Session Start Skill

Run at the start of every session. Surfaces state from the previous session and orients the orchestrator before any work begins.

When to Use

• First action in every new session (before responding to any work request)
• After /clear (context was wiped — re-orient)
• When returning from a multi-day break

Steps Claude Follows

1. Readiness Report (from last session’s stop hook)

Read .claude/state/readiness-report.md. If it contains ## Warnings, surface each warning to the user with the suggested fix. If “All Clear”, note it briefly.

2. Session Snapshot (from /save before last /clear)

Read .claude/state/session-snapshot.md. If the snapshot is less than 48 hours old (check the timestamp in the file), surface the ## Ideas / Notes section and the WRK summary. If older than 48h, skip.

3. Quota Status

Read config/ai-tools/agent-quota-latest.json. Surface the weekly utilization for each provider:

• =90%: warn (route tasks to alternative provider)

• 70-89%: note (approaching limit)
• <70%: no comment needed

If the cache file is older than 4 hours, note that quota data may be stale.

4. Top 3 Unblocked Items

Scan .claude/work-queue/pending/ for items where blocked_by: [] (or no blocked_by field). Show the top 3 by priority (high > medium > low), then by ID (lowest first). For each item show: ID, title, complexity, tags (first 2).

5. Computer Context

If a computer: field exists in recent working/ items or in .claude/state/, note which machine was last active. Prompt user to confirm if working on a different machine today (multi-machine handoff check).

Output Format

## Session Briefing — 2026-02-19

**Readiness:** [All Clear | N warnings listed]
**Snapshot:** [Summary of Ideas/Notes if fresh | None]
**Quota:** Claude 11% | Codex 43% | Gemini 0%

**Top 3 unblocked items:**
1. WRK-XXX — Title (complexity, tags)
2. WRK-YYY — Title (complexity, tags)
3. WRK-ZZZ — Title (complexity, tags)

Path Discipline Reminder

Surface this reminder silently at session start (do not print unless a violation is detected):

• Prefer relative paths inside repo-scoped scripts (e.g., ./scripts/foo.sh, not /mnt/...)
• Never hardcode /mnt/local-analysis/workspace-hub/ in generated scripts — use ${REPO_ROOT}, $(git rev-parse --show-toplevel), or $(pwd) instead
• Absolute paths are allowed only when a tool explicitly requires them (e.g., file_path in Read/Edit/Write tool calls)
• If you catch yourself writing a hardcoded /mnt/ path in a script, stop and replace it with the env-var form before proceeding

Hook Integration

This skill is also invocable as a startup context. Add to CLAUDE.md:

At the start of every new session, run /session-start to orient before responding.