| name | Trim |
| version | 1.0.4 |
| description | Reduces an always-on LifeOS context file that has grown too big via a human-gated pass — deterministic GC of stale entries first, then semantic merges and relocations — never dropping a directive and committing every change reversibly. USE WHEN /trim, trim the context, trim OPERATIONAL_RULES, this file is too big, reduce a doctrine file, prune an always-loaded file, fold the proposal inbox, shrink CLAUDE.md or DA_IDENTITY. NOT FOR general code refactoring, trimming video or audio media (use AudioEditor for audio files), or removing AI writing patterns from prose. |
Trim
Shrink an always-on context file that has gotten too big. /trim <file> walks the reduction, safest cuts first, never dropping a rule.
Workflow Routing
| Trigger | Workflow |
|---|
/trim <file>, "trim OPERATIONAL_RULES", "this file is too big", "reduce a doctrine file", "fold the proposal inbox" | Workflows/Trim.md |
Quick Reference
- Target resolution: a bare name (
OPERATIONAL_RULES) resolves against the always-on set — the system prompt, CLAUDE.md, its @-imports, and the hook-injected memory files. No arg → wc -c that set and take the largest.
- Order is safest-first: (1) show state, (2) deterministic GC (zero-risk), (3) semantic trims (human-gated), (4) safety gate, (5) re-measure. Full steps:
Workflows/Trim.md.
- One tool it orchestrates — never reimplement:
LIFEOS/TOOLS/ProposalGC.ts (removes superseded/duplicate/absorbed entries). Sizes come from wc -c.
- Three semantic moves: MERGE overlapping rules, TIGHTEN verbose ones, RELOCATE rarely-used detail to an on-demand reference (leave a stub + pointer).
- The invariant: a trim never drops a distinct directive. If a merge would, keep the original.
Gotchas
- USER files commit to the USER_DATA repo, not
~/.claude. LIFEOS/USER/** (OPERATIONAL_RULES, PROJECTS, the identity files) is a symlink into a separate private repo. Commit with git -C ~/.config/LIFEOS/USER …. A ~/.claude commit captures nothing under LIFEOS/USER/ — a false safety net.
- The file can change mid-edit. The autonomic memory loop appends proposals to these files while you work. If a Write/Edit reports "modified since read", RE-READ before writing — a concurrent correction may have landed (this is how a real deploy-command fix was nearly reverted). Never write from a stale read.
- Semantic merges must never drop a directive. Before applying any merge/tighten, confirm every proper noun, path, tool name, and imperative from the originals survives in the result. If one is missing, the merge is wrong — keep the original. Deterministic GC (superseded/dup/absorbed) is always safe; semantic edits are the risky class.
bun/bunx only, never npm/npx.- Deterministic first, always. Run ProposalGC before proposing any semantic edit — the free, zero-risk removals often clear enough that no judgment-call edit is needed.
Examples
/trim OPERATIONAL_RULES
# → shows 40,456 B → ProposalGC dry-run (0 removable) → ranks semantic trims (fold the
# 40-entry proposal tail, relocate skill-scoped directives to a reference)
# → applies approved ones behind the safety gate → commits to USER_DATA → 28,043 B (−31%)
/trim
# → no arg: wc -c the always-on set, take the largest, then the same walkthrough
Execution Log
After completing the workflow, append a single JSONL entry:
echo '{"ts":"'$(date -u +%Y-%m-%dT%H:%M:%SZ)'","skill":"Trim","workflow":"Trim","input":"8_WORD_SUMMARY","status":"ok|error","duration_s":SECONDS}' >> ~/.claude/LIFEOS/MEMORY/SKILLS/execution.jsonl
