start

/start — cold-boot vjt-claude

Brings up the persistent IRC session from a fresh claude launch. This is the cold path, run manually. The post-/clear hot path is /resume (fired by the watchdog) — do not run /start there.

/start = cold-only setup (sections 1–3) then the shared warm-bringup protocol (section 4). The shared steps — reseed today's activity-log body, adopt/attach the Monitor, WIP sweep, report — live in one place: .claude/skills/resume/SKILL.md. /start does NOT restate them; it defers, so the two skills can't drift. Idempotent: safe to re-run mid-session.

1. Memory housekeeping — archive entries >14d (cold-only)

Never full-read the activity log. It can be hundreds of KB; a full Read burns input tokens for nothing. It is NOT auto-loaded (MEMORY.md holds only a one-line pointer), so this is the only place that burns it. Keep it cheap:

• grep -nE '^### [0-9]{4}-[0-9]{2}-[0-9]{2}' memory/project_activity_log.md — lists every ### YYYY-MM-DD heading with its line number. Cheap skeleton.
• Compare each heading against currentDate.
• Archive every day-heading + bullets older than 14 days. Per feedback_archive_dont_delete, move them to project_activity_log_archive.md (read-on-demand, not in MEMORY.md) — don't discard. Only read the body of the days being trimmed (Read with offset/limit around their line range), never the whole file.

This >14d archive sweep is the one thing /resume deliberately skips (the log is read-on-demand, so an untrimmed log costs zero per-clear tokens — it's disk hygiene, not token hygiene). That's why it lives here, on the cold path, not in the hot path. The today's-heading reseed is NOT done here — it's part of the shared protocol (section 4), so it isn't done twice.

Then lint the memory index (cold-only): run bash memory/check-index.sh. It reports MEMORY.md byte budget, over-length index lines, dead pointers, and true orphans (files neither indexed nor [[linked]]). If it flags anything, fix per feedback_memory_anti_clutter — trim fat descriptions to hooks, group rosters into one pointer, index/link orphans. MEMORY.md is auto-loaded every session and the loader truncates it past ~24.4KB, so keeping it lean is what stops silent index loss.

2. Ensure all three services are up — systemd --user (cold-only)

As of 2026-04-20 bot.py plus the watchdog sidecar run as systemd user services with loginctl enable-linger vjt, so they auto-start at boot, auto-restart on crash, decoupled from Claude Code lifecycle.

• vjt-claude-bot.service — the IRC bridge (bot.py). RestartSec=30 + StartLimitBurst=5/300s to survive NickServ reconnect storms. ExecStop sends QUIT :systemd stop via FIFO for a clean disconnect.
• vjt-claude-aup-watchdog.service — tails the active Claude Code jsonl, injects /clear into the claude tmux pane on AUP refusal / TURNS cap / idle, follows every /clear with a /resume scrub prompt. See project_aup_watchdog.md.

Unit files live in-repo at systemd/, copied into ~/.config/systemd/user/. Check and (idempotently) start:

systemctl --user is-active vjt-claude-bot.service vjt-claude-libera-bot.service vjt-claude-aup-watchdog.service
# any not "active":
systemctl --user start vjt-claude-bot.service vjt-claude-libera-bot.service vjt-claude-aup-watchdog.service

Nothing to mkfifo / nohup / pgrep — systemd owns all three. Structured logs via journalctl --user -u <service>.

3. Greet on own JOIN (cold-only)

Home-channel JOINs (#sniffo, #olografix) + ChanServ INVITE #it-opers are declared in bot.startup and fired by the bot after NickServ identifies. Do NOT re-send them from the skill — redundant. See project_bot_auth_and_startup.md.

The skill owns the greet on own JOIN event — when the Monitor shows JOIN vjt-claude <chan>, send one short line, channel-register appropriate:

• #sniffo, #it-opers, default → Porco Dio register (e.g. SAY #sniffo porco dio raga, vjt-claude in sala).
• #olografix → supercazzola, less blasphemy (e.g. SAY #olografix tarapìa tapìoco come se fosse antani...).

See project_greet_on_join.md.

4. Run the shared warm-bringup protocol (/resume)

Now execute the four steps documented in .claude/skills/resume/SKILL.md, in order:

1. Reseed today's activity-log heading body (append today's heading if missing, read its body in full — the mandatory seed for today's events / in-flight threads / PENDING markers).
2. Adopt or attach ALL THREE Monitors — Azzurra, Libera (two separate bots, each its own stdout stream) AND Telegram (@gazzurbo_bot long-poller). pgrep the surviving tail -F … bot.stdout.log + … bot.libera.stdout.log pipelines and tg_poll.sh; adopt each if present, else attach fresh via start-monitor.sh / start-monitor-libera.sh / var/tg_poll.sh. Libera and Telegram usually do NOT survive /clear — attach them.
3. WIP sweep — git status --short + bot.log inbound/outbound dedup; resume only true gaps.
4. Report ready.

On a true cold boot these usually find nothing to adopt (fresh attach) and no WIP (clean sweep) — but run them anyway: a manual idempotent re-run of /start mid-session may find a surviving Monitor or pending work. Follow /resume's rules exactly (especially: when in doubt between replying and silence on the WIP sweep, stay silent; a dupe is worse than a miss). Read that file for the exact commands and edge cases — it is the single source of truth for these steps.

Standing context

After bringup, the session is expected to:

• Follow the reply policy (feedback_irc_reply_policy.md — reply only when addressed or truly adding value).
• Apply channel registers (feedback_olografix_tone.md, project_sniffo_supercazzola_on_join.md, global CLAUDE.md's Porco Dio principle).
• Append to the activity log on meaningful events (decisions, shipped code, people appearing, plans formed). NOT on casual banter.
• Clown vjt on irssi ESC+N misfires (bare digit + "ops sorry" — feedback_mock_esc_number_misfire.md).

The parent CLAUDE.md covers all of this — read it if unsure.

Sending to IRC

Commands go into the FIFO /home/vjt/code/IRC/vjt-claude/bot.send. Verbs: SAY, ACT, NOTICE, JOIN, PART, WHOIS, QUIT, RAW.

printf 'SAY #sniffo tarapìa tapìoco, porco dio!\n' > /home/vjt/code/IRC/vjt-claude/bot.send

Relaunching from scratch

Always start fresh (plain claude, NOT claude --continue). The architecture bootstraps from disk — CLAUDE.md, MEMORY.md, activity log — not from conversational history. --continue would carry over lossy compaction summaries and defeat the sliding-window design. The jsonl archives preserve deep history regardless.

Shutting down

Use /close for a graceful exit (supercazzola QUIT, Monitor stop, memory consistency check). Run it before /exit when you're actually killing Claude Code. If just leaving the session running, skip.