Claude SDK Agent

Back to agent catalog | main doc

Type: Custom A2A Framework: Anthropic SDK / OpenAI-compatible (httpx) LLM: LiteMaaS (llama-scout-17b) Supervisor: No Sandbox Model: Tier 3 (plain Deployment, no supervisor) Status: Deployed, tested (Kind + HyperShift)

1. Overview

Code review agent using the Anthropic Python SDK with automatic format switching. When ANTHROPIC_BASE_URL points to a non-Anthropic endpoint (e.g., LiteMaaS), uses httpx with OpenAI chat/completions format. A2A protocol implemented manually via Starlette (Anthropic SDK has no built-in A2A wrapper unlike Google ADK).

2. Architecture

3. Files

deployments/openshell/agents/claude-sdk-agent/
├── agent.py              # Starlette A2A server + Anthropic/OpenAI client
├── Dockerfile            # python:3.12-slim
├── deployment.yaml       # Deployment + Service + AgentRuntime CR
├── policy-data.yaml      # OPA policy
├── sandbox-policy.rego   # OPA Rego rules
└── requirements.txt      # anthropic, starlette, uvicorn, httpx

4. Deployment

Same as adk-agent (docker build + kind load, or OCP binary build).

5. Capabilities

Capability	Supported	Notes
A2A protocol	Yes	Manual Starlette implementation
Multi-turn context	No	Stateless — each request independent
Tool calling	No	LLM prompt-based only
Subagent delegation	No	Single-purpose agent
Memory/knowledge	No	No persistent state
Skill execution	Via prompt	Skill markdown injected into system prompt
HITL approval	L0	OPA policy mounted, not enforced

6. Rossoctl Integration

6.1 Communication Adapter

A2A JSON-RPC (already implemented).

6.2 Session Management

None — stateless. Backend stores turns in PostgreSQL.

6.3 Observable Events

Event	Source	Rossoctl UI Component	Phase
LLM response	A2A response artifacts	AgentChat	Current
Error (LLM unavailable)	Generic error message	AgentChat	Current

6.4 FileBrowser Integration

N/A — no workspace.

7. LLM Compatibility

Provider	Protocol	Works?	Notes
LiteMaaS	OpenAI-compat	Yes	Auto-detected via base_url
Anthropic API	Claude messages	Yes	Native SDK
Budget Proxy	OpenAI-compat	Yes	Default config

8. Skill Execution

The Claude SDK agent is the most capable skill executor in the PoC — it supports all skill types and has the most passing skill tests. Skills are injected by embedding the skill markdown (from .claude/skills/<name>/SKILL.md) into the A2A prompt.

Supported Skills

Skill	Test	Status	How It Works
PR Review	test_pr_review__claude_sdk_agent	PASS	Skill markdown from github:pr-review injected into prompt
RCA	test_rca__claude_sdk_agent	PASS	Skill markdown from rca:ci injected; analyzes CI failure logs
Security Review	test_security_review__claude_sdk_agent	PASS	Prompt-based K8s manifest security analysis
Real GitHub PR	test_review_real_github_pr__claude_sdk	PASS	Fetches PR #1300 diff via GitHub API, reviews with LLM
RCA CI Logs	test_rca_style_log_analysis__claude_sdk	PASS	Analyzes real CI log output for root causes
Code Generation	test_code_generation__claude_sdk	PASS	Generates Python code from natural language prompt
TDD	Not tested	—	Can execute test:review skill via prompt injection
Docs Review	Not tested	—	Can execute docs:review skill via prompt injection

Running Skills Manually

# Port-forward to Claude SDK agent
kubectl port-forward -n team1 svc/claude-sdk-agent 8002:8000 &

# PR Review skill
curl -s -X POST http://localhost:8002/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0", "id": "1", "method": "message/send",
    "params": {"message": {"role": "user",
      "parts": [{"type": "text", "text": "Review this code change for issues:\n\n```diff\n- password = request.form[\"password\"]\n+ password = hashlib.md5(request.form[\"password\"]).hexdigest()\n```"}]
    }}
  }' | python3 -m json.tool

# RCA with real CI log
curl -s -X POST http://localhost:8002/ \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0", "id": "2", "method": "message/send",
    "params": {"message": {"role": "user",
      "parts": [{"type": "text", "text": "Analyze this CI failure:\n\nERROR: Connection refused to postgres:5432\nFATAL: role \"test_user\" does not exist\n\nContext: This started after we upgraded the PostgreSQL Helm chart from 14.x to 15.x"}]
    }}
  }' | python3 -m json.tool

kill %1

Prerequisites

• LiteLLM model proxy running in team1
• OPENSHELL_LLM_AVAILABLE=true for E2E tests
• .env.maas with LiteMaaS credentials (or substitute your LLM endpoint)

9. Testing Status

Test File	Tests	Pass	Skip	Notes
test_02_a2a_connectivity	2	2	0	Hello + agent card
test_05_multiturn	3	2	1	Sequential + isolation pass; continuity skips
test_07_skill_execution	7	5	2	PR review, RCA, security, real GH PR, RCA logs

10. Sandbox Deployment Models

Model	Supported	Notes
Mode 1: Rossoctl Deployment	Current	Standard Deployment + Service
Mode 1 + Supervisor	Possible	Would enable OPA enforcement
Mode 2: Sandbox CR	Not applicable	Custom code, not a CLI agent