Releases·Tool Release
I measured what happens when a coding agent reads the same file repeatedly
A developer released sqz, a local content-addressed cache for Claude Code sessions that replaces exact repeated file content with tiny references. The author observed agents rereading unchanged ~2,000-token files multiple times, paying full token cost each time. Across 3,003 compressions it saved 178,442 tokens with a 24.7% average reduction and no extra LLM call.
8 engagement1 sourceSun, Sep 20, 2026, 06:37 PM
AI Overview
- sqz uses a content-addressed cache to replace exact repeated file content with tiny references, saving 178,442 tokens across 3,003 compressions (24.7% average reduction), with no extra LLM call
- Developer built sqz after observing Claude Code reading the same ~2,000-token file multiple times per session
- sqz preserves exact content rather than summarizing, to avoid incorrect summaries of important material
Entities
Related
Releases·Tool Release
llama.cpp fork deduplicates byte-identical messages in agent loops, cutting prompt tokens (e.g. 18,092 to 6,820)
A llama.cpp fork adds an opt-in pass to llama-server's chat parser that replaces byte-identical repeated messages in agent loops with one-line references, keeping the first copy intact. It is stateless and deterministic so prompt caches keep hitting, and it reports dedup stats in response timings. For agent workloads that repeatedly re-read files or re-send tool outputs, it cuts prompt sizes substantially.
A llama.cpp fork adds an opt-in pass to llama-server's chat parser that replaces byte-identical repeated messages in agent loops with one-line references, keeping the first copy intact. It is stateless and deterministic so prompt caches keep hitting, and it reports dedup stats in response timings. For agent workloads that repeatedly re-read files or re-send tool outputs, it cuts prompt sizes substantially.
10 engagement1 sourceredditSat, Sep 26, 2026, 02:02 PM
Products·launched
Lossless context compressor for coding agents with prompt cache awareness
A lossless context compression tool for coding agents that preserves tool output (file contents, JSON, stack traces, diffs) exactly. It sits between the agent and the LLM, rewriting only blocks that can be proven exactly reconstructible, and attaches a BLAKE3 certificate for verification. Solves the problem of token waste in agent loops without losing critical information.
A lossless context compression tool for coding agents that preserves tool output (file contents, JSON, stack traces, diffs) exactly. It sits between the agent and the LLM, rewriting only blocks that can be proven exactly reconstructible, and attaches a BLAKE3 certificate for verification. Solves the problem of token waste in agent loops without losing critical information.
8 engagement1 sourceredditTue, Jul 21, 2026, 04:25 AM
Products·launched
Open-source MCP server + hook for lossless context compression in Claude Code
This is an open-source MCP server and hook that losslessly compresses Claude Code context, reducing token usage by 2-8x while ensuring byte-exact reconstruction. It solves the problem of context windows filling up with irrelevant data and compaction losing important details like stack traces and IDs. Aimed at developers using Claude Code for long sessions.
This is an open-source MCP server and hook that losslessly compresses Claude Code context, reducing token usage by 2-8x while ensuring byte-exact reconstruction. It solves the problem of context windows filling up with irrelevant data and compaction losing important details like stack traces and IDs. Aimed at developers using Claude Code for long sessions.
7 engagement1 sourceredditMon, Aug 10, 2026, 02:23 PM
Products·launched
Chisle: Token-saving tool for Claude Code that compresses context and avoids unnecessary code generation
Chisle is a tool that integrates with Claude Code to reduce token usage by compressing tool output before it enters the context window and by prompting the model to avoid writing unnecessary code. It solves the problem of high token costs for developers using AI coding assistants, particularly when the model reads large amounts of tool output.
Chisle is a tool that integrates with Claude Code to reduce token usage by compressing tool output before it enters the context window and by prompting the model to avoid writing unnecessary code. It solves the problem of high token costs for developers using AI coding assistants, particularly when the model reads large amounts of tool output.
2 engagement1 sourceredditTue, Jul 28, 2026, 12:34 PM
Social·Community
Made a pattern that cuts agent token waste by 90% — Git as cache, markers auto-generate, no sightseeing
A developer describes a pattern for reducing token waste in coding agents like Cline, where the dominant cost is re-reading already-processed files across turns rather than code generation. The approach treats Git as a cache and auto-generates markers so the agent avoids redundant file reads. It offers an alternative to vector databases, AST indexers, and memory banks for long agent sessions.
A developer describes a pattern for reducing token waste in coding agents like Cline, where the dominant cost is re-reading already-processed files across turns rather than code generation. The approach treats Git as a cache and auto-generates markers so the agent avoids redundant file reads. It offers an alternative to vector databases, AST indexers, and memory banks for long agent sessions.
15 engagement1 sourceredditMon, Sep 7, 2026, 11:27 AM