Repository navigation
Releases
Agent Skills v1.9.1
v1.9.1 — 2026-11-12
- Harden Ubuntu 24.04 STIG remediation and network and GRUB safeguards, with expanded failure-path tests.
- Tighten skill-name, script, secret-scan and CI example checks, and reject malformed release-note versions.
- Harden Confluence page fetching and update Kubernetes, PostgreSQL and Next.js deployment guidance.
- Pin validation images and link tagged GitLab releases to the durable repository source archive.
- Fix release publishing to trust the Runner-provided GitLab CA and read the release-note file correctly.
v1.0.0
v1.6 — quality-sync Error Handling + confluence-to-nextjs scripts
Quality-bar sync from the ai-optimization skill dashboard's automated rubric pass — Error Handling sections across 5 skills, plus a real script extraction for confluence-to-nextjs.
Error Handling sections added
arcgis-enterprise-k8s— 6-row troubleshooting table: PVC stuckPending, invalid license, LoadBalancer stuck<pending>, TLS/FQDN mismatch after deploy, upgrade rollout hang,webgisdrexport failure.confluence-to-nextjs— 7 failure modes: 401/403, wrong page ID (404), 429 rate limiting on bulk exports, unmapped<ac:structured-macro>types, broken image/attachment links, malformed storage HTML breaking JSX, unmigrated internal<ac:link>targets.k8s-nextjs-deploy— new "Deploying a new Next.js app — step by step" (6 steps, namespace → pull secret → Deployment → Service → Ingress → verify), each with an explicit success check.kilo-gateway— 7-row signal/cause/recovery table: 404 misrouting, 401 key issues, 402 balance depletion, 429 free-tier limiting, BYOK silent failure, missingusagefields, hung streams.nextjs-monorepo-ci— new "Adding a new app to the pipeline — step by step" (7 steps: scaffold → validate → build → obfuscate → package → notify → IndexNow key), matching the pipeline's own stage order.postgres-ops— 7 failure modes tied to the diagnostic workflows already documented: connection exhaustion blocking diagnosis itself,EXPLAIN ANALYZEexecuting writes for real, missingpg_stat_statements,CREATE INDEX CONCURRENTLYleftINVALID,pg_upgrade --checkextension mismatches, pgBouncer transaction-pooling prepared-statement errors, empty replication-lag reads.
confluence-to-nextjs script extraction
- Added
scripts/fetch-page.sh— parameterized Confluence REST API v2 fetch (pinsbody-format=storage, fails loudly with the response body on 401/403/404 instead of writing an error payload silently to disk). - Added
scripts/slugify.mjs— the heading-to-anchor-ID slug generator from Step 2, plusdedupeSlugs()for the numeric-suffix collision handling described in Anti-pattern 6. - Step 1 and Step 2 in
SKILL.mdnow reference these scripts instead of inline snippets.
v1.5 — Magnific generative-media API skill
New skill — magnific
The Magnific API (https://api.magnific.com/v1, part of Freepik): image
generation, AI upscaling, video, audio, and stock content through one
async task API.
SKILL.mdfront-loads the submit →task_id→ poll-or-webhook model,
x-magnific-api-keyauth, a robust poll helper, a working Svix-style
HMAC-SHA256 webhook verifier (Python + Node), the full model catalog,
a pre-flight checklist, and a 9-item anti-patterns gallery (empty
generatedarray, sync Remove-Background,task_statuson Style
Transfer, NSFW defaults, 25.3 MP upscaler cap, Runway pixel-ratio
strings).- Five
references/files document all ~60 endpoints verbatim
(parameters, enums, defaults, request/response examples):
foundations,image-generation,image-editing,
video-generation,audio,analytics-and-stock. - Covers Mystic, Flux ×7, Seedream, Z-Image, the Creative/Precision
upscalers, Relight, Style Transfer, Remove Background, Image Expand,
Kling, MiniMax, WAN, Runway, LTX, Seedance, PixVerse, OmniHuman, VFX,
music/SFX/isolation, team analytics, and stock icons/videos/templates
— plus rate limits, the credit pricing model, and the hosted MCP
server.
v1.4 — Loki memcached cache right-sizing
loki-best-practices — memcached cache right-sizing
New operational knowledge in the loki-best-practices skill:
- The
grafana/lokiHelm chart shipschunksCache.allocatedMemory: 8192and
resultsCache.allocatedMemory: 1024(MB). That number is both memcached's
-mflag and the pod's memory request and limit (≈allocatedMemory × 1.2Mi),
so the two caches reserve ~11Gi per cluster regardless of log volume — the
single biggest memory reservation in the namespace and a common scheduling blocker. - Fix: right-size to
chunksCache.allocatedMemory: 1024+resultsCache.allocatedMemory: 256
(~1.5Gi total) via a two-linehelm upgrade --reuse-values -foverlay; the memcached
-mand the podresourcesmove together.
Captured in section 2b. Deployment & configuration and added to the Most common
pitfalls list (now 8). Drawn from a real fleet incident — a 16Gi worker pinned at
89% MEM requests was almost entirely the Loki chunk-cache pod.
v1.3 — Quality pass across all 12 skills
What's new in v1.3
Quality pass across all 12 skills — every skill now meets the full quality bar.
New: Anti-patterns galleries
Five-to-seven concrete failure modes per skill — things that look in-scope but silently fail, corrupt data, break production, or waste money. Added to all skills that were missing it:
postgres-ops— VACUUM FULL locking, directUpdate-Databasein prod, work_mem misconfiguration, untested pg_dump restores, pgBouncer prepared-statement trap, pg_upgrade --link hazard, missingCONCURRENTLYon index creationconfluence-to-nextjs—body-format=viewvsstorage, unslugified anchor IDs, transclusion macros, duplicate headingidcollisions, stale sitemapubuntu24-stig— FIPS mode on live EC2, wrong SCAP benchmark path,notapplicablemasking real failures,journalctlpermission blocking monitoring agents, PAM lockout without SSM fallback, AIDE baseline on dirty state,auditctl -Rnot persistedarcgis-enterprise-k8s—kubectl scalebypassing ArcGIS state machine, post-deploy FQDN change, wrong storage class, namespace delete leaving CRDs, skipping PVC snapshot before upgrade, L4/L7 mix, initial admin credential rotation
New: Pre-flight checklists (SDK skills)
Ordered verification steps to run before writing any API code — added to anthropic-sdk, openai-sdk, kilo-gateway.
New: Worked end-to-end examples (SDK skills)
Complete runnable code with a triage decision table showing keep/continue/done paths — added to anthropic-sdk, openai-sdk, kilo-gateway.
Model updates in anthropic-sdk
claude-opus-4-7→claude-opus-4-8(current Opus model)claude-fable-5added to the model selection table (best for autonomous agentic workflows)
Install
for skill in postgres-ops nextjs-monorepo-ci k8s-nextjs-deploy confluence-to-nextjs ubuntu24-stig login-gov arcgis-enterprise-k8s loki-best-practices session-handoff anthropic-sdk openai-sdk kilo-gateway; do
npx skills add "andreab67/agent-skills@$skill" -g -y
donev1.2 — provider SDK skills (anthropic-sdk, openai-sdk, kilo-gateway)
New skills
anthropic-sdk
Anthropic Python SDK: Messages API, streaming, multi-turn tool use loop, count_tokens() pre-flight, prompt caching with cache_control: {type: ephemeral}, vision input. Includes Opus 4.7 / Sonnet 4.6 / Haiku 4.5 pricing table and a cost estimation helper.
npx skills add andreab67/agent-skills@anthropic-sdk -g -yopenai-sdk
OpenAI Python SDK: chat completions, streaming, function calling with tool_call loop, JSON structured output, embeddings, tiktoken count_message_tokens() helper. GPT-4o / mini / o1 / o3 pricing table and cost estimation.
npx skills add andreab67/agent-skills@openai-sdk -g -ykilo-gateway
Kilo.ai unified gateway — OpenAI-compatible API at https://api.kilo.ai/api/gateway routing to 500+ models across 14+ providers. BYOK setup (zero markup), provider/model-name format, Python/TypeScript/cURL/LangChain examples, prompt-cache X-KiloCode-TaskId header.
npx skills add andreab67/agent-skills@kilo-gateway -g -yFull changelog
See CHANGELOG.md for all release notes.
v1.1 — session-handoff skill
Adds
session-handoff
Manually-invoked skill that rescues critical session context to persistent auto-memory before /clear, /compact, or any conversation wipe.
The whole skill rides on one heuristic: would re-reading the repo recover this? If yes, drop. If no, it's a rescue candidate. That single test eliminates ~80% of false saves.
Highlights:
- Six rescue categories mapped to the
user/feedback/project/referenceauto-memory schema - Hard exclusions for secrets, code-derivable facts, git history, fixed-bug recipes, and CLAUDE.md duplicates — refused even when explicitly asked
- Pre-flight read of
MEMORY.mdto reuse slug prefixes, detect duplicates, and surface contradictions with stale entries - Review-before-write via
AskUserQuestion: multi-select for ≤ 4 candidates, table format for > 4, with explicit drop reasons shown - Conflict resolution that updates existing entries in place rather than writing duplicates
- Anti-patterns gallery (7 examples a senior engineer would refuse)
- Worked end-to-end example: 90-minute design session → 11 candidates → 6 keep, 5 drop, with the exact triage table
- Optional
UserPromptSubmithook (hooks/session-handoff-gate.js) that intercepts the first/clearand routes to the skill — second/clearpasses through
Install
npx skills add andreab67/agent-skills@session-handoff -g -yFull changelog
- 0ceff0a feat: add session-handoff skill — rescue critical context before /clear
v1.0 — first tagged release
First tagged release. The repo now ships 8 production-grade agent skills for Claude Code and other LLM agents, and is open for outside contributions.
Skills in this release
| Skill | Purpose |
|---|---|
postgres-ops |
Production PostgreSQL SRE: incident diagnosis, query plans, HA/DR, migrations, pgBouncer, observability, STIG |
nextjs-monorepo-ci |
GitLab CI/CD for Next.js monorepos: 6-stage pipeline (validate→security→build→obfuscate→package→notify) with Kaniko, Semgrep, gitleaks, Harbor |
k8s-nextjs-deploy |
Kubernetes patterns for Next.js: Deployment/Service/Ingress, Harbor pull secrets, Traefik + cert-manager, Linkerd, failure diagnosis |
confluence-to-nextjs |
Migrate Atlassian Confluence pages to a self-hosted Next.js KB with REST API v2 + storage-format → JSX conversion |
ubuntu24-stig |
DISA STIG V1R1 hardening for Ubuntu 24.04 on AWS EC2 (OpenSCAP, MAC-2_Sensitive remediation, auditd, FIPS, AIDE) |
login-gov |
login.gov OIDC integration: private_key_jwt, ACR values (auth-only/IAL2/facial-match/PIV-CAC), Partner Portal setup |
arcgis-enterprise-k8s |
ArcGIS Enterprise on Kubernetes 12.0: deployment profiles, EKS/AKS/GKE/OpenShift sizing, TLS, upgrades |
loki-best-practices (new) |
Grafana Loki SRE: reason-keyed ingest-discard triage, WAL/flushing, schema/Helm pitfalls, OTel push debugging, query perf, S3/IRSA, ring HA, + the 10-command diagnostic cheatsheet |
Install
for skill in postgres-ops nextjs-monorepo-ci k8s-nextjs-deploy confluence-to-nextjs ubuntu24-stig login-gov arcgis-enterprise-k8s loki-best-practices; do
npx skills add andreab67/agent-skills@$skill -g -y
doneContributing
The repo is now open for outside contributions. See CONTRIBUTING.md for skill-layout conventions, voice guidelines, and PR scope rules. The bar for a new skill is "this would have saved me 3 hours during an incident" — open an issue first to scope the trigger surface before writing the full SKILL.md.
License
BSD 3-Clause.