Coding-agent context cuts that look cheaper can raise executed prefill more than sixfold under prefix caching

Teams still price context policy by prompt tokens deleted. Prefix caching prices the tokens that actually miss the cache.

Yu and colleagues at UCSD posted arXiv:2610.34763 with a quality-gated physical accounting protocol for coding agents. The spine is blunt: under prefix caching, deleting history can raise uncached prefill even while the logical prompt shrinks. Mask-5 is the exhibit that makes the intuition expensive.

The paper sits next to Compact-Bench and CTXAgent on the same desk for a reason. Retention asks what a summary keeps. Self-management asks who decides to cut. Physical accounting asks whether the cut that looks thrifty on a prompt dashboard still spends less on the GPU once the cache is in play.

Physical accounting splits prompt volume from uncached prefill

The protocol treats executed work as more than the prompt token counter. Under a caching runtime, prefill covers only the suffix after the longest cached prefix. The authors split executed prefill into logical prompt minus reused prefix, then track KV residency, time, and an energy proxy, and they refuse savings that fail a five-point non-inferiority quality gate.

On 450 localisation cases derived from SWE-rebench V2, Eager preload of ten retrieved blocks and Compact (Eager plus anchored history compaction) both pass the gate. In a separate 900-cell resource study, Eager cuts model calls by 35.7% and the GPU-energy proxy by 33.2% against on-demand Grep; Compact saves nothing further once Eager is in place. Enabling prefix caching leaves Eager's prompt volume similar (+8.3%, CI spanning zero) while cutting its uncached prefill by 67.7%.

The cautionary policy is Mask-5 on mini-SWE-agent across 100 SWE-bench Verified instances: keeping only the last five observations cuts Eager's prompt volume by 33.3%, loses about 135.9k reusable tokens per task, raises uncached prefill by 608.5%, raises the energy proxy by 125.6%, and misses the quality gate.

Compaction and Raw-RAG debates still sell logical token cuts

Compact-Bench and related work made retention loss visible. CTXAgent asked who should own compaction. Operators still ship "trim history" and "mask old observations" as cost wins because dashboards show prompt tokens.

Prefix caching is now default on many serving stacks. That flips the economics: a rewrite that breaks a long cached prefix forces the runtime to re-prefill work that used to be free. Yu et al. give infra and harness owners a measurement language that matches the GPU bill rather than the prompt counter.

UCSD preprint with SWE-rebench and SWE-bench Verified cells

Paper: arXiv:2610.34763, Yu et al., UCSD. Evaluation includes 450 SWE-rebench V2 localisation cases, a 900-cell resource study comparing Grep, Eager, and Compact, and a mini-SWE-agent Mask-5 study on 100 SWE-bench Verified instances. This is a research measurement protocol and policy comparison. Take any public code from the paper's artefact links when present, and treat the work as paper-first until a pinned repo is verified for your stack.

The contribution is methodological as much as comparative: quality-gated physical metrics that stay honest when logical prompt volume and uncached prefill move in opposite directions.

Prefer trajectory-shortening policies that keep the cache warm

Reach for Eager-style preload when localisation agents repeatedly re-search the same neighbourhood and when model-call count dominates cost. Prefer Compact only after Eager is already paying for itself; on this suite Compact cleared quality without further resource wins.

Avoid Mask-5-style hard history deletes on prefix-caching runtimes unless you re-measure uncached prefill and quality together. If your serving path has prefix caching off, the 67.7% Eager uncached-prefill cut will not appear, and the ranking can change. Re-run the gate on your tokenizer, cache keying, and agent observation format before promoting a trim policy.

Infra owners should also keep call count on the same dashboard as prompt tokens. Eager's win on this suite is partly fewer model calls from preloaded blocks, which is a trajectory effect rather than a summariser effect.

Logical savings can hide cache breaks and quality misses

Any policy that rewrites early tokens, renumbers turns, or drops reusable observations can destroy prefix reuse. Prompt-volume dashboards will celebrate while GPU time and energy climb. Quality gates that look only at final pass rates can also miss regressions the authors' five-point non-inferiority rule would catch.

Absolute percentages are tied to mini-SWE-agent, the SWE-bench Verified slice, SWE-rebench localisation, and the authors' energy proxy. A different harness observation schema or cache implementation needs fresh cells. Compact's "no further save" finding is conditional on Eager already being the baseline.

Read the Mask-5 and Eager tables before you trim history

Harness and infra owners should start with the Mask-5 uncached-prefill spike and the Eager-versus-Grep call and energy cuts, then instrument prefix-hit rate beside prompt tokens in staging. The operating sentence to keep: shorter logical prompts can raise executed prefill when they break a warm cache.