Usecase Overview
- Business Goal: Enable Agent descriptions carried by plugins to complete registration, signature verification, Agent ID generation, and admission policy binding within 5 seconds, allowing the platform to maintain real-time control of Agent ledger and immediate reference in orchestration platforms.
- Success Metrics: Registration success rate ≥98%; signature/Schema verification coverage 100%;
agent.registry.latency_p95<5s; duplicate/conflict blocking rate 100%; audit write latency <1s. - Scenario Association: Corresponds to sub-scenario A of
SCN-AGENT-REG-MGMT-001, provides metadata foundation toSCN-AGENT-REG-TENANT-001,SCN-AGENT-REG-LIFECYCLE-001, and also provides input to capability map synchronization ofSCN-AGENT-TASK-EXEC-001.
Summary: This Seed focuses on Registry API, signature verification, permission policy binding, and sandbox verification, incorporating plugin-built-in Agents into controlled catalogs in an automated manner with continuous observability.
Context & Assumptions
- Prerequisites
- Scenario document
docs/scenarios/agent-orchestration/SCN-AGENT-REG-AUTO-001.mdhas defined processes, metrics, and participants. UC-AGENT-REG-AUTO-001node indocs/_data/docmap.yamlis fully consistent with this Seed frontmatter.- Feature Flags
agent-registry-v1,plugin-autoreg-webhook,agent-sandboxare registered in the configuration center. - Plugin build pipeline generates
agent.manifest.jsonand calls Registry API on startup. - IAM, Secret Manager, Telemetry pipeline are online with necessary permissions.
- Scenario document
- Input/Output
- Input: plugin ID/version, signature, Agent capability description, permission/rate declarations, dependency tool list, sandbox validation parameters.
- Output: Agent ID, metadata records, permission policy ID, sandbox validation results, audit events,
agent.registry.*metrics.
- Boundaries
- Not responsible for tenant custom Agents (handled by
UC-AGENT-REG-TENANT-001). - Does not handle runtime monitoring or sharing policies (see lifecycle/sharing sub-usecases).
- Does not include plugin internal business logic or model inference pipelines.
- Not responsible for tenant custom Agents (handled by
Solution Blueprint
System Decomposition
| Layer | Main Components/Modules | Responsibilities | Code Entry |
|---|---|---|---|
| integration | Agent Registry API Gateway | Receive plugin manifest, authentication, rate limiting, Agent ID generation | services/agent/registry/http.ts |
| integration | Manifest Validator & Schema | Validate fields, signature, plugin version compatibility, generate audit payload | backend/config/agent/registry/schema.yaml + services/agent/registry/validator.ts |
| integration | IAM Policy Publisher | Map permission/rate policies to IAM, generate credential references | services/iam/policy/publisher.ts |
| integration | Sandbox Validation Runner | Trigger scripts/ops/agent-sandbox-validate.mjs, upload execution reports | scripts/ops/agent-sandbox-validate.mjs |
| integration | Audit & Telemetry Sink | Write agent.registry.state.changed events, metrics, logs | services/observability/audit_pipeline.ts |
Flow & Sequence
- Step 1 – Manifest Intake: Plugin startup hook submits manifest + signature to
POST /internal/agent/registry; gateway performs OAuth2 + plugin certificate authentication. - Step 2 – Schema & Signature Validation: Validator validates fields, capability references, version compatibility, signature certificates; failures return debuggable error codes and write alerts.
- Step 3 – Agent ID Issuance: Generate Agent ID, record plugin mapping, write to metadata tables and audit logs; pass manifest parsing results to IAM publisher.
- Step 4 – Policy Binding & Sandbox: IAM publisher generates permission/rate policies and credentials; trigger Sandbox Runner to execute regression use cases, get health reports.
- Step 5 – Activation & Broadcast: Write registration success event to
agent.registry.state.changed, synchronize to Agent orchestration platform and monitoring dashboards; if Sandbox fails, mark status aspending_fixand notify Vendor.
Contracts & Interfaces
- Inbound APIs / Events
POST /internal/agent/registry— Body includesagent_id?,plugin_id,version,capabilities[],permissions,rate_limits,signature; requiresagent.registry.writescope and plugin certificate; supportsdry_run.POST /internal/agent/registry/{agent_id}/sandbox— Manually rerun sandbox validation; can specifyprofile,timeout,telemetry_tags.EVENT agent.registry.registered/agent.registry.failed— Connect to monitoring, alerting, and Copilot.
- Outbound Calls
IAM Policy Service /internal/policies— Create permission, rate policies and credentials; failures require Registry state rollback.Agent Catalog Sync /internal/catalog/agents— Publish metadata to orchestration platform.Telemetry Bus agent.registry.*,Audit Service /events— Write metrics/logs.
- Configs & Scripts
backend/config/agent/registry/schema.yaml— Define required fields and compatibility policies.config/feature_flags/agent-registry.yaml— Control auto-approval/sandbox policies.scripts/ops/agent-sandbox-validate.mjs— Run regression, record health reports.
Implementation Checklist
| Item | Description | Completion Status | Owner |
|---|---|---|---|
| Manifest Schema & Lint | Support capabilities, permissions, dependencies, tenant label fields, provide CLI validation | [ ] | Agent Platform Guild |
| Signature Verification | Integrate certificate trust chain, expiration alerts, replay protection | [ ] | Plugin Guild |
| Agent ID & Metadata Store | Create/migrate table structure, indexes, idempotent writes | [ ] | Agent Platform Guild |
| IAM Policy Binding | Auto-generate permission/rate policies, credential callback, conflict rollback | [ ] | Agent Platform Guild |
| Sandbox Automation | Extend script to cover core plugins, upload metrics & reports | [ ] | Ops Reliability Center |
| Audit & Alerting | Event/metrics writing, Grafana dashboards, PagerDuty alerts | [ ] | Ops Reliability Center |
Testing Strategy
- Unit Tests
- Manifest Schema validator (required/enum/JSON schema).
- Signature verification, replay protection, duplicate Agent blocking logic.
- IAM publisher serialization and idempotency for rate/permission mapping.
- Integration Tests
- Use sandbox plugins to trigger
POST /internal/agent/registry, verify interaction with IAM, Sandbox Runner. - Inject exceptions (signature errors, missing fields, IAM failures) to ensure rollback and alert triggering.
- Use sandbox plugins to trigger
- End-to-End Validation
- In staging use real plugin packages to execute auto-registration, observe Agent ID generation, orchestration platform sync, sandbox reports.
- Run
scripts/ops/agent-sandbox-validate.mjs --agent <id> --profile fullto validate regression pipeline.
- Non-functional Tests
- Load test Registry API (100 concurrent RPS) ensuring <5s SLA.
- Chaos: network jitter, Secret Manager failures, confirm alert/rollback paths.
Observability & Ops
- Metrics
agent.registry.latency_p95,agent.registry.success_total,agent.registry.signature_failure_total,agent.registry.duplicate_block_total,agent.registry.sandbox_failure_total.
- Logs
- Each registration writes plugin ID, Agent ID, version, signature fingerprint, policy ID, sandbox results; log levels INFO/ERROR;落地 Elastic + S3.
- Alerts
- Signature failure rate >2% (5-minute window), registration error rate >5%, Sandbox Pending >10 minutes, IAM publishing failure.
- Notification channels: Ops Pager, Agent Platform Teams channel, Vendor email broadcast.
- Dashboards
- Grafana「Agent Registry」 panel, Datadog
agent.registry.*, Elastic Dashboard (audit logs), Sandbox report view.
- Grafana「Agent Registry」 panel, Datadog
Rollback & Failure Handling
- Immediately delete newly created Agent records, revoke IAM policies, clean up audit references after failure, return debuggable error codes.
- Provide
scripts/ops/agent-registry-cleanup.mjs --agent <id>to clean residual metadata and credentials. - Sandbox failure: mark Agent status
pending_fix, block orchestration platform reference, and automatically openagent-registry-sandbox-failurePager. - When IAM or Audit unavailable: retry in memory queue, after 5-minute timeout enter isolated queue and alert on-call.
Follow-ups & Risks
| Risk/Item | Impact | Mitigation | Owner | ETA |
|---|---|---|---|---|
| Legacy plugins not upgraded manifest Schema, causing mass registration failures | Affect Vendor rollout节奏 | Provide CLI validator, add schema lint to release script; allow schema_version gradual strategy | Plugin Guild | 2025-03-05 |
| Sandbox resource insufficient causing queuing >10 minutes | Registration pipeline delay, SLA violation | Scale sandbox pool, introduce priority queue, allow "post-sandbox" when necessary with Ops approval | Ops Reliability Center | 2025-03-01 |
References & Links
- Scenario Document:
docs/scenarios/agent-orchestration/SCN-AGENT-REG-MGMT-001.md - Sub-scenario:
docs/scenarios/agent-orchestration/SCN-AGENT-REG-AUTO-001.md - Docmap:
docs/_data/docmap.yaml(SCN-AGENT-REG-MGMT-001 → UC-AGENT-REG-AUTO-001) - Repo Metadata:
docs/_data/repos.yaml(key: powerx) - Contracts & Standards:
docs/standards/powerx/backend/integration/09_agent/Agent_Manager_and_Lifecycle_Spec.md - Related Scripts:
scripts/ops/agent-sandbox-validate.mjs,scripts/ops/agent-registry-cleanup.mjs
