Compressing file reads isn't the bill — tool-schema filtering is
Aug 19 Chen/Shi (Paritok): instrumented Claude Code/Codex sessions — tool filtering saves ~21–57K tokens/turn (linear); content compression ~2%/turn but compounds ~3,350 N²; single-shot SWE-bench quality ≠ multi-turn cost.
Source: An Empirical Cost Attribution of Context-Compression Gateways in Multi-Turn Coding Agents
Chen and Shi (Paritok, Aug 19; arXiv 2609.22114) instrument a production non-destructive compression gateway between coding agents (Claude Code, Codex) and frontier models, then toggle three levers independently: tool-schema filtering, content compression of file reads/tool output, and history summarization. The field assumption they puncture: a strong single-shot compressor (Paritok-4B retains 86.5% SWE-bench Lite quality at 25.7% size) does not imply proportional multi-turn dollar savings. Turn count — not compression rate — is first-order; cache-read pricing (~10× cheaper re-sent prefix) changes which savings matter.
Isolated A/Bs: tool filtering removes a fixed ~21K–57K tokens every turn (linear in (N)) and is the only unambiguously positive lever — ~20:1 to 33:1 vs content savings on MCP-heavy sessions. Content compression shaves only ~2% of the cache-priced prefix per turn, but compressed blocks accumulate and re-send, so cumulative saving fits ≈ 3,350 (N^2) and overtakes the tool filter near (N ≈ 6) before the context window caps it. Non-destructive recall is required for exact edits and costs a bounded one-segment waste per recall — not a multiplicative blowup — so content compression nets positive on long, low-recall sessions and loses on precise multi-line edits that pull originals back. Practical hierarchy: filter tools always (never stub core exec), compress content when sessions are long and recalls rare, measure end-to-end dollars at the right cache tier, and stop citing single-shot quality as a multi-turn cost claim.
Why it matters for code review
Continuous verification under volume dies on invoice before it dies on judgment. Factories that keep agents warm overnight need the same cost decomposition: tool bloat paid every turn will outrun a glamorous compressor, and FOMO for “always-on” without turn-count discipline just buys a more expensive morning queue. Review the gateway like merge rate — which lever moved dollars under real multi-turn trajectories, not which benchmark retained solve quality in one shot.