Summary
Several disk-hygiene docs are inaccurate about platform capability and the guard's env mechanism.
Grouping the doc corrections so the set stays internally consistent (some should land together with
the #376 fix).
D1 — setup mislabels Windows execution posture as "full"
skills/setup/SKILL.md:39-42 frames Windows as execution-capable ("Windows (full, lstat
reparse + Win32, never UAC)") and separates macOS as "audit/report only by design."
- But
hygiene.py:1013-1016 (execution_blockers) returns execution-platform-unsupported for
every non-Linux OS, and SKILL.md:151 confirms only Linux executes.
A Windows user is told they have "full" capability when apply always returns
execution-platform-unsupported. Windows should be described as audit-only for execution, like macOS.
D2 — multiple docs restate the false "hook process receives CLAUDE_PLUGIN_DATA" premise
Beyond the already-known CHANGELOG 0.3.0 line, the same incorrect claim appears at:
SKILL.md:59-61 ("The guard validates --data-root against the value its own hook process
received … even when the shell environment lacks CLAUDE_PLUGIN_DATA")
safety-model.md:55-59
These are false for skill-frontmatter hooks (the #376 mechanism) and should be corrected together
with the CHANGELOG and the code fix so the security narrative is consistent.
D4 — PowerShell-lane docs omit two real gaps
safety-model.md:63-66 and SKILL.md:222-225 describe the PowerShell lane as gating "known
deletion spellings" without noting that (a) the kill switch does not close it (see the kill-switch
issue) and (b) move/overwrite spellings are not covered (see the mutation-spelling issue).
Severity
MEDIUM (correctness of the platform + security narrative). D2 should ship with the #376 fix.
Work-class: C2 (mechanical) — attended triage 2026-07-23, operator-ratified. 🤖
Summary
Several disk-hygiene docs are inaccurate about platform capability and the guard's env mechanism.
Grouping the doc corrections so the set stays internally consistent (some should land together with
the #376 fix).
D1 —
setupmislabels Windows execution posture as "full"skills/setup/SKILL.md:39-42frames Windows as execution-capable ("Windows (full,lstatreparse + Win32, never UAC)") and separates macOS as "audit/report only by design."
hygiene.py:1013-1016(execution_blockers) returnsexecution-platform-unsupportedforevery non-Linux OS, and
SKILL.md:151confirms only Linux executes.A Windows user is told they have "full" capability when
applyalways returnsexecution-platform-unsupported. Windows should be described as audit-only for execution, like macOS.D2 — multiple docs restate the false "hook process receives CLAUDE_PLUGIN_DATA" premise
Beyond the already-known CHANGELOG 0.3.0 line, the same incorrect claim appears at:
SKILL.md:59-61("The guard validates--data-rootagainst the value its own hook processreceived … even when the shell environment lacks
CLAUDE_PLUGIN_DATA")safety-model.md:55-59These are false for skill-frontmatter hooks (the #376 mechanism) and should be corrected together
with the CHANGELOG and the code fix so the security narrative is consistent.
D4 — PowerShell-lane docs omit two real gaps
safety-model.md:63-66andSKILL.md:222-225describe the PowerShell lane as gating "knowndeletion spellings" without noting that (a) the kill switch does not close it (see the kill-switch
issue) and (b) move/overwrite spellings are not covered (see the mutation-spelling issue).
Severity
MEDIUM (correctness of the platform + security narrative). D2 should ship with the #376 fix.
Work-class: C2 (mechanical) — attended triage 2026-07-23, operator-ratified. 🤖