Zeus is a Claude Code plugin that turns a 12-lecture harness-engineering curriculum into 25 actionable skills covering the full agent lifecycle: kickoff, planning, execution, verification, handoff. Grounded in the 7-gate completion cascade and the 5-layer defense model.
- 7-gate completion cascade (
references/seven-gates.md) — the canonical answer to "what does done mean?". G1 (code written) → G7 (handoff state clean), each producing verifiable evidence. - 5-layer defense model (
references/five-layers.md) — failure taxonomy: every agent failure attributes to one of five harness layers. - 12 lectures (
references/twelve-lectures.md) — the source curriculum from harness-engineering. - Skill writing convention (
references/skill-style.md) — the contract every zeus skill follows.
# In a Claude Code session — add the marketplace source, then install:
/plugin marketplace add rainhan99/zeus
/plugin install zeus@zeus-marketplaceOr from a local clone:
git clone https://github.com/rainhan99/zeus.git
# In a Claude Code session:
/plugin marketplace add ./zeus
/plugin install zeus@zeus-marketplaceAfter install, every new conversation auto-loads using-zeus via the SessionStart hook, anchoring the 7-gate cascade and 5-layer model in working memory.
For users who already know the lifecycle, four thin wrappers expose the main phases directly:
| Command | Wraps | When to use |
|---|---|---|
/zeus:brainstorm <idea> |
zeus:brainstorming skill |
Design a new feature — produces a spec doc that hard-gates implementation until approved |
/zeus:plan [spec] |
zeus:writing-plans skill |
Turn an approved spec into bite-sized tasks with TDD steps + footer-signed approval |
/zeus:execute [plan] |
zeus:executing-plans skill |
Walk the approved plan task by task with fresh verification + G4 DoD sweep |
/zeus:ship |
zeus:finishing-a-development-branch skill |
Present merge / PR / push / keep options once G7 closes |
The slash commands are thin — every iron law (spec approval, plan footer signature, TDD per task, Logic Completeness Manifest, DoD gate) is still enforced by the underlying skills. Natural-language triggers (e.g., "I want to add X") continue to work via the SessionStart-injected routing table.
| Skill | Purpose |
|---|---|
using-zeus |
Auto-injected at session start. Anchors the 7-gate cascade, 5-layer model, project entry protocol, and user decision guardrail. |
| Skill | Purpose |
|---|---|
kickoff-agents-md |
Generates AGENTS.md by scanning repo signals + interviewing the user. |
kickoff-definition-of-done |
Refines DoD into command-verifiable items. |
kickoff-feature-list |
Produces .zeus/features.md — per-feature roadmap with stable IDs. |
| Skill | Purpose |
|---|---|
brainstorming |
Design-before-code. Explores options, produces a spec for user approval. |
writing-plans |
Converts approved specs into sequenced, testable task plans. |
decompose-large-projects |
Splits large projects into independent sub-projects. |
| Skill | Purpose |
|---|---|
executing-plans |
Sequential plan executor with 3-failure architectural escalation. |
test-opment |
G2 gatekeeper. TDD red→green→refactor with ecosystem-standard tooling. |
subagent-driven-development |
Subagent-per-task executor with two-stage review. |
dispatching-parallel-agents |
Multi-problem parallelizer with independence verification. |
using-git-worktrees |
Workspace isolation via git worktrees. |
| Skill | Purpose |
|---|---|
verification-before-completion |
G3 gatekeeper. No completion claims without fresh verification evidence. |
e2e-gate |
G4+G5 gatekeeper. DoD sweep + end-to-end pipeline. |
systedebugging |
4-phase root-cause debugging with senior-architect fix discipline. |
requesting-code-review |
G6 request side. Dispatches structured two-stage review. |
receiving-code-review |
G6 response side. Verify before implementing, push back when wrong. |
| Skill | Purpose |
|---|---|
memory-management |
Self-contained project memory in .zeus/memory/ with LLMLingua-inspired tiered compression. |
session-init |
Session startup: loads memory, restores handoff, surfaces lessons. AGENTS.md → CLAUDE.md fallback chain. |
long-task-continuity |
Periodic checkpointing and re-anchoring fs. |
session-handoff |
G7 gatekeeper (write side). Produces handoff memo. |
clean-state |
G7 gatekeeper (clean side). Ensures branch is shippable. |
| Skill | Purpose |
|---|---|
observability |
G7 gatekeeper (audit side). Structured run log with gate status and feature attribution. |
finishing-a-development-branch |
Terminal state after G7. Four options: merge / PR / keep / discard. Agent never auto-merges. |
writing-skills |
Meta-skill for creating new zeus skills. Skeleton template + validation. |
- User decision guardrail — when user instructions violate codebase logic, the agent warns twice with specific conflicts before proceeding. Override decisions are recorded as lessons.
- Ecosystem-standard tooling — ESLint, Ruff, cargo clippy, golangci-lint, etc. Never roll custom scripts when standard tools exist.
- Senior-architect fix discipline — root-cause analysis, best-practice research, codebase-wide pattern scan. Proactively ask user to fix all similar issues at once.
- Instant lesson capture — user corrections are immediately written to
.zeus/memory/lessons/, never compressed, never evicted, loaded every sess## Repository layout
zeus/
├── .claude-plugin/ # plugin and marketplace manifests
├── skills/ # 25 zeus skills
├── hooks/ # SessionStart bootstrap
├── references/ # canonical reference docs (7 gates, 5 layers, 12 lectures, skill style)
└── templates/ # boilerplate for kickoff-agents-md
If you're upgrading from a zeus version that wrote artifacts under docs/specs/, docs/plans/, or kept FEATURES.md at the project root, run scripts/migrate-to-dotzeus.sh once per project:
bash scripts/migrate-to-dotzeus.sh
git commit -m "chore: migrate zeus artifacts to .zeus/"The script is idempotent — re-running it after the move is a no-op. It uses git mv to preserve history. After migration, all plugin-generated artifacts live under .zeus/ (.zeus/specs/, .zeus/plans/, .zeus/features.md, .zeus/memory/, .zeus/state/). The only zeus-related files at the project root are your CLAUDE.md or AGENTS.md contract.
MIT — see LICENSE.