Agent Skills v3.1.0
A lightweight, context-efficient Claude Code plugin for Java Spring backend development.
Built on Harness Engineering principles: Agent = Model + Harness. The plugin provides the harness — hook-bootstrapped skill system with verify/fix loops, observability traces, context budget management, and state persistence on disk.
Quick Start
Step 1 — Install plugin (one-time per machine)
/plugin marketplace add taipt1504/agent-skills
/plugin install devco-agent-skills
Step 2 — Setup project (one-time per project, lead dev)
/setup
Interactive setup walks you through mode selection (standard/yolo/strict), workflow options (auto-verify, auto-review, max retries), and team features. Or use --mode standard to skip prompts.
Step 3 — Commit and share
git add .claude/ && git commit -m "chore: add Claude Code project context"
Teammates clone → run claude → auto-prompted to install plugin. Zero setup needed.
Step 4 — Start coding
/plan # for any non-trivial task
Architecture
Design Philosophy: Harness Engineering
The plugin treats the agent as Model + Harness. CLAUDE.md is passive (project conventions, ~800 tokens). The harness — hooks, skills, agents, and state files — does the heavy lifting:
| Harness Component | Implementation |
|---|
| Tool Orchestration | 13 hooks across 7 lifecycle events, skill-router auto-matching |
| Knowledge Curation | 19 domain skills, 3-tier lazy loading, ≤800 tokens each |
| Context Management | Token budget estimation, progressive unload warnings at 70/85/95% |
| State Persistence | workflow-state.json, verify-fix-state.json, build-checkpoint.json — all on disk |
| Verification Loops | Ralph Pattern: detect failure → extract error signature → retry with circuit breakers |
| Observability | execution-trace.jsonl (per-call), session-metrics.json (aggregated) |
Hook-Bootstrapped Enforcement
SessionStart hook → injects bootstrap/SKILL.md
→ Agent learns: search skills → announce → use → follow workflow
→ Auto-detects: Java/Spring type, Summer Framework, project structure
→ Lazy loads: domain skills on demand, ≤800 tokens each
Context Budget
| Scope | Token Limit |
|---|
| Bootstrap skill | ≤ 1,700 |
| CLAUDE.md | ≤ 1,000 |
| Each domain skill | ≤ 800 |
| Each rule | ≤ 500 |
| Auto-loaded per session | ≤ 5,000 |
| Max with lazy-loaded skills | ≤ 15,000 |
5-Phase Workflow
PLAN → SPEC → BUILD (TDD) → VERIFY → REVIEW
| Phase | Command | Agent | Auto-Transition |
|---|
| PLAN | /plan | planner (opus) | → remind /spec |
| SPEC | /spec | spec-writer (opus) | → remind /build |
| BUILD | /build | implementer (sonnet) | → AUTO /verify |
| VERIFY | /verify | pipeline | → AUTO /dc-review |
| REVIEW | /dc-review | reviewer (opus) | → TASK COMPLETE |
Skip condition: ≤5 lines, 1 file, no new behavior → BUILD directly.
Verify/Fix Loop (Ralph Pattern)
When BUILD or VERIFY fails, the harness handles recovery automatically:
1. Hook detects gradle failure → extracts normalized error signature (MD5)
2. Attempt < maxRetryOnFail AND not same error repeated:
→ Agent runs /build-fix → re-runs /verify (automatic)
3. Same error ≥ noProgressThreshold (3) → ESCALATE to user
4. Total attempts ≥ maxRetryOnFail (3) → FORCE_ACCEPT with warning
State persists in .claude/verify-fix-state.json — survives context resets.
Structure
taipt1504