Feature/agentcore ops review

[pamvas]

Description

Adds agentcore-ops-review, a read-only operational review skill for Amazon Bedrock AgentCore, aligned with the AWS Well-Architected Framework. It discovers a customer's AgentCore estate — runtimes, endpoints, versions, memories, gateways and targets, browsers, code interpreters, and workload identities — and assesses it across four check areas plus cross-pillar runtime observability:

• Runtime Resilience (Reliability) — failed/stuck runtimes and endpoints, single-AZ VPC placement, endpoint version drift, DEFAULT-endpoint-only deployments (AC-RUN-1..4)
• Gateway Health (Reliability) — gateway status, target redundancy/health, policy engine attachment, stale target sync (GW-01)
• Memory & Knowledge Effectiveness (Performance Efficiency) — strategy-aware extraction errors, empty long-term memory, high ingestion error rate, never-populated memories (AC-MEM-1..4)
• Resource Utilization & Operational Hygiene (Operational Excellence) — idle resources, consolidation opportunities, low utilization (AC-UTIL-1..3)
• Runtime Observability — SessionCount, Invocations, Throttles, CPUUsed-vCPUHours, MemoryUsed-GBHours from the AWS/Bedrock-AgentCore CloudWatch namespace

The skill is entirely read-only: no Create*/Update*/Delete*, no InvokeAgentRuntime data-plane calls, and it reads no memory record content, prompts, or responses. Incomplete signals (missing permissions, no queryable metric) degrade to a documented "visibility limit" rather than producing a false finding.

IAM / two modes. bedrock-agentcore: is not in the standard AIDevOpsAgentAccessPolicy. The skill supports an observability-only mode (CloudWatch only, no additions) and a full control-plane mode (read-only bedrock-agentcore: List/Get set + ec2:DescribeSubnets). This PR adds a parameter-gated policy block (EnableAgentCoreOpsReview) to cloudformation/devops-agent-skill-policies.yaml, following the existing per-skill convention, so the control-plane permissions are opt-in.

Differentiation. Complements bedrock-adoption-readiness (foundation-model workloads) — this skill reviews the AgentCore runtime/memory/gateway layer. The observability-configuration boundary is deferred to agentcore-observability-setup; this skill consumes that telemetry for posture and does not re-report wiring gaps.

Files:

• skills/agentcore-ops-review/ — SKILL.md, README.md (with non-production disclaimer), CHANGELOG.md, references/ (check definitions, report template, sample IAM policies), evals/, .skilleval.yaml
• cloudformation/devops-agent-skill-policies.yaml — opt-in EnableAgentCoreOpsReview parameter, condition, policy, and summary

Type of change

• New skill
• New custom agent
• Update to an existing skill or agent
• Documentation or infrastructure change

Testing

Manual (DevOps Agent, live account): Uploaded to an Agent Space (Chat + Evaluation) with the full control-plane policy attached. Ran natural-language prompts (without naming the skill) and confirmed via the reasoning trace that the skill activated and executed the checks. Against a real estate (2 runtimes, 6 memories, 1 gateway, 1 browser, 1 code interpreter, 3 workload identities) it correctly flagged a Critical gateway finding (READY but zero targets, no policy engine), a 239-day idle runtime, ~22% estate utilization, and strategy-aware memory findings. Unreadable memory record counts and absent runtime metrics were reported as visibility limits — not false findings — confirming graceful degradation.

Agent Skill Eval: skill-eval audit → 98/100, Grade A, 0 critical, 0 warning (the single INFO, STR-016 README-alongside-SKILL.md, is suppressed via .skilleval.yaml per this repo's convention, as the README is required). Functional (5 cases / 18 assertions) and trigger (3 no-trigger + activation) checks pass; the trigger set correctly abstains on unrelated and foundation-model queries. A full skill-eval report report.json with "passed": true will be attached once run with the Claude CLI.

License confirmation

• By submitting this pull request, I confirm that my contribution is made under the terms of the Apache License 2.0.