PR Review
Reviewed head: ac9b6fe27d186bc79b1da6ad855f3b71c41b4e36

Summary

• Adds DraftStreamer to stream partial Claude responses to Telegram via the new sendMessageDraft API, showing tool activity and response text in real-time
• Integrates stream deltas from the SDK (include_partial_messages, StreamEvent handling) and wires into agentic_text for private chats only (opt-in via ENABLE_STREAM_DRAFTS)
• 27 unit tests covering accumulation, throttle, truncation, self-disable, and flush

What looks good

• Clean self-disabling pattern on API error — graceful fallback to normal delivery without crashing the request
• Tail-truncation with ellipsis prefix is correct for partial streaming (avoids abrupt cutoff mid-word being the last thing the user sees)
• private chat restriction is a sensible first-ship constraint; group threads need more testing with draft semantics

Issues / questions

1. [Important] src/bot/utils/draft_streamer.py — generate_draft_id uses hash() which is randomized per-process in Python 3 (PYTHONHASHSEED). Two rapid calls for the same chat_id are extremely unlikely to collide, but more importantly the & 0x7FFFFFFF | 1 mask could return the same ID on a process restart for a resumed session. Is draft continuity across restarts intentional or is a new ID on each restart fine? If fine, a simple secrets.randbits(30) | 1 is cleaner and removes the hash dependency.

2. [Important] src/bot/orchestrator.py (_make_stream_callback) — when draft_streamer is active and verbose_level >= 1, tool lines go to the draft but NOT to tool_log. On error/fallback the progress message won't contain any tool activity. Is that intentional? If the draft streamer disables mid-run (e.g. sendMessageDraft flakes), the user may see nothing until the final response.

3. [Important] src/bot/orchestrator.py (agentic_text) — draft_streamer.flush() is called in finally but exceptions are silently swallowed. If flush() raises (e.g. the bot object is torn down), the suppression is fine, but a logger.debug inside the except would help diagnose silent draft gaps in production.

4. [Nit] src/bot/utils/draft_streamer.py — _MAX_TOOL_LINES = 10 uses self._tool_lines[-_MAX_TOOL_LINES:] (shows most recent 10), but the "... +N more" prefix is prepended before those lines. The net display is "... +N more\nline1\n...\nline10", which reads a bit awkwardly if only 1–2 lines were trimmed. Consider the threshold before showing the overflow notice.

5. [Nit] src/config/settings.py — the blank line between agentic_mode and reply_quote was removed (looks unintentional, adds noise to the diff).

Suggested tests (if needed)

• Test that when DraftStreamer self-disables mid-stream, append_text / append_tool calls afterwards are no-ops (this path is implied but worth asserting explicitly)
• Test orchestrator fallback: when draft streamer disables, tool_log remains empty and the final progress message still shows something useful (or confirm empty is expected)

Verdict
⚠️ Merge after fixes — the self-disable graceful fallback is solid, but the tool_log gap on streamer failure (#2) and the draft ID semantics (#1) deserve a quick answer before merge.

— Friday, AI assistant to @RichardAtCT

PR Review (incremental — new commit since last review)
Reviewed head: 26d6e0a8a1c8eb62f2e3632365c5e57cc52f6f89

Summary

• Fix commit addressing prior review feedback: randomness, tool_log fallback, overflow UX, flush error logging

What looks good

• secrets.randbits(30) | 1 is the right call — drops the hash(f"...") fragility and the chat_id coupling that wasn't needed
• Always populating tool_log regardless of streamer state (orchestrator.py:710,723) correctly solves the fallback gap
• Debug-logging the flush failure instead of silently swallowing it is good hygiene

Issues / questions

1. [Nit] draft_streamer.py:_compose_draft — overflow >= 3 suppresses the "...+N more" notice for 1–2 overflow lines, but those lines are still being dropped from visible (you still slice to [-_MAX_TOOL_LINES:]). For overflow=1 or 2, the user sees the last 10 but silently loses earlier lines with no indicator at all. Might be fine UX-wise, but worth a comment explaining the intent.

Verdict
✅ Ready to merge — the fixes are clean and the new tests (small overflow + mid-stream disable) cover the changed behaviour well.

— Friday, AI assistant to @RichardAtCT

Merged manually after rebasing to resolve conflicts with voice transcription PR. Included in main.