v0.3.0 — Inline Keyboards & Mini App

What's New

Telegram Inline Keyboards

Transform the bot from command-driven to tap-to-act. Every response and listing now includes interactive buttons:

• Agent responses — [Regenerate] [Continue] [Switch Model] buttons on every reply
• Session management — [Resume] [Fork] buttons per session in /sessions
• Job management — [Run] [Pause] [Enable/Disable] [Delete] buttons per job in /jobs
• Model picker — Tap Switch Model for [Opus] [Sonnet] [Haiku] quick selection
• Delete confirmation — [Yes, delete] [Cancel] safety prompt before removing jobs
• 12 callback actions with compact action:payload encoding (fits Telegram's 64-byte limit)
• Shared ChatState class between commands and callback handler for consistent state

Telegram Mini App Dashboard

Embedded mobile dashboard accessible via the "Dashboard" menu button in Telegram:

• Chat — Full agent chat with SSE-streamed responses, themed message bubbles
• Jobs — View, trigger, pause, toggle, and delete scheduled jobs
• Memory — Browse and search agent memory entries
• Sessions — Combined mini-app and Telegram session history with tap-to-resume
• Native Telegram theming via --tg-theme-* CSS variables (auto light/dark)
• HMAC-SHA256 initData authentication for API routes per Telegram spec
• Mounted at /tma, exempted from IP allowlist, secured by Telegram auth

Agent Reliability

• Agent now always provides a final response when hitting turn or budget limits instead of silently stopping
• Stop reason tracking (end_turn, max_turns, budget_exceeded, stale_timeout) shown in run metadata

Files Changed

19 files changed, 1,352 insertions, 51 deletions

Full Changelog: v0.2.0...v0.3.0

v0.2.0 - Job Pausing, Model Selection & YouTube Improvements

What's New

Job Pause/Resume

• Pause until a date/time — job stays scheduled but executor skips runs until the datetime passes, then auto-clears
• Skip N runs — job skips the next N scheduled executions, decrementing each tick, then auto-resumes
• Both modes can be combined; manual triggers (job_trigger) always bypass pause conditions
• New job_pause and job_resume MCP tools for Telegram control
• Web UI: inline pause form, amber status badges, and resume button

Per-Job Model Selection

• Jobs can now specify a Claude model (e.g. claude-sonnet-4-5-20250929) instead of always using the default
• Supported across MCP tools (job_create, job_update, job_get), web UI, and executor
• YouTube Watchlist Processor job seeded with explicit Sonnet model

YouTube Pipeline Improvements

• Real-time progress notifications via Telegram during playlist processing
• Per-video status details ([NEW], [SKIP], [MOVE], [FAIL]) in processing results
• Model passthrough for transcript analysis (configurable via YOUTUBE_ANALYSIS_MODEL)
• Reduced retry limit to prevent re-processing stuck videos
• New scripts/audit-playlists.ts for database synchronization audits

Agent Infrastructure

• model override support in AgentRunOptions — queries can now specify a model per-run
• Keep-alive callback resets the stale timer during long-running tool calls (e.g. YouTube processing)

Files Changed

17 files changed, 411 insertions, 36 deletions

Full Changelog: v0.1.0...v0.2.0

v0.1.0 - Initial Release

psibot v0.1.0

Personal AI assistant: Telegram bot + web dashboard, powered by Claude Agent SDK.

Core

• Telegram bot with grammy - command handlers, voice messages, photo handling, auth middleware
• Web dashboard with Hono + HTMX - SSE-based agent chat streaming, job/memory/log viewers
• Claude Agent SDK integration with MCP servers and subagent spawning
• SQLite database with WAL mode, FTS search, and session tracking

Agent Capabilities

• MCP tool servers: memory, knowledge, browser automation, job scheduling, Telegram media, git worktrees
• Media tools: Gemini image generation/editing, Edge TTS (text-to-speech), mlx-audio STT (speech-to-text)
• Subagents: image-generator, audio-processor, coder, researcher - each with specialized models
• YouTube processing: video transcript extraction, structured analysis, vector embeddings for semantic search, playlist batch processing via OAuth

Infrastructure

• macOS daemon via LaunchAgent with CLI management (psibot install/start/stop/restart/status/logs)
• Heartbeat system: periodic maintenance cron with quiet hours and budget limits
• Job scheduler: croner-based cron and one-off scheduled tasks with budget controls
• Webhook mode: Telegram webhooks via Tailscale Funnel
• Session management: resume/fork sessions, session hashing for message stats

User Experience

• Setup/uninstall scripts (scripts/setup.sh, scripts/uninstall.sh) for automated first-time installation from a fresh clone