Multi-agent architecture: review + fix + test agents

[haasonsaas]

Problem

The AI code review market is moving to agentic architectures. CodeRabbit and GitHub Copilot both went agentic in 2025. The pattern:

• Review agent posts findings → user replies "@diffscope fix this" → coding agent opens a fix PR
• Review agent detects missing tests → test generation agent writes them
• Triage agent reads the ticket → review agent validates against acceptance criteria
• Security agent runs a dedicated deep-dive → feeds into the main review

DiffScope's existing multi_pass_specialized config runs sequential LLM calls. These should be independent agents with tool use, composable into workflows.

Proposed Solution

Agent Types

1. Review Agent (exists today — enhance with tool use)

   • Can read files from the repo (not just the diff)
   • Can query the symbol graph
   • Can run linters/scanners
   • Posts structured findings

2. Fix Agent (new)

   • Triggered by user command: @diffscope fix on a review comment
   • Reads the review finding + surrounding code
   • Generates a fix as a commit on the PR branch
   • Or opens a separate fix PR

3. Test Agent (new)

   • Triggered when review detects missing test coverage
   • Generates unit tests for changed functions
   • Opens as a commit or separate PR

4. Triage Agent (new)

   • Reads linked tickets (Jira/Linear/GitHub Issues)
   • Summarizes requirements for the review agent
   • Validates completion after review

Orchestration

PR Opened
  ├── Triage Agent → fetches ticket context
  ├── Security Agent → dedicated security scan  
  ├── Review Agent → code review (receives context from triage + security)
  └── Summary Agent → generates PR description + diagrams
  
User replies "@diffscope fix #3"
  └── Fix Agent → generates fix commit for finding #3

User replies "@diffscope test"
  └── Test Agent → generates tests for changed code

Interactive Commands (via PR comments)

• @diffscope review — trigger full review
• @diffscope fix <comment-id> — auto-fix a specific finding
• @diffscope test — generate tests for changed code
• @diffscope explain <file:line> — explain code in context
• @diffscope ignore <comment-id> — suppress a finding

Architecture

• Each agent is a separate async task with its own LLM context
• Agents communicate via structured messages (not shared context)
• Orchestrator manages agent lifecycle and result aggregation
• Webhook receiver dispatches to appropriate agents based on event type

Competitive Context

• CodeRabbit: Agentic commands — generate tests, create issues, write docstrings from PR comments
• GitHub Copilot: Spawns Copilot coding agent to fix flagged issues
• Greptile v4: Agent architecture launched late 2025
• Ellipsis: @ellipsis-dev can be assigned tasks like a human team member

Priority

Tier 1 — Strategic. This is where the entire market is heading. The multi-agent architecture is the highest-leverage investment for DiffScope's competitive position.