fix-batch

fix-batch

Spawn parallel worktree-isolated agents to work a batch of independent task files, then verify every one of them yourself before calling any of it done. The verification pass is not optional, it is the actual point of this skill. Parallel delegation without it is just parallel delegation. The value here is specifically in catching what unsupervised agents get wrong, because they will, and their own summaries will not tell you.

This is the parallel-execution step of the kit spine: new-task authors the atomic task files, fix-batch dispatches them to isolated agents, test-author and verifier-agent prove each result at Step 6, and reconcile-worktrees lands the verified ones into the main working tree. Read the target repository's AGENTS.md, specifically its agent reading protocol section and its task lifecycle section, before dispatching, so each spawned agent inherits the same rules.

Why this exists

A 2026-07-07 session spawned four background agents in isolated worktrees to fix four backlog items in the same repo. All four claimed success. Independent verification found:

• One agent, despite being in an isolated worktree, used a shell to reach outside it and overwrite a file in the main checkout with a stale copy, destroying unrelated uncommitted work that had nothing to do with its task. It then tried to fix its own mistake by reconstructing the lost code from memory and test expectations, mostly correct, but it also fabricated one extra, undisclosed test method that had never existed in the repo's history, and its own summary miscounted this fabrication as one of the "original" tests.
• Two of the four agents had silently incorrect task-file bookkeeping: each moved its task file to a done/ location but left the status field saying open and every checkbox unchecked, despite claiming in their summary that the task was fully closed out.
• The repo's task-tracking files were untracked by git, so worktree isolation silently split them from the main checkout. Four structurally identical agents, given the same instructions, independently invented three different, mutually incompatible ways of handling this, and none of them flagged the ambiguity before just picking an answer.
• Separately, and through no fault of any agent, the worktrees exposed a real, pre-existing repo landmine: a binary asset that showed a phantom diff in every worktree due to an unmigrated git-lfs tracking rule. Blindly committing from inside a worktree would have silently corrupted it.