⚙️ FEATURE: Sub-agent peer review — route written code to a verification sub-agent

[FernandoCelmer]

Problem

pycodeloop's DelegateTool spawns sub-agents for task delegation, but there is no built-in pattern for peer review: after a coder sub-agent writes code, a second read-only sub-agent verifies it (review, security check, test coverage analysis).

"Code as Agent Harness" (arXiv 2605.18747) identifies multi-agent peer review via shared code artifacts as a key scaling pattern for reliable long-horizon agent execution.

Expected behavior

from pycodeloop import CodeLoop, PeerReviewWorkflow

result = await PeerReviewWorkflow(
    coder=CodeLoop(config=coder_config),
    reviewer=CodeLoop(config=reviewer_config),  # read-only tools
).run("Implement JWT authentication")
# reviewer output appended to session trace

Or via built-in Config:

Config(peer_review=True, reviewer_system_prompt="You are a security reviewer...")

Suggested implementation

• Add PeerReviewWorkflow class in pycodeloop/workflows/peer_review.py
• Coder agent runs first with full tool access
• After coder finishes, collect list of files written from JSONL trace
• Reviewer agent receives those files as context (read-only tools) + reviewer system prompt
• Reviewer output stored as {"type": "peer_review", "findings": [...]} in JSONL trace
• If reviewer finds blocking issues, feed back to coder for a fix cycle

References

• "Code as Agent Harness" — arXiv 2605.18747, Section 5 (Multi-agent peer review)