# Setup — Claude Code Aktueller Zustand (2026-03-18)

## System

| Property | Value |
|---|---|
| OS | Pop!_OS + COSMIC + Wayland |
| Shell | bash |
| Python | 3.12+ (uv, ruff, pytest) |
| Node | via nvm |
| Claude Code | v2.1.76-78 |
| Modell Hauptsession | claude-opus-4-6 |
| Modell Subagents | claude-sonnet-4-6 |
| Modell Background | claude-haiku-4-5 |
| Effort | high |
| Experimental | CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 |
| Voice | enabled |
| Startups | 534 (seit 28. Januar 2026) |

---

## CLAUDE.md Ecosystem

| Datei | Details |
|---|---|
| `~/.claude/CLAUDE.md` | Global Master, 280+ Zeilen, 16 KB |
| Project-level CLAUDE.md | 10 Dateien |
| Hierarchie | Global > Project Root > .claude/ Scoped |
| Bekanntes Problem | Bloat, Widersprüche, Attention-Budget-Kollaps |

---

## Hook-System

6 Events, 7 Scripts.

### UserPromptSubmit

| Script | Funktion |
|---|---|
| `hook_inject.py` (Sidecar-NG) | Keyword-Prefilter + LLM-Fallback → additionalContext |

### PostToolUse

| Script | Funktion |
|---|---|
| `usage-tracker.sh` | JSONL Usage Logging |
| `event_logger.py` | Universal Event Logging |
| `coach-git-test.sh` | Git/Test Micro-Lessons (5 min Rate-Limit) |
| `gsd-context-monitor.js` | Context Warning bei ≤35% / 25% |

### SessionStart

| Script | Funktion |
|---|---|
| `gsd-check-update.js` | GSD Version Check |
| `coach-before.sh` | Pre-Session Coaching (10 Regeln, 40 Tips, Streak Counter) |

### Stop

| Script | Funktion |
|---|---|
| `coach-after.sh` | Post-Session Compliance Analysis |

### StatusLine

| Script | Funktion |
|---|---|
| `statusline.sh` | Sidecar-NG Status Display |

### PreToolUse (Project-Level)

| Script | Funktion |
|---|---|
| `safety_gates.py` | 6 deterministische Safety Gates |

---

## MCP Server

| Server | Tools |
|---|---|
| Gemini MCP | ask-gemini, brainstorm, fetch-chunk |
| OpenCode MCP | 13 Tools: ask, run, reply, check, fire, context, review_changes, ... |
| PromptGarage MCP | project-level (nicht global): search, show, use, random_prompt, stats |

---

## Agents (21 Total)

### Implementation

| Agent | Aufgabe |
|---|---|
| sonnet-builder | Code-Edits, Refactoring, Bug-Fixes |
| opus-thinker | Architektur-Analyse, Deep Thinking |
| sonnet-scout | Research, File Finding |
| sonnet-pruefer | Tests, Reviews, Regressions |

### GSD (16 Agents)

| Gruppe | Agents |
|---|---|
| Core | gsd-planner, gsd-executor, gsd-debugger, gsd-verifier |
| Planning | gsd-plan-checker, gsd-codebase-mapper, gsd-phase-researcher |
| Research | gsd-project-researcher, gsd-research-synthesizer, gsd-roadmapper |
| Quality | gsd-integration-checker, gsd-nyquist-auditor |
| UI | gsd-ui-auditor, gsd-ui-checker, gsd-ui-researcher |

### Kontext

| Agent | Aufgabe |
|---|---|
| samuel-kontext | ADHS-informiertes Profil, Kommunikationsstil |

---

## Skills (69 Total)

### Global (1)

| Skill | Funktion |
|---|---|
| `/eval` | Evaluation Framework |

### Sidecar-NG Core (33)

**Tier 1** — Häufig genutzt:
`/recap`, `/check-state`, `/focus`, `/review`, `/test`, `/checkpoint`, `/debug-loop`, `/learn`, `/explore-first`, `/feierabend`, `/research`

**Tier 2** — Orchestrierung:
`/chef`, `/chef-async`, `/chef-lite`, `/chef-subagent`, `/delegate`, `/research-subagent`, `/research-oc`, `/research-swarm`, `/quickwin`

**Tier 3** — Spezialisiert:
`/batch`, `/auto`, `/bigwin`, `/crew`, `/swarm`, `/test-crew`, `/selfimprove`, `/setup-git`, `/snapshot`, `/eval-loop`, `/codex`, `/voice-smart`, `/validate-config`

### GSD (36)

`/gsd:plan-phase`, `/gsd:execute-phase`, `/gsd:new-project`, `/gsd:progress`, `/gsd:do`, und weitere.

---

## Sidecar-NG

| Property | Value |
|---|---|
| `hook_inject.py` | 1.172 LOC |
| `safety_gates.py` | 209 LOC |
| Knowledge Base | 6 Files, max 15K char, priority-ordered |
| Keyword Pre-Filter | ~20 Rules, first-match-wins, 0 ms |
| LLM Fallback Chain | Cerebras → Groq → SambaNova → Together → OpenRouter → Mistral → NVIDIA → GitHub → Gemini → MiniMax |
| Rate Limiting | 3s global, 60s per-provider nach 429 |
| Logging | `inject_log.jsonl`, `safety_gates_log.jsonl` |

---

## Coaching System

| Komponente | Details |
|---|---|
| `coach-before.sh` | 10 rotierende Regeln, 40 Micro-Tips, Streak Counter |
| `coach-after.sh` | Agent Compliance Analyse, Peak Concurrent Agents |
| `coach-git-test.sh` | Real-time Git/Test Lessons |
| `coaching_log.md` | Session Review Log |
| Weekly Cron | Sonntag 20:00 |

---

## Permissions

### Allowed

| Kategorie | Tools |
|---|---|
| Core | Edit, Write, Read, Bash(*), Glob, Grep, Agent, NotebookEdit |
| Web | WebFetch, WebSearch |
| MCP | Gemini (3 tools), OpenCode (13 tools) |
| Other | Skill Tool |

### Denied

`rm`, `rmdir`, `shred`, `dd`, `mkfs`, `apt`

---

## Projekte (38+)

| Kategorie | Projekte |
|---|---|
| Hauptprojekte | Sidecar-ng, MyAIGame, ClaudeCodePanel, ShortcutCounter, StudyAgent |
| Weitere | FireTracker, BookmarkBrain, PromptGarage, AgentArena, Subconsciousness, TypoTuner, Voicemode, Sonnenuhr, SandrasWritingWorld, shared-gui, ... |

---

## Kosten-Tracking

| Property | Value |
|---|---|
| Speicherort | `~/.claude/usage/` (Daily JSONL) |
| Höchste Einzelkosten | Dokumente/Pläne: $44.90 |
| Zweithöchste | AiSystemForVibeCoding: $39.74 |
| Geschätzte Gesamtkosten | $100+ (analysierte Periode) |

---

## Backup

| Datei | Details |
|---|---|
| `claude-backup-2026-03-16.tar.zst` | 837 MB |
| `ClaudeSystemBackup/` | Session-Logs |