Agent harnesses (OpenClaw, Hermes, Agent Zero, MemUbot…) re-send your context package on every single turn. Bloated context is a per-turn tax — in tokens, latency, and degraded instruction-following.
TokensBudget analysis
Per-file and total token estimates via two independent heuristics, clearly labeled as estimates, so you know which file is eating the window.
RedundancyDuplicate detection
Exact duplicate lines and paraphrased near-duplicate sentences (3-gram overlap) across files — the copy-paste drift every long-lived agent accumulates.
StructureSection lint
Checks for identity, tools, constraints, examples, and output-format sections, plus wall-of-text blocks and contradictory "always X" / "never X" pairs.
Trim planRanked savings
Full mode ranks concrete trims by estimated token savings — cut the biggest waste first, not the easiest.
ProjectionsWindow share
Your package's estimated share of the context windows of popular open models (Llama 3.x, Qwen 2.5, Mistral, DeepSeek). Specs flagged verify-before-relying.
Multi-fileOne paste, many files
Separate files with marker lines and the linter splits them automatically:
--- file: AGENTS.md ---
...
--- file: SOUL.md ---
...