💡 Intelligent CI Failure-to-Agent Router

[github-actions[bot]]

Summary

Build a routing layer that automatically dispatches CI failures and bot review comments to the most appropriate agentic resolver — Claude Code dev-lead for complex logic fixes, Copilot cloud agent for simple one-click Action fixes, and CodeRabbit Autofix for style and lint issues — based on failure type classification. This reduces human toil in the fix-review-push cycle and optimizes cost by matching problem complexity to agent capability.

Market Signal

GitHub launched "One-click fixes for failing Actions with Copilot cloud agent" (May 18, 2026) and a REST API to start Copilot cloud agent tasks programmatically (May 13). CodeRabbit launched Autofix (April 2026) that spawns coding agents to commit fixes. Claude Code Action supports intelligent mode detection. The industry is moving from single-agent to multi-agent orchestration — GitHub, Anthropic, and CodeRabbit are all shipping agent-to-agent handoff capabilities.

User Signal

Open issue #291 "fix(claude-code): wire agentic responses to trusted bot reviews and CI failures" directly requests this capability. The org already runs three agent systems (Claude Code via dev-lead.yml, Copilot code review, CodeRabbit) but they operate independently with no coordination. PR #323 "auto-trigger PR review when all readiness criteria are met" shows the pattern of event-driven agent triggering. The current approach sends all failures to Claude Code Opus, which is expensive for simple lint fixes.

Technical Opportunity

The dev-lead.yml workflow already responds to PR events. GitHub's check_suite and check_run webhook events provide structured failure data including the check name, conclusion, and output. The Copilot cloud agent REST API (GA May 13) enables programmatic task dispatch. A classifier step could route: lint/format failures → CodeRabbit Autofix, simple CI failures → Copilot cloud agent, complex logic/test failures → Claude Code dev-lead.

Assessment

Dimension	Score	Rationale
Feasibility	med	Requires failure classification logic and multi-agent API integration
Impact	high	Reduces human toil, optimizes agent costs, addresses open issue #291
Urgency	high	Issue #291 is an active feature request; Copilot APIs just became available

Adversarial Review

Strongest objection: Multi-agent routing is over-engineering. The org currently uses Claude Code dev-lead for everything and it works. Adding routing logic increases complexity, creates failure modes, and may cost more by invoking multiple agent systems.
Rebuttal: The org already pays for three agent systems running independently on every PR. Simple fixes like a missing import or lint violation don't need Opus-class reasoning — routing them to cheaper, faster agents saves both time and money. Issue #291 confirms this is a real user need. The routing logic is a single classifier step in the existing workflow, not a new system. Copilot's "one-click fixes for failing Actions" and CodeRabbit's Autofix show the industry is converging on specialized, cost-optimized agent routing.

Suggested Next Step

Prototype a failure classifier in the dev-lead.yml workflow that categorizes CI check failures (lint, type, test, build, security) and routes simple failures to Copilot cloud agent via REST API while reserving Claude Code for complex multi-file fixes. Related: Issue #291.