Vanta Context Sync

Keep Vanta's context files true to the code after every meaningful change. Stale docs cost more tokens than no docs — a subagent that trusts a wrong file map wastes a whole round-trip. This skill is the discipline that stops drift between what the code does and what the docs claim.

When to run

• Finished a build slice or phase (and it verifies — tests pass, typecheck clean)
• Added or removed a module, tool, provider, or directory level
• Changed a command, env var, or API surface
• Locked a design decision
• Hit a gotcha worth recording for next time

Do NOT run for

• Work in progress — wait until the slice actually verifies
• Pure refactors with no structural or decision change
• Trivial edits (typo, comment) that don't change the file map or behavior

Checklist (create a TodoWrite item per step)

1. Verify reality first. ls the changed directories and grep the actual exports, routes, and commands. Update docs from what is, never from memory. A file map listing a deleted file is worse than none.
2. Nearest-level CLAUDE.md. Update the file map, decisions, and gotchas for the layer you touched — vanta-ts/CLAUDE.md for agent work, root CLAUDE.md for kernel work. One line per file/fact.
3. New directory level → new CLAUDE.md. If you created a new layer or package, add a CLAUDE.md there: one-line role, file map, key decisions, how-to-extend, commands. Link up to the parent.
4. Root CLAUDE.md status + maps. Update the ## Status line (current phase, test count, what's next) and the kernel module/API tables if the kernel changed.
5. README "What works now." Only if the change is user-facing. Keep the run commands runnable — actually run them.
6. PRD checkboxes (docs/prd.md). Tick the phase done-criteria you met. If scope shifted, note it explicitly — don't rewrite the roadmap silently.
7. Locked decisions. Add to the layer CLAUDE.md "Key decisions (don't re-litigate)" section: the choice + a one-line why. This is what stops future re-debate.
8. Project memory. Update MEMORY.md index + project_argo.md with new state, new gotchas, and the next step. Convert relative dates to absolute.
9. Self-check. Re-read each edited file: does the map match ls? Do the commands run? Any reference to a renamed or deleted thing? Fix inline.

Density rules (non-negotiable — token economy is the whole point)

• Tables and one-liners over prose. No "this section describes…" filler.
• Decisions get a why; facts don't.
• If a CLAUDE.md exceeds ~120 lines, it's carrying explanation that belongs in code comments or the PRD — trim it.
• Global ~/.claude/CLAUDE.md conventions are inherited, never repeated.
• The test: could a fresh subagent do the next slice from these files without opening more than 2-3 source files? If not, the docs are incomplete. If they're reading docs longer than the code, they're bloated.