apex-forge

中文文档

An execution framework for AI coding agents. 43 skills, three complexity tiers, one install. Remixed from 8 open-source projects.

---

AI coding agents are unreliable. They guess instead of verifying. They skip tests. They claim "done" without proof. They can't coordinate on multi-step tasks. And you end up installing 5 different tools to get a complete workflow.

Apex Forge fixes this by enforcing discipline at the protocol level, not by hoping the agent behaves.

---

git clone https://github.com/d-wwei/apex-forge.git ~/.claude/skills/apex-forge
cd ~/.claude/skills/apex-forge
bun install && bun run build:all
bunx playwright install chromium
./dist/apex-forge init

Or just tell the AI: "initialize apex forge"

You now have 3 binaries (apex-forge, apex-forge-browse, apex-forge-mcp), 43 skills, and a protocol that auto-injects on every session.

After init, your project gets an .apex/ directory:

.apex/
  state.json      -- current stage and session state
  tasks.json      -- task list and state machine
  memory.json     -- memory store (persists across sessions)
  analytics/      -- usage data
  screenshots/    -- browser screenshots

Already in .gitignore. Won't be committed.

---

Every scenario below shows two ways to work: talk to the AI in natural language, or run a command directly. Mix and match.

Tell the AI: "I want to add user auth with JWT"

Or drive each stage manually:

# 1. Requirements exploration (hard gate: no code gets written here)
/apex-forge-brainstorm
# -> produces docs/brainstorms/auth-requirements.md

# 2. Implementation plan (file paths, function signatures, test scenarios)
/apex-forge-plan
# -> produces docs/plans/auth-plan.md

# 3. TDD execution (tests first, complex tasks auto-split to sub-agents)
/apex-forge-execute

# 4. Quality review (18 perspectives: security, correctness, framework-specific