Semantic blast radius is the practical answer to a common AI coding question: did the agent change more than it was supposed to? A file diff can show that five files changed. It cannot tell you whether those files alter billing behavior, auth assumptions, generated docs, or release posture.
flowchart TD
A[Agent implementation] --> B[nool debug blast-radius]
B --> C{Risk threshold exceeded?}
C -- No --> D[nool propose]
C -- Yes --> E[Split task or justify]
E --> F[nool task block or revised intent]
F --> D
D --> G[nool solidify]
The verified command is simple: `nool debug blast-radius <targets> --compact`. Targets can be file paths or knot IDs. Use `--symbol <SYMBOL>` when you need a symbol-level view, and `--threshold <THRESHOLD>` when you want warnings tuned to the risk profile of the change.
Blast radius should run after implementation but before proposal acceptance. It is especially useful when an agent touches shared models, auth, billing, migrations, generated code, CI, or routing. If the risk is high, that does not automatically mean reject the change. It means the proposal needs a real justification, stronger tests, or a narrower task split.
For broader context, pair it with `nool query blast-radius <knot>` when analyzing causal descendants of already-recorded work, and `nool discover similar <topic>` when checking whether related work has happened before.
The goal is not to make agents timid. The goal is to make broad impact explicit. A wide blast radius is sometimes correct. A hidden blast radius is where expensive bugs are born.
A good convention is to make high-risk output change the process. If blast radius crosses a module boundary, add a finding or a proposal justification. If it crosses a security or billing boundary, require a human review before `nool solidify`. If it crosses many unrelated boundaries, split the task. The command provides the signal; the team policy decides the response.
That policy should be written down as findings and thread decisions, not remembered informally. Otherwise every new agent relearns the same risk threshold from scratch.
Repeated rediscovery is preventable waste.