StackMap
Subscribe

headroom vs SoL-Pi

Context compression layer for agents: squeezes tool outputs, logs, files and RAG chunks 20-95% before the LLM — reversible, local-first; library, proxy, one-command agent wrap, or MCP. — versus — NVIDIA's Pi extension with four auto-research-discovered efficiency mechanisms - action fusion, observation packing, evidence-preserving log reduction, online context compaction - all opt-in.

The curated verdict

Both compress tool outputs and context before the model sees them; headroom is a harness-agnostic proxy layer, SoL-Pi is Pi-native with evidence-preservation guarantees.

headroomSoL-Pi
Stars73k2.3k
Forks5.6k181
LanguagePythonTypeScript
LicenseApache-2.0MIT
Last activity2 days agoyesterday
Topicscoding, gatewaycoding, agents
Curated connections66

headroom — the curator's take

The dedicated answer to context bloat: content-aware compressors route JSON, logs and code differently (60-95% on JSON, 15-20% on real coding sessions), originals stay cached so the model can retrieve what compression dropped — the reversibility is what makes aggressive ratios safe. Adoption cost is near zero: `headroom wrap claude` and you're running, or use it as proxy/library/MCP. Prompt-cache-aware alignment avoids torching your cache hit rate. NOT free lunch: a lossy-in-context layer between agent and model is another thing to debug when the model 'misses' something — budget for retrieval round-trips — and the 20% coding figure is the honest number, not the 95% headline.

SoL-Pi — the curator's take

Install SoL-Pi if you run Pi as your coding agent and want fewer turns and less context replay without the agent skipping work: Action Fusion runs an edit's validation command in the same tool call, ObservationPack turns repeated big tool results into paged handles, the Reducer only compacts a log when every retained quote still matches the archive. It is an extension over unmodified Pi 0.85.1 - no patches, everything disabled until you write sol-pi.json. Not for Claude Code/Codex users (it uses Pi's extension APIs); headroom, rtk and token-optimizer are the harness-agnostic ways to get the same effect. The paper is the interesting part: the mechanisms were found by scaled auto-research loops over the harness itself, which is autoresearch/evo pointed at agent efficiency.