Skip to content

Declared and behavior-verified capabilities

The registry included in 0.1.0 enumerates versioned adapter descriptions and keeps declarations separate from scoped observations. It does not execute a vendor runtime, discover plugins or enable harness policies during inspection.

agentloop harness list-adapters
agentloop harness capabilities --adapter python-wrapped --json-out runs/python-capabilities.json
agentloop harness capabilities --adapter omnigent-otel

The registry currently describes four built-in adapters:

Adapter Kind and scope
python-wrapped Existing Python AdapterCapabilities; explicitly wrapped callable boundaries
langgraph-1.2.11 Existing LangGraphHarness(boundaries=ALL) declarations; requires explicit boundary registration at runtime
harbor-atif Read-only ATIF source projection; no runtime interception
omnigent-otel Read-only OTLP/session/policy observations; no dispatch control

LangGraph enumeration constructs its import-safe capability descriptor without loading the LangGraph SDK. This describes the all-boundaries configuration, not the narrower default graph configuration or every vendor-internal call. Version 1.2.11 is the target upstream version; enumeration does not assert that it is installed or behavior-tested.

Manifest 1.0

AdapterManifest records identity/version, upstream harness/version, transport, control/observation kind, observation/declaration sources, code fingerprint and typed capability declarations. Each declaration records its name, boolean/null support, explicit boundary set, reason and live-probe requirement. Unknown structural fields and duplicate dimensions fail. Canonical manifest SHA-256 binds these exact declarations; source fingerprints identify available built-in module bytes and can be unavailable in a frozen distribution.

Dimensions cover observation, dispatch enforcement, callable lifecycle and compatibility. Declarations derive supported model/tool/completion hooks and sync/async/generator lifecycles from the existing contracts. No universal pre-request gate, arbitrary interruption, model override, approval, compaction, resume or fork is inferred from those contracts. Absent declarations stay null.

Observation and state

CapabilityObservation records capability/verdict, versioned probe identity, tested adapter version/transport, manifest hash, relative evidence file/hash, environment fingerprint, timezone-qualified test timestamp and tested boundaries. Its source distinguishes offline behavioral, live behavioral and external reports.

State Meaning
declared_supported Support is declared; no current qualifying probe
declared_unsupported The adapter explicitly disclaims support
verified_supported A trusted current scoped behavioral probe passed
verified_failed A trusted current scoped behavioral probe failed
unknown Missing declaration/evidence, stale identity, wrong scope or external-only result
skipped Probe did not run; a reason is mandatory

A declaration of true with a failed probe produces drift. A false declaration with a passing probe also produces drift; the original declaration is retained. Missing probe/version/evidence/environment/timestamp, changed manifest, transport or boundary mismatch cannot produce verified support. A credential/SDK skip is neither a pass nor evidence of unsupported behavior.

Model override, approval, context handling, resume and fork need a relevant live probe. An offline fake cannot upgrade those states. A successful Python wrapper probe does not transfer to ACP, native TUI, native server or another SDK version.

Trust and runtime controls

Catalog files are bounded strict JSON data. Loading a manifest never imports its adapter name, follows entry points or runs source code. No plugin discovery is currently justified; an explicitly installed future Python plugin would be fully trusted code, separate from artifact parsing.

Observation serialization preserves recorded claims. Deserializing a historical report gives external evidence, not a fresh trusted runtime probe. The internal probe bridge is reserved for executable driver outcomes; it is not an authenticity guarantee against malicious code in the current Python process. Hashes establish supplied-byte consistency, not honesty.

require_verified_boundary gates new registry-dependent controls on current built-in, control-kind, declared support and trusted scoped probe evidence. A read-only source or catalog cannot enable enforcement. The LangGraph gate also checks the currently installed upstream version. Existing wrapper configuration and enforcement APIs retain their behavior; registry inspection never switches their disabled/shadow/enforce mode.

Instrumentation coverage is explicit boundaries only. Hidden SDK calls, unwrapped tools and externally running agents are outside that claim. A recorded policy decision or approval event is not a behavioral denial probe.

Python and validation

from agentloop.interoperability.registry import capability_report, get_adapter

report = capability_report(get_adapter("python-wrapped"))
assert not report["drift"]
# This is a declaration report, not proof that any behavioral probe ran.
uv run --frozen python -m pytest tests/test_capability_registry.py tests/test_harness.py tests/test_enforcement_conformance.py -q -ra

Tests cover scoped declarations, both drift directions, missing/stale evidence, serialization, data-only loading, unsupported enforcement, import safety and existing wrapper defaults. The executable offline bench provides scoped wrapper probes. No live provider/transport was exercised here.

The design follows the declaration/observation separation in Omnigent a2956be0e97bb175a60b274053d836f07d494c6c, using AgentLoop's existing implementations. No upstream code was copied or production control enabled. Core Python 3.10 remains supported; these commands are included in 0.1.0.