whether prose can. An author who means the chain and not a route
# writes it in its own sentence, where neither pattern looks.
NOTATION_SLASH = re.compile(
r"(?<![\w./*-])/(%s)(?![\w-])(?!/|\.\S)" % NAME_ANY, re.I)
NOTATION_ARROW = re.compile(r"(?:->|→)\s*(%s)\b" % NAME_HYPH, re.I)
# CLAUSE_BODY is what may sit between `Not` and the arrow, and it is NOT
# `[^.;]`. That class cannot cross a `.`, so every boundary clause naming a
# DOTTED FILENAME between the two — `.pre-commit-config.yaml`, `AGENTS.md`,
# `.vale.ini` — was invisible to both patterns below, and the two resulting
# failures were different sizes (issue #110):
# * with a BACKTICKED target the clause was MISDIAGNOSED. The backtick sweep
# still extracted the target, so the route was checked, but the gate
# reported "no boundary clause" on a clause that was present and working.
# Three authors in two retrofit waves reworded a correct clause to satisfy
# the regex, one of them stripping the very filename that discriminates the
# skill from its neighbour.
# * with a BARE target the clause was UNCHECKED. ARROW_BOUNDARY is the only
# extractor for a bare arrow target, so `Not AGENTS.md -> no-such-skill`
# produced no target, no dangling report and no missing-clause SUGGESTION.
# Silence, not noise — the worse of the two failure modes.
# A dot inside a filename is followed by a non-space; a sentence-ending dot is
# followed by whitespace or by end of string. So the class admits a `.` only
# when the next character is not whitespace, which crosses `AGENTS.md` and
# still stops at a real sentence end.
CLAUSE_BODY = r"(?:[^.;]|\.(?=\S))"
ARROW_BOUNDARY = re.compile(
r"\bnot\b%s*?(?:->|→)\s*(%s)\b" % (CLAUSE_BODY, NAME_HYPH), re.I)
BACKTICK = re.compile(r"`(%s)`" % NAME_HYPH, re.I)
# A boundary clause takes two shapes and BOTH count: the prose markers, and
# ADR-0020's compressed arrow form `Not <thing> -> <name>`.
BOUNDARY_MARKER = re.compile(r"\b(?:do\s+not|instead|rather\s+than|not\s+for)\b", re.I)
BOUNDARY_ARROW = re.compile(r"\bnot\b%s*?(?:->|→)" % CLAUSE_BODY, re.I)
# Sentence boundaries decide the CORROBORATION scope above, so getting one wrong
# is not cosmetic — it moves a target between SUGGESTION and blocking ERROR. Two
# shapes common in these descriptions defeat the naive "period, space, capital"
# rule, in OPPOSITE directions:
# OVER-SPLIT. `e.g. "set up the manifest"` ends no sentence, but the quote
# looks like one starting. The clause is cut in half, the corroborating
# target lands on the far side of the cut, and a genuinely dangling target
# silently demotes to SUGGESTION — the gate takes a measurement and then
# throws it away, which is the vacuous-green shape this file exists to stop.
# UNDER-SPLIT. A real sentence opening with a code span or a lowercase skill
# name ("... Composes it. `gitea-prs` also uses it.") is not seen as a start
# at all, so two sentences merge and a resolving target vouches for an
# unresolvable one it never stood beside — a hard FAIL with no escape hatch,
# which is exactly the failure the corroboration rule was added to prevent.
# Both are closed here: the five abbreviations that actually occur in routing
# prose are excluded as sentence ends, and the opener class admits a backtick or
# a lowercase letter. Verified zero-delta on the current corpus (37 ERROR / 58
# SUGGESTION / 2 dangling before and after) — this protects the descriptions
# issue #99 is about to rewrite, not the ones already measured.
# re.I here too, and NOT as a tidy-up: this was the one pattern in the file
# built without it, contradicting the uniformity note on CONT_*/ARROW_* above.
# Without the flag `E.g.` and `I.e.` — the sentence-initial spellings, which is
# where an abbreviation most often lands — matched none of the lookbehinds, so
# the clause split at the abbreviation, the corroborating target was stranded on
# the far side of the cut, and a genuinely dangling target silently demoted from
# blocking ERROR to SUGGESTION. That is the OVER-SPLIT failure described
# directly above, still live for exactly the capitalised half of the input.
SENTENCE_SPLIT = re.compile(
u'(?<!\\be\\.g\\.)(?<!\\bi\\.e\\.)(?<!\\betc\\.)(?<!\\bvs\\.)(?<!\\bcf\\.)'
u'(?<=[.!?])\\s+(?=[A-Za-z`"“(])', re.I)
# The token that may follow a route target without turning it into a compound
# modifier: punctuation, end of sentence, a conjunction, a boundary word, or a
# head noun that names the artifact itself ("the git-workflow skill"). Anything
# else — `hooks`, `template`, `formatting`, `logic` — is attributive prose.
FOLLOWER = re.compile(r"[`\s]*([a-z][a-z0-9]*)")
FOLLOWER_OK = frozenset("""
and or nor but for to when if unless while after before with from in on at by
of as than then instead rather directly first only always never also even
both either neither so because since per via plus alone here there this that
these those it its they them is are was were be been being has have had will
would can could should must may might does do did
skill skills agent agents plugin plugins command commands
""".split())
def normalize_target(target):
"""Comparison key: namespace stripped, lowercased.
Extraction is case-insensitive (re.I) but the universe is built from
lowercase directory names, so `Git-Commits` at the start of a sentence
resolved to nothing until this normalization existed.
"""
return target.split(':')[-1].lower()
def has_boundary_clause(description):
return bool(BOUNDARY_MARKER.search(description)
or BOUNDARY_ARROW.search(description))
def _first(match):
"""(name, start, end) offsets for the first group that matched."""
for index in range(1, (match.re.groups or 0) + 1):
if match.group(index):
return match.group(index), match.start(index), match.end(index)
return None, None, None
def _terminal(text, pos):
"""True if the token at pos does not make the preceding name a modifier."""
follower = FOLLOWER.match(text, pos)
return not follower or follower.group(1) in FOLLOWER_OK
def _notation(text, start, arrow):
"""True if the name is written in route NOTATION rather than in prose.
Two forms qualify: `/name` (Claude Code's invocation syntax, detected from
the character before the name) and `-> name` (ADR-0020's compressed boundary
form, passed in by the caller that matched the arrow). A backticked name
does NOT qualify — a code span is how a tool, a file and a skill are all
cited, so it carries no intent the follower test hasn't already read.
"""
return arrow or (start > 0 and text[start - 1] == '/')
def _add(out, text, name, start, end, strict=None, arrow=False):
"""Record one target as (name, may_dangle, notation).
NOTATION IS DECIDED FIRST, and when it is set the follower test is skipped.
The header above promises that route notation "always blocks", and for the
`/name` form that was false: `-> name` reached this function with
strict=True from its two call sites, but `/name` did not, so it fell to
_terminal() and a follower outside FOLLOWER_OK set may_dangle=False. The
target then reached unresolved_targets() unblockable — and, before the
companion fix there, unreported as well. `... use /no-such-skill
afterwards.` exited 0 in total silence, on the one form ADR-0020 offers an
author who wants a route checked unconditionally.
"""
if not name:
return
notation = _notation(text, start, arrow)
if strict is None and notation:
strict = True
out.append((name,
_terminal(text, end) if strict is None else strict,
notation))
def _scan(text, route_re, cont_re, out):
for match in route_re.finditer(text):
name, start, end = _first(match)
if not name:
continue
_add(out, text, name, start, end)
# "use git-history or git-branches instead" / "use gitea-issues /
# gitea-prs" — keep consuming conjoined targets after the first.
pos = match.end()
while True:
cont = cont_re.match(text, pos)
if not cont:
break
_add(out, text, *_first(cont))
pos = cont.end()
def _extract_sentence(sentence):
"""[(name, may_dangle, notation)] for the routing targets in ONE sentence.
Kept separate from _extract() because corroboration is scoped to a single
sentence: a target's evidence is what stands beside it, not what the rest of
the description happens to mention.
"""
out = []
boundary = bool(BOUNDARY_MARKER.search(sentence))
_scan(sentence,
ROUTE_ANY if boundary else ROUTE_MARKED,
CONT_ANY if boundary else CONT_MARKED,
out)
for match in ARROW_MARKED.finditer(sentence):
# `-> name` and `-> /name` are route notation, not prose: nothing
# reads as a compound modifier after an arrow, so no follower test.
_add(out, sentence, *_first(match), strict=True, arrow=True)
for match in ARROW_BOUNDARY.finditer(sentence):
_add(out, sentence, match.group(1), match.start(1), match.end(1),
strict=True, arrow=True)
# `/name` wherever it sits, in ANY sentence — not only where a route verb or
# an arrow happens to precede it, and NOT only inside a boundary sentence.
# See the EXPLICIT ROUTE NOTATION note in the header for the eight phrasings
# this recovers and for why silence was the failure mode. The sweep takes no
# follower test: _add() reads the notation first and marks it.
for match in NOTATION_SLASH.finditer(sentence):
_add(out, sentence, match.group(1), match.start(1), match.end(1))
if boundary:
# The arrow and backtick forms are ambiguous in ordinary prose, so they
# stay scoped to a sentence that carries a boundary marker.
for match in NOTATION_ARROW.finditer(sentence):
_add(out, sentence, match.group(1), match.start(1), match.end(1),
strict=True, arrow=True)
for match in BACKTICK.finditer(sentence):
_add(out, sentence, match.group(1), match.start(1), match.end(1))
return out
def _extract(description):
"""[(name, may_dangle, notation)] for every routing target."""
out = []
for sentence in SENTENCE_SPLIT.split(description):
out.extend(_extract_sentence(sentence))
return out
def boundary_targets(description):
"""Every routing target, for reporting and for confirming a route."""
return sorted({name for name, _, _ in _extract(description)})
def _arrow_targets(description):
"""Names extracted from ARROW notation specifically.
Kept apart from boundary_targets() because the arrow form is the one shape
that ALWAYS names a target: ADR-0020's `Not <thing> -> <name>`. A clause
written that way from which nothing could be extracted is a parse failure
that deserves its own message, and telling it apart needs the arrow targets
alone rather than every target in the description.
"""
out = []
for sentence in SENTENCE_SPLIT.split(description):
for match in ARROW_MARKED.finditer(sentence):
name, _, _ = _first(match)
if name:
out.append(name)
for match in ARROW_BOUNDARY.finditer(sentence):
out.append(match.group(1))
return out
def boundary_clause_status(description):
"""'absent', 'unparsed' or 'present' — three outcomes, not two.
Issue #110's standing request: the gate must distinguish "no boundary
clause" from "boundary clause I could not parse". Reporting the first for
the second sends the author hunting for a problem that is not there, and
three of them reworded a correct clause to satisfy a regex instead.
'unparsed' is the narrow, certain case: an ADR-0020 arrow clause was
detected and NO target came out of it. The arrow form always names one, so
zero targets means the name is written in a shape the extractor cannot see
— a single-word bare target (`Not X -> forge`, which has to be written
`` `forge` `` or `/forge`) is the live example, since single-word names are
deliberately not matchable bare.
A PROSE clause yielding no target is NOT reported: "Do not use for anything
else" is a complete and legitimate boundary clause that names nowhere to go.
"""
if BOUNDARY_ARROW.search(description) and not _arrow_targets(description):
return 'unparsed'
if has_boundary_clause(description):
return 'present'
return 'absent'
def multi_target_arrow_clauses(description):
"""[(first, second)] for arrow clauses naming more than one target.
Issue #107: only the FIRST target after an arrow is resolved. The
conjunction continuation (CONT_*) is wired to the prose route verbs and
never to arrows, so `Not X -> a or b` resolved `a`, left `b` neither
resolved nor reported, and then printed "1 of 1 boundary target(s) resolve"
on a clause naming two — a gate under-reporting its own coverage, which is
the one failure mode ADR-0020 says a gate must not have.
The clause is REJECTED rather than the arrow scan extended. Extending it
would widen the resolver's deliberately conservative false-positive tuning
across every arrow in the corpus; rejecting costs nothing and makes the
one-arrow-per-target convention — already what every retrofitted gitea
skill does in practice — explicit instead of folkloric. The caller emits a
SUGGESTION telling the author to split.
"""
hits = []
for sentence in SENTENCE_SPLIT.split(description):
matches = (list(ARROW_MARKED.finditer(sentence))
+ list(ARROW_BOUNDARY.finditer(sentence)))
for match in matches:
first, _, _ = _first(match)
if not first:
continue
cont = CONT_ANY.match(sentence, match.end())
if not cont:
continue
second, _, _ = _first(cont)
if second:
hits.append((first, second))
return hits
def unresolved_targets(description, known):
"""Targets resolving to nothing, split into (blocking, reported).
`blocking` earns a hard error; `reported` is SUGGESTION tier — named on
every run, never fatal. Three conditions gate the promotion, and all of them
are documented at length in the ATTRIBUTIVE USE and CORROBORATION notes
above:
1. the target must be terminal, not a compound modifier ("pre-commit
hooks" is prose about a tool, not a route),
2. it must be written in route notation (`/name`, `-> name`), OR
3. its own sentence must name another target that DOES resolve.
Everything else is reported and left alone. `known` is the resolved
universe from known_targets(); passing an empty set is not meaningful —
callers check for that first and decline out loud instead.
A NON-TERMINAL target is reported, never dropped. FOLLOWER_OK is a closed
whitelist of maybe eighty words, so the follower rule says "this token is
outside a list I keep" and not "this is prose" — and the old `continue`
turned that into invisibility at every tier. The gate then failed OPEN on
its own unfamiliarity: any target followed by a word nobody thought to
enumerate was neither blocked nor mentioned, so the check that did not run
said nothing about not running. The follower rule may withdraw the power to
BLOCK a commit — that is what it was added for, and the ATTRIBUTIVE USE note
above is the argument for it — but it may not withdraw visibility, which is
the same rule the corroboration tier already follows.
"""
blocking, reported = set(), set()
for sentence in SENTENCE_SPLIT.split(description):
found = _extract_sentence(sentence)
resolved = {normalize_target(name) for name, _, _ in found
if normalize_target(name) in known}
for name, may_dangle, notation in found:
key = normalize_target(name)
if key in known:
continue
if not may_dangle:
reported.add(name)
continue
if notation or (resolved - {key}):
blocking.add(name)
else:
reported.add(name)
return sorted(blocking), sorted(reported - blocking)
# --- Frontmatter ----------------------------------------------------------
# Tolerant on the way in, HARD-FAILING on the way out. A UTF-8 BOM, a leading
# blank line, trailing whitespace after either `---`, or CRLF line endings all
# defeated the old `^---\n(.*?)\n---`, and the miss was SILENT: every ADR-0020
# check was skipped and the file reported green (measured: a 550-character
# description with a 1,000-word body exited 0 behind a BOM). A file that cannot
# be measured must never report green, so every caller of these two ERRORs on a
# miss instead of moving on.
#
# The CLOSING marker is anchored at column 0 — deliberately NOT `[ \t]*---`.
# YAML block-scalar content must be indented deeper than its key, so an
# indented `---` inside a folded description is CONTENT; letting it close the
# frontmatter truncated the description mid-value and silently reclassified the
# rest as body, which is a vacuous green in both directions at once. Leading
# whitespace is still tolerated on the OPENING marker, where no such content
# can exist.
FRONTMATTER_RE = re.compile(
r'^[ \t\r\n]*---[ \t]*\r?\n(.*?)\r?\n---[ \t]*(?:\r?\n|\Z)', re.DOTALL)
def strip_bom(text):
return text[1:] if text.startswith(u'') else text
class FrontmatterError(Exception):
pass
def description_value(fm_text):
"""The description VALUE, with YAML folding resolved.
PyYAML is a HARD requirement, preflighted in bash. The hand-rolled fallback
this replaced diverged from a real parser across the FAIL boundary — one
corpus description measured 270 characters parsed and 412 unparsed, and a
quoted `"description"` key or an explicit `description: null` returned empty
from it, silently skipping the description AND routing checks. A gate that
disagrees with itself depending on which reader ran is worse than no gate.
This is the ONLY reader any of the three scripts may use to decide whether a
description is present. A line regex cannot: `description:` with no value
followed by `model: sonnet` lets `\\s*` cross the newline and captures the
NEXT key, which reads as a non-empty description, skips the "missing or
empty" failure, and then early-returns out of every ADR-0020 gate on the
genuinely empty folded value. That combination exited 0 with zero output on
a BLOCKING pre-push gate.
"""
try:
data = yaml.safe_load(fm_text)
except Exception as exc:
# Every FrontmatterError message is a COMPLETE clause, never a detail a
# caller wraps in one. Callers used to prefix a hard-coded "frontmatter
# is not valid YAML (...)", which is true only of this branch: the two
# type failures below come from frontmatter that parsed fine, and
# telling their author the YAML is invalid sends them hunting for a
# syntax error that is not there — on a blocking gate with no baseline.
raise FrontmatterError('frontmatter is not valid YAML (%s)'
% re.sub(r'\s+', ' ', str(exc)).strip())
if not isinstance(data, dict):
raise FrontmatterError('frontmatter is not a YAML mapping')
value = data.get('description')
if value is None:
return ''
if not isinstance(value, str):
# NOT str()-coerced. `description: true` became the 4-character "True"
# and sailed through the 400-character gate; a list or mapping was
# measured as its Python repr. Neither is a description a host can
# preload, so this is a parse failure, reported as one.
raise FrontmatterError(
'description is a %s, not a string' % type(value).__name__)
return re.sub(r'\s+', ' ', value).strip()
def hand_invoked(fm_text):
"""True when the frontmatter marks this file as reached only by hand.
`disable-model-invocation: true` removes a skill from the model-visible
listing entirely — it is not preloaded, and the Skill tool refuses to call
it — so its description is never matched against user intent. ADR-0020 and
skill-author's contract give such a skill ONE plain human-facing sentence:
no trigger list, no boundary clause. No validator knew the field existed
(issue #108), so the boundary-clause SUGGESTION fired on exactly the shape
the contract mandates, and its remedy — "add a boundary clause so the router
knows where NOT to send this skill" — was addressed to a router that cannot
see the skill at all. An author who followed the advice made the file worse.
Only the ROUTING rules are lifted. The body word budget still applies: the
body is loaded on invocation like any other, and competes with the caller's
live conversation the same way. So does the 400-character description FAIL —
a hand-invoked description is not preloaded, but it is still the one line
the user reads when choosing from the `/` menu, and the ceiling is the
outlier stop rather than the style target.
A parse failure returns False rather than raising. This is a MODIFIER on
other checks, not a check of its own: the frontmatter's validity is decided,
and failed, by description_value() on the same text, and raising a second
exception here would report one broken file twice with two different
diagnoses.
"""
try:
data = yaml.safe_load(fm_text)
except Exception:
return False
if not isinstance(data, dict):
return False
value = data.get('disable-model-invocation')
if isinstance(value, str):
# PyYAML already resolves the unquoted YAML 1.1 booleans, so this only
# catches a QUOTED "true" — which a host reads as truthy and which no
# gate should treat as opting back in to the routing rules.
return value.strip().lower() in ('true', 'yes', 'on')
return value is True
# --- Body-shape checks (skills only; agents have no references/ dir) -------
# Deterministic and countable, so they are enforced here. Whether a given
# gotcha is WARRANTED is semantic and stays the auditor's judgment, which is why
# both gotcha checks are SUGGESTION tier. A missing reference file is not a
# style opinion — it is a broken pointer — so that one is ERROR tier.
#
# Both read a FENCE-MASKED copy of the body. Scanning the raw body made a
# ```-fenced example a hard ERROR — and the skills most likely to carry one are
# skill-author and factory-audit, which DOCUMENT the references/ convention — and
# let a `## Gotchas` heading inside a fenced block stand in for the real
# section. Masking preserves every byte offset (content becomes spaces,
# newlines stay), so a span found in the mask slices the original.
GOTCHA_MAX_ENTRIES = 5
GOTCHA_MAX_BODY_FRACTION = 0.25
# The heading has to BE "Gotchas", not merely contain the word: `## Gotcha
# handling` and `## Why gotchas matter` are prose sections, and treating one as
# the Gotchas section measured a span that was never a gotcha list.
GOTCHA_HEADING = re.compile(r'^(#{1,6})[ \t]+(?:[^\n]*?[ \t])?gotchas?[ \t]*:?[ \t]*$',
re.I | re.M)
# Column 0 only. `^[ \t]{0,3}` counted a two-space-indented CHILD bullet as a
# top-level entry, so a five-entry section with sub-bullets reported nine.
GOTCHA_ENTRY = re.compile(r'^(?:[-*+]|\d+[.)])[ \t]+', re.M)
FENCE_OPEN = re.compile(r'^[ \t]{0,3}(`{3,}|~{3,})')
REFERENCE_POINTER = re.compile(
r'(?<![\w./-])(?:\./)?references/([A-Za-z0-9][A-Za-z0-9._/-]*\.md)')
# A pointer named in a sentence that says the file is GONE is a historical
# mention, not a dispatch entry: "the old `references/legacy.md` was removed in
# v2" is prose and must not be a hard ERROR. Narrow on purpose — a live
# dispatch table never describes its own target as removed, so this costs no
# recall.
REFERENCE_PAST = re.compile(
r'\b(?:removed|deleted|renamed|superseded|replaced|obsolete|deprecated'
r'|former|formerly|gone|no longer|used to)\b', re.I)
# A pointer QUALIFIED by another skill's name — "factory-audit's
# references/skill-validation-scripts.md" — names a file that is deliberately NOT in
# this skill's directory. Requiring it on the local disk left NO legal spelling
# for a cross-skill reference at all: the only alternative, a full repo path
# (`plugins/kyberforge/.apm/skills/factory-audit/references/...`), is itself a
# FAIL under factory-audit's own file-structure rubric, because a path climbing
# out of the skill directory stops resolving once the plugin is cache-installed.
# The possessive form is the sanctioned spelling, and it is skipped here. It is
# not checked further — this function has no way to locate another skill's
# directory, and inventing one would reintroduce exactly the cross-plugin path
# assumption the rubric forbids.
REFERENCE_QUALIFIER = re.compile(u"[A-Za-z0-9][A-Za-z0-9._-]*`?['’]s[ \t]+`?$")
def mask_fenced(text):
"""Body with fenced code blocks blanked out, byte offsets preserved."""
out = []
fence = None
for line in text.splitlines(keepends=True):
stripped = line.rstrip('\r\n')
opener = FENCE_OPEN.match(stripped)
marker = opener.group(1) if opener else None
if fence is None:
if marker:
fence = marker
out.append(' ' * len(stripped) + line[len(stripped):])
continue
out.append(line)
else:
out.append(' ' * len(stripped) + line[len(stripped):])
if (marker and marker[0] == fence[0] and len(marker) >= len(fence)
and not stripped.strip()[len(marker):].strip()):
fence = None
# An UNCLOSED fence has no cost-free answer, only a choice of which way to
# be wrong. Masking to end-of-body blanks the rest of the body, silently
# disabling the ERROR-tier references/ check and the gotcha counts.
# Returning the raw text instead exposes the unclosed example's own
# content, so a fenced example naming a nonexistent references/ file
# becomes a hard ERROR it would not have been had the fence been closed —
# confirmed, not hypothetical. The loud-false-positive direction is the one
# chosen: this script's rule is that a file it cannot measure must never
# report green, and masking-onward is exactly that failure. Both outcomes
# need an already-malformed file, and the false positive costs one fence.
if fence is not None:
return text
return ''.join(out)
def gotcha_stats(body):
"""(entry count, section word count) for the first Gotchas section, or None.
The section runs to the next heading at the same level or shallower.
Entries are top-level list items; a section written as subheadings instead
of a list counts those. Headings and entries are read from the fence mask;
the word count is taken from the original slice, because fenced lines are
real body words and the fraction is measured against the whole body.
"""
masked = mask_fenced(body)
match = GOTCHA_HEADING.search(masked)
if not match:
return None
level = len(match.group(1))
rest = masked[match.end():]
nxt = re.search(r'^#{1,%d}[ \t]+' % level, rest, re.M)
end = match.end() + (nxt.start() if nxt else len(rest))
section = masked[match.end():end]
entries = len(GOTCHA_ENTRY.findall(section))
if entries == 0 and level < 6:
entries = len(re.findall(r'^#{%d,6}[ \t]+' % (level + 1), section, re.M))
return entries, len(body[match.end():end].split())
def missing_reference_pointers(body, skill_dir):
"""references/<file>.md named in the body but absent from disk."""
masked = mask_fenced(body)
missing = set()
for match in REFERENCE_POINTER.finditer(masked):
start = masked.rfind('\n', 0, match.start()) + 1
end = masked.find('\n', match.end())
if end < 0:
end = len(masked)
# The pointer's OWN SPAN is excised before the sweep. Run over the
# whole line, the past-tense test matched the very path it was judging,
# so a file exempted itself by its NAME: `references/deprecated-api.md`,
# `references/removed-flags.md` and `references/gone.md` produced no
# ERROR at all, while `references/missing.md` — an identical break —
# errored. The exemption is about what the SENTENCE says about the
# pointer, never about what the pointer is called.
line = masked[start:match.start()] + masked[match.end():end]
if REFERENCE_PAST.search(line):
continue
if REFERENCE_QUALIFIER.search(masked[start:match.start()]):
continue
if not os.path.isfile(os.path.join(skill_dir, 'references', match.group(1))):
missing.add('references/' + match.group(1))
return sorted(missing)
# ===== END ADR-0020 SHARED BOUNDARY RESOLVER =====
# --- Per-file checks ------------------------------------------------------
for path in files:
if not os.path.isfile(path):
# NOT a silent skip. A broken symlink named SKILL.md is storable in git,
# so pre-commit really can hand one to this hook, and a directory named
# SKILL.md reaches it the same way — both used to exit 0 with zero
# output, which is precisely the "stay quiet about a measurement it did
# not take" failure this script forbids itself two screens up. This is
# the ONLY place that diagnosis is made; the bash pre-loop that used to
# duplicate it printed a second ERROR line for the same broken file.
if os.path.isdir(path):
why = "is a directory, not a file"
elif os.path.islink(path):
why = "is a symlink that does not resolve to a file"
elif os.path.exists(path):
why = "is not a regular file"
else:
why = "does not exist"
error("%s: %s, so none of the ADR-0020 gates could run on it. A path "
"this gate was handed and could not read does not get to pass in "
"silence." % (path, why))
continue
try:
raw = read_text(path)
except EncodingError as exc:
error("%s: %s. Neither the spec line/word ceilings nor any of the "
"ADR-0020 gates could run on this file." % (path, exc))
continue
# SPEC CONFORMANCE (family 1). Whole file, frontmatter included, counted
# with the SAME primitives factory-audit/scripts/lib-checks-skill.sh uses for these
# two constants — see the note in bash above for what the previous awk pass
# got wrong.
lines = len(raw.splitlines())
words = len(raw.split())
if lines > MAX_LINES:
error("%s has %d lines, exceeding the %d-line ceiling "
"(agentskills.io skill-authoring.md)" % (path, lines, MAX_LINES))
if words > MAX_WORDS:
error("%s has %d words (proxy for tokens), exceeding the %d-word ceiling "
"(~5,000 tokens, agentskills.io skill-authoring.md)"
% (path, words, MAX_WORDS))
content = strip_bom(raw)
fm_match = FRONTMATTER_RE.match(content)
if not fm_match:
error("%s: no parseable YAML frontmatter block. Expected a `---` line, "
"the fields, then a closing `---` line (a BOM, leading blank "
"lines, trailing spaces after either marker and CRLF endings are "
"all tolerated). None of the ADR-0020 gates could run on this "
"file — a file that cannot be measured does not get to pass."
% path)
continue
try:
desc = description_value(fm_match.group(1))
except FrontmatterError as exc:
# `exc` carries the whole clause — invalid YAML, a non-mapping block, or
# a description of the wrong type. Do not prefix a diagnosis here; the
# last one named a syntax error for two failures that have none.
error("%s: %s. None of the ADR-0020 gates could run on this file."
% (path, exc))
continue
# Required-field presence (folded in from the former standalone
# `skill-frontmatter` hook). description_value() above already proved the
# frontmatter is valid YAML and a mapping, so a second yaml.safe_load here
# cannot raise.
fm_data = yaml.safe_load(fm_match.group(1)) or {}
name_val = fm_data.get('name')
if not isinstance(name_val, str) or not name_val.strip():
error("%s: name field is missing or empty (required frontmatter field)."
% path)
version_val = None
metadata_val = fm_data.get('metadata')
if isinstance(metadata_val, dict):
version_val = metadata_val.get('version')
if version_val is None:
error("%s: metadata.version field is missing (required frontmatter "
"field, e.g. \"1.0.0\")." % path)
# Same shape as check-skill-version-bump.sh: ASCII digits, at most nine per
# part (bash arithmetic), no leading zero (semver 2.0.0 item 2).
elif not re.fullmatch(r'(0|[1-9][0-9]{0,8})\.(0|[1-9][0-9]{0,8})\.(0|[1-9][0-9]{0,8})',
str(version_val).strip().strip('\'"')):
error("%s: metadata.version is malformed (%r) -- expected a "
"three-part semver, e.g. \"1.0.0\"." % (path, version_val))
body = content[fm_match.end():]
skill_dir = os.path.dirname(os.path.abspath(path))
# ADR-0020's hand-invocation carve-out. See hand_invoked() for what it lifts
# and, more importantly, what it does not (issue #108).
by_hand = hand_invoked(fm_match.group(1))
# An absent or empty description is an ERROR here too, not a silent skip.
# All three ADR-0020 scripts have to agree on this input: the description is
# the one field preloaded into every session, so a SKILL.md that ships
# without one is the worst case the contract exists for, and a gate that
# merely declines to measure it reports green.
if not desc:
error("%s: description field is missing or empty. It is the only part of a "
"skill preloaded into every session, so a skill without one can never "
"be routed to — and none of the ADR-0020 description or boundary gates "
"have anything to measure." % path)
if desc:
dlen = len(desc)
if dlen > DESC_MAX_CHARS:
error("%s: description is %d characters, exceeding the %d-character ceiling "
"(ADR-0020). It is preloaded into every session. Keep a trigger clause, "
"at most one capability clause, and a boundary clause; move capability "
"enumeration, output-format detail, composition notes and implementation "
"detail to the body or README.md."
% (path, dlen, DESC_MAX_CHARS))
elif dlen > DESC_SUGGEST_CHARS and not by_hand:
# The 250-character TARGET is a routing-quality budget: it exists to
# keep the preloaded listing small and the trigger clause sharp. A
# hand-invoked description is in no listing, so there is no budget to
# spend and no shape to enforce. The 400-character FAIL above still
# applies — see hand_invoked().
suggest("%s: description is %d characters, over the %d-character target "
"(ADR-0020, hard fail at %d)."
% (path, dlen, DESC_SUGGEST_CHARS, DESC_MAX_CHARS))
body_words = len(body.split())
if body_words > BODY_MAX_WORDS:
error("%s: body is %d words, exceeding the %d-word ceiling (ADR-0020). This counts "
"the body ONLY — it is a separate measurement from the %d-word whole-file "
"spec ceiling above. Move lookup tables, spec restatements, output schemas, "
"templates and rationale prose to references/ behind an explicit "
"\"If X, read references/file.md\" trigger."
% (path, body_words, BODY_MAX_WORDS, MAX_WORDS))
elif body_words > BODY_SUGGEST_WORDS:
suggest("%s: body is %d words, over the %d-word target (ADR-0020, hard fail at %d)."
% (path, body_words, BODY_SUGGEST_WORDS, BODY_MAX_WORDS))
# Reference pointers must exist. ERROR, not SUGGESTION: a dispatch table
# naming a file that is not on disk is a hard break, and nothing else in
# the gate/audit/vale stack notices it.
for ref in missing_reference_pointers(body, skill_dir):
error("%s: body points at %s, which does not exist on disk. A dispatch "
"table or \"read X\" trigger naming a missing file sends the "
"agent nowhere." % (path, ref))
# Gotchas discipline. SUGGESTION on both counts: the measurement is
# deterministic, but whether a given gotcha earns its place is judgment.
stats = gotcha_stats(body)
if stats is not None:
entries, section_words = stats
if entries > GOTCHA_MAX_ENTRIES:
suggest("%s: Gotchas section has %d entries, over the %d-entry guideline. "
"A list that long is usually a missing references/ file or a design "
"problem written up as a warning."
% (path, entries, GOTCHA_MAX_ENTRIES))
if body_words and section_words > body_words * GOTCHA_MAX_BODY_FRACTION:
suggest("%s: Gotchas section is %d of %d body words (%d%%), over the %d%% "
"guideline. Move the durable parts to references/ and keep the "
"section for live traps."
% (path, section_words, body_words,
round(100.0 * section_words / body_words),
round(100.0 * GOTCHA_MAX_BODY_FRACTION)))
# Boundary clause. SUGGESTION, not ERROR: detecting the absence is
# deterministic, but whether this particular skill warrants one is the
# auditor's call. Both accepted shapes count — the prose markers and
# ADR-0020's compressed `Not <thing> -> <name>` arrow.
#
# THREE outcomes, not two. Reporting "no boundary clause" for a clause that
# is present and merely unparsed is a wrong finding, not a strict one, and
# it cost three authors a reworded clause before it was diagnosed (#110).
#
# Skipped entirely for a hand-invoked skill: the contract gives it one plain
# sentence with no boundary clause, so the finding is wrong and its remedy
# names a router that cannot see the skill (#108).
if desc and not by_hand:
status = boundary_clause_status(desc)
if status == 'absent':
suggest("%s: description has no boundary clause (ADR-0020). Add the prose form "
"(\"Do not use for X — use `y` instead\") or the compressed form "
"(\"Not X -> y\") so the router knows where NOT to send this skill."
% path)
elif status == 'unparsed':
suggest("%s: description has an arrow boundary clause (\"Not X -> y\") from which "
"no target could be read, so the dangling-target check did not run on it "
"(ADR-0020). The clause is present — this is a PARSE failure, not a "
"missing clause. Most often the target is a single word, which is "
"deliberately not matchable bare because `research`, `triage` and `forge` "
"are all ordinary English: write it as `name` or /name." % path)
# One arrow, one target. A second name after the arrow is resolved by
# nothing and reported by nothing, so the clause claims coverage it does
# not have (#107).
for first, second in multi_target_arrow_clauses(desc):
suggest("%s: an arrow boundary clause names more than one target ('%s', then "
"'%s'), and only the first is resolved — the second is checked by "
"nothing (ADR-0020). Split it into one arrow per target: "
"\"Not X -> %s. Not Y -> %s.\"" % (path, first, second, first, second))
targets = boundary_targets(desc)
if targets:
known = known_targets(skill_dir)
if known:
blocking, reported = unresolved_targets(desc, known)
for target in blocking:
error("%s: description routes to '%s', which does not resolve to a skill "
"or agent in this monorepo, in this package, or in a package it "
"declares in apm.yml dependencies.apm (ADR-0020). A boundary clause "
"that names a non-existent target sends the router nowhere."
% (path, target))
for target in reported:
suggest("%s: description routes to '%s', which does not resolve to a skill "
"or agent in this monorepo, in this package, or in a package it "
"declares in apm.yml dependencies.apm (ADR-0020). SUGGESTION rather "
"than a hard failure because nothing else in the sentence resolves, "
"so this is equally likely to be a tool, a file format or an English "
"compound. If it IS a route, write it as `/%s` or `-> %s` and it will "
"be checked properly." % (path, target, target, target))
else:
info("%s: boundary-target resolution DID NOT RUN — no skill universe "
"could be determined for this path (no authoring root above it, no "
"apm package root, no declared apm dependencies, no deployed "
".claude/ or .agents/ tree). Unchecked target(s): %s"
% (path, ", ".join(targets)))
sys.exit(1 if failed else 0)
PYTHON
then
FAIL=1
fi
exit $FAIL
Reference in New Issue View Git Blame Copy Permalink