Crate llmtrim 

llmtrim CLI/proxy — the user-facing surface over the llmtrim_core engine.

This crate owns everything the embeddable core deliberately excludes: the command-line interface (main.rs), the MITM interceptor (serve), the local daemon/autostart/setup machinery, the SQLite token ledger (tracking), the live A/B benchmark + quality harness, and the terminal UI. The compression itself lives in llmtrim_core and is called through its public compress* API.

Modules

autostart

Run the interceptor at login.

bench

Benchmark harness (quality axis).

breakdown

Interactive cost-breakdown TUI.

compact

Claude Code compaction request detection and ordered model planning.

daemon

Background-daemon control for the interceptor: a pidfile under ~/.llmtrim, plus detached-spawn / liveness / stop. Pure std (no async, no GUI) — the rich CLI face of the always-on proxy. status reads this plus the savings ledger.

discover

llmtrim discover — scan the before/after capture corpus and report WHERE compressible tokens are still escaping compression, so the next compression target is chosen from real traffic instead of guesswork.

doctor

llmtrim doctor — end-to-end install diagnosis (optionally repairs with --fix).

ensure

Bring this install to the recommended current state — one verb for setup, update, and repair.

guard

guard — a Claude Code UserPromptSubmit hook that stops the first turn of a resumed, cold-cache, large-context session so the user sees what it costs before paying it.

mcp

mcp — Model Context Protocol server over stdio.

monitor

Terminal rendering for the monitor command — savings snapshot, live dashboard, time-series reports, and machine-readable export.

quality

Quality evaluation harness.

recall

Ephemeral recovery storage for future lossy first-arrival tool-output shaping.

reroute

Subscription reroute: send intercepted Anthropic /v1/messages traffic through CLIProxyAPI instead of Anthropic.

route_agents

Claude Code custom agents for request-local subscription routing.

serve

serve — MITM HTTPS interceptor (the universal integration), modeled on llm-interceptor but compressing instead of just logging.

setup

llmtrim setup — the one-command bootstrap. llmtrim is only a MITM proxy, so integration is purely at the environment level: it ensures the local CA, then sets HTTPS_PROXY + NODE_EXTRA_CA_CERTS for the user (POSIX: a managed shell-profile block; Windows: HKCU\Environment) so newly-launched tools route through the interceptor and trust the CA — no IDE settings touched, no sudo — enables run-at-login, and starts the daemon. On POSIX the block wires the vars only while the daemon is up (gated on llmtrim _alive), so a shell opened after stop routes directly instead of at a dead proxy; Windows syncs the registry on daemon start/stop instead.

statusline

statusline — one elegant line for Claude Code’s custom status line.

tracking

Re-exports the ledger data layer from llmtrim-ledger.

transport

Transport — send the compressed request to the LLM API and return the response.

tray

Launch the desktop tray app.

ui

Shared terminal UI — one visual voice for every human-facing command.

update

End-user updates: a channel-aware update command + an occasional, cached, opt-out “new release available” check.

window_sub

Private, per-Claude-Code-window subscription overrides.

wrap

llmtrim wrap <agent> [-- <args>] — thin convenience launcher.