Skip to content

v0.10.3

Choose a tag to compare

@GlitterKill GlitterKill released this 04 Apr 13:22
· 313 commits to main since this release

What's Changed

Added

  • Jina embeddings v2 base-code model support for semantic symbol search alongside nomic-embed-text.
  • Graph-guided cluster expansion — cluster neighbor expansion now uses graph edges for diversity instead of flat member lists.
  • Confidence-aware context planning — the Planner adjusts rung paths based on confidence tiers (high confidence → cheapest plan, low → deeper rungs).
  • Evidence-aware context symbol ranking — retrieved symbols ranked by retrieval evidence quality, not just graph proximity.
  • Semantic-first context seedingsdl.context seeding prefers semantic (embedding) retrieval over keyword-only search when available.
  • sdl.context quality benchmark harness (tests/) for measuring context retrieval accuracy across task types.
  • Compound identifier search in context seeding — multi-word queries generate camelCase variants ("beam search" → "beamSearch") for better symbol discovery.
  • Focus-path inference and language-affinity scoring in context ranking.
  • Response deduplication across context evidence items.

Changed

  • Broad-mode context retrieval overhauled: improved identifier extraction, capped cluster expansion to prevent token blowout, and compacted context responses.
  • Planner broad-mode path selection improved — selects rungs more appropriate for exploratory queries.
  • Task-type ranking weights tuned per task type (debug, review, implement, explain) for more relevant symbol ordering.
  • ETag cache optimizations — reduced redundant card lookups across workflow steps.
  • Context answers preserved under broad-mode truncation (previously dropped when response was trimmed to budget).
  • Memory opt-in model: memory tools hidden from action search when memory.enabled is false; runtime behavior fully gated behind config.

Fixed

  • Token tracking accuracy corrected for multi-step workflow executions.
  • Bloom filter microbenchmark replaced with deterministic correctness assertions (eliminated flaky CI failures).
  • Audit test updated for ranking refactor; hoisted Set allocation out of hot loop.
  • Context contextMode propagation — broad-mode options no longer stripped by gateway schema validation.

Full Changelog: v0.10.2...v0.10.3