Claude Code plugin that attacks session token waste via output compression, CLAUDE.md slimming, tool-output filtering, telemetry, and scope-drift guardrails.
Key Takeaways
Benchmarks show Governor cuts output tokens 55.5% vs control (vs 44.9% for Caveman) on three technical prompts with Sonnet, no tools.
Tool-output filtering reduces noisy pytest -vv output from ~54,314 estimated tokens to ~1,726, a 96.8% reduction while preserving failure lines.
/governor:compress rewrites CLAUDE.md into dense prose with protected-span validation; low-savings results are rejected and backups restored automatically.
Telemetry ledger at ~/.claude/plugins/governor/ tracks blocked tokens, failed tool calls, compactions, and memory compression savings in JSONL.
Portable rule snippets extend compact-mode behavior to Codex, Gemini CLI, Cursor, Windsurf, and Cline via agent-specific rule files.
Hacker News Comment Review
The single substantive comment challenges the project’s framing: token savings are table stakes; the real unknown is whether compression degrades the model’s effective reasoning quality.
No evidence or benchmarks on intelligence regression are provided in the source, which is a gap given the aggressive filtering and rewriting involved.
Notable Comments
@esafak: “we need to get past touting token savings. You can save more if you just throw everything away!”