Installed product: OpenTenBase AgentX; supported backends: pg; edition: Community; license: Apache-2.0. Other backends and Enterprise-only capabilities discussed below describe differences; they do not extend this package's scope.
This guide applies to both editions. Database entities retain identity, authorization, exact revisions, source lineage, attempts, leases and audit facts. An integration setting does not grant permissions. Install the complete additive migration chain for the exact package and verify deployment before starting new workers. For v5.0.5, PostgreSQL ends at 121 and Oracle/YashanDB at 119; verify the complete chain and native controls in migration. The package's generated baseline manifest determines the actual migration list.
In Dashboard Platform → Agent extensions, select an existing Provider Profile and run each capability probe independently. Text, streaming, structured output, Tool proposals, reasoning parameters, usage, image input and client cancellation are bound to the exact profile revision. Approve supported parameters only from the matching probe; changing the profile requires new evidence. Hidden Provider reasoning is never used as the visible answer. Client cancellation records local stream closure separately from Provider-confirmed cancellation.
Portal and Channels share knowledge retrieval and answer preparation. General model answers and partial knowledge supplements carry their own source labels. Ambiguous vendor/company/database questions request clarification. Answer-quality cases and results preserve the entry kind, case/profile revision, policy digest and authorized citations. A result cannot be carried over to a changed profile.
For final-answer evaluation, use Workspaces → Work continuity → Retrospective
and answer evaluation → Answer evaluation. A platform manager registers a
versioned case and rubric, queues a run for an authorized active Agent, and
explicitly starts any model request. Read the actual final output and digest;
a different authorized Human reviews facts and citation support. Queue/preparation
success, phrase assertions and an explicitly reported passed flag are separate
from this review. Profile, policy or source changes block stale runs; unknown
sends are not automatically repeated. See answer quality
for exact methods, bodies, states and recovery.
Use the directory to search authorized Tool/Skill metadata, then load the selected
exact version and content digest. Search results contain no execution grant.
Use GET /api/agent-extensions/directory and POST .../directory/load for the
authenticated API workflow. Skill text and Tool schemas load only after current
domain, credential and resource checks.
Use the context budget operation before deriving an existing continuity assembly. The conservative UTF-8 upper bound reserves output and instructions. Extractive derivations retain constraints, decisions and incomplete work verbatim; a budget too small for those obligations is rejected. Reading a derivation rechecks every original source, digest, domain and assembly expiry.
The Work continuity Agent assets area supports exact Knowledge/Memory/Skill/
Graph/Source bindings, query-ranked Skill metadata selection, model-aware source
budgets and configured document/pinned GitHub synchronization. Source managers
need platform.manage; binding/sync operators need agents.operate in addition
to current source/Domain permissions. Import remains a pending candidate and
never executes code or automatically publishes. Portal does not grant platform
management. See Agent assets and
memory and working notes for full workflows.
The SLO view supports authorized cursor pagination and status filters. Manifest preview validates size, depth, schemas and known executors, and returns a preview digest. Import that exact preview to create a draft; it does not publish or run it. A checkpoint fork requires the exact current checkpoint, reason and idempotency key. It preserves lineage and budgets; uncertain external effects require review.
Offline replay reconstructs committed state and compares event/checkpoint hashes. Migration preflight compares exact source/target versions, stable node mappings, pending nodes, leases, schemas, budgets and compensation requirements. Both return diagnostics without dispatching an executor or changing an active Run.
Register each endpoint with its Security Domain, protocol, exact version, bounded timeout/response size and reason. Use HTTPS and dedicated credentials for remote services. Credential values are encrypted; discovery, execution and revocation recheck current domain authority. Protocol and SDK versions are separate facts.
| Integration | Supported workflow |
|---|---|
| MCP | Streamable HTTP 2025-06-18 / 2025-11-25; bounded paginated discovery, independent approval of an exact read-only Tool/schema, authenticated calls, cancellation and uncertain-send reconciliation. The separate 2026-07-28 adapter supports bounded server/discover, tools/list and read-only tools/call; see versioned MCP. |
| A2A | Bounded HTTP JSON 1.0.1 tasks bound to an authorized Graph Version and compiled Plan; submit, inspect, page, resume and cancel with durable task/run mapping. |
| OTLP | Registered OTLP/HTTP JSON 1.9.0 collector; committed trace metadata only, durable delivery attempts, bounded retries and collector receipts. Prompts and output bodies are excluded. |
| Framework | LangGraph 1.2.12 only, using the packaged hash-pinned wheel closure and verified immutable Linux rootfs. Outcomes require independently observed isolated-worker evidence. |
MCP discovery never approves execution. A newly observed contract closes existing approval; approve it again after review. Uncertain sends are reconciled rather than automatically repeated. The A2A entry preserves original credentials and Graph authorization. A2A and DB4A2A keep their protocol/transport paths while sharing durable Task Plan and Step links, fenced execution attempts, cancellation and unknown-send reconciliation; see continuity operations.
Start scripts/agent_integration_worker.py under the operator's service manager
to deliver registered telemetry and reconcile uncertain calls. Prepare a new
framework rootfs with scripts/prepare_framework_rootfs.py --rootfs /var/lib/chuanxu/framework-v5.0.5, then configure its verified Linux
sandbox target. The framework receives bounded messages; platform entities remain
the source of execution and authorization facts. Generic framework execution,
live effect replay, live migration and global autonomy remain disabled.
Upload an owned image or document in an authorized Security Domain. The bounded
parser accepts PNG/JPEG, UTF-8 plain text/Markdown and text-extractable PDF; the upload
limit is 4 MiB. Images additionally require a current verified IMAGE_INPUT probe.
Queries use the authorized active model and store input/output digests and status.
Revocation prevents subsequent reads/use. Documents are untrusted data, not Tool
instructions or grants; scanned PDF OCR and arbitrary audio/video are unsupported.
Execution progress exposes durable state, queue/first-token latency and cancellation
outcome. Read-only diagnostics run only on an explicit request and store redacted
results for capability, Agent or Provider checks. They never repair configuration
or trigger global autonomous actions. Keep the distinction between FAILED,
CANCELLED and UNOBSERVED when a remote outcome has not been confirmed.