CubeCloud Skills Bundle：一条命令为 VS Code Copilot Chat 装齐 275 个技能

One-command setup for a full VS Code Copilot Chat agent-skills stack on Windows — 275 bundled skills (302 on disk), 20 CLIs, 12 MCP servers, and a 74-site design-system library, all security-gated.

Why this exists

VS Code Copilot Chat gets dramatically more powerful when you give it skills — markdown instruction packs that teach it domain-specific workflows (TDD, systematic debugging, design systems, Azure patterns, PR review). But assembling a trustworthy stack by hand is painful: you have to find skills, vet them for safety, wire up the MCP servers, install the CLIs, and repeat on every machine.

CubeCloud Skills Bundle does all of it in one command. Every skill passes through NVIDIA SkillSpector before landing on your machine — vulnerable skills are blocked by design, not after the fact.

What you get

Architecture

graph TB
    subgraph L4["LAYER 4 — Project Bootstrap"]
        SPEC["specify init --integration copilot"]
        AGENTS["10 .agent.md commands"]
        TEMPLATES[".specify/ template tree"]
        SPEC --> AGENTS
        SPEC --> TEMPLATES
    end

    subgraph L3["LAYER 3 — Agent Commands (per-project)"]
        SD["/speckit.specify → /speckit.plan → /speckit.tasks"]
        IM["/speckit.implement → /speckit.converge"]
        EXT["extensions · presets · bundles"]
    end

    subgraph L2["LAYER 2 — Skills (275 bundled)"]
        direction LR
        METH["superpowers (12)"] --> COPILOT["VS Code Copilot Chat"]
        DSN["design systems (74)"] --> COPILOT
        REV["code review (6)"] --> COPILOT
        DBG["debugging (4)"] --> COPILOT
        LOOP["loop engineering (5)"] --> COPILOT
        AZURE["Azure/cloud (27)"] --> COPILOT
        CRAFT["crafted (16)"] --> COPILOT
        DEC["decision models: jev + laya (7)"] --> COPILOT
    end

    subgraph L1["LAYER 1 — CLIs + MCP Servers + Decision Sidecars"]
        direction LR
        CLI["20 CLIs on PATH"] --> MCP["12 MCP servers"]
        MCP --> VSCODE["VS Code mcp.json"]
        SC["laya-serve: decision sidecar on 127.0.0.1:8770 (opt-in, AMD ROCm)"] --> COPILOT
    end

    subgraph L0["LAYER 0 — Design Systems"]
        D74["74 DESIGN.md files"] --> LIB["design-md-library skill"]
    end

    L4 --> L3
    L3 --> L2
    L2 --> L1
    L1 --> L0

    style L4 fill:#1a3a2a,stroke:#4caf50,color:#e0ffe0
    style L3 fill:#1a2a3a,stroke:#2196f3,color:#e0f0ff
    style L2 fill:#2a1a3a,stroke:#9c27b0,color:#f0e0ff
    style L1 fill:#3a3a1a,stroke:#ff9800,color:#fff0e0
    style L0 fill:#3a1a1a,stroke:#f44336,color:#ffe0e0

Loading

Layer 4 is where spec-kit lives — the only component that bridges "new project idea" → "structured, agent-ready project." Without it, every project starts as a blank prompt. With specify init, Copilot gets 10 native workflow commands, an SDD template stack, and an extension ecosystem.

Quick start

git clone https://github.com/JZKK720/cubecloud-skills-bundle-kit.git ~/dev/setup
powershell -NoProfile -ExecutionPolicy Bypass -File ~/dev/setup/setup/setup-global-skills.ps1

Then restart VS Code, open Copilot Chat, and type # to see your MCP tools appear.

Total time: ~15 minutes (or ~10 min with -SkipForks).

Prerequisites

The script checks for and (where possible) installs these. Pre-install on a fresh machine:

winget install Python.Python.3.12
winget install Python.Python.3.13
winget install OpenJS.NodeJS
winget install Git.Git
winget install Microsoft.VisualStudioCode

Python 3.12 is used for guarded headroom-ai installs to avoid the unstable 3.13 build path.

What's included

Skills (275 bundled — 274 active + 1 disabled)

Why the README says 275 but the installer prints 302. The installer ends with (Get-ChildItem ~/.agents/skills -Directory).Count, which counts everything in that folder — including the 27 skills VS Code installs itself for the Azure, Entra, and Foundry extensions. Those are already on disk on any normal install and the bundle neither writes nor removes them (README §"Azure & cloud" below). 275 + 27 = 302. Both numbers are correct; they just measure different sets.

A further 45 gate-blocked entries are parked as comments inside skills-list.csv (not installed, not counted above) with a per-skill reason code — see the blocked-skills table.

Superpowers methodology (12 skills) from obra/superpowers: test-driven-development · systematic-debugging · writing-plans · executing-plans · subagent-driven-development · requesting-code-review · receiving-code-review · using-git-worktrees · finishing-a-development-branch · writing-skills · using-superpowers · dispatching-parallel-agents

ECC agent engineering (35 skills) from evan-ai/ECC — curated from 278, framework-agnostic only: safety-guard · token-budget-advisor · intent-driven-development · verification-loop · eval-harness · agent-self-evaluation · prompt-optimizer · rules-distill · knowledge-ops · codebase-onboarding · repo-scan · code-tour · search-first · blueprint · strategic-compact · enterprise-agent-ops · production-audit · error-handling · delivery-gate · coding-standards · context-budget · security-review · security-scan · security-bounty-hunter · brand-discovery · brand-voice · frontend-design-direction · make-interfaces-feel-better · continuous-agent-loop · cost-tracking · cost-aware-llm-pipeline · automation-audit-ops · connections-optimizer · mcp-server-patterns · backend-patterns

oz-skills (14 active) from the JZKK720 fork mirror: analysis-artifacts · ci-fix · create-pull-request · dbt-model-index · docs-update · github-bug-report-triage · github-issue-dedupe · mcp-builder · scheduler · seo-aeo-audit · slack-qa-investigate · terraform-style-check · web-accessibility-audit · web-performance-audit

ui-skills (7 active) from ibelick/ui-skills: ui-skills-root · baseline-ui · create-design-md · fixing-accessibility · fixing-metadata · fixing-motion-performance · improve-ui

agent-skills (23 active, unique entries) from addyosmani/agent-skills: api-and-interface-design · browser-testing-with-devtools · ci-cd-and-automation · code-review-and-quality · code-simplification · context-engineering · debugging-and-error-recovery · deprecation-and-migration · documentation-and-adrs · doubt-driven-development · frontend-ui-engineering · git-workflow-and-versioning · idea-refine · incremental-implementation · interview-me · observability-and-instrumentation · performance-optimization · planning-and-task-breakdown · security-and-hardening · shipping-and-launch · source-driven-development · spec-driven-development · using-agent-skills

Azure & cloud (27 skills) — bundled with VS Code Azure extensions, discovered automatically: ai-mlstudio · airunway-aks-setup · appinsights-instrumentation · azure-ai · azure-aigateway · azure-cloud-migrate · azure-compliance · azure-compute · azure-cost · azure-deploy · azure-diagnostics · azure-enterprise-infra-planner · azure-kubernetes · azure-kusto · azure-messaging · azure-prepare · azure-quotas · azure-reliability · azure-resource-lookup · azure-resource-visualizer · azure-storage · azure-upgrade · azure-validate · entra-agent-id · entra-app-registration · microsoft-foundry · python-appservice-deploy

These Azure/Foundry entries are standard extension-provided skills, not custom bundle ports.

Crafted individual skills:

• self-learning — capture hard-won workflows as reusable skills
• improve — audit a codebase into prioritized implementation plans
• loopy — discover, run, and publish repeatable agent loops
• loop-engineering — design, scaffold, audit, and operate agent loop infrastructure (scheduling, isolation, scoring, governance). Complements loopy (loop content) with loop infrastructure.
• watch-skill — video intelligence for agents: watch, remember, verify. 23 MCP tools for video analysis, transcription, OCR, and THE LOOP (browser/UI verification).
• wigolo — local-first web intelligence: search, fetch, crawl, extract, research. 10 MCP tools, no API keys needed for core tools. Complements firecrawl (paid) with a free local alternative.
• ponytail — force the laziest solution that actually works (YAGNI)
• hallmark — anti-slop UI design for landing pages and redesigns
• taste-skill — frontend taste/polish, anti-templated output
• karpathy-guidelines — coding behavioral guidelines
• agent-reach — research across 15 platforms (Twitter, Reddit, YouTube, GitHub, LinkedIn, Xueqiu, Bilibili, XiaoHongShu, and more)
• graphify — turn any folder into a knowledge graph
• browser-harness — full CDP browser automation (navigate, extract, interact)
• loop-engineering (5 skills) — agent self-management: triage, constraints, budget, verification, minimal-fix
• ARIS ports (3 skills) — academic research workflow: literature search, novelty check, idea generation
• changelog-generator — auto-generate user-facing changelogs from git history

Custom implementations maintained in this setup are agent-reach and gstack-review.

Hand-ported:

• gstack-review — pre-landing PR review with structural-issue checklist + 8 specialist lenses (security, testing, maintainability, performance, data-migration, api-contract, red-team). Adapted from garrytan/gstack (MIT).

• design-md-library — indexes the 74 DESIGN.md files in the awesome-design-md fork mirror so agents can self-serve "make it look like Stripe" requests

• idea-to-design — clean port of obra/superpowers brainstorming (upstream blocked by SkillSpector for tool parameter abuse in stop-server.sh). Methodology only: collaborative design dialogue, hard gate before implementation, spec self-review, user review gate. No browser server, no scripts.

• webapp-testing — clean port of JZKK720/oz-skills webapp-testing (upstream blocked by SkillSpector for shell=True tool parameter abuse in scripts/with_server.py). Methodology only: reconnaissance-then-action pattern, static vs dynamic decision tree. No bundled scripts; agent writes native Playwright or uses browser MCP tools.

• archify — clean port of tt-a1i/archify (MIT, v2.17). Upstream is blocked by SkillSpector (HIGH MP3), so the port keeps the authoring contract, type router, invariant list, Mermaid conversion rules and the validate → deliver → visual-check sequence, and drops the renderer/schemas/examples. Gates 0/100 SAFE. Methodology only — the upstream CLI lives in the fork mirror; the SKILL.md says so explicitly rather than implying an unrun validation passed.

  Optional: the archify CLI. The renderer itself can also be exposed as a bare command:

  powershell -NoProfile -ExecutionPolicy Bypass -File .\setup-global-skills.ps1 -IncludeArchifyCli

  This writes one 86-byte ~/.local/bin/archify.cmd shim that forwards to the fork mirror's bin/archify.mjs. It is opt-in because every other phase of the installer is unconditional and this one adds a command to your PATH. archify is a zero-runtime-dependency Node CLI — bin/archify.mjs imports only node: stdlib, and the mirror has no node_modules and needs none — so the shim needs no package manager, no registry, and no global npm install. (An npm install is not even possible: upstream package.json sets "private": true.) It requires the archify fork mirror, so it is incompatible with -SkipForks; the installer warns and continues rather than failing. The installer runs archify doctor before writing the shim, so it will not advertise a command it has not verified.

  archify doctor
  archify render architecture diagram.json out.html
  archify validate architecture diagram.json --quality showcase --json

tech-leads-club/agent-skills (12 active) from tech-leads-club/agent-skills (MIT):

• Domain-driven design, installed as a set (6) — domain-analysis · domain-identification-grouping · coupling-analysis · decomposition-planning-roadmap · tactical-ddd · modular-design-principles. These reference each other, so a partial subset would leave dangling references.
• Planning and evaluation (6) — create-technical-design-doc (1055 lines) · create-rfc · harness-eval · spec-driven-eval · not-your-babysitter · learning-opportunities

marketingskills (8 active) from coreyhaines31/marketingskills (MIT): product-marketing · copywriting · content-strategy · competitor-profiling · launch · pricing · marketing-loops · marketing-psychology

General marketing discipline in, platform-specific execution out. The repo's tools/ tree (60+ optional integration scripts) is deliberately not installed.

openai/skills (5 active) from openai/skills (Apache-2.0 per skill): figma · screenshot · pdf · transcribe · security-threat-model

The repo root has no LICENSE, but every skill directory ships its own LICENSE.txt and the adopted set is Apache-2.0, matching the README. Provenance is recorded deliberately because the root looks ambiguous.

firstmate (4 active) from kunchenguid/firstmate (MIT): bearings · diagnostic-reasoning · bootstrap-diagnostics · project-management

Four of 22 — the rest are bound to firstmate's own supervisor daemon and bin/fm-*.sh scripts, with one gate block (stow) and three high-severity findings.

humanlayer/skills (3 active) from humanlayer/skills (MIT): show-me · improve-claude-md · narrow-react-prop-types

Prose de-slopping (2 active):

• humanizer — blader/humanizer (MIT, 291 lines). Wikipedia's "Signs of AI writing": numbered tells ranked strongest-first with explicit weak-alone marking, plus voice preservation and a fiction exemption.
• no-ai-slop — petergyang/no-ai-slop (MIT, 61 lines). Edit and detect modes, banned word list, pattern catalogue, eval.md self-check. Gates 0 findings at any severity.

Two independent implementations kept as complements. The bundle's write-concisely covers density and brand-voice covers deriving a voice from source material, but neither de-slopped existing prose.

Disabled by default:

• caveman — token compression, opt-in only

Blocked by SkillSpector (not installed — by design):

Added v1.8.0 — blocked or high-risk entries found in the 14-repo evaluation:

Also skipped for HIGH findings even though the gate returned 0: the-fool, the-jury (HIGH AR2 anti-refusal), mermaid-studio (HIGH PE3), codenavi and react-composition-patterns (HIGH MP3), cloudflare-deploy / sentry (HIGH E2), vercel-deploy / react-native-expert (HIGH PE3), core-web-vitals / seo / web-best-practices (HIGH P2), security-best-practices / react-best-practices (HIGH OH1), create-adr / tlc-spec-driven / figma-implement-design (HIGH E4), speech (HIGH P6), jupyter-notebook (HIGH RA1), cli-creator (HIGH AS1), sentry (HIGH SC2), marketing-plan (HIGH AR1), harness-adapters (HIGH AS1), quota-array-dispatch / updatefirstmate (HIGH RA1), build-iterated-agentic-loop / design-control-loop (HIGH E2).

AR2 (anti-refusal) is a consistent tell across catalogs: skills that instruct adversarial or persistent behaviour tend to phrase it as an override of the model's own judgement. The policy is to exclude on AR2, and where the value is high enough, port the methodology without the override.

These manifest entries are advertised by the install manifest but blocked by the SkillSpector hard gate at install time (verified 2026-09-13). They are not a defect — the gate is doing its job. The 9 ce-* entries and the 2 taste-* entries below have been pruned from the manifest so install-missing-skills.ps1 no longer retries them on every run; they are listed here so the gate result stays visible.

Curated out by scope (not blocked — deliberately excluded)

These pass the gate but are excluded by the same curation rule the ECC block already applies ("framework-agnostic only; framework-specific, ECC-internal, and domain-niche excluded"). Excluding them keeps the bundle coherent.

CLIs installed

MCP servers

Configured in VS Code User mcp.json:

• markitdown — convert anything to Markdown
• skillspector — skill security scanning
• firecrawl — web scraping/crawling
• scrapling — stealthy fetching
• gbrain — persistent memory
• graphify — codebase knowledge graphs
• headroom — context compression (headroom_compress, headroom_retrieve, headroom_stats)
• loop-engineering — loop pattern lookup, skills, state (@cobusgreyling/loop-mcp-server)
• watch-skill — video analysis, transcription, OCR, THE LOOP verification (watch-skill serve)
• wigolo — local-first web search, fetch, crawl, research (npx wigolo)
• skillopt — SkillOpt research engine: validation-gated skill optimization via MCP tools skillopt_list_configs, skillopt_train, skillopt_eval (microsoft/SkillOpt, MIT). Shells out to the repo's scripts/train.py / eval_only.py — requires the SkillOpt fork mirror.

OpenCodeReview (AI code review)

The bundle includes the ocr CLI (@alibaba-group/open-code-review, Apache-2.0) — a deterministic + agent hybrid code review engine battle-tested at Alibaba's scale. Two skills ship with it:

• open-code-review — invokes ocr review with the right flags, prerequisite checks, and a comment-triage rubric (High/Medium/Low). Requires a configured LLM.
• open-code-review-delegate — delegation mode: the host agent drives the review using its own LLM; OCR handles only deterministic file selection and rule resolution. No OCR-side LLM needed.

ocr review --audience agent -b "context"           # review working copy
ocr review --audience agent -b "context" --from main --to feature  # branch range
ocr review --preview                                # dry-run (no LLM cost)
ocr delegate preview                                # delegation mode preview

SkillOpt-Sleep (nightly self-evolution)

The setup also schedules SkillOpt-Sleep — a nightly offline cycle that harvests past session transcripts, mines recurring tasks, replays them, and stages a validated skill edit for review (adopt is manual by design). Installed as a Windows Scheduled Task (schtasks) at 03:17 daily via skillopt-sleep schedule --project ~/dev --backend mock. The mock backend spends no model budget; re-schedule with --backend claude or --backend copilot once credentials are configured:

skillopt-sleep status      # show state + latest staged proposal
skillopt-sleep adopt       # apply the latest staged proposal (with backup)
skillopt-sleep unschedule --all   # remove the nightly task
skillopt-sleep schedule --project ~/dev --backend claude --hour 3 --minute 17

Logs land in ~/dev/.skillopt-sleep/cron.log. Sleep only STAGES proposals — no skill changes until you run adopt.

Security model

Every skill is scanned by NVIDIA SkillSpector before install:

skillspector scan --no-llm <dir>

• Exit 0 → safe → install proceeds
• Exit 1 → do_not_install → HARD BLOCK, skill is not installed
• Exit 2 → error → investigate before retrying

skills-ref validate runs as an advisory check (logs spec drift, doesn't block).

Full verdict history is in upstream/SCAN_LOG.md.

Repository layout

cubecloud-skills-bundle-kit/          # git clone target — run installers from this root
├── setup/                            # the one-command installer + config
│   ├── setup-global-skills.ps1       # master installer (run this)
│   ├── install-skill.ps1             # security-gated skill install helper
│   ├── skills-list.csv               # manifest of 275 entries (274 active + 1 disabled)
│   ├── mcp.json.template             # 12 MCP server config
│   ├── skillspector-ollama-models.yaml  # token-budget metadata for local Ollama scans
│   └── SETUP_GUIDE.md                # detailed guide
├── bin/                              # 27 audit / fix / install helper scripts
├── upstream/                         # governance docs (SCAN_LOG, audits) + 51 hand-authored port skills
├── docs/  plans/                     # design notes and integration plans
├── Update-Skills.ps1                 # thin CSV-driven wrapper (legacy)
├── AGENTS.md  CHANGELOG.md  LICENSE
└── README.md

# At runtime the installer ALSO creates (outside the repo):
~/dev/bin/install-skill.ps1           # deployed helper copy
~/dev/upstream/SCAN_LOG.md            # security scan log
~/dev/forks/JZKK720/                  # 50 read-only fork mirrors (gitignored, re-cloned)

Path convention: run installer commands from the repo root (setup\setup-global-skills.ps1), and audit helpers from the same root (.\bin\<script>.ps1). The ~/dev/setup/... form is only correct if you cloned the repo into ~/dev/setup/, in which case the script is at ~/dev/setup/setup/setup-global-skills.ps1.

How to use after setup

In Copilot Chat, try:

• "use the improve skill to audit this codebase"
• "use systematic-debugging to investigate this error"
• "use the design-md-library to build me a page that looks like Stripe"
• "use gstack-review to review my PR"
• "use agent-reach to research what people are saying about X on Reddit"
• "use specify init my-app --integration copilot to scaffold a new SDD project"
• "then /speckit.specify Build a photo organizer with album grouping and drag-and-drop"
• "use loop-engineering to set up automated daily triage on this repo"
• "use watch-skill to analyze this meeting recording"
• "use wigolo to research what's new in React 19"
• "use domain-analysis to find the bounded contexts in this codebase"
• "use create-technical-design-doc to write the design doc for this feature"
• "use humanizer to make this README not sound like AI wrote it"
• "use figma to implement this design at 1:1 parity"
• "use product-marketing to sharpen our positioning for the OEM page"
• "run **archify validate architecture diagram.json --quality showcase"** (requires -IncludeArchifyCli)

To update later

# Standard safe update pass (guarded headroom update + skills-ref repair + gbrain update + smoke)
powershell -NoProfile -ExecutionPolicy Bypass -File .\bin\safe-update-pass.ps1

# Or re-run the installer (skips already-installed items)
powershell -NoProfile -ExecutionPolicy Bypass -File ~/dev/setup/setup/setup-global-skills.ps1

# Manual fallback (if you need step-by-step control)
uv tool upgrade --all
npm update -g
bun install -g github:garrytan/gbrain

# Re-scan all skills quarterly
skillspector scan ~/.agents/skills/ --recursive --no-llm

To add a new skill

cd ~/dev/bin
.\install-skill.ps1 -Repo "owner/repo" -Name "skill-name"
# Disabled:
.\install-skill.ps1 -Repo "owner/repo" -Name "skill-name" -Disabled
# Custom path:
.\install-skill.ps1 -Repo "owner/repo" -Name "skill-name" -SkillRelPath "path/to/skill"

Using Jev (typed decisions)

Jev is TypeSafe's System One model: it takes a state plus typed questions and returns choice (with per-option probabilities), score, or noul (0–1). It does not generate text or write code — it is not a replacement for your coding-agent LLM, and the vendor says so explicitly. You call it from code your agent writes.

Two skills cover it: typesafe-ai (the official vendor skill — API, primitives, patterns) and jev-harness (build measurable decision pipelines, optionally evolved against outcomes).

from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

client = TypeSafeClient(api_key=KEY, base_url="https://openrouter.ai/api")
resp = client.system_one(
    model="jev-1.13",
    state="Help! My payouts have been failing for 3 days.",
    questions={
        "is_urgent": Noul(instructions="Does this convey urgency?"),
        "department": Choice(
            instructions="Which team should handle this?",
            criteria={"billing": "Payments", "technical": "Bugs"},
        ),
        "frustration": Score(
            instructions="How frustrated is the customer?",
            criteria=["Calm", "Frustrated", "Very angry"],
        ),
    },
)
resp.answers["department"].choice        # 'billing'
resp.answers["department"].confidence    # 0.84
resp.answers["is_urgent"].noul           # 0.97

Routes. Point the official TypeSafe SDK at whichever provider you have:

Choice.criteria is a mapping (id → description); Score.criteria is a sequence (2–10 levels). The method is system_one() — not ask().

No API key? The jev skill documents a keyless simulation route (route "B") that runs the same state, questions and criteria through your own agent. It returns jev_called: false with null probabilities — useful for building the integration, and explicitly not real Jev. The skill also warns that a missing key is never a reason to paste a secret into chat.

Platform limitations (Windows)

Changelog

v1.9.16 (2026-10-05)

Full upstream-sync day + a research-driven writing-quality suite upgrade.

• 13 mirror re-clones: agent-skills, awesome-design-md, compound-engineering-plugin, EverOS, Gskills, humanizer, humanlayer-skills, jev-skill, marketingskills, markitdown, ponytail, superpowers, ui-skills. All were shallow-stale (the v1.9.2 anti-bloat SKIP trade-off); all re-cloned from current upstream and re-pushed to the JZKK720 forks (--force-with-lease after a 403-then-non-FF sequence — each fork tip was a strictly older snapshot, no independent work overwritten).
• 27 skill refreshes applied via -Refresh: superpowers/TDD family (executing-plans 65→374 lines), agent-skills methodology family, marketing family, the jev-triage/documents/eval trio, and humanizer → v3.1.0 (35 patterns, from 26).
• 4 new skills installed: stop-slop (hardikpandya, 17.8k⭐, false-agency detection), humanize (aashaexo/soundshuman — as a prose-only port in upstream/humanize/; direct install gate-BLOCKED on AR2 x2 false positives + the sloplint.js CLI), unslop (asavvin-pixel, structure-over-vocabulary, "clean slop" second-pass category, UMD/DeepMind study), deslop (stephenturner, scientific-writing skew). All MIT, all gate-PASS (humanize re-gated at 20/100 LOW SAFE after the port).
• 1 skill parked with reason: slopgent (ehmo/slopkit) — its judge-benchmark corpus of deliberately-wrong force-push replies trips HIGH YR1 x2; fixtures, not malware, but unfixable without forking the benchmark data. Row commented out; mirror kept.
• verify-state.ps1 PATH bootstrap now includes %LOCALAPPDATA%\Programs\witr (witr 0.3.4 was present on disk but invisible to a stale-shell PATH — the 2026-10-05 FAIL was environmental).
• Totals: manifest 261 (260 active + 1 disabled), local/* 49, ~/.agents/skills 288, ~/.claude/skills 521, mirrors 55, upstream/ 52.

v1.9.14 (2026-09-23)

Two skills added, the Jev-vs-LLM question answered, and a real Jev call verified.

Does the Jev API replace our LLM models? No — and the vendor says so

From TypeSafe — Jev with coding agents:

Jev is not a drop-in replacement for the LLM behind Claude Code, Cursor, opencode, Copilot… There is no model: "jev-latest" setting that turns your coding agent into a Jev-powered agent, because the two systems solve different problems.

Jev is a System One model — state + typed questions in, choice/score/noul out. It cannot generate text, so it structurally cannot drive a coding agent. Copilot and Claude stay exactly as they are; Jev is called by the code your agent writes.

No MCP server was added, deliberately. MCP grants an agent tool access; Jev is a typed call made by application code, not a tool an agent invokes. The count stays at 11.

Added — typesafe-ai (official skill)

The vendor's own skill, MIT, 1.9k★. Teaches the Jev API, the three primitives, the patterns, and the habit of reading live docs over stale knowledge. SkillSpector score 3 / LOW / SAFE, 100% coverage, zero Claude-Code coupling.

Added — jev-harness (TianyuCodings, 143★)

Builds task-specific Jev decision pipelines: code prepares observations and enforces actions, Jev supplies judgments, an optional reflection model improves the pipeline offline. This is pipeline optimisation, not model-weight training.

Windows caveat, precisely. Only functional Python nodes are macOS-only — python_nodes.py raises PythonSandboxUnavailable unless sys.platform == 'darwin', with no host-exec fallback by design. Everything else works here: verified that import auto_jev, spec, runtime and PipelineRuntime all succeed on win32, and the shipped configs/flow_seed.json (a real 8-node parallel expression+Jev flow) validates on Windows. auto-jev is not on PyPI — source-only, Python ≥3.11.

Gated: score 37 / MEDIUM / CAUTION — all four findings verified false positive. The gate blocks only on do_not_install (exit 1); this scanned exit 0, consistent with the 7 MEDIUM/CAUTION skills already shipped.

Added — official SDKs

• Python typesafe-sdk 0.7.1 → import typesafe_sdk
• JavaScript @typesafe-ai/sdk 0.6.0

Trap: the PyPI name typesafe-ai is a name-squatting redirect shim, not the SDK. It warns at import: "Install typesafe-sdk and use import typesafe_sdk instead."

Verified — a real call, not just an install

Everything else here was verified without a key. This closes that gap — a genuine call, OpenRouter-billed, routed to TypeSafe:

client = TypeSafeClient(api_key=KEY, base_url="https://openrouter.ai/api")
resp = client.system_one(model="jev-1.13", state=..., questions={...})

is_urgent    noul   = 0.97
department   choice = billing   (confidence 0.84)
frustration  score  = 1.17      (0 Calm · 1 Frustrated · 2 Very angry)
model        = "typesafe/jev-1.13-20260917"     ← resolved ID, not the requested alias
usage        = 409 input / 73 output

The route that works

Pointing the official SDK at OpenRouter is 4 lines — "Change the base URL and the API key. Everything else stays the same." Preferred over the @openrouter/sdk alpha endpoint, which works but is alpha.

Corrections to guidance published earlier in this release

• "Python runtime is macOS-only" was too broad — only functional Python nodes are. See the caveat above.
• Absence from the models list is not a health check. GET /api/v1/models returns 454 models with 0 typesafe/jev entries, yet the model routes fine.
• ask() does not exist — the method is system_one().
• Choice.criteria is a mapping, Score.criteria is a sequence. Easy to invert.

Counts

Manifest 255 → 257; local/* 46 → 48; ~/.agents/skills 282 → 284; ~/.claude/skills 515 → 517; upstream/ 49 → 51.

v1.9.13 (2026-09-23)

New skill: github-repo-metadata, harvested from a workflow that took several wrong turns.

Updating a repo's description looks blocked when gh is absent, GITHUB_TOKEN is unset, and an unauthenticated PATCH returns 401 — the natural conclusion is "no auth, ask the user for a token." That is wrong. A machine that has pushed over HTTPS holds a usable token in the Windows credential store (credential.helper=manager); git credential fill retrieves it with no env var and no CLI. The check order is now explicit: gh → env token → credential store → only then genuinely blocked.

Two traps recorded, both of which cost real time:

• GET /repos/... is cached. A PATCH returning 200 can read back the old value seconds later. This looked like a silent no-op; it was a stale read.
• PowerShell $home is read-only (alias for HOME). Assigning to it throws. Use $homeUrl.

Gated: SkillSpector score 5 / LOW / SAFE, 100% coverage. One MEDIUM flag (confidence 0.5) on the api.github.com call — correct, expected, no secret embedded. Logged to SCAN_LOG.md rather than suppressed.

Security: the token was read in memory only — never printed, logged, or persisted. Only length and token class were reported. A post-run sweep for token-shaped strings in the working tree returned 0 matches.

Counts: 254 → 255 manifest entries; on-disk 281 → 282.

v1.9.12 (2026-09-22)

bin/verify-state.ps1 did not verify state. The designated entry point for "is my machine correctly set up?" was under-checking and could not fail on the thing it exists to check. Four defects, all fixed:

• Checked 12 of 18 CLIs. witr, semantica, loop, watch-skill, wigolo, and ocr were absent from the list, so a broken install of any of them reported a clean pass.
• Never exited non-zero on a CLI miss. Missing CLIs printed MISS and the script still fell through to exit 0 — only the mcp.json branches had an exit path. A CI consumer saw green while the toolchain was incomplete. Now prints RESULT: OK / RESULT: FAILED with a matching exit code.
• Never cross-checked the MCP servers the bundle installs. It validated mcp.json as JSON and listed whatever was present, but never compared against setup/mcp.json.template, so a bundle server missing from config was invisible. Now reads the expected set from the template (one source of truth) rather than hardcoding a second copy.
• Never checked the parked set was actually parked. Parking works by renaming SKILL.md → SKILL.md.disabled; a bare SKILL.md left behind is still discoverable, and nothing caught it.

One trap worth recording: a parked directory existing under ~/.claude/skills/ is not a leak. The Claude mirror parks in place rather than in a skills._disabled/ sibling, so directory presence proves nothing — the marker file is what matters. This was mistaken for a defect during verification and disproved before it reached a commit.

The MCP comparison is deliberately tolerant: gallery-installed servers are keyed owner/name and version-pinned, while the template uses short names with @latest. firecrawl and firecrawl/firecrawl-mcp-server are the same server, so raw key equality would report every gallery install as missing — as it did on first run.

v1.9.11 (2026-09-22)

Documentation and release-metadata fixes. No skill content, no script logic changed.

• The public quick start was broken, and always had been. It instructed ~/dev/setup/setup-global-skills.ps1, but the installer has lived at setup/setup-global-skills.ps1 since the initial commit (023c906) — git log --diff-filter=D -- setup-global-skills.ps1 confirms it was never at repo root. A fresh clone following the docs hit a missing-file error. Fixed in both the Quick start and "To update later".
• SETUP_GUIDE.md had a typo'd clone URL — cubecloud-skillsboundle-setup (misspelled "bundle"), which is not a repository. Corrected on both the git remote add and git clone lines.
• The repository-layout block described the runtime tree, not the repo. It showed a ~/dev/ tree with the installer at top level. Rewritten to the real clone tree, with runtime-only paths (~/dev/bin/, ~/dev/upstream/SCAN_LOG.md, ~/dev/forks/) listed separately and an explicit path convention.
• v1.9.8–v1.9.10 were lightweight tags (no release message), unlike the annotated v1.9.3–v1.9.7. Re-created with one-line messages. Pointed-at commits unchanged — metadata only.
• The "254 vs 281" mismatch was never explained. The README said 254 skills; a completed install prints 281. Both were correct but nothing said why, so it read as a broken counter. Now documented inline: 253 bundle-installed + 27 VS Code Azure/Entra/Foundry extension skills
  • 1 comment-documented port (gstack-review) = 281 on disk.
• Layout counts corrected: 27 bin/ scripts (was 19), 48 upstream/ dirs, and the fifth setup/ file (skillspector-ollama-models.yaml).

v1.9.10 (2026-09-22)

Parked 13 Copilot-incompatible skills so Copilot stops discovering them. Not a quality judgement — this is a VS Code Copilot / Windows stack, and the parked set targets Claude Code.

Three independent blockers on the forensics family, each verified:

1. Coupling — all 12 use allowed-tools: Bash(*), ... mcp__codex__codex, $ARGUMENTS substitution, and a Bash() preamble. Copilot cannot resolve allowed-tools, does not substitute $ARGUMENTS, and has no Codex MCP. This is the same coupling the manifest already refuses for gstack-review.
2. Security gate — evidence-ledger scans 76 / HIGH / DO_NOT_INSTALL; six more are MEDIUM/CAUTION. install-skill.ps1 blocks on this by design.
3. Missing dependency — the family invokes tools/adjudicate_findings.py plus check_ai_style.py and check_presentation.py. None exists on disk (searched all of ~/dev), so the pipeline cannot run even with the coupling stripped.

Also: 11 of 12 are absent from SCAN_LOG.md and none has an upstream or fork mirror — no source of record to track.

Parked, not deleted. Files moved to ~/.agents/skills._disabled/<name>/ with SKILL.md → SKILL.md.disabled, and the same rename on the Claude side. Reversing is a directory move; the semantics match install-skill.ps1 -Disabled exactly (the same mechanism that parks caveman).

Verified before parking: the family has exactly one inbound reference from outside itself — claudex-loop → crucible, a description string naming the legacy /crucible trigger, not a dependency. The 12 cross-reference only each other. Parking breaks nothing.

Counters: installed 294 → 281 active (~/.agents/skills/), 14 parked (skills._disabled/). Manifest unchanged at 254 — parking is a filesystem state, not a row change.

v1.9.9 (2026-09-22)

Tracked 5 bundled capabilities that paired a hand-authored SKILL.md with a CLI and/or MCP server but had no manifest row. Manifest-tracking only: 0 of 294 installed skills modified.

All five were live and servable but untracked — orphaned the same way as v1.9.8's three.

Why local/* and not an upstream row (verified, not assumed):

• No upstream SKILL.md exists for any of them. microsoft/markitdown is the right repo but ships zero SKILL.md files (checked the fork; .github/skills is a 404). The two markitdown skills are hand-authored docs about markitdown, same lineage as the User/prompts trio dated 2026-06-27. An upstream row would resolve to nothing.
• local/* rows are refresh-protected. install-missing-skills.ps1:174-176 skips them with "skip $name (protected local port)", so hand-authored content is never overwritten — the failure mode that once truncated book-to-skill (702→414 lines) and diagram-design (565→273).
• Subdirs survive. install-skill.ps1:128 copies with -Recurse, so references/ and references/steps/ install intact.

upstream/ mirrors created for all five, byte-identical (SHA-256 verified) with full subdir trees. Also refreshed a stale gstack-review mirror — tracked copy was 88 lines (pre-refresh), the installed port is 112 with a section 7 mapping review layers to reference files; re-synced to 13/13.

Counters: manifest 249 → 254 (253 active + 1 disabled); local/* 40 → 45. Installed unchanged at 294 — all five were already on disk.

v1.9.8 (2026-09-22)

Re-added 3 skill rows that commit 3757869 removed on a premise that later became false.

Commit 3757869 (2026-08-11 00:18) deleted local/loop-engineering, local/watch-skill and local/wigolo, recording the reason as "hand-authored placeholder entries that never had a SKILL.md." That was accurate when written. It stopped being accurate 9 hours later, when those three SKILL.md files were authored (2026-08-11 09:21) and installed:

Why this mattered. The three were live and being served to Copilot from ~/.agents/skills/, but untracked. That orphaned them: install-missing-skills.ps1 skips untracked skills entirely, so they could never be refreshed, and a future cleanup pass would have seen three unknown directories and deleted them. The manifest's own comment was actively misleading a reader into thinking removal was current policy.

• Fix: rows re-added as local/<name>|<name>||upstream/<name>, and the now-false comment replaced with one that records the real history.
• upstream/ mirrors created for all three (upstream/<name>/SKILL.md), byte-identical to the installed copies (SHA-256 verified). This is the convention all 40 local/* rows follow — the resolver at install-missing-skills.ps1:120-124 maps local/* to upstream/$Name, so without a mirror the rows would report no source.
• Deliberately left alone: the CLI/MCP half of 3757869's rationale still holds and needed no change. All three CLIs install in Phase 2 (uv/npm), and all three MCP servers are already present in mcp.json.template and in the installed mcp.json — verified, not assumed.
• archify untouched. It was named in the same removal commit but was re-integrated separately at line ~372 via JZKK720/archify + upstream/archify.

Counters: manifest 246 → 249 (248 active + 1 disabled); local/* 37 → 40. Installed count unchanged at 294 — these three were already on disk; only their manifest tracking was restored.

v1.9.7 (2026-09-22)

Two data-only fixes — llm-council could never resolve, and 4 installed Ollama models were unregistered. No script changed; 0 of 294 installed skills touched.

1. llm-council reported (no source) while its source sat right there. The row had an empty relpath, and with an empty relpath both resolvers default to skills/$Name (install-missing-skills.ps1:127, install-skill.ps1:81) — but this repo keeps SKILL.md at the root. Fixed with relpath .. The CSV comment already said "SKILL.md at repo root"; the relpath was simply left blank.

   Chose a data fix over a script fix on purpose: adding a root-SKILL.md fallback would have edited install-skill.ps1, which is deployed in 3 places that must stay hash-identical (setup/, bin/, ~/dev/bin/). One manifest cell beats chasing hash parity across three copies. Confirmed no-op for content before editing — installed and mirror are identical (468 lines / 2462 words, same mtime, neither has argument-hint). Now reports skip llm-council (up to date).

2. 4 installed Ollama models were unregistered, so scans still fell back to the guessed 128000-token budget: nemotron-safe:latest (1048576), qwen-safe:latest (262144), qwen3.8-flash-next:125b-a6b-q4_K_M (262144), glm-ocr:q8_0 (131072). Each context length re-measured with ollama show. Also dropped the meaningless max_output_tokens from the embedding entry. After: 0 unregistered, valid YAML (13 models).

3. 7 rows remain mirrorless (self-learning, loopy, karpathy-guidelines, i-have-adhd, dependency-doctor, project-graveyard, hyperframes). All 7 are already installed, so this is provenance, not functionality. Left unmapped deliberately: cloning a mirror exposes its row to -Refresh (the v1.9.6 lesson), so each needs its own decision.

v1.9.6 (2026-09-22)

Guard expansion — -Refresh could have truncated 16 ce-*/lfg skills by 17–83%. No skill content changed: 0 of 294 installed skills were touched (hash-verified).

1. 16 skills exposed to a destructive refresh. Cloning the compound-engineering-plugin fork mirror — done to give 16 rows that reported skip (no source) a real source — had a side effect: they became refresh candidates. Measured installed-vs-source lines: ce-ideate 434→74, ce-test-browser 242→51, ce-debug 339→117, ce-handoff 142→55, lfg 149→58.

2. The refresh source is stale in both directions. It resolves to the fork mirror, pinned at 84bdf8c (2026-08-27), which is itself behind upstream (upstream ce-debug is 123 lines vs the mirror's 117). A refresh would therefore match neither the installed state nor current upstream. Unlike hallmark / taste-skill / create-technical-design-doc (name-only divergence, explained by an upstream rename), these are genuine content divergences — so the rename rule does not apply. All 16 are now $protected, blocked with a named reason until a human decides whether the larger copies are deliberate enrichment or a stale install. Preview after: Would refresh: 0 Up-to-date/protected: 246 Blocked: 0.

3. Latent parse break removed. sync-fork-upstreams.ps1 maps compound-engineering-plugin already; a duplicate key produces 哈希文本中不允许重复的键, which fails the entire script parse. Removed, with comments recording that the key exists and that hyperframes intentionally has no mapping (JZKK720/hyperframes 404s).

v1.9.5 (2026-09-21)

full-audit.ps1 correctness fixes — it leaked MCP servers, its MCP verdicts were timing-dependent, and its summary hid the real breakdown behind one number.

1. Process leak that wedged the next run. The MCP smoke test started each server as cmd.exe /c "<launcher> > out 2> err". Those commands are launchers: uvx spawns markitdown-mcp, npx spawns node → firecrawl-mcp. Process.Kill() kills only the shell, so the server survived as an orphan still holding the redirect file handles. Measured: 19 stale server processes and 12 locked audit_mcp_*.log files; the next run's Remove-Item threw RemoveFileSystemItemIOError and the audit appeared to hang. A recursive Stop-Tree now reaps descendants on both the timeout and early-exit paths. After: 0 orphans, 0 lock errors.

2. Verdicts that flipped run to run. The old code grepped stderr for "error|Error|traceback|..." on the early-exit branch — and Windows' launcher-failure text ('x' is not recognized as an internal or external command) contains none of those words, so a failed launch was reported as PASS. The verdict now comes from the server's own liveness and its output streams. After: identical verdicts and identical stderr byte counts across consecutive passes.

3. A summary that read as flakiness. WARN/ADVISORY: N was a regex union. The real split is 1 WARN + 73 ADVISORY — nearly all of it advisory skills-ref results across ~294 skills. They are now reported separately: PASS: 52 | FAIL: 0 | WARN: 1 | ADVISORY: 73.

Also: smoke-test logs are per-run (audit_mcp_<name>_<runId>.log) so a leftover process can never break the next run's cleanup on a name collision.

Runtime note — the audit takes ~18 minutes, by design. Measured 1108 s. It is not hung and a timeout is not a failure: ~122 s is the MCP smoke test, and the dominant cost is skills-ref validate spawned once per skill across ~294 skills. Budget 20+ minutes.

v1.9.4 (2026-09-21)

Six typed-decision skills added (jev-skill), one of them a gate-remediated port; plus 51 orphaned diagram-design reference docs returned to version control.

1. jev-skill family (wuyoscar/jev-skill, MIT, 324★) — Jev returns typed decisions (choice / noul / score) over a caller-supplied state, questions and criteria. It does not browse, execute tools or generate prose; the host agent still plans, acts and verifies. Six skills: jev (router), jev-triage, jev-documents, jev-eval, jev-ui, jev-simulation.

   Five install directly from upstream and pass the gate cleanly. The general jev router is HARD-BLOCKED upstream — exit 1, risk 80/100, DO NOT INSTALL. Exact cause: the Executable scripts: Yes flag plus HIGH E2 Env Variable Harvesting at scripts/jev.py:131 and :180 — that script reads OPENROUTER_API_KEY/TYPESAFE_API_KEY from the environment to make authenticated outbound calls. That is its declared function, so the rule keys on real behaviour and is not a false positive.

   Two of its other findings are false positives, and are recorded as such rather than "fixed": P3 at SKILL.md:151 scores the prohibition "Do not silently send private documents to an external API" (substring match, polarity inverted), and P6 at 26% confidence reads a prompt-injection test recipe as prompt extraction. The credential read was deliberately not obfuscated to dodge the scanner.

   Remediation: drop scripts/ only — both HIGH findings and the executable flag come from that one directory. Installed as local/jev → upstream/jev; re-verified exit 0 / MEDIUM / CAUTION through the real installer. Same pattern as the timesfm port in v1.9.3. Trade-off, documented in the port itself: the Run (Jev API mode) CLI commands no longer work from the installed copy — simulation mode (B) (no key, no CLI, no network call) is the default, and real-Jev users fetch the CLI from upstream with explicit consent.

2. 51 orphaned diagram-design reference docs tracked. upstream/diagram-design/ had 1 tracked file against 52 on disk; .gitignore did not cover it, so this was genuine drift. SKILL.md links those docs, so a fresh clone would have silently lost them.

3. Fork mirror #50: JZKK720/jev-skill created, cloned and wired into sync-fork-upstreams.ps1.

Counters: manifest 240 → 246 (245 active + 1 disabled); local/* 36 → 37; fork mirrors 49 → 50; installed 288 → 294 in ~/.agents/skills/ and 506 → 514 in ~/.claude/skills/. full-audit.ps1 → PASS 52 | FAIL 0.

v1.9.3 (2026-09-15)

Two skills added — one upstream install, one gate-remediated port — plus a local-Ollama scan launcher.

1. hyperframes (heygen-com, Apache-2.0) — HTML-native video rendering, installed as the router only. It is the entry point to 20 published skills and pulls each workflow on demand via npx hyperframes skills update <workflow>, so the skill index stays lean and installed skill count can grow at runtime by design, not drift. Gate PASS (exit 0, risk 18/100 LOW, 0 executable components); the 52 advisory MEDIUMs (MP2 ×31 router-enumeration + RP1 ×21 unpinned npx) sit on SKILL.md only. Needs Node ≥ 22 + FFmpeg, both already present.
2. timesfm-forecasting — added as a remediated local/* port rather than an upstream install, because upstream hard-blocks the gate (exit 1, HIGH TM1 ×2) and both findings trace to one construct: a --skip-check flag that bypasses the mandatory RAM/GPU/disk preflight. Removing the flag (and its dead branch) flips the gate to exit 0 and loses nothing but the ability to skip a safety check. Isolation-proved: deleting only that file also yields exit 0.
3. bin/skillspector-local.ps1 — runs the optional LLM-assisted scan against a local Ollama model, so findings can be cleared or confirmed with no cloud key and without shipping skill source to a third party. Deliberately declares no param() block: SkillSpector's short flags (-f/-o/-r) collide with PowerShell's binder three ways, and any new param can re-break them.

Also corrected this release: that registry file shipped wrong context lengths while claiming in its own comment to be measured (5 of 10 entries), and the hyperframes note mis-stated its own gate result (claimed 3 MEDIUMs at lines 106/107/110; actual is 52 MEDIUMs, 3 inside that window). Both are fixed. The local/* row shape is now documented in the manifest too — all 36 rows are 4 fields, so the trailing upstream/<path> is documentation and field 4 is parsed as $disabled.

Counters: manifest 238 → 240 (239 active + 1 disabled); local/* 35 → 36; installed 288 ~/.agents/skills/ + 508 ~/.claude/skills/. full-audit.ps1: PASS 52 | FAIL 0.

v1.9.2 (2026-09-14)

Root-caused and fixed the mirror bloat that v1.9.1 identified but left alone.

1. sync-fork-upstreams.ps1 no longer un-shallows mirrors. setup-global-skills.ps1 creates every mirror with git clone --depth 1 on purpose, but sync fetched with a plain git fetch upstream — no --depth — which un-shallows the clone and re-downloads full history. That one line is how 41 mirrors reached 1,517 MB of .git against 947 MB of working trees. The fetch is now git fetch --depth 1 --no-tags upstream.
2. Measured before/after on a live mirror. With the old plain fetch, a single run grew a freshly-shallow mirror from 1.6 MB → 365 MB. With the fix, a full sync run grew it +0.01 MB; the forks root held at 987 MB, and all 20 syncable mirrors still report is-shallow-repository = true. The only full-history repo is the bundle itself.
3. A shallow fast-forward failure is now a SKIP, not a FAILED. A shallow clone shares no ancestor with its remote, so merge --ff-only says "refusing to merge unrelated histories" even for a clean fast-forward. Deepening to fix that is the exact bloat being removed, so it now reports SKIP (shallow, no common ancestor - re-clone to update) and counts as skipped. Re-running setup-global-skills.ps1 re-clones cleanly. This is a real trade-off — those mirrors report SKIP rather than advancing when genuinely behind — chosen deliberately over silent disk growth.

The largest remaining .git is markitdown at 4.99 MB.

v1.9.1 (2026-09-14)

Corrected a v1.9.0 size claim, and reclaimed 1,482 MB of mirror overhead.

1. The v1.9.0 "~11 MB" figure was wrong. The impeccable mirror was described as a ~11 MB sparse checkout; its real total was 370.5 MB — 1.0 MB worktree plus 369.5 MB of .git. The sparse checkout was real, but an earlier --deepen 100 had pulled a 301 MB history pack, and the docs counted only the worktree. Corrected in README.md, CHANGELOG.md, setup/skills-list.csv, and bin/sync-fork-upstreams.ps1.
2. Root cause of mirror bloat, found while measuring. setup-global-skills.ps1 clones each mirror with --depth 1, but sync-fork-upstreams.ps1 fetches with a plain git fetch upstream, which un-shallows and accumulates full history. Measured directly: a single sync run grew the shallow impeccable mirror from 1.6 MB to 365 MB. That is why 41 mirrors accumulated 1,517 MB of .git and only 947 MB of working trees.
3. 1,482 MB reclaimed by removing .git from the 21 mirrors whose .git was ≥ 5 MB. Nothing reads mirror git history — verified by grepping every installed skill for git log|show|blame|diff against a mirror path (zero hits) — and install-skill.ps1 clones from GitHub, never from ~/dev/forks/. Worktrees were preserved exactly: file counts verified identical across all 21 (0 losses), and both skills that actually read a mirror still work (design-md-library → 74 DESIGN.md files; archify → bin/archify.mjs present).
4. The trim is durable, and the mirror count is unchanged. A mirror without .git is SKIPped by sync as "not a git checkout" — so it can never be re-fattened. sync-fork-upstreams.ps1 now reports Synced: 0, skipped: 49, Failed: 0: 20 mirrors stay syncable (all small ones) and 29 are static (the 21 trimmed plus the 8 one-file port stubs). Forks root: 2,470 MB → 988 MB.
5. Known limitation introduced. JZKK720/impeccable is now static, so it will not fast-forward to upstream. Its content already matches upstream at the revision the 21 ports were cut from, so ports are correct; a future re-port needs a fresh git clone --depth 1 --filter=blob:none --sparse of pbakaus/impeccable directly. Also noted: the fork's main sits at f88b2837 (skill v4.1.1) while upstream is at cb56ed6c1.

v1.9.0 (2026-09-14)

Traced the impeccable upstream, repaired its provenance, and completed the port set (6 → 21).

The 6 impeccable-* ports shipped in v1.8.0 were correct but unreproducible. The manifest cited ~/dev/forks/impeccable/skill/reference/*.md as their source — a path that has never existed on this machine. The fork JZKK720/impeccable was live on GitHub but had never been cloned, so nothing on disk could confirm what the ports were derived from, and sync-fork-upstreams.ps1 silently SKIPped the repo as "no upstream mapping". Same failure shape as the two orphaned local/* sources fixed in v1.8.2: an advertised source that does not exist, behind a gate that reports success while doing nothing.

1. Mirror created and tracked. ~/dev/forks/JZKK720/impeccable is a sparse, blob-filtered, depth-1 checkout of the fork at pbakaus/impeccable upstream/main (cb56ed6c1), mapped in sync-fork-upstreams.ps1 as "impeccable" = "pbakaus/impeccable". Full clone would be ~358 MB; the sparse path keeps only skill/reference, .agents/skills/impeccable/reference, and both SKILL.md files.

   Correction (v1.9.1). v1.9.0 described this mirror as ~11 MB. That was wrong: the checkout was sparse, but its .git still held full history — an earlier --deepen had pulled a 301 MB pack — for a real total of 370.5 MB (1.0 MB worktree + 369.5 MB .git). The v1.9.0 figure counted only the worktree. See item 5.

2. Port set completed 6 → 21. Upstream exposes 35 compiled reference docs; v1.8.0 ported the 6 whose names matched existing skills. A coupling audit found 21 of the 35 are pure prose — zero instructions to run a bundled script and no dependency on upstream-only artifacts. The missing 15 were added: adapt, adapt-native, audit, audit-native, android, ios, colorize, craft-floor, extract, harden, operate, optimize, overdrive, quieter, shape. The other 14 were excluded at the document level, not by keyword: live, live-setup, init, new-work, critique, visualize (live browser iteration), hooks, doctor, routing (script-driven), layout, typeset, polish (script-coupled), craft (deprecated alias), document (28 refs to PRODUCT.md / .impeccable/design.json / "the live panel").

3. Why ports, not a direct install — verified, not assumed. Upstream ships one catch-all skill that mandates node scripts/context.mjs. Its .agents/skills/impeccable payload is 153 files, including a 513 KB live-browser bundle, a 260 KB detector ruleset, a 121 KB question server, a 97 KB hook library, and hook-before-edit.mjs, which intercepts every file edit. That is a code-execution surface, not a prose library; installing it would drag the application into a skills directory. Ports stay.

4. Ports regenerated from the current upstream, not the stale revision. The mirror was four skill-versions behind (f88b2837 = v4.1.1 → cb56ed6c1). Four ported docs had already changed — adapt (+6), audit (+2/-1), audit.native (+1/-1), harden (+9) — all new touch/gesture prose (pointercancel, lostpointercapture, interrupted gestures). All 15 were regenerated from upstream/main so they ship current rather than recreating the drift.

5. A BOM bug the gate caught. The first generation pass used PowerShell's Set-Content -Encoding UTF8, which writes a UTF-8 BOM on PowerShell 5.1. That byte order mark is an invisible character at SKILL.md:1, so all 15 ports failed both SkillSpector (HIGH: Hidden Instructions, 60% confidence) and skills-ref (must start with YAML frontmatter). Writing BOM-free UTF-8 through UTF8Encoding($false) cleared both at once.

6. document dropped after porting. It passed the coupling screen but carried 28 references to upstream-only artifacts and "the live panel". Porting it would have shipped prose that cannot stand alone, so it was removed and the count corrected 16 → 15.

7. Mojibake in the ported bodies. After the BOM fix the gate was green, but a byte-level audit found 70 corrupted characters — U+9225 (鈥) where em-dashes belonged and U+922B (鈫) where arrows (→) belonged, in 15 of the 21 ports. Cause: PowerShell 5.1 decodes a BOM-less file as ANSI, so both Get-Content -Raw on the upstream source and the non-ASCII literals in the generation script itself were mangled. Fixed by reading sources through an explicit UTF-8 decoder and keeping the authored descriptions pure ASCII; arrows and em-dashes now survive as real U+2192 / U+2014. Verified 70 → 0 corrupted characters. Worth noting the gate did not catch this — SkillSpector and skills-ref both passed on the mojibake, so it needed a byte-level check.

Counters: manifest 223 → 238 (237 active + 1 disabled); local/* rows 20 → 35; fork mirrors 48 → 49; installed 271 → 286 in ~/.agents/skills/ and 491 → 506 in ~/.claude/skills/. All 15 new ports pass the SkillSpector gate (0 blocked) and skills-ref validate (15/15 valid); 21/21 impeccable-* sources resolve, and installed copies hash-match their sources 15/15. No empty fork mirrors, no skill missing a SKILL.md.

v1.8.2 (2026-09-14)

Release verification caught two real doc/repo defects. No feature changes.

1. Two local/* port sources were never committed to git. local/evidence-graph-review and local/ops-authorization-methodology were advertised by the manifest, but their upstream/ directories did not exist and had never been committed. Because the installer treats a missing local source as a silent SKIP, any fresh machine quietly skipped both and still reported success. Both installed copies were real content (82 and 87 lines) and were recovered byte-exactly into upstream/. All 20 local/* rows now resolve — verified 20/20.
2. Stale local/* count (21 → 20) in the CHANGELOG and evaluation plan, left over from an edit made after the count was taken. Also corrected a 36 new rows reference to 35.

Method worth keeping: this was found by checking every manifest row's source path against the filesystem, not by reading the docs. A counter-only audit passes while the repo is still broken.

v1.8.1 (2026-09-14)

Optional archify CLI. The archify renderer can now be exposed as a bare command:

powershell -NoProfile -ExecutionPolicy Bypass -File .\setup-global-skills.ps1 -IncludeArchifyCli

It writes one 86-byte ~/.local/bin/archify.cmd shim forwarding to the fork mirror's bin/archify.mjs. archify is a zero-runtime-dependency Node CLI (bin/archify.mjs imports only node: stdlib; the mirror has no node_modules and needs none), so this needs no package manager, no registry, and no global npm install. An npm install is not even possible — upstream package.json sets "private": true.

Verified before shipping: archify doctor → exit 0 (15/15 checks), render → 820,813-byte HTML, validate --quality showcase → exit 0 with receipt. The installer runs doctor first, so it never advertises a command it has not verified.

Optional laya-serve sidecar (opt-in separately): after pip install laya works, run bin/install-laya-serve.ps1 to ship an always-hot decision API on http://127.0.0.1:8770 (copies upstream/laya/laya-serve.py to C:\rocm-sdk\, writes start/stop shims + per-user autostart). Copilot-written tools then POST decisions there at ~30-50 ms without loading torch themselves. See CHANGELOG 1.9.15.

Opt-in because every other phase of the installer is unconditional and this one adds a command to your PATH. Requires the archify fork mirror, so it warns and skips under -SkipForks.

This is CLI-only. The archify skill remains the clean methodology port (upstream/archify/) — upstream the skill is gate-blocked (HIGH MP3), and that gate does not apply to the CLI.

Also fixed: two local/* port sources were never committed to git. The manifest advertised local/evidence-graph-review and local/ops-authorization-methodology, but neither upstream/ directory existed on disk and neither had ever been committed. The installer's guard is a silent SKIP (local source missing or no SKILL.md), not an error — so on any fresh machine those two rows were quietly skipped and the install still reported success.

Both installed copies are real content (82 and 87 lines), so the text was recoverable, not lost. Recovered byte-exactly from ~/.agents/skills/ into upstream/ after confirming neither copy carried an install-time argument-hint: that must not be written back.

Found by checking every manifest row's source path against the filesystem — a count-only audit would not have caught it.

v1.8.0 (2026-09-13)

35 new skills from a 14-repo evaluation, one gate-blocked skill ported, and a manifest row repaired.

223 skills in manifest · 18 CLIs · 11 MCP servers · 48 fork mirrors · 271 active total

Fourteen candidate repositories were cloned, surveyed for shape and licence, and every skill in scope was run through the same skillspector scan --no-llm gate the installer uses — 104 candidate skills gated in total. Full write-up: docs/plans/2026-09-13-skills-bundle-integration-evaluation.md.

Added (35)

• Domain-driven design, as a set (6) — from tech-leads-club/agent-skills (MIT): domain-analysis, domain-identification-grouping, coupling-analysis, decomposition-planning-roadmap, tactical-ddd, modular-design-principles. These six reference each other, so they are installed together rather than as a partial subset.
• Planning and evaluation (6) — same source: create-technical-design-doc (1055 lines — the deepest planning artifact in any candidate), create-rfc, harness-eval (audits an agent harness for broken paths, duplication and dead refs — this bundle's own manifest and MCP config are exactly that surface), spec-driven-eval, not-your-babysitter (evidence-before-completion discipline), learning-opportunities.
• Marketing (8) — from coreyhaines31/marketingskills (MIT): product-marketing, copywriting, content-strategy, competitor-profiling, launch, pricing, marketing-loops, marketing-psychology. General discipline in, platform-specific execution out.
• OpenAI curated skills (5) — from openai/skills (Apache-2.0, per-skill): figma (the workflow contract for the Figma MCP this bundle already configures), screenshot (OS-level desktop capture — the bundle had browser capture but no desktop capture), pdf, transcribe, security-threat-model.
• Agent self-management (4) — from kunchenguid/firstmate (MIT): bearings, diagnostic-reasoning, bootstrap-diagnostics, project-management. Four of 22 — the rest are bound to firstmate's own daemon and shell scripts.
• Visual explanation and instruction files (3) — from humanlayer/skills (MIT): show-me, improve-claude-md, narrow-react-prop-types.
• Prose de-slopping (2) — humanizer from blader/humanizer (291 lines, based on Wikipedia's "Signs of AI writing") and no-ai-slop from petergyang/no-ai-slop (61 lines, gates 0 findings at any severity). Two independent implementations kept as complements — the bundle's write-concisely covers density and brand-voice covers deriving a voice, but neither de-slops existing prose.
• archify (1) — see below.

Ported: archify

tt-a1i/archify (MIT) failed the gate on the exact entry the roadmap wanted:

HIGH: MP3 - Memory Manipulation

This re-confirms the block recorded during the v1.6.3 pruning (then classified HIGH YR4). A clean methodology-only port now supplies the skill, following the same pattern as book-to-skill, diagram-design, and webapp-testing. The port keeps the authoring contract, type router, invariant list, Mermaid conversion rules and the validate → deliver → visual-check sequence, and gates 0/100 SAFE. It drops the renderer, schemas, examples and brand-mark database — which is where the gate fired.

The cost of that decision is real. A port is methodology only. archify's genuine value is its validated renderer (bin/archify.mjs validate, 9 showcase checks), and that is not installed. The ported SKILL.md tells the agent to use the fork-mirror checkout for the CLI and to report truthfully when it is absent, rather than implying an unrun validation passed.

Fixed: two latent installer bugs

1. i-have-adhd would have failed on every fresh machine. Upstream moved its skill from .cursor/skills/i-have-adhd/ to skills/i-have-adhd/, but the manifest still pointed at the old path — so that row would die with SKILL.md not found. Repointed; both copies gate PASS.

2. Two fork names would have collided with unrelated forks. The installer clones JZKK720/<last-path-segment>, and the naive names were already taken:

   Candidate	Naive name	Already occupied by	Forked as
   tech-leads-club/agent-skills	agent-skills	addyosmani/agent-skills — this bundle's active mirror	tech-leads-club-agent-skills
   openai/skills	skills	vercel-labs/skills	openai-skills

   A naive clone would have overwritten the addyosmani mirror that 23 existing rows depend on. Both were forked under explicit non-colliding names.

Excluded, and why

• calesthio/OpenMontage is AGPL-3.0 — 138 skills of video/3D/motion content, and a hard no. AGPL's network-copyleft obligation attaches to distribution, and this bundle redistributes to end-user machines and mirrors publicly.
• Tencent/WeKnora and THU-MAIC/OpenMAIC are products, not skill packages — a Go RAG platform and a Next.js learning app respectively. Their skills/ trees are their own runtime content.
• openai/skills .system (5 skills) — two blocked, three high-risk, and all five describe Codex's own skill/plugin system, which is not this harness.
• 45 gate-blocked entries parked as comments with per-skill reason codes, so install-missing-skills.ps1 stops retrying them. New this round: shopify-developer (CRITICAL YR1 credential-exfiltration webhook), security-ownership-map (HIGH P6, blocked in two independent catalogs), figma-use (HIGH P6), migrate-to-codex, ad-creative, stow (HIGH E4), plus the .system pair.
• HIGH-severity passes skipped by policy — the-fool and the-jury (HIGH AR2 anti-refusal), mermaid-studio (HIGH PE3), codenavi (HIGH MP3), cloudflare-deploy / sentry (HIGH E2), and others. AR2 is a consistent tell across catalogs: skills that instruct adversarial or persistent behaviour tend to phrase it as an override of the model's own judgement.

Verified

• install-missing-skills.ps1 → Installed: 0 / Blocked: 0 / Skipped: 223 on re-run (idempotent)
• full-audit.ps1 → PASS: 51 | FAIL: 0
• sync-fork-upstreams.ps1 → all 9 new mirrors tracked, Failed: 0
• skills-ref valid count 167 → 196; all 271 installed skills have a SKILL.md
• humanizer is byte-identical to upstream (29102 bytes, SHA-256 match) — installed from upstream rather than ported specifically to avoid the re-encoding that corrupted the book-to-skill port
• Every one of the 9 install-skill.ps1 copies still hashes identically (4C94C9CC…)

v1.7.0 (2026-09-13)

Skills can now be refreshed in place. Syncing a fork mirror does not update skills that were already installed — those are point-in-time copies. Until now there was no way to bring them forward without reinstalling by hand.

# preview (writes nothing)
powershell -NoProfile -ExecutionPolicy Bypass -File .\bin\install-missing-skills.ps1 -Refresh -WhatIf

# apply
powershell -NoProfile -ExecutionPolicy Bypass -File .\bin\install-missing-skills.ps1 -Refresh

First run brought 79 skills up to date with their current sources, including all 14 that had fallen behind (performance-optimization +151 lines, repo-scan +100, api-and-interface-design +73, …) and 34 skills whose upstream description: now declares activation triggers — descriptions drive Copilot's skill discovery, so those were user-visible.

Three protections. A refresh is a content update, not a reinstall: it never re-runs the SkillSpector gate (the source was gated at install and every fork mirror is git-tracked, so changes are reversible).

1. local/* rows are never refreshed. These are hand-authored ports living inside this repo under upstream/. Refreshing them is circular — the bundle's own upstream/ copy can be leaner than what was installed.
2. codex-review / codex-build are pinned. Upstream reduced them to ~17-line shims delegating to sibling claudex-loop; a refresh would replace a working standalone workflow with a pointer. The user decides.
3. argument-hint: is preserved. The bundle injects this at install time and upstream does not ship it; a naive overwrite drops it. Refreshes re-attach it after the description block.

Verified: 51 → 51 skills retain argument-hint:, 0 hints lost, both pinned skills byte-identical, and repeated runs report Refreshed: 0 (idempotent).

Incident, recorded honestly. The first refresh shipped with a too-narrow protection rule: it pinned only the two named skills and skipped local/* rows in prose but not in code. That overwrote two ports — book-to-skill (702 → 414 lines) and diagram-design (565 → 273). diagram-design was recovered from a later 578-line v2.5 copy found in another repo. book-to-skill's original 702-line content is unrecoverable — no backup, no git history, no dangling blobs. The rule was then implemented in code and re-verified. If you maintained a local book-to-skill enhancement, it is gone and will need rewriting.

v1.6.3 (2026-09-13)

Upstream follow-through: renames, additions, and two more stale fork mirrors repaired.

187 skills in manifest · 18 CLIs · 11 MCP servers · 39 fork mirrors · 236 active total

• Fixed: crucible was renamed upstream. Commit a027430 rebranded skills/crucible → skills/claudex-loop and added skills/claudex-route, so the old manifest entry 404'd. claudex-loop keeps the legacy /crucible trigger and generalises it to bidirectional Claude/Codex work. Both installed; the retired entry is parked as a comment.
• Fixed: two stale JZKK720 fork mirrors. crucible (12 commits) and ECC (277 commits) were behind upstream on GitHub while the local clones were current, so installs failed with "SKILL.md not found" because the installer clones from JZKK720/.... Both were clean fast-forwards and were pushed.
• Fixed: gstack-review was in the wrong place and missing its references. It lived only in ~/.copilot/skills/ as a bare SKILL.md, without the reference/ tree the manifest documents. Rebuilt in ~/.agents/skills/ with all 12 reference files (checklist, design-checklist, TODOS-format, greptile-triage, and the 8 specialists/ lenses), and SKILL.md gained a section 7 mapping each review layer to its file so the copied docs are actually discoverable. Stale duplicate removed.
• Fixed: two empty fork mirrors. ui-skills and loop-engineering were 0-byte placeholder directories. Both cloned for real (230 and 731 files) and added to sync-fork-upstreams.ps1 so they stay tracked.
• Added: constraint-driven-development (addyosmani/agent-skills) — writes a project's quality bar into CONSTRAINTS.md and catches a weakened bar in the diff; budget-negotiator and install-loop (cobusgreyling/loop-engineering) — completing the loop family at 7/7; counterparty-channel-discipline and operator-approval-loop (ECC) — mention gating, silent observation, hashed draft approvals and a pre-draft baseline gate for agents talking to external counterparties.
• Pruned: 17 gate-blocked entries (ce-* ×9, taste-* ×2, commit-archaeologist, scope-creep-detector, archify, huashu-design, plus the earlier caveman-era set) moved from active manifest rows to comments, so the installer stops retrying them. A full install-missing-skills.ps1 pass now reports Installed: 0 / Blocked: 0 / Skipped: 188.
• Documented: a new "curated out by scope" table covering rails-patterns, the two ECC legal skills, tasteforge-video, and the 137-skill Gskills mirror (118 GCP-specific, mirror-only by decision).
• Counters corrected: 194 → 187 manifest entries; ~/.agents/skills/ 228 → 236 directories.

v1.6.2 (2026-09-13)

Verification pass + pipeline repair. Counters corrected against measured reality, and four real defects in the install/update tooling fixed.

194 skills in manifest · 18 CLIs · 11 MCP servers · 39 fork mirrors · 228 active total

• Fixed: install-skill.ps1 was corrupt. The deployed ~/dev/bin/install-skill.ps1 had stray prose prepended above its block comment, producing 12 parse errors. Every install silently failed and install-missing-skills.ps1 reported Installed: 0 without erroring. Redeployed from setup/install-skill.ps1; all three copies now hash identically.
• Fixed: sync-fork-upstreams.ps1 had 4 bugs. (1) crashed on non-git mapped directories, aborting the whole run; (2) a $LASTEXITCODE leak forced the wrong branch, breaking context-engineering-kit; (3) git progress on stderr was promoted to a terminating error, producing false FAILED reports for successful fast-forwards; (4) five repos were missing from $upstreamMap and were silently skipped forever. 16 mirrors fast-forwarded as a result.
• Fixed: portability. 26 hardcoded C:\Users\<dev>\... paths across 9 scripts replaced with $env:USERPROFILE / $env:APPDATA / $env:TEMP equivalents; same for mcp.json.template.
• Fixed: manifest paths. huashu-design (SKILL.md is at repo root, needs SkillRelPath=".") and archify (needs archify/archify) had empty SkillRelPath, so their installs could not locate SKILL.md.
• Fixed: MCP config. The skillopt server pointed at a non-existent path in both the template and the live mcp.json. loop-engineering shipped in the template but never reached the live config. SkillOpt fork mirror restored; both entries now resolve. Live MCP servers 21 → 22.
• Restored: agent-reach and semantica were advertised as installed CLIs but were absent from PATH. Both reinstalled.
• Documented: the 13 manifest entries blocked by the SkillSpector hard gate, with per-skill reason (see table above).
• Counters corrected: README claimed 153 skills / 173 active total; measured reality is 194 manifest entries with 228 directories in ~/.agents/skills/.

v1.6.1 (2026-08-18)

2 new skills (clean ports): evidence-graph-review (methodology extraction from reverse-skill, MIT), ops-authorization-methodology (auth/scope check for security tasks, MIT) — both clean methodology-only ports, no scripts, no toolchain coupling.

155 skills · 18 CLIs · 11 MCP servers · 39 fork mirrors · 173 active total

• evidence-graph-review: Evidence Graph Review — audit evidence-to-finding-to-path traceability, confidence grading, scope gating, and stop/replan conditions when authorization or evidence quality is insufficient. Methodology-only; no scripts, no tool execution.
• ops-authorization-methodology: Ops Authorization Methodology — confirm scope.md, network_profile mode, in-scope/out-of-scope assets, deliverables, and stop/replan conditions. No scripts, no toolchain coupling.

v1.6.0 (2026-08-13)

153 skills · 18 CLIs · 11 MCP servers · 39 fork mirrors · 172 active total

• +3 skills (clean ports): anti-slop (miqdadbadjuber/anti-slop, MIT), diagram-design (cathrynlavery/diagram-design, MIT), book-to-skill (virgiliojr94/book-to-skill, MIT)
  • book-to-skill and diagram-design upstreams were blocked by SkillSpector (CRITICAL/HIGH — bundled Python scripts + HTML assets); re-authored as clean methodology-only local/* ports using markitdown for extraction
• Fix: Phase 4 local/* handling now reads the 5th CSV column (upstream/<name>) for source paths, matching install-missing-skills.ps1

v1.5.0 (2026-08-11)

144 skills · 18 CLIs · 11 MCP servers · 39 fork mirrors · 172 active total

• +5 skills: archify, book-to-skill, huashu-design, scope-creep-detector, dependency-doctor, project-graveyard (awesome-llm-apps set)
• +2 CLIs: semantica (uv tool, v0.6.0), witr (GitHub release binary, v0.3.3 — no Go needed)
• +3 fork mirrors: alchaincyf/huashu-design, semantica-agi/semantica, pranshuparmar/witr
• Phase 2: automatic witr release-binary download added (no Go toolchain required)
• Bug fixes: removed 4 dead local/* placeholder entries from skills-list.csv; fixed local/* handling in Phase 4
• 25 compound-engineering skills integrated (blocked by SkillSpector risk 81/100 — available as fork mirror)

v1.4.0 (2026-07-15)

139 skills · 16 CLIs · 10 MCP servers · 36 fork mirrors

• +35 ECC skills (agent engineering, codebase intelligence, development governance, security, design/UI, operations/cost, engineering practices)

v1.3.0 and earlier

See git tag --list for older releases.

License