ECC is an engineering process layer for coding agents. It packages a plan, test, implement, review, verify, remember, improve loop as installable components (agents, skills, hooks, rules, and memory) so the process does not have to be rebuilt in every prompt. Plans become editable artifacts before implementation starts, test-driven development runs as a gated red, green, refactor workflow with captured evidence, and a fresh-context reviewer looks for regressions in the code the main session wrote.
The components divide the work by context behavior. Skills are reusable workflows (TDD, security review, deep research) loaded when a task needs them. Agents are scoped subagents with their own context and tool permissions for planning, review, build repair, security, architecture, and domain work (language-specific reviewers and build resolvers for Go, Python, TypeScript, Java, Kotlin, Rust, Ruby, C++, F#, and others). Rules are always-loaded project or language standards installed selectively. Hooks run outside the model context on harness events and can enforce deterministic checks, including a GateGuard that blocks destructive shell commands before they run. Instincts are patterns learned from sessions with confidence scores, recalled when relevant and clustered into new skills over time.
The Memory Vault distills sessions into summaries, instincts, and skills rather than storing transcripts, and gives Claude Code, Codex, Hermes, OpenClaw, Kimi, and other harnesses one local, inspectable Markdown format for durable context and handoffs between them. Memory is treated as unreviewed context, not executable policy, and an optional MCP server exposes the same save, search, read, and doctor operations.
Claude Code is the primary harness, delivered as a native plugin. Codex has a native plugin, Cursor and OpenCode have beta adapters, GitHub Copilot is instruction-only, and Gemini, Zed, Antigravity, Qwen, Hermes, OpenClaw, Kimi, CodeBuddy, and JoyCode have experimental adapters with no claim of feature parity. The repository root is the source of truth, and each adapter packages or maps the same workflows rather than keeping separate copies. ECC does not pin a model provider: it works through each harness's own configuration, so an official provider, a gateway, or a self-hosted open-weight model (a Kimi checkpoint served on your own GPU capacity, for example) all work unchanged.
AgentShield, a companion scanner, treats the agent configuration itself as an attack surface. It audits instruction files, settings, MCP configs, hooks, agent definitions, and skills for secrets, permission problems, hook injection, and risky MCP servers, and an optional mode runs three Claude Opus agents as attacker, defender, and auditor. The toolkit also includes context controls (a context-budget check, strategic compaction suggestions, and advice to keep MCP server count low), a skill generator that mines a repository's git history into SKILL.md files and instinct collections, and a hosted GitHub App that does the same analysis for larger repositories with automatic pull requests and team sharing.
Features
- Planning before building: plans become editable artifacts the agent confirms before implementation starts
- Gated TDD workflow: red, green, refactor steps with captured failing-test and passing-test evidence, then build, lint, type, and test verification
- Fresh-context review: a separate reviewer agent checks the main session's code for regressions and blind spots
- Agents: 71 scoped subagents for planning, architecture, review, build repair, security, E2E testing, documentation, and language-specific review
- Skills: 302 reusable workflows covering TDD, research, security, docs, frontend, data, ML, operations, and framework-specific patterns
- Hooks and GateGuard: deterministic checks on harness events outside the model context, including a guard that blocks destructive shell commands
- Rules: always-loaded common and per-language standards you install selectively to limit context cost
- Continuous learning: instincts learned from sessions with confidence scores, with import, export, and clustering into skills
- Memory Vault: one local Markdown memory format with project, team, and user scopes, shared across harnesses, with an optional MCP server
- Multi-harness adapters: native Claude Code and Codex plugins, beta Cursor and OpenCode adapters, and experimental targets for Gemini, Zed, Antigravity, Qwen, Hermes, OpenClaw, Kimi, CodeBuddy, and JoyCode
- AgentShield: scans agent configs for secrets, permission issues, hook injection, and MCP risk, with terminal, JSON, Markdown, and HTML output and a CI exit code
- Skill generation: mines a repository's git history into SKILL.md files and instinct collections, locally or through a GitHub App
