@neat.is/claude-skill 0.4.16 → 0.4.18

package/SKILL.md

@@ -4,21 +4,35 @@ This skill exposes NEAT's live semantic graph to Claude Code over MCP. Once inst
 
 ## What you get
 
-Nine MCP tools, served by `@neat.is/mcp` over stdio:
+Sixteen MCP tools, served by `@neat.is/mcp` over stdio — ten read-only graph queries plus six `/neat extend` tools for instrumentation. The canonical list lives in `MCP_TOOL_NAMES` (`@neat.is/types`); the server registrations are the source for every description below.
+
+### Read tools
 
 | Tool | What it does |
 |------|--------------|
-| `get_graph` | Snapshot of the full graph for a project — every node, every edge, with provenance. |
-| `get_node` | One node by id, with its incoming and outgoing edges. |
-| `get_dependencies` | Transitive closure of `DEPENDS_ON` / `CONNECTS_TO` from a starting node. |
-| `get_root_cause` | Walks incoming edges from a failing node and returns the first divergence — typically a version mismatch or a config gap. |
-| `get_blast_radius` | BFS outbound from a starting node — every service / database / config that would feel a change here. |
-| `get_recent_errors` | Last N error events, ordered by `lastObserved`. |
-| `semantic_search` | Embedding-based search over node names + descriptions. |
-| `check_policies` | Runs the project's `policy.json` rules against the live graph. Returns active violations. |
-| `get_compatibility` | Compatibility matrix lookups — what's known about a `<driver, engine>` pair. |
-
-All nine read from the live graph the daemon maintains in memory. No fs reads of `graph.json` at request time.
+| `get_root_cause` | Trace a failing node up its dependency graph to the underlying cause. Use when something is breaking and you want the upstream culprit. |
+| `get_blast_radius` | List every node downstream of a node — what would break if it failed or was redeployed. |
+| `get_dependencies` | Transitive outgoing dependencies, BFS to depth N, each carrying distance, edge type, and provenance (EXTRACTED vs OBSERVED). |
+| `get_observed_dependencies` | Only the runtime (OBSERVED via OTel) outgoing dependencies — compare what code declares against what production does. |
+| `get_incident_history` | Recent OTel error events recorded against a node, most recent first. |
+| `get_divergences` | Places where the code (EXTRACTED) and production (OBSERVED) disagree, ranked by confidence × severity. The most NEAT-shaped query — reach for it on "is anything weird?" |
+| `get_graph_diff` | Diff a saved graph snapshot against the current live graph — added/removed/changed nodes and edges. |
+| `get_recent_stale_edges` | Most recent OBSERVED → STALE transitions — integrations that have gone quiet. |
+| `check_policies` | Inspect or dry-run the project's `policy.json`. Returns current violations, or violations a hypothetical action would cause. |
+| `semantic_search` | Search nodes by natural-language query (embedding vectors when available, substring fallback otherwise). |
+
+### Extend tools (`/neat extend`, ADR-081 / ADR-086)
+
+| Tool | What it does |
+|------|--------------|
+| `neat_list_uninstrumented` | List libraries that need instrumentation beyond the auto-instrumentations bundle. |
+| `neat_lookup_instrumentation` | Look up the registry entry for a library — canonical instrumentation package, version, registration snippet. |
+| `neat_describe_project_instrumentation` | Describe the current OTel state: which hook files exist, whether `.env.neat` is present, which OTel deps are installed. |
+| `neat_dry_run_extension` | Preview what an apply would do — the exact file diff, deps to add, install command — without changing anything. |
+| `neat_apply_extension` | Install an instrumentation package and splice its registration into the OTel hook file. Idempotent. |
+| `neat_rollback_extension` | Undo the last apply for a library — removes the dep and registration. |
+
+The ten read tools read from the live graph the daemon maintains in memory. No fs reads of `graph.json` at request time. The extend tools modify instrumentation files, `package.json`, and the lockfile only; NEAT never calls an LLM and the agent reasons over their output (ADR-084).
 
 ## Install
 
@@ -54,7 +68,7 @@ This merges the `neat` server into `~/.claude.json` without touching other entri
 
 - Auto-detection of an alternate Claude Code config path. The installer assumes `~/.claude.json`.
 - Per-project skill overrides. The skill is user-scoped; project-level MCP config can be added later as a follow-up.
-- Tool-level disable flags. All nine tools are wired in; if you want to hide one, edit the snippet by hand.
+- Tool-level disable flags. All sixteen tools are wired in; if you want to hide one, edit the snippet by hand.
 
 ## Where to look when it doesn't work
 

package/claude_code_config.json

@@ -5,7 +5,7 @@
       "command": "npx",
       "args": ["-y", "@neat.is/mcp"],
       "env": {
-        "NEAT_API_URL": "http://localhost:8080"
+        "NEAT_CORE_URL": "http://localhost:8080"
       }
     }
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@neat.is/claude-skill",
-  "version": "0.4.16",
+  "version": "0.4.18",
   "description": "Claude Code skill drop-in for NEAT — wires the @neat.is/mcp server into Claude's MCP config",
   "license": "BUSL-1.1",
   "homepage": "https://neat.is",