Compress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
Python agentaianthropicclaude-codecompressioncontext-engineeringcontext-windowcursorfastapilangchainllmmcpopenaiprompt-engineeringproxypythonragtoken-optimizationtokenstypescript
fix: hold Cursor's read_file byte-exact too
Tejas Choprad02e7df5889b
diff --git a/headroom/config.py b/headroom/config.pyindex 5bc864f70..330e20ae4 100644--- a/headroom/config.py+++ b/headroom/config.py@@ -303,6 +303,18 @@ DEFAULT_BYTE_EXACT_EXCLUDE_TOOLS: frozenset[str] = frozenset({"Read","read",+ # Cursor's `Read` equivalent. Named here for a reason worth stating,+ # because it is counter-intuitive: adding a read tool to the proxy's+ # EXCLUDE set to keep it away from lossy compression makes things WORSE+ # on its own. Exclusion is exactly what routes a tool INTO the+ # excluded-tool fold, and the pluggable provider on that path is handed+ # content with no tool name, so it cannot tell a file read from a grep+ # and will happily rewrite raw source. Protecting a read tool takes+ # both halves: excluded from the lossy path, and named here to stay out+ # of the fold. Caught by scripts/verify-lossless-coverage.py in the+ # partner-trial repo, which reported `read_file` reaching the provider+ # seam once the plugin started excluding it.+ "read_file",})