dsh-graphify-skill

A DeepSeek Harness bundle that installs graphify's knowledge graph into an agent session as two parts that are set up together: the MCP server that can query the graph, and the skill that tells the agent when and how to query it.

What you get

Part	Where it lands	What it does
MCP server entry	DSH loader tree (this bundle's cordis.patch.yml)	Registers graphify's stdio server as @deepseek-ai/dsh-mcp-client. The tools appear as mcp__graphify__query_graph, mcp__graphify__get_node, mcp__graphify__get_neighbors, mcp__graphify__shortest_path, mcp__graphify__get_community, mcp__graphify__god_nodes, mcp__graphify__graph_stats, and the three PR tools mcp__graphify__list_prs / get_pr_impact / triage_prs. Six graphify:// resources come with them.
graphify skill	<DSH_HOME>/skills/graphify/SKILL.md (+ references/)	Tells the agent to treat a codebase question as a graph query first when graphify-out/ exists, which tool to call first, how to read EXTRACTED / INFERRED / AMBIGUOUS edges, and how to fall back to the graphify CLI when the server is down.

Both halves come from this repository, so they stay on one version. Nothing is published to npm: the package is a bundle, and its install-time files are cordis.patch.yml plus skills/, which dsh plugin add (a pnpm install of the git repo) already puts on disk.

Install

Install the MCP entry first, then the skill beside it — both in the same step, so the skill and the server describe the same graphify release. <checkout> is the profile's copy of this repository (~/.dsh/profiles/<profile>/node_modules/dsh-graphify-skill).

# 1. The MCP entry and its future updates
dsh plugin --profile <profile> add github:Ravomi/dsh-graphify-skill

# 2. The skill into DSH's skill root, from the copy just installed
CHECKOUT="${DSH_HOME:-$HOME/.dsh}/profiles/<profile>/node_modules/dsh-graphify-skill"
mkdir -p "${DSH_HOME:-$HOME/.dsh}/skills"
cp -R "$CHECKOUT/skills/graphify" "${DSH_HOME:-$HOME/.dsh}/skills/graphify"

Then install the graphify runtime itself, which the server runs:

uv tool install 'graphifyy[mcp]'

graphifyy ships the graphify-mcp entry point (a stdio server equivalent to python -m graphify.serve); the bundle calls it by name.

Serving a graph

The stdio child inherits the active workspace as its working directory, so with no configuration the server serves <workspace>/graphify-out/graph.json — the graph that /graphify or graphify extract . wrote for the repository you are in.

Two environment variables override that, for a graph somewhere else:

Variable	Default	Meaning
GRAPHIFY_GRAPH	unset (workspace-relative)	Absolute path to a graph.json the server serves instead.
GRAPHIFY_MCP_COMMAND	graphify-mcp	Program to launch. Set it when graphify lives in a virtualenv that is not on PATH, e.g. /home/me/.local/share/uv/tools/graphifyy/bin/graphify-mcp.

Start DSH from the repository you want to explore, or set GRAPHIFY_GRAPH before starting it.

Build a graph first

There is nothing to query until the graph exists:

cd /path/to/repository
graphify extract . --code-only    # AST only, no API key, no LLM calls

Drop --code-only to also index docs, papers, and images (that pass sends them to a model). The skill and the tools are both useless against an empty graph — the server answers "graph.json not found" rather than inventing an answer.

Verify the install

In a DSH session:

• the tool list contains mcp__graphify__graph_stats — call it; it returns node and edge counts;
• the skill catalog contains graphify with the description from skills/graphify/SKILL.md.

If the tools are missing, gs in the DSH logs names the failed server; the two usual causes are graphify-mcp not on PATH (set GRAPHIFY_MCP_COMMAND) and the workspace having no graphify-out/graph.json (build it, or set GRAPHIFY_GRAPH).

Updating

• MCP entry: dsh plugin --profile <profile> update dsh-graphify-skill.
• Skill: re-run step 2 after the update, so the skill body catches up with the bundle. skills/graphify/ is the canonical copy in this repository; the copy under ~/.dsh/skills is a deployment of it, never the other way round.

Uninstall

dsh plugin --profile <profile> remove dsh-graphify-skill
rm -rf "${DSH_HOME:-$HOME/.dsh}/skills/graphify"

License and provenance

The skill text and the references under skills/graphify/ are graphify's packaged skill-agents.md and its references/ sidecar, reproduced from Graphify-Labs/graphify under Apache-2.0 (see LICENSE). The bundle glue — package.json's dsh.bundle and cordis.patch.yml — is this repository's own work.