Executive Summary
Plugins should carry Agent description files upon startup or upgrade and automatically complete registration through the Registry API. The platform needs to complete signature/field verification, generate Agent ID, write metadata and audit within 5 seconds, while triggering security reviews or directly authorizing sandbox runs based on policies. This sub-scenario ensures Vendor-delivered Agents are immediately纳入资产台账 and can be referenced by the main orchestration platform, avoiding zombie or duplicate instances.
Scope & Guardrails
- In Scope: Agent description generation, Registry API, signature and compatibility verification, Agent ID generation, approval policies, sandbox verification, audit and metrics.
- Out of Scope: Plugin business logic, Agent task execution, Tenant custom configuration, Marketplace review.
- Environment & Flags:
agent-registry-v1,plugin-autoreg-webhook,vendor-sandbox; depends on Secret Manager, Plugin Manifest Builder, Audit Service.
Participants & Responsibilities
| Scope | Repository | Layer | Responsibilities & Deliverables | Owners |
|---|---|---|---|---|
| registry-api | powerx | service | Registry API, signature & Schema verification, Agent ID allocation, audit logging | Agent Platform Guild |
| plugin-manifest | powerx-plugin | integration | Build Agent description, carry version/permissions, sandbox regression scripts | Plugin Guild |
| security-hooks | powerx | service | Plugin Allowlist, risk policies, auto-approval/review logic | Agent Platform Guild |
End-to-End Flow
- Stage 1 – Manifest Build & Dispatch: Plugins generate
agent.manifest.jsonat compile time or startup, containing capabilities, interfaces, permissions, version, dependencies and signature, and call Registry API via startup hooks. - Stage 2 – Validation & Correlation: Registry verifies signature, Schema, and plugin version alignment, blocks missing fields or duplicate Agents, and correlates records to plugin version ledger and Vendor version history.
- Stage 3 – ID Issuance & Policy Hooks: Generate Agent ID via snowflake or UUID, write to metadata repository and trigger IAM Policy Publisher to generate permission/rate policies.
- Stage 4 – Sandbox & Security Checks: Automatically trigger sandbox verification, auto-review or manual review based on risk policies,沉淀沙箱报告与 Audit Trail.
- Stage 5 – Activation, Broadcast & Telemetry: Synchronize Agent information to orchestration platform, Catalog, monitoring metrics, publish
agent.registry.state.changedevent, and return traceable audit ID to Vendor.
Key Interactions & Contracts
- APIs / Events:
POST /internal/agent/registry,GET /internal/agent/{id},POST /internal/agent/registry/{id}/validate,EVENT agent.registry.registered,EVENT agent.registry.failed. - Configs / Schemas:
config/agent/registry/schema.yaml,docs/standards/powerx/backend/integration/09_agent/Agent_Manager_and_Lifecycle_Spec.md,plugins/<name>/agent.manifest.jsontemplate. - Security / Compliance: Plugin signature and certificate verification, Manifest version compatibility strategy, same-name Agent conflict protection, audit trails and risk control events.
Usecase Links
UC-AGENT-REG-AUTO-001— Plugin automatic registration flow (integration layer,docs/use_cases/_from_hub/SCN-AGENT-REG-MGMT-001/UC-AGENT-REG-AUTO-001.md).
Implementation Checklist
| Item | Description | Owner | Status |
|---|---|---|---|
| Manifest Schema & CLI | Maintain config/agent/registry/schema.yaml and lint/CLI tools, covering capabilities, permissions, dependencies, tenant labels | Plugin Guild | [ ] |
| Registry API Gateway | services/agent/registry/http.ts: authentication, rate limiting, callbacks, replay protection | Agent Platform Guild | [ ] |
| Signature & Security Hooks | services/security/signature_verifier.ts, Allowlist, risk policies, audit extensions | Agent Platform Guild | [ ] |
| IAM Policy Binding | services/iam/policy/publisher.ts: permission/rate policy generation, conflict rollback | Agent Platform Guild | [ ] |
| Sandbox & Telemetry | scripts/ops/agent-sandbox-validate.mjs, services/observability/audit_pipeline.ts: verification, metrics, events, alerts | Ops Reliability Center | [ ] |
Acceptance Criteria
- Manifest submission to successful registration averages <5 seconds, failures return debuggable error codes.
- Description file signature and required field verification coverage is 100%, duplicate or missing fields must be blocked.
- Within 1 second after successful registration, Agent can be queried in Agent ledger, orchestration platform and plugin version records.
Testing Strategy
- Unit: Manifest Schema validator, signature verifier, duplicate blocking logic, IAM publisher idempotency all need 90%+ coverage.
- Integration: Use sandbox plugins to call
POST /internal/agent/registrycovering success/failure paths; simulate signature expiration, missing fields, IAM failure, sandbox errors. - End-to-End: Run
scripts/ops/agent-sandbox-validate.mjs --agent <id>,scripts/qa/plugin-autoreg.mjs --plugin insight-bot@1.2.0, observe Audit/metrics/events. - Non-functional: Concurrent load testing of Registry API (100 RPS) and Chaos (Secret Manager/Audit unavailable) to verify degradation and rollback.
Observability & Ops
- Metrics:
agent.registry.latency_p95,agent.registry.success_rate,agent.registry.signature_failure_total,agent.registry.duplicate_block_total,agent.registry.sandbox_failure_total. - Logs/Audit: Each registration records plugin ID, Agent ID, manifest hash, signature fingerprint, policy ID, sandbox results; INFO/ERROR level written to Elastic + Audit Log.
- Alerts: Signature failure rate >2%, registration error rate >5%, sandbox Pending >10 minutes, Audit write failures; push to PagerDuty + Teams #agent-registry.
- Dashboards: Grafana「Agent Registry」, Datadog
agent.registry.*, Vendor self-service reports (output byscripts/qa/plugin-autoreg.mjs).
Rollback & Failure Handling
- Registration failure: Revoke newly created Agent records, clean up policies/credentials and audit; return
4xx/5xxerror and traceId. - Signature/Schema compatibility issues: Blocked by CLI/lint in CI stage, Registry provides
dry_run=truediagnostic mode. - Sandbox failure: Mark Agent as
pending_fix, block orchestration platform references, notify Vendor and allowPOST /internal/agent/registry/{id}/validatererun. - Audit/Telemetry unavailable: Temporarily store events in local queue, batch replay after recovery; trigger manual runbook on timeout.
Follow-ups & Risks
| Risk/Item | Impact | Mitigation | Owner | ETA |
|---|---|---|---|---|
| Manifest Schema incompatible with Vendor legacy versions | Mass registration failures | Maintain schema_version strategy + compatibility layer, provide upgrade guide and CLI validation | Plugin Guild | 2025-03-01 |
| Signature certificate rotation not timely | Security risk/registration rejection | Build certificate expiration alert, force Vendor upload T-7 days in advance | Agent Platform Guild | 2025-02-28 |
| Sandbox resource bottleneck | Registration queuing | Scale up resource pool, introduce priority queue and offline batch mode | Ops Reliability Center | 2025-03-05 |
Appendix
docs/meta/scenarios/powerx/agent-and-automation/agent-orchestration/agent-registration-and-management/primary.mddocs/scenarios/agent-orchestration/SCN-AGENT-REG-MGMT-001.md
