Session Init Skill

A skill for checking the environment and understanding the current task status at session start.

Invocation Conditions

This skill is called internally from session / SessionStart-type flows. The user-facing entry point is /session or the normal session start flow.

Legacy trigger phrases:

• "Start session"
• "Begin work"
• "Start today's work"
• "Check status"
• "What should I do?"
• "start session"
• "what should I work on?"

Overview

The Session Init skill automatically checks the following at Claude Code session start:

1. Git state: Current branch, uncommitted changes
2. Plans.md: In-progress tasks, requested tasks
3. AGENTS.md: Role assignments, review of prohibitions
4. Previous session: Review of carry-over items
5. Latest snapshot: Summary of progress snapshot and diff from previous session

Execution Steps

Step 0: File State Check (Automatic Cleanup)

Check file sizes before session start:

# Check Plans.md line count
if [ -f "Plans.md" ]; then
  lines=$(wc -l < Plans.md)
  if [ "$lines" -gt 200 ]; then
    echo "⚠️ Plans.md has ${lines} lines. Recommend organizing with 'organize it'"
  fi
fi

# Check session-log.md line count
if [ -f ".claude/memory/session-log.md" ]; then
  lines=$(wc -l < .claude/memory/session-log.md)
  if [ "$lines" -gt 500 ]; then
    echo "⚠️ session-log.md has ${lines} lines. Recommend organizing with 'organize session log'"
  fi
fi

Display a suggestion if cleanup is needed (does not affect work).

Step 0.5: Legacy Local Memory Compatibility Handling (Optional)

The current standard is the Unified Harness Memory in Step 0.7. Checking legacy local memory compatibility is generally unnecessary; reference it individually only for special migration checks.

Note: In normal operation, skip this step and treat the common DB Resume Pack as the sole resume entry point.

Step 0.7: Unified Harness Memory Resume Pack (Required)

Retrieve resume context from the Codex / Claude / OpenCode shared DB (~/.harness-mem/harness-mem.db).

Required call:

harness_mem_resume_pack(project, session_id?, limit=5, include_private=false)

Operational rules:

• Always specify current project name for project
• Obtain session_id in order: $CLAUDE_SESSION_ID → .session_id in .claude/state/session.json
• Use of the first result from harness_mem_sessions_list(project, limit=1) is limited to read-only (resume check); do not use for writes via record_checkpoint / finalize_session
• Inject retrieval results into session start context
• On retrieval failure, check daemon state with harness_mem_health(), note the failure, and continue
• Recovery order: scripts/harness-memd doctor → scripts/harness-memd cleanup-stale → scripts/harness-memd start

Step 1: Environment Check

Run the following in parallel:

# Git state
git status -sb
git log --oneline -3

# Plans.md
cat Plans.md 2>/dev/null || echo "Plans.md not found"

# AGENTS.md key points
head -50 AGENTS.md 2>/dev/null || echo "AGENTS.md not found"

Step 2: Understand Task Status

Extract from Plans.md:

• cc:WIP - Tasks continuing from previous session
• pm:requested - Tasks newly requested by PM (compatible: cursor:requested)
• cc:TODO - Unstarted but assigned tasks

Step 3: Output Status Report

## 🚀 Session Started

**Date/Time**: {{YYYY-MM-DD HH:MM}}
**Branch**: {{branch}}
**Session ID**: ${CLAUDE_SESSION_ID}

---

### 📋 Today's Tasks

**Priority tasks**:
- {{pm:requested (compatible: cursor:requested) or cc:WIP tasks}}

**Other tasks**:
- {{List of cc:TODO tasks}}

---

### ⚠️ Notes

{{Key constraints/prohibitions from AGENTS.md}}

---

**Ready to start work?**

Output Format

At session start, present the following information concisely:

Related Commands

• /work - Task execution (parallel execution supported)
• /sync-status - Plans.md progress summary
• /maintenance - Automatic file cleanup

Notes

• Always check AGENTS.md: Understand role assignments before starting work
• If Plans.md is missing: Guide to /harness-init
• If previous work was interrupted: Confirm whether to continue