llm-multiverse/implementation-plans/issue-086.md

Issue #86: Implement assistant agent type

Metadata

Field	Value
Issue	#86
Title	Implement assistant agent type
Milestone	Phase 10: Remaining Agent Types
Status	PLANNED
Language	Python
Related Plans	issue-069.md, issue-081.md
Blocked by	#69, #52

Acceptance Criteria

• Assistant system prompt emphasizing helpful conversation, summarization, Q&A
• Agent manifest with allowed tools: web_search, memory_query only
• No file system or code execution access
• No network egress beyond search
• Uses the same agent loop infrastructure as researcher
• Optimized for conversational response quality

Architecture Analysis

Service Context

• Orchestrator service — new agent type alongside researcher and coder
• Uses the same PromptBuilder with a custom ASSISTANT_SYSTEM_PROMPT
• Routes via AGENT_TYPE_ASSISTANT (proto value 5) in service.py

Existing Patterns

• researcher.py — base pattern (web_search + memory_read only)
• coder.py — extended pattern with tool-output source tracking
• prompt.py — parameterized PromptBuilder with system_prompt field
• config.py — AgentConfig dataclass reused per agent type

Dependencies

• common_pb2.AGENT_TYPE_ASSISTANT (value 5) — already defined in proto
• No new proto changes needed
• No new external libraries

Implementation Steps

1. Types & Configuration

• Add assistant: AgentConfig field to Config dataclass
• Add YAML loading for assistant section in Config.load()

2. Core Logic

• Create AssistantAgent class in assistant.py
• Mirror researcher pattern exactly (web_search + memory_read tools only)
• No _TOOL_OUTPUT_SOURCES tracking (assistant has no fs/code tools)
• Source tracking: web_search → RESULT_SOURCE_WEB, else RESULT_SOURCE_MODEL_KNOWLEDGE
• Agent ID prefix: ast-{uuid}
• Agent-specific messages: "Assistance incomplete: {reason}", "Assistance failed."

3. System Prompt

• Add ASSISTANT_SYSTEM_PROMPT to prompt.py
• Conversational tone, emphasizing helpful responses, summarization, Q&A
• Capabilities: web_search, memory_read only
• No file system or code references
• Encourage clear, well-structured answers

4. Service Integration

• Import AssistantAgent in service.py
• Instantiate with self._config.assistant
• Route AGENT_TYPE_ASSISTANT subtasks to assistant.run()

5. Tests

• Create tests/test_assistant.py mirroring test_researcher.py
• Key assistant-specific tests:
  • Uses AGENT_TYPE_ASSISTANT for tool discovery/execution
  • Agent ID prefix is ast-
  • Uses ASSISTANT_SYSTEM_PROMPT
  • Only web_search and memory_read tools available
  • Source tracking: web → WEB, default → MODEL_KNOWLEDGE
  • No tool-output source (no fs/code tools)
• Standard tests: happy path, tool discovery, empty tools, execution, memory, max iterations, timeout, failures, gateway errors, confidence, unknown tools, discovery errors, memory unavailable, parse errors, reasoning, execution errors, compaction

Files to Create/Modify

File	Action	Purpose
services/orchestrator/src/orchestrator/assistant.py	Create	AssistantAgent class
services/orchestrator/tests/test_assistant.py	Create	Unit tests
services/orchestrator/src/orchestrator/prompt.py	Modify	Add ASSISTANT_SYSTEM_PROMPT
services/orchestrator/src/orchestrator/config.py	Modify	Add assistant AgentConfig
services/orchestrator/src/orchestrator/service.py	Modify	Route AGENT_TYPE_ASSISTANT

Risks and Edge Cases

• Assistant is closest to researcher pattern — minimal risk
• Must ensure no fs/code tools leak through (enforced by tool broker manifest, not agent code)

Deviation Log

Deviation	Reason