Skip to main content

Changelog

All notable changes to PicoClaw are documented here.

v0.2.7​

Released: 2026-04-22

Highlights​

  • Launcher auth flow upgrade: Unified dashboard /login / /setup / /logout flow, added OAuth --no-browser login, and migrated launcher auth to password mode (#2339, #2549, #2608)
  • Agent Loop two-phase refactor: Split and reorganized the core loop pipeline to reduce maintenance complexity and prepare for parallel execution capabilities (#2564, #2585)
  • Native Gemini provider: Added native Gemini provider support and separated thought-message handling from output-message handling (#2475)
  • Toolchain and session stability improvements: Added /context command and context usage indicator; /clear now supports Seahorse DB cleanup; scheduled tasks now run in isolated sessions (#2537, #2495, #2474)
  • Web UI UX improvements: Refactored tools page into tabbed Library/Settings views, improved Markdown highlighting, and completed disabled-state reason hints (#2539, #2529, #2523, #2430, #2526)
  • Search and network robustness improvements: Added configurable Sogou search backend and improved proxying, error classification, and fallback behavior (#2524, #2542, #2547, #2517)

Features​

Core & Agent​

  • Loop refactor: Completed Phase 1/2 loop file split and execution-pipeline reordering (#2564, #2585)
  • LLM-as-Judge: Added new evaluation mode (#2484)
  • Parallel execution evolution: Continued enabling parallel-execution scenarios in the agent loop (#2503)

Providers & Auth​

  • Native Gemini protocol support: Native provider plus thought/output separation (#2475)
  • OAuth login UX: Added --no-browser option for headless/no-GUI environments (#2549)
  • Auth compatibility improvements: Disabled token-auth fallback on unsupported platforms; fixed Google Antigravity provider normalization and credential consistency (#2466, #2599)

Channels, Tools & Interaction​

  • Multi-instance channel config evolution: Moved toward a multi-instance configuration model, fixed nested channel_list patch persistence, and improved list editing (#2481, #2530, #2595)
  • Tool call consistency: Fixed cross-turn tool-call ID reuse; kept tool call summaries and assistant output in sync (#2528, #2449)
  • Disabled-state messaging: Expanded composer/tooltip disabled-reason hints and fixed related regressions (#2523, #2430, #2526)

Bug Fixes​

  • Runtime recovery: Fixed MCP/discovery tool recovery after reload; fixed recovery after image-input-unsupported errors (#2489, #2525)
  • Seahorse search safety: Fixed FTS5 MATCH input sanitization issue (#2436)
  • Release/update stability: Fixed retry behavior for transient release-info fetch failures; improved pre-start error logging (#2511, #2414)
  • Config and network edge cases: Fixed allowFrom containing empty-string values; improved network error classification and fallback handling (#2507, #2547)
  • Frontend stability: Fixed config refresh resetting search drafts; cleaned recovered session transcripts and improved chat UI (#2536, #2605)
  • Engineering quality fixes: Fixed govet shadow warnings and adjusted isolation tests on unsupported operating systems (#2613, #2434)

Build & Ops​

  • Android release pipeline: Added Android arm64 cross-compilation and integrated Android bundle publishing into GoReleaser (#2486, #2497)
  • CI workflow updates: Switched to pnpm/action-setup and updated README installation steps (#2512, #2552)
  • Dependency maintenance: Upgraded multiple frontend/backend dependencies (React, TanStack, shadcn, sqlite, MCP SDK, router/lint, and others)
  • Container behavior parity: Changed self-built images to run as root to match release image behavior (#2435)

Documentation​

  • Docs structure reorganization: Reorganized docs by type and added layout guide plus session/routing docs (#2567, #2571)
  • Protocol docs updates: Updated native Gemini protocol docs and cross-provider JSON escaping guidance (#2601, #2420)
  • Localization and misc fixes: Added Korean README; fixed Conventional Commits link in CONTRIBUTING (#2418, #2494)

Full changelog​

v0.2.6​

Released: 2026-04-08

Highlights​

  • Subprocess Isolation: New pkg/isolation runtime sandboxes child processes (exec tool, CLI providers, process hooks, MCP stdio servers) using bwrap on Linux and a restricted token + Job Object on Windows (#2423)
  • Seahorse Short-Term Memory (LCM): New per-agent SQLite-backed memory engine with FTS5 full-text retrieval, hierarchical summarization, and short_grep / short_expand tools (#2285)
  • Enhanced Hooks System: New respond action lets before_tool hooks return tool results directly, enabling plugin-tool injection, result caching, and tool mocking. Comprehensive new upstream protocol docs (#2215)
  • Microsoft Teams Channel: New teams_webhook output-only channel posts Adaptive Cards to Teams via Power Automate workflow webhooks, with multi-target routing and automatic markdown table conversion (#2244)
  • Custom HTTP Headers for Providers: HTTP-based LLM providers now accept a custom_headers field per model entry to inject arbitrary headers into every request (#2402)
  • MCP Artifact Storage: Oversized text results from MCP tools are now persisted as artifacts instead of bloating the context (#2308)

Features​

Channels​

  • Teams Webhook: New output-only channel via Power Automate workflow webhooks; supports multiple named webhook targets selectable per message; renders messages as Adaptive Cards with native Markdown table support (#2244)
  • Feishu: Reply context enrichment now also covers card and file replies, with a 30s message cache and 600-character cap (#2144)

Core & Agent​

  • Subprocess Isolation: New pkg/isolation runtime with enabled and expose_paths config under a top-level isolation block. Linux uses bwrap for filesystem and IPC namespace isolation; Windows uses restricted token + low integrity + Job Object. macOS not yet implemented (#2423)
  • Seahorse Short-Term Memory Engine: SQLite-backed per-agent store at <workspace>/sessions/seahorse.db with FTS5 indexing on summaries and messages; two-tier hierarchical summarization (leaf + condensed); auto-registers short_grep and short_expand tools when active (#2285)
  • Hook respond Action: New HookActionRespond lets a before_tool hook return a HookResult directly, skipping tool execution; comprehensive hook-json-protocol.md and plugin-tool-injection.md upstream specs added; bypasses approve_tool checks so trust hooks accordingly (#2215)
  • Per-candidate Provider for Fallbacks: model_fallbacks now supports independent provider config per candidate (#2143)

Providers​

  • Custom HTTP Headers: New custom_headers field on ModelConfig for HTTP providers — injected into every request, useful for auth proxies, observability headers, vendor-specific routing (#2402)

MCP​

  • Artifact Storage: Oversized text results from MCP tools are persisted to artifact storage and referenced by handle, preventing context bloat (#2308)
  • Isolated Command Transport: MCP stdio servers now go through the unified isolation startup path

Tools & Memory​

  • LOCOMO Membench Tool: New benchmark utility under pkg/membench for evaluating long-conversation memory engines (#2353)
  • write_file: Clarified nested-JSON escape semantics and added tests (#2320)

Web UI​

  • WebSocket URL: Now derived from window.location in the browser instead of being hardcoded by the backend, fixing reverse-proxy and remote access scenarios (#2405)
  • Launcher HTTP Flow: Standard /login / /setup / /logout HTTP endpoints for the dashboard, fixed Windows PID lock for WebSocket (#2339)

Bug Fixes​

  • WebUI: Fixed WebUI being unable to connect to the gateway started by WebUI itself (#2267)
  • Gateway: Hardened PID liveness handling, ownership validation, and WebSocket proxy state (#2403, #2422)
  • Tools: message tool no longer suppresses reply to the originating chat (#2180)
  • Docker: Added -console flag and open network access for launcher (#2314); self-built images now run as root for parity with release images (#2435)
  • CLI: Fixed duplicate v in CLI help banner version line (#2316)
  • Seahorse: Disabled context manager on FreeBSD/ARM and other unsupported platforms (#2417, #2384); corrected BM25 rank semantics in comments (#2360)
  • Tests: Skip TestPrepareCommand_AppliesUserEnv on unsupported operating systems (#2434)

Build & Ops​

  • Dependencies: modernc.org/sqlite bumped from 1.47.0 to 1.48.0 (#2289); github.com/pion/rtp bumped from 1.8.7 to 1.10.1 (#2290)
  • Assets: Updated WeChat QR code image (#2385)
  • Docs: Korean README translation added

Documentation​

This release ships with new docs site coverage for all the major features above:

Full changelog​

v0.2.5​

Released: 2026-04-03

Highlights​

  • VK (VKontakte) Channel: New channel for Russia's largest social network, with Long Poll API, group chat support, and voice capabilities (#2276)
  • LLM Rate Limiting: Built-in rate limiter for provider API calls with configurable limits and automatic fallback (#2198)
  • Pluggable Context Management: New ContextManager abstraction enabling custom context strategies (#2203)
  • New Providers: Venice AI (#2238), Mimo (#1987), LMStudio (#2193) with aligned local provider defaults
  • Enhanced Tools: Exec tool with background execution and PTY support (#1752), reaction tool with reply-aware sends (#2156), load_image for local file vision (#2116), read_file by line range (#1981)
  • Web UI Overhaul: Skill marketplace hub (#2246), dashboard token auth (#1953), first-time tour guide, image messages in Pico chat (#2299), service log level controls (#2227)

Features​

Providers​

  • Venice AI: New Venice AI provider integration (#2238)
  • Mimo: New Mimo provider support (#1987)
  • LMStudio: Local provider with aligned default auth/base handling (#2193)
  • Azure OpenAI: Use OpenAI Responses API for Azure endpoints (#2110)
  • Rate Limiting: Per-model rate limiting with automatic fallback cascade (#2198)
  • User Agent: Configurable userAgent per model in model_list (#2242)
  • Model Availability: Enhanced probing with backoff and caching (#2231)

Channels​

  • VK (VKontakte): Full channel implementation — text, media attachments, voice (ASR/TTS), group triggers, and user allowlists (#2276)
  • Telegram: Quoted reply context and media in inbound turns (#2200); refined duplicate-message protection; fixed HTML links broken by italic regex (#2164)
  • DingTalk: Honor mention-only groups and strip leading mentions (#2054)
  • Feishu: Skip empty random_reaction_emoji entries
  • WeChat: New protocol support (#2106); persist context tokens to disk (#2124)
  • Multi-message: Channels support multi-message sending via split marker (#2008)
  • Channel.Send: Now returns delivered message IDs (#2190)
  • Fail-fast: Gateway fails fast when all channel startups fail (#2262)

Core & Agent​

  • ContextManager: Pluggable context management abstraction for custom strategies (#2203)
  • Token Estimation: Fixed double-counting of system message tokens; added reasoning content guards
  • Context Overflow: Improved detection and classification of context overflow errors (#2016)
  • Light Provider: Use light provider for routed model calls (#2038)
  • Prompt Tokens: Log prompt token usage per request (#2047)
  • Timezone: Load zoneinfo from TZ and ZONEINFO env variables (#2279)

Tools​

  • Exec Tool: Background execution and PTY support (#1752)
  • Reaction Tool: Emoji reactions and reply-aware message sends (#2156)
  • load_image: Load local image files for vision model analysis (#2116)
  • read_file: Read files by line range (#1981)
  • web_search: Date range filter support
  • Cron: Unified agent execution path; publish responses to outbound bus (#2147, #2100)
  • MCP: Support DisableStandaloneSSE for HTTP transport (#2108)

Web UI & Launcher​

  • Skill Marketplace: Browse, search, and install skills from the hub (#2246)
  • Dashboard Auth: Token-protected launcher dashboard with SPA login (#1953)
  • Tour Guide: First-time interactive tour for new users
  • Image Messages: Support image messages in Pico chat (#2299)
  • Log Level Controls: Adjust service log levels from the web UI (#2227)
  • Config Page: Version display moved to header; channel configs load without exposing secrets (#2273, #2277)
  • Dashboard Token: Persisted in launcher config (#2304)

Bug Fixes​

  • Telegram: Fixed edit timeout (#2092); DM policy security hardening (#2088)
  • Config: Fixed FlexibleStringSlice crash on startup (#2078); disabled tool feedback by default (#2026); array placeholder fix
  • Gateway: Fixed reload causing pico channel to stop working (#2082); gateway port check and fatal log recording (#2185)
  • Retry: Proper Retry-After header handling for 429 responses with overflow-safe clamping (#2176)
  • Loop: Fixed polling behavior (#2103)
  • WeChat: Fixed pico token empty when gateway started by launcher (#2241)
  • Web: Hydrate cached Pico token for websocket proxy (#2222); skills page dark mode theme colors (#2166); Discord token persist from channel settings (#2024)
  • Container: Graceful shutdown on SIGINT/SIGTERM
  • BM25: Precomputed index for repeated searches (#2177)
  • Panic Recovery: Unified all panic events to panic log file (#2250); added missing recover in subturn (#2253)

Build & Ops​

  • Windows: Fixed make build error, support custom build env (#2281)
  • Logging: PICOCLAW_LOG_FILE env var for file-only logging; component-based logger with highlighted output; default log level changed to warn
  • Config: Refactored config and security structure for simplification (#2068); ModelConfig.Enabled field; placeholder text supports string or list
  • Self-update: Robust selection and extraction with nightly default (#2201)
  • Security: Open-by-default warning and * allow_from wildcard support

Full changelog​

v0.2.4​

Released: 2026-03-25

Highlights​

  • Largest Update Ever: 539 files changed, ~86,000 lines of code added, surpassing 26K Stars
  • Agent Architecture Overhaul: Turn & Sub-turn lifecycle, EventBus, Hook system, Steering dynamic intervention
  • Deep WeChat/WeCom Integration: Enhanced message processing and context management
  • Security System Upgrade: Comprehensive security configuration and permission management
  • New Providers & Channels: Extensive new model and channel support

Agent Architecture Refactor (Phase 1)​

  • Turn & Sub-turn Lifecycle: Turn as atomic unit for context compression and session management, supporting Sub-turn with max concurrency 5 and max nesting depth 3, supporting evaluator-optimizer mode
  • EventBus: 18 event types defined, covering full Agent lifecycle including conversation turns, tool calls, LLM requests/responses, sub-task creation/termination, Steering injection, context compression
  • Hook System: Supports Observer, Interceptor, and Approval hook types with 5 checkpoints (before_llm, after_llm, before_tool, after_tool, approve_tool), supporting both in-process and external process hooks (stdio/gRPC, Python/Node.js compatible)
  • Steering Dynamic Intervention: Inject new instructions into running Agent during tool call gaps
  • Context Budget Management: Proactively check Token consumption before each LLM call, automatically compress history when approaching limits while preserving conversation coherence based on Turn boundary detection
  • AGENT.md Structured Definition: Define Agent identity, capabilities, and personality through AGENT.md file with YAML frontmatter

Channels & Integration​

  • Deep WeChat/WeCom Integration: Enhanced message processing and context management capabilities
  • Security System Upgrade: Comprehensive security configuration with optimized permission management
  • New Providers & Channels: Expanded Provider and Channel options

Full changelog​

v0.2.3​

Released: 2026-03-17

Highlights​

  • System Tray UI: Desktop tray support across all platforms (macOS, Linux, FreeBSD)
  • Exec Controls: Configurable exec settings with cron command gating
  • Web Gateway: Hot reload and polling state sync for gateway management
  • SpawnStatusTool: New tool for reporting subagent statuses

Features​

  • Configurable cron command execution settings exposed in the web UI
  • WebSocket proxied through the web server port for unified ingress
  • Refactored gateway helpers with server.pid health check

Bug Fixes​

  • GLM: Fixed nil input handling in tool_use blocks for GLM provider
  • Gateway: No longer auto-starts when the server is not running
  • Workspace: Normalized whitelist path checks for symlinked allowed roots

Full changelog​

v0.2.2​

Released: 2026-03-11

Highlights​

  • Voice Transcription: Echo voice audio transcription for Discord, Slack, and Telegram
  • Agent Management UI: New web UI for agent management and launcher integration
  • Security Hardening: Hardened unauthenticated tool-exec paths

Features​

  • Exec allow_remote config support in the web settings page
  • Session key sanitization for slash characters in forum topic keys
  • Refactored skill loader markdown metadata parsing

Bug Fixes​

  • Gateway: Fixed gateway binary path resolution and --config flag passing
  • Migration: Skip meta JSON files during session migration
  • Slack: Fixed double messages in threads

Full changelog​

v0.2.1​

Released: 2026-03-09

Highlights​

  • Web & UI Architecture Upgrade: Launcher migrated to a modular web frontend/backend architecture, greatly improving management UX with a brand-new launcher banner (#1275, #1008).
  • Vision Pipeline (Multimodal): Full image/media support in the Agent core — media references are converted to Base64 streams with auto file type detection, enabling multimodal LLM vision capabilities.
  • JSONL Memory Storage Engine: Append-only JSONL session storage with physical-level history compression and crash recovery for safer session persistence.
  • MCP Deep Integration: Full Model Context Protocol (MCP) support, including external MCP Server connectivity and parallel LLM tool call execution (#1070).
  • New Providers & Search Sources: Added Minimax, Moonshot (Kimi), Exa AI, SearXNG, Avian, and more.

Features​

Providers​

  • New LLMs: Minimax (#1273), Kimi (Moonshot) & Opencode, Avian, LiteLLM alias support (#930).
  • Advanced Model Features: Anthropic extended thinking stream (#1076), Anthropic OAuth setup-token login (#926).
  • New Search Sources: SearXNG, Exa AI, GLM Search Web (#1057).
  • Model Routing: Language-agnostic model complexity scoring for smarter model dispatch in the Agent Loop.

Channels​

  • WeCom (Enterprise WeChat): New WeCom AIBot integration with streaming task context, reasoning channel display, timeout and cleanup enhancements.
  • Discord: Proxy support, channel reference parsing, message link expansion, quote/reply context understanding (#1047).
  • Telegram: Custom Bot API Server (base_url) support (#1021), command scope optimization, long message chunking.
  • Feishu (Lark): Enhanced card (Markdown), media, @mention, editing; random reaction emoji (#1171).
  • New Channels: Native Matrix and IRC channel support.

Core & Agent​

  • Multimodal: Media field in Message structure, resolveMediaRefs for end-to-end multimodal pipeline.
  • Memory & Context: Reasoning content retention in multi-turn history; configurable summary threshold and token ratio (#1029).
  • System & Config: .env environment variable loading and config override; PICOCLAW_HOME for custom config/workspace path; /clear command (#1266).
  • Interaction: Agent fallback to reasoning content output (#992).

Tools & Skills​

  • MCP Architecture: Workspace-level env file loading, relative path resolution, full MCP registration and lifecycle management.
  • Tool Control: Global enable/disable toggle for tools (#1071).
  • File Sending: New send_file tool for outbound multimedia file sending (#1156).

Bug Fixes​

Stability & Concurrency​

  • Fixed MCP TOCTOU race conditions and resource leaks; deadlock during tool registration; aggregated errors on connection failure.
  • JSONL migration crash consistency (fsync hardening, safe metadata writes); memory lock limits and scanner buffer increase to prevent OOM.
  • WeCom app/bot dedupe data race (switched to ring queue); Discord cross-guild message link leaks; Matrix group mention regex.

Tools & Network​

  • Fixed web_fetch max payload limit issue; parallel execution request interrupt and response body leak (#940).
  • Fixed ClawHub 429 retry; skill discovery registry flag (#929).
  • Fixed exec tool shell command timeout config; command injection interception (kill -9 pattern filter).

Experience & Compatibility​

  • Telegram HTML parsing: degradation and multi-queue re-chunking for oversized content.
  • OpenAI compatibility: improved non-JSON response (HTML error page) handling and streaming parsing.
  • Fixed hot-reload self-kill protection false positive.

Build, Deploy & Ops​

  • Docker: Base image switched from Debian to Alpine; Docker Compose v2 syntax migration; MCP-ready Docker image config.
  • Cross-compilation: Added Linux/s390x and Linux/mipsle (#1265); macOS binary notarization via goreleaser (#1274).
  • CI/CD: Nightly build workflow (#1226); govulncheck and dupl checks; auto-upload to Volcengine TOS (#1164).
  • Observability: Cron job lifecycle logging (#1185).

v0.2.0​

Released: 2026-02-28

New Features​

  • Architecture Overhaul:
    • Protocol-based Providers: Introduced a protocol-based provider creation mechanism, unifying integration logic for OpenAI, Anthropic, Gemini, Mistral, and other major protocols.
    • Channel System Upgrade: Rebuilt channel architecture with automated placeholder scheduling, typing indicators, and emoji reactions.
  • Ecosystem & Toolchain:
    • PicoClaw Launcher & WebUI: New web interface for visual configuration management and gateway process scheduling.
    • Skill Discovery: ClawHub integration for online skill searching, caching, and installation.
  • New Providers:
    • Cerebras, Mistral AI, Volcengine (Doubao), Perplexity, Qwen (Tongyi Qianwen), and Google Antigravity native support.
    • Reasoning model thought/reasoning content output with dedicated channel display.
  • New Channels:
    • Native WhatsApp: High-performance native WhatsApp channel implementation.
    • WeCom: Enterprise WeChat and self-built WeCom application channels.

Optimizations​

  • Low-resource Adaptation: Optimized binary size with comprehensive ARM build support (ARMv6, ARMv7, ARMv8.1).
  • Concurrency: Refactored MessageBus to resolve deadlocks; message buffer increased from 16 to 64.
  • Search Enhancement: Tavily search proxy support; new OpenAI-compatible WebSearch interface.
  • Robust Configuration: model_list templates for zero-code model addition with duplicate model name validation.

Bug Fixes & Security​

  • Sandbox Security: Fixed ExecTool working directory escape risk and TOCTOU race conditions; hardened system instruction injection defense.
  • Stability:
    • Fixed memory leaks (particularly in WhatsApp channel).
    • Fixed Gemini API 400 errors and prompt cache invalidation issues.
  • Code Quality: Enabled golangci-lint across the project; cleaned up redundant code and lint errors.

Documentation​

  • Multi-language: Added Vietnamese and Portuguese (Brazil) README translations; improved Japanese translation.
  • Architecture Guides: New detailed Chinese and English documentation on channel and protocol architecture.

v0.1.2​

Released: 2026-02-17

New Features​

  • LINE Channel: Added LINE Official Account channel support.
  • OneBot Channel: Added OneBot protocol channel support.
  • Health Check: Added /health and /ready endpoints for container orchestration (#104).
  • DuckDuckGo Search: Added DuckDuckGo search fallback with configurable search providers.
  • Device Hotplug: USB device hotplug event notifications on Linux (#158).
  • Telegram Commands: Structured command handling with dedicated command service and telegohandler integration (#164).
  • Local AI: Added Ollama integration for local inference (#226).
  • Skill Validation: Added validation for skill info and test cases (#231).

Improvements​

  • Security: Blocked critical symlink workspace escape (#188); tightened file permissions and enforced Slack ACL checks (#186).
  • Docker: Added curl to Docker image; improved Docker build workflow triggers.
  • 32-bit Support: Added build constraints for Feishu to support 32-bit architectures.
  • Loong64: Added Linux/loong64 build support (#272).
  • GoReleaser: Migrated to GoReleaser for release management (#180).

Bug Fixes​

  • Fixed duplicate Telegram message sending (#105).
  • Fixed code block index extraction bug (#130).
  • Fixed OneBot connection handling.
  • Prevented panic on MessageBus publish after close (#223).
  • Removed duplicate file extension in DownloadFile (#230).
  • Fixed OpenAI OAuth authorize URL and params (#151).

v0.1.1​

Released: 2026-02-13

New Features​

  • Dynamic Context Compression: Intelligent conversation context compression (#3).
  • New Channels: Feishu (#6), QQ (#5), DingTalk Stream Mode (#12), Slack Socket Mode (#34).
  • Agent Memory System: Agent memory and tool execution improvements (#14).
  • Cron Tool: Scheduled task support with direct shell command execution (#23, #74).
  • OAuth Login: SDK-based subscription provider OAuth authentication (#32).
  • Migration Tool: picoclaw migrate command for OpenClaw workspace migration (#33).
  • Telegram Upgrade: Migrated to Telego library for better Telegram integration (#40).
  • CLI Provider: CLI-based LLM provider for subprocess integration (#73, #80).
  • Provider Selection: Explicit provider field support for model configuration (#48).
  • macOS Support: Added Darwin ARM64 build target (#76).

Improvements​

  • Consolidated media handling and improved resource cleanup (#49).
  • Better version info with git commit hash (#45).
  • Workspace directory boundary enforcement for system tools (#26).

Bug Fixes​

  • Fixed heartbeat service startup bug (#64).
  • Fixed LLM error from CONSCIOUSLY message history cleanup (#55).
  • Fixed OpenAI device-code flow string interval (#56).
  • Fixed Telegram channel permission check (#51).
  • Fixed atomic safety for AgentLoop running state (#30).

v0.0.1​

Released: 2026-02-09

Initial release of PicoClaw. Please refer to the README for basic usage instructions.