Skill Resolver

Deep dispatcher for agent action. This file answers: which skill, tool, MCP, CLI, service, or protocol should handle the task?

Read wiki/meta/agent-operations-hub.md first each session. The hub tells you when this resolver is needed. Use wiki/RESOLVER.md for page filing. Use Capability Routing Map when the question spans overlapping capability families such as browser vs Playwright, wiki search vs live search, or CLI vs MCP.

This file is intentionally more detailed than config/cursor/rules/tool-hierarchy.mdc. The compact rule is the pocket map. This file is the full registry and decision procedure.

Operating Contract

1. Skills are procedures. If a skill matches, read its SKILL.md before acting.
2. Tools are capabilities. If a tool/service is needed, use the highest-ranked interface that can do the job.
3. Protocols are obligations. If the task triggers capture ingest, no-one-off work, startup, quality gates, process hygiene, or config sync, follow that protocol even if the user did not spell it out.
4. Kevin's wiki is first. For people, tools, concepts, decisions, and preferences, search the wiki before web or API calls.
5. Routing changes must compound. If you discover a better route, update this file, tool-hierarchy.mdc when compact routing changes, Active Stack (What to Actually Use) when it should be recommended, and trigger evals when applicable.

How To Use This Resolver

If two skills match, read both and let them chain. Example: content publish work often routes social-draft -> humanizer -> kevin-voice; frontend work often routes taste -> browser verification.

Default Workflow Spine

The current active stack is the working default. Use it unless the task clearly says otherwise. Source: Active Stack (What to Actually Use), 2026-06-17

Skill, Tool, and Page Layers

Executable skill prompts live under skills/<category>/<slug>/SKILL.md and carry provenance with origin: / source_slug: frontmatter. The wiki no longer mirrors every skill as wiki/skills/<slug>.md; use Skill Registry for the generated inventory and the *-skills family pages for durable synthesis. Tool pages still own product facts, setup, tradeoffs, and alternatives. This resolver decides which executable skill or tool layer fires for a concrete task. Source: Capability Routing Map, 2026-06-17; User request, 2026-06-27; User request, 2026-06-28

Repository Context Switch

Skill sources live under ~/Documents/GitHub/kevin-wiki/skills/{engineering,productivity,personal,misc,in-progress,deprecated}/. Promoted skills are projected into skills/.runtime/all, which is symlinked into agent runtimes. Claude, Dedalus, Cursor plugin, Codex plugin, and skills.sh imports are folded into the shared categories with origin: metadata. Config, hooks, MCP, and tool hierarchy live under config/.

---

Dedalus monorepo (wiki mirror)

config/dedalus/ is a historical reference copy of the Dedalus agent harness and product Cursor rules. Kevin's affiliation ended 2026-07-06. Do not assume access, refresh from upstream, or route current work there unless Kevin explicitly supplies a legacy task and authority. See Agent Config Sync § Dedalus Repo Config. Source: User request, 2026-07-14

Layout

Path-scoped rules (agents/rules/)

Dedalus Cursor rules (cursor-rules/)

Routing (explicit historical Dedalus task only)

1. Load config/dedalus/agents/rules/ethos.md + the path rule for files you touch.
2. Pick the matching standalone Dedalus-origin category skill when the task matches the slug (below) - before reaching for unrelated generic skills. If the slug is listed as folded, use the generalized skill that carries the Dedalus reference.
3. For Dedalus frontend in Cursor, apply config/dedalus/cursor-rules/ alongside personal config/cursor/rules/.

Skill slugs → Dedalus-origin or folded category SKILL.md

Quick map

Standalone origin: dedalus slugs (20)

ci-cd-best-practices · comment · commit · db · db-write · dcs-test · exemplar-audit · fail-closed-case-matching · go-style · perf · pr · pr-review · promote · read-and-review · shadcn · stage-refactor-checklist · stripe · style-guide · test-writing · update-docs

Folded Dedalus-reference slugs (21)

bugs · column-aligned-fields · counterfactual · fix-types · generate-interface · greentext · hotfix-preview · invariant-first-coding · issue · lerp · loading-screens · make-interfaces-feel-better · minimal-first-implementation · postmortem · refine · rtfm · rust-neckbeard · simple · state-machine-dfa · tests · torvalds

---

Service Registry (per-service interface routing)

For each service, prefer the first interface that can do the job. Drop to lower ranks only when the higher-ranked interface cannot handle the specific task.

Broad Capability Discovery

Use Capability Harvest Pattern when Kevin asks to harvest, refresh, or expand the universe of available skills, MCPs, service CLIs, source artifacts, templates, design techniques, browser/testing workflows, eval/review/security patterns, runtime substrates, or automations.

Do not mass-install harvested capabilities. Promote a capability only after audit, then update the owning surface: a skill for repeatable procedure, a tool/design/style/workflow/architecture page for durable knowledge, an automation for cadence, a trigger eval for natural-language routing, or a source-compile rule for recurring artifacts. If it becomes a default, also update this resolver, tool-hierarchy.mdc, Active Stack (What to Actually Use), and trigger evals. The 2026-06-24 promotion installed audited official/reputable skills for Azure, Firebase, Google Workspace, Vercel AI SDK/CLI, Better Auth, Stitch, and Excalidraw; registry slugs that no longer existed upstream were skipped or mapped to current upstream equivalents. The 2026-06-25 expansion added non-installable patterns and source feeds as first-class harvest surfaces. Source: User request, 2026-06-24; User request, 2026-06-25

Horizontal Service CLI Registry

Use skills/engineering/service-cli-registry/ when a task touches a SaaS or infrastructure product and no more specific installed skill owns it. It covers the current CLI seed catalog in Service CLI Coverage: GitLab, AWS, Google Cloud, Docker, Kubernetes, Helm, Terraform, Pulumi, Ansible, Sentry CLI, Datadog CI, New Relic, Grafana, Snowflake, Databricks, BigQuery, Auth0, Okta, Twilio, SendGrid, Contentful, Sanity, Strapi, Algolia, Hugging Face, Netlify, Fly.io, Railway, Heroku, Neon, PlanetScale, Turso, Upstash, MongoDB, MySQL, PostgreSQL, CircleCI, and Buildkite.

The skill is a routing/safety gate. It does not approve live mutations. Before any external command, identify account/context, run read-only discovery first, use dry-run/plan/preview where available, and ask before destructive, expensive, production, or customer-data actions. Source: User request, 2026-06-22

Vercel

Next.js 16.3 retired knowledge-only Next skills in favor of local bundled docs and managed AGENTS.md; keep skills for workflow execution and verification. Source: Next.js 16.3 AI Improvements, 2026-06-26

Stripe

Supabase

Clerk

No CLI available.

Better Auth

Firebase

TanStack

TanStack is a frontend library ecosystem, not a SaaS control plane. Do not route it through service CLI/MCP patterns; verify with typecheck, tests, and rendered app behavior. See TanStack Ecosystem. Source: User request, 2026-06-27; Ecosystem Capability Coverage

Vite / Astro / Zustand

Route these as frontend/build-system libraries, not SaaS services. If the task touches deployment, route the deployment provider separately.

Vue / Rust

Figma

PostHog

Sentry

Datadog

Linear

Slack

Shopify

ClickHouse

GitHub

AWS

If aws sts get-caller-identity returns a token error, ask user to run aws sso login.

Google Cloud

Run gcloud auth list and gcloud config get-value project before touching resources.

Azure

Specialized Azure references installed from the 2026-06-24 promotion now live under skills/engineering/azure/references/skills/.

Infrastructure CLIs

Require context checks before mutation: docker context ls, kubectl config current-context, kubectl config view --minify, terraform workspace show, terraform plan, pulumi stack, pulumi preview, hcloud server list, Kafka cluster/topic read-only inspection, and Ansible check mode where possible.

Cloudflare

Canonical ecosystem page: Cloudflare Ecosystem.

Google Workspace

Browser

Route by task type, with Browser Testing Skills as the default browser automation surface for agent loops. See wiki/concepts/complementary-browser-testing.md.

Use skills/engineering/browser-use/ as an opt-in playbook when a task specifically asks for Browser Use or depends on its library. It is not the default browser route because the installed skill was flagged high Gen risk by skills.sh. Source: skills.sh/browser-use/browser-use, 2026-06-22

Feedback loop: agent-browser discovers browser behavior and visual bugs → codify important behavior as Playwright assertions → CI gates future PRs → agent loop moves on.

Browser playbook persistence

Agent IDE context cost

React hygiene (post-agent UI passes)

Multi-agent orchestration & fleet (ALL agents — Cursor, Codex, Claude Code, Hermes, OpenClaw)

Applies to every agent, not just Cursor. Pick the topology in Multi-Agent Topologies (orchestrator-worker / fan-out / pipeline / debate / evaluator-optimizer) — go multi-agent only when the task is wide, valuable, and parallelizable, and wire observability first. Fleet substrate: Agent Machines (run as many agents as you want, on any model) + Agent Control Plane Architecture (fleet state, not chat state). Loops to run live: Browser Testing Skills · Loopy · Agent Self-Improvement Eval Library / agent-eval-library · Borrowed Intelligence (Drain It into Plans) · The Eval Loop (Slop Is an Output Problem) · The Brain-Agent Loop. One harness across all agents: Cross-Agent Harness Portability.

Agent-native service / API / app / site access

Enforced by config/cursor/rules/agent-native-cli.mdc. Concept: wiki/concepts/agent-native-cli.md. Both tools below are by Matt Van Horn (author of /last30days).

Headless / second-Mac agent auth & session sync

Data movement / ingestion (ETL / ELT)

Agent sandbox / microVM compute

Frontier model picks (recent)

Knowledge-base format (agent context)

Codex workflow → reusable skill

Code host for agents

Document → markdown / structured extraction

Agent memory (knowledge graph)

Agent frameworks / platforms (recent)

Creative / content tooling (recent)

---

Routing rules (always-on vs on-demand)

Browser Automation and Frontend Verification

Route by task type with Browser Testing Skills as the default agent browser surface (canonical table: § Service Registry → Browser, above; concept: wiki/concepts/complementary-browser-testing.md).

Feedback loop: agent-browser finds visual bugs → codify important behavior as Playwright assertions → CI gates future PRs.

Generative UI

Prefer: json-render > AI SDK structured output > raw JSX (never)

AI SDK agent apps, workflows, and harness embedding

Prefer: verify current AI SDK docs/source -> pick the smallest matching primitive -> prove with deterministic tests when possible.

Full pages: Vercel AI SDK, AI SDK HarnessAgent, ai-test-kit.

Rich-text / document editor

Prefer: Plate (React + shadcn/ui) > Tiptap (framework-agnostic, mature collab) > Lexical (perf-critical) > BlockNote (batteries-included Notion clone). Full notes: Plate (PlateJS).

Document, Spreadsheet, Deck, PDF, and Template Artifacts

Prefer the artifact-specific Codex skill when the requested deliverable is a file, not a web app. These are full multi-file skill packages: keep their bundled scripts, references, assets, templates, and agents/openai.yaml metadata with the skill source.

Sites and Hosted Web Artifacts

Use these Codex site skills only when the active harness exposes the Sites plugin or the task explicitly asks for a site deployment/version. For normal app implementation in a repo, keep using the repo's framework and frontend skills first.

Cursor Built-In Onboarding, Automation, and Review Subagents

These are harness-specific imported Cursor skills. They are routed here so they remain visible in the generic brain, but use them only when the active harness supports the referenced Cursor feature.

Code Review

Prefer: no-sus-code-doctor (strict self-audit / 100 score) > autoreview (pre-merge closeout / second-model review, installed) > gstack-review (general production-bug review) > counterfactual > cross-modal-review

Design Review

Prefer: design-engineering-polish (animation / polish) > frontend-design (UI creation / direction) > web-design-engineer (visual web artifact build) > gstack-design-review (comprehensive audit) > frontend-design-taste

Frontend build and design taste

Prefer: taste (router with folded taste-* references), or web-design-engineer (full build workflow + 25 named style-recipes + anti-cliché system) · web-design-guidelines (review) · vercel/nextjs folded React/Next performance references. See also Design Review + Animation and 3D.

QA and Testing

Prefer: gstack-qa > agent-browser > dogfood > tests > Playwright (CI only). For beta readiness / first-run onboarding UX: beta-dx-walk.

Planning and plan catalog (borrowed intelligence)

Drain the strongest model now into self-contained plans; execute later with a cheaper/owned model. Principle: wiki/concepts/borrowed-intelligence.md. Skill (installed): skills/engineering/improve/SKILL.md — /improve audits + writes plans; inventory row in Skill Registry and engineering-family context in Agent Engineering Skills. Plan bank: repo-root plans/ (plans/README.md).

Research

Prefer: last30days > agent-reach > WebSearch

Skill discovery is the default for any missing capability. Before hand-rolling a workflow (design, testing, SEO, deploy, docs, review, marketing…), search the open agent-skills ecosystem first: npx skills find <query> → check the leaderboard / install count (prefer ≥1K, reputable source: vercel-labs, anthropics, addyosmani, coreyhaines31, microsoft) → npx skills add <owner/repo@skill> -g -y (or vendor into skills/<category>/<slug>/) → run npm run skill-registry → route it here + in Active Stack (What to Actually Use) + a trigger-evals.json case. Keep installed skills current with npx skills check / npx skills update. Browse skills.sh; full protocol in Security and Review Skills.

Career and Content

Prefer: content-strategy for planning; writing-fragments -> writing-beats / writing-shape for long-form; social-draft -> humanizer -> kevin-voice for publishable social copy; copywriting for marketing pages.

Security

Prefer: security-doctor-chain for multi-tool pre-ship hardening; deepsec for a single repo-wide scan; then bugs > security-best-practices > security-threat-model.

User-facing content moderation (consideration)

Not ranked above custom eval rubrics — a free harm-classifier option for apps that show user input or model output to humans.

Sound effects / audio generation

Prefer: elevenlabs-sfx (generate) → web-kits/audio (play in browser) → curated libraries

Animation and 3D

Prefer: GSAP (scroll) > Motion (component) > React Spring (physics) > Lottie (AE)

Operational

Identity and Setup

Disambiguation Rules

When multiple skills could match:

1. Confirm wiki/meta/agent-operations-hub.md § workspace switch so you search the right skill tree.
2. Check tool-hierarchy.mdc first when choosing tools or service interfaces.
3. Prefer the most specific skill that owns the workflow. Example: gstack-design-review for audits, not generic frontend-design.
4. Prefer product-specific Dedalus skills in the Dedalus workspace when a matching slug exists.
5. Chain skills when the output of one is a required input to the next. Example: capture -> resolve -> update pages -> build-index -> qmd -> log.
6. Ask Kevin only when the missing decision is high-risk, destructive, strategic, or not discoverable from the repo/wiki.

Self-Update Protocol

This resolver is a living document. Update it when:

1. A new skill is installed or materially changed.
2. A tool ranking changes.
3. A workflow becomes active enough to recommend by default.
4. A route is missing or ambiguous in real use.
5. A Dedalus skill/rule changes upstream.
6. A compact always-on rule disagrees with this deep resolver.

After updating:

1. Update config/cursor/rules/tool-hierarchy.mdc when compact routing changes.
2. Update Active Stack (What to Actually Use) when the item should be recommended in live workflows.
3. Update Capability Routing Map when a capability family hierarchy changes.
4. Add or adjust config/trigger-evals.json when a trigger phrase should route differently.
5. Run npm run routing-doctor (includes check-resolvable, check-tool-discovery, trigger evals, and rule typing).
6. Run npx tsx scripts/build-index.ts and qmd update && qmd embed.
7. Append wiki/log.md.

---

Timeline

• 2026-07-14 | Routed Impeccable, Plotly, Evil Charts, and Is This Tech Dead? by task shape; refreshed ACP/Hermes memory as operator-stack references. Source: User request, 2026-07-14
• 2026-07-07 | Added Sonner to component-source routing as the React toast/notification route beside chart sources like Evil Charts. Source: User request, 2026-07-07; Component Library Sources
• 2026-07-02 | Added graphify-sidecar as the route for safe Graphify repo graphs, focused subgraph queries, path/explain/affected questions, PR risk analysis, and Graphify reflection while keeping qmd as the wiki retrieval layer. Source: User request, 2026-07-02
• 2026-06-29 | Added StyleX as the frontend/design-system reference route for agent-authored UI that needs typed token/design-decision constraints, compile-time CSS enforcement, or styled-components migration strategy. Source: User request, 2026-06-29
• 2026-06-29 | Routed agent-eval-library for self-improvement benchmarks and evaluator hardening, and loopy for bounded loop-skill crafting. Source: User request, 2026-06-29; Agent Self-Improvement Eval Library; Loopy; Red Queen Gödel Machine
• 2026-06-26 | Expanded component-source routing to Radix/Base primitives, shadcn, Blocks.so, Chanh Dai, Evil Charts, Skiper UI, Cult UI, Magic UI, React Bits, Fancy Components, Origin UI, and Shaders.com. Source: User request, 2026-06-26; Component Library Sources
• 2026-06-25 | Added Google Workspace CLI as the upstream Google Workspace service interface reference and promoted Component Library Sources plus Skiper UI / Cult UI into frontend/design routing through animated-component-libraries. Source: User request, 2026-06-25; googleworkspace/cli; Skiper UI; Cult UI
• 2026-06-25 | Added check-tool-discovery: opted-in tool/design pages with routing_summary must expose a ## Routing Summary and appear in central or category-specific discovery surfaces, without requiring a central resolver table. Source: User request, 2026-06-25
• 2026-06-18 | Added the operating contract, resolver usage procedure, default workflow spine, skill/tool mirror rule, repository context switch, and stricter self-update protocol tied to active-stack, capability routing, and trigger evals. Source: User request, 2026-06-18; Active Stack (What to Actually Use); Capability Routing Map
• 2026-06-22 | Routed the recently-added tools (Activation Gate): new Service Registry clusters — agent sandbox / microVM compute (Firecracker, Cloud Hypervisor, Rust VMM, E2B Code Sandboxes), frontier model picks (Claude Fable 5, GLM-5.2, OpenRouter Fusion, ZCode), Open Knowledge Format (OKF), Codex Record & Replay, Cursor Origin, doc-extraction (MarkItDown - Universal File-to-Markdown Converter, LangExtract), Neo4j Agent Memory, agent frameworks (Vercel Eve, Open Agents.dev, Matt Pocock's Skills (Skills for Real Engineers), Hermes Agent (Nous Research), OpenEnv), and creative tooling (Higgsfield MCP, GEO-SEO-Claude); plus Planning (Loop Library) and Frontend design rows (Kigen Palette Generator, Arcotype Font Browser, Moodboard Sites, Design Engineering Resources). Source: User request, 2026-06-22
• 2026-06-30 | Promoted LiteParse to first route in document extraction and imported its upstream Agent Skill after reviewing the launch artifact, package versions, GitHub source, and run-llama/llamaparse-agent-skills. Source: X bookmark artifact audit, 2026-06-30