emmy-knowledge-store

name	emmy-knowledge-store
description	Use when an Emmy task may benefit from the shared Confluence knowledge store: product context, architecture, ATO, onboarding, runbooks, decisions, troubleshooting, stale context, prior team knowledge, or durable task learnings that should be captured or published through cms-atlassian-confluence.

Emmy Knowledge Store

Use this skill to help agents work with the Emmy Agent Knowledge Store as a shared team memory. The default posture is: consult freely, capture proactively, publish deliberately.

Source Of Truth

• Entry point: https://confluenceent.cms.gov/spaces/SFIV/pages/1381557075/Emmy+Agent+Knowledge+Store
• Confluence space: SFIV
• MCP server: cms-atlassian-confluence
• Required user environment variable: CONFLUENCE_PERSONAL_TOKEN

Dynamic Store Map

Treat the entry-point page as the only stable root. Do not depend on a hard-coded list of child pages. At the start of a knowledge-store task, inspect the current entry-page children and use their titles, labels, short page content, and own children as routing metadata.

Load the smallest useful slice of the store:

1. Read the entry page or get its current children.
2. Build a lightweight map of top-level child pages without content first.
3. Read only likely candidate pages, using search and labels when hierarchy alone is not enough.
4. Recurse into children when a page acts like an index, queue, registry, category, source collection, or canonical topic area.
5. Treat newly added child pages as valid store structure without requiring a skill update unless the source-of-truth root, MCP server, or write policy changes.

For broad navigation, duplicate detection, or page-organization questions, read references/search-and-labels.md. For whole-store cleanup, duplicate audits, or deciding whether to add new child pages, use $emmy-knowledge-cleanup.

Ambient Loop

1. Consult current Confluence content when the task may depend on Emmy product, architecture, ATO, operational, onboarding, troubleshooting, or prior decision context.
2. Apply what you learn to the user's task. Cite page titles and URLs or page IDs when Confluence content materially shaped the answer or implementation.
3. Capture durable knowledge discovered during the task: verified decisions, reusable facts, canonical links, runbooks, failure resolutions, or resolved ambiguity.
4. Publish or propose the capture. Only write to Confluence when the user explicitly asks to persist, update, curate, or capture knowledge-store content. Otherwise, summarize the proposed capture in the final answer.

Read references/ambient-protocol.md when deciding whether to consult, propose, queue, or publish knowledge.

When To Consult

Consult the knowledge store before making claims about current Emmy context, prior decisions, architecture, ATO state, team runbooks, onboarding steps, or operational practices. Also consult when local files, memory, or copied notes may be stale or conflict with a current team source.

Do not consult for trivial local-only edits, purely mechanical formatting, commands that do not depend on Emmy context, or tasks where the user explicitly asks you not to use external knowledge.

For broad or uncertain searches, read references/search-and-labels.md.

Specialized Source Ingestion

Do not use this ambient skill for source ingestion from local files, URLs, external Confluence pages, or Emmy repository checkouts. When the user explicitly asks to ingest a source into the knowledge store, use $emmy-artifact-ingest or the emmy-artifact-ingestor plugin agent.

When answering from knowledge that came through the ingestion workflow, preserve source provenance in citations or summaries. For repository-derived knowledge, re-check the local repository evidence when branch, commit, or dirty-worktree applicability matters.

Specialized Queue Curation

Do not use this ambient skill for interactive review of pending Knowledge Ingestion Queue entries. When the user asks to process, review, curate, verify, or promote queue entries with a human reviewer, use $emmy-queue-curate or the emmy-queue-curator plugin agent.

Queue curation validates one needs-review entry at a time and promotes only after explicit final human approval. If the reviewer does not know, wants a teammate to answer, or says the entry needs discussion, the v1 curator should pause without writing Confluence discussion artifacts.

Capture And Write Policy

Durable knowledge includes verified team decisions, stable procedures, reusable troubleshooting findings, canonical source links, current architecture or ATO facts, and context that would help another agent avoid repeating the same investigation.

Never store secrets, tokens, personal credentials, speculative claims, transient chat notes, local scratch paths, private user data, or unverified conclusions.

Writing rules:

• No silent canonical updates. Canonical pages may be updated only when the user explicitly asks to update or persist knowledge-store content.
• Low-confidence, contested, or not-yet-curated knowledge belongs in an Agent Capture Queue entry when the user authorizes persistence.
• Create one capture page per candidate instead of appending many agents' notes to the same page.
• Before any Confluence write, read references/write-safety.md.
• For multiple agents or concurrent developers, read references/fleet-agent-protocol.md.
• When drafting page content or a capture candidate, use references/knowledge-entry-template.md.

Runtime Requirements

• uvx must be available on PATH so the plugin can run uvx mcp-atlassian.
• The agent environment needs network access to https://confluenceent.cms.gov.
• Each user must provide their own CMS Confluence Data Center personal access token in CONFLUENCE_PERSONAL_TOKEN.
• The shared Atlassian MCP config also forwards JIRA_PERSONAL_TOKEN when present, sets JIRA_URL to https://jiraent.cms.gov, and includes the Jira tool allowlist for future Jira-specific workflows.
• This shared MCP config is a collision-safe superset. Follow this skill's narrower tool-use policy; MCP tool availability is not permission to use Jira from this workflow.

If the MCP server is unavailable or Confluence authentication fails, stop and tell the user that the plugin needs a valid CONFLUENCE_PERSONAL_TOKEN in the agent environment. Do not ask the user to paste tokens into the chat.

MCP Tool Use

Start from the entry-point page and inspect its current children unless the user gives a more specific page URL, page ID, title, or Confluence query. Use Confluence MCP tools such as confluence_search, confluence_get_page, confluence_get_page_children, confluence_get_page_history, and confluence_get_page_diff for reading.

Use confluence_create_page, confluence_update_page, confluence_add_comment, confluence_reply_to_comment, confluence_get_labels, confluence_add_label, confluence_upload_attachment, confluence_upload_attachments, confluence_get_attachments, confluence_download_attachment, confluence_download_content_attachments, and confluence_get_page_images only when they fit an authorized write, comment, label, or attachment task.

Do not use Jira tools for this skill, even though the shared MCP config exposes them. Do not delete, move, or reorganize pages unless a future plugin or an explicit task adds that capability.