npm package discovery and stats viewer.

Discover Tips

  • General search

    [free text search, go nuts!]

  • Package details

    pkg:[package-name]

  • User packages

    @[username]

Sponsor

Optimize Toolset

I’ve always been into building performant and accessible sites, but lately I’ve been taking it extremely seriously. So much so that I’ve been building a tool to help me optimize and monitor the sites that I build to make sure that I’m making an attempt to offer the best experience to those who visit them. If you’re into performant, accessible and SEO friendly sites, you might like it too! You can check it out at Optimize Toolset.

About

Hi, 👋, I’m Ryan Hefner  and I built this site for me, and you! The goal of this site was to provide an easy way for me to check the stats on my npm packages, both for prioritizing issues and updates, and to give me a little kick in the pants to keep up on stuff.

As I was building it, I realized that I was actually using the tool to build the tool, and figured I might as well put this out there and hopefully others will find it to be a fast and useful way to search and browse npm packages as I have.

If you’re interested in other things I’m working on, follow me on Twitter or check out the open source projects I’ve been publishing on GitHub.

I am also working on a Twitter bot for this site to tweet the most popular, newest, random packages from npm. Please follow that account now and it will start sending out packages soon–ish.

Open Software & Tools

This site wouldn’t be possible without the immense generosity and tireless efforts from the people who make contributions to the world and share their work via open source initiatives. Thank you 🙏

© 2026 – Pkg Stats / Ryan Hefner

ajiaskill

v0.2.4

Published

Personal skill manager and platform for coding agents: catalog, MCP, native-first routing, automatic learning, and gated promotion.

Readme

ajiaskill

A local skill platform for coding agents. State the feature or task; the host detects the right core skill and ajiaskill serves it. One catalog, one personal profile, one MCP server, and six hosts share the same setup.

Status: Phase 4: automatic learning and host integrations. Catalog, search/use, MCP, safety, promotion, rollback, native inventory, remote source add, cross-agent foundry, profiles, automatic learning, and host integrations exist. Automatic skill detection, native-first routing, body and optional dense retrieval, an optional 0-2 gate, review-to-learn authoring, and one personal profile across six hosts exist. MCP is the default trigger; host plugins and hooks are opt-in through the setup commands.

What it does

  • Automatic skill detection: the user states the task or feature, not a skill name. The host reads it; native descriptions, the MCP trigger policy, and ajia_search / ajia_use select the skill. ajia route --gate adds an optional 0-2 model gate; a hook can make the choice deterministic.
  • Unified personal settings: one ~/.ajia home and one profile control sources, hosts, routing thresholds, and safety policy across all six hosts. Core skills are the shared baseline; switching profiles does not reinstall them.
  • Catalog: indexes SKILL.md files from local and git sources.
  • Search: ranks by name, description, path, and use count, then reranks the top candidates with a full-body BM25 pass and RRF. Duplicate names keep the core copy.
  • Use: copies the exact bytes to a read-only store and returns the full SKILL.md to the agent. Use it in place; copying it into a host skill directory is installation.
  • MCP (default): the ajiaskill server exposes ajia_search and ajia_use as tools. The agent calls them when no native or core skill fits.
  • Router (optional): ajia route maps a message to one skill with rules and metadata. ajia-router is the fallback skill. Host hooks are reference-only.
  • Remote search: queries skills.sh.
  • Remote source path: ajia remote add turns a skills.sh result into a pinned git source, syncs it, and finds the skill in the catalog.
  • Remote install: wraps npx skills add when you explicitly want the upstream installer.
  • Scan: NVIDIA SkillEvaluator and SkillSpector plus a built-in tripwire.
  • Promote: manual, gated, backed up, and recorded in a lock file.
  • Sandbox: prints a no-network, read-only Docker command for source scripts.
  • Foundry: writes cross-agent candidates under ~/.ajia/foundry, and indexes Hermes /learn skills as candidates.
  • Profiles: one personal profile controls sources, hosts, routing thresholds, and safety policy across every host.
  • Generated metadata: Hermes created_by and curator state, plus foundry candidates, appear in the catalog.
  • Automatic learning: ajia review and the host integrations review a finished turn and queue an evidence-backed lesson. ajia learn is the only author: it scores the lesson against a rubric, deduplicates against the catalog, writes a learn-owned foundry candidate, scans it, and either stages approval or installs it. Durable memory is injected on later turns.

Automatic skill detection

You describe the feature or task. You do not name a skill, and you do not wire a command per feature. The host reads the message and detects the skill while you use its features normally.

  1. Native first: install.sh puts core skills in host roots. The host loads their descriptions and may use one directly.
  2. Policy: the MCP server sends an intent map in its initialize response.
  3. Search fallback: when no native skill clearly fits, the host calls ajia_search, then ajia_use.
  4. Optional gate: ajia route --gate asks a model to select 0-2 candidates from the retrieved set.
  5. Deterministic path: ajia route or a host hook can inject one skill on every matching turn.

OpenCode walkthrough

MCP is the tool surface, not the chooser: it never reads the user message, and the host and model decide. That split is the unique part. In OpenCode the path is:

  1. Native check: OpenCode loads the descriptions of installed core skills from its host roots (mapped in hosts.json). A native match wins; ajia_use refuses to return a duplicate fallback.
  2. ajia_search: with no native match, the model calls ajia_search with the task and host: opencode. The server filters by profile source, ranks the catalog (lexical, body BM25, optional dense), and returns metadata only, including id, description, license, risk, and native_hosts.
  3. Choose: the model picks a result id. The server never selects a skill.
  4. ajia_use: the server copies the exact bytes to a content-addressed read-only store, records a 30-day trial, and returns the full SKILL.md plus base_dir.
  5. Follow in place: the model follows the returned body. Nothing is installed or promoted; ajia promote remains a separate, explicitly approved step.

If no candidate is safe, the correct result is none. See the OpenCode skill discovery workflow for the full standalone diagram.

Unified personal settings

One ~/.ajia home is shared by every host:

  • profiles.json controls sources, hosts, routing thresholds, and safety policy.
  • hosts.json maps the skill roots for OpenCode, Pi, OMP, Codex, Hermes, and Claude.
  • routing.json and the core skills provide one selection baseline.
  • install.sh links the same core skills into the host roots.
  • ajia profile use work switches the personal policy without reinstalling anything.

The result is one personal setup, not six independent skill collections. The core skills stay the same; the profile changes what is visible, how routing behaves, and which safety rules apply.

Quick start

python3 bin/ajia init
python3 bin/ajia source sync
python3 bin/ajia search "review ui accessibility"
python3 bin/ajia use <id>
python3 bin/ajia route "review this PR" --json --host opencode

source sync indexes the local ajia-core source and the foundry. Hermes (~/.hermes/skills) is configured but off by default; enable it only if you use Hermes. The 24 remote sources are also off by default.

npm

Published on npm as ajiaskill. It ships the CLI, the MCP server, the core skills, the OpenCode plugin, and the Pi/OMP extensions. Python 3.10+ is required; the Node wrapper calls python3 unless AJIA_PYTHON is set.

npm install -g ajiaskill
ajiaskill init
ajia search "review ui accessibility"
ajiaskill mcp --host codex

Without a global install:

npx -y ajiaskill init
npx -y ajiaskill mcp --host codex
npx -y ajiaskill-mcp --host codex

ajiaskill and ajia are the same CLI. ajiaskill-mcp starts the MCP server directly. The local setup command prints absolute Python commands for this checkout; npm users can substitute npx -y ajiaskill mcp --host <host>.

Enable a remote source when you need it:

python3 bin/ajia source enable pstack
python3 bin/ajia source sync pstack

# Hermes is optional; enable it only when you use Hermes.
python3 bin/ajia source enable hermes
python3 bin/ajia source sync hermes

Router (optional)

The default automatic path is MCP. ajia route remains for manual routing, the ajia-router fallback skill, and a future deterministic mode.

ajia route "review this PR" --json --host opencode
ajia route "make this page prettier" --json --host opencode
ajia route "what is the weather" --explain

The router uses routing.json plus optional sidecar metadata:

| Field | Use | |---|---| | capabilities | Canonical verbs such as review.code or design.ui. | | triggers | Phrases, keywords, and regex. | | examples | Positive and negative sentences. | | negative_triggers | Veto or downrank, with an unless context. | | file_globs | Context evidence. | | host_fit | Soft host preference. | | requires | Hard tool and environment gate. | | priority | Tie-break only. |

The score is:

0.55 * lexical + 0.25 * rule + 0.10 * context + 0.10 * priority

The router injects at 0.60 or more with a 0.08 margin, or at a 0.90 rule hit. It suggests between 0.45 and 0.60. It does nothing below 0.45. It logs decisions to ~/.ajia/routing.jsonl.

Search and routing also read the top candidates' SKILL.md bodies and add a body-aware BM25 signal. Search fuses lexical and body ranks with RRF. The router adds a bounded body boost to the lexical component.

Optional embeddings

Set an OpenAI-compatible embeddings endpoint to add a dense retrieval arm:

export AJIA_EMBED_URL="http://127.0.0.1:11434/v1/embeddings"
export AJIA_EMBED_MODEL="nomic-embed-text"
# export AJIA_EMBED_API_KEY="..."
# export AJIA_EMBED_TIMEOUT="8"

Skill texts are cached by content hash in ~/.ajia/catalog/embeddings.json. Search fuses lexical, body, and dense ranks with RRF; the router adds a dense signal when it is available. If the endpoint is unset, unreachable, or returns invalid data, both fall back to lexical + body. The endpoint receives skill names, descriptions, and a body excerpt, so use a local model if that text must not leave the machine. http is allowed only on loopback; use https elsewhere.

Optional 0-2 gate

Set a chat-completions endpoint to let a model select 0-2 candidates after retrieval:

export AJIA_GATE_URL="http://127.0.0.1:11434/v1/chat/completions"
export AJIA_GATE_MODEL="llama3.1:8b"
# export AJIA_GATE_API_KEY="..."
# export AJIA_GATE_TIMEOUT="8"

ajia route "draw the architecture of this repo" --gate --json

The gate sees the top candidates with a body excerpt and returns {"skills":[...]}. It can return none. Any error, timeout, or invalid id keeps the retrieval result. bin/ajia-hook adds --gate automatically when AJIA_GATE_URL is set.

Composition and full menu

ajia route "scan this skill and promote it" --multi 3 --json
ajia search "all" --all --source ajia-core --json

--multi splits a task into clauses, routes each, and returns a de-duplicated skill plan. It is a heuristic, not an LLM decomposer. --all returns the full de-duplicated menu, ranked by a local trust proxy: use count, known license, and core source.

The router returns native when the host already has the skill, inject when it can use a source or core body, suggest when ambiguous, and none when no match is safe. Search applies a greedy description-similarity filter so near-duplicate results do not fill the list. The primary automatic path is still MCP. ajia-router is the manual fallback.

Remote backends

ajia remote search "test driven development" --limit 5
ajia remote add obra/superpowers --skill test-driven-development
ajia use remote-obra-superpowers/skills/test-driven-development
ajia scan remote-obra-superpowers/skills/test-driven-development
ajia promote remote-obra-superpowers/skills/test-driven-development --dry-run

remote search uses the skills.sh API. remote add normalizes the source, queries the GitHub repository license, records it as the source license fallback, clones under ~/.ajia/sources, rebuilds the catalog, and returns the catalog id. A per-skill license wins over the source license.

remote install prints the npx skills add command and only runs it with --yes. The upstream installer bypasses the ajia scan and promotion gate.

Install

./install.sh --dry-run
./install.sh

The installer links each skill into ~/.agents/skills and ~/.claude/skills, and links the CLI into ~/.local/bin/ajia. Use --no-cli to skip the CLI link, or --project . to copy skills into a project.

Uninstall:

./install.sh --uninstall

The uninstaller removes only symlinks that point into this repository.

MCP (default)

MCP is the default trigger on all six hosts. Register it once:

python3 bin/ajia setup opencode
python3 bin/ajia setup pi
python3 bin/ajia setup hermes
python3 bin/ajia setup codex
python3 bin/ajia setup omp
python3 bin/ajia setup claude

Each command prints the exact registration command and the trigger policy. The MCP server is registered as ajiaskill; its tools are read-only. They are live on OpenCode, Pi, Hermes, Codex, OMP, and Claude. Each registration passes --host so search results can mark native skills and ajia_use can refuse a duplicate fallback.

The server sends a trigger policy on initialize: diagrams and visual explainers -> archify; one-page HTML answers -> answer-me-with-html; browser tasks -> browser-skill; app verification -> create-verification-skill; frontend design -> impeccable. The policy is guidance, not injection; the model still decides. Install core skills natively (see install.sh) to make them auto-discoverable in the host.

Start a new host session to load ajia_search and ajia_use if they are not visible yet.

Updating

Core skills ship inside the package, so updating them means updating ajiaskill itself.

ajia update --check     # report only
ajia update             # check, install, and re-index
npm install -g ajiaskill@latest   # equivalent when installed from npm

ajia mcp runs the same check on startup. Within a 24-hour cache window (~/.ajia/update-check.json) it queries the npm registry; if a newer version exists it runs the matching upgrade (npm install -g ajiaskill@latest for an npm install, git pull --ff-only for a checkout), re-indexes the catalog, and adds one line to the MCP initialize instructions so the agent can tell you. The MCP server cannot reload its own Python, so restart the host to pick up new code; skill bodies and the catalog are already current.

| Control | Effect | |---|---| | AJIA_AUTO_UPDATE=0 | Disable the startup check and upgrade. | | AJIA_UPDATE_DRY_RUN=1 | Check and report, never install. | | AJIA_UPDATE_REGISTRY=<url> | Override the registry endpoint (loopback HTTP allowed). |

Automatic upgrades mean new skill prose is trusted without the scan/promote gate. If that trade-off is wrong for you, set AJIA_AUTO_UPDATE=0 and use ajia scan / ajia promote for reviewed updates.

Host integrations (optional)

MCP is the default. The installer links skills and the CLI; it does not enable host plugins or hooks. Enable them explicitly:

ajia setup opencode-plugin --install   # OpenCode V2 plugin
ajia setup claude-hooks --install      # Claude Code hooks
ajia setup codex-hooks --install       # Codex hooks; trust via /hooks
ajia setup pi-plugin --install         # Pi extension
ajia setup omp-plugin --install        # OMP extension

bin/ajia-hook supports pre-turn and post-turn review for Codex, Claude, Pi, OMP, and Hermes. The integrations call ajia route before a turn, ajia review to queue a lesson after it, and ajia learn to author the skill. They fail open on timeout or error, and never install a skill without the engine's rubric, dedup, scan, and approval gates. Re-run the same setup command after an upgrade: it replaces its own hook or link in place instead of adding a duplicate. See adapters/README.md for per-host caveats.

Safety stack

| Layer | Tool | Gate | |---|---|---| | Tier 1 | NVIDIA SkillEvaluator | PII, unicode, script lint, license, then SkillSpector. | | Static scan | NVIDIA SkillSpector | 71 patterns; JSON and SARIF. Primary gate. | | Host gate | Hermes built-in scanner | Quarantine, trust tier, content hash. | | Second opinion | Snyk mcp-scan | Prompt injection and tool poisoning. Optional. | | Runtime | Docker, OpenShell, CUA | No host files, no secrets, no network by default. | | Pin | git SHA, tree hash, lock | Exact bytes; re-scan every change. |

Scanners are triage, not certificates. Pin the exact bytes and re-scan.

Scan and promote

ajia scan <id> --json
ajia promote <id> --dry-run
ajia promote <id> --yes
ajia uninstall <id>

Promotion rules:

  • block verdict: refuse.
  • warn verdict: require --force after review.
  • Unknown license: require --allow-unknown-license.
  • Existing non-owned target: refuse.
  • Promoted copies are named ajia-<skill> and carry .ajia-promoted.json.
  • Replaced copies are backed up under ~/.ajia/backups/.
  • The lock file is ~/.ajia/ajia.lock.json.

Commands

| Command | Purpose | |---|---| | ajia init | Create ~/.ajia and the default sources.yaml. | | ajia source list | Show sources and their on/off state. | | ajia source add <id> <path-or-url> | Add a local or git source. | | ajia source enable/disable <id> | Toggle a source. | | ajia source sync [id...] | Fetch sources and rebuild the catalog. | | ajia catalog build | Rebuild the catalog from enabled sources. | | ajia catalog stats | Show skill counts and catalog digest. | | ajia search "<query>" --json | Search the local catalog. | | ajia search "<query>" --all | Return the full de-duplicated menu, ranked by trust. | | ajia use <id> --json | Use a skill without installing it. | | ajia route "<task>" --json | Route a message to one skill, with native-first decisions. | | ajia route "<task>" --gate | Apply the optional 0-2 LLM gate after retrieval. | | ajia route "<task>" --multi 3 | Compose up to three skills from task clauses. | | ajia generated [--json] | List generated and foundry candidates. | | ajia profile list/show/use | Inspect or switch local profiles. | | ajia foundry init/sync/list | Manage the cross-agent foundry. | | ajia native [--host] | List native and ajia-owned skills per host. | | ajia rollback <id> | Restore a promoted skill from its backup. | | ajia remote search "<query>" | Search skills.sh. | | ajia remote add <source> --skill <skill> | Add a skills.sh source, sync it, and find the skill. | | ajia remote install ... | Print or run the npx skills installer. | | ajia mcp | Run the stdio MCP server. | | ajia setup <host> | Print the MCP registration command with the host id. | | ajia scan <id> --json | Run SkillEvaluator security, PII, and lint, plus SkillSpector and the tripwire. | | ajia promote <id> --dry-run | Show the promotion plan. | | ajia promote <id> --yes | Apply the promotion. | | ajia uninstall <id> | Remove an owned promoted copy. | | ajia sandbox <id> --run '<cmd>' | Print or run a Docker sandbox command. | | ajia learn '<source>' | Print the Hermes /learn path and the foundry path. | | ajia learn --new "<idea>" | Start a guided user-authored skill draft; prints the brief when non-interactive. | | ajia learn --new ... --name <slug> --description <text> --body <file> | Write a user-owned foundry candidate, scan it, and print the promote plan. | | ajia learn --edit <slug> --body <file> | Show a diff for a user-owned skill; add --yes to back up, apply, and re-scan. | | ajia learn --queue | List proposed review lessons. | | ajia learn --from-lesson <id> --plan <json> | Author one lesson into a skill; --approve overrides the patch cap. | | ajia learn --audit | Report learned skill quality, use, and patch churn. | | ajia review --transcript <file> [--plan <json>] | Review a transcript and queue a lesson; never writes or installs a skill. | | ajia learning status/pause/resume | Inspect or pause automatic learning. | | ajia learning quarantine list/show/release | Inspect or release quarantined learned skills. | | ajia learning budget | Show review, lesson, and authoring budgets. | | ajia learning suggest | Suggest repeated tasks from local routing/search logs. | | ajia learning ledger/rollback/adopt/pin | Inspect, restore, or protect learned skills. | | ajia learning approve/reject <slug> | Install or reject a pending learned skill. | | ajia learning archive/restore/prune | Curate learned skills without deleting them. | | ajia memory show/add/context | Read, append, or inject durable user or project memory. | | ajia policy | Print the 2-line trigger policy. | | ajia status | Show sources, skills, and trials. | | ajia doctor | Check the local setup. | | ajia update [--check] [--json] | Check the npm registry, apply an upgrade, and re-index. |

Foundry and profiles

The cross-agent foundry is the reliable path:

mkdir -p ~/.ajia/foundry/my-skill
$EDITOR ~/.ajia/foundry/my-skill/SKILL.md
ajia foundry sync
ajia generated
ajia use foundry/my-skill

Any agent can write a foundry candidate. The source is enabled by default and indexed with generated_by: foundry. Nothing is installed until promotion.

Hermes /learn is interactive. hermes -z and hermes chat -q prune skill_manage, so they cannot author skills. Use ajia learn --execute as a best-effort PTY driver, or run /learn in hermes chat and then ajia source sync hermes. Hermes skills with created_by learn or agent carry generated_by and curator_state in the catalog.

Profiles live in profiles.json (or ~/.ajia/profiles.json):

ajia profile list
ajia profile show
ajia profile use work

A profile selects sources, hosts, router thresholds, and safety policy. The default personal profile allows every enabled source. The work profile is a smaller example.

Learning runtime (M4)

ajia review is the detector. It takes a transcript, or a --plan produced by a host model, validates an evidence-backed lesson, and appends it to ~/.ajia/learning/lessons/. It cannot write a SKILL.md or install anything.

ajia learn is the only author. It takes a proposed lesson, scores the skill against a five-part rubric, deduplicates against the whole catalog, scans it, and writes a learn-owned foundry candidate. Installation is gated by auto_install and require_approval.

A user can also ask for a skill directly. ajia learn --new guides the five authoring questions and writes a user-owned candidate; ajia learn --edit shows a diff, backs up, applies, and re-scans. User-owned skills are never auto-patched by review. The authoring rules follow the Anthropic Agent Skills and OpenAI skill-creator guidance: third-person description with trigger phrases, under 500 lines, progressive disclosure, exact steps for fragile work, verification, and gotchas.

ajia learn --queue
ajia learn --from-lesson L-20261103T... --plan skill.json
ajia learn --from-lesson L-20261103T...
ajia learn --audit
ajia learning status
ajia learning quarantine list
ajia learning rollback <slug>

The default policy is strict: auto_install: false, require_approval: true, min_confidence: 0.7, 10 reviews and 3 authored skills per day, a 120-line body limit, an 8/10 rubric, and 3 patches per skill per week. Set auto_install: true only after a clean audit. ajia learning pause stops review and authoring without removing installed skills.

Review and authoring share one ledger and one sidecar. New skills record origin: learn, generated_by: learn, review_id, lesson_id, rubric, and quality_score. Existing review-created skills stay origin: review for provenance and remain patchable by ajia learn.

M0 ships the engine, usage ledger, scan, quarantine, backup, ledger, rollback, and immutable core/user guards. M1 adds the OpenCode V2 plugin: install it with ajia setup opencode-plugin --install and restart OpenCode. M2 adds Claude Code and Codex hooks: ajia setup claude-hooks --install or ajia setup codex-hooks --install. M3 adds Pi/OMP extensions (ajia setup pi-plugin --install, ajia setup omp-plugin --install), repeated-task suggestions, near-duplicate refusal, and archive/restore/prune curator commands. M4 adds the review-to-learn split, evidence-backed lessons, the quality rubric, whole-catalog dedup, patch caps, transcript redaction, daily budgets, and an offline injection/safety eval. The hooks review on Stop/agent_end, use the host model for both the lesson and the authoring call, and inject ~/.ajia/memory. Codex asks you to trust the new hooks in /hooks.

Native-first

hosts.json holds the user and project skill roots for all six hosts. ajia native scans them, marks ajia-owned symlinks and promoted copies, and reports what is truly native. ajia search --host and ajia route --host use that inventory. A native skill wins, and the source fallback is refused.

Core skills

Workflow and output style:

  • grill-me — turn a vague idea into shared understanding.
  • caveman — terse caveman voice; every technical fact, no ceremony.
  • ponytail — lazy senior dev mode; the smallest complete change.
  • karpathy-guidelines — surgical changes, explicit assumptions, verifiable success.
  • i-have-adhd — lead with the next action and keep progress visible.

Platform:

  • ajia-find — fallback catalog search/use.
  • ajia-router — fallback router for manual use.
  • ajia-scan — the safety stack and report reading, including PII and lint.
  • ajia-promote — the gated promotion flow.
  • ajia-sandbox — no-network script execution.
  • ajia-learn — Hermes /learn foundry.
  • create-verification-skill — generate a project-local skill that drives the real app and captures evidence.

Output and interfaces:

  • archify — validated architecture, flow, sequence, ERD, and timeline diagrams as standalone HTML; export images or print to PDF.
  • answer-me-with-html — one-page visual HTML explainer from a short Markdown draft.
  • impeccable — frontend design, redesign, audit, polish, and print/PDF export.
  • browser-skill — drive the user's logged-in Chromium browser through bsk.

The bundled third-party skills keep their upstream notices in THIRD_PARTY_NOTICES.md. Some need an external tool: archify and answer-me-with-html run bundled Node scripts, impeccable fetches its engine on first run, and browser-skill needs bsk and its browser extension.

Layout

bin/ajia                 # CLI, router, MCP server, learning engine
bin/ajia-hook            # pre-turn and post-turn hook for five hosts
npm/                     # Node entry points for the npm package
package.json             # npm manifest, bin mapping, and publish file list
hosts.json               # native and project skill roots per host
profiles.json            # sources, hosts, routing, safety
routing.json             # routing metadata for core and common skills
adapters/                # OpenCode plugin and Pi/OMP extensions
skills/                  # core skills
docs/plans/              # autolearn plan (Markdown + HTML)
scripts/validate-skills.py
scripts/eval-routing.py
scripts/eval-learning.py
scripts/test_mcp_hosts.py
tests/test_ajia.py
tests/test_opencode_plugin.mjs
install.sh

Validate

python3 -m unittest discover -s tests -v
python3 scripts/validate-skills.py
python3 scripts/eval-routing.py --verbose
python3 scripts/eval-routing.py --cases evals/routing-heldout.jsonl --body --verbose
python3 scripts/eval-learning.py
scripts/test_mcp_hosts.py
./install.sh --check

License

MIT for the original content. Adapted skills keep their upstream notices in THIRD_PARTY_NOTICES.md.