ops-memory

> STOP. READ THIS ENTIRE SKILL.MD BEFORE CALLING ANY COMMAND. # ops-memory The memory/database setup is no longer opaque. `ops-memory` is the one place to ask "is /memory healthy, how is it built, and what is quietly rotting?" — and to take a backup. It **composes the owning skills**; it never touches ArangoDB or Qdrant directly. ## Boundary (non-negotiable) - **No direct DB access.** ops-memory shells out to `ops-arango` (Arango admin + read-only `coverage`), `ops-qdrant` (read-only vector health), and `memory` (`/recall`). It has no `arango`/`qdrant` client and no `httpx` of its own. - **Detection + read-only backups only.** `backup --now` and `fix --apply` are the only side effects, and both are explicit, gated subcommands. - **Semantic-sync repair is memory-repo work.** `fix` plans/triggers the sanctioned memory-owned primitive `scripts/validation/dewey_embedding_repair.py` (one collection+operation lane at a time, with a rollback manifest); it never reimplements sync. Flag→op map: `embedding_array_violation`→`inline-vectors`, `bm25_only`→`missing-qdrant-embeddings`, `partial_sync`→`qdrant-pointer-metadata`. - All child calls use argument lists (never `shell=True`); every child payload is validated by a typed dataclass seam (`ArangoCheck`, `ArangoCoverage`, `QdrantHealth`) and drift raises `SeamViolation` (fail closed). ## Command surface ``` ops-memory health [--json] # merged ArangoDB + Qdrant health ops-memory metrics [--json] [--sample N] [--stale-days D] # per-collection matrix + flags ops-memory topology [--chart] [--json] # stack wiring; --chart via phart-dag-chart ops-memory explain "<question>" # NL router over the above ops-memory recall "<query>" [--k N] # passthrough to memory /recall ops-memory backups [--json] # list 12TB ArangoDB backups + retention ops-memory backup [--now] # create a dump on the 12TB drive (ops-arango) ops-memory fix [--collection N --operation OP --apply] # plan/run one Dewey repair lane ops-memory config doctor [--json] # non-interactive readiness ``` ## The metrics matrix (why this skill exists) `metrics` joins per-collection facts from `ops-arango coverage` (counts, index inventory, **ArangoSearch view membership**, named-graph membership, LIMIT-bounded Qdrant-pointer sampling, index-backed latest timestamp) with `ops-qdrant` health, and derives one row per collection. Each row reports its `recall_lanes` — any of **bm25** (an ArangoSearch view links it), **dense** (it has Qdrant vectors), **graph** (edge / named-graph member reachable by multi-hop traversal) — plus a `recall_connected` boolean and these flags: | Flag | Meaning | |---|---| | `not_recall_connected` | **no bm25, dense, or graph lane — invisible to `/memory recall`** | | `no_arangosearch` | document collection linked in no ArangoSearch view — no BM25/text lane | | `bm25_only` / `no_qdrant_embedding` | document collection with **0%** Qdrant pointers — no dense/semantic lane | | `partial_sync` | some, not all, docs vector-synced (0% < frac < 90%) | | `no_secondary_index` | only the primary/edge index — no query index | | `slow_scan_risk` | ≥50k docs **and** no secondary index — query-performance risk | | `embedding_array_violation` | Arango is holding embedding arrays (contract breach; see best-practices-arangodb rule 22) | | `empty` | zero documents | | `stale` | latest **indexed** timestamp older than `--stale-days` | The report also carries a `recall_connectivity` summary (`connected`, `not_connected`, `non_empty`). Staleness is asserted **only** where a timestamp index makes it cheap to prove; otherwise the row reports `latest_timestamp: unknown` (never fabricated). Edge collections are exempt from `bm25_only`/`no_arangosearch` (they carry no embeddings and no text view) and are counted recall-connected through the graph lane. ## Backups on the 12TB drive `ops-arango dump` already writes to `/mnt/storage12tb/backups/arangodb` with retention (keep last 7). ops-memory surfaces that as one call: ``` ops-memory backups # what backups exist, sizes, newest first ops-memory backup --now # take a fresh dump on the 12TB drive now ``` ## Progressive disclosure - Layer 1: this frontmatter (routing). - Layer 2: this SKILL.md (command map + boundary). - Layer 3: `scripts/ops_memory.py` (typed seams, flag logic), plus `docs/PROJECT_KNOWLEDGE.md` (verified topology + known gaps). ## Readiness Complex orchestrator. `sanity.sh` is the offline behavioral + safety gate (topology/config-doctor contracts, no-direct-DB-access, no `shell=True`, and a fail-closed negative control). `sanity-e2e.sh` is the opt-in **live** gate — it calls the real ops-arango/ops-qdrant/phart entrypoints, asserts the merged schemas and `seam_validation=PASS`, and fails closed on a missing downstream receipt. Never mark a feature READY on exit-0 alone. See `fixtures/agentic_eval.json` for the positive/negative/adversarial cases.