| name | agent-hub |
| description | Plug-and-play orchestration skill. User states a goal in natural language; the model runs commands/agent-hub.md (dispatcher), classifies work via agent/operating-model.md, then opens only the matching command or architecture doc (map, architect, memory, recovery, verify, next-pass, or one NN-layer). Layered docs for intent, policy, context, planning, routing, delegation, execution, MCP/tool boundaries, verification, memory, recovery, observability, human gates, release. Use for repo orchestration reviews, boundary work, memory/recovery design, permission-sensitive refactors. |
| license | See repository LICENSE |
agent-hub (machine-readable entrypoint)
What this is
agent-hub is a documentation-driven control plane for agents: layered architecture, indexes, diagrams, templates, command specs, and explicit editing rules. Application code under review may live outside this folder; this package defines how to navigate and change it safely.
Plug-and-play: Start from commands/agent-hub.md when the user’s goal is broad or ambiguous. Infer intent, then route to the smallest command or single architecture/NN-*.md.
Canonical path: SKILL.md at the root of this package (this file).
Read first (minimum)
agent/README.md — operating scope for agents using this package.agent/operating-model.md — classify tasks and pick a route.architecture/00-scope-and-rules.md — global rules and doc authority.
Then branch by task using indexes/orchestration-map.md and the matching architecture/NN-*.md file. For delegated specialist work, see subagents/README.md and run routing via commands/agent-hub.md.
What not to assume
- Do not assume a particular framework, vendor SDK, or single entry file name in application code.
- Do not assume one universal verification command covers all quality gates; verification is multi-entry (skill, command, workflow). See
architecture/07-verification-and-reflection.md.
- Do not assume memory is one subsystem; session, durable extraction, and shared team memory differ in triggers and writers. See
architecture/08-memory-pipelines.md.
- Do not assume UI control flow equals MCP transport, client logic, or tool execution. See
architecture/06-mcp-and-tool-boundaries.md.
Where to look
| Concern | Primary doc |
|---|
| Orchestration overview | indexes/orchestration-map.md |
| Intent and commands | architecture/01-intent-and-commands.md |
| Policy and permissions | architecture/02-policy-and-permissions.md |
| Context and state | architecture/03-context-and-state.md |
| Planning and routing | architecture/04-planning-and-routing.md |
| Delegation and execution | architecture/05-delegation-and-execution.md |
| MCP and tool boundaries | architecture/06-mcp-and-tool-boundaries.md |
| Verification and reflection | architecture/07-verification-and-reflection.md |
| Memory pipelines | architecture/08-memory-pipelines.md |
| Recovery and resume | architecture/09-recovery-and-resume.md |
| Observability and human gates | architecture/10-observability-and-human-gates.md |
| Release and scope | architecture/00-scope-and-rules.md |
| Specialist subagents | subagents/README.md |
| Cursor launcher setup (paste into chat) | COMMANDS.md |
High-risk areas (typical patterns)
- Central mutable session or bootstrap state: guarded boundary; do not add fields casually. See
architecture/03-context-and-state.md.
- Main query or agent loop modules: high blast radius; prefer phased extraction. See
architecture/05-delegation-and-execution.md.
- Permission and human-gate hooks: easy to double-resolve or bypass; coordinate with
architecture/02-policy-and-permissions.md and 10-observability-and-human-gates.md.
- MCP clients and connection managers: easy to conflate with UI; see
06-mcp-and-tool-boundaries.md.
Choosing commands vs docs vs agent guidance vs subagents
- Universal dispatcher:
commands/agent-hub.md (routing modes, when to assign a specialist first).
- Quick launcher: other
commands/*.md, or optional workspace .cursor/commands/agent-hub-*.md / agent-hub.md when present (see indexes/command-index.md). To create those files automatically, the user can attach COMMANDS.md and follow its paste prompt.
- Specialist persona:
subagents/*.md for scoped analysis or planning; they link into canon instead of replacing it.
- Deep boundary rules:
architecture/*.md.
- Behavioral contract (stop conditions, escalation):
agent/*.md.
- Full audit:
commands/architect.md workflow, then templates/inspection-pass-template.md.
Commands index
See indexes/command-index.md for canonical commands/*.md versus optional editor launchers.
Decisions
indexes/decisions-index.md — architecture decision log pointers.docs/decisions/README.md — how to record decisions.
