Changelog
Release notes and version history for TRW. Every release builds on accumulated intelligence from prior sessions.
Current Versions
0.32.0
trw-mcp
0.4.0
trw-memory
0.3.0
trw-eval
v24.4
Framework
Releases
Executable Assertions & Self-Verifying Memory
- Machine-verifiable grep/glob assertions attached to memory entries — learnings auto-verify against your codebase
- Assertion-aware recall: failing assertions penalize utility scores so stale knowledge ranks lower
- Security-first verification engine: path traversal rejection, pattern caps, binary file detection
- 93 new tests across both packages covering models, SQLite round-trip, all verification types
Device Auth & One-Line Install
- One-line install: curl -fsSL trwframework.com/install.sh | bash — detects Python, installs from PyPI, configures your project
- RFC 8628 device authorization flow — authenticate CLI via browser with QR code support
- OAuth social login with Google and GitHub, including Google One Tap for instant sign-in
- trw-mcp auth commands: login (device flow), logout, and status with multi-org support
Observation Masking & Open Source Prep
- 3-tier context budget middleware: progressive verbosity (full/compact/minimal) reduces tool response tokens as sessions grow — 52% cost reduction
- Open-source publication prep: BSL 1.1 license, public README, secrets scrubbed
- Redundancy detection via SHA-256 hashing suppresses repeated identical tool responses
- CLAUDE.md cache invalidation on upgrade — version bumps force re-render across all projects
Framework v24.4 & Structured Logging
- Framework version bump to v24.4_TRW — coordinated release across all 5 monorepo packages
- Structured logging overhaul: structlog wired across all modules with consistent snake_case event naming
- 150 cross-package integration tests covering tool-to-state-to-persistence boundaries
- Agent Teams worktree merge fix — worktree branches merge before cleanup, preventing work loss
Client Profiles & Adaptive Ceremony
- 5 built-in client profiles: claude-code, opencode, cursor, codex, aider — each with calibrated ceremony weights
- Profile-aware scoring and instruction sync — ceremony adapts to each platform's context window and capabilities
- Response optimizer middleware: compacts JSON, rounds floats, strips nulls for LLM context efficiency
- 7 delivery gate structural fixes and DRY helper extraction for events.jsonl processing
Multi-Platform Support & 24 Tools
- Multi-platform ceremony: full, light, and off modes — works with Claude Code, Cursor, opencode, Codex, and aider
- 24 MCP tools, 6 resources, 24 skills, and 18 bundled agents — complete orchestration toolkit
- Learning injection into AGENTS.md — high-impact learnings auto-promote to platform-generic instruction files
- Installer UX overhaul with real backend health checks and cross-platform MCP server restart after upgrade
Agent Teams Safety & Type System
- Worktree pre-spawn safety: git status check before worktree creation blocks on uncommitted changes
- 79 TypedDict classes across 18 submodules replacing dict[str, object] at all major cross-module boundaries
- AARE-F scoring truthfulness fix: removed stub dimensions, recalibrated weights, scoring ceiling now reaches 100
- Test file ownership enforcement — zero-overlap rules prevent multiple agents from editing the same test file
Documentation Overhaul
- Complete rewrite of 7 documentation pages modeled after Anthropic's docs style — progressive disclosure, scannable tables
- Tools reference expanded from 15 to 24 tools, Skills from 17 to 24, Agents from 11 to 18
- SEO overhaul: WebSite schema, enhanced Organization schema, custom OG copy for key pages, expanded sitemap
- Full API reference with 20 endpoints, auth badges, curl examples, rate limits, and error formats
Interactive Memory Visualizations
- 9 interactive memory sections: growth timeline, tier migration, smart search, knowledge graph, forgetting curve
- Cross-widget dashboard interactions: searching highlights matching tiers, click-to-cycle metrics
- Homepage growth timeline showing framework phases compounding over tasks with animated SVG appreciation cards
trw-memory: Standalone Memory Engine
- Standalone persistent memory engine extracted from trw-mcp — SQLite primary + YAML secondary backends
- Hybrid search: BM25 sparse + dense vector (sqlite-vec) with Reciprocal Rank Fusion
- Knowledge graph, tiered storage (hot/warm/cold), semantic dedup, and LLM-powered consolidation
- Integrations: LangChain, LlamaIndex, CrewAI, and OpenAI-compatible adapters