feat: v2.0.4 "Deep Reference" — cognitive reasoning engine + 10 bug fixes

New features:
- deep_reference tool (#22): 8-stage cognitive reasoning pipeline with FSRS-6
  trust scoring, intent classification (FactCheck/Timeline/RootCause/Comparison/
  Synthesis), spreading activation expansion, temporal supersession, trust-weighted
  contradiction analysis, relation assessment, dream insight integration, and
  algorithmic reasoning chain generation — all without calling an LLM
- cross_reference (#23): backward-compatible alias for deep_reference
- retrieval_mode parameter on search (precise/balanced/exhaustive)
- get_batch action on memory tool (up to 20 IDs per call)
- Token budget raised from 10K to 100K on search + session_context
- Dates (createdAt/updatedAt) on all search results and session_context lines

Bug fixes (GitHub Issue #25 — all 10 resolved):
- state_transitions empty: wired record_memory_access into strengthen_batch
- chain/bridges no storage fallback: added with edge deduplication
- knowledge_edges dead schema: documented as deprecated
- insights not persisted from dream: wired save_insight after generation
- find_duplicates threshold dropped: serde alias fix
- search min_retention ignored: serde aliases for snake_case params
- intention time triggers null: removed dead trigger_at embedding
- changelog missing dreams: added get_dream_history + event integration
- phantom Related IDs: clarified message text
- fsrs_cards empty: documented as harmless dead schema

Security hardening:
- HTTP transport CORS: permissive() → localhost-only
- Auth token panic guard: &token[..8] → safe min(8) slice
- UTF-8 boundary fix: floor_char_boundary on content truncation
- All unwrap() removed from HTTP transport (unwrap_or_else fallback)
- Dream memory_count capped at 500 (prevents O(N²) hang)
- Dormant state threshold aligned (0.3 → 0.4)

Stats: 23 tools, 758 tests, 0 failures, 0 warnings, 0 unwraps in production

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Sam Valladares 2026-04-09 16:15:01 -05:00
parent 61091e06b9
commit 04781a95e2
28 changed files with 1797 additions and 102 deletions

View file

@ -5,6 +5,38 @@ All notable changes to Vestige will be documented in this file.
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
## [2.0.4] - 2026-04-09 — "Deep Reference"
Context windows hit 1M tokens. Memory matters more than ever. This release removes artificial limits, adds contradiction detection, and hardens security.
### Added
#### cross_reference Tool (NEW — Tool #22)
- **Connect the dots across memories.** Given a query or claim, searches broadly, detects agreements and contradictions between memories, identifies superseded/outdated information, and returns a confidence-scored synthesis.
- Pairwise contradiction detection using negation pairs + correction signals, gated on shared topic words to prevent false positives.
- Timeline analysis (newest-first), confidence scoring (agreements boost, contradictions penalize, recency bonus).
#### retrieval_mode Parameter (search tool)
- `precise` — top results only, no spreading activation or competition. Fast, token-efficient.
- `balanced` — full 7-stage cognitive pipeline (default, no behavior change).
- `exhaustive` — 5x overfetch, deep graph traversal, no competition suppression. Maximum recall.
#### get_batch Action (memory tool)
- `memory({ action: "get_batch", ids: ["id1", "id2", ...] })` — retrieve up to 20 full memory nodes in one call.
### Changed
- **Token budget raised: 10K → 100K** on search and session_context tools.
- **HTTP transport CORS**: `permissive()` → localhost-only origin restriction.
- **Auth token display**: Guarded against panic on short tokens.
- **Dormant state threshold**: Aligned search (0.3 → 0.4) with memory tool for consistent state classification.
- **cross_reference false positive prevention**: Requires 2+ shared words before checking negation signals.
### Stats
- 23 MCP tools, 758 tests passing, 0 failures
- Full codebase audit: 3 parallel agents, all issues resolved
---
## [2.0.0] - 2026-02-22 — "Cognitive Leap"
The biggest release in Vestige history. A complete visual and cognitive overhaul.