context-bench

Your AI editor knows what you're working on — automatically.

A self-learning Claude Code plugin that detects your current project from your first message and injects the right file context — without any manual configuration.

How It Works · Installation · How It Learns · Architecture · Toggle

---

The Problem

Every Claude Code session starts the same way: you tell it what project you're working on, what files matter, what the architecture looks like. Then the next session, you do it again.

That's wasted time and context.

context-bench solves this by learning from your sessions. After a few interactions, it knows:

• What keywords belong to which project
• Which files you typically reference
• Which framework markers (pyproject.toml, package.json, Cargo.toml) identify your stack
• When a project is no longer active (confidence decay)

It then injects relevant file paths as invisible context before you even finish typing your first message.

---

How It Works

┌──────────────────────────────────────────────────────────┐
│                  Claude Code Session                      │
│                                                           │
│  User: "fix the parser in my strategy builder"            │
│         │                                                 │
│         ▼                                                 │
│  [UserPromptSubmit hook fires]                            │
│         │                                                 │
│         ▼                                                 │
│  context-bench:                                           │
│    ├─ Keywords: "parser" + "strategy" + "builder"         │
│    ├─ Match: "nq-strategy-builder" (confidence 0.85)      │
│    └─ Inject: /path/to/parser.py, /path/to/knowledge.py   │
│         │                                                 │
│         ▼                                                 │
│  Claude responds with full project context                │
│                                                           │
│  ... user edits files ...                                 │
│         │                                                 │
│  [PostToolUse hook fires] → tracks changed files          │
│                                                           │
│  ... session ends ...                                     │
│         │                                                 │
│  [SessionEnd hook fires] → updates confidence +0.15       │
│                                                           │
└──────────────────────────────────────────────────────────┘

The three hooks

Hook	Trigger	What it does	Latency
UserPromptSubmit	User sends a message	Matches keywords → injects relevant file paths as context	< 200ms
PostToolUse	Tool call completes	Records which files were read/written	< 50ms
SessionEnd	Session closes	Updates confidence scores, decays old topics, creates new ones	< 100ms

---

Installation

As Claude Code Plugin (recommended)

claude plugin add nessos666/context-bench

That's it. The plugin registers all three hooks automatically.

Standalone (manual)

git clone https://github.com/nessos666/context-bench
cd context-bench
./install.sh

install.sh registers the three hooks in ~/.claude/settings.json and copies context_bench.py to ~/.context-bench/.

Local testing

claude --plugin-dir ./context-bench

---

How It Learns

context-bench uses a confidence-based learning system — no ML model, no API calls, just deterministic updates:

Confidence scoring

Event	Confidence change	Why
Match found + files changed	+0.15	Strong signal — you're working on this project
Match found + no files changed	−0.05	Weak signal — maybe a false positive
No match + files changed	New topic at 0.5	Discovered a new project organically
Topic idle > 30 days	−0.01/day	Gradual decay for inactive projects
Topic below 0.3	Auto-removed	Confidence too low to be useful

Bootstrap (first run)

On first run (no projects.json yet), context-bench scans the current directory for framework markers:

Marker	Detected as
pyproject.toml / requirements.txt / setup.py	Python project
package.json	Node.js project
go.mod	Go project
Cargo.toml	Rust project
pom.xml	Java project

Bootstrap completes within 200ms to avoid blocking hooks.

Data model

{
  "version": 1,
  "projects": [
    {
      "id": "nq-strategy-builder",
      "keywords": ["strategy", "builder", "nq", "fvg", "backtest"],
      "root": "/home/user/projects/strategy-builder",
      "paths": ["engine/parser.py", "sb/cli.py"],
      "confidence": 0.85,
      "uses": 47,
      "last_used": "2026-05-17T10:30:00"
    }
  ]
}

---

Architecture

context_bench.py (single file, zero dependencies)
├── cmd_prompt()  → UserPromptSubmit  → match topic, inject context
├── cmd_track()   → PostToolUse       → record changed files
└── cmd_learn()   → SessionEnd        → update confidence, create topics, decay

Single file. Zero dependencies. Zero API calls.

---

Toggle (disable/enable)

Temporarily disable context-bench without uninstalling:

touch ~/.context-bench/DISABLED   # disable
rm ~/.context-bench/DISABLED      # enable again

When disabled, all hooks exit immediately (no context injected, no tracking). The learn hook still cleans up session files to prevent leaks.

Claude Code skill shortcuts

~/.claude/commands/ctx-bench-aus.md:

Run: mkdir -p ~/.context-bench && touch ~/.context-bench/DISABLED
Output: "context-bench deactivated"

~/.claude/commands/ctx-bench-an.md:

Run: rm -f ~/.context-bench/DISABLED
Output: "context-bench active"

---

Uninstall

Plugin mode:

claude plugin remove context-bench

Standalone:

./uninstall.sh

---

Requirements

• Python 3.9+ (stdlib only, no dependencies)
• Claude Code with hooks support

---

Project Structure

├── context_bench.py              # Single-file plugin (hooks + learning)
├── hooks/hooks.json              # Hook registration
├── install.sh                    # Hook installer
├── uninstall.sh                  # Hook remover
├── .github/workflows/tests.yml   # CI pipeline
├── tests/test_context_bench.py   # Test suite
└── examples/
    ├── node-project.json         # Example project manifest
    ├── python-project.json
    └── rust-project.json

---

Testing

pytest tests/ -v

# Manual test
python context_bench.py prompt "fix the parser in my strategy builder"

---

License

MIT

Built because repeating your project context every session is a waste of tokens — and brainpower.
github.com/nessos666