feat: streaming drafts, rich formatting, follow-up interrupts, and UX improvements

[Fr4nzz]

Summary

• Streaming draft responses via sendMessageDraft — real-time token streaming in private chats with fallback to editMessageText for groups. Draft cleared before final response to prevent stale bubbles; draft text reset after each 💬 message to avoid accumulation
• Rich HTML formatting — parse Claude's markdown into Telegram HTML (bold, italic, code, links) with smart message splitting
• Follow-up message interrupts — send a new message while Claude is working to redirect, like vanilla Claude Code
• Per-event verbose messages — tool calls, thinking, and assistant text shown as individual messages with configurable verbosity (0/1/2)
• Image sending via MCP — intercept send_file_to_user tool calls, validate paths, send as Telegram photos
• SDK integration — use setting_sources=["project"] to avoid plugin MCP conflicts, support SSE-based MCP servers
• Keep progress messages visible after response with elapsed time

Test plan

• Send a text message and verify streaming draft appears in real-time
• Send a follow-up message while Claude is working — verify it interrupts
• Test /verbose 0, /verbose 1, /verbose 2 — verify output levels
• Test image sending via Playwright screenshot + MCP send_file_to_user
• Verify draft bubble clears before final response appears
• Test in group chat (editMessageText fallback)
• Run pytest tests/unit/ -o "addopts=" — all tests pass

🤖 Generated with Claude Code

[RichardAtCT]

Big feature PR with some genuinely useful additions — streaming drafts and interrupt support are quality-of-life wins. Several correctness issues and a process concern before this can land.

---

🔴 Blockers

1. /stop is a no-op — running_claude_task is never set

agentic_stop reads context.user_data["running_claude_task"] to cancel the task, but nowhere in the diff is that key actually written. The command will always hit the else branch and say "Nothing running." You need something like:

task = asyncio.create_task(self._run_agentic(...))
context.user_data["running_claude_task"] = task

before any await in the message handler. Without this, the entire interrupt mechanism is non-functional.

2. concurrent_updates(True) makes the race worse without the fix above

Enabling concurrent updates means multiple messages process simultaneously — right intent for interrupt support. But if /stop can't actually cancel the running task, you now have concurrent handlers stepping on each other with no abort path. Don't ship concurrent_updates(True) until the task lifecycle is wired up correctly.

---

🟡 Consistency / Design Issues

3. user_data vs chat_data — conflicts with PR #165

agentic_cleanup reads last_tool_message_ids and last_tool_chat_id from context.user_data. PR #165 (also open, also touches orchestrator.py) moved session state to chat_data. These two PRs will conflict on merge and produce inconsistent state. Agree on chat_data as the canonical store before either lands.

4. tool_log used as a mixed-type container

tool_log.append({"_tool_message_ids": tool_message_ids})

Injecting a sentinel dict into a list of tool call entries is a code smell — any downstream iterator has to defensively check for this magic key. Track tool_message_ids as a separate variable in the closure scope instead.

---

🟡 Documentation Drift

5. Verbose level 3 undocumented / CLAUDE.md out of date

CLAUDE.md config docs apparently still document verbose range as 0–2. Update to reflect level 3 and describe what it enables.

---

🟡 Process

6. Overlap with PR #148 (same author)

PR #148 is already open with very similar orchestrator.py changes from the same author. These will be nearly impossible to merge independently. Recommend either:

• Explicitly closing feat: enhanced UX with per-event messages and file handling #148 and noting this PR supersedes it, or
• Rebasing this on top of feat: enhanced UX with per-event messages and file handling #148 after that merges

As-is, whichever merges second will have significant conflicts.

---

🟢 Test gaps

No tests visible for:

• agentic_stop / agentic_cleanup handlers
• draft_streamer.py (private chat vs group fallback)
• html_format.py markdown→HTML edge cases (nested formatting, split threshold)

---

Minor

• except Exception: pass in _send_tool_msg and agentic_cleanup silently eats errors — log at WARNING with structlog
• agentic_stop and agentic_cleanup missing type hints on update/context — mypy strict violation

---

Summary: Streaming draft and HTML formatting look solid in concept, but /stop is broken as written (task never stored), and concurrent_updates(True) actively makes things worse without it. Fix the task lifecycle, resolve the user_data/chat_data conflict with #165, and sort out the #148 overlap before merging.

— Friday, AI assistant to @RichardAtCT (posted as @RichardAtCT — FridayOpenClawBot access pending)