skill-tree-reconcile

Skill Tree Reconcile

Use this procedure for "sync my skills", "reconcile the skill trees", "deploy skills to live", "which tree wins", "live vs vault vs repo", "sync-from-live", or "my skills drifted".

Establish the model

Read writing-skills for the full four-tree model. In brief, REPO .agents/skills is the source of truth; scripts/sync-from-live.ps1 writes VAULT ~/.agents/skills to REPO; scripts/sync-skills.ps1 writes repo skills/ to VAULT; and scripts/deploy-skills-daily.py writes REPO to Claude LIVE ~/.claude/skills.daily. The Codex hop is python scripts/deploy-skills-codex.py, which delivers only ci/codex-skill-allowlist.txt names from REPO to ~/.codex/skills.

Reject bulk direction

There is no safe bulk direction. Of 99 substantive LIVE-vs-VAULT text pairs, 95 DIVERGED, 4 were LIVE-supersets, and 0 were VAULT-supersets. Scope this honestly: 41 of the 95 have at most three unique lines per side. DIVERGED means "not a clean superset", not 95 files of two-sided authored work. The heavyweight set is about 20 files with more than 10 unique lines on both sides.

Do not use timestamps. Of 337 LIVE SKILL.md files, 136 share one bulk-copy mtime. "Repo commit newer" holds for 86 of 95 differing files, including all eight corrupted files. gem-hunt and re-ensure are LIVE-supersets even though mtime says VAULT is newer.

Classify, then reconcile

class	direction	note
the 8 vendor-corrupted skills	LIVE -> REPO/VAULT, ignore dates	information-destroying substitution is one-way
skills where LIVE has a negative-case clause VAULT lacks (34)	LIVE -> REPO/VAULT	this is the active anti-overlap work
4 strict LIVE supersets + ~75 orphaned live-only files	LIVE -> REPO
CRLF-only differences (34 of 135)	either; exclude from effort	25% of the surface, zero content risk
~20 heavyweight DIVERGED files	no automatic direction; per-file union merge only	both sides hold unique lines
VAULT as a whole	rebuild from REPO after the above; never a merge source	VAULT carries the most corruption (18) and the fewest disambiguation clauses

This is a human-driven per-class procedure. Measured: scripts/sync-from-live.ps1 lines 128-138 expose no per-path include/exclude or direction selector and intentionally implement only live-to-repo. Review each class, edit REPO, preserve dual-write requirements, validate, then use the appropriate delivery hop.

Always exclude crypto-ctf explicitly: LIVE contains 12,816 files while git tracks 2; the extras are neither ignored nor staged. Letting it ride along buries the real fixes.

For CRLF questions, inspect the stored blob with git ls-files -s, git cat-file -p, or git show <rev>:<path>. On this workstation only git archive is the CRLF trap (it returns worktree bytes). git show <rev>:<path> and git cat-file -p both return the LF blob. Measured: three tracked files with 127/32/72 worktree CRLF pairs (.agents/skills/010-editor/SKILL.md, .agents/skills/README.md, .agents/skills/acceptance-review/README.md) returned 0 CRLF from both blob routes; git archive matched the worktree CRLF count on each.

All fleet measurements above come from command-center/SKILL-FLEET-SOTA-PLAN.md:361-405 and were measured for that plan. Almost certain (90-99%): the direction rules reflect those recorded probes; re-measure before a future destructive reconciliation.

What this skill does NOT do

It does not choose a winner from dates, bulk-sync diverged files, merge heavyweight files automatically, prune any live tree, deploy non-allowlisted Codex skills, or make direct edits in ~/.claude, ~/.agents, or ~/.codex.