Skip to content

disk-hygiene: apply/handoff ask gate and unattended-stop are advisory, not deterministic — apply can run with no human #381

Closed as not planned

Description

@kyle-sexton

Summary

disk-hygiene's manual-handoff and apply safety bar depends on a PreToolUse ask decision plus a
model-advisory instruction — neither is a deterministic human gate. In auto-approve / background /
--dangerously-skip-permissions sessions, both can be satisfied with no human in the loop.

Details

  • skills/clean/scripts/destructive_guard.py:314-323 returns permissionDecision: "ask" for engine
    apply; the PowerShell deletion lane likewise returns ask. An ask can be auto-resolved when
    the session is non-interactive or permissions are pre-granted.
  • skills/clean/SKILL.md:46 — "Automated, scheduled, remote, unattended … always audit and stop" —
    is enforced only by model compliance. Nothing in the guard or engine detects session
    attendedness.

On Linux (the one platform where apply actually deletes), a plan can therefore be applied with no
human confirmation. This session (a background job) is a concrete example where deletion spellings
executed without a human prompt materializing.

Suggested fix

  • Make the deterministic gate independent of ask semantics: e.g. the engine apply requires an
    explicit human-confirmation token that cannot be auto-minted by the model, issued only through an
    interactive prompt.
  • Give the guard a session-attendedness signal (if Claude Code exposes one) so unattended sessions
    are deterministically forced to audit-and-stop rather than relying on the model to comply.
  • Add a regression test for the unattended/auto-resolve path.

Severity

MEDIUM-HIGH (security).

Activity

  1. added
    needs-humanHuman-in-the-loop required; autonomous sessions must not resolve items carrying this.
    on Jul 23, 2026
  2. kyle-sexton commented on Jul 27, 2026

    @kyle-sexton
    MemberAuthor

    This was generated by AI (triage lane — gate-audit claim). Claiming this item for one gate-audit cycle: an independent panel re-classifies whether the needs-human / status: needs-decision gate is still warranted, or whether the item is C2/C3 and liftable to agent-ready. Outcome comment follows in this cycle.

  3. kyle-sexton commented on Jul 27, 2026

    @kyle-sexton
    MemberAuthor

    This was generated by AI (triage lane — gate audit). Independent panel: two fresh-context agents (one fable, one opus), each given only the issue content and the repo's own plugins/autonomy/reference/guardrails/work-classes.md, plus the triage lane's own read. This lane never executes work items, authors fixes, or merges.

    Gate audit: GATE UPHELD — C4, security surface

    Panel: 2-of-2 STAYS-GATED, both C4. Lane read concurs.

    This asks to replace an advisory ask / model-compliance gate with a new deterministic human-confirmation contract inside a PreToolUse destructive guard (destructive_guard.py still emits "ask"). That is a change to the plugin's safety boundary, not a briefed fix, and a mistake silently re-opens unattended deletion — the exact failure the issue exists to close.

    It also depends on a capability nobody has verified: the issue's own second bullet hedges "if Claude Code exposes one" for a session-attendedness signal. Designing a safety gate around an unverified harness capability is the cross-cutting-invention floor.

    Decision required

    What deterministic mechanism gates unattended apply?

    RECOMMENDED — fail closed, and do not depend on attendedness detection. Require an explicit human-issued confirmation token that the model cannot mint; when it is absent, apply hard-stops to audit-only. This works regardless of whether Claude Code ever exposes an attendedness signal, so the design does not rest on an unverified upstream capability. Accept the consequence: apply becomes unavailable to unattended automation, which is the correct direction for a destructive operation.

    Alternative — verify first, then branch on attendedness. Per the research-first posture, establish whether Claude Code exposes any session-attendedness/interactivity signal to hooks, then design around the verified answer, keeping ask for attended sessions. Better ergonomics if the signal exists; wasted design cycle if it does not.

    Either path stays human-reviewed and human-merged — this edits the destructive guard.

  4. kyle-sexton commented on Jul 27, 2026

    @kyle-sexton
    MemberAuthor

    This was generated by AI (triage lane — claim release). Gate-audit cycle complete for this item; the triage claim posted earlier this cycle is released. No lease is held — another instance or lane may pick this up immediately. The outcome comment above this one carries the classification and either the lift rationale or the decision summary.

  5. kyle-sexton commented on Aug 13, 2026

    @kyle-sexton
    MemberAuthor

    Frontier disposition (drain session)

    Ruling: Close as not planned for this drain wave (frontier consensus).

    Dual-panel drain ruling: advisory, superseded, folded into a program umbrella, or blocked on a host/design carrier without an actionable in-repo slice this session. Re-open with a concrete acceptance brief if a consumer needs the work.

    KEEP set for this wave: standing telemetry; host-physical Class C (#631 family); program umbrellas #530/#531/#532/#535; live design carriers #1500/#1553/#1668/#1695/#1716/#1681/#1718; #1360.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area: securitySecurity-relevant: vulnerability, hardening, or disclosure follow-up.needs-humanHuman-in-the-loop required; autonomous sessions must not resolve items carrying this.

    Type

    Fields

    No fields configured for Task.

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions