Skip to content

pr_body.md

Latest commit

 

History

History
34 lines (27 loc) · 1.79 KB

pr_body.md

File metadata and controls

34 lines (27 loc) · 1.79 KB
Raw

Implements the Unified Mask provider cache mark ledger (issue #29).

Changes

  • CacheMarkLedger service (packages/core/src/session/cache-mark-ledger.ts):

    • getMark(providerID): Retrieve current mark for a provider
    • applyUsage(providerID, usage): Update mark based on cache_read (forward-only)
    • getProjectionMark(providerID): Get mark for compression (fail-safe: returns null on error)
    • isMarkStale(providerID, maxIdleMs): Check if mark should reset due to inactivity

  • Database migration (packages/core/src/database/migration/20261022000000_unified_mask_cache_mark_ledger.ts):

    • Creates provider_cache_mark table (provider_id PK, mark_tokens, prefix_hash, last_used_at)
    • Index on last_used_at for idle detection
    • Uses IF NOT EXISTS for idempotent migrations

  • MemoryContext integration (packages/core/src/session/memory-context.ts):

    • Added providerMark parameter to project() function
    • Only compresses content beyond the mark (suffix-based compression)
    • Content before mark remains byte-identical
    • Fail-safe: any ledger error → full prefix compression
    • Updates marks after projection via usage context

  • SessionRunner integration (packages/core/src/session/runner/llm.ts):

    • Reads provider mark before projection
    • Applies usage after projection to update mark

  • Tests (packages/core/test/cache-mark-ledger.test.ts):

    • Ledger basics: getMark=0 for new, applyUsage updates mark, forward-only, new provider=0
    • getProjectionMark returns mark, isMarkStale for non-existent/recent providers
    • Integration: suffix-only compression, prefix byte-identical, provider isolation
    • A/B acceptance: mark=0=Track A, mark>0=Track B, fail-safe
    • MemoryContext integration: providerMark parameter, falls back on ledger failure

Refs #29