context-graph: single-command installer for Claude Code - #296
Merged
Conversation
The Claude Code plugin registration step (claude plugin marketplace add + claude plugin install) is what actually wires hooks into Claude Code -- running `agent-context-graph bootstrap` alone does not, even though its `doctor` check comes back all green. That gap made the memgraph-platform ai-memory.sh demo's "wire it into a real harness" snippet look complete without actually capturing anything. install.sh folds the whole flow (start Memgraph if none is reachable, register the marketplace, install the plugin, install the CLI with all three connectors, set identity, verify with doctor) into one idempotent script. Verified locally end-to-end in a sandboxed $HOME/$CLAUDE_CONFIG_DIR (cold start and already-installed paths) and added a manual-dispatch CI job that runs it for real against the published marketplace/plugin. Along the way, fixed a real readiness race the local run surfaced: a plain TCP-accept check declares the MAGE image ready well before it can actually serve a Bolt handshake, so the immediately-following bootstrap/doctor step failed with a real connection error. Gated readiness on an actual mgconsole query instead, matching the pattern skills-graph/actions-graph's own CI jobs already use for the same image.
Turns out the docs' claim that Codex has "no non-interactive plugin-install step" was stale: codex-cli 0.147.0 has `codex plugin add <plugin>@<marketplace>`, which installs *and* enables the plugin in one non-interactive call (no auth needed, verified locally with CODEX_HOME sandboxed and no OPENAI_API_KEY set, same as the Claude Code path). Fixed that claim in the two READMEs that made it and in the plugin's own README. install.sh now branches on CONTEXT_GRAPH_RUNTIME (claude-code, the default, or codex) for the marketplace/plugin step; the rest of the script (Memgraph, uv, agent-context-graph bootstrap/doctor) was already runtime-agnostic. Verified both runtimes end-to-end in isolated HOME/CLAUDE_CONFIG_DIR/ CODEX_HOME sandboxes, plus a regression run of the default claude-code path. CI: test-context-graph-install.yaml now matrixes over both runtimes, installing @anthropic-ai/claude-code or @openai/codex per leg. OPENAI_API_KEY is wired in for the codex leg (already an available repo secret) even though it isn't actually required -- doctor's hook smoke test calls the runtime adapter in-process rather than shelling out to claude/codex, so it never touches a model either way.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
memgraph-platform PR #110's "wire it into a real harness" snippet (
uv tool install+agent-context-graph bootstrap+config set identity.user_id) looks complete —doctorcomes back all green — but it's missing the one step that actually wires hooks into Claude Code:claude plugin marketplace add+claude plugin install. Without it, the CLI is installed and Memgraph is reachable, but Claude Code never learns the plugin exists, so no real session writes anything.context-graph/scripts/install.sh(new) — one idempotent script: starts a local Memgraph if none is reachable, registers thememgraph/ai-toolkitmarketplace, installs thecontext-graphplugin (skipping steps already done), installsagent-context-graphwith all three connectors, sets identity (git config user.namefallback), and verifies withdoctor.context-graph/README.md— leads with the one-liner; the old 5-step manual flow is preserved in a collapsed<details>for people who want to see/run it by hand, plus a note that Codex has no non-interactive plugin-install step today so it isn't covered end-to-end..github/workflows/test-context-graph-install.yaml(new) — manual-dispatch job that installs the Claude Code CLI and runsinstall.shfor real against the published marketplace/plugin (noANTHROPIC_API_KEYneeded — plugin management and the bootstrap hook smoke test are both local/no-LLM).Bug found while testing
The initial version of
install.shgated "Memgraph is ready" on a plain TCP-accept check. That's not sufficient for the MAGE image: the socket opens well before Memgraph can actually complete a Bolt handshake (it's loading heavier ML query-module deps first), so the immediately-followingbootstrap/doctorstep failed with a realCould not connect to Memgraph databaseerror. Fixed by gating readiness on an actualmgconsolequery viadocker exec, matching the patternskills-graph's andactions-graph's own CI jobs already use for this same image.Test plan
install.shend-to-end locally in a sandboxed$HOME/$CLAUDE_CONFIG_DIR(so it wouldn't touch my real Claude Code config): cold start (no Memgraph, no marketplace, no plugin) — hit the TCP-vs-Bolt-ready race above, fixed it, reran clean: everydoctorlineOK(uv, memgraph, agent-context-graph, all three connectors,runtime:claude-codestrict hook smoke).bash -nclean.test-context-graph-install.yamlworkflow isworkflow_dispatch-only (mirrorstest-live-graph-model.yaml's convention: real external state, not something to run on every push); trigger it manually to verify on a fresh runner.