Set up Researcher agent: infrastructure, MCP tools, and

Tinker

Home Epics Set up Researcher agent: infrastructure, MCP tools, and guardrailsEdit

Set up Researcher agent: infrastructure, MCP tools, and guardrails

Cancel

Title *

Project *

Description

Set up the basic Researcher agent infrastructure with agent record, MCP tools, and system prompt.

## Agent Setup

**Create researcher agent in database:**
- Name: "tinker-researcher" or similar
- Role: `researcher` (new enum value)
- Status: active
- Capabilities: create_proposal, search_memory, list_tickets, read_code

## MCP Tools for Researcher

**Add to agent-bridge.go:**
```go
// Researcher can read but not modify
- search_memory(query)
- list_tickets(filters)
- get_diff(file_path)
- get_test_coverage(file_path)
- create_proposal(title, type, reasoning, confidence, metadata)
- bulk_create_proposals(proposals)
```

**Guardrails enforced at MCP level:**
- No create_ticket (only create_proposal)
- No update_memory/delete_memory (only propose via create_proposal)
- No write_file/edit_file operations
- Read-only access to codebase

## System Prompt

Researcher gets specific instructions:
```
You are the Researcher - a 24/7 autonomous analyst.

Your job:
1. Read code, tickets, memories continuously
2. Find patterns, problems, opportunities
3. Create proposals (NOT tickets - proposals need approval)

What you CAN do:
- Search memories for patterns
- Read tickets and find recurring issues
- Analyze code for quality issues
- Create proposals for human review

What you CANNOT do:
- Modify code directly
- Create tickets without approval
- Delete/modify memories
- Execute any write operations

Every proposal must include:
- title (clear, concise)
- proposal_type (new_ticket, memory_cleanup, refactor, test_gap, feature)
- reasoning (why this matters)
- confidence (high/medium/low)
- priority (high/medium/low)
- evidence_links (tickets, memories, files that support this)

Expect 90% of proposals to be rejected. That's OK - volume over quality.
```

## Trigger Mechanisms

**Researcher can be started via:**
1. WebSocket message to agent channel
2. Scheduled job (every 15-30 minutes)
3. Manual API trigger
4. On-demand via MCP tool

## Acceptance Criteria
- [ ] Researcher agent record exists
- [ ] `researcher` role enum added
- [ ] MCP tools available (read-only + create_proposal)
- [ ] System prompt configured
- [ ] Can be triggered via WebSocket/API
- [ ] Guardrails prevent unauthorized modifications