Skip to content

landscape.md

Latest commit

 

History

History
52 lines (38 loc) · 4.21 KB

landscape.md

File metadata and controls

52 lines (38 loc) · 4.21 KB
Raw

Landscape: who else works on LLM cache / prefix optimization

The user asked how many open-source projects target the same goal. Short answer: several touch prompt caching or Claude-Code-to-DeepSeek routing, but none is a Claude Code plugin whose specific job is to rewrite CC's request bytes to maximize DeepSeek's automatic prefix cache. That's the gap Permafrost fills.

Surveyed June 2026. Grouped by what they actually do.

A. Prefix-cache stabilization (closest neighbors)

Project What it does vs. Permafrost
Headroom Context-compression proxy + headroom wrap claude. Its CacheAligner detects volatile content in the system prompt and warns; deterministic tool sort; byte-faithful forwarding; session-sticky anthropic-beta. Headroom's compression is the headline; its cache aligner is detector-only. Permafrost is cache-first and relocates volatile content instead of only warning, specialized for DeepSeek's prefix cache, shipped as a CC plugin.
DeepSeek-Reasonix A whole DeepSeek-native coding agent (Go), engineered around the prefix cache: frozen system prefix, sorted tools, turn-tail injection, compaction as the one reset point. Reasonix is a standalone agent, not a way to make Claude Code cache-friendly. Permafrost ports its techniques into a proxy that sits in front of CC.

B. Claude Code ↔ DeepSeek routing (no cache work)

Project What it does vs. Permafrost
claude-code-router (CCR) Local proxy that routes CC requests to different backend models by task. Routing, not cache stability. Could sit alongside Permafrost.
CLIProxyAPI Wraps CC/Gemini/Codex CLIs as OpenAI/Anthropic-compatible API services. Transport/compat layer; no prefix-cache alignment.
claudecode-deepseek-stack (MG-Cafe) Two env vars to point CC at DeepSeek. Exactly the setup Permafrost starts from — then fixes the cache busting that setup leaves on the table.

C. Provider-native / response caching

Project What it does vs. Permafrost
LiteLLM (Claude Code prompt-cache routing) Provider-sticky routing + a warm-up call that pre-caches the system prompt; manages Anthropic cache_control. Targets Anthropic's explicit-breakpoint cache, not DeepSeek's automatic prefix cache; doesn't rewrite CC's volatile prefix.
NadirClaw OpenAI-compatible router; LRU response cache (identical prompt → skip API). Response-level exact-match cache, a different mechanism from input prefix caching.
OpenRouter prompt caching Provider-sticky routing to keep implicit caches warm; advises "keep the start of messages consistent." Hosted routing guidance; Permafrost enforces that consistency on the wire.

D. Claude-Code-specific cache fixes (adjacent, Anthropic-only)

Project What it does vs. Permafrost
claude-code-cache-fix Fixes a prompt-cache regression on resumed CC sessions (Anthropic billing). Anthropic's own cache, resume-specific; not DeepSeek, not prefix-byte alignment.
flightlesstux/prompt-caching Zero-config automatic prompt caching for CC (Anthropic). Manages Anthropic cache_control placement; orthogonal to DeepSeek's marker-free auto cache.

E. Output/context shrinkers (reduce tokens, not stabilize prefix)

RTK, lean-ctx, hosted compressors — they cut how many tokens you send. Complementary to Permafrost: fewer tokens and a frozen prefix compound.

The one-line differentiation

Permafrost is the only one that is (a) a Claude Code plugin, (b) aimed at DeepSeek's automatic prefix cache specifically, and (c) actively rewriting the request bytes (sort, strip, relocate, canonicalize) to keep the tools+system anchor frozen — rather than routing models, compressing tokens, managing Anthropic breakpoints, or merely detecting volatility.