reference

Forge Studio

Agent = Model + Harness. The harness is everything except the model: behavioral steering, context management, memory, evaluation, orchestration, and multi-agent decomposition. Research shows changing only the harness can produce a 6x performance gap.

Forge Studio implements harness principles as composable Claude Code plugins.

9 plugins. 35 skills. 22 hooks. 4 agents.

---

Install

# Add the marketplace
/plugin marketplace add ArtMin96/forge-studio

# Install by layer — pick what you need

# Behavioral Steering (recommended: start here)
/plugin install behavioral-core@forge-studio

# Context Management
/plugin install context-engine@forge-studio

# Memory Architecture
/plugin install memory@forge-studio

# Evaluation & Quality Gates
/plugin install evaluator@forge-studio

# Orchestration
/plugin install workflow@forge-studio

# Multi-Agent Decomposition
/plugin install agents@forge-studio

# Reference & Tips
/plugin install reference@forge-studio

# Execution Trace Collection
/plugin install traces@forge-studio

# Token-Optimized Output (always-on compressed communication)
/plugin install caveman@forge-studio

After installing, start a new session for plugins to load.

Recommended CLAUDE.md

A lean CLAUDE.md template is included at templates/CLAUDE.md. Designed to work with forge-studio plugins — covers personality, judgment, context management, self-evaluation, and project config without repeating what hooks enforce.

cp templates/CLAUDE.md ./CLAUDE.md
# Edit the Project Config and Conventions sections for your project

Recommended settings.json

A power-user settings.json template is included at templates/settings.json. Enables extended thinking, maximum effort, LSP tools, and bypass permissions with a deny list for destructive commands.

# Copy to your global Claude Code config
cp templates/settings.json ~/.claude/settings.json

Key choices:

• Bypass permissions + deny list — allows everything except destructive commands. Two safety layers: the deny list here and behavioral-core hooks.
• No co-authored-by — removes the "Co-Authored-By: Claude" trailer from commits
• Always thinking + high effort — maximizes reasoning quality at the cost of more tokens
• LSP + tool search — enables IDE-level code navigation and on-demand tool loading
• Auto-compact at 75% — compacts context earlier than the default 95%, preventing quality decay
• 90-day transcript retention — extends the default 30-day cleanup period

See Settings Best Practices for detailed documentation.

---

Architecture

┌─────────────────────────────────────────────┐
│                User / IDE                   │
├─────────────────────────────────────────────┤
│           Harness (Forge Studio)            │
│                                             │
│  behavioral-core ──── Steering & discipline │
│  context-engine ───── Context window mgmt   │
│  memory ───────────── Cross-session recall  │
│  evaluator ────────── Quality gates & review│
│  workflow ─────────── Orchestration patterns│
│  agents ───────────── Multi-agent triad     │
│  reference ────────── Power-user tips       │
│  caveman ──────────── Token-optimized output│
│                                             │
├─────────────────────────────────────────────┤
│              Claude Model                   │
└─────────────────────────────────────────────┘

See docs/architecture.md for the full design rationale.

---

Plugin Reference

behavioral-core — Behavioral Steering

Re-injects behavioral rules on every message to prevent drift in long sessions. Rules live in rules.d/ as individual files, priority-ordered by numeric prefix (10-no-sycophancy, 20-no-filler, etc.). Add, remove, or reorder rules by managing files. Hooks enforce ~100% compliance where system prompt instructions degrade to ~80%.

Skill	Purpose
/rules-audit	Audit session for sycophancy, apologies, scope creep, filler
/scope <task>	Define task boundaries and acceptance criteria
/timebox [N]	Set a message budget (default 15) for the current task

context-engine — Context Management

Progressive 5-stage context pressure tracking replaces fixed message-count thresholds. Warns at ~50% (re-read files), ~65% (consider compact), ~75% (recommend compact), ~85% (recommend handoff), ~92% (critical — handoff now). Automatically uses actual context percentage when Claude Code exposes it.