Instruction file imported from darkforce09/TBD-reforger (.cursor/rules/no-duplicate-slice-agents.mdc). Copyright stays with the author.

No duplicate slice / worktree agents (HARD)

Paid for three times already (T-860 wave 220). Launching a "fresh" or "redo" agent while another is still bound to the same ticket/worktree races the tree, doubles cost, and enrages the operator. This is not a soft preference.

Forbidden

❌ Task/subagent for T-XXX while any prior Task for T-XXX is still running
❌ "Fresh redo" / "interrupt and take over" / second generalPurpose on the same
   `.ai/artifacts/worktrees/T-XXX` because the first looks slow or stuck
❌ Parallel agents that both write `tools_v2/xtask/src/main.rs` or the same owns paths
   for the same ticket (even if you think the first is idle)
❌ Interpreting a short transcript, long silence, or "only untracked files"
   as permission to spawn a replacement

Required instead

1. One agent per ticket worktree until it completes or the operator says kill it.
2. If it looks stuck: wait, or ask the operator one line — do not spawn a twin.
3. To nudge a live agent: resume that same agent id only (no interrupt unless the operator explicitly says interrupt/kill/replace).
4. Before any new Task for T-XXX, check that no in-flight subagent for T-XXX exists in this conversation. If one does → do not dispatch.
5. Factory command center: gate/land/verify on completed commits only; never "help" a live slice agent by opening a second coding agent on its tree.

Operator override (only legal exception)

The operator must explicitly say to kill/replace that agent, e.g. "kill T-860 and start a new one" / "interrupt and replace". Silence, anger about delay, or your own impatience is not override.