Wire-level token minification losslessly compresses Claude Code tool_result
payloads before they reach the model — cutting token cost without changing
semantics (the model sees the same information, fewer tokens).
A telemetry-locked wire-level proxy (a 3rd-party tool pinned in
external-tools.json, run via uv) that sits between Claude Code and the API
and compresses traffic. Started by .claude/hooks/fleet/headroom-proxy-start/
at SessionStart, which also sets ANTHROPIC_BASE_URL to the local proxy. Held
telemetry-OFF and fail-closed; the sfw CDN allowlist is the runtime backstop.
(Replaced the former in-repo @socketsecurity/token-minifier.)
A PostToolUse(mcp__.*) rewriter (.claude/hooks/fleet/minify-mcp-out/) that
losslessly compresses MCP tool_result blocks on the way back to the model.
Both are lossless and on by default; neither changes tool behavior.