emmy-artifact-ingest

Emmy Artifact And Repository Ingest

Use this skill for the specialized knowledge-source ingestion workflow. Do not use it for ordinary knowledge-store lookup, general Confluence updates, or ambient capture suggestions.

The default ingest mode is Review Queue: create source records and needs-review knowledge entries. Promote to canonical pages only when the user explicitly asks for promotion or names a canonical destination page.

Source Of Truth

• Knowledge-store entry point: https://confluenceent.cms.gov/spaces/SFIV/pages/1381557075/Emmy+Agent+Knowledge+Store
• MCP server: cms-atlassian-confluence
• Source registry area: Knowledge Source Registry
• Ingestion queue area: Knowledge Ingestion Queue

Supported Inputs

Classify the source first:

• Local file: upload the file as a Confluence attachment, extract knowledge, and cite the source record plus file metadata.
• URL: extract directly from the reachable content and do not upload a copy.
• External Confluence page: read through Confluence MCP, record page metadata, and do not upload a copy.
• Emmy repository checkout: extract from a local Git checkout and do not upload files. Treat the local checkout as the lookup source, not the github.com URL.

Read references/source-ingest-protocol.md before ingesting any source.

Provenance Contract

Every extracted claim must be traceable to one source record. Every source record must link forward to all queue entries or canonical pages created from that source.

For repository-derived claims, include repo, branch, commit, relative_path, line_range or symbol, and applicability. Read references/branch-applicability.md before writing repository-derived knowledge.

Use references/source-provenance-template.md for source records and queue entries. Use references/project-repositories.md to resolve known Emmy repository identifiers and local checkouts.

Ingest Workflow

1. Confirm the user explicitly asked to ingest knowledge into the store.
2. Classify the input as local file, URL, external Confluence page, or repository checkout.
3. Inspect the knowledge-store entry page children for the source registry and ingestion queue.
4. Create missing registry or queue pages only when the user authorized persistence.
5. Create or reuse a source record for the source.
6. Extract only evidence-backed, reusable knowledge.
7. Create needs-review queue entries by default, linking each claim to the source record.
8. If canonical promotion is explicitly requested, run the normal emmy-knowledge-store write-safety workflow before updating the page.
9. Re-read or diff changed Confluence pages and report source record plus queue or canonical page URLs.

Stop rather than guessing when a source cannot be accessed, a local repository checkout cannot be found, Confluence authentication fails, or the source appears to contain secrets or sensitive data that should not be stored.

Tool Use

Use Confluence MCP tools for page search/read/create/update/comment/label and attachment upload tasks. Use local shell and structured parsers when available to inspect files or Git repositories. Do not use Jira tools. Do not delete, move, or reorganize Confluence pages.