paseo/REPORT-daemon-e2e-audit.md

Daemon E2E Test Coverage Audit

Date: 2025-12-25 Test File: packages/server/src/server/daemon.e2e.test.ts

Test Run Summary

5 passed, 2 skipped in 40.8s

All active tests pass. The 2 skipped tests are Claude permission flow tests (documented issue with Claude SDK behavior in daemon context).

---

Current Test Coverage

Test: creates agent and receives response

• Provider: Codex
• Coverage:
  • createAgent() → creates agent, returns AgentSnapshotPayload
  • sendMessage() → sends text message to agent
  • waitForAgentIdle() → waits for agent to complete
  • Verifies agent_state message with status: "idle"
  • Verifies agent_stream events: turn_started, turn_completed, timeline (assistant_message)

Test: permission flow: Codex > approves permission and executes command

• Provider: Codex
• Coverage:
  • waitForPermission() → receives permission_requested event
  • respondToPermission() → sends allow response
  • Verifies permission_resolved event with behavior: "allow"
  • Verifies timeline has tool_call with status: "granted"
  • Verifies file system side-effect (file created)

Test: permission flow: Codex > denies permission and prevents execution

• Provider: Codex
• Coverage:
  • waitForPermission() → receives permission_requested event
  • respondToPermission() → sends deny response with message
  • Verifies permission_resolved event with behavior: "deny"
  • Verifies timeline has tool_call with status: "denied"
  • Verifies file system side-effect (file NOT created)

Test: persistence flow > persists and resumes Codex agent

• Provider: Codex
• Coverage:
  • Agent creation and messaging
  • deleteAgent() → deletes active agent
  • Verifies agent_deleted event
  • resumeAgent(handle) → resumes from persistence handle
  • Verifies resumed agent can receive messages
  • Verifies conversation context preserved

Test: multi-agent orchestration > parent agent creates child via agent-control MCP

• Provider: Codex (parent and child)
• Coverage:
  • Parent agent uses agent-control MCP to call create_agent
  • Verifies tool call in timeline: tool: "create_agent", server: "agent-control"
  • Verifies both parent and child visible in agent_state messages
  • Verifies child agent ID extracted from tool output

Skipped: permission flow: Claude > approves/denies permission

• Reason: Claude SDK doesn't request permissions in daemon context
• Note: Direct claude-agent.test.ts permission tests pass; issue is daemon-specific

---

DaemonClient API Coverage

Method	Tested	Notes
connect()	✅	Used in beforeEach via createDaemonTestContext()
close()	✅	Used in afterEach via cleanup()
createAgent()	✅	All tests
deleteAgent()	✅	Persistence test
listAgents()	❌	Not explicitly tested
listPersistedAgents()	❌	Not tested (Codex doesn't implement it)
resumeAgent()	✅	Persistence test
sendMessage()	✅	All tests
cancelAgent()	❌	Not tested
setAgentMode()	❌	Not tested
respondToPermission()	✅	Permission tests
waitForAgentIdle()	✅	All tests
waitForPermission()	✅	Permission tests
on() (event subscription)	❌	Not tested directly
getMessageQueue()	✅	Used for assertions
clearMessageQueue()	✅	Used to isolate test phases

---

Message Protocol Coverage

Inbound Messages (Client → Daemon)

Message Type	Tested
create_agent_request	✅
delete_agent_request	✅
send_agent_message	✅
agent_permission_response	✅
resume_agent_request	✅
cancel_agent_request	❌
set_agent_mode	❌
list_persisted_agents_request	❌
refresh_agent_request	❌
initialize_agent_request	❌
git_diff_request	❌
file_explorer_request	❌
git_repo_info_request	❌
clear_agent_attention	❌
list_provider_models_request	❌

Outbound Messages (Daemon → Client)

Message Type	Tested
agent_state	✅
agent_stream	✅ (partial - timeline, turn events, permissions)
session_state	❌ (received but not explicitly verified)
agent_deleted	✅
agent_permission_request	✅
agent_permission_resolved	✅
list_persisted_agents_response	❌
status	❌
activity_log	❌
assistant_chunk	❌
audio_output	❌ (realtime mode)
transcription_result	❌ (realtime mode)
artifact	❌
conversation_loaded	❌
git_diff_response	❌
file_explorer_response	❌
git_repo_info_response	❌
list_provider_models_response	❌

---

Agent Provider Coverage

Provider	Basic Flow	Permissions	Persistence	Multi-agent
Codex	✅	✅	✅	✅
Claude	❌	⏸️ (skipped)	❌	❌

---

Coverage Gaps & Recommendations

Priority 1: High Value / Low Effort

1. cancelAgent() test

   • Cancel an agent mid-execution
   • Verify agent stops and reaches idle/error state
   • Currently untested DaemonClient method

2. setAgentMode() test

   • Create agent in one mode, switch to another
   • Verify mode change reflected in agent state
   • Currently untested DaemonClient method

3. listAgents() test

   • Connect client, call listAgents()
   • Verify session_state returns current agents
   • Currently the method exists but is never called in tests

Priority 2: Provider Parity

4. Claude basic flow test

   • Currently no passing Claude tests
   • Add simple "creates agent and receives response" for Claude
   • Investigate why Claude permissions behave differently in daemon

5. Claude persistence test

   • Resume Claude agent from persistence handle
   • Verify conversation context preserved

Priority 3: Edge Cases

6. Error handling test

   • Agent encounters error during execution
   • Verify error state, lastError field populated
   • Verify recovery (can send new message after error)

7. Connection handling test

   • Reconnect after disconnect
   • Multiple simultaneous clients
   • Rate limiting/timeout behavior

8. Message ordering test

   • Send multiple messages rapidly
   • Verify ordering preserved
   • Verify no race conditions

Priority 4: Feature Coverage (Lower Priority)

9. Image attachment test

   • sendMessage() with images option
   • Requires multimodal agent support

10. Git integration tests

    • git_diff_request/response
    • git_repo_info_request/response
    • Requires git repository setup

11. File explorer tests

    • file_explorer_request/response
    • Navigate filesystem via daemon

---

Summary

Current coverage: Core happy-path scenarios for Codex agents (create, message, permission, persistence, multi-agent).

Gaps:

• No Claude provider tests pass (permissions issue)
• Several DaemonClient methods untested (cancelAgent, setAgentMode, listAgents)
• No error/edge case testing
• No tests for auxiliary features (git, file explorer, models)

Recommendation: Focus on Priority 1 & 2 items before expanding to edge cases. The Claude permission investigation should be a separate task as it involves SDK behavior analysis.