Architecture Memory Update

---
name: architecture-memory-update
description: Update or seed opted-in architecture memory from Git.
---

Explicit requests only. Resolve `<workflow>` to `<plugin-root>/skills/architecture-memory/scripts/workflow.js` from this file. Append `--project-root <project>` to each command.

Run `node <workflow> update` before code or document reads. `current` and `ignored` finish without edits; report the unchanged checkpoint. `applying` resumes with `apply`. For `draft`, reuse saved classifications and the returned draft; committed evidence is fixed at `source_revision` even when HEAD or working files change.

Run `changes` for pending paths. Classify each affected responsibility: boundaries, runtime/storage units, selected components, deployment, integrations, quality constraints and decision artifacts. Use `source --path <path> --diff` for relevant hunks and `source --path <path>` for captured context. With a null checkpoint, use the [initialization evidence pass](../architecture-memory-init/references/initialization.md#evidence-pass) on the captured inventory, reconcile current-state documents, and reconstruct no ADR history.

Locate affected memory with [retrieval](../architecture-memory/references/retrieval.md). Edit only its corresponding draft sections under [recording](../architecture-memory/references/recording.md). Preserve user reasons and operating facts; code changes establish neither acceptance nor rationale. Update stable routing paths on renames. Use [decision templates](../architecture-memory/references/decision-templates.md) only for an explicit rationale-bearing decision in the evidence or current conversation. Non-ancestor checkpoints describe snapshot differences, not ordered history.

Persist each classification with `classify --path <path> --effect architecture|none --reason "<short evidence-based reason>"`. Use `--prefix <directory/>` for a known homogeneous group, including both sides of renames; a directory name alone does not establish irrelevance. Remaining paths disappear from subsequent `changes` pages only when classified. Keep unresolved scope pending.

Once the draft reflects every affected responsibility, run `apply`. It validates and publishes the draft, then advances the captured checkpoint, including when classification requires no document edits. Do not rewrite checkpoint fields or repeat the same structural check separately. Preserve document verification fields on partial edits.

For paging, custom roots and failures, read [workflow](../architecture-memory/references/workflow.md). Report changed topics and checkpoint in the user's language. Report incomplete publication accurately; resume it before claiming completion.