Governor: A Claude Code Plugin to Cut Token Waste via Output Compression, Context Slimming, and Tool Filtering

Governor is a Claude Code plugin that tackles token and context bloat beyond simple style compression. It reduces output tokens, shrinks recurring memory files like CLAUDE.md, filters noisy build/test logs, and adds planning guardrails to prevent scope drift. It's installed as a Claude Code plugin, exposing a /governor:* command namespace.
Quick Start
bash install.sh --forceRestart Claude Code, then run:
/governor:status
/governor:audit
/governor:compress CLAUDE.mdGovernor auto-starts in compact professional mode. Toggle with /governor:on and /governor:off.
Benchmarks
Small local benchmarks with Sonnet (non-interactive, no tools):
- Output tokens (three technical prompts): Control avg 989 tokens → Governor avg 440 tokens (55.5% savings). Caveman-style gives 44.9% savings.
- Memory compression: 1877-token file → 838 tokens (55.4% savings) with medium compression.
- Tool output filtering: Synthetic
pytest -vvnoisy output of 54,314 estimated tokens → 1,726 tokens (96.8% blocked) with preserved failure lines.
Key Features
- Always-on compact mode via
SessionStartandUserPromptSubmithooks. - Professional memory compression:
/governor:compress CLAUDE.mdrewrites verbose files while preserving code blocks, inline code, paths, URLs, headings, tables, warnings, and env vars. A quality guard rejects low-savings compressions and restores the backup. - Tool-output filtering: Summarizes noisy Bash/test/build output but preserves failure signals.
- Telemetry ledger:
/governor:statusreports blocked tokens, failures, compactions. - Prompt guidance: Soft suggestions for vague broad prompts.
- Plan and drift guard:
/governor:plancreates an implementation contract;/governor:guardchecks current changes against it. - Portable rule snippets: Compact-mode rules included for Codex, Gemini, Cursor, Windsurf, and Cline.
Commands
/governor:on//governor:off– Enable/disable response compression/governor:status– Usage dashboard and waste heat map/governor:audit– Find bloated memory/rule files/governor:compress CLAUDE.md– Compress memory files/governor:full– Next Bash command returns full output/governor:plan//governor:guard– Scope management/governor:benchmark– Explain benchmark results
Installation
# Local development
claude --plugin-dir .
One-line local install
bash install.sh --force
From repository
gh repo clone 0xhimanshu/governor
cd governor
bash install.sh --force
Heavy Claude Code users should try Governor if they struggle with context bloating, broad task drift, or noisy logs eating their quota. It's a serious alternative to style-only token savers.
📖 Read the full source: HN AI Agents
👀 See Also

Six open-source tools that address OpenClaw's security, cost, and complexity issues
A developer tested six community tools to solve OpenClaw's security gaps flagged by Cisco, spiraling costs, and complex setup. ClawSec provides security scanning and integrity verification, Antfarm enables deterministic multi-agent workflows, and LanceDB Pro improves memory retrieval with hybrid vector search.

BuddyBoard: A Competitive Leaderboard for Claude Code's /buddy Feature
BuddyBoard is a community-built tool that creates a competitive leaderboard for Claude Code's /buddy feature, generating trading cards with stats, rarity tiers, and a BuddyDex tracking 1,728 possible combinations. Run with npx buddy-board to submit your buddy to the global ranking.

AGENTS-COLLECTION: 129 Claude Code Agents Organized in One Repository
A developer has compiled 129 Claude Code agents into a single repository in ~/.claude/agents/ format, ready for installation with a simple copy command. The collection includes the full agency-agents system with 68 personality-driven agents across multiple disciplines, plus additional agents for multi-agent team workflows.

Custom llama.cpp Backend Offloads LLM Matrix Multiplication to AMD XDNA2 NPU on Ryzen AI MAX 385
A developer built a custom llama.cpp backend that dispatches GEMM operations directly to the AMD XDNA2 NPU on Ryzen AI MAX 385 (Strix Halo), achieving 43.7 t/s decode at 0.947 J/tok with Meta-Llama-3.1-8B-Instruct Q4_K_M. The NPU decode path saves ~10W versus Vulkan-only while matching decode throughput.