yaaif-kg-enhance

Files (1)

• SKILL.md

SKILL.md (first lines)

Call every Context Store tool through `yaaif_local_tool_call` with `name` set to the tool and the tool's fields under `arguments`. Surface tool errors verbatim and never simulate a result. If a `context_kg_*` call fails because the tool is unavailable, stop and say Context Store tools are not mounted on agent-service (CONTEXT_STORE_BASE_URL and the Context Store plugin must be active).

## Goal

Find new relationships between existing catalog IDs (`object_type:`, `mcp_tool:`, `mcp_server:`, `dimension:`, `metric:`) and add them as AI enrichment edges after the user approves. Never propose `skill:` nodes. Optional `concept:` nodes are allowed only when they label a discovered relationship, each needs at least one bridge triple to an existing catalog ID (`APPLIES_TO`, `ENABLES_OBJECT_TYPE`, `USES_TOOL`, `USES_SERVER`, `USES_FIELD`), and at most 3 per proposal. Prefer triples only.

Copy node IDs from tool output (`node_id`). Never invent IDs or plant/demo codes. Preferred predicates: `REFERENCES`, `READS_OBJECT_TYPE`, `WRITES_OBJECT_TYPE`, `BELONGS_TO_SERVER`, `HAS_DIMENSION`, `HAS_METRIC`, `RELATED_TO`. Every triple endpoint must already exist in the graph. Tag AI edges with `source: ai_enrichment` and a short rationale in metadata.

## Start

Ask once for `enrichment_mode` (`durable` survives Rebuild, `ephemeral` is wiped on the next Rebuild), default `durable`, and an optional scope (type, MCP server, field). Then start scanning with no further confirmation. Pass `enrichment_mode` to propose and apply.

## Phases

Call `context_kg_set_run_phase` on entering each phase so the Entity Graph page shows progress.

1. `scanning`: read `context_kg_get_data_model`, `context_kg_get_mcp_metadata`, `context_kg_get_type_navigation`, `context_kg_get_snapshot`, and use `context_graph_search`, `context_graph_neighbors`, `context_graph_query`, `context_graph_path` as needed. Look for FK-like fields without `REFERENCES`, tools without object-type edges, and types that share keys but lack navigation. Report real counts and 2-5 candidate gaps (subject, predicate, object).
2. `proposing`: call `context_kg_propose_enrichments` with the triples and a summary. Read `floating_ai_node_count`, `catalog_bridge_count`, and skill-node counts from the response. If `floating_ai_node_count > 0`, any `skill:` node appears, or unknown IDs were rejected, revise and re-propose. Do not ask for approval on an invalid proposal.
3. `awaiting_approval`: set the phase, show the proposal as a short list with the counts, and ask the user to approve, narrow the scope, or cancel. Stop and wait. On cancel set `done`.
4. `applying`: only after explicit approval, call `context_kg_apply_enrichments` with the `proposal_id` in batches of at most 25 nodes or triples. Summarize after each batch and continue while `remaining_*` is above 0.
5. `verifying`: call `context_kg_get_enrichment_diff`, summarize new edges, confirm floating and skill AI node counts are 0, set `done`, and tell the user to refresh the Entity Graph page.

## Rules

Additive only; never delete platform seed edges. Never claim completion without an apply after approval. Keep replies short: phase, bullets, next action.

From yaaif/cursor-plugin, Apache-2.0.

More from yaaif/cursor-plugin