mirror of
https://github.com/qwai-tech/obsidian-plugin-intelligence-assistant.git
synced 2026-07-22 16:20:32 +00:00
ecc5d1eca6
This-session changes (reviewed, lint-clean, build+deploy verified): - vector-store: fix misleading "no embedding" comment in search; count and warn on unembedded chunks (was silently scored 0); cache per-chunk vector norms to avoid recomputing cosine norms across queries; expose unembeddedChunkCount in stats. - planner: add structured `expectsWriteProposal` signal so the forced write-proposal retry no longer depends on English prompt-string matching. - chat.service: remove unreachable agent-mode branch (streamResponse only runs in chat mode) and the now-unused toolRegistry dependency; sync call sites. - remove dead code: ChatState (superseded by ChatViewState) and AgentMemoryManager (superseded by agent-memory-service). - docs: ADR-001 documenting the two-layer safety model (write-proposal gate vs policy budget) and why the vendored engine is intentionally permissive. - tests: adverse-condition agent reliability (circuit breaker / maxSteps / abort / failure recovery), real builtin-tools E2E against an in-memory vault, and a live agent E2E (real model + real tools) that auto-skips without a key. Also bundles pre-existing uncommitted WIP not authored in this session (per explicit request to commit everything): LLM providers, web-search-service, document-grader, rag-manager, rag-tab, and their tests. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |
||
|---|---|---|
| .. | ||
| config | ||
| utils | ||
| build.js | ||
| clean.js | ||
| deploy.js | ||
| dev.js | ||
| hot-reload.js | ||
| version.js | ||