-
-
Notifications
You must be signed in to change notification settings - Fork 80.7k
DeepSeek cache hit rate <10% after 6.x upgrade - boundary-aware caching breaks prefix matching #94518
Copy link
Copy link
Closed
MuftahFree/thermidor
#81Labels
P1High-priority user-facing bug, regression, or broken workflow.High-priority user-facing bug, regression, or broken workflow.clawsweeper:linked-pr-openClawSweeper found an open linked pull request for this issue.ClawSweeper found an open linked pull request for this issue.clawsweeper:needs-live-reproClawSweeper needs live local, crabbox, or manual validation to confirm this issue.ClawSweeper needs live local, crabbox, or manual validation to confirm this issue.clawsweeper:needs-product-decisionClawSweeper marked this issue as needing a product or behavior decision.ClawSweeper marked this issue as needing a product or behavior decision.clawsweeper:no-new-fix-prClawSweeper does not recommend queueing a new automated fix PR for this issue.ClawSweeper does not recommend queueing a new automated fix PR for this issue.impact:auth-providerAuth, provider routing, model choice, or SecretRef resolution may break.Auth, provider routing, model choice, or SecretRef resolution may break.impact:otherThis issue has meaningful maintainer-visible impact outside the owned taxonomy.This issue has meaningful maintainer-visible impact outside the owned taxonomy.issue-rating: 🐚 platinum hermitGood issue quality with a plausible reproduction path needing some confirmation.Good issue quality with a plausible reproduction path needing some confirmation.
Description
Metadata
Metadata
Assignees
Labels
P1High-priority user-facing bug, regression, or broken workflow.High-priority user-facing bug, regression, or broken workflow.clawsweeper:linked-pr-openClawSweeper found an open linked pull request for this issue.ClawSweeper found an open linked pull request for this issue.clawsweeper:needs-live-reproClawSweeper needs live local, crabbox, or manual validation to confirm this issue.ClawSweeper needs live local, crabbox, or manual validation to confirm this issue.clawsweeper:needs-product-decisionClawSweeper marked this issue as needing a product or behavior decision.ClawSweeper marked this issue as needing a product or behavior decision.clawsweeper:no-new-fix-prClawSweeper does not recommend queueing a new automated fix PR for this issue.ClawSweeper does not recommend queueing a new automated fix PR for this issue.impact:auth-providerAuth, provider routing, model choice, or SecretRef resolution may break.Auth, provider routing, model choice, or SecretRef resolution may break.impact:otherThis issue has meaningful maintainer-visible impact outside the owned taxonomy.This issue has meaningful maintainer-visible impact outside the owned taxonomy.issue-rating: 🐚 platinum hermitGood issue quality with a plausible reproduction path needing some confirmation.Good issue quality with a plausible reproduction path needing some confirmation.
Type
Fields
Priority
None yet
Environment
Problem
After upgrading from 2026.5.28 to 2026.6.x, the DeepSeek prompt cache hit rate collapsed from ~95% to <10%. Token costs spike ~10x.
Root Cause
The 6.x boundary-aware caching system splits the system prompt into "stable prefix + variable suffix". This works for Anthropic/OpenAI but breaks DeepSeek prefix-matching cache: splitting the prefix prevents DeepSeek from matching any previously cached prefix.
Already Tried
Related Closed Issues
Request
Please add a config option like disableBoundaryAwareCache in model compat to let DeepSeek users opt out of the split, or apply a provider-specific fix (similar to opencode PR #31867 - move dynamic content to user message instead of system context).