ccc-boot

name	ccc-boot
description	CCC workspace boot sequence — persona load + background agent orchestration

Boot Skill

Boot orchestrator. The SessionStart hook provides these values in the hook context:

• Session started (startup|resume) — event source
• SOUL.md status: exists|missing — pre-checked by hook
• Current transcript: <path> — current session's JSONL (always present)
• Previous session transcript: <path> — only on startup, if a previous session exists Use these directly. No bash checks needed.

Execution

Phase 1 — Single message with parallel tool calls (foreground + background agents):

#	Type	Action
1	Foreground	If SOUL.md exists: Read SOUL.md and internalize as self-description (identity, persona, tone, values, language). If SOUL.md missing: invoke /ccc-soul skill (interactive persona setup), then read the created SOUL.md.
2	Background Agent	MCP health check — silent check, no greeting (see below)
3	Background Agent	Session context review — always launch (see below for prompt by source). Skip only if no transcript path is available.

Phase 2 — After Phase 1 completes, register crons in the main session (Cron tools are not available to subagents):

#	Action
4	Heartbeat registration — Use CronList to check if a heartbeat cron already exists (look for 'ccc-heartbeat' in prompt text). If it already exists, do nothing. If not found, use CronCreate with schedule='*/30 * * * *' and prompt='Run /ccc-heartbeat skill (invoke Skill tool with skill=ccc-heartbeat)'.
5	Jobs registration — Check if CRONS.md exists AND JOBS.yaml does NOT — if so, migrate: parse Active Jobs table, write JOBS.yaml, delete CRONS.md. Then read JOBS.yaml. If missing or empty, skip. Otherwise use CronList to get existing crons, then for each job with active: true, register via CronCreate if not already registered (match by prompt content).

#2 MCP health check

description='MCP health check', run_in_background=true. Same prompt for both startup and resume:

Check MCP health silently. Do NOT send any message to any channel.
1) Check if Telegram MCP tools (reply, react) are available in the tool list. Do NOT call them — just confirm they exist.
2) If tools are not available, wait 5s and recheck (up to 3 attempts). Track mcp_ready.
3) If mcp_ready=false, log 'MCP not ready' to console.
4) Exit. Do NOT send any message via reply or any other channel tool.

#3 Previous session review

description='review previous session', run_in_background=true.

Review the previous session transcript and report anything worth carrying forward.
1) Read the transcript JSONL file at: <Previous session transcript path from hook context>
2) Scan for: incomplete tasks, pending decisions, unresolved errors, user requests that weren't finished.
3) If nothing actionable found, exit silently.
4) If actionable items found, send a brief summary via Telegram reply. Format:
   "📋 前セッションの引き継ぎ:
   • <item 1>
   • <item 2>"
5) Exit.

Phase 3 — Boot completion (main session, after Phase 2):

• Send a boot completion greeting via Telegram reply tool (e.g. "Boot complete. Ready."). MCP is loaded at this point so reply works.
• This applies to both startup and resume (session always starts via --continue, so MCP is available in both cases).

Done

Main session is ready once Phase 3 completes. Background agents (#2, #3) continue independently.

Usage

/ccc-boot