legatum

name	legatum
description	State-transfer for impending session death — bequeath context before /clear, or auto-fire at gear shifts (checkpoint mode). NOT for graceful session-end (use telophase) or day-level close (eow/daily).
user_invocable	true

Legatum — Session State Transfer

A dying session bequeaths its knowledge forward. Two modes:

Core framing: This is state transfer, not documentation. The session holds volatile state — decisions, understanding, corrections — that dies with /clear. The legatum's job is to move what matters to durable storage. Everything else is theatre.

• Full (/legatum or /wrap) — end-of-session close. All steps.
• Checkpoint (/legatum checkpoint or /wrap checkpoint, auto-triggered at gear shifts) — capture learnings + sweep TODOs, skip session-end bookkeeping. Preserves context — no /compact, no closure framing.

Session scope = files modified + tool calls + conversation turns since this session began (or since last checkpoint).

Triggers

• "wrap", "wrap up", "let's wrap", "legatum" → full mode
• "checkpoint", "wrap checkpoint" → checkpoint mode
• Auto-trigger (Claude-initiated): When detecting a significant gear shift (different project, different domain, switching from building to admin, etc.), run checkpoint mode silently before proceeding. Don't ask — just capture.
• "what did we learn" → checkpoint mode

Mode Detection

If invoked as checkpoint or auto-triggered at a gear shift → checkpoint mode. If invoked at session end → full mode.

Checkpoint mode runs: Step 0 (pre-wrap), Step 0.5 (friction review, but don't truncate log), Step 1 (TODO sweep), Step 4 (learning capture). Checkpoint mode skips: Step 2 (session log), Step 3 (NOW.md rewrite — delta update only if needed).

Workflow

Skip gate

now-age

If NOW.md is <15 minutes old AND user did not explicitly invoke /legatum, skip to Step 4 briefly, then Output. Explicit invocation always runs all steps.

Step 0: Pre-Wrap Check

Run prewrap and answer these five questions. Complete blocking actions (garden post, arsenal) before outputting the block — the block is a receipt, not a plan.

prewrap
peira status 2>/dev/null || true

Questions (explicit yes/no for Q4–5, silence is not "no"):

1. Unverified? Any tool output this session that wasn't checked?
2. Deferred? Anything mentioned as "later/next/TODO" not yet captured? Route by type: has a deadline → TODO.md. Has a context trigger ("next time I'm in X") → memory/priming.md. Neither → daily note.
3. Uncommitted? Dirty repos touched this session? → offer to commit (leave other repos alone)
4. Garden posts + consulting arsenal? Pause and replay the session arc (or arc since last checkpoint). What did we learn, not just do? What surprised us? What principle emerged that wasn't obvious at the start? Give yourself 30 seconds of generative thinking before answering — the best posts come from connections between topics, not from any single task.
   • Garden test: Non-obvious insight, clear thesis, Terry's lane, no unverified facts? Publish immediately via publish new → write → publish publish --push. Multiple posts per session is normal for meaty sessions.
   • Arsenal test: Concretely applicable to a bank/client AI engagement? If yes → add bullet to [[Capco Transition]] now.

CLAUDE.md modified? One-line tightening check: does it belong in CLAUDE.md or in a skill / MEMORY.md / ~/docs/solutions/?

Background dispatches — fire with run_in_background: true when the session touched the relevant area:

Audit	When
MEMORY.md hook coverage	MEMORY.md modified AND <145 lines
Skill staleness	Any skill edited or added
Solutions KB dedup	~/docs/solutions/ modified
Vault orphan links (nexis)	Monthly only

Output — light or full:

If all checks clean and no blocking actions: ✓ Clean — [prewrap summary]. Garden: no — [reason]. Arsenal: no — [reason].

Otherwise, full block:

─── Pre-Legatum ─────────────────────────────────
⚠  [only if action needed]
→  Deferred: [items or "none"]
✓  [clean checks summary]
Garden:      published → <slug>, <slug>, ... | no — [reason]
Arsenal:     added → [[Capco Transition]] | no — [reason]
Dispatched:  <audit> (<task-id>) | none
──────────────────────────────────────────────────

Then proceed to remaining steps.

Step 0.5: CLI Friction Review

cat ~/.claude/cli-friction.jsonl 2>/dev/null | wc -l

If ~/.claude/cli-friction.jsonl has entries: read the file, group errors by CLI tool, and for each tool with 2+ friction events (or 1 event with an obvious fix), suggest a concrete improvement (alias, positional arg, better error message). Output as a fenced block. If any fix is trivial (< 20 lines), implement it or add to TODO.md with agent:claude. Full mode: truncate the file after processing (> ~/.claude/cli-friction.jsonl). Checkpoint mode: leave the file intact (accumulate across checkpoints, truncate only at session end).

Step 1: TODO Sweep

Read ~/epigenome/TODO.md. Skip if missing.

• Complete: Done items → [x] with note and done:YYYY-MM-DD. Hard test: truly done, or just "dev done"? Move checked items to ~/epigenome/TODO Archive.md.
• Create: New commitments or interrupted WIP → add with verb + concrete next action. Tag agent: if Claude can resume.

Step 2: Session Log (full mode only)

Append to ~/epigenome/chromatin/Daily/YYYY-MM-DD.md:

### HH:MM–HH:MM — [Brief title]
- Key outcome or decision (1-3 bullets)
- Abandoned: X because Y  ← if a path was explored and dropped

Step 3: NOW.md + Trackers (full mode only)

now-age

NOW.md — read from disk first. If recent (<1h), update only deltas. If light session (<3 files, no decisions) and still accurate, skip.

Max 15 lines. Resume points must pass cold-start test. Use [decided] vs [open]. Prune [decided] items that no longer gate future action.

Vault flush: Update canonical tracker notes (e.g. [[Capco Transition]]) if the session advanced them.

Project CONTEXT.md — if cwd is ~/code/<project>/ and meaningful progress was made, update State/Last session/Next/Open questions sections. Commit after writing.

Step 4: Learning Capture (always runs)

Single pass. If nothing surfaces: "Nothing to generalise."

Scope: Since last checkpoint (if any), otherwise since session start.

Failure mode check (scan before writing):

Smell	Fix
"Updated X" with no what/why	State what changed and why — cold reader test
"TODO: consider whether..."	Decide now or delete. Legatum is not a parking lot
>3 "filed to X" items	Over-capturing. Prioritise
Recording what happened, not what was learned	Logs → daily note. Insights → skill/memory/garden
Legatum takes >10 minutes	You're writing, not closing. Stop
"Fixed the bug" without generalising	Capture the pattern, not the instance

Belief corrections? Did a prior get challenged this session — something Terry (or I) assumed that turned out wrong? Not just career — technical assumptions, how things work, self-knowledge, anything. If yes → append to [[Priors Worth Correcting]] with: old belief, the correction, and when the old belief would reassert.

Scan → Route → Implement:

• Scan for non-obvious patterns, friction, corrections, gotchas, and new user preferences or personal context
• Route each finding to the most specific destination:
  • Tool gotcha → ~/docs/solutions/
  • Cross-session context → MEMORY.md
  • Workflow change → the relevant skill's SKILL.md
  • Same mistake twice → escalate per ~/docs/solutions/enforcement-ladder.md
• Default: implement now. Skill edits, MEMORY.md additions, solutions files, small hooks — do them, don't propose them. "Needs design input" is not a valid reason to defer a 20-line improvement. Propose only if: touches shared infrastructure, irreversible, or genuinely ambiguous (and state which). If you wrote "propose" in the output, ask: could I have just done it? If yes, go back and do it.
• Persistence gate: Consult custodia before routing. One home per insight, never point to a pointer, count layers.

MEMORY.md ≥145 lines + entries added this session → demote lowest-recurrence entry to ~/docs/solutions/memory-overflow.md now. Don't ask — pick it yourself.

Decay tracker: If any MEMORY.md entries prevented mistakes this session, update memory/decay-tracker.md with today's date. This is the empirical signal for what to keep vs demote.

Legatum audit: log, don't essay. Instead of launching verbose audit agents, do a quick self-scan and append one-liners to ~/docs/solutions/operational/wrap-violations.jsonl. Pattern detection happens in /weekly, not per-session.

Self-scan checklist (30 seconds, no agents):

1. Did any CLAUDE.md rule get violated AND cause a worse outcome? (Not "technically violated but fine")
2. Did anything compound (insight, tool, framework) that wasn't captured?
3. Were garden posts published? If 3+, flag for quality cull.
4. Ad-hoc script spiral? Were 5+ consecutive tool calls spent working around a broken CLI/tool (different keychain lookups, manual API calls, alternative scripts for the same goal)? If yes → file the root-cause fix as a priming entry or TODO, and log to wrap-violations.jsonl with "rule": "ad-hoc-spiral".

If a violation caused harm: append to wrap-violations.jsonl:

{"date": "YYYY-MM-DD", "rule": "rule name", "harm": "what went wrong", "hookable": true/false}

If something compounded but wasn't captured: just capture it now (skill, garden, arsenal, vault). Don't log it — act on it.

Garden cull: If 3+ posts published this session, launch one haiku agent: "Review these posts. Flag weak ones (no thesis, generic, restating others) for removal or merge." Present in output.

Garden quality cull (weekly only): If this is a /weekly session, also scan all posts from the week — not just this session's.

/weekly pattern detection: Read wrap-violations.jsonl, group by rule. Any rule with 3+ violations → escalate per enforcement ladder. Any rule with 0 violations over 4 weeks → candidate for removal (dead rule).

All file writes must complete before the output.

Output

Full mode: Bordered prose. Handoff note to tomorrow-you — arc, what's staged/unfinished, learnings captured. 2-3 sentences for light sessions, up to 6 for heavy. Don't hard-wrap.

─── Legatum ────────────────────────────────────

[Prose summary]

Filed: [exact file path or "nothing to generalise"]
Session: [honest 1-line judgment — real output vs theatre, what moved vs what just felt productive]
─────────────────────────────────────────────────

Checkpoint mode: Lighter border. What was captured, then move on. No handoff framing.

─── Checkpoint ─────────────────────────────────
[1-2 sentences: what was captured/filed]
─────────────────────────────────────────────────

Boundaries

• Do NOT perform external sends (messages, emails, posts) during legatum.
• Do NOT run deep audits or long research — legatum is a close-out, not a new workstream.
• Full mode: Stop after writes + summary unless explicitly asked to continue.
• Checkpoint mode: Continue with the next task after output. No stopping.