Use for the named Markdown-plus-Milvus memory project. Read pinned source facts before enabling capture, indexing or reset.
Instructions
- Confirm the repo/commit against
references/source.json. For a checkout, from repo root runpython3 .agent-skills/memsearch/scripts/preflight.py --repo /absolute/nonsymlink/checkout --json;--repomust be an explicit absolute nonsymlink directory. For installed skill-root inventory, from that skill directory runpython3 scripts/preflight.py --json. This standalone stdlib helper requires Python 3.10+ on POSIX (macOS/Linux; use WSL2 on Windows because native Windows fails closed) and makes no subprocess/network calls, writes, or secret reads. It compares SHA-256 of selected pinned files and checks PATH command-name presence only (not versions); it never compares Git HEAD/the whole revision or verifies runtime. Exit 0 means valid input and selected-file checks pass (command presence is informational); 1 means a selected file is missing/mismatched; 2 means bad input. Without--repo, it reports unverified inventory. - Inspect current Markdown root, Milvus URI/collection, config precedence and plugin capture state. Preserve Markdown as canonical data; treat Milvus as derived index.
- Choose exactly one operation.
index,watch, plugin capture,compact, local model fetch andresethave distinct data/provider effects. Do not run a destructive or background operation as a diagnostic. - Require explicit approval before installation/plugin hooks, network/model downloads, enabling capture/watch, indexing or writing memory, remote embedding/LLM calls,
compact, or deleting/resetting indexed data. Confirm target files, provider, cost and collection. Protect private TOML and never print secrets. - Verify by inspecting resulting state only after authorized action. State separately whether Markdown changed, index changed, provider was contacted, and hooks/service were activated.
Examples
- “Explain how memsearch separates Markdown memory from its Milvus index.”
- “Index this private notes folder with remote embeddings.” Pause for approval of data, provider and cost.
- “Reset the Milvus collection.” Confirm exact URI/collection and explicit destructive approval; leave Markdown intact.
Best practices
- Keep canonical memory files and the derived vector index separate in reports and backups.
- Treat plugin capture and watchers as ongoing data collection, not a passive install.
- Project config is intentionally restricted; use the global config for endpoint/credentials only with care.
- Preflight proves file-pin integrity only, not index freshness or retrieval quality.
References
-
Offline preflight contract (added by repository maintainer)
-
Agent-memory architecture for vendor-neutral choices.
-
Authored routing and boundary scenarios, not executed model-quality results.