Add hull integrity monitoring and relief on station

[harrymunro]

Summary

Ships in a Nelson squadron now have damage control for the one thing that actually sinks AI agents: running out of context window.

The admiral monitors hull integrity across the entire squadron by reading exact token counts from Claude Code session JSONL files. No estimation heuristics, no paid APIs, no external dependencies — the numbers come straight from the API usage data that Claude Code already records on every assistant turn.

Each subagent (ship) has its own JSONL file at {session-id}/subagents/agent-{agentId}.jsonl, separate from the flagship's session file. The admiral scans the whole directory in one pass to build a squadron readiness board — a single view of every ship's context health.

When a ship's hull integrity drops to Red, the admiral orchestrates relief on station — a tactical handover borrowed from Royal Navy patrol doctrine. The damaged ship writes a rich turnover brief to file (not a message, to keep the replacement's context clean), a fresh ship reads it and picks up exactly where the other left off. The flagship monitors itself too, and can hand over to a new session before it founders.

What's new

3 damage control procedures:

• Hull integrity — Four-tier threshold system (Green/Amber/Red/Critical) with a squadron readiness board the admiral maintains at each quarterdeck checkpoint
• Relief on station — Full handover sequence including flagship self-monitoring, chained reliefs (A → B → C with bounded cumulative briefs), and a crew variant for captain-level recovery
• Session hygiene — Clean start procedure that clears stale damage reports and turnover briefs, with an archive option for previous sessions

2 admiralty templates:

• Damage report — JSON format for hull integrity reporting, stored at .claude/nelson/damage-reports/{ship-name}.json
• Turnover brief — Rich handover document with standard and flagship variants, including relief chain history

1 utility script:

• count-tokens.py — Three modes of operation:
  • --session — Read a single JSONL file for exact token counts
  • --squadron — Scan a session directory for the flagship + all subagent JSONL files and produce a full readiness board
  • --file — Heuristic fallback for plain text (character count / 4)

SKILL.md integration:

• Session hygiene hooked into Step 1 (Issue Sailing Orders)
• Hull integrity checks added to Step 4 (Quarterdeck Rhythm)
• Turnover brief reference added to Step 3 (Battle Plan)
• Three new entries in the Damage Control table

The key insight

Claude Code already records exact token counts (input_tokens + cache_creation_input_tokens + cache_read_input_tokens) on every assistant turn in the session JSONL. We don't need to estimate anything — we just read the last assistant message's usage data.

Ships can't easily self-monitor (they don't know their own agent ID to locate their JSONL), but that's the correct pattern: the flagship monitors everyone. The admiral runs --squadron against the session directory at each quarterdeck checkpoint and gets the full picture.

Live readiness board from the session that built this PR:

Ship	Tokens	Hull	Status
Flagship	104,365	47%	Red
HMS Kent	26,952	86%	Green
HMS Argyll	29,341	85%	Green
HMS Daring	34,693	82%	Green
HMS Astute	57,269	71%	Amber

Test plan

• --session produces correct JSON with exact token counts from JSONL
• --squadron scans flagship + all subagent files and returns a readiness board
• --file falls back to heuristic for plain text
• Hull integrity thresholds trigger correctly (tested Green at 98%, Critical at 27%, Red at 47%)
• Subagent JSONL files are separate from flagship — no cross-contamination
• scripts/check-references.sh passes — all cross-references valid
• No existing files broken — only additive changes to SKILL.md and CLAUDE.md
• Live test in a Nelson mission with actual relief-on-station handover