DocsHelp
Error recovery
What to do when a test fails, a build breaks, work goes missing or a ship run stops. The scaffolded AGENTS.md carries most of this table, so the agent has it in every session.
The scaffolded AGENTS.md has a “When something goes wrong” table for the first six situations below; the last row is about the ship runner.
| Situation | Recovery |
|---|---|
| A test fails after a code change | Don’t iterate blindly; use the ab-systematic-debugging skill |
| Merge conflict | Read both sides, understand the intent, then resolve |
| Broken build after a dependency update | Pin the previous version and put the upgrade in BACKLOG.md |
| Corrupted worktree | Create a fresh one from main and cherry-pick the finished commits |
| A helper returns bad results | Verify the findings before acting on them |
| Lost uncommitted changes | Check git stash list, git reflog, git fsck --lost-found |
| A ship run stalls or stops | Read the runner’s last output and exit code. .agent-blueprint/ holds a reason when the skill stopped the run; after a stop by the runner it can still say running or done. The ab-forensics skill reads the logs, run state and git history |