[Vasiliy-Bondarenko]

Why

In a few days of use, nearly every command Blast Radius held was Claude cleaning up after itself: rm -rf on its own scratchpad, dumps it had just downloaded into tmp/, install logs, caches. A pane for each one gets in the way and trains people to press Proceed without reading.

What

A new autoSkip option, off by default. With it off, the mod behaves as before; a test checks that classify() reads every test case the same as the committed version.

• rules: an rm runs without the pane when every path (after $VAR, ~ and symlink resolution) is inside this session's scratchpad, or doesn't exist. Every rm on the line must pass. Before the rm, the line may only run commands that read or write plain files. Lines with heredocs, $(...) or braces are left to the next step.
• rules+model: other rm commands are described to an OpenRouter Decisions model (~perplexity/pplx-decider-latest by default), with a short legend. The request carries facts, not the command text: size, up to five file names, git tracked/ignored, birth time against the session start, earlier mentions in the session and on the same line. If the model's rating reaches caution (low, medium or high: 0.8, 0.9 or 0.97), the command runs, and a toast says why. Otherwise the pane opens with the rating in its note. If the model fails or takes longer than 8 seconds, the pane opens.
• Some deletions never go to the model: top-level folders, files git tracks, .git, .env, secrets, backups, .sql dumps, migrations, more than 100 MB or 5,000 files, and paths with .. or braces that only the shell can resolve.
• rm inside ssh '...' or docker exec/kubectl exec goes through the same steps, but the rules never skip it.

Also:

• New holdMinutes option: how long a held command waits for an answer. Default 10.
• A refusal after an error while holding now names the error. In practice the error came from the mod being reloaded while it held a call.

Tests

• node --test tests/autoskip.test.mjs: 135 offline checks, all passing. They run in a sandbox file tree with tracked, ignored, old and new files and a symlink. The 63 cases are modelled on real sessions (paths made up), built to get past simple rules, or held out.
• DECIDER_LIVE=1 OPENROUTER_API_KEY=... node --test tests/decider.live.test.mjs: sends every case to the model once. It fails if any dangerous case would run without the pane at any caution level. A run costs well under a cent.
• The prompt went through six versions, each followed by a live run. Run-to-run noise was ±0.006 on average. The README has the details. With the final prompt, the model on its own rated no dangerous case above 0.35 on the known cases, or above 0.04 on the 15 held-out ones. At medium, it rated 18 of 21 known and 6 of 8 held-out temporary cases as disposable.
• claude plugin validate passes.

Not done: a recorded run in a live Claude Code session with auto-skip turned on, and new screenshots.

I have read the CLA Document and I hereby sign the CLA

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[Vasiliy-Bondarenko]

recheck

[Vasiliy-Bondarenko]

I have read the CLA Document and I hereby sign the CLA
recheck