Skip to main content

Multi-Turn Conversation

Test file: rossoctl/tests/e2e/openshell/test_05_multiturn_conversation.py Tests: 12 | Pass: 9 | Skip: 3 (Kind, fresh cluster)

What This Tests​

Validates that agents handle sequential A2A messages with context continuity, isolation between independent conversations, and service persistence across delayed requests.

Architecture Under Test​

Test Matrix​

Testweather_agentadk_agentclaude_sdk_agentweather_supervisedos_claudeos_opencodeos_generic
3 sequential messages✅✅✅✅ (exec)———
Context isolation✅✅✅⏭️ needs A2A———
Context continuity⏭️ stateless⏭️ ADK gap⏭️ stateless⏭️ Phase 2———
Service persistence✅✅✅————

Skip reasons:

  • stateless — Agent doesn't return contextId (Rossoctl backend will manage context externally)
  • ADK gap — ADK to_a2a() doesn't support client-sent contextId (upstream limitation)
  • needs A2A — Supervised agent netns blocks port-forward; ExecSandbox gRPC needed for multi-turn
  • Phase 2 — Requires PVC-backed session store + Rossoctl backend integration
  • — — Builtin sandboxes use terminal sessions (persistent by nature), not A2A

Test Details​

test_multiturn__agent__responds_to_3_sequential_messages (parametrized: 3 A2A agents)​

  • What: Send 3 sequential messages and verify responses
  • Asserts:
    • Each turn has "result" in response
    • Text not empty for each turn
    • contextId extracted and passed to next turn
  • Debug points: Response structure, text length, contextId
  • Agent coverage: weather_agent, adk_agent, claude_sdk_agent
  • Skip condition: LLM agents skip if OPENSHELL_LLM_AVAILABLE != true

test_multiturn__weather_supervised__kubectl_exec​

  • What: Supervised agent: test via kubectl exec (netns blocks port-forward)
  • Asserts: exec succeeds, stdout contains "alive"
  • Debug points: exec returncode, stderr
  • Agent coverage: weather_supervised
  • Note: Basic liveness test only; multi-turn requires ExecSandbox gRPC

test_context_isolation__agent__independent_requests_isolated (parametrized: 3 A2A agents)​

  • What: Two independent requests should have different contextIds
  • Asserts: contextId from request A != contextId from request B
  • Debug points: contextIds, response text
  • Agent coverage: weather_agent, adk_agent, claude_sdk_agent
  • Skip condition: Agent doesn't return contextId OR LLM not available

test_context_isolation__weather_supervised__netns_blocks_test​

  • What: Supervised agent: context isolation test requires A2A
  • Asserts: N/A (always skips)
  • Agent coverage: weather_supervised
  • Skip reason: netns blocks port-forward; TODO: ExecSandbox gRPC integration

test_context_continuity__agent__context_preserved_across_turns (parametrized: 3 A2A agents)​

  • What: If agent returns contextId, it should persist across turns
  • Asserts:
    • Turn 1 returns contextId
    • Turn 2 with same contextId returns same or related contextId
  • Debug points: contextId values across turns
  • Agent coverage: weather_agent, adk_agent, claude_sdk_agent
  • Skip condition:
    • Agent is stateless (no contextId)
    • ADK upstream doesn't preserve client-sent contextId
    • TODO: Rossoctl backend session store will manage context externally

test_context_continuity__weather_supervised__requires_grpc​

  • What: Supervised agent: context continuity requires ExecSandbox gRPC
  • Asserts: N/A (always skips)
  • Agent coverage: weather_supervised
  • Skip reason: Phase 2 integration needed

test_service_persistence__agent__responds_after_delay (parametrized: 3 A2A agents)​

  • What: Send message, wait 10s, send again — agent should still respond
  • Asserts:
    • First request succeeds
    • Second request (after delay) succeeds
  • Debug points: Response text, delay duration
  • Agent coverage: weather_agent, adk_agent, claude_sdk_agent
  • Note: Validates that agents are long-running services (not ephemeral pods)

Context Management Models​

Agent TypeContext StoragecontextId SupportFuture State
weather_agentNone (stateless)NoRossoctl backend session store
adk_agentIn-memoryReturns contextId, ignores client'sUpstream PR or backend override
claude_sdk_agentNone (stateless)NoRossoctl backend session store
weather_supervisedNone (stateless)NoBackend + ExecSandbox gRPC
openshell_claudePVC /workspace/.claude/Native (via Claude API)Phase 2 provider integration
openshell_opencodePVC /workspace/Via LiteLLMPhase 2 provider integration

Agent-Specific Prompts​

Tests use type-appropriate prompts for each agent:

AGENT_PROMPTS = {
    "weather-agent-supervised": [
        "What's the weather in Seattle?",
        "What about New York?",
        "And in Tokyo?",
    ],
    "adk-agent-supervised": [
        "Hello, who are you?",
        "What can you help with?",
        "Thank you!",
    ],
    "claude-sdk-agent": [
        "Hello, who are you?",
        "What capabilities do you have?",
        "Goodbye!",
    ],
}

Future Expansion​

Agent TypeWhen AddedWhat's Needed
openshell_claudePhase 2ExecSandbox gRPC adapter for terminal session → A2A translation
openshell_opencodePhase 2ExecSandbox gRPC adapter for terminal session → A2A translation
Context continuityPhase 2Rossoctl backend session store (PVC-backed PostgreSQL + checkpointer)
ADK contextUpstream PRADK to_a2a() must respect client contextId

Common Failure Modes​

SymptomCauseFix
Empty responseAgent initialization delayAdd retry with exponential backoff
contextId changesAgent restarts between turnsAdd PVC-backed checkpointer
Connection refusedPort-forward killedReestablish port-forward between turns
Timeout on turn 3Agent resource exhaustionCheck pod memory/CPU limits