diff --git a/.changes/disk-hygiene/docs-5609-bypass-probe-record-4a0eac9f.md b/.changes/disk-hygiene/docs-5609-bypass-probe-record-4a0eac9f.md new file mode 100644 index 0000000000..9f2796dbf8 --- /dev/null +++ b/.changes/disk-hygiene/docs-5609-bypass-probe-record-4a0eac9f.md @@ -0,0 +1,7 @@ +--- +bump: patch +--- + +### Changed + +- The safety model records the interactive bypassPermissions and auto-mode probes of the apply guard: its `ask` reaches a person or is denied in every probed mode, so no mode-independent gate is added (#5609). diff --git a/plugins/disk-hygiene/skills/clean/reference/safety-model.md b/plugins/disk-hygiene/skills/clean/reference/safety-model.md index 52c385b916..1d8f9ca1d0 100644 --- a/plugins/disk-hygiene/skills/clean/reference/safety-model.md +++ b/plugins/disk-hygiene/skills/clean/reference/safety-model.md @@ -705,14 +705,20 @@ semantics. The README states each surface in one table. The launch shape is asse `hooks/run-python-hook.test.sh` and `test_hygiene.py`, the no-interpreter posture by the former, and both are verified as step 1 of `/disk-hygiene:setup check`. -**Claim:** the guard's PreToolUse `ask` on `hygiene.py apply --execute` held in the three modes -probed. Headless default mode denied the call, `--bg` default mode parked at the permission prompt, -and headless `--permission-mode bypassPermissions` listed the call in `permission_denials`; in each -the target survived. Not probed: an interactive `bypassPermissions` session (where a person could -answer the prompt), auto mode against this guard, and the Windows PowerShell tool. **Basis:** the -probe table in and +**Claim:** the guard's PreToolUse `ask` on `hygiene.py apply --execute` held in every mode probed. +Headless default mode denied the call, `--bg` default mode parked at the permission prompt, headless +`--permission-mode bypassPermissions` and headless auto mode listed the call in +`permission_denials` with the target intact, and an interactive `bypassPermissions` session showed +the person a prompt and ran the delete only after they approved it. So the guard needs no gate that +does not depend on the permission mode: in each mode its `ask` reaches a person or is denied. Not +probed: the Windows PowerShell tool. **Basis:** the probe table in + and -(Claude Code 2.1.285, Linux/WSL2, Bash tool, fresh scratch target). The official docs do not name a +(Claude Code 2.1.285, Linux/WSL2, Bash tool, fresh scratch target); + (auto +mode, 2.1.307) and + +(interactive bypass, 2.1.308). The official docs do not name a hook `ask` under `bypassPermissions`: says "Allow rules have no effect in `bypassPermissions`", and its "Actions no mode auto-approves" list names "Tools matched by an explicit ask rule" and critical-path `rm` and `rmdir`, not a hook `ask`; @@ -722,9 +728,9 @@ statement about permission modes. Upstream resets bypass mode; closed as not planned 2026-04-20) and [#79356](https://github.com/anthropics/claude-code/issues/79356) (a hook `ask` and `permissions.ask` not enforced on the PowerShell tool in default mode; closed as not planned 2026-09-21) leave the -behavior undocumented. **As of:** 2026-09-30, both doc pages fetched that day. **Recheck:** a -Claude Code changelog entry that names PreToolUse `ask` or `bypassPermissions`, a change to the -"Actions no mode auto-approves" list, or the outcome of the unprobed-mode probes tracked in #5609. +behavior undocumented. **As of:** 2026-10-22 for the probes; 2026-09-30 for the two doc pages. +**Recheck:** a Claude Code changelog entry that names PreToolUse `ask` or `bypassPermissions`, a +change to the "Actions no mode auto-approves" list, or the PowerShell-tool probe tracked in #5609. A depth-limited scan records every directory it declined to enter in `truncated_paths`. Truncated directories have no captured descendant set, so the preview blocks them (and anything beneath them)