> For the complete documentation index, see [llms.txt](https://docs.dos.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.dos.ai/changelog.md).

# Changelog

All notable changes to the DOS.AI platform are documented here. This file is the source of truth - synced to `dosai.changelog` on deploy.

Format: `- **type** [product] Title - Description` Types: `feature`, `fix`, `improvement`, `breaking` Products: `dosclaw`, `dashboard`, `gateway`, `dosafe`, `inference`

***

## 2026-06-16

* **feature** \[dosclaw] Multi-Platform Commerce Tools - One agent can now serve more than one storefront at once (e.g. KiotViet POS + WooCommerce); product lookups aggregate across every connected platform with platform-qualified codes (`woocommerce:11`, `kiotviet:DH000002`), and stock/order/create-order tools route to the right platform automatically; single-platform agents are unchanged

## 2026-06-15

* **feature** \[gateway] Response Caching - Exact-match response cache for non-streaming requests; identical requests return instantly with no model call and no token charge. On by default for `temperature: 0`; control per request with the `X-DOS-Cache: on|off` header; cache hits return `X-DOS-Cache: hit` / `X-Provider: cache`
* **feature** \[gateway] Guardrails (PII) - Opt-in deterministic PII detection via the `X-DOS-Guardrails: block|redact|flag` header; detects email, phone, and Luhn-validated card numbers in chat/completions input
* **feature** \[gateway] Batch API - OpenAI-compatible `/v1/files` + `/v1/batches` for async batch processing of `dos-ai` chat completions, processed on idle GPU capacity
* **feature** \[dosclaw] Commerce Connectors - Connect commerce platforms (WooCommerce and KiotViet) to an agent directly from the dashboard; the bot gains in-chat tools to search products, check stock, look up order status, and create orders against the live store
* **feature** \[dosafe] C2PA Provenance + Integrity Clash - Image AI detection now parses real C2PA manifests; flags an "Integrity Clash" when a manifest claims camera origin but the forensic ensemble independently reports the pixels as AI, instead of being laundered past the detector by forged camera credentials
* **improvement** \[dosafe] In-the-Wild Detection Calibration - Recalibrated the standalone image `is_ai_generated` threshold against an in-the-wild benchmark to catch more real-world AI images

## 2026-06-13

* **feature** \[dosclaw] Scope Guard - Business-bounded bots (customer-service, shop, sales, real-estate, finance) now politely decline off-topic general-AI requests (coding, homework, essays, translation, chit-chat) with one short redirect instead of acting as a free ChatGPT; education-tutor and content-creator templates are exempt
* **fix** \[dosclaw] Reasoning Leak on Bank Reset - Serial-less device-reset requests in bank-officer groups no longer leak the model's raw chain-of-thought into the customer reply; the bot now asks for the serial in one short line

## 2026-06-12

* **feature** \[dashboard] Agents Funding Summary - Billing Overview now shows an Agents card: total agents, how many are included in the plan vs billed from credits (\~$1/day each), per-agent status, and a "deleted in Xd" countdown for suspended agents
* **feature** \[dashboard] Subscription Cancel Flow - Billing overview shows the current plan and adds a cancel flow; pay-by-credits wording clarified
* **fix** \[dosclaw] Device Reset Messaging - Tingee loa reset now correctly tells the customer the unlink is immediate (no restart, no waiting) and that the device must be re-linked to use again; actionable failure reasons are relayed instead of a generic system error
* **fix** \[dosafe] QR Decode Accuracy - Swapped the QR decoder (goqr → gozxing/ZXing) after the old library silently mis-decoded a real device QR into the wrong serial

## 2026-06-11

* **feature** \[dosafe] Image Source-URL Analysis - Image AI detection now classifies the source URL (domain reputation + filename/path AI markers like `ai-generated`, `midjourney`, `dall-e`) and feeds it to the meta-judge, catching AI images that the pixel ensemble alone scores as borderline
* **feature** \[dosclaw] No Invented Links - KB grounding now forbids agents from fabricating third-party links in answers
* **feature** \[gateway] Billing Drift Alerts - Real-time DOSiren alert when a paid webhook fails to grant credit, plus a config-drift canary on `/system/health` to catch silent billing-config mismatches

## 2026-06-10

* **feature** \[dosclaw] Cleaner Web-to-KB Ingestion - Ingesting a web URL into an agent's knowledge base now packs chunks, strips markup noise and empty heading-anchor links, and understands GitBook `llms-full.txt` for higher-quality retrieval

## 2026-06-09

* **feature** \[dashboard] Pricing v2 - New plan tiers (Plus $9 / Pro $19) with a 1-day free trial plus a 4-hour countdown, per-billing-cycle plan credits, and free `dos-ai` model usage; existing trials are grandfathered so the shorter trial is not applied retroactively
* **fix** \[gateway] Stripe Subscription Reference - Read the subscription reference from `invoice.parent` to match Stripe's API so subscription credit grants resolve correctly

## 2026-06-08

* **feature** \[dosclaw] Free dos-ai for Agents - DOSClaw agent traffic now runs the `dos-ai` model for free, with per-tier rate limits and a request queue (10s timeout) when the concurrency limit is hit
* **feature** \[dosclaw] Web-Link Knowledge Sources - Agents can ingest web URLs into their knowledge base (with a refresh cron) and manage those web sources from a new dashboard UI
* **feature** \[dosclaw] KB-Only Mode - Per-agent `kb_only` flag denies `web_fetch`/`web_search` so a bot answers strictly from its own knowledge base

## 2026-06-06

* **feature** \[gateway] Provider Marketplace (Phase A/B) - Catalog of upstream providers and a registry-based routing layer (flag-gated, default off) toward a multi-provider LLM marketplace
* **feature** \[dashboard] OIDC Login (app.dos.ai) - Migrated app sign-in to an OIDC authorization-code + PKCE flow; logout now ends the IdP session via SLO so the next sign-in shows the login form instead of silently re-authenticating
* **fix** \[dosclaw] Workspace-Only File Tools - OpenClaw template now enforces `tools.fs.workspaceOnly=true`, restricting agent file operations to the workspace
* **fix** \[gateway] PII Masking in Memory - Vietnamese phone numbers are masked before being written to durable agent memory

## 2026-06-04

* **feature** \[dosclaw] DOS Grounding Plugin - New first-party OpenClaw plugin injects KB context before the prompt is built, giving deterministic knowledge-base grounding across all gateway channels (not just Zalo); blocks fabricated order lookups

## 2026-06-03

* **feature** \[dosclaw] Self-Serve Zalo Personal (zalouser) - Connect a personal Zalo account to an agent via QR code from the dashboard
* **feature** \[dosclaw] Cross-Channel Customer Identity - Per-customer memory is now keyed by phone number across channels, with Telegram `request_contact` to link a customer's identity across Zalo/Telegram
* **feature** \[dosclaw] Device Reset Tool - `reset_device` MCP tool for Tingee loa reset, wired per-agent and gated to designated groups
* **fix** \[dosclaw] Stronger Knowledge Grounding - `search_knowledge` now returns an explicit instruction to answer only from results (curbs hallucination) with a broader-triggering tool description

## 2026-06-02

* **feature** \[dashboard] User Timezone + Zalo OA Schedule - User timezone setting with timezone-aware Zalo OA auto-active scheduling
* **feature** \[dosclaw] Zalo OA Profile Context - Injects the Zalo OA follower's profile into the agent context and strips OA-incompatible markdown; server-side per-customer memory auto-capture and recall on Zalo personal, gated to 1:1 DMs to prevent group leakage
* **fix** \[dosclaw] Zalo OA Reply Hygiene - Hard-strips CJK characters from Zalo OA replies and unwraps code-wrapped replies

## 2026-06-01

* **feature** \[dosclaw] Per-Customer Memory (All Channels) - Agents gain per-customer memory across every channel via MCP tools (`search_memory` + `remember`), so a bot remembers a customer's stated facts between conversations
* **feature** \[dashboard] Self-Service Zalo OA - Connect and manage a Zalo Official Account integration, including an auto-active schedule editor, from the dashboard
* **improvement** \[gateway] Reasoning-Aware Timeouts - Non-streaming timeout now accounts for reasoning tokens (ported from DOSRouter) so long reasoning turns are not cut off

## 2026-05-31

* **feature** \[dosclaw] Zalo OA Scheduled Auto-Active - Zalo OA bots can open on a scheduled window with no `/start` needed
* **fix** \[dosclaw] Memory Capture Accuracy - Memory now captures only customer-stated facts, never the assistant's own claims
* **fix** \[dosclaw] Version Floor - Enforces an OpenClaw image version floor (currently 2026.5.12) so agents can't be pinned to an image version that crash-loops

## 2026-05-30

* **feature** \[dosclaw] Multi-OA Zalo Official Account Channel - Multi-tenant Zalo OA channel with stateless OAuth, webhook verification, activation gating, and a 24h idle sliding-window expiry
* **feature** \[gateway] 3-Tier Detector Fallback - Detection requests fall back across Lambda → local → Cloud Run for resilience
* **improvement** \[gateway] Self-Host-Primary Embeddings - Embeddings serve from self-hosted vLLM first with a cloud fallback

## 2026-05-27

* **feature** \[dosclaw] Google Sheet → Knowledge Base Sync - Sync a Google Sheet into an agent's knowledge base, resolving FAQ columns by header name (diacritic-insensitive) on a recurring cron
* **improvement** \[gateway] On-Demand Detector GPU - Pre-warms the GPU detector only while the primary detector is down, avoiding idle GPU billing

## 2026-05-25

* **feature** \[dosclaw] Agent Location Selection - Choose an agent's hosting location on create, with project-scoped grants/RBAC and a per-backend agent cap
* **feature** \[dosclaw] Human Handoff (Multi-Target) - `handoff_to_human` MCP tool hands a conversation off to a human, fanning out to multiple targets (Zalo + Telegram) and attaching the customer's Zalo group invite link
* **feature** \[dosclaw] Chat-Save to Knowledge - Endpoint to save knowledge from chat with peer-owner auth; SOUL/IDENTITY/AGENTS.md and knowledge files persist across the container lifecycle
* **fix** \[dosclaw] Agents Run 24/7 - Local agents stay running by default; idle-sleep is now opt-in

## 2026-05-24

* **feature** \[gateway] Unified DOSafe API Key - DOSafe routes now accept the unified `dos_sk_` key
* **feature** \[dashboard] DOSafe Usage in Dashboard - View DOSafe API usage from the dashboard

## 2026-05-23

* **feature** \[dosclaw] Custom SOUL/Identity + Knowledge - Per-agent custom SOUL/identity with hybrid-by-size knowledge (small knowledge inlined, larger sets via RAG search)
* **feature** \[dosclaw] Image Anti-Fabrication Rule - All agent SOUL templates get an image anti-fabrication rule so bots don't invent image contents

## 2026-05-22

* **feature** \[dosclaw] RAG Knowledge Base (Phase 1) - Knowledge base with chunking, embedding, and kNN search; uploads replace files in place on re-upload
* **feature** \[dosafe] Partner Server Deployment Package - On-prem/partner server deployment package and backend config for running the stack at a partner site

## 2026-05-20

* **feature** \[dosclaw] Bring-Your-Own-Subscription Plugins - Per-agent plugin slots plus a model-auth status endpoint for agents using their own provider subscription
* **feature** \[gateway] Billing Reconciliation - Billing reconciliation cron with an unbilled-usage status and Telegram alerting on unbilled fallback

## 2026-05-19

* **feature** \[dashboard] Connect ChatGPT (BYOS) - Connect your own ChatGPT/OpenAI account to an agent via device-code OAuth, with a per-agent Model Auth status card
* **fix** \[gateway] Emergency Paid Fallback - Falls back to a paid provider with a TTFB timeout when self-hosted vLLM is down

## 2026-05-16

* **feature** \[dosclaw] FinOne Merchant Pairing - Bind and verify a merchant's FinOne userId during agent provisioning; the pairing survives container rebuild/upgrade
* **feature** \[dosafe] BlazeFace Liveness Pipeline - Face verifier swaps to MediaPipe BlazeFace (Apache 2.0) for detection and the CVPR 2024-winning Swin-V2 liveness model

## 2026-05-14

* **feature** \[dosclaw] VN Merchant Accounting Agent - New `finance-assistant` template and a public FinOne skill giving VN merchants a bot with accounting MCP tools (transactions, budgets, goals, recurring, insights, education)

## 2026-05-13

* **feature** \[dosclaw] Google Workspace OAuth in Agents - Bundles Google Workspace OAuth into agent containers so bots can act on Gmail/Calendar/Drive/Sheets
* **feature** \[dosafe] Plagiarism Detection - `/detect-plagiarism` wires corpus search + upsert for plagiarism checking
* **feature** \[dosafe] Vietnamese AI-Text Detector (Pilot) - Pilots a Vietnamese AI-text detector (mE5-VN, CC BY 4.0)
* **feature** \[dosafe] Partner Webhooks - Risk assessments fire events to B2B partner webhooks

## 2026-05-12

* **improvement** \[dosafe] Speaker Verification Upgrade - Swaps the speaker-verification ensemble to ERes2NetV2 (Apache 2.0)

## 2026-05-10

* **feature** \[dosafe] MamBo-3 Voice Antispoof - New MamBo-3 voice anti-spoof backend (MIT license), replacing the prior in-house model
* **fix** \[gateway] Encrypt Slack Tokens - Slack workspace bot tokens are now encrypted at rest

## 2026-05-08

* **feature** \[gateway] Prometheus Metrics - `/metrics` endpoint with hot-path counters and a structured-logging foundation for the gateway

## 2026-05-04

* **fix** \[gateway] Streaming Billing Hardening - Paid streaming requests hold against the actual provider, use a bounded SSE reader, and split PayOS/VNPay pending-vs-settled idempotency keys to prevent double-charges
* **fix** \[gateway] Webhook PII Redaction - Webhook body logs are redacted to stop leaking PII, and webhook-secret defaults are removed (fail-closed on missing secret)

## 2026-05-03

* **feature** \[gateway] Per-User JWT Rate Limit - Per-user token-bucket rate limiting on the JWT auth path

## 2026-04-30

* **feature** \[gateway] Fleet Upgrade Safety - Post-incident upgrade safety controls and a `/system/agents/fleet-upgrade` admin endpoint with a job-level mutex to prevent simultaneous fleet upgrades

## 2026-04-29

* **feature** \[gateway] Multi-Tenant Gateway (Phase 1.5) - Consolidates per-product routing and quota into the gateway

## 2026-04-27

* **feature** \[gateway] Supabase JWT Auth + Per-Product Quota - Supabase JWT auth (ES256/JWKS verified) with per-product monthly quota, a `/v1/p/<product>/` route prefix, and per-product audit logging

## 2026-04-26

* **feature** \[dosafe] Document Forgery Detection - `/v1/dosafe/detect-doc-forgery` route backed by a multi-layer ensemble (Qwen3.6 + classical forensics + rules)
* **feature** \[dosafe] Multimodal Identity Verify - `/v1/dosafe/identity/verify-multimodal` endpoint with active-liveness challenge nonce, plus a multi-frame anti-spoof voting endpoint
* **feature** \[dosclaw] Skills Installed State - `/v1/agents/:id/skills/installed` endpoint with cross-device hydrate and an Installed tab in the dashboard
* **improvement** \[dosclaw] OpenClaw Changelog Auto-Sync - OpenClaw GitHub releases auto-sync into the changelog every 6h

## 2026-04-24

* **fix** \[dashboard] Cross-Subdomain Logout - Session cookie set on `.dos.ai` for all subdomains, with duplicate `Set-Cookie` clears and a logout layout-race fix

## 2026-04-22

* **feature** \[inference] dos-ai → Qwen3.6-35B-A3B (NVFP4) - Upgraded the self-hosted `dos-ai` model to Qwen3.6-35B-A3B in NVFP4 on vLLM
* **feature** \[dosafe] Detector v2 Consolidation - Merged image-detector, face anti-spoof, and voice anti-spoof into one consolidated detector service

## 2026-04-21

* **feature** \[dashboard] Onboarding Email Drip - Wires a 5-email onboarding drip sequence via Brevo templates

## 2026-04-20

* **feature** \[dosafe] 1:N Speaker Identification - Voice ID gains a 1:N speaker-identification endpoint returning top-K candidates

## 2026-04-18

* **feature** \[dosafe] Face Anti-Spoof v2 - New face anti-spoof v2 service (beta flag)
* **feature** \[gateway] Agent Suspension Lifecycle - Flag-gated agent suspension + purge lifecycle; closes a self-credit exploit with atomic credit RPCs
* **feature** \[dashboard] Playground - Interactive playground chat UI with streaming

## 2026-04-17

* **feature** \[dosafe] Compliance Call Streaming - SSE streaming for progressive call-compliance results, with speaker diarization producing a dialogue-formatted transcript and anti-spoof running independently of enrollment

## 2026-04-15

* **feature** \[dashboard] Transactional Email (Brevo) - Brevo integration for transactional emails
* **fix** \[dashboard] SSO Login Reliability - Reworked the SSO handler (hash-based token flow, direct JWT decode, redirect-loop break, login-spinner timeouts) to fix sign-in

## 2026-04-14

* **feature** \[dashboard] Agent Skills Catalog - New OpenClaw skills catalog tab in the agent dashboard with skill detail views (SKILL.md content, stats, Install button) and a redesigned card-grid agents list

## 2026-04-13

* **feature** \[dosafe] Alibaba Cloud eKYC Fallback - Adds an Alibaba Cloud eKYC fallback for face verification

## 2026-04-12

* **feature** \[gateway] Cache-Aware Sticky Routing - DOSRouter pins model to session when context exceeds 3K tokens (single message) or 5K tokens (cumulative) to maximize provider-side prefix cache hits; sticky TTL is per-provider (5min for API providers, 10min for self-hosted vLLM)
* **feature** \[gateway] Per-Provider Cache TTL - Sticky routing TTL matches each provider's prefix cache lifetime: Anthropic/OpenAI/DeepSeek (5 min), vLLM/self-hosted (10 min); configurable via `providerCacheTTLMs` map
* **fix** \[dashboard] Cross-Account Logout Loop - Logout now passes `prompt=login` to id.dos.me to force login form display instead of auto-SSO, preventing cross-account session loops

## 2026-04-11

* **feature** \[gateway] DOSRouter Upstream Sync to v0.12.146 - 17/19 ClawRouter releases ported; includes usage cost breakdown, eco/premium tier fallback, session pinning, agentic 3-state, model roster updates
* **feature** \[gateway] DOSRouter Full Port Expansion - Wallet module (EVM + Solana), payment module (x402 protocol), image generation endpoint, full CLI (serve, classify, models, stats, logs, cache, report, wallet, chain, doctor)
* **feature** \[gateway] DOSRouter Open-Sourced - Standalone Go LLM router at github.com/DOS/DOSRouter with 15-dimension scoring, tier-based routing, structured fallback chains

## 2026-04-08

* **feature** \[dosclaw] OpenClaw v2026.4.5 - Major engine upgrade with video/music generation, enhanced memory, and improved channel experience
* **feature** \[dosclaw] Video Generation - Agents can create videos using the built-in `video_generate` tool with xAI Grok, Alibaba Wan, and Runway providers
* **feature** \[dosclaw] Music Generation - Built-in `music_generate` tool with Google Lyria and MiniMax providers; async task tracking and follow-up delivery
* **feature** \[dosclaw] ClawHub Skill Store - Search, browse, and install skills directly from the Control UI Skills panel
* **feature** \[dosclaw] Memory Dreaming - Agents automatically distill important conversations into long-term memory via background dreaming phases; Dream Diary surface in the UI
* **feature** \[dosclaw] Thinking Level Control - Per-session thinking depth picker in chat header; choose how much reasoning the agent applies per conversation
* **feature** \[dosclaw] Structured Progress - Long-running agent tasks now show step-by-step progress updates in compatible UIs
* **improvement** \[dosclaw] Multilingual Control UI - Added 12 new locales: Chinese (Simplified/Traditional), Portuguese, German, Spanish, Japanese, Korean, French, Turkish, Indonesian, Polish, and Ukrainian
* **improvement** \[dosclaw] Prompt Caching - Smarter cache reuse across follow-up messages for faster and more cost-efficient conversations
* **improvement** \[dosclaw] Cleaner Replies - Internal tool tags and planning text no longer leak into user-visible messages
* **improvement** \[dosclaw] Discord Media Limit - Inbound/outbound media cap raised to 100MB, matching Telegram
* **fix** \[dosclaw] Telegram - Voice note transcription restored in DMs; model picker fixed; topic replies; reaction persistence across restarts; better image handling
* **fix** \[dosclaw] Discord - Image generation replies now include actual images; reply threading fixed; voice auto-join more reliable
* **fix** \[dosclaw] WhatsApp - Reconnect loop fixed for quiet chats
* **fix** \[dosclaw] Group Chat - Replies now use natural chat-style formatting instead of document-style spacing
* **breaking** \[dosclaw] Config Cleanup - Legacy config aliases removed; existing configs auto-migrate via `openclaw doctor --fix`

## 2026-04-07

* **feature** \[gateway] Streaming Pre-Deduct - Paid provider streaming requests now hold estimated cost before streaming via atomic `hold_credits()` RPC; reconciled on completion or kept on interruption
* **feature** \[dosclaw] Agent Metrics Rewrite - Metrics tab shows billing-based usage (spend, requests, tokens, daily charts) instead of slow openclaw exec; load time reduced from \~10s to \~2s
* **feature** \[dashboard] Usage Cost Tracking - Total Spend, Daily Spend chart, and Avg Cost/Request added to the Usage page
* **feature** \[dosclaw] Agent Usage Self-Lookup - New `GET /v1/agents/:id/usage/summary` endpoint; bots can answer "how much did I cost?" via SOUL.md curl instruction
* **improvement** \[gateway] Atomic Hold RPCs - `hold_credits()` and `release_hold()` PostgreSQL functions for race-free balance holds during streaming
* **improvement** \[dosclaw] Container Env Vars - `AGENT_ID` and `AGENT_SLUG` now injected into all agent containers for self-identification

## 2026-04-04

* **feature** \[dosclaw] Shared Bot Multi-Agent Routing - Messages from shared Telegram/Discord bots now route through the correct agent's OpenClaw container based on chat ID; agent name/ID injected for personality routing
* **improvement** \[gateway] Alert Backend Labels - All container and Status API alerts now include source (LOCAL/AZURE); vLLM health monitoring every 3 minutes via Cloudflare tunnel URLs; recovery alerts bypass dedup cooldown
* **fix** \[dosclaw] Agent Version Display - Agent settings now always show current running version with fallback to "latest" instead of blank

## 2026-04-03

* **feature** \[dosclaw] OpenClaw v2026.4.2 - Agents pinned to v2026.4.2 with SearXNG web search plugin enabled; version selector filters to multi-arch tags only
* **feature** \[dosclaw] Agent Usage Billing - Per-agent usage billing endpoints; agents now track and report token consumption
* **feature** \[dashboard] Crypto Payment - Stripe crypto checkout alongside cards; Google Pay, Apple Pay, and Link payment methods supported
* **fix** \[gateway] Billing & Provider Routing - Retail pricing uses customer-requested model identity; cloud-only models fail honestly instead of falling back to vLLM; promo pricing cost floor during upstream billing
* **fix** \[dashboard] Billing Page - SVG brand icons for payment methods, Link display fix, success toast; fixed fetchBillingData crash from missing brand/last4 fields
* **fix** \[gateway] vLLM Health Check URLs - Cloud Run health checks use configured Cloudflare tunnel URLs instead of localhost (unreachable from Cloud Run)

## 2026-04-02

* **feature** \[dosclaw] Agent Memory Search - Shared Qwen3-Embedding-4B AWQ embedding service; agents now perform semantic memory search via local vLLM instead of keyword-only recall
* **feature** \[dosclaw] SearXNG Web Search - Self-hosted SearXNG enabled for `web_search` tool in agent containers; no external API key required
* **fix** \[dosafe] Audio Speech Detection - Replaced spectral band analysis (too many false positives on music) with energy CoV + pause ratio; music-only clips now excluded from AI probability blend
* **fix** \[dosafe] Video Frame Extraction - LLM visual analysis now sends extracted frames as base64 images instead of raw video URL; text-only fallback on HTTP 500

## 2026-04-01

* **feature** \[gateway] LLM API Marketplace - `GET /v1/catalog` retail endpoint with DB-driven pricing; DeepSeek V3 + Qwen 397B / 122B / 27B now live via DashScope / Alibaba Cloud
* **fix** \[gateway] Billing Race Condition - Replaced race-prone async deduction with atomic `deduct_usage()` PostgreSQL RPC; pre-flight balance gate blocks requests before proxying; streaming billing fixed (was charging 0 tokens on SSE responses)
* **feature** \[gateway] Multi-Backend Agent Routing - `BackendRouter` dispatches agent ops to local or Azure backend per agent; Azure VM CPU/memory/disk monitored via `/metrics`, alerts on 3 consecutive threshold breaches
* **feature** \[gateway] Embeddings Endpoint - `/v1/embeddings` routes to dedicated pooling backend, falls back to DashScope `text-embedding-v4` when local vLLM unavailable
* **feature** \[dashboard] dos-ai Launch Promo - $0.01 / 1M tokens (down from $0.10); promo badge with strikethrough original price on models page and model detail
* **feature** \[dashboard] SEO - JSON-LD structured data, Open Graph, robots.txt, sitemap.xml, `llms.txt`, `ai.txt` for AI crawlers
* **feature** \[dosclaw] Lite Agent Tier - New 2 GB / 0.5 vCPU instance size for free tier users
* **feature** \[dosafe] RDAP + URL Path Detection - RDAP domain registration data and URL path heuristics added to entity risk assessment
* **fix** \[dosclaw] Pairing & Channels - Shared bot auto-approves pairing; Discord Gateway WebSocket reconnect fixed; slug→UUID deep link resolution fixed
* **fix** \[dashboard] Signup Bonus - Fixed $5 credit not granted on Google OAuth login

## 2026-03-31

* **feature** \[gateway] Enriched Entity Check - `firstSeenAt`, `reportCount`, `relatedEntities` added to `/v1/dosafe/check` response
* **feature** \[dosclaw] Custom Bot Token Validation - Token validated against Telegram API before saving; Disconnect button for shared/custom bot links; clear existing webhook on new custom token connect
* **feature** \[dosclaw] Channel Linked State UI - Discord/Telegram channel cards show live connection state (shared bot vs custom token)
* **feature** \[dashboard] i18n Agent Detail - 200+ hardcoded strings translated; agent detail page fully internationalized across all 7 supported languages
* **improvement** \[dashboard] Language Switcher Redesign - Globe icon + locale code dropdown with checkmark, matching DOSafe design
* **fix** \[dosclaw] Agent Stability - Fixed OOM crash loops, memory persistence on restart, deep link slug→UUID resolution

## 2026-03-30

* **feature** \[dashboard] Comprehensive i18n - All dashboard pages internationalized (agents list, agent detail, plans, billing, settings)
* **feature** \[dashboard] Support Banner - Announcement banner in topbar with Telegram and Discord community links
* **feature** \[dashboard] AI Detector Redirect - AI Detector page redirects to dosafe.io (DOSafe owns detection features)
* **fix** \[dosclaw] Agent Deep Links - Slug→UUID resolution fixed in DeepLink handler; pairing code shown explicitly in Telegram connect UI
* **fix** \[gateway] Image Upload Size - nginx `client_max_body_size` increased to 50 MB for image detection uploads
* **perf** \[dashboard] Agent Detail Load - Parallel data fetching reduces agent detail page load time

## 2026-03-29

* **feature** \[dosclaw] Instance ID Badge - Each agent now shows a short instance ID (e.g. #568bc2) in the header for quick identification when reporting issues
* **fix** \[dosclaw] Open Console Instant Load - Console button now appears immediately on page load instead of after a 4-5s delay
* **fix** \[dosclaw] Provisioning Progress Bar - Redeploy now shows the same step-by-step progress bar as initial deploy
* **fix** \[dosclaw] Agent Health Stability - Gateway marked healthy only after 3 consecutive successful health checks, preventing premature "Open Console" during boot
* **feature** \[dashboard] Language Switcher - Added support for Chinese, Japanese, Korean, Thai, and Indonesian in addition to Vietnamese and English

## 2026-03-28

* **feature** \[dosclaw] WhatsApp Dual-Mode - Support both shared bot and native QR code connection for WhatsApp
* **feature** \[dosclaw] Discord Connect UI - New Discord integration panel with bot invite flow and connection status
* **fix** \[dosclaw] Shared Bot Status - Channel cards now correctly show shared bot connection state for all supported platforms
* **fix** \[dosclaw] WhatsApp QR Generation - Fixed CLI flags and stream header parsing for reliable QR code display
* **fix** \[dosclaw] Container Lifecycle - Confirm dialogs for stop/restart; fixed upgrade race condition and lxcfs mount errors on restart
* **fix** \[dosclaw] Agent Entrypoint - Phased background installs to prevent OOM during boot; pinned dependency versions for reproducibility
* **improvement** \[dosclaw] Agent Identity - Agent name and persona now managed via IDENTITY.md for cleaner runtime separation

## 2026-03-26

* **feature** \[dosclaw] Instance Sizing - Choose Standard (1 vCPU, 2 GB, $5) or Plus (2 vCPU, 4 GB, $10) when creating agents
* **fix** \[dosclaw] Version Selector Filter - Hide architecture-specific tags (-arm64, -amd64), show only multi-arch versions
* **fix** \[dosclaw] Token Preservation - Fixed bug where disabling/enabling channels erased encrypted bot tokens

## 2026-03-25

* **feature** \[dashboard] Alert Notifications - Telegram alerts for provisioning failures and container resource limits
* **feature** \[dosclaw] Auto-capture Owner Chat ID - Bot automatically captures owner's Telegram chat ID on first message for alert delivery

## 2026-03-24

* **feature** \[dosclaw] Version Pinning - Pin agents to a specific OpenClaw version or follow latest
* **feature** \[dosclaw] Agent Upgrade - One-click upgrade to latest OpenClaw image with graceful container recreation
* **feature** \[dosclaw] Brave Search Integration - Connect Brave Search API key to enable web\_search tool
* **feature** \[dosclaw] GitHub Integration - Connect GitHub token for code-related agent skills

## 2026-03-23

* **feature** \[dosclaw] Standalone Agent Containers - No HiClaw Manager dependency, agents run independently with trusted-proxy auth
* **feature** \[dosclaw] Bot Personality System - First-run onboarding: name, creature, vibe, emoji
* **feature** \[dosclaw] Agent Templates - 5 templates (Personal Assistant, Sales TikTok, CS Shopee, Content Creator, Custom)
* **feature** \[dosclaw] Credit-based Pricing - Free=1 trial bot 7d, Plus=1 free, Pro=3 free, extra bots charged from credits
* **feature** \[dosclaw] Console Proxy - Trusted-proxy auth for browser-based agent console
* **improvement** \[dashboard] Agent Creation Flow - 2-step UI: template selection then configure screen
* **breaking** \[dosclaw] Removed CoPaw Runtime - All agents now use OpenClaw runtime only

## 2026-03-21

* **feature** \[gateway] Go API Gateway on Cloud Run - Replaced Cloudflare Worker with full Go backend at api.dos.ai
* **feature** \[gateway] LLM Inference Proxy - vLLM primary + fallback providers, SSE streaming
* **feature** \[gateway] API Key Auth - dos\_sk\_\* format, SHA-256 hash lookup, sliding window rate limiting
* **feature** \[gateway] Credit Billing - Balance check, token-based deduction, credit management

## 2026-03-20

* **feature** \[dosclaw] Agent Knowledge Base - File upload with RAG indexing for agent context
* **feature** \[dosclaw] Agent Backup/Restore - Export and import agent configurations
* **feature** \[dosclaw] Default Permissions - Allow-all or ask-permission modes per agent

## 2026-03-15

* **feature** \[dosafe] SPAI Detector - Spectral analysis for modern AI image detection (CVPR 2025)
* **feature** \[dosafe] Context-Aware Ensemble - Dynamic model weights based on image metadata (editor, compression, beauty app)
* **feature** \[dosafe] Exact Match Safety Cap - Reverse search exact matches cap AI score (3+ matches -> 45%)
* **improvement** \[gateway] LLM Prompt V3 - Calibrated for DINOv3 + SPAI + CommFor ensemble

## 2026-03-10

* **feature** \[dosafe] Web Search + LLM Analysis - 4th data source for entity risk assessment (Serper/SerpApi + vLLM)
* **feature** \[dosafe] Partner API - Public API at dosafe.io/api/v1 with API key auth and scoped permissions

## 2026-03-08

* **feature** \[dosclaw] Onboarding Wizard - Full CLI setup: security, model, channels, skills, hooks, hatching
* **feature** \[dosclaw] Plan Enforcement - Soft warning for plan limits instead of hard block
* **feature** \[dashboard] Channel Selector - Telegram, Discord, WhatsApp (coming soon) for agent creation

## 2026-03-01

* **feature** \[gateway] DOSafe Integration - Entity check, bulk check, URL check, text/image AI detection endpoints
* **feature** \[gateway] Agent Management API - Full CRUD + lifecycle (start, stop, upgrade, restart)
* **feature** \[gateway] Anonymous IP Quota - 20 checks/day for DOSafe public endpoints
* **improvement** \[gateway] User Quota Persistence - Moved from in-memory to Supabase

## 2026-02-15

* **feature** \[dashboard] Organization Management - Create and manage organizations
* **feature** \[dashboard] API Key Management - Create, list, revoke keys with usage tracking
* **improvement** \[dashboard] Dark Mode - Fixed flash on select dropdowns
* **fix** \[dashboard] Token Refresh - Fixed billing precision issues

## 2026-02-01

* **feature** \[dashboard] Supabase Auth Migration - OAuth, password login, logout, identity management
* **feature** \[dashboard] Avatar Upload - Profile photo upload with organization support
* **feature** \[dashboard] Billing System - Credit-based billing with Stripe integration
* **improvement** \[dashboard] Server-side OAuth - Redirect to avoid /login flash


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question.

Perform an HTTP GET request on the current page URL with the `ask` query parameter:

```
GET https://docs.dos.ai/changelog.md?ask=<question>
```

The question should be specific, self-contained, and written in natural language.
The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.