CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

Deprecated

• telegram_schedule_message agent tool: Now logs a runtime deprecation warning and surfaces deprecated: true plus a deprecationNotice field in its result. The tool only queues plain text and cannot execute tools, trading functions, or multi-step workflows when the message is delivered, which silently breaks any automation that relies on it. Use telegram_create_scheduled_task (with a tool_call or agent_task payload) for any automation that must run at a scheduled time. The tool description now leads with [DEPRECATED — use telegram_create_scheduled_task instead] so the LLM picks the correct tool by default (closes #459).

Added

• Bot API HTTPS proxy (mtproto.bot_api_proxy): Optional HTTP/HTTPS or SOCKS5 proxy URL for Telegram Bot API HTTPS calls to api.telegram.org. MTProto proxies cannot tunnel HTTPS, so this lets the deals bot reach the Bot API in regions where Telegram is also blocked at the IP level. Wired through to Grammy's client.baseFetchConfig.agent via https-proxy-agent / socks-proxy-agent (closes #439).

Fixed

• Spurious [Bot] Polling error: Aborted delay on Ctrl+C: DealBot.start() no longer logs the polling promise rejection that Grammy raises when bot.stop() aborts the in-flight long-poll delay. Stopping the agent (especially with the MTProxy path active) is now silent on the polling channel, while real polling failures during normal operation continue to be logged (closes #460).
• WorkflowScheduler cron deduplication (AUDIT-M7): tick() now tracks runningWorkflowIds (in-memory Set) to skip workflows whose previous execution is still in progress, and persists last_fired_bucket (floor(ms/60000)) to the DB so the same minute bucket never fires twice — even after a process restart. DB migration 1.26.0 adds the last_fired_bucket column to workflows (closes #327).

Changed

• Autonomous TON spending defaults tightened (AUDIT-M3): DEFAULT_POLICY_CONFIG.tonSpending reduced by 10× (perTask 1 → 0.1 TON, daily 5 → 0.5 TON, requireConfirmationAbove 0.5 → 0.05 TON) to limit financial exposure for users who run the agent with a linked wallet and do not customise the policy config. Users who relied on the previous permissive defaults must explicitly raise the limits in their config.yaml under the autonomous.policy.ton_spending key (closes #286).

Added

• Prediction engine: Behavior event tracking, Markov-style next-action predictions, topic-to-tool suggestions, WebUI prediction APIs, and dashboard suggestions with feedback.
• web_download_binary tool: Download public HTTP(S) binary files into workspace downloads/ with MIME validation, a 10 MB size cap, redirect support, and optional request headers for authorized URLs.
• Upstash Vector setup guide (docs/upstash-vector-setup.md): Step-by-step walk-through for provisioning the Upstash index with the dimension the embedding provider produces, connecting Teleton through the WebUI / config.yaml / environment variables, verifying the health-check log, and recovering from a dimension mismatch. Referenced from the README, GETTING_STARTED.md, docs/configuration.md, and docs/semantic-memory.md (closes #248).

Fixed

• Autonomous task escalations reach the user (AUDIT-H2): notify() in src/autonomous/integration.ts now pushes a Telegram DM to every admin via deps.bridge.sendMessage, records an in-app warning in the notifications table, and emits escalation / update events on notificationBus for real-time WebUI badges. Delivery failures are caught per channel so log.warn remains the last-resort fallback (closes #262).
• Vector memory sync: Detect Upstash Vector index/embedding dimension mismatches before upsert, surface the configured index dimension in semantic memory status and sync responses, and log an actionable warning at startup (closes #246).
• Autonomous policy bypass via pause/resume (AUDIT-C3): AutonomousLoop now persists and hydrates PolicyEngine state (rate-limit sliding windows, loop-detection recent actions, uncertainty counter) through a new policy_state table so that scripted pauseTask() + resumeTask() cycles can no longer reset the 100 tool-calls-per-hour limit or the 5-identical-actions loop detector. Adds migration 1.23.0 and regression tests covering 10 pause/resume cycles (closes #256).

0.8.1 - 2026-03-05

Added

• TON Proxy module: Built-in Tonutils-Proxy lifecycle manager — auto-download binary from GitHub, start/stop, health checks, auto-restart on crash, PID-based orphan cleanup, WebUI API routes for hot-toggle
• SDK signed transfers: createTransfer(), createJettonTransfer(), getPublicKey(), getWalletVersion() — sign TON/jetton transfers without broadcasting for x402 payment protocol
• Plugin hooks system: 13 typed hooks via sdk.on() — message:receive, response:before/after/error, tool:error, prompt:after, agent:start/stop, plus 5 original lifecycle hooks with configurable priority
• User-configurable hooks: Keyword blocklist and context triggers for automated responses
• QR code login: WebUI setup wizard supports QR code authentication as alternative to phone+code
• Two-phase observation masking: Old tool results fully masked, previous iteration results truncated at 4K while preserving summary fields, current iteration intact

Changed

• WebUI Config page: Reorganized into dedicated tabs (Agent, Telegram, TON Proxy, Sessions, Tool RAG)
• RAG performance: Knowledge + feed hybrid searches run concurrently via Promise.all (~200-500ms saved per message); parsed transcripts cached in memory with invalidation on delete/archive
• 15 LLM providers: Documentation updated across all .md files to reflect Cerebras, ZAI, MiniMax, Hugging Face additions
• 70+ models in shared catalog (up from 60+)

Fixed

• Tool RAG scoring: Keyword search scores normalized to 1.0 weight when no embedding provider is configured (was incorrectly using 0.4)
• Transcript deduplication: loadContextFromTranscript() deduplicates toolResult messages by toolCallId, preventing API 400 errors on corrupted transcripts
• TON Proxy orphan process: Manager now writes PID file and checks port occupancy before start, killing orphan processes from previous sessions
• Security: Sanitize hook context, fix effectiveIsGroup self-reference crash (TDZ)
• CI: Coverage thresholds lowered with margin for Node 20 CI variance
• ESLint: Strict config with quality tooling and CI hardening

0.8.0 - 2026-03-02

Added

• 4 new LLM providers (11 → 15): Cerebras (ultra-fast inference, free tier), ZAI/Zhipu (2 free models), MiniMax (M2.5, 204K ctx), Hugging Face (routing to 18 models via single token)
• Bot SDK for plugins: sdk.bot with inline query handling, callback routing, colored/styled buttons (success/danger/primary), lazy-loaded, rate-limited, namespace-isolated per plugin
• 29 new SDK methods: Full Telegram surface (77 tools), TON jetton analytics, dual DEX aggregator, .ton DNS management, scheduled messages, Stars/gift marketplace, getDialogs/getHistory, kickUser
• dns.setSiteRecord(): Set ADNL records on .ton domains for TON Site hosting
• GramJS Layer 223: Participant ranks and message from_rank surfaced in agent display

Changed

• Moonshot provider: Refactored from hardcoded model dict to pi-ai native kimi-coding provider (30 lines removed). Backward-compat alias maps kimi-k2.5 → k2p5
• Configurable keys: Provider list derived from getSupportedProviders() instead of hardcoded copy

Fixed

• Docker build: Remove deleted scripts/ references from Dockerfile; skip husky prepare in runtime stage
• Release workflow: Publish-npm and create-release skip gracefully when version already published (idempotent re-push)
• Security: NFKC normalization + Unicode Tag Block filtering, SQL comment stripping on plugin DB proxy, download size guard (50MB), deep-clone frozenConfig
• Performance: Single shared embedding for context + tool RAG, edges-first chunk reordering, feed truncation (2000 chars)
• UTC session reset, transcript permissions, masked API key display

0.7.5 - 2026-02-28

Added

• YOLO Mode (Coding Agent): 4 new exec tools for full system access on Linux — exec_run (bash commands), exec_install (apt/pip/npm/docker), exec_service (systemd management), exec_status (server health). Disabled by default (mode: off), requires explicit mode: yolo opt-in. Admin-only scope, configurable timeout (120s), output limit (50KB), full audit trail in SQLite
• admin-only access policy: New DM and group policy option — only Telegram admins can interact with the agent. Now the default for new installations (previously open)
• DNS set-site tool: dns_set_site links a .ton domain to a TON Site via ADNL address for decentralized website hosting
• GramJS Layer 222 fork: Switch from npm telegram to TONresistor/gramjs fork — native Layer 222 constructors, no more TL schema patching
• 4 NFT marketplace tools (73 → 77): get-unique-gift, get-unique-gift-value, send-gift-offer, resolve-gift-offer
• Gift service messages: Real-time handling of gift offers received/declined and gifts received — agent can react automatically
• TON balance query: telegram_get_stars_balance now supports ton=true for internal TON ledger balance
• Live token usage tracking: WebUI dashboard displays real-time token consumption with cache hit rates
• Channel username tools (70 → 73): check-channel-username, set-channel-username, create-channel-username
• Toncenter API key: Centralized TonClient caching with optional Toncenter API key for higher rate limits
• DB migration 1.12.0: exec_audit table for command execution history (indexed by timestamp, user)
• DB migration 1.13.0: Per-session token usage tracking (input/output tokens accumulated per chat)
• Session auto-pruning: Sessions older than 30 days are automatically cleaned up at startup

Changed

• Tool RAG enabled by default: Semantic tool selection now active for all providers, reducing ~120 tools to ~25 per LLM call
• 35+ tool descriptions enriched: Cross-references and clearer context for better RAG matching accuracy
• Default access policy: DM and group policies default to admin-only instead of open — secure by default
• CLI wizard: New "Coding Agent" setup question, policy choices reordered (Admin Only first)
• WebUI wizard: New "System Execution" select with YOLO mode + VPS warning
• Dashboard: Policy selects updated with admin-only option and clearer labels
• Gift catalog rework: get-available-gifts now supports pagination, sorting (price, resale count), search by title, and resale filter
• Resale identifiers: buy-resale-gift migrated from odayId to slug, set-collectible-price from odayId to msgId
• Resale error handling: STARGIFT_RESELL_TOO_EARLY parsed with human-readable wait time, STARGIFT_INVALID with guidance
• Styled keyboard: Native Layer 222 constructors for KeyboardButtonStyle, KeyboardButtonCopy, KeyboardButtonCallback — no more (Api as any) casts
• WebUI dashboard: Redesigned with provider switch, tools & plugins panels
• WebUI config page: Harmonized UX across all settings panels
• Ston.fi DEX: Migrated to SDK v2 with hardened SendMode and transaction locking

Fixed

• Typing indicator: Persistent typing during agent processing with retry and dedup hardening
• Auth flow: Guard SentCodePaymentRequired type (Layer 222 narrowing) in both CLI and WebUI auth
• send-gift: Use getInputEntity() instead of getEntity() for correct InputPeer type

Removed

• Postinstall patch system: scripts/patch-gramjs.sh and scripts/postinstall.mjs — no longer needed with Layer 222 fork

0.7.4 - 2026-02-25

Added

• Configurable keys overhaul: Array type support (admin_ids, allow_from, group_allow_from), labels and option labels on all keys, new keys for Telegram rate limits, Deals params, Embedding model, Cocoon port, Agent base_url
• ArrayInput component: Tag-style input for managing array config values in the dashboard
• Memory sources browser: List indexed knowledge sources with entry counts, expand to view individual chunks with line ranges
• Workspace image preview: Serve raw images with correct MIME type, 5MB limit, SVG sandboxing
• Tool RAG persistence: RAG config (enabled, topK, alwaysInclude, skipUnlimitedProviders) now persists to YAML
• Tasks bulk clean: Clean tasks by terminal status (done, failed, cancelled) instead of just done
• GramJS bot session persistence: Save/load MTProto session string to avoid re-auth on restart

Changed

• Remove "pairing" DM policy: Simplified to open/allowlist/disabled — pairing was unused
• Dashboard Config page reorganized with Telegram settings section, Cocoon port panel, extended Tool RAG controls
• Setup wizard flow reordered, wallet and modules steps cleaned up
• Dashboard and Config pages restructured for better UX
• Soul editor textarea fills available height

Fixed

• Select dropdown renders via portal (z-index stacking fix)
• Model selection moved into Provider step (no longer separate Config step)
• Async log pollution during CLI setup suppressed
• Telegram commit notification extra blank lines removed
• owner_id auto-syncs to admin_ids on save

0.7.3 - 2026-02-24

Added

• Claude Code provider: Auto-detect OAuth tokens from local Claude Code installation (~/.claude/.credentials.json on Linux/Windows, macOS Keychain) with intelligent caching and 401 retry
• Reply-to context: Inject quoted message context into LLM prompt when user replies to a message
• Fragment auth: Support Telegram anonymous numbers (+888) via Fragment.com verification
• 7 new Telegram tools (66 → 73): transcribe-audio, get/delete-scheduled-messages, send-scheduled-now, get-collectible-info, get-admined-channels, set-personal-channel
• Voice auto-transcription: Automatic transcription of voice/audio messages in handler
• Gated provider switch: Dashboard provider change requires API key validation before applying
• Shared model catalog: 60+ models across 11 providers, extracted to model-catalog.ts (eliminates ~220 duplicated lines)

Fixed

• TEP-74 encoding: Correct jetton transfer payload encoding and infrastructure robustness
• Replace deprecated claude-3-5-haiku with claude-haiku-4-5
• Seed phrase display in CLI setup
• Bump pi-ai 0.52 → 0.54, hono 4.11.9 → 4.12.2, ajv 8.17.1 → 8.18.0

0.7.2 - 2026-02-23

Fixed

• Plugins route: WebUI now reflects runtime-loaded plugins instead of static config

0.7.1 - 2026-02-23

Added

• Agent Run/Stop control: Separate agent lifecycle from WebUI — start/stop the agent at runtime without killing the server. New AgentLifecycle state machine (stopped/starting/running/stopping), REST endpoints (POST /api/agent/start, /stop, GET /api/agent/status), SSE endpoint (GET /api/agent/events) for real-time state push, useAgentStatus hook (SSE + polling fallback), and AgentControl sidebar component with confirmation dialog
• MCP Streamable HTTP transport: StreamableHTTPClientTransport as primary transport for URL-based MCP servers, with automatic fallback to SSEClientTransport on failure. mcpServers list is now a lazy function for live status. Resource cleanup (AbortController, sockets) on fallback. Improved error logging with stack traces

Fixed

• WebUI setup wizard: Neutralize color accent overuse — selection states, warning cards, tag pills, step dots all moved to neutral white/grey palette; security notice collapsed into <details>; "Optional Integrations" renamed to "Optional API Keys"; bot token marked as "(recommended)"
• Jetton send: Wrap entire sendJetton flow in try/catch for consistent PluginSDKError propagation; remove SendMode.IGNORE_ERRORS (errors are no longer silently swallowed); fix || → ?? on jetton decimals (prevents 0 decimals being replaced by 9)

0.7.0 - 2026-02-21

Added

• WebUI Setup Wizard: 6-step guided onboarding flow (Welcome, Provider, Telegram, Config, Wallet, Connect) with shared Shell sidebar layout, React context state management, server-side validation mirror, and "Start Agent" button with seamless setup-to-dashboard transition
• Local LLM Provider: New "local" provider for OpenAI-compatible servers (Ollama, vLLM, LM Studio, llama.cpp) with auto-model discovery from /models endpoint, CLI --base-url option, and WebUI provider card
• getEffectiveApiKey() helper for consistent API key resolution across all LLM call sites
• ASCII banner for teleton setup --ui matching teleton start
• 86 setup route tests + 39 validation tests (898 total tests)

Fixed

• Security audit remediation (27 fixes): MCP env var blocklist, sendStory symlink-safe path validation (realpathSync), DB ATTACH/DETACH proxy for plugin isolation, BigInt float precision (string-based decimals), debounce clamp, SendMode.IGNORE_ERRORS removed, URL quote escaping, wallet JSON validation, pino redact, and more
• fetchWithTimeout (10s) + http/https scheme validation on local model discovery
• Model array capped to 500 entries to prevent unbounded growth
• Early exit when provider=local but base_url missing
• Non-interactive onboarding: relaxed --api-key for local/cocoon providers
• WebUI UX: CSS specificity fixes, bot token inline field, wallet address prominent display, TonAPI/Tavily as plain optional fields

Note — 2026-02-21

Git history rewritten to fix commit attribution (email update from tonresistor@github.com to the account owner's actual email). All commit hashes changed; code, dates, and messages are identical. Tags re-pointed to new hashes. Force-pushed to origin. No code or functionality was affected.

0.6.0 - 2026-02-20

Added

• Cocoon Network proxy-only LLM provider with XML tool injection
• Moonshot (Kimi K2.5 / K2 Thinking) LLM provider
• Mistral LLM provider
• Pino structured logging — migrated from console.* across entire codebase
• MCP client support with CLI management commands (teleton mcp add/remove/list)
• Plugin Marketplace with secrets management and download functionality
• WebUI: Config + MCP pages, custom Select component, centralized CSS
• WebUI: accordion UI, dashboard settings
• Tool RAG, web tools, and admin enhancements

Changed

• Type safety overhaul: reduced as any from 135 to 32 instances
• Setup wizard migrated to @inquirer/prompts with auto-resolve owner
• All dependencies upgraded to latest versions

Fixed

• Data integrity and cleanup from full audit

0.5.2 - 2026-02-16

Added

• Auto-install npm dependencies for plugins on load

Fixed

• Robust local embedding model loading (ONNX cache dir fix for global installs)

Removed

• Dead dependencies from package.json
• Obsolete TGAPI.md documentation file

0.5.1 - 2026-02-16

Changed

• CI/CD pipelines for SDK, WebUI, and Docker builds

0.5.0 - 2026-02-16

Added

• Data-bearing tool categories with strict DB row types
• Plugin event hooks: onMessage and onCallbackQuery
• WebUI: inline dropdown task details with overflow fix
• WebUI: auth system, dashboard, tool config, plugins page, and documentation pages
• Plugin SDK expansion to 53 methods

Changed

• RAG rebalancing for improved search relevance
• Core hardening and open-source cleanup
• Plugin SDK extraction to standalone package

Fixed

• Key caching, transaction reliability, debouncer, and market extraction

0.4.0 - 2026-02-14

Added

• Plugin SDK with namespaced services (sdk.ton, sdk.telegram, sdk.db)
• DeDust prices and token-info tools
• /task admin command connected to scheduled task system
• Local embeddings with hybrid vector search (sqlite-vec + FTS5)
• Casino extracted as external plugin

Changed

• DEX tools reorganized by provider with scope security enforcement
• Memory init deduplicated, using isVectorSearchReady()
• System prompts hardened with memory size management
• Crypto-safe randomId used across codebase

Fixed

• sqlite-vec startup logs no longer print before ASCII banner
• ChatId validation prevents entity resolution crashes on display names
• DELETE+INSERT for vec0 tables (upsert is unsupported)
• Auto-migrate legacy plugin data from memory.db on first startup
• Plugin SDK hardened: escape hatch removed, timeouts and cleanup added
• Sender ID always included for unambiguous user identification

Removed

• Built-in casino module (replaced by external plugin)

0.3.0 - 2026-02-13

Added

• Local ONNX embeddings (Xenova/all-MiniLM-L6-v2)
• Hybrid vector + FTS5 search for RAG

Fixed

• Docker image name corrected in README
• Guard against undefined model from pi-ai getModel()
• Bot messages ignored in DMs to prevent bot-to-bot loops

0.2.5 - 2026-02-12

Added

• Per-group module permissions with /modules admin command
• Swap tools allowed in groups with module level display

Fixed

• /clear command crashing on missing vec0 table
• Post-audit hardening: timeouts, seqno race, cached endpoints
• Bot token made mandatory when deals module is enabled

Removed

• Unused @tonkite/highload-wallet-v3 dependency

0.2.4 - 2026-02-10

Fixed

• Memory database properly closed on shutdown
• Atomic deal state guards prevent race conditions

0.2.3 - 2026-02-10

Fixed

• MarketPriceService crash on fresh installs

0.2.2 - 2026-02-10

Fixed

• Peer cache used in bridge.getMessages for reliable entity resolution

0.2.1 - 2026-02-10

Changed

• Tool registration decentralized into co-located ToolEntry arrays

Fixed

• Cached peer entity used in get-history for reliable channel resolution
• Mention detection fallback and duplicate message guard

0.2.0 - 2026-02-10

Changed

• Deals and market extracted into standalone modules
• Gemini schema sanitizer for Google provider compatibility
• Casino extracted into self-contained plugin module

Removed

• Dead casino files (game-executor, validators)

0.1.21 - 2026-02-09

Added

• Prompt injection defense and tool context scoping

Fixed

• clearHistory order, cached endpoint, tasks index
• install.sh reads from /dev/tty and uses lowercase Docker image name

Removed

• Jackpot system removed entirely

0.1.20 - 2026-02-09

Added

• getTonPrice() caching with 30-second TTL
• Completed deals logged to business journal
• Transcript files older than 30 days cleaned up at startup

Fixed

• Shallow copy returned from getTonPrice cache

0.1.19 - 2026-02-08

Fixed

• Folder IDs start at 2 (IDs 0-1 reserved by Telegram)
• GetDialogFilters returning object instead of array
• DialogFilter title wrapped in TextWithEntities for GramJS layer 222+
• Atomic status preconditions added to deal verify-payment

0.1.18 - 2026-02-08

Added

• Optimized runtime logs and TonAPI rate limiting

0.1.17 - 2026-02-08

Added

• /boot admin command for agent bootstrap

Fixed

• Deals and Market merged into single module option
• Imperative placeholders removed from MEMORY.md template

0.1.16 - 2026-02-08

Fixed

• Agent empty response when memory_write is the only tool call
• @ston-fi bundled with all transitive deps via external blacklist

0.1.15 - 2026-02-08

Fixed

• @ston-fi bundled with all transitive dependencies

0.1.10 - 0.1.14 - 2026-02-08

Fixed

• Repeated @ston-fi bundling and dependency resolution fixes
• postinstall script removed to avoid preinstall blocker

0.1.9 - 2026-02-08

Fixed

• @ston-fi/api bundled to avoid pnpm-only install blocker

0.1.8 - 2026-02-08

Fixed

• scripts/ directory copied in Dockerfile build stage

0.1.7 - 2026-02-08

Fixed

• Docker build issues resolved

0.1.6 - 2026-02-08

Added

• First public npm release with Docker support

Fixed

• Docker build failing due to husky in production install
• Docker tags lowercased, release decoupled from Docker

0.1.4 and earlier - 2026-02-08

Added

• Initial release of Teleton Agent
• Autonomous Telegram AI agent with TON blockchain integration
• Multi-provider LLM support (Anthropic, OpenAI, Google, xAI, Groq, OpenRouter)
• Deals system with inline bot, payment verification, and auto-execution
• Styled inline buttons and custom emoji via MTProto layer 222 patch
• Interactive setup wizard with wallet safety and model selection
• Admin commands: /model, /policy, /pause, /resume, /wallet, /stop, /loop
• TonAPI key support for higher rate limits
• Professional distribution (npm, Docker, CI/CD)
• Pre-commit hooks and linting infrastructure