Compaction cache misses: why every auto-compaction re-bills the whole context

August 15, 2026 · View on GitHub

By zoahdev 路 2026-08-15 路 Upstream patch: https://github.com/zoahdev/deepseek-harness/tree/fix/compaction-inherit-header-config

TL;DR

The summarizer request only inherited provider/model and forced maxTokens: 8192, so its parameters diverged from normal turns - the provider prefix cache missed completely and ~122k tokens were re-billed on every compaction.

1. Evidence (real session data)

Before the fix:

{"inputTokens":122558,"outputTokens":1584}

No cacheReadTokens - a 100% miss. After the fix:

{"inputTokens":402,"outputTokens":3548,"cacheReadTokens":220928}

99.7% of the prefix is reused; only the compaction instruction itself is new.

2. Root cause

summarizeWithLlm (compaction-basic) built the summarizer call as:

{ provider, model, messages, system, tools, maxTokens: config.maxTokens(8192), ... }

It picked only provider/model from the last routed request header, dropped reasoningEffort and friends, and forced a maxTokens that normal turns never send. Providers whose cache key includes reasoning_effort / max_tokens (DeepSeek KV cache, OpenAI-style gateways) treated the compaction call as brand new.

3. Fix

fix/compaction-inherit-header-config: inherit the routed header config wholesale (same semantics as agent-loop's request proposal), drop adapter-defaulted fields as usual, and stop forcing maxTokens. The compaction call now keeps the exact cache key of normal turns.

4. How to self-check

In a long session, look at the compaction event usage:

  • cacheReadTokens present and much larger than inputTokens -> cache is being reused;
  • inputTokens near the context size with cacheReadTokens: 0 -> the cache is being defeated; check whether you run a patched build.

5. References