Skip to content

CHANGELOG.md

Latest commit

 

History

History
68 lines (55 loc) · 3.69 KB

CHANGELOG.md

File metadata and controls

68 lines (55 loc) · 3.69 KB
Raw

Changelog

Unreleased

Added

  • Empty-database regression coverage for the Hermes capsule learning report.

Changed

  • Clarified public README, vision, status, roadmap, and security docs around source-tree versions, published package versions, and UATP 7.4 agent execution traces.

Fixed

  • Hermes capsule learning report now treats a missing local capsules table as an empty dataset instead of failing CI.

1.1.0 - 2026-04-08

Added

  • Signal detection: 7 signal types (correction, acceptance, refinement, soft_rejection, abandonment, requery, code_execution) with regex patterns, word-boundary matching, terse correction detection
  • Confidence calibration: Autoresearch loop using local Gemma to tune confidence weights against 1042 DPO pair outcomes. MAE reduced from 0.567 to 0.176
  • Extended thinking capture: Chain-of-thought extraction from Claude Code transcripts (thinking blocks) and Hermes sessions (reasoning column)
  • Tool call graphs: Structured tool invocation sequences with reasoning-before-action and result previews
  • Economics capture: Token usage, cache hit rates, cost data, eval duration/tokens-per-second for Ollama
  • DPO pair extraction: 1987 training pairs (720 correction chains + 1267 labeled singles) from 79 capsules
  • Cross-model comparison: Analysis tool comparing correction rates, acceptance rates, confidence across models
  • Tool pattern analysis: Bigram/trigram tool sequences correlated with quality outcomes
  • Hermes Agent plugin: Auto-captures CLI and gateway sessions on end, fires via on_session_end hook
  • Gateway capture hooks: Telegram/Discord/Slack sessions now captured at reset, /new, /resume, /branch
  • Standalone Ollama proxy: 345-line proxy at scripts/proxy/ with zero UATP dependencies, Ed25519 signing via PyNaCl
  • Model backfill: Script to recover model names from Claude Code JSONL transcripts (985 capsules updated)
  • Capsule rescore: Re-runs improved signal detector on all existing capsules (14 → 5638 signals)
  • Knowledge engine: Gold standard knowledge system with capsule-backed retrieval and autoresearch optimization
  • Quality grade from signals: Grades now derived from actual correction/acceptance rates, not step count heuristics

Changed

  • Confidence weights calibrated: assistant_base 0.85→0.55, user_base 0.70→0.52
  • ConfidenceExplainer reads from single _CONFIDENCE_WEIGHTS source of truth
  • Session deduplication prevents duplicate capsules for same session
  • Plugin timeout increased to 30s with graceful fallback
  • Thinking stored as separate field, not concatenated into content

Fixed

  • Hardcoded user_id="kay" in rich_hook_capture.py → os.getenv("USER")
  • Empty assistant messages (pure tool dispatch) filtered from reasoning steps
  • Signal detector false positives on short words ("yes" in "yesterday")

1.0.1 - 2026-03-15

Added

  • Ed25519 local signing (Python + TypeScript SDKs)
  • ML-DSA-65 post-quantum dual signing (beta)
  • RFC 3161 timestamping via DigiCert TSA
  • DSSE bundle export (Sigstore-compatible)
  • FastAPI capsule backend with FTS5 search
  • Next.js dashboard (beta)
  • Claude Code hook capture (basic)
  • Layered capsule structure (Events/Evidence/Interpretation/Judgment)
  • Self-inspection (semantic drift, confidence-evidence alignment)
  • Audit-ready enrichment framework

0.2.1 - 2026-03-01

  • Initial PyPI release
  • Basic SDK: create, sign, verify capsules
  • Local key management