SkillPedia Skills Encyclopedia
中文 EN
今日更新 2026-11-18 UTC

今天还没有新提交

查看更新动态

技能 2026-10-09 23:41 UTC

agent-skills

由提交人 simota 提交,主要更新类型为:问题缺陷修复与稳定性提升。提交说明:fix: second-pass review of refs, scripts, routing (#54)。解决了已知运行时异常与边界错误,提升系统稳定性。

simota 4537a3e simota/agent-skills

Commit提交信息原文

fix: second-pass review of refs, scripts, routing (#54)

* fix(launch): make report and audit scripts portable

puppeteer-pdf.js resolved puppeteer only from the script's own directory,
so the documented `npm install puppeteer` in the user's project never
worked. token-economy.py defaulted to one maintainer's macOS transcript
path, rejected Claude Code's dash-prefixed project directory names, and
ignored --repo-root, so the audit step HARNESS_EVOLUTION documents exited
2 everywhere else. The client report's category chart had two colors for
up to nine categories, and its Chart.js CDN URL was unversioned.

The catalog page also stopped rendering when DOM storage is disabled
(localStorage is null), because the storage fallback re-threw anything
but SecurityError. Each fix carries a regression test.

* fix(skills): settle overlapping ownership claims

Several skills claimed the same work in ways their own boundary docs
already ruled out: Quill wrote ADRs although BOUNDARIES gives ADR creation
to Atlas (architecture) and Scribe (other decisions); Sweep's EXECUTE phase
deleted code although Sweep "Writes Code: Never" and its protocol says
Builder executes; Tuner and Schema both owned index design. Nexus also had
chain variant families (QA, TEST, UX_DESIGN) with no matrix task type, so
SELECT could never reach them.

Align each with the existing authority: Quill keeps ADR upkeep, Sweep
hands batches to Builder, Tuner recommends from observed plans while
Schema designs up front, a TEST matrix row anchors the test chains and
UX_DESIGN variants fold into existing types. Also record Orbit's Cast
SPEAK use as a runtime dependency, collect Gauge's S10/S11 inputs during
SCAN, and name real phases in four OPUS rows.

* fix(routing): check subcommand pointers corpus-wide

RO-9 only checked routing-matrix hints, so the same defect survived
elsewhere: scope-boundary lines and handoffs named subcommands their
target never defined (Shift `migrate`, Siege `mutate`, Oracle `eval`,
Builder `automate`) or put a recipe's mode where the subcommand goes
(echo `jtbd` instead of `demand jtbd`, Magi[expert] instead of
Magi[advisor expert]). Dispatch falls back to the default recipe on an
unknown first token, so each one silently ran the wrong work.

Fix the 25 pointers and add RO-10, which checks the bold
**Skill `token`** convention and recipe-mode-in-subcommand-position across
every skill, reference and shared protocol, with failing-case tests.

* fix(refs): correct m–z reference facts and pointers

A full mechanical pass over the m–z reference files (code-block parsing,
pointer resolution, outdated-tech and citation greps) found claims that
would mislead the skills that load them: web_fetch called unsupported on
Opus 5 and budget_tokens called deprecated where it is removed, OWASP
2025 CWE counts and a misplaced prompt-injection entry, FID and RFC 7807
presented as current, an XFetch formula with beta inverted, Aurora and
pg_stat_statements queries using columns that do not exist, GDPR/CCPA
response deadlines misstated, and an unterminated bash example.

Also re-point references to deleted or renamed files (sigil, tome, magi,
native) and add the REFINE phase Weave's AUTORUN schema omitted.

* fix(refs): correct a–l reference facts and pointers

The a–l reference files carried stale and wrong guidance the skills load
as authority: EU AI Act high-risk dates superseded by the Digital Omnibus
(Annex III now 2 Dec 2027), a HIPAA Security Rule date HHS has moved,
Node 20 defaults and actions majors past their runtime EOL, Node release
lines misdated, a JWT example using a library without EdDSA, FIPS 140-2
for KMS HSMs, an A/B sample-size table 2.5x off its own formula,
outdated Sonnet pricing, renamed OTel GenAI attributes, and an OWASP
table headed 2025 but numbered 2021.

Also fix 21 funnel pointers to pre-merge file names, a grep -E pattern
using a lookahead it cannot parse, and two reference/SKILL.md threshold
contradictions (ink icon budget, gear setup-node).

* fix(breach): align AI Act dates with the Omnibus

Canon's regulatory reference now records the Digital Omnibus deferral of
Annex III high-risk obligations to 2027-12-02, but Breach still told red
teams that high-risk and systemic-risk GPAI testing was due together by
2026-08-02 and quoted the prohibited-practice penalty tier. Split the two
regimes and point at Canon as the source. Also fix Gear's capability line,
which still named setup-node v4 for the Node 24 migration its own core
contract says needs v5+.

* fix(breach): cite Canon regulatory without a path

A SKILL.md may only name reference paths that resolve from its own
directory (lint-contracts CD-4); refer to Canon's regulatory recipe
instead.

* fix(routing): resolve canon[legal] to real recipes

Canon has no `legal` subcommand, so the 23 `canon[legal]` pointers in
shared protocols and Nexus recipes all dispatched to Canon's default
OWASP recipe. Point each at the recipe its context means (claims for
advertising claims, legal-gap for multi-document pre-launch review,
tos/privacy/cookie for specific documents), and name qualified counsel
where no recipe covers the work (patents, licensing, brand safety).
Stage pointers to nonexistent draft/theme recipes now name the format
recipes.

Extend RO-10 to backticked `skill[token]` invocations, which the corpus
uses only for real subcommands, with a failing-case test.

* fix(skills): restore words lost to skill renames

Retiring skills replaced each name with its successor across the corpus,
including where the name was an ordinary word or product: "WHERE canon"
for WHERE clause, "launch-now-decrypt-later" for harvest-now, "credential
launch", Relay (GraphQL) became Gateway, Grafana Tempo became Weave,
Logstash grok patterns became builder patterns, ZAP's and Screaming
Frog's Spider became Vector, and absorbed pairs left lists like
"Flux/Flux/Magi". Security, database and observability guidance read
wrong as a result.

Word-diff every rename commit and each archived original against the
live text, restore the original word wherever it was not the skill, and
collapse duplicated skill names left behind by absorptions.

* fix(scripts): tighten RO-10 modes and legacy fallback

RO-10 harvested any backticked two-word phrase in a SKILL.md as a
`subcommand mode` declaration, so Judge's trigger phrase `security
review` -> security focus made `review` a mode and an ordinary
Judge[review] annotation failed validation. Read modes only from the
Recipes and Subcommand Dispatch sections, skipping trigger cells and
phrases that are themselves mapped.

token-economy.py fell back to this repo's legacy transcript directory
even for an explicit --repo-root, reporting another repository's usage
as the target's with exit 0. Only the default repo uses the fallback.

---------

Co-authored-by: Claude <noreply@anthropic.com>