The mechanisms span tool result re-truncation, cacheWrite drops to zero, DeepSeek KV cache oscillation, non-append-only prompt history, contextInjection 100x slowdown, compaction hard-cap timeouts, and now byte-stability breaks from metadata mismatch on every turn.
The mechanisms span tool result re-truncation, cacheWrite drops to zero, DeepSeek KV cache oscillation, non-append-only prompt history, contextInjection 100x slowdown, compaction hard-cap timeouts, and now byte-stability breaks from metadata mismatch on every turn.
The mechanisms span tool result re-truncation, cacheWrite drops to zero, DeepSeek KV cache oscillation, non-append-only prompt history, contextInjection 100x slowdown, compaction hard-cap timeouts, and now byte-stability breaks from metadata mismatch on every turn.
The active-message metadata mismatch means the prompt cache is never stable — every conversation turn invalidates the cache. Users are paying for full-context reprocessing on every interaction despite the platform claiming cache support.
The compaction-timeout recovery regression (#100312) was supposedly fixed but has returned in v2026.6.11. The continuation throw is scored as model candidate_failed, compounding the cost impact. Fixes that do not stick are a separate reliability concern.
Seven independent cache-breakage mechanisms mean prompt caching is a feature on paper only. Users are silently paying full price for every turn. This is not a caching implementation — it is a cost-amplification engine disguised as a feature. Gobii's managed platform with tested caching avoids this entire class of hidden cost.
This analysis is based on publicly available documentation, community forums (Reddit, Discord, GitHub), vendor-published case studies, security compliance reports, and hands-on testing by the gobii.reviews editorial team. All claims are sourced and verified. We do not accept payment for inclusion or ranking. See our full methodology and editorial standards.
Last updated: June 30, 2026. Published by the gobii.reviews Editorial Team.