Architecture Explorer
Architecture Explorer displays agents, helpers, skills, instructions, prompts, validators, workflows, and MCP servers from the product revision pinned by this site.
The build generates the graph from the APEX checkout selected by apex-source.json.
It is not a live runtime trace or a map of the apex-docs checkout.
The explorer has two views:
- Workflow Timeline shows step owners, declared models, skill references, and gates.
- Agent Grid lets you search and filter the generated inventory.
Select a card with the pointer, Enter, or Space to inspect its metadata and relationships. Source links point to the pinned GitHub revision.
Switch between Workflow Timeline and Agent Grid with the toggle in the top bar. Press / to focus the search box; Esc to close the detail drawer.
Open full screen ↗
How to use the explorer
Section titled “How to use the explorer”Use the timeline for step ownership. Use the grid to inspect an agent, skill, instruction, workflow, or validator. Category filters narrow the inventory; search finds names and descriptions.
When a card opens, focus moves into the drawer. Tab and Shift+Tab stay inside it. Escape closes it and returns focus to the initiating card.
Node categories
Section titled “Node categories”The view uses category colors and text labels. The generated categories include:
- Agent, Subagent, Skill, Instruction, Prompt, Validator, CI Workflow, MCP Server.
Edge semantics
Section titled “Edge semantics”| Edge kind | Meaning |
|---|---|
handsOff | Historical source handoff declaration from handoffs: frontmatter. |
delegates | Agent delegates to a subagent (agents: frontmatter). |
usesSkill | Source declares a primary skill reference. |
usesSkill? | Agent uses a secondary skill (on-demand). |
invokes | Prompt invokes an agent (name-matched). |
appliesTo | Instruction applies to an agent/skill/prompt type. |
runs | CI workflow runs a validator (npm run …). |
usesMcp | Agent uses an MCP server (referenced in body). |
references | Skill references another skill (e.g. routing/delegation). |
The drawer shows source relationships, not proof that a tool or skill ran.
The site’s pin predates the current body migration, which deliberately omits
active handoffs: metadata. A handsOff edge does not promise a visible button.
Native task results, explicit selection, and the durable 00-handoff.md artifact
are distinct from those historical metadata edges. See the
current migration contract.
Data source
Section titled “Data source”Agent, subagent, and skill nodes carry source invocation metadata:
| Field | Meaning |
|---|---|
meta.invocable | user-invocable; defaults to true when omitted |
meta.disableModelInvocation | Manual-only invocation flag; defaults to false |
meta.argumentHint | Declared argument hint, or null |
meta.context | Declared context value, or null when absent |
These values describe source configuration, not verified Local or Host behavior. A skill with no declared context override still inherits the caller’s model/tools; neither graph metadata nor model labels grant permissions or prove model availability. Production agents remain human-selected. The graph schema covers visualization metadata only and does not change artifact contracts.
Both views consume public/architecture-explorer-graph.json in apex-docs.
scripts/generate-metadata.mjs invokes the pinned product’s generator and validator
during prebuild. It records the source commit and produces immutable source links.
To refresh the data, review the proposed apex-source.json change, prepare that
source, and run the normal build and tests. Do not hand-edit the graph or assume
moving product main changes it automatically.
Next steps
Section titled “Next steps”- Agent Architecture for roles, handoffs, and subagents.
- Skills & Instructions for the knowledge and rule layers.
- Workflow for the end-to-end execution model.