Forge is an end-to-end workflow for complex coding tasks. It turns vague requests into verified changes through discovery, planning, delegated execution, and review. This repo packages the skill and its sub-agents for every harness that can load local skills and agent definitions.
skills/forge/— the skill (SKILL.md) and the discovery protocol (discovery.md). These two files are the contract for how Forge operates.agents/— sub-agent definitions. Each file's body is a prompt, not documentation — every paragraph lands in a model's context, so keep them tight and behavior-shaping..claude-plugin/— Claude Codeplugin.json+marketplace.json..codex-plugin/— Codexplugin.json..cursor-plugin/— Cursorplugin.json..codex/agents/— native Codex sub-agent TOML files (one per agent inagents/)..codex/INSTALL.md— per-harness setup docs.
Editing one file usually means editing several. The common landmines:
- Adding a new sub-agent — create
agents/<name>.mdAND a matching.codex/agents/<name>.toml(verbatim body + per-agentsandbox_mode), then wire it into: SKILL.md (lifecycle if it has a phase, briefing, parallelize principle if it runs concurrently),README.mdsub-agents list,.codex/INSTALL.mdsymlink loop, and.cursor-plugin/INSTALL.mdinstall/uninstall lists. - Version bumps —
.claude-plugin/plugin.json,.claude-plugin/marketplace.json,.codex-plugin/plugin.json, and.cursor-plugin/plugin.jsonmust stay in sync.
- No tests, CI, or dependencies. This is a markdown + JSON distribution.
CLAUDE.mdis a symlink toAGENTS.md. Don't turn it into a separate file.