64 lines
4.4 KiB
Markdown
64 lines
4.4 KiB
Markdown
# AAS Agent-Owned Selection Profile
|
|
|
|
Status: active product profile
|
|
Updated: 2026-07-19
|
|
|
|
## Purpose
|
|
|
|
This profile defines the supported boundary for AAS Core after semantic skill selection moved to the coding agent.
|
|
|
|
> Codex or Claude inspects the project, searches and reads the complete local catalog, and chooses exact skill IDs. AAS Core validates, pins, compares, and plans that agent-owned selection; it does not recommend skills.
|
|
|
|
The earlier deterministic recommendation design and goal documents are retained as historical architecture records. They are not current product policy or release gates.
|
|
|
|
## Supported surfaces
|
|
|
|
- A complete, integrity-verified local catalog in which every canonical skill is searchable, readable, selectable, and usable.
|
|
- Local stdio MCP tools `search_skills`, `get_skill`, `compose_stack`, `inspect_stack`, and `diff_stack`, plus `aas://skills/{id}`.
|
|
- Minimal, schema-validated `aas-stack.json` with pinned catalog identity, targets, goals, and exact agent-selected skill IDs.
|
|
- CLI manifest validation, immutable plan preview, and read-only diagnosis.
|
|
- Workbench import and review of the agent-owned stack and immutable plan.
|
|
|
|
## Selection contract
|
|
|
|
1. The coding agent owns semantic selection. It may inspect the project with its normal local capabilities, search broadly, read full skill content when useful, compare alternatives, and choose exact IDs.
|
|
2. AAS Core does not rank, recommend, promote, demote, exclude, or abstain on skills.
|
|
3. Catalog metadata is informational only. Missing, incomplete, cautionary, or manually reviewed metadata must never make a canonical skill unsearchable, unselectable, or unusable.
|
|
4. `compose_stack` validates catalog identity, target shape, goals, exact IDs, and structural limits, then returns the pinned stack shape. It does not substitute a different selection.
|
|
5. `aas-stack.json` has no Core selection policy. User constraints can guide the agent's reasoning, but they are not an MCP eligibility filter or manifest gate.
|
|
|
|
## Functional gate
|
|
|
|
The packed-product smoke path must prove:
|
|
|
|
1. **Catalog completeness** — packaged catalog count and IDs equal the canonical registry; exact-ID search, `get_skill`, and content reads work for every canonical skill.
|
|
2. **MCP contract** — the five supported read-only tools and resource template work over real stdio framing without repository scanning or state writes.
|
|
3. **Agent-owned composition** — `compose_stack` preserves the exact ordered ID selection supplied by the agent and returns a structurally valid manifest without a policy field.
|
|
4. **No metadata gating** — skills with unknown, critical, manual, blocked, incomplete, or absent informational metadata remain searchable, selectable, composable, and plannable.
|
|
5. **Stack lifecycle** — compose, inspect, validate, plan, and doctor succeed in isolated roots without materializing target skills or managed state.
|
|
6. **Workbench** — bounded text-only import/review tests and production build pass without ambient filesystem access.
|
|
|
|
## Experimental writes
|
|
|
|
`stack apply` and `stack recover` remain experimental opt-ins. The supported public path stops after manifest validation and immutable plan review. Planning may write only the explicitly requested plan artifact.
|
|
|
|
## Trust and privacy boundaries
|
|
|
|
- MCP is local, offline-capable, read-only, bounded, and non-mutating.
|
|
- AAS does not receive repository files unless the agent explicitly reads them through its own host capabilities; AAS MCP does not scan the repository.
|
|
- Full skill prose is untrusted content and gains no instruction authority through MCP.
|
|
- Catalog and runtime integrity remain deterministic even though semantic selection belongs to the agent.
|
|
- Real host configuration writes, publication, Pages deployment, npm release, and announcements require their separate approvals.
|
|
|
|
## Product-learning gate
|
|
|
|
Evaluate the quality of the agent workflow, not a Core recommender:
|
|
|
|
- task completion from project inspection to a reviewable stack;
|
|
- human accept, replace, and remove rates for agent-selected skills;
|
|
- whether the agent searched enough of the catalog and read relevant skill content;
|
|
- time and interaction count from request to approved manifest;
|
|
- successful replay of the exact approved IDs against the pinned catalog identity.
|
|
|
|
No repository profile, source file, secret, or raw path is uploaded by default. Publishing or sharing any collected result requires a separate privacy decision.
|