Squeezr — AI Context Compression

Everything else in Squeezr shrinks what goes into the model. Output reduction is the first lever on what comes out — and output tokens cost about 5× input on Opus-class models. The proxy never generates output, so it works by reshaping the request. Off by default — enable it in [output] or with SQUEEZR_OUTPUT_SHAPER=1. Both levers are cache-safe.

Verbosity steering

Squeezr appends a small, byte-stable instruction block to the end of the system prompt (after the cache_control breakpoint, so the cached prefix is never touched). It asks the model to skip preamble/postamble, not restate code or output already visible in the conversation, and not narrate after a successful tool call.

Four cumulative levels (level, default 2):

1 — no preamble/postamble; start with the substance.
2 (default) — + never restate code/output already in the conversation; reference by path:line; don't narrate after a successful tool call.
3 — + conclusions only, smallest edit over full rewrites, minimal prose.
4 — minimum tokens, fragments fine, no rationale.

Effort routing

Agentic loops are mostly mechanical continuations: the last message is a clean tool_result (a file read, a passing test). Thinking bills as output tokens, so on those turns Squeezr lowers an already-present thinking budget down to mechanical_thinking_floor (default 1024). It never injects a lever the client didn't send and never toggles thinking on/off — so it can't 400 or bust the cache. On errors and new user questions full effort is kept.

Enable it

[output]
enabled = true            # master switch (default false)
verbosity_steering = true # terse block appended to the system tail (cache-safe)
level = 2                 # 1 (light) .. 4 (minimum tokens)
effort_routing = true     # lower an existing thinking budget on mechanical turns
mechanical_thinking_floor = 1024

Or, without editing the config:

SQUEEZR_OUTPUT_SHAPER=1

Measuring it

The dashboard's Output card reports the echo ratio — the % of the model's output that merely restated context it was already given (lower is better) — plus how many turns were steered and how many had effort lowered. Reading the response to measure this never changes the bytes Squeezr forwards, so it stays cache-safe.