v2.2.0 — Long-Running Task Infrastructure

What's New

Tasks sent through the bot can now run for hours without being killed. Previously, a 10-minute hard timeout would terminate complex tasks (SSH operations, multi-file refactors, deployments) with a cryptic "exit code null" error.

Changes

• 4-hour safety cap replaces the 10-minute hard timeout — /cancel is now the primary kill mechanism for user sessions
• Progress pings every 5 minutes — "Still working... (X min elapsed). Use /cancel to stop."
• Process health checks — detects zombie processes (close event never fired) every 30 seconds
• Graceful shutdown — if the gateway restarts during a long task, it notifies the user instead of silently killing the process
• /status shows task duration — see how long the current task has been running
• Better error messages — timeout, cancellation, and signal kills each get distinct, user-friendly messages instead of "exit code null"
• Partial result recovery — if Claude produced output before being killed, that output is returned instead of an error

Unchanged

• Lite mode timeout (2 min) — unchanged
• Heartbeat timeout (2 min) — hardcoded, unchanged
• Cron job timeout (5 min) — hardcoded, unchanged
• /cancel and AbortSignal flow — unchanged
• Queue persistence and drain — unchanged

Files Modified

• src/config.ts — fullTimeoutMs → 4h, added progressPingIntervalMs
• src/claude-runner.ts — health check, timeout/signal handling, partial results
• src/index.ts — progress pings, task tracking, graceful shutdown, /status enhancement

v2.1.2 — Telegram Command Menu & PAI Skill Discovery

What's New

Telegram Command Menu — Type / in your chat and you'll see a full menu of available commands. Built-in commands (like /new, /sessions, /cancel) are listed alongside any PAI skills you have installed.

PAI Skill Discovery — The gateway scans ~/.claude/skills/ at startup and registers discovered skills as Telegram bot commands. For example, if you have a Research skill, typing /research quantum computing routes it through Claude Code in full mode.

Details

• New src/skills.ts module — scans skill directories, parses SKILL.md YAML frontmatter, normalizes to Telegram-safe command names
• registerBotCommands() wraps Telegram's setMyCommands API
• Skill routing rewrites /skillname args to "Use the SkillName skill: args" before passing to Claude
• Built-in commands always take priority over skills (no collisions)
• /help now shows discovered skills section
• Telegram practical limit: ~50 total commands (13 built-in + up to 37 skills in menu). All discovered skills remain routable even if not in the menu.
• Registration is non-fatal — if it fails, the bot works normally without the command menu
• full-system.md includes a skill routing hint so Claude invokes skills via the Skill tool

Full Changelog

v2.1.1...v2.1.2

v2.1.1 — Clean Telegram Response Formatting

Fix: PAI Formatting Stripped from Responses

v2.1.0 added a system prompt override to suppress PAI output formatting (═══ PAI headers, TASK/CONTENT/VERIFY sections). However, the system prompt only applies to new sessions — resumed sessions already have PAI formatting rules baked into Claude's context from ~/.claude/CLAUDE.md.

What Changed

Added a post-processing stripPaiFormatting() function that runs on every response before sending to Telegram. It strips:

• ═══ PAI ═══ and ════ PAI | NATIVE MODE ═══ header lines
• 🗒️ TASK: metadata lines
• 🔧 CHANGE: and ✅ VERIFY: sections (including sub-bullets)
• 📋 SUMMARY: sections
• 🗣️ BotName: prefix lines (used as fallback if no other content found)
• Mode label emoji prefix (💡/🔧) that was prepended to every message

Extracts the actual conversational content — 📃 CONTENT: text takes priority, then free-form text, then the bot summary line as fallback.

Before

🔧 ═══ PAI ═══════════════════════════
🗣️ DORA: Nice! Tsunami is a solid pick — enjoy your sushi!

After

Nice! Tsunami is a solid pick — enjoy your sushi!

Upgrade

Drop-in replacement. No config changes needed.

Full Changelog: v2.1.0...v2.1.1

v2.1.0 — Session Memory, Image Handling, PAI Integration

What's New

Session Memory System

Three-layer persistence that keeps your conversations coherent across messages:

• Transcript safety net — per-session JSONL recording. When Claude prunes a session and --resume fails silently, the bot detects it and re-injects conversation context automatically.
• Permanent memory — MEMORY.md at ~/.claude/MEMORY/TELEGRAM/ stores cross-session knowledge (8K char cap). Injected into new sessions.
• Daily interaction logs — written to ~/.claude/MEMORY/RELATIONSHIP/YYYY-MM/YYYY-MM-DD.md. PAI's LoadContext hook reads these automatically, so Telegram conversations surface in desktop Claude Code sessions with zero configuration.

Mode Stickiness

All sessions now auto-lock to full mode for context continuity. Previously, conversational messages like "Yes" or "What were we talking about?" were routed to stateless lite mode, losing all context. Both models use the same sonnet — zero cost difference. Use /lite to opt out if needed.

Image Support

Send photos or image documents via Telegram — they're downloaded, saved to a temp file, and forwarded to Claude with a hint to view them using the Read tool.

Resume Failure Detection

When --resume fails silently (session was pruned by Claude Code), the bot detects it by comparing returned session_id to the stored one. The next message gets conversation context prepended automatically.

Clean Response Format

System prompts now override PAI output formatting. No more ═══ PAI ═══ headers or emoji bullet sections in Telegram responses — just clean conversational text.

Files Changed

• New: src/transcript.ts, src/memory.ts
• Modified: src/full.ts, src/index.ts, src/sessions.ts, src/telegram.ts
• Prompts: prompts/full-system.md, prompts/lite-system.md
• Docs: CLAUDE.md, README.md, package.json

Upgrade Notes

• No breaking changes — drop-in replacement for v2.0.0
• New data/transcripts/ and data/memory/ directories are created automatically on first run
• If you have PAI installed, Telegram memory integrates automatically via ~/.claude/MEMORY/

Full Changelog: v2.0.0...v2.1.0

Beta 1.0.0

PAI Mobile — Beta 1.0.0

The first public beta of PAI Mobile, a Telegram gateway that connects your phone to Claude Code.

What's included

• Dual-mode routing — keyword classifier auto-routes simple questions (lite) vs complex tasks (full Claude Code with tools)
• Multi-session management — create, switch, and manage multiple conversation sessions
• Heartbeat system — periodic AI check-ins that alert you only when something is actionable
• Cron scheduler — natural language scheduled tasks ("daily 7am", "weekdays 9am", "every 2h")
• Outbound queue — bundles proactive messages, rate limits, retries
• macOS service management — launchd auto-start, crash recovery, one-command install
• Security hardening — chat ID whitelist, permission modes, env isolation for system sessions, rate limiting

Getting started

See the README for step-by-step setup instructions.

Feedback

This is a beta release for the PAI community. Open an issue or drop feedback in PAI Discussions.