Automating the AI Agent Feedback Loop with a CI Monitor Extension

I Built a CI Monitor That Let Me Walk Away From My Terminal

Hector Flores developed a Copilot CLI extension to bridge the gap between AI agent sessions and GitHub Actions pipelines. During a single 78-turn session, the agent autonomously resolved eight deployment failures without human intervention. This transformation shifted the developer’s role from a manual human webhook to a high-level quality reviewer.

Why This Matters

In traditional agentic development, a critical gap exists because agents can write and push code but cannot inherently see the downstream CI/CD results. This forces developers into a babysitting role, manually copying failure logs from browser tabs back into terminal sessions to provide the agent with context. This repetitive cycle of manual intervention acts as a bottleneck for scaling AI-driven engineering tasks.

The ci-monitor extension solves this by creating a closed-loop system where CI status and deployment health checks are piped directly back into the agent’s active session. This technical reality moves beyond the ideal model of prompt engineering and into a production-ready workflow where the agent iterates on its own deployment failures based on real-time feedback. By treating the agent as a participant in the full delivery cycle, engineering teams can focus on evaluating product quality rather than debugging deployment scripts.

Key Insights

• The agent autonomously handled 8 deployment failures involving systemd and CLI flags during a single 78-turn session (Hector Flores, 2026).
• The gh pr checks —watch command allows the extension to block until all GitHub Action checks complete without requiring manual polling logic.
• The session.send() SDK primitive enables background processes to inject messages into active Copilot CLI sessions as if they were user inputs.
• Retrieving the last 3,000 characters of failed job logs via gh run view —log-failed provides diagnostic context without exceeding AI context windows.
• PR comment polling every 60 seconds enables developers to route feedback from the GitHub UI directly back to the active AI agent.
• Hot-reloading via extensions_reload allows for immediate iteration on extension behavior without restarting the agent session.

Working Examples

The core onPostToolUse hook that intercepts git push commands to trigger background CI monitoring.

onPostToolUse: async (input) => { if (/\bgit\b.*\bpush\b/i.test(input.toolArgs?.command)) { startMonitoring(cwd, (msg) => { session.send({ prompt: msg }); }); } }

Function utilizing the GitHub CLI to block the extension until PR checks complete.

async function waitForChecks(cwd, requirePending = false) { try { await runGh(['pr', 'checks', '--watch', '--interval', '15'], cwd, 35 * 60 * 1000); return true; } catch { return false; } }

Practical Applications

• System: GitHub Actions + Copilot CLI. Use case: Intercepting git push to trigger background CI monitoring and automated log retrieval. Pitfall: Monitoring global runs instead of branch-specific ones, leading to incorrect status reporting.
• System: PR Comment Feedback. Use case: Routing manual tester feedback from PR comments back into the agent session for iterative debugging. Pitfall: Flooding the session with non-actionable comments without applying filtering or summary logic.
• System: Deployment Verification. Use case: Using ---DEPLOY-SUMMARY-START--- markers in workflow logs to extract structured health check data for the agent. Pitfall: Relying on brittle parsing of unstructured log lines which may change between different deployment versions.

References:

• https://dev.to/htekdev/i-built-a-ci-monitor-that-let-me-walk-away-from-my-terminal-101d
• https://htek.dev/prompts/ci-monitor-extension.md