feat(registry): agent asset catalog on AgentCore — provisioning, port/adapter, API, CLI (#246)

agent/src/registry/__init__.py

@@ -0,0 +1,13 @@
+"""Agent asset registry client + reference grammar (#246).
+
+The agent is a *read-only* consumer of the registry: it receives already-resolved
+assets in its task payload and, where it needs to look one up directly, talks to
+the substrate through the ``RegistryClient`` port (never a raw AWS SDK client).
+
+Public surface:
+  - ``parse_ref`` / ``ParsedRef`` (``ref``) — the strict ``registry://`` grammar,
+    mirrored byte-for-byte by ``cdk/src/handlers/shared/registry/ref.ts`` and the
+    ``contracts/registry-resolution/`` parity corpus.
+
+See ``docs/design/REGISTRY.md`` and ``ISSUE_246_AGENTCORE_FINDINGS.md``.
+"""

agent/src/registry/agentcore_client.py

@@ -0,0 +1,144 @@
+"""Read-side AgentCore implementation of the ``RegistryClient`` port (#246).
+
+The agent only reads: ``get_record`` and ``resolve``. This is the one Python file
+that talks to the AgentCore control plane (via boto3) — everything upstream uses
+the port, so a substrate swap is confined here. Mirrors the read half of
+``cdk/src/handlers/shared/registry/agentcore-client.ts``.
+"""
+
+from __future__ import annotations
+
+import json
+import re
+from typing import TYPE_CHECKING, Any
+
+from registry.client import RegistryResolutionError, ResolvedAsset
+from registry.resolver import select_highest
+
+if TYPE_CHECKING:
+    from registry.ref import ParsedRef
+
+_NAME_SEP = "/"
+# Option-A record name is `kind/namespace/name`; the name may itself contain `/`.
+_NAME_MIN_PARTS = 2
+# The reverse-DNS key under which ABCA runtime config rides in a native `_meta`
+# block (matches RUNTIME_META_KEY in registry/types.ts).
+_RUNTIME_META_KEY = "dev.abca.runtime"
+# Frontmatter key carrying the runtime payload (JSON) in a native AGENT_SKILLS
+# SKILL.md — mirrors SKILL_RUNTIME_FM_KEY in registry/agentcore-client.ts.
+_SKILL_RUNTIME_FM_KEY = "x-abca-runtime"
+_SKILL_RUNTIME_RE = re.compile(rf"^{_SKILL_RUNTIME_FM_KEY}:\s*'(.+)'\s*$", re.MULTILINE)
+_RESOLVABLE_STATUSES = ("APPROVED", "DEPRECATED")
+
+
+class AgentCoreRegistryClient:
+    """Read-only registry access backed by AgentCore."""
+
+    def __init__(self, registry_id: str, client: Any) -> None:
+        # ``client`` is a boto3 ``bedrock-agentcore-control`` client, injected so
+        # the agent's scoped-session helper (aws_session) owns credential wiring.
+        self._registry_id = registry_id
+        self._client = client
+
+    # --- name (Option A) decode -------------------------------------------------
+
+    @staticmethod
+    def _decode_name(record_name: str) -> tuple[str, str, str]:
+        parts = record_name.split(_NAME_SEP)
+        kind = parts[0] if parts else ""
+        namespace = parts[1] if len(parts) > 1 else ""
+        name = _NAME_SEP.join(parts[_NAME_MIN_PARTS:]) if len(parts) > _NAME_MIN_PARTS else ""
+        return kind, namespace, name
+
+    @staticmethod
+    def _id_from_arn(arn: str) -> str:
+        return arn.split("/")[-1] if "/" in arn else arn
+
+    # --- record extraction ------------------------------------------------------
+
+    def _extract_runtime(self, raw: dict[str, Any]) -> dict[str, Any]:
+        descriptors = raw.get("descriptors", {}) or {}
+        descriptor_type = raw.get("descriptorType")
+        if descriptor_type == "CUSTOM":
+            body = json.loads(descriptors.get("custom", {}).get("inlineContent", "{}"))
+            return body.get("runtime", {})
+        if descriptor_type == "AGENT_SKILLS":
+            # SKILL.md is Markdown frontmatter, not JSON — recover the runtime
+            # from the `x-abca-runtime` frontmatter key (mirrors the TS adapter).
+            skill_md = (
+                descriptors.get("agentSkills", {}).get("skillMd", {}).get("inlineContent", "")
+            )
+            m = _SKILL_RUNTIME_RE.search(skill_md)
+            return json.loads(m.group(1)) if m else {}
+        # MCP: JSON server.json with the runtime in a `_meta` block.
+        inline = descriptors.get("mcp", {}).get("server", {}).get("inlineContent") or "{}"
+        body = json.loads(inline)
+        return body.get("_meta", {}).get(_RUNTIME_META_KEY, {})
+
+    def _list_records(self, kind: str, namespace: str) -> list[dict[str, Any]]:
+        out: list[dict[str, Any]] = []
+        next_token: str | None = None
+        while True:
+            kwargs: dict[str, Any] = {"registryId": self._registry_id, "maxResults": 50}
+            if next_token:
+                kwargs["nextToken"] = next_token
+            page = self._client.list_registry_records(**kwargs)
+            for summary in page.get("registryRecords", []) or []:
+                dkind, dns, _ = self._decode_name(summary.get("name", ""))
+                if dkind != kind or dns != namespace:
+                    continue
+                record_id = (
+                    self._id_from_arn(summary["recordArn"])
+                    if summary.get("recordArn")
+                    else summary.get("recordId")
+                )
+                if not record_id:
+                    continue
+                full = self._client.get_registry_record(
+                    registryId=self._registry_id, recordId=record_id
+                )
+                out.append(full)
+            next_token = page.get("nextToken")
+            if not next_token:
+                break
+        return out
+
+    # --- port surface -----------------------------------------------------------
+
+    def get_record(
+        self, kind: str, namespace: str, name: str, version: str
+    ) -> dict[str, Any] | None:
+        for raw in self._list_records(kind, namespace):
+            _, _, dname = self._decode_name(raw.get("name", ""))
+            if dname == name and raw.get("recordVersion") == version:
+                return raw
+        return None
+
+    def resolve(self, ref: ParsedRef) -> ResolvedAsset:
+        ref_str = f"registry://{ref.kind}/{ref.namespace}/{ref.name}@{ref.constraint.raw}"
+        records = self._list_records(ref.kind, ref.namespace)
+        candidates = [
+            r
+            for r in records
+            if self._decode_name(r.get("name", ""))[2] == ref.name
+            and r.get("status") in _RESOLVABLE_STATUSES
+        ]
+        by_version = {r.get("recordVersion", ""): r for r in candidates}
+        winning = select_highest(list(by_version.keys()), ref.constraint)
+        if winning is None:
+            raise RegistryResolutionError(
+                "NO_MATCHING_VERSION",
+                ref_str,
+                f"no approved version of {ref.kind}/{ref.namespace}/{ref.name} "
+                f"satisfies {ref.constraint.raw}",
+            )
+        winner = by_version[winning]
+        warnings = ["DEPRECATED"] if winner.get("status") == "DEPRECATED" else []
+        return ResolvedAsset(
+            kind=ref.kind,
+            namespace=ref.namespace,
+            name=ref.name,
+            version=winning,
+            runtime=self._extract_runtime(winner),
+            warnings=warnings,
+        )

agent/src/registry/client.py

@@ -0,0 +1,60 @@
+"""The read-side ``RegistryClient`` port for the agent (#246).
+
+The agent is a read-only consumer: it resolves refs and fetches records, but never
+publishes or governs. It talks to the substrate through this Protocol, never a raw
+AWS SDK client — the one implementation is ``AgentCoreRegistryClient``
+(``registry.agentcore_client``), so a substrate swap is confined there.
+
+The write-side verbs (publish / submit / approve) live only on the TypeScript port
+(``cdk/src/handlers/shared/registry/client.ts``); the agent has no business
+mutating the registry.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import TYPE_CHECKING, Any, Protocol
+
+if TYPE_CHECKING:
+    from registry.ref import ParsedRef
+
+
+@dataclass(frozen=True)
+class ResolvedAsset:
+    """One resolved asset: enough to load it without knowing where bytes live."""
+
+    kind: str
+    namespace: str
+    name: str
+    version: str
+    runtime: dict[str, Any]
+    warnings: list[str] = field(default_factory=list)
+
+
+class RegistryResolutionError(Exception):
+    """Raised when a ref cannot be resolved. ``reason`` matches the TS token set
+    (``NO_MATCHING_VERSION`` / ``REMOVED`` / ``INVALID_CONSTRAINT`` /
+    ``INVALID_REGISTRY_REF``) so both languages agree on *why*."""
+
+    def __init__(self, reason: str, ref: str, message: str) -> None:
+        super().__init__(message)
+        self.reason = reason
+        self.ref = ref
+
+
+class RegistryClient(Protocol):
+    """Read-only registry access. See the TS port for the full (write) surface."""
+
+    def get_record(
+        self, kind: str, namespace: str, name: str, version: str
+    ) -> dict[str, Any] | None:
+        """Fetch a single record by exact coordinates, or ``None`` if absent."""
+        ...
+
+    def resolve(self, ref: ParsedRef) -> ResolvedAsset:
+        """Resolve a parsed ref to a single APPROVED (or DEPRECATED+warn) asset.
+
+        Fail-closed: raises ``RegistryResolutionError`` with a specific reason on
+        any unresolved ref.
+        """
+        ...

agent/src/registry/ref.py

@@ -0,0 +1,110 @@
+"""Strict ``registry://`` reference grammar for the agent asset registry (#246).
+
+::
+
+    registry://<kind>/<namespace>/<name>@<constraint>
+      kind       = [a-z][a-z0-9_]*            snake_case: mcp_server, cedar_policy_module
+      namespace  = [a-z][a-z0-9-]*
+      name       = [a-z0-9][a-z0-9._-]*
+      constraint = [^~]?MAJOR.MINOR.PATCH[-prerelease]   exact / caret / tilde only
+
+The ``@<constraint>`` pin is MANDATORY (fail-closed: no implicit "latest").
+
+This module mirrors ``cdk/src/handlers/shared/registry/ref.ts`` byte-for-byte and
+is exercised by the ``contracts/registry-resolution/`` parity corpus. Keep the two
+in lockstep — a change here without the matching TS change (or vice versa) is a
+parity break that CI must catch.
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass
+
+# MVP asset kinds the registry loads end-to-end or stages.
+REGISTRY_KINDS = ("mcp_server", "cedar_policy_module", "skill")
+# Reserved kinds accepted by the grammar but rejected at publish (no loader yet).
+RESERVED_KINDS = ("plugin", "subagent", "prompt_fragment", "capability")
+
+# Structural split — scheme + 3 path segments + the (mandatory) constraint.
+# ``\Z`` (absolute end of string), not ``$``: Python's ``$`` also matches just
+# before a trailing newline, so ``$`` would accept ``…@1.0.0\n`` that the JS
+# mirror (ref.ts, no ``m`` flag) rejects — a byte-for-byte parity break.
+_REF_SHAPE = re.compile(
+    r"^registry://([a-z][a-z0-9_]*)/([a-z][a-z0-9-]*)/([a-z0-9][a-z0-9._-]*)@(.+)\Z"
+)
+# exact / caret / tilde over MAJOR.MINOR.PATCH with an optional prerelease.
+# Rejects ``*``, ``latest``, ``>=``, ``<=``, x-ranges, and bare prerelease modifiers.
+_CONSTRAINT = re.compile(
+    r"^([\^~]?)(0|[1-9]\d*)\.(0|[1-9]\d*)\.(0|[1-9]\d*)(?:-([0-9A-Za-z.-]+))?\Z"
+)
+_OP_BY_PREFIX = {"": "exact", "^": "caret", "~": "tilde"}
+
+
+class RefError(ValueError):
+    """Raised when a ``registry://`` ref is malformed.
+
+    ``reason`` is one of ``INVALID_REGISTRY_REF`` / ``INVALID_CONSTRAINT`` — the
+    same reason tokens the TS resolver reports, so both sides agree on *why* a
+    ref failed, not just *that* it failed.
+    """
+
+    def __init__(self, reason: str, message: str) -> None:
+        super().__init__(message)
+        self.reason = reason
+
+
+@dataclass(frozen=True)
+class ParsedConstraint:
+    op: str  # exact | caret | tilde
+    major: int
+    minor: int
+    patch: int
+    prerelease: str | None
+    raw: str
+
+
+@dataclass(frozen=True)
+class ParsedRef:
+    kind: str
+    namespace: str
+    name: str
+    constraint: ParsedConstraint
+
+
+def parse_constraint(raw: str) -> ParsedConstraint | None:
+    """Parse + validate a constraint string in isolation. ``None`` if invalid."""
+    m = _CONSTRAINT.match(raw)
+    if not m:
+        return None
+    prefix, major, minor, patch, prerelease = m.groups()
+    return ParsedConstraint(
+        op=_OP_BY_PREFIX[prefix],
+        major=int(major),
+        minor=int(minor),
+        patch=int(patch),
+        prerelease=prerelease,
+        raw=raw,
+    )
+
+
+def parse_ref(ref: str) -> ParsedRef:
+    """Parse a strict ``registry://kind/namespace/name@constraint`` reference.
+
+    Raises ``RefError`` (with a ``reason``) on a malformed ref or a floating /
+    unsupported constraint — pins are mandatory.
+    """
+    shape = _REF_SHAPE.match(ref)
+    if not shape:
+        raise RefError(
+            "INVALID_REGISTRY_REF",
+            f"not a valid registry ref (expected registry://kind/namespace/name@constraint): {ref}",
+        )
+    kind, namespace, name, raw_constraint = shape.groups()
+    constraint = parse_constraint(raw_constraint)
+    if constraint is None:
+        raise RefError(
+            "INVALID_CONSTRAINT",
+            f"unsupported version constraint '{raw_constraint}' (use exact, ^, or ~)",
+        )
+    return ParsedRef(kind=kind, namespace=namespace, name=name, constraint=constraint)