Botte Proxy
The proxy forwards HTTP requests to a configured upstream endpoint. Chat requests are processed by the universal compressor, prefix alignment and, when enabled, output shaping. Compatibility depends on the client's paths and request format; the proxy is not a general guarantee for every agent or provider API.
Compression and measurements
compress_messages()calls the universal compressor withreversible=False. The proxy exposes no restore route or original-content retrieval guarantee.- Text and log summaries can change representation. System messages also use the text transform. Retain original instructions and evidence, and compare actual behavior before routing a workflow through the proxy.
- Input counts use a rough
len(text) // 4estimate. Dashboard savings and cost projections are estimates, not proof of billed token reduction or preserved answer quality. No fixed percentage gain is guaranteed.
Endpoint and listener
Compression runs locally, then the request is sent to the configured target. A remote or cloud target receives that data. Check the target and its handling of content before forwarding it.
The CLI binds 0.0.0.0 by default. For a listener on this machine only, specify
--host 127.0.0.1. The target prefix is concatenated with the incoming request
path; avoid duplicating /v1 between the target and client path.
# Client sends /v1/chat/completions; target is the upstream origin.
python -m skills.botte_proxy.cli proxy --host 127.0.0.1 --port 8787 --target http://localhost:11434
Point a compatible OpenAI client at http://127.0.0.1:8787/v1. Credentials,
when required by the selected upstream, can be supplied through an already
configured BOTTE_API_KEY; do not print them in reports.
The running process serves /stats and /dashboard. The separate CLI stats
command reads its own process's in-memory counters; it does not query another
running proxy. Use the running server's /stats endpoint for that process.
The compression routine itself does not call a model. The upstream LLM and the host agent may consume model tokens; inspect provider usage and task quality before claiming a saving.
