You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Should the unproven "redirects under the work directory are blocked" reason be verified or deleted? #15666
Should the dotfiles reason "guardrails blocks redirects under the work directory" be verified and kept, or deleted?
From the talk
Chapter 2, slide 11: prose is the judgment layer; a stated reason must be true.
Today and the gap
The user instructions say a file under a repository's .work/ must be written with the Write tool because guardrails blocks shell redirects there. A search of plugins/guardrails/hooks for .work found only delete and git guards. UNVERIFIED.
Recommendation
Verify with a live probe; delete the reason if no guard exists, or point it at the guard that does (menu priority P1). Basis: the search above. The issue-filing wording and the worktree rule in the same pair were already decided in the interview.
Alternatives
Add the guard the rule describes; keep the rule without the reason.
Done would add
Decided here, built in the dotfiles repository: the line corrected or removed.
Understood: the dotfiles instructions say shell redirects under .work/ are blocked by guardrails, but a search of plugins/guardrails/hooks found only delete and git guards. The reason is unverified. The fix would be a one-line edit in the dotfiles repository, or a new guard here if you choose that.
The issue lists alternatives (add the guard, or keep the rule without the reason) next to its recommendation, so the outcome needs an owner pick. No likely duplicate or merged fix found.
Should the reason be verified by a live probe, then deleted if no guard exists or repointed at the guard that does?
RECOMMENDED: yes. A stated reason must be true, and the search found no matching guard. Adding a new guard is a separate feature.
If no guard exists, keep the Write-tool rule without a reason, or drop the rule entirely?
RECOMMENDED: keep the rule without the reason. It is harmless, and removing it is a separate decision.
To answer, post a comment that starts with this line:
Resolved (owner, 2026-11-13: "go w/ recommendations"): delete the line, rule and reason both ("Write a file under a repo's .work/ with the Write tool, not a shell redirect: guardrails blocks redirects there so the secret scan sees the content"), and fix the two stale comments in plugins/guardrails/hooks/secret-pattern-detection.sh (:336,403) that describe exemptions "as block-hook-bypass exempts", a guard that no longer exists.
The reason was verified, and it is false today. The issue's search found no guard because the guard was removed, not because it never existed:
block-hook-bypass blocked shell writes so the Write-tool checks would run. Guardrails 0.52.0 removed it on 2026-10-16, the day the line was written: "A file a shell command writes is checked when it is committed, by the pre-commit-content git hook" (plugins/guardrails/CHANGELOG.md:1077-1079).
.work/ is ignored by git, so a file a redirect writes there is never committed and never scanned. A Write-tool write is still scanned (plugins/guardrails/README.md:86).
No exposed surface: a secret in .work/ leaves the machine only by a commit (.work/ is ignored) or a GitHub post, which guardrails scans. It would matter if a .work/ file were copied into a tracked path in a repository without the commit-time scan.
Agents already ignore the line: about 950 shell redirects into .work/ in local transcripts (a regex upper bound).
Basis in the talk: slide 20, the cartoon where the robot says "I made it up" and is asked "Source?"; slide 11, transcript 1:04:06: prose is the judgment layer.
Not chosen: keeping the rule with a corrected reason (a habit that protects no exposed surface); re-adding a narrow guard (guardrails removed the broad one on precision grounds, CHANGELOG.md:1062-1066).
Not verified: no live probe was run; the reading is static, over the guardrails hooks, the user hooks and settings.
Build: a one-line dotfiles edit riding with dotfiles #1556, and one comment-only guardrails item here, filed when the map closes.
Activity
github-actions commentedon Nov 12, 2026
Size: Small
Complexity: Low
Uncertainty: Knowable
Flags: intent
Blocked by: none
Understood: the dotfiles instructions say shell redirects under
.work/are blocked by guardrails, but a search ofplugins/guardrails/hooksfound only delete and git guards. The reason is unverified. The fix would be a one-line edit in the dotfiles repository, or a new guard here if you choose that.The issue lists alternatives (add the guard, or keep the rule without the reason) next to its recommendation, so the outcome needs an owner pick. No likely duplicate or merged fix found.
Should the reason be verified by a live probe, then deleted if no guard exists or repointed at the guard that does?
RECOMMENDED: yes. A stated reason must be true, and the search found no matching guard. Adding a new guard is a separate feature.
If no guard exists, keep the Write-tool rule without a reason, or drop the rule entirely?
RECOMMENDED: keep the rule without the reason. It is harmless, and removing it is a separate decision.
To answer, post a comment that starts with this line:
Automated triage. A maintainer reviews it.
melodic-automation-lanes commentedon Nov 12, 2026
The triage verdict left this issue at
needs-owner-decision:Reason:
intentkyle-sexton commentedon Nov 13, 2026
Resolved (owner, 2026-11-13: "go w/ recommendations"): delete the line, rule and reason both ("Write a file under a repo's
.work/with the Write tool, not a shell redirect: guardrails blocks redirects there so the secret scan sees the content"), and fix the two stale comments inplugins/guardrails/hooks/secret-pattern-detection.sh(:336,403) that describe exemptions "as block-hook-bypass exempts", a guard that no longer exists.The reason was verified, and it is false today. The issue's search found no guard because the guard was removed, not because it never existed:
block-hook-bypassblocked shell writes so the Write-tool checks would run. Guardrails 0.52.0 removed it on 2026-10-16, the day the line was written: "A file a shell command writes is checked when it is committed, by thepre-commit-contentgit hook" (plugins/guardrails/CHANGELOG.md:1077-1079)..work/is ignored by git, so a file a redirect writes there is never committed and never scanned. A Write-tool write is still scanned (plugins/guardrails/README.md:86)..work/leaves the machine only by a commit (.work/is ignored) or a GitHub post, which guardrails scans. It would matter if a.work/file were copied into a tracked path in a repository without the commit-time scan..work/in local transcripts (a regex upper bound).Basis in the talk: slide 20, the cartoon where the robot says "I made it up" and is asked "Source?"; slide 11, transcript 1:04:06: prose is the judgment layer.
Not chosen: keeping the rule with a corrected reason (a habit that protects no exposed surface); re-adding a narrow guard (guardrails removed the broad one on precision grounds,
CHANGELOG.md:1062-1066).Not verified: no live probe was run; the reading is static, over the guardrails hooks, the user hooks and settings.
Build: a one-line dotfiles edit riding with dotfiles #1556, and one comment-only guardrails item here, filed when the map closes.
kyle-sexton commentedon Nov 13, 2026
Done (2026-11-13): Resolved; see the resolution comment.