Forge Studio

Agent = Model + Harness. Research shows changing only the harness produces a 6x performance gap (Meta-Harness, 2026). Forge Studio implements harness principles as composable Claude Code plugins.

12 plugins. 40 skills. 32 hooks. 4 agents. 14 behavioral rules.

---

Install

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

# Install by layer — pick what you need
/plugin install behavioral-core@forge-studio    # Behavioral steering (start here)
/plugin install context-engine@forge-studio      # Context window management
/plugin install memory@forge-studio              # Cross-session recall
/plugin install evaluator@forge-studio           # Quality gates & review
/plugin install workflow@forge-studio            # Orchestration patterns
/plugin install agents@forge-studio              # Multi-agent decomposition
/plugin install reference@forge-studio           # Power-user tips
/plugin install traces@forge-studio              # Execution trace collection
/plugin install diagnostics@forge-studio         # Codebase health scanning
/plugin install caveman@forge-studio             # Token-optimized output
/plugin install token-efficiency@forge-studio    # Duplicate read detection
/plugin install research-gate@forge-studio       # Read-before-edit enforcement

Start a new session after installing for plugins to load.

Templates

cp templates/CLAUDE.md ./CLAUDE.md       # Lean project instructions
cp templates/settings.json ~/.claude/settings.json  # Power-user settings

See docs/settings.md for settings documentation.

---

Plugins

Plugin	Purpose	Hooks	Skills
behavioral-core	Behavioral steering via modular rules.d/ rules, destructive command blocking, scope discipline	4	3
context-engine	Progressive context pressure, session handoffs, edit safety, environment bootstrap, compaction recovery, task tracking	12	6
memory	Three-tier memory: pointer index → topic files → searchable transcripts	0	3
evaluator	Static analysis gates (PHP/JS/TS), adversarial review, verification, test nudge	5	7
workflow	Daily lifecycle (morning → eod → weekly), task routing, explore/plan/implement cycle	0	7
agents	Planner/Generator/Reviewer triad with tool-isolated capability boundaries	0	4
reference	Hidden Claude Code features: thinking modes, parallel patterns, CLI piping	0	3
traces	JSONL execution traces, compiled views, failure mining, harness evolution	3	4
diagnostics	Documentation drift, registration gaps, convention violations, stale memory	0	1
caveman	Always-on compressed output (~65% token savings). Survives compaction.	2	1
token-efficiency	Duplicate read detection, session token audit	1	1
research-gate	Blocks Edit/Write on unread files + exploration depth warnings	4	0

Key Skills

Skill	Plugin	What it does
/morning	workflow	Daily planning: review yesterday, check handoffs, prioritize today
/plan <task>	workflow	Create implementation plan with files, changes, risks
/implement	workflow	Execute plan step-by-step with scope checks
/verify	evaluator	Evidence-based completion check before claiming done
/scope <task>	behavioral-core	Define task boundaries and acceptance criteria
/handoff [topic]	context-engine	Generate session transfer document
/resume	context-engine	Pick up from latest handoff
/explore <what>	workflow	Subagent exploration without polluting main context
/dispatch	agents	Analyze task, recommend single-agent vs fan-out vs pipeline
/trace-compile	traces	Compile raw JSONL traces into summary and error views
/trace-evolve	traces	Mine failure patterns, propose harness improvements
/healthcheck	evaluator	Run quality pipeline (Pint + Larastan + optional tests)
/audit-context	context-engine	Analyze token overhead from CLAUDE.md, plugins, MCP servers
/entropy-scan	diagnostics	Full 6-check codebase health scan

Agents

Agent	Plugin	Tools	Role
planner	agents	Read, Glob, Grep, Bash	Read-only exploration + design
generator	agents	Read, Write, Edit, Bash, Glob, Grep	Implementation
reviewer	agents	Read, Grep, Glob, Bash	Read-only critique
adversarial-reviewer	evaluator	Read, Grep, Glob	Skeptical security/edge-case review

---

Active Hooks

Hooks fire automatically. No commands needed.

Session Lifecycle

Event	Plugin	What it does
SessionStart	context-engine	Environment snapshot: OS, memory, languages, tools, git state
SessionStart	context-engine	MCP server instruction token monitoring
SessionStart	caveman	Load compressed communication rules
PreCompact	context-engine	Save scope, plan, handoff, git state to recovery file
PostCompact	context-engine	Re-inject scope, plan, tasks, modified files from recovery
PostCompact	caveman	Re-inject compressed communication rules
SessionEnd	traces	Write session summary to trace file

Every Message

Event	Plugin	What it does
UserPromptSubmit	behavioral-core	Re-anchor all behavioral rules from rules.d/
UserPromptSubmit	context-engine	5-stage progressive context pressure warnings
UserPromptSubmit	context-engine	Track system-reminder injection patterns
UserPromptSubmit	context-engine	Remind about incomplete tasks

Before Tool Use

Event	Plugin	What it does
PreToolUse:Bash	behavioral-core	Block destructive commands (rm -rf, git push --force, etc.)
PreToolUse:*	behavioral-core	Remind of active scope boundaries
PreToolUse:Edit|Write	research-gate	Block edit/write if file not Read in session (exit 2)
PreToolUse:Edit|Write	research-gate	Warn if insufficient exploration before first edit
PreToolUse:Bash	evaluator	Evaluation gate: warn if plan exists but /verify not run

After Tool Use

Event	Plugin	What it does
PostToolUse:Write|Edit	behavioral-core	"Does this change do ONLY what was asked?"
PostToolUse:Read	context-engine	Warn on files >500 lines
PostToolUse:Bash|Grep	context-engine	Warn on large output or near-truncation
PostToolUse:Edit|Read	context-engine	Track edits per file; warn after 3 without re-reading
PostToolUse:Edit	context-engine	Detect thrashing (5+ edits to same file, oscillating regions)
PostToolUse:EnterPlanMode	context-engine	Inject plugin-aware plan mode guidance
PostToolUse:Write|Edit (.php)	evaluator	Run PHPStan on changed file
PostToolUse:Write|Edit (.js/.ts)	evaluator	Run tsc + ESLint on changed file
PostToolUse:Edit|Write	evaluator	Nudge to run tests after every 3rd edit
PostToolUse:Bash	evaluator	Reset test-nudge counter when tests detected
PostToolUse:Write|Edit	traces	Log file change to session trace
PostToolUse:Bash	traces	Log command, exit code, output to session trace
PostToolUse:Read	research-gate	Record file read for edit gate
PostToolUse:Read|Grep|Glob	research-gate	Track exploration depth
PostToolUse:Read	token-efficiency	Warn on duplicate reads
TaskCreated	context-engine	Log task for progress guardian

---

Customization

• Behavioral rules: Add/remove files in plugins/behavioral-core/hooks/rules.d/. Numbered for priority.
• Context thresholds: Set FORGE_CONTEXT_STAGE1-STAGE5 in settings.json env (message counts) or FORGE_CONTEXT_PCT1-PCT5 (percentages).
• Exploration depth: Set FORGE_EXPLORE_DEPTH (default 6, IDE-Bench recommends 8+).
• Test nudge interval: Set FORGE_TEST_NUDGE_INTERVAL (default 3 edits).
• Large file warning: Set FORGE_LARGE_FILE_LINES (default 500).
• Disable traces: Set FORGE_TRACES_ENABLED to "0".
• Disable evaluation gate: Set FORGE_EVALUATION_GATE to "0".
• Disable research gate: Set FORGE_RESEARCH_GATE to "0".
• Disable a plugin: /plugin disable {name}@forge-studio

---

Docs

Doc	Purpose
Architecture	Design rationale, 7-component model, hook mechanics
Research	Research papers and findings backing each design decision
Harness Spec	Mechanical invariants, architectural primitives, protocols
Settings	settings.json best practices
Token Optimization	Where tokens go and how to reduce spend
Execution Traces	Trace collection, analysis, harness evolution
Research Gate	Read-before-edit enforcement design and data
Workflow Lifecycle	Morning-to-weekly development cycle