|ClaudeOps|DEPLOYED
Conductor: better goal extraction + last-message outcome signal
Goal extractor: - Pull <command-args> body when the first message is a slash command (was capturing the framework caveat instead of the user's prompt).
Goal extractor:
- Pull <command-args> body when the first message is a slash command
(was capturing the framework caveat instead of the user's prompt).
- Strip <local-command-*>, <system-reminder>, <command-name>,
<command-message> wrappers from plain prompts.
- Walk to the next user message if the current one has no real content.
Outcome:
- Surface last_assistant_message from the most recent Stop event as
last_message in /api/conductor/sessions and detail endpoint.
- Show under the goal in the table as "↳ <last words>" so the row
answers both "what was it trying to do?" and "where did it leave off?".
- Dedicated section in the detail drawer.
Backfill:
- regoal_sessions.py walks ~/.claude/projects/*/<session>.jsonl, runs
the new extractor, and overwrites session_goals.first_prompt for any session whose existing goal looks like junk. Already ran: 117 sessions updated with real prompts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Files Changed
Commit:76092d8