context-graph: test-graph-model — verify the graph model against a real Claude Code session (map #288)

[antejavor]

Summary

Implements the destination of wayfinder map #288: a repeatable way to verify, against a real Claude Code session (not synthetic e2e data), that a broad slice of the actions-graph/skills-graph model actually holds up.

• scripts/dev-memgraph.sh test-graph-model — drives a scripted, non-interactive claude -p session restricted to the Task tool (forcing genuine subagent delegation), where the one spawned subagent both reads a real SKILL.md file and searches the repo — one LLM session, eleven assertions:
  • Session -[:HAS_AGENT]-> Agent, Action -[:SPAWNED]-> Agent (the real spawning tool call), Agent -[:HAS_ACTION]-> Action (Grilling: formalize the subagent-nesting parent-inference rule (including parallel/batch tool calls) #277/Grilling: does a subagent instance become a first-class Agent node, or stay a flat Action pair with a populated PARENT_OF? #278)
  • FOLLOWED_BY: the subagent's own chain is correctly ordered, and never crosses into the top-level session's own separate chain (Grilling: does a subagent instance become a first-class Agent node, or stay a flat Action pair with a populated PARENT_OF? #278)
  • PARENT_OF: ToolCall→ToolResult correlation at both the top level and inside the subagent
  • USED_TOOL: the subagent's tool calls link to real Tool nodes
  • USED_SKILL: attaches to the Agent, not the Session, when the skill read happens inside a subagent (Grilling: does Skill Usage need to attach to the specific nested Agent, not just the flat Session? #280) — verified live for the first time
• .github/workflows/test-live-graph-model.yaml — manual workflow_dispatch job wrapping the same check. Deliberately not on every push/PR: it drives a real, billed LLM session, and subagent-spawning is inherently non-deterministic — worth having on tap for deliberate use when changing something heavily in this area, not a PR gate.

Two real bugs found, not just tooling built

Running this against an actual live session (rather than reasoning from docs) surfaced two genuine, previously-unknown gaps:

1. agent-context-graph hook run claude-code silently does nothing without explicit --connector skills-graph --connector actions-graph --connector sessions-graph flags — undocumented outside the installed marketplace plugin's own generated command.
2. actions-graph's agent_spawning_tool_names defaulted to {"Task"}, but real PreToolUse/PostToolUse hook payloads report the spawning tool as tool_name: "Agent" — "Task" is only how Claude Code's CLI/UI refers to it. This meant Tier 1 of Grilling: formalize the subagent-nesting parent-inference rule (including parallel/batch tool calls) #277's inference rule had never actually matched anything against a real session. Fixed to {"Agent", "Task"}, with a regression test locking in the real name. Correction posted directly to Grilling: formalize the subagent-nesting parent-inference rule (including parallel/batch tool calls) #277, since it invalidates that ticket's doc-derived assumption.

Also worked around a PATH ambiguity discovered along the way: a stale, globally uv tool installed agent-context-graph can shadow this repo's own workspace source. The hook command now always resolves via uv run --package agent-context-graph, deterministically, regardless of what else is installed.

Test plan

• scripts/dev-memgraph.sh test-graph-model run against a real Claude Code session — all 11 checks passing (plus 3 earlier clean runs of the narrower pre-expansion check)
• Full suite (55 tests) green via scripts/dev-memgraph.sh test
• ruff check / ruff format --check clean
• CI workflow itself (test-live-graph-model.yaml) needs its first live workflow_dispatch trigger post-merge to fully validate the fresh-runner path (install steps, PATH wiring) — validated locally against every piece it depends on, but a GitHub-hosted runner run is the real proof