Session Start Routine

Session Start Routine

Pre-flight checklist for every new session. Run these checks before beginning work.

Checklist

1. Load context

• Load memory context files from the repo memory directory
• Check MEMORY.md for active project state and recent feedback

2. Check prior session state

• Read today's session signals: .claude/state/session-signals/YYYY-MM-DD.jsonl
• Look for sessions that ended mid-task (incomplete commits, unreleased wip: labels)
• Check git status for uncommitted work from prior sessions

3. Check for in-flight work

• Scan for wip labels on GitHub issues (gh issue list filtered by wip label)
• Check other terminals: ps aux | grep claude | grep -v grep
• Check /tmp/.claude-wip-* markers if present

4. Validate environment

• Verify tools: uv --version, gh auth status, git status
• Check disk space: df -h /mnt/local-analysis
• Verify governance hooks are registered in settings

5. Check governance limits

• Review tool-call counter: .claude/state/session-governor/tool-call-count
• Confirm counter is below 200 ceiling (or reset if new day)

When to use

• Start of every interactive session
• After a context reset or conversation compression
• When resuming work after a break
• When the user asks for a terse live "status update" from mobile/Telegram

Mobile status-update mode

When the user asks only "Status update" (especially from Telegram), produce a compact evidence-backed snapshot rather than a broad narrative.

1. Gather live state before summarizing:
   • git status --short --branch in the active repo.
   • Active/background agent processes or Hermes process-manager state.
   • hermes cron/cronjob list for active, paused, failed, and next scheduled jobs.
   • GitHub counts for wip, status:plan-review, status:plan-approved, and recent closed issues.
   • Most recent relevant cron output files when scheduled jobs produced user-facing findings.
2. Separate counts clearly:
   • live label counts are not the same as artifact-audited approval readiness;
   • closed/recent results are not the same as currently running lanes;
   • paused autofeed monitors are not active throughput.
3. For Telegram/mobile delivery:
   • avoid pipe tables; use numbered bullets and labeled key/value lines;
   • link GitHub issues as [#{number}](url);
   • keep to the shortest useful update, with risks and next best action last.
4. Never clean/reset/stage dirty control-plane files during a status-only request; report dirty state and recommend a narrow reconciliation pass if needed.