PR Review
Reviewed head: 6658f4feb0f6b306a40618bc3d4f553c69e056e9

Summary

• Raises truncation thresholds so long assistant responses/code blocks are preserved more often.
• Aligns callback diff preview budget with Telegram’s 4096-char limit by increasing diff snippet cap.
• Updates formatter tests to validate split-vs-truncate behavior for larger code blocks.

What looks good

• src/bot/features/conversation_mode.py now treats truncation as an extreme fallback and relies on caller-level message splitting for normal cases, which matches the comment and intended flow.
• tests/unit/test_bot/test_formatting.py adds a useful distinction between truly huge blocks (truncate) and medium blocks (split), which is the right behavioral contract.

Issues / questions

1. [Important] poetry.lock — this PR changes lock metadata from python-versions = "^3.10" to "^3.11" and removes several py3.10-era deps. Is dropping Python 3.10 support intentional for this repo right now? If not, this lockfile refresh could unintentionally narrow runtime/CI compatibility.
2. [Important] src/bot/utils/formatting.py (ResponseFormatter.max_code_block_length) + src/bot/features/conversation_mode.py (max_content_length) — thresholds were increased significantly (3k→15k, 3k→50k). Could we add/confirm an integration test that exercises an end-to-end oversized response through the actual Telegram send path to ensure no edge case still trips message too long errors after HTML escaping / headers / keyboards are added?

Suggested tests (if needed)

• Add one integration-style test around the final outbound Telegram payload size (post-formatting, with HTML markup and any suffix text/buttons metadata) to verify no single outbound message exceeds platform limits.
• Add a regression test for handle_git_callback diff previews near 3500 chars including HTML-sensitive characters (<, >, &) to validate effective rendered length remains safe.

Verdict

• ⚠️ Merge after fixes