Skip to content

CHANGELOG.md

Latest commit

 

History

History
369 lines (278 loc) · 24.7 KB

CHANGELOG.md

File metadata and controls

369 lines (278 loc) · 24.7 KB

Changelog

All notable changes to LocalGPT are documented in this file.

Unreleased

[0.3.6] - 2026-04-12

Added

  • Memory Wiki — structured knowledge management with claims, evidence chains, and staleness tracking.
  • Active memory recall — agent automatically recalls relevant memories before replying.
  • Dreaming memory consolidation — background process consolidates and deduplicates memories, wired into heartbeat runner.
  • Compaction checkpoints — session restore points created during context compaction.
  • Model cost tracking — per-session token usage and cost tracking.
  • Elevated permission mode — approval gate for dangerous tools with StreamEvent::ApprovalRequired, HTTP approve endpoint, and loopback auto-approval.
  • TLS for HTTP server — auto-certificate generation, cert CLI commands, and --no-tls flag.
  • Durable outbound message queue — SQLite-backed outbox with retry, wired into Telegram bridge send path.
  • Claude CLI effort configuration — configurable effort level (low/medium/high/max) with /effort slash command for runtime switching.
  • Gen MCP multimodal input (AI3) — image-guided world generation tools.
  • Gen MCP HTTP transport — compliance metadata, starter templates, and AGENTS.md.
  • Gen MCP tool annotations — Dockerfile, Smithery config, and privacy policy for distribution.
  • MCP conformance tests — test suite, case study docs, and composable agent guide.
  • Heartbeat gen experiment dispatch (H2.5) — localgpt-gen subprocess for automated scene generation.
  • SSRF protection — server-side request forgery guard, stricter CORS, NPC brain loop, gallery thumbnails.
  • Per-job MCP server filtering — cron jobs can specify which MCP servers to use.
  • Telegram heartbeat alerts — heartbeat alerts routed to Telegram forum topics.
  • Gen conditional MCP relay — relay only starts when using CLI backends.

Fixed

  • Config show — resolved paths displayed correctly, secrets redacted in output.
  • Clippy and cargo-deny — warnings and stale advisory ignores resolved.

0.3.5 - 2026-03-22

Added

  • WorldGen pipeline (WG1–WG7) — procedural world-building: blockout layout planning (gen_plan_layout, gen_apply_blockout), navmesh with A* pathfinding, hierarchical entity placement (hero/medium/decorative), screenshot evaluation loop, blockout editing with incremental regeneration, scene decomposition with semantic roles, depth map rendering, and styled 2D preview generation.
  • Gen gameplay runtime (P1–P5) — NPC dialogue UI panel, teleport fade and collectible pickup effects, physics constraints, interaction triggers, and terrain runtime systems.
  • Gen AI infrastructure — AI1 local 3D asset generation scaffolding, AI2 AI NPC intelligence with dialogue and behaviors.
  • Gen world inspector — egui overlay panel with 3D viewport picking, WebSocket protocol for remote inspection, scene detection, transform streaming, entity focus. Native clients: SwiftUI (iOS/macOS), Compose (Android), GLB scene snapshot export.
  • Gen headless mode — headless generation pipeline with experiment queue for CI/batch workflows. CLI relay routes MCP tool calls to an existing Bevy window.
  • Gen world forking — gen_fork_world tool for world remix with attribution tracking.
  • Gen noclip mode — N key toggles noclip, V key toggles first/third-person POV.
  • Gen physics — avian3d 0.6 / bevy-tnua 0.30 physics integration with new physics tools.
  • Gen gameplay triggers — PlaySoundAction in triggers, key collection in collectibles.
  • Gen multi-file worlds — multi-file world generation with drift detection.
  • Gen mesh segmentation — optional mesh segmentation on glTF loading (WG6.3).
  • Plugin system — plugin CLI with localgpt plugin enable/disable/list, daemon-level plugin filtering.
  • Pluggable memory backends — sqlite, markdown, and none backends selectable via config.
  • Post-compaction audit logging — Merkle hash chain for tamper-evident audit trail after session compaction.
  • Encryption at rest — encrypted sessions and config secrets, localgpt encrypt CLI commands.
  • Slack bridge — Socket Mode integration for Slack workspaces.
  • Telegram forum topics — support for Telegram forum/topic-based conversations.
  • Telegram reactions — acknowledgment reactions on incoming user messages.
  • Session branching — fork conversation sessions into parallel branches.
  • Self-repair and stuck recovery — agent detects and recovers from stuck states automatically.
  • LLM-based query expansion — memory search queries expanded via LLM for better recall.
  • Memory citation format — search results include source citations.
  • Multimodal embeddings — Gemini embedding provider with multimodal support.
  • Gemini embedding provider — gemini as a new embedding backend.
  • Session transcript indexing — past session content indexed for memory search.
  • Browser automation tool — Chrome DevTools Protocol integration for web interaction.
  • Docker/Podman sandbox — container-based sandbox backend alongside Seatbelt/Landlock.
  • Web UI settings panel — config editing via POST /api/config and web UI.
  • Channels status API — /api/channels/status endpoint and web UI panel.
  • localgpt tool CLI — MCP server management from the command line.
  • Webhook signature verification — HMAC-SHA256 verification for incoming webhooks.
  • Media processing cache — file-hash-based cache for document/audio processing.
  • Image optimization — automatic image resizing/optimization for vision models.
  • OpenRouter alias — openrouter/* provider prefix for convenience.
  • Streaming abort preservation — partial output preserved when streaming is aborted.
  • Post-compaction context injection — AGENTS.md content injected after session compaction.
  • Document and audio tools — document_load and transcribe_audio wired as agent tools.
  • Gemini CLI streaming — gemini-cli tool extraction support in streaming UI.

Changed

  • Legacy avatar system removed — unified to Player + FreeFly camera modes.
  • gen_package_world removed — skill directories are now the sole distribution unit.

Fixed

  • Gen physics — Tnua initiate_action_feeding called every frame (not just on jump), physics engine plugins registered correctly, SpatialQuery made optional in camera follow system.
  • Gen display — player mesh hidden in first-person mode, all keyboard controls shown in startup help, 4 P0 creator workflow blockers resolved.
  • Gen stability — wgpu surface crash on Wayland avoided, audio buffer underruns fixed, MCP relay and GPU lock hardened, CLI backend MCP tool calls routed to existing Bevy window.
  • Windows — HANDLE type casting and BOOL API mismatch fixed for windows-sys 0.59, all compiler warnings resolved.
  • CLI — short flags removed from tool add to avoid clap conflict.

0.3.4 - 2026-03-15

Added

  • 3D World Templates — website section for browsable world templates.

Fixed

  • Gen ECS — resolved Bevy ECS query conflict B0001 in interaction systems.

0.3.3 - 2026-03-14

Added

  • Gen P1–P5 MCP tool specs — NPC dialogue, interaction triggers, terrain, collectibles, portals, and physics constraint tools.
  • Gen click triggers, portals, and sky lighting — click-based interaction triggers, portal visuals, and sky lighting system.
  • Gen 3D UI enhancements — improved interaction effects with Apple SceneKit animation fixes.
  • iPad/RealityKit — RealityKit world generation with MCP tools on iPad, with macOS conditional compilation.
  • Strict CORS — enforced strict CORS origins for the HTTP server.
  • Session management enhancements — previews and history playback in the CLI.

Changed

  • Physics made optional — avian3d physics behind a feature gate for lighter builds.
  • Interaction params — replaced string-typed interaction parameters with enums and added terrain features.

Fixed

  • Gen NPC idle — avoided query conflict in npc idle system.
  • Core clippy — addressed clippy warnings across codebase.
  • iPad — Swift warnings for RealityKit resolved, macOS conditional compilation added.

0.3.2 - 2026-03-08

Added

  • Gen MCP server — MCP server mode for external CLI backends with core tools, memory-only mode with write tools, and Ctrl+C handling.
  • Gen new primitives — pyramid, tetrahedron, icosahedron, and wedge parametric shapes.
  • Gen HTML export — export worlds as interactive browser-based experiences with Three.js viewer.
  • Gen avatar movement and camera control — WASD/arrow key movement, mouse look, first-person/third-person camera modes.
  • CLI TUI slash commands and line editing — /help, /clear, /exit, /model, /session commands with readline-style editing (Ctrl+A/E, arrow keys, history).
  • Gen streaming chat — real-time tool call display and streaming responses in Gen mode for better visibility into scene building.
  • Gen batch entity operations — gen_spawn_entities tool for efficient multi-entity creation in a single call.
  • Gen on-demand export — gen_export_world tool with mesh asset localization for portable scene exports.
  • Gen human-readable filenames — export filenames now use YYYY-MM-DD_HH-MM-SS format instead of timestamps.

Changed

  • Gen mode architecture — delegated to localgpt-gen binary via subprocess for cleaner separation.
  • Gen glTF export format — switched to JSON format with version compatibility for better interoperability.

Fixed

  • Core provider error messages — API errors now show actual provider name instead of generic label for easier debugging.
  • Agent streaming — correct tool message ordering in streaming path. (#88)
  • Gen compiler warnings — silenced warnings in plugin.rs.

Contributors

Thanks to @itripn (Ron Forrester) for fixing tool message ordering in the streaming path!

0.3.1 - 2026-03-03

Added

  • Gen undo/redo system — full undo/redo support with stable entity IDs, LLM tools (gen_undo, gen_redo), and persistence via history.jsonl. Covers entities, lights, behaviors, camera, and environment.
  • Gen audio undo/redo — audio emitter commands now support full undo/redo with gen_audio_emitter and gen_modify_audio operations reversible.
  • Gen behavior system — declarative entity animations: orbit, spin, bob, look_at, pulse, path_follow, bounce. Behaviors stack and persist through save/load.
  • Gen world save/load — complete worlds serialized as skills (SKILL.md + world.ron). Includes gen_save_world, gen_load_world, and gen_clear_scene tools.
  • Gen avatar and tours — avatar and tours sections in world.ron for user presence and guided waypoint sequences with descriptions and movement modes.
  • Gen parametric shapes — unified world data model with shape field on entities. Supported: box, sphere, cylinder, capsule, plane, torus.
  • Gen material properties — full PBR material support: alpha_mode, unlit, double_sided, reflectance, emissive. Exposed in spawn and modify tools.
  • Gen light properties — range, outer_angle, inner_angle for spot lights; direction for directional/spot lights. All persisted and exposed in entity_info.
  • Gen entity info enrichment — entity_info now includes shape type, emissive, light data (direction, range, angles), audio emitter type, and behavior info.
  • Gen glTF tracking — source path tracked through save/load cycle for asset provenance.
  • Document loaders — new module for loading documents (PDF, DOCX, etc.) for RAG workflows.
  • Audio transcription — new module for transcribing audio files via Whisper-compatible APIs.
  • TTS module — text-to-speech integration for voice output.
  • MMR re-ranking — Maximal Marginal Relevance re-ranking for diverse memory search results.
  • CLI completion — shell completion generation for bash, zsh, fish via localgpt completion.
  • CLI cron command — manage cron jobs from CLI (localgpt cron list/add/remove).
  • CLI hooks command — manage lifecycle hooks from CLI.
  • SpacetimeDB multiplayer — web app for multiplayer 3D world collaboration.
  • Mobile workspace editor — iOS/Android UI for editing workspace files with security hardening.

Changed

  • Mobile apps restructured — apps/ios renamed to apps/apple for multi-platform support (iOS + macOS).
  • OAuth providers removed — all OAuth provider implementations removed for simplified authentication.

Fixed

  • Gen visibility persistence — visibility state now persists correctly through save/load and undo snapshots.
  • Gen camera FOV — saves actual camera FOV instead of hardcoded 45 degrees.
  • Gen light saving — light components save on any entity type, not just GenEntityType::Light.
  • Gen loop detection — raised threshold and fixed command channel race condition.
  • Gen world load — undo stack cleared when loading world without saved history.
  • Codex CLI provider — updated for new CLI interface.

0.3.0 - 2026-02-27

A major release bringing the agent platform to production readiness with MCP tool integration, OpenAI-compatible API, cron scheduling, security hardening, mobile apps, and multi-agent orchestration.

Added

  • MCP client support — connect to external MCP tool servers via stdio or HTTP/SSE transports. Tools are auto-discovered and namespaced as mcp_{server}_{tool}. Configure in [mcp] config section.
  • OpenAI-compatible HTTP API — /v1/chat/completions (streaming + non-streaming) and /v1/models endpoints. Enables integration with Cursor, Continue, Open WebUI, and the Python openai library.
  • Cron job scheduling — run prompts on cron expressions (0 */6 * * *) or simple intervals (every 30m). Each job gets a fresh agent session with overlap prevention and configurable timeout.
  • Per-IP rate limiting — token bucket rate limiter on all API routes. Configurable requests_per_minute and burst in [server.rate_limit].
  • Oversized payload guard — RequestBodyLimitLayer prevents OOM from large POST bodies (default: 10MB, configurable via server.max_request_body).
  • Configuration hot-reload — daemon watches config.toml for changes and propagates updates to running services without restart. Also responds to SIGHUP on Unix.
  • Session pruning — auto-cleanup of old session files at daemon startup and hourly. Configurable session_max_age (default: 30 days) and session_max_count (default: 500).
  • localgpt doctor command — diagnostics that validate config, check provider reachability, test memory database, verify MCP connections, and report disk space. Supports --fix and --json flags.
  • Multi-agent orchestration with spawn_agent tool for hierarchical delegation to specialist subagents.
  • OpenAI-compatible provider for third-party APIs (OpenRouter, DeepSeek, Groq, vLLM, LiteLLM).
  • Multi-provider failover with automatic retry across configured fallback models.
  • Lifecycle hook system — beforeToolCall, onMessage, onSessionStart extensibility points.
  • Stuck loop detection — prevents infinite tool-call loops by detecting repeated identical calls (configurable max_tool_repeats).
  • Bearer token authentication for all HTTP API routes via server.auth_token.
  • Session file permissions hardened to 0o600.
  • Codex CLI provider integration (codex/* models).
  • Apple Foundation Models integration for on-device AI on iOS.
  • iOS app with SwiftUI, MVVM architecture, and UniFFI bindings.
  • Android app initial project structure.
  • Profile isolation via -p/--profile CLI option for complete XDG path separation.
  • Hybrid web search with configurable providers (searxng, brave, tavily, perplexity) and native-search passthrough.
  • xAI provider support (xai/*, grok-*) with native web_search tool passthrough.
  • Vertex AI provider — access Claude and Gemini models via Google Cloud with service account authentication (vertex/* models).
  • Gemini API key provider — simple GEMINI_API_KEY authentication without OAuth (gemini/* models).
  • Bridge health monitoring — automatic health status tracking (healthy/degraded/unhealthy) for all connected bridges.
  • CLI bridge (localgpt-bridge-cli) — interactive terminal that connects to a running daemon via bridge IPC socket.
  • Skill routing rules — useWhen and dontUseWhen conditions for context-aware skill activation.
  • Temporal decay for memory search — optional scoring penalty for older memories (temporal_decay_lambda config).
  • LLM reasoning text preservation — thinking/reasoning content emitted alongside tool calls is now preserved in responses.

Changed

  • Actor-based agent execution with Arc<MemoryManager> for improved thread safety.
  • web_fetch extraction upgraded to use the readability crate with fallback text sanitization.
  • BridgeManager moved from core to server crate for cleaner dependency graph.
  • Model routing updated to support Claude 4.6 models.
  • Replaced unsafe string byte-slicing with floor_char_boundary to prevent UTF-8 panics.
  • Config templates expanded with providers.xai, [tools.web_search], [cron], and [mcp] examples.

Fixed

  • Mobile init EPERM by passing Config to MemoryManager.
  • iOS XCFramework library identifiers, actor isolation, and C++ linking.
  • Silent NO_REPLY tokens filtered from user-facing chat responses.
  • Daemon foreground mode logging level.

Contributors

Thanks to all contributors who helped shape this release! Special thanks to @jcorbin for generalizing daemon process handles, improving heartbeat reliability, incremental session saves, provider tooling improvements, and web search summaries; @TranscriptionFactory for tool filter infrastructure.

0.2.0 - 2026-02-14

A milestone release introducing LocalGPT Gen for 3D scene generation, XDG Base Directory compliance, Docker Compose support, and workspace restructuring.

Added

  • LocalGPT Gen — a new localgpt-gen subcrate for AI-driven 3D scene generation (Phase 1+2). (55aa127)
  • Secure Docker Compose setup for running LocalGPT in containers. (#2)
  • XDG Base Directory layout for all paths, following platform conventions on Linux, macOS, and Windows. (#18)
  • Local server config guidance for OpenAI-compatible server setups. (#25)
  • Security section added to README covering sandbox, signed policy, and injection defenses.

Changed

  • Extracted localgpt-gen into its own workspace member and bumped to v0.2.0.
  • Replaced OpenClaw auto-migration with a detection notice (no longer silently migrates config).
  • Configured cargo-release for joint crates.io publishing.

Fixed

  • Added glibc compatibility shim for ort-sys on glibc < 2.38 (fixes builds on older Linux distros).

Contributors

Thanks to @ttulttul (Ken Simpson) for the secure Docker Compose setup, and @cnaples79 (Chase Naples) for documenting local server configuration!

0.1.3 - 2026-02-12

A major release focused on security hardening, new provider support, and the Telegram bot interface.

Added

  • Telegram bot interface with one-time pairing auth, slash commands, streaming responses with debounced edits, and full tool support. Runs as a background task inside the daemon. (#15)
  • Telegram HTML rendering for agent responses with markdown-to-HTML conversion. (#16)
  • GLM (Z.AI) provider support, adding Z.AI's GLM models as a new LLM backend. (#21)
  • Security policy module with HMAC signing and tamper-detecting audit chain.
  • Kernel-enforced shell sandbox for LLM-issued commands using macOS Seatbelt and Linux Landlock/seccomp.
  • Prompt injection defenses with per-turn injection, suspicious-content warnings surfaced to users, and configurable strict policy mode.
  • Windows build support by gating Unix-only sandbox and nix APIs behind cfg(unix).

Changed

  • Renamed localgpt security CLI subcommand to localgpt md.
  • Updated LocalGPT.md init template to match standing-instructions framing.
  • Upgraded all dependencies to latest versions.

Fixed

  • Security block incorrectly sent as a prompt to Claude CLI instead of as a user message.
  • Clippy warnings for Rust 1.93.
  • Landlock/seccompiler API usage updated for latest crate versions.

Contributors

Thanks to @haraldh for building the Telegram bot interface and HTML rendering, and @austingreisman for adding GLM (Z.AI) provider support!

0.1.2 - 2026-02-09

This release enables tool calling for Ollama and OpenAI-compatible providers, and improves memory search quality.

Added

  • Ollama tool calling support, allowing Ollama models to execute agent tools. (#14)
  • Desktop feature flag for headless builds (compile without GUI dependencies).
  • GitHub Actions CI workflow with license audit via cargo-deny.

Fixed

  • OpenAI provider tools were silently dropped during streaming — the default chat_stream fallback now forwards tools and handles ToolCalls responses correctly. (#11)
  • Memory search improved with token AND matching and rank-based scoring for more relevant results. (#10)
  • Linux desktop builds now include x11 and Wayland features.

Contributors

Thanks to @JarvisDeLaAri for enabling Ollama tool calling, and @Ax73 for fixing OpenAI provider tool support!

0.1.1 - 2026-02-07

Introduces the desktop GUI, GGUF embedding support, and workspace concurrency safety.

Added

  • Desktop GUI built with egui, providing a native app experience.
  • GGUF embedding support via llama.cpp for fully local semantic search.
  • Streaming tool details and slash commands in the egui and web UIs.
  • Concurrency protections for workspace file access.

Fixed

  • UTF-8 boundary panics in memory search snippets resolved; indexing simplified.

0.1.0 - 2026-02-04

Initial release of LocalGPT — a local-only AI assistant with persistent markdown-based memory.

Added

  • Interactive CLI chat with streaming responses and tool execution.
  • Multi-provider LLM support: Anthropic, OpenAI, Ollama, and Claude CLI.
  • Markdown-based memory system with MEMORY.md, daily logs, and HEARTBEAT.md.
  • Semantic search using SQLite FTS5 and local embeddings via fastembed.
  • Autonomous heartbeat runner for background task execution on configurable intervals.
  • HTTP/WebSocket API with REST endpoints and real-time chat.
  • Embedded Web UI for browser-based interaction.
  • OpenClaw compatibility for workspace files, session format, and skills system.
  • Agent tools: bash, read_file, write_file, edit_file, memory_search, memory_get, web_fetch.
  • Session management with persistence, compaction, search, and export.
  • Image attachment support for multimodal LLMs.
  • Tool approval mode for dangerous operations.
  • Zero-config startup defaulting to claude-cli/opus.
  • Auto-migration from OpenClaw config if present.