Forge: Turn Issues into Shipped Fixes with Agent Approval

Where this fits

A developer drops a bug report into the issue tracker, then asks their agent to "investigate and fix it." The agent starts editing code immediately—often on the wrong branch or without a clear plan. Forge solves this by forcing a two-phase workflow: first classify, fetch, and propose a plan; then, only after explicit user approval, execute the implementation lifecycle. If you've ever wanted a safety gate between investigation and action, Forge is the missing handrail.

Why agents benefit

• Enforces a plan gate – The agent must present a clear, branch-correct plan before any code changes. This prevents destructive premature edits and keeps the human in the loop when it matters.
• Structured lifecycle – Steps 1–12 are numbered and split: gate (1–6) then autonomous execute (7–12). The agent knows exactly where it is and what to do next, reducing improvisation.
• Autonomous review and refactor loop – After implementation, the agent reviews its own work, refactors, and even spins off sub-tasks—without needing constant human nudging.
• Branch and context awareness – Forge fetches the issue, checks out the right branch, and loads relevant context, so the agent operates with full situational awareness.
• Controlled automation – The automode flag offers a sanctioned bypass of the gate when you trust the agent fully, giving you flexibility without losing safety.

Practical scenarios

Critical bug triage

Your agent receives a link to a production incident. Forge classifies it, fetches the issue, proposes a hotfix plan, and waits for your "yes, implement." You review the diff scope, approve, and the agent runs the entire fix, tests, and refactor without further prompts.

Multi-ticket sprint cleanup

You have a backlog of minor issues. Run /forge on each with automode enabled. The agent works through them sequentially: propose, implement, verify, close. You only step in when a plan looks risky.

Post-PR review refactoring

A reviewer flags structural issues in an open PR. You run /forge pr-123 and the agent analyzes the diff, proposes a refactoring plan, then after approval implements it, updates tests, and pushes—keeping the PR branch intact.

Add it to your agent workflow

Install the skill with:

npx skills add radimsem/forge-skills --skill forge

Then, in your agent interface (e.g., Claude Code, Cursor), simply mention an issue or run /forge #123. The agent will start Part 1, present a plan, and hold for approval. Once you confirm, it executes the full lifecycle automatically.

Compare before adopting

Before committing, check a few things:

• Quality signals – The repo has 15 stars, but the README is detailed and the design is deliberate. Look for recent commits or issues.
• Maintenance freshness – Is it actively updated? A stalemated skill can break with newer agent versions.
• Alternatives – Many skills skip the approval gate. If you need that control, Forge is rare. But if you prefer full autonomy, a simpler skill might suffice.
• Workflow fit – Does the numbered, gate-first pattern match how you want agents to work? If you already have strong review discipline, the gate might be redundant.

Why it is worth tracking

The approval-gate pattern is a meaningful shift toward safe agent automation. Even with modest community traction, Forge demonstrates a workflow that many developers will eventually need. As agents become more autonomous, having a skill that insists on a human checkpoint before mutating code is a model to watch. Track this skill to see how the pattern evolves—especially if you're building automation that must stay trustworthy. For a deeper look, see the skill page on OpenAgentSkill.