entroly

Optimizer interference and provider prompt caches

Reducing the number of tokens in one request is not the same as reducing total provider cost. An optimizer can rewrite old conversation history, shorten the current request, and still lose money when the rewrite invalidates a discounted provider prefix cache.

Entroly separates four quantities:

1. Gross reduction: tokens removed from the provider-bound request.
2. Recovery tax: omitted tokens later re-expanded or retrieved.
3. Provider cache usage: cache-read, cache-write, and uncached input tokens reported by the provider.
4. Prefix continuity: a local estimate of how much append-only prefix the upstream agent offered versus how much the optimized request preserved.

The Context Health dashboard calls the first two quantities retrieval-adjusted net tokens. It does not call them economic net savings. Economic attribution requires provider usage, pricing provenance, and a paired baseline with the same workload.

Hash-only continuity measurement

PrefixContinuityGuard renders only provider prompt-bearing fields into a transient canonical surface. It divides that surface into fixed-size blocks and retains SHA-256 digests plus byte counts. Prompt text, code, paths, raw request or conversation identifiers, and model names are not retained.

For each conversation transition it compares:

• the common prefix available in consecutive raw agent requests; and
• the common prefix present in consecutive Entroly outbound requests.

If the outbound prefix is materially shorter, Entroly reports an estimated optimizer-interference risk. This is a local estimate, not proof of a provider cache miss. Provider usage metadata remains authoritative.

Automatic guard

The proxy captures the original provider request before optional image or history optimization. If emergency session rescue is required, its recoverable output becomes the new baseline. Optional image changes, context injection, or history pruning are compared with that baseline. When the provider has reported a warm cache and the optional candidate would materially shorten the reusable prefix, Entroly forwards the safer baseline.

The guard cannot undo:

• outbound redaction;
• fail-closed safety controls;
• recovery persistence; or
• emergency session rescue required to stay inside the context window.

Once session rescue compresses historical evidence, its deterministic frozen representation remains byte-stable on later turns.

Evidence levels

Dashboard field	Evidence basis
Retrieval-adjusted net	Optimization ledger events minus measured recovery adjustments
Live request cache-hit rate	Bounded in-memory provider observations
Cached input-token ratio	Content-blind process-local provider totals by default
Priced and durable usage	Opt-in local usage ledger and pricing catalog
Prefix continuity	Local hash-only estimate
Economic net	Unavailable without a paired baseline

The dashboard does not infer causal savings from a cache correlation and does not convert an unavailable price into $0.00.

This site is open source. Improve this page.