Summary

• _log_history_lines() was entirely commented out and returned '', so no message history was ever logged in debug mode
• _log_format_message_line() referenced message.metadata.tokens but BaseMessage has no metadata attribute — the old wrapper-object data model it relied on no longer exists
• The function also called _log_extract_message_content() which was removed in a prior refactor

Restores both functions using the current BaseMessage API (.text property + MessageHistory.get_messages()). Token counts are estimated via the standard ~4 chars/token heuristic; the ~ prefix makes the approximation explicit. Actual billed token counts remain tracked by the TokenCost service post-response as before.

Fixes #4150

Changes

• service.py: add missing import shutil (was referenced in commented-out code but never imported)
• _log_format_message_line: replace '??? (TODO)' with f'~{len(content) // 4}'.rjust(5)
• _log_history_lines: fully restored using get_messages() / m.text; dead references to old m.metadata / m.message / _log_extract_message_content removed

Testing

Enable logging.DEBUG and run any agent task. Verbose logs now show:

📜 LLM Message history (3 messages, ~1842 tokens):
🧠[ ~312]: You are a browser automation agent...
💬[~1420]: Current page: github.com | Task: find open...
🔨[ ~110]: I'll click the search bar first...

Summary by cubic

Restores verbose message history logging and shows approximate token counts (~4 chars/token) in debug output. Fixes #4150.

• Bug Fixes
  • Re-implemented _log_history_lines using history.get_messages() and BaseMessage.text; respects terminal width.
  • Shows "~" token estimates per message and total; actual billing still via TokenCost.
  • Removed stale metadata/helper refs and added missing shutil import.

^{Written for commit 4dcf983. Summary will update on new commits.}