Skills / git context controller
git context controller
Structured context management framework for LLM agents. Implements Git-like operations (COMMIT, BRANCH, MERGE) to manage long-horizon agent memory.
Installation
Kompatibilitaet
Beschreibung
Git Context Controller (GCC) v2
Lean, git-backed context management for LLM agents.
GCC v2 stores hash + intent + optional decision notes instead of verbose markdown. Full context is reconstructed on demand via git show. Dual mode: git-backed or standalone.
Based on the research paper: Git Context Controller
Why GCC?
LLM agents lose context as conversations grow. GCC solves this by giving agents structured memory:
- Lean storage -- ~50 tokens per entry vs ~500 in v1
- Git-backed truth -- real commits, not narrative copies
- Safe experimentation -- branches via git worktrees for real isolation
- Cross-session recovery -- reconstruct context from hashes on demand
- aiyoucli bridge -- auto-feeds commit data to vector memory when available
How It Works
Git Mode (with repo)
.GCC/
├── index.yaml # Single source of truth (~50 tokens/entry)
├── branches/ # Branch-specific notes
├── worktrees/ # Worktree tracking
└── .bridge-log # Sync state with aiyoucli
index.yaml entry:
- id: C001
hash: 85c8539 ← pointer to git truth
intent: "release prep" ← why
note: "descartamos semantic-release por overhead" ← optional decision
branch: main
date: "2026-02-25T21:40:00Z"
Standalone Mode (no repo)
Falls back to markdown files compatible with v1:
.GCC/
├── index.yaml # Timeline
├── main.md # Roadmap (v1 compat)
├── log.md # OTA traces (v1 compat)
└── branches/
Context Reconstruction
Agent needs context
│
▼
gcc_context.sh --summary ← ~50 tokens, zero git calls
gcc_context.sh --last 5 ← reconstructs via git show
gcc_context.sh --decisions ← only entries with notes
gcc_context.sh --hash abc123 ← full diff for one commit
Installation
As a Claude Code Skill
# Via skills.sh
npx skills add faugustdev/git-context-controller
# Manual installation
cp -r gcc/ your-project/.claude/skills/gcc/
Standalone
git clone https://github.com/faugustdev/git-context-controller.git
# Initialize GCC in your project
cd your-project
/path/to/scripts/gcc_init.sh
Quick Start
Once installed, GCC activates automatically. Use commands or natural language:
| Action | Command | Natural Language |
|---|---|---|
| Save progress | /gcc commit | "save this milestone" |
| Try alternative | /gcc branch experiment | "branch to try a different approach" |
| Integrate results | /gcc merge experiment | "merge the experiment results" |
| Quick status | /gcc context --summary | "where were we?" |
| Recent work | /gcc context --last 10 | "show recent activity" |
| View decisions | /gcc context --decisions | "what did we decide about X?" |
| Deep dive | /gcc context --hash abc123 | "details on that commit" |
| Clean up | /gcc cleanup | "clean up old worktrees" |
| Sync memory | /gcc bridge --sync | "sync to aiyoucli" |
Scripts
GCC v2 includes utility scripts for mechanical operations:
| Script | Purpose |
|---|---|
| gcc_init.sh | Initialize GCC (auto-detects git/standalone) |
| gcc_commit.sh | Real git commit + lean index entry |
| gcc_context.sh | Reconstruct context from hashes |
| gcc_bridge.sh | Feed commit data to aiyoucli vector memory |
| gcc_cleanup.sh | TTL-based worktree cleanup + index pruning |
gcc_commit.sh
gcc_commit.sh "implement retry logic"
gcc_commit.sh "release prep" "descartamos semantic-release por overhead"
gcc_commit.sh --staged "hotfix: null check"
gcc_context.sh
gcc_context.sh --summary # one-line per entry (cheapest)
gcc_context.sh --last 5 # last 5 with git details
gcc_context.sh --hash abc123 # full diff for specific commit
gcc_context.sh --decisions # only entries with notes
gcc_context.sh --full # everything
gcc_bridge.sh
gcc_bridge.sh --status # check bridge connectivity
gcc_bridge.sh --sync # sync all unsynced entries to aiyoucli
gcc_cleanup.sh
gcc_cleanup.sh --dry-run # show what would be cleaned
gcc_cleanup.sh --force # clean without asking
gcc_cleanup.sh --prune-index 50 # keep last 50 timeline entries
Commands Reference
COMMIT
Persists a milestone. In git mode, executes a real commit.
/gcc commit <summary>
BRANCH
Creates an isolated workspace. In git mode, uses real git worktrees.
/gcc branch <name>
MERGE
Integrates a branch back. In git mode, real git merge + synthesis note.
/gcc merge <branch-name>
CONTEXT
Retrieves historical memory reconstructed from git.
/gcc context [--summary | --last N | --hash HASH | --decisions | --full]
| Flag | Returns | Cost |
|---|---|---|
| --summary | One-line per entry | ~0 extra tokens |
| --last N | Last N entries with git details | ~200 tokens/entry |
| --hash HASH | Full diff for one commit | Variable |
| --decisions | Only entries with notes | Minimal |
| --full | Everything | All entries |
Configuration
Controlled via index.yaml:
config:
proactive_commits: true # Auto-suggest commits after milestones
worktree_ttl: 24h # Auto-cleanup expired worktrees
bridge_to_aiyoucli: auto # auto | off | manual
Project Structure
git-context-controller/
├── SKILL.md # Skill definition (v2)
├── README.md # This file
├── LICENSE # MIT License
├── CONTRIBUTING.md # Contribution guidelines
├── .gitignore
├── scripts/
│ ├── gcc_init.sh # Initialization (git detect + index.yaml)
│ ├── gcc_commit.sh # Commit + lean entry
│ ├── gcc_context.sh # Context reconstruction
│ ├── gcc_bridge.sh # aiyoucli memory bridge
│ └── gcc_cleanup.sh # Worktree TTL + index pruning
├── references/
│ └── file_formats.md # Format specifications (v2)
└── examples/
└── sample_session.md # Example session (v2)
Contributing
Contributions are welcome. See CONTRIBUTING.md for guidelines.
References
- Paper: Git Context Controller: Structured Context Management for LLM Agents
- Concept: Emergent Mind - GCC Topic
License
Aehnliche Skills
last30days skill
AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary
context mode
Context window optimization for AI coding agents. Sandboxes tool output, 98% reduction. 12 platforms
claude seo
Universal SEO skill for Claude Code. 19 sub-skills, 12 subagents, 3 extensions (DataForSEO, Firecrawl, Banana). Technical SEO, E-E-A-T, schema, GEO/AEO, backlinks, local SEO, maps intelligence, Google APIs, and PDF/Excel reporting.
pinme
Deploy Your Frontend in a Single Command. Claude Code Skills supported.
claude ads
Comprehensive paid advertising audit & optimization skill for Claude Code. 250+ checks across Google, Meta, YouTube, LinkedIn, TikTok, Microsoft & Apple Ads with weighted scoring, parallel agents, industry templates, and AI creative generation.
claude code
Claude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.