Humans in the loop¶
Live
Forge is autonomous between your decisions, not instead of them. This page lists where a person decides and what Forge does when it should not.
What you approve¶
- The plan. Nothing is built until you reply
/proceed(or the project uses auto-approve). The plan includes a cost and risk band. - Spending over budget. If a build is projected to exceed the remaining monthly budget, it is blocked until an approver overrides with
/proceed --force. The approver's name is recorded. - Pull requests. Forge opens pull requests and never merges them. You merge them. By default Forge writes only to your
developandtestbranches and opens the pull request intomain. See Branches, PRs and deploys. - Anything it cannot safely decide. See below.
When Forge stops and asks¶
- The epic and the code disagree about what exists (a hard domain mismatch).
- An agent hits a decision it should not make alone. It posts a question on the issue and waits for
/answer. - A migration slice contains a hard blocker, such as WebForms or AppDomain use. The slice is paused with numbered options and the exact reply to resume. See Modernization.
- Automatic retries are used up. The epic escalates with a diagnosis rather than looping.
- A deploy failure shows a credentials or permissions problem. Forge escalates immediately instead of retrying.
Controls you have at any time¶
| Command | Effect |
|---|---|
/pause and /resume |
Stop Forge from starting new work, then continue. Running agents finish. |
/abort |
Stop everything on the epic, or on one story when run on the story issue. |
/intervene <guidance> |
Give an agent guidance on a cross-repo epic. |
/skip |
Skip a story. |
/answer |
Answer an agent's question. |
Full list: Slash commands.
Your own rules¶
Project guardrails and repo skills apply to every agent, and reviewers flag violations. Only people write guardrails.
What a human always reviews¶
Security review and code review by agents do not replace a person reading the pull request. They reduce what the person has to catch.