ForgeLoopField notes
Better systems around coding agents.
Practical engineering notes for teams moving from generated code to verifiable software delivery. No invented benchmarks. No shortcut around review.
Runner operations
Recover a Self-Hosted AI Coding Runner
Use a practical restart recovery checklist for self-hosted AI coding runners: verify durable state, credentials, heartbeats, leases, and safe task resumption.
Read field noteGitHub workflows
From GitHub issues to AI pull requests
Build an issue-to-PR workflow with explicit intake rules, bounded tasks, verification evidence, and a deliberate approval point.
Read field noteArchitecture
Coding agents vs. delivery harnesses
Understand the difference between an AI coding agent and the orchestration, policy, evidence, and recovery system around it.
Read field noteVerification
Acceptance criteria coding agents can verify
Write observable acceptance criteria for AI coding tasks, including failure paths, permissions, test evidence, and explicit scope boundaries.
Read field noteRunner operations
Self-hosted AI runners: what stays local?
Understand what a self-hosted coding runner controls, what data can reach a model provider, and why local execution still needs security boundaries.
Read field noteGitHub workflows
Auto-merge for AI PRs: checks before trust
Design guarded auto-merge around required checks, exact commits, repository policy, and human escalation rather than model confidence.
Read field noteCost control
Budgeting AI coding agents without guesswork
Track model usage, distinguish unknown costs from zero, and bound retries and task scope before starting autonomous coding work.
Read field noteReliability
Designing repair loops that know when to stop
Make AI repair loops bounded and useful with structured failure context, retry classifications, durable attempts, and human escalation.
Read field noteArchitecture
Git worktrees for parallel coding agents
Use separate worktrees and explicit path ownership to reduce parallel agent conflicts, while keeping integration and host security separate.
Read field noteTesting
Test AI workflows without spending API credits
Separate free deterministic tests from paid model validation using fake providers, local servers, persistence checks, and explicit evidence boundaries.
Read field noteIntegrations
Reliable GitHub webhook intake for AI agents
Design webhook intake with signature verification, idempotency, authorization, and clear issue eligibility before dispatching agent work.
Read field note