Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

OpenFang changelog: what's new each month

Every stable OpenFang release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.

May 2026

7 releases: 0.6.3 → 0.6.9

May brought major orchestration capabilities, dashboard improvements, and provider ecosystem expansions. The month focused on agent lifecycle management, multi-provider routing, and infrastructure hardening.

Action needed
  • Workspace state directory moved: private state now stays under ~/.openfang/, user workspace untouched.

Highlights

  • Agents can now activate peers and clone templates with fresh state via new API endpoints.
  • Reasoning models retain thinking blocks across conversation turns for all three supported drivers.
  • Providers screen redesigned with search, filtering, and category grouping for easier discovery.
  • Telegram and Discord message metadata now surface to agents with stable user identifiers and image captions.
  • WebSocket chat reconnects to your active agent after page refresh, with full session history.

New

  • Agent activation and cloning endpoints let orchestrator agents manage peer lifecycle.
  • Server-assigned message IDs (UUID v4) for tracking and debugging across sessions.
  • Create directory tool with path traversal protection via workspace sandbox.
  • Skill management tools (skilllist, skilldescribe, skillexecute) replace filesystem fallback.
  • TTS and image generation base URLs now configurable for local or custom endpoints.
  • Lenient binding parser tolerates malformed entries without dropping the entire table.
  • LaTeX math rendering in chat messages via KaTeX.
  • Agent discovery from disk: chat tab auto-registers agents in ~/.openfang/agents/ at boot.

Improved

  • Reasoning field support for vLLM 0.19+ while maintaining legacy compatibility.
  • Discord image attachments converted to text captions for non-vision models.
  • Channel ID binding validation with single source of truth and 14 new tests.
  • Subprocess timeout settings hot-reload without restart.
  • Slack Socket Mode deduplication with 60s TTL cache prevents double-processing.
  • Firefox sidebar navigation fixed by replacing x-if + x-transition with x-show + x-cloak.
  • Shift+Enter now inserts newlines in chat input; plain Enter sends.
  • Hand-owned agents no longer error on re-activation after stop.

Fixed

  • Telegram transient rate-limit reactions no longer cached as permanent rejection.
  • OpenRouter free alias now routes to tool-supporting backends.
  • Fish shell config installation no longer breaks on CachyOS and Arch.
  • Workspace state directory properly split between private state and user workspace.
  • Dashboard WebSocket authentication aligned with HTTP middleware.
  • Three rustls-webpki security issues: CRL parsing panic, URI name constraints, wildcard constraint handling.

April 2026

7 releases: 0.5.7 → 0.6.2

April brought major security fixes, stability improvements, and expanded integrations across OpenFang. The month culminated in version 0.6.0 with new delivery targets and scheduling APIs.

Action needed
  • Auth now fail-closed: non-loopback requests rejected with 401 when no API key is configured (loopback still works for local setup)
  • System prompt messages filtered from GET /api/agents/:id/session by default (use ?includesystem=true to opt-in)

Highlights

  • Security fixes for command execution bypasses in allowlist mode and PowerShell wrapper validation.
  • Dashboard configuration now persists to disk, surviving daemon restarts.
  • New delivery targets for schedules: channels, webhooks, local files, and email.
  • Agent context and multimodal message handling improved for real-time data and vision models.
  • Expanded provider support including AWS Bedrock, Gemini embeddings, and self-hosted Revolt instances.

New

  • GitHub Copilot OAuth authentication
  • Russian localization
  • AWS Bedrock provider integration
  • WeCom Stream mode support
  • Session interaction endpoints for dashboard
  • Gemini embedding driver
  • Delivery targets API supporting channels, webhooks, files, and email
  • Agent name prefix option for multi-agent channels

Improved

  • Custom hands now persist across daemon restarts
  • Discord gateway heartbeat detection prevents zombie connections
  • System prompts no longer leak in Web UI by default
  • Global fallback providers are now used at runtime
  • Agent context.md re-read on every turn for live data updates
  • Multimodal messages combine text and images correctly
  • MCP stdio servers receive proper HOME and TMP environment variables
  • Cron job results broadcast to chat WebSocket clients in real time

Fixed

  • Command allowlist bypass in rm and shell wrapper execution
  • Moonshot session repair non-unique tool call ID format
  • WebSocket 404 race condition on agent upgrade
  • Feishu webhook panic paths from poisoned mutex and malformed data
  • TLS panics on Discord, Mattermost, and Feishu connections
  • Schedule tools and HTTP routes now actually execute jobs

March 2026

34 releases: 0.3.5 → 0.5.6

March brought significant stability improvements, security hardening, and expanded provider support to OpenFang. The month focused on fixing edge cases in multi-byte character handling, tool schema compatibility across LLM providers, and authentication workflows.

Action needed
  • apikey config moved from [api] section to root level; auto-migration on startup.
  • Default token quota changed from 1M/hour to unlimited to prevent fresh install quota errors.
  • Hardcoded ~/.openfang paths now respect OPENFANGHOME environment variable.
  • Hand continuous polling interval changed from 60s to 3600s to prevent credit waste.

Highlights

  • Security: Allowlist mode now uses direct argv execution preventing shell injection attacks; all public endpoints require authentication.
  • Tool compatibility: Fixed tool schema validation issues across Gemini, OpenAI, and other providers; added 11 recovery patterns for models without native function calling.
  • Provider expansion: Added 50+ new models including reasoning models (o1, o3, o4), OpenRouter free tier variants, and new providers like NVIDIA NIM and Chutes.ai.
  • Channel reliability: Fixed Telegram HTML sanitization, WhatsApp media handling, Slack threading, and Bluesky notification polling.
  • UTF-8 safety: Resolved 16+ panics from byte-slicing on multi-byte characters affecting CJK text and emoji.

New

  • Hand installation via CLI and API with HAND.toml file support.
  • Image generation pipeline with DALL-E/GPT-Image support.
  • WeCom channel adapter for enterprise messaging.
  • Docker sandbox runtimes for isolated skill execution.
  • Shell skill runtime for command execution.
  • Slack unfurl links support in messages.
  • SSRF allowlist configuration for self-hosted environments.
  • Token quota management with pre-emptive compaction before LLM calls.

Improved

  • Memory recall no longer duplicates injected memories in prompts.
  • Telegram messages now include persistent typing indicators refreshing every 4 seconds.
  • Concurrent bridge dispatch for faster multi-channel message handling.
  • Provider key detection now uses environment variable auto-detection instead of requiring manual configuration.
  • Web wizard displays actual Claude Code CLI installation status dynamically.
  • Model switcher dropdown shows provider:model format for clarity.
  • Config auto-backup to config.toml.bak before any modifications.
  • Embedding auto-detection now probes 6 API providers before falling back to local options.

Fixed

  • Gemini tool schemas now strip unsupported JSON Schema fields (const, format, oneOf, type arrays).
  • OpenAI tool parameters now correctly formatted as JSON objects instead of strings.
  • gzip/deflate/brotli decompression enabled for webfetch preventing finance site failures.
  • Multi-byte UTF-8 character handling fixed across 9 string slicing sites preventing CJK and emoji panics.
  • Telegram forum topics now correctly route replies back to the original thread.
  • Voice messages no longer dropped when sent quickly; concurrent bridge dispatch with per-agent locking.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. OpenFang: pricing, features and alternatives · All changelogs