ai-plugins

A multi-harness marketplace of AI coding-assistant plugins for Claude Code, Codex, and other harnesses that adopt plugin or marketplace concepts.

Personal development system

This marketplace has one audience and one installable plugin: development-system. It supports Codex and Claude Code with one initialization command and one project configuration file.

The default preset is direct-to-trunk delivery with Tiber and on-demand linked worktrees for concurrent mutable work. Optional agentic-system and eval-reporting capabilities are selected in .development-system.toml; the plugin owns its bundled MCP surface.

The strong recommendation is to install only development-system. Additional plugin marketplaces expand the supply-chain trust surface. The SessionStart hook warns about conflicting plugins, incompatible harness settings, and user-managed MCPs that need compatibility review.

Plugin catalog

Plugin	Harnesses	Description	Version
development-system	Codex, Claude Code	One configurable development workflow with on-demand skill routing.	3.2.1

Using the marketplace (Claude Code)

Add this repository as a marketplace, then install a plugin from it:

# From inside Claude Code:
/plugin marketplace add jwilger/ai-plugins      # GitHub owner/repo shorthand
# ...or a local checkout:
/plugin marketplace add ./ai-plugins

/plugin install development-system@ai-plugins

The marketplace is referenced by its name (ai-plugins) in install commands, regardless of the URL you added it from. List and manage with /plugin list, /plugin marketplace update ai-plugins, and /plugin marketplace remove ai-plugins.

Using the marketplace (Codex)

Codex-facing marketplace metadata lives in .agents/plugins/marketplace.json, and each plugin has a .codex-plugin/plugin.json manifest. In a local checkout, install or sync the plugin from the matching directory under plugins/ using the Codex plugin flow available in your Codex environment.

Install development-system from the local marketplace, then start a new thread and run its setup skill from the target repository's primary checkout.

Developing in this repo

A Nix flake provides a reproducible devshell with Node, npm, jq, prettier, ripgrep, fd, just, and bats.

nix develop        # enter the devshell
# or, with direnv:
echo "use flake" > .envrc && direnv allow

Any globally installed npm tooling (npm install -g …) is redirected into a git-ignored ./.dependencies/ directory by the devshell, so it never pollutes your home directory. Delete that directory any time for a clean slate.

This repo also has a committed package.json/package-lock.json for the local Promptfoo eval runner. node_modules/ is ignored and restored with npm ci; the eval scripts run that automatically when the Promptfoo, Codex SDK, or Claude Agent SDK packages are missing.

See AGENTS.md for how to author, validate, and publish a plugin.

Eval reports

The repo-owned eval dashboard is generated under site/evals/ by node scripts/evals/build-site.mjs. It is a local/static artifact for review and workflow uploads; the durable record is repo-owned and does not depend on promptfoo-hosted sharing.

Local runs reuse existing Claude Code/Anthropic and Codex/ChatGPT subscription sessions. They do not require provider API keys or fresh approval for the repository-owned evals authorized in AGENTS.md. Unattended trusted automation may instead use protected provider credentials when interactive harness sessions are unavailable; untrusted pull-request checks remain secret-free and validate only the eval configuration and dry-run wiring.

The dashboard includes latest-run status, provider/case/sample pass rates, threshold status, exact installed provider compositions, and separate case-target plugin/skill summaries so regressions can be traced back to both the loaded marketplace surface and the behavior each scenario exercises.

The canonical promptfoo behavior evals run through Promptfoo's native coding agent providers: anthropic:claude-agent-sdk for Claude Code and openai:codex-sdk for Codex. The runner generates the promptfoo config from the current marketplace manifests and labels no-plugin, targeted-plugin, and full-marketplace behavior modes. Codex uses a separate generated home for each mode. For both harnesses, targeted mode installs the deterministic, deduplicated union of plugins declared by the selected behavior cases; EVAL_CASE_FILTER therefore narrows both the cases and their installed plugin set. Full-marketplace mode still installs the complete harness-specific catalog, while no-plugin mode installs none. The generated config records each provider's exact installed composition separately from the plugins targeted by an individual case. An unfiltered targeted run equals the full catalog in both harnesses today because the marketplace has one public plugin and the selected cases target it. The two modes remain distinct controls for filtered runs and future catalog changes. Promptfoo is pinned at 0.121.18; the Promptfoo, Codex SDK, and Claude Agent SDK packages are pinned in package.json and package-lock.json. The runner disables prompt response caching and hosted sharing so a behavior run is a fresh local record.

Default eval harness posture:

• Claude Code: anthropic:claude-agent-sdk, Sonnet 5 via the sonnet alias, local Claude Code authentication via apiKeyRequired: false, and all local plugins with skills: all. The intended human-facing Claude Code posture remains Sonnet high effort with Opus 4.8 advisor where that harness exposes those controls; Promptfoo's current Claude Agent SDK provider does not expose those knobs in this repo's generated config.
• Codex execution: openai:codex-sdk, gpt-5.6-terra with model_reasoning_effort=medium, read-only sandbox, no approvals, streaming, deep tracing disabled, and isolated generated homes containing no plugins, the selected cases' deterministic plugin union, or the complete harness-specific catalog according to the behavior mode. Model-graded assertions independently default to gpt-5.6-sol with high reasoning through the same SDK, so OpenAI model access goes through local Codex auth rather than OPENAI_API_KEY. Override the two roles separately with CODEX_EVAL_MODEL / CODEX_EVAL_REASONING_EFFORT and CODEX_GRADER_MODEL / CODEX_GRADER_REASONING_EFFORT.

The focused GPT-5.6 model-family benchmark compares Sol, Terra, and Luna without running the full marketplace eval suite. Its trace-enforced Codex app-server wrapper and skills-only/no-plugin homes are benchmark controls; the canonical behavior runner above continues to use the native Codex SDK provider and the configured behavior-mode matrix.

The canary suite is separate from behavior evals. Canaries may explicitly ask the harness to prove plugin and skill loading. Behavior prompts stay natural and do not tell the model to use this repository's plugins.

Repeated samples are a deliberate measurement choice, not a blanket rule. The default one-sample matrix treats every case as a binary pass/fail observation and estimates population quality across distinct cases. Set EVAL_SAMPLES above one only when measuring per-input reliability, pass@k capability, pass^k reliability, judge variance, or a small stochastic difference; fractional per-case thresholds apply only to those repeated runs. PR dry-runs do not run live samples.

Pull-request CI validates the eval configuration with --dry-run but does not claim behavior evidence. Provider-backed behavior evidence comes from local, scheduled, manual, or main runs where Claude Code and Codex authentication are available.

To produce the same artifacts locally:

just evals  # runs provider-backed evals, shares the result, and prints the URL
nix develop -c scripts/evals/run.sh
nix develop -c scripts/evals/run.sh --suite canary
nix develop -c node scripts/evals/build-site.mjs

Eval runs have no implicit whole-run deadline: a large matrix must not lose nearly-complete work to an arbitrary wall-clock cutoff. Set EVAL_TIMEOUT only when the caller deliberately wants a bounded run. Timed-out or interrupted runs write evals/out/status.json so the dashboard can show why no fresh result completed.

just evals uploads the latest eval result through promptfoo share. For a local-only report, run scripts/evals/run.sh and then nix develop -c node_modules/.bin/promptfoo view. If a behavior eval exits with Promptfoo's normal failure status after writing artifacts, just evals still attempts to share the report and then returns the original eval status. If the eval run is interrupted, terminated, or times out, just evals stops without sharing. Interrupted, terminated, and timed-out runs all retain any partial artifacts under evals/out/timeout-artifacts/ for debugging.

The repository retains an optional Promptfoo MCP definition under plugins/development-system/components/agentic-systems-engineering/. That path is an internal component identity, not a public marketplace plugin or skill name; users install only development-system. A project that explicitly wires the optional server must provide promptfoo@0.121.18 on PATH. When the project uses flake.nix, prefer pkgs.promptfoo when nixpkgs provides the required version so updates flow through the flake lockfile; otherwise use the project's local package-manager sandbox. The server supports agent-assisted config validation, focused eval runs, result inspection, and fixture development. It supplements the canonical runner; it does not replace the repo-owned artifact path above. Promptfoo's separate mcp provider is for testing MCP servers as systems under test and should be added only when a plugin or project exposes an MCP server to evaluate.

If Codex or Claude Code reports No such file or directory for a bundled MCP client, upgrade or reinstall development-system so the active root manifest and launchers are refreshed. For an explicitly configured Promptfoo server, also verify that the pinned runtime above is available on PATH. The bundled launchers bootstrap through /bin/sh before resolving their commands.

Reporting eval cases

When a plugin, skill, prompt, or workflow behaves incorrectly or only partially works, file an Eval case issue in this repository. Eval cases are the intake path for future regression fixtures in evals/fixtures/.

Include the sanitized input, actual behavior, expected behavior, expected eval outcome (pass, fail, partial, adversarial, or unsure), and the assertion or rubric that would catch the behavior. Do not include secrets, credentials, auth headers, cookies, session ids, private keys, private client data, private repository names, internal hostnames, or raw proprietary source excerpts.

Repository layout

.
├── .agents/
│   └── plugins/
│       └── marketplace.json  # Codex-facing marketplace manifest
├── .claude-plugin/
│   └── marketplace.json      # Claude Code marketplace manifest
├── .github/
│   ├── ISSUE_TEMPLATE/       # eval-case intake form
│   └── workflows/            # CI and eval workflows
├── docs/
│   └── superpowers/plans/    # implementation plans for larger changes
├── evals/
│   ├── fixtures/             # behavior eval scenarios
│   └── promptfoo/            # promptfoo loaders and assertions
├── plugins/                  # one subdirectory per plugin
├── scripts/
│   ├── evals/                # eval config generator, runner, and dashboard builder
│   └── tests/                # Bats tests
├── site/
│   └── evals/                # generated dashboard target, ignored except .gitkeep
├── flake.nix                 # Nix devshell
├── AGENTS.md                 # guidance for AI agents working in this repo
└── README.md                 # this file

License

See individual plugins for their licenses.