Agent Forge

Agent Forge keeps reusable agent skills, OpenCode plugins, Claude Code plugins, MCP servers, and their shared code in one Git repository.

Keep reusable instructions in one place. Retain each runtime's native plugin format. Produce installable bundles from those sources.

Install Hardlook

Hardlook is a self-contained, evidence-first code review skill for OpenCode, Cursor, Pi, Claude Code, and Codex. It supports routine review, independent deep review, and structural audits. OpenCode support targets the latest stable V1 and V2 releases only.

npx skills add samjingwen/agent-forge --skill hardlook

Choose your harness. Start a new session. Ask for "Hardlook review this change", "Hardlook deep", or "Hardlook structure". For OpenCode's slash-menu command, follow the Hardlook setup instructions. Existing Hardlook installations continue to work. Use this repository for future installs and updates.

Hardlook was migrated from samjingwen/hardlook. The migration preserved its Git history, tests, and license notices. The repository and private package use the name agent-forge. Former samjingwen/agent-toolkit GitHub URLs redirect to this repository.

Install Plainspoken

Plainspoken rewrites dense or ambiguous English using ASD-STE100-inspired rules. It is intended for tool descriptions, error messages, prompts, status reports, and inter-agent instructions. It preserves meaning, uncertainty, conditions, and requirement strength rather than merely shortening text.

npx skills add samjingwen/agent-forge --skill plainspoken

Ask your agent to "Use Plainspoken to disambiguate this text" or "Apply Simplified Technical English and show the before/after table". The skill returns rewritten text alone by default. Its optional Python linter checks structural patterns, not semantic fidelity or certified STE compliance. Plainspoken adapts danyuchn/asd-ste100-skill. It preserves the upstream MIT license and attribution.

Repository layout

Path	Purpose
skills/<name>/SKILL.md	Canonical skill instructions and supporting files
plugins/opencode/<package>/	Native OpenCode plugin packages
plugins/claude-code/<bundle>/	Claude wrapper sources for generated bundles
mcp-servers/<package>/	MCP server implementations that can run independently
packages/<package>/	Shared libraries used by plugins and servers
config/<runtime>/	Runtime configuration examples and templates
catalog/claude-bundles.json	This repository's Claude bundle declarations
catalog/external-tools.json	Inventory of third-party tools and upstream sources
evals/	Behavior cases and evaluation fixtures
tests/<name>/	Component behavior and installation tests
docs/<name>/	Component usage and maintainer documentation
tooling/	Repository checks, packaging, and configuration helpers
distribution/claude-code/<bundle>/	Generated, self-contained Claude plugins
.claude-plugin/marketplace.json	Generated Claude marketplace catalog

Plugin, MCP-server, and shared-library paths above are conventions for implemented components, not required empty scaffolding. Create those directories when real components need them; no .gitkeep placeholders are required. See architecture for structural validation and missing-directory behavior.

Use one .git at the repository root. Package directories are ordinary folders, not nested Git repositories. pnpm-workspace.yaml supports JavaScript and TypeScript packages. Individual modules can retain Gradle or uv projects when appropriate.

Requirements

Installed Hardlook helpers require Node.js 22 or later. They have no npm runtime dependencies. Private repository development requires Node.js 22.13 or later for the supported ESLint toolchain. Hardlook's CLI also requires Git.

Plainspoken's optional linter and its tests require Python 3.11 or later. Neither requires third-party Python packages. Using Plainspoken's rewrite instructions in an agent does not require Python.

Development

Run these commands from the repository root:

npm ci --ignore-scripts
npm run quality
npm run check
npm test
npm run config:opencode
npm run build:claude

Development dependencies use exact version pins and come from package-lock.json. The installed skill does not depend on them. You can use pnpm run for the same scripts. For new Worktrunk worktrees, see dependency and download-cache reuse. The blocking setup hook reuses compatible ignored dependencies without sharing writable directories or copying private/runtime state.

For Hardlook alone, run npm run test:hardlook. For metadata-only harness availability checks, run npm run doctor:hardlook. Run npm run test:quality for executable quality-tool and hook-consumer tests. Run npm run test:plainspoken for Plainspoken's helper and bundle tests. Use npm run lint:plainspoken -- --json <file> to check a selected text file without changing it.

npm run quality checks without changing files. npm run format is the separate formatting fix. See code quality for language coverage, ownership exclusions, CI, and direct project verification.

check performs basic layout, JSON, catalog, and path checks. It also checks that each discovered skill folder contains SKILL.md. It does not validate the complete skill specification, YAML, runtime APIs, or plugin behavior.

test exercises:

• Hardlook's behavior and adapters.
• Plainspoken's structural linter.
• Both skills' installation and relocation boundaries.
• Real passing and failing consumers of the quality tools.

These tests make no live model calls. They do not prove model-review accuracy or rewrite fidelity.

config:opencode prints JSON with absolute paths to the canonical skills directory. The JSON also contains absolute paths to the OpenCode package directories it discovers one level under plugins/opencode/. This configuration output uses native V2 syntax; use the bundled Hardlook adapter instructions for V1 setup. Review the output. Merge the relevant settings into your OpenCode configuration. The command does not install plugins or change global settings.

build:claude creates bundles only when the catalog contains bundle entries. No Claude bundles are currently configured. Hardlook and Plainspoken are installed directly as skills.

Read Hardlook development, architecture, and official references before changing the relevant component. For Plainspoken, read its usage and development notes.

Optional Claude bundle example

This example packages the existing Hardlook skill as a hypothetical hardlook-review Claude plugin. The wrapper and catalog entry below are not currently configured. Use the direct skill installation above unless you want to add a bundle.

1. Reuse the canonical skill and its supporting files in skills/hardlook/, including SKILL.md.

2. Create plugins/claude-code/hardlook-review/.claude-plugin/plugin.json:

   {
     "name": "hardlook-review",
     "version": "0.1.0",
     "description": "Hardlook code review skill for Claude Code.",
     "author": { "name": "Sam Jing Wen" }
   }

   Add any Claude-specific agents, hooks, or MCP configuration under plugins/claude-code/hardlook-review/. Do not add a skills/ directory to the wrapper. The builder supplies it.

3. Add the bundle to catalog/claude-bundles.json:

{
  "schemaVersion": 1,
  "marketplace": {
    "name": "agent-forge",
    "owner": { "name": "Sam Jing Wen" }
  },
  "bundles": [
    {
      "name": "hardlook-review",
      "source": "plugins/claude-code/hardlook-review",
      "skills": ["hardlook"]
    }
  ]
}

4. Run npm run check. Then run npm run build:claude. Validate the resulting plugin in Claude Code. Try the plugin in Claude Code:

claude plugin validate ./distribution/claude-code/hardlook-review
claude --plugin-dir ./distribution/claude-code/hardlook-review

The builder copies the wrapper and selected skills into distribution/claude-code/hardlook-review/. It generates the root marketplace entry with source: "./distribution/claude-code/hardlook-review".

Before publishing the repository as a marketplace, commit generated distributions and the generated marketplace. Consumers need those files in the repository they fetch. Edit canonical sources. Rebuild the generated files. Do not edit generated files by hand.

License

MIT. Hardlook's adapted review techniques retain their bundled attribution and license notices. Plainspoken preserves its upstream MIT license and attribution notices.