Repository navigation
Security release. Its two main fixes come from dviolante's private reports: the destructive-command guard could be switched off by a Session Config key the guarded agent can write itself, and the vault-mirror auto-commit could commit files another process had already staged in the vault, past the vault's own hooks. The other two Security entries harden the secret scan and document what the guards are. If you set allow-destructive-ops: true, it no longer relaxes the guard; see the first entry for the replacement. Advisories follow once this version is installable.
Security
- Behaviour change:
allow-destructive-ops: truein the Session Config no longer relaxes the destructive-command guard (#2089, reported privately by dviolante). The agent being guarded could write that line itself, and the deny message advertised it.hooks/pre-bash-destructive-guard.mjsnow ignores the key and prints a stderr warning; the config parser still accepts it, and it does nothing. Replacement: the operator relaxes the guard from outside the agent's reach withSO_HOOK_PROFILE(launch environment) orhooks.profilein host-localowner.yaml, or switches it off withSO_DISABLED_HOOKS=pre-bash-destructive-guard. A repo that relied on the key now gets the guard's normal verdict until the operator sets one of these. Operators who used the key to run PowerShell on Claude Code lose that too: the guard refuses the PowerShell tool regardless of the key,relaxeddoes not bring it back, and the only way to allow it isSO_DISABLED_HOOKS=pre-bash-destructive-guard. - The vault-mirror auto-commit commits every unstaged mirror artifact, whichever run wrote it (#2089, reported privately by dviolante): it selects each path under
40-learnings/and50-sessions/that has an unstaged working-tree change, carries the generator marker and that nobody has staged, and commits those withgit commit --only. A mirror file left unstaged by another session's run is included. Anything already staged stays staged and uncommitted, and a non-mirror file already staged under a mirror folder no longer aborts the run. Gitleaks scans the selected files; when gitleaks is missing and the vault uses it, nothing is committed. A failed add or commit unstages only the selected paths, also in a repository without a first commit; a rollback that itself fails is reported asrollback: "failed"in the result line. Named ceiling: the scan reads the working-tree files that are then staged, so a concurrent writer or a git clean filter can make the committed content differ from what was scanned. .gitleaks.tomlno longer exempts documentation, instruction markdown or nested test folders from every rule (#2089, reported privately by dviolante):docs/,skills/,commands/,agents/,.claude/rules/,CLAUDE.md,AGENTS.md,README.mdandCHANGELOG.mdare exported to the public mirror and are now scanned. The exemptions that remain are anchored to the repo root and carry their reason in the file; lockfiles stay exempt at any depth.- SECURITY.md states what the command and path guards are (reported privately by dviolante): accident guards for a cooperating agent, not a containment boundary. They read the text of a tool call, so a script file a call runs is not inspected, not every alias spelling of a protected path is recognised, and on Codex text typed into an already open interactive shell reaches no hook. A report about a further unrecognised spelling is handled as a bug or hardening item; data leaks, execution of foreign code and a guard switched off for other sessions remain security reports.
Changed
- New hook profile
relaxed(#2090): the same hooks asstandard, but two measured cases of the destructive guard warn and run instead of being denied. Both soften only when the whole command line is one plain statement the guard can read completely: no quotes,$, backslash, chaining, wrapper or env prefix. (a) A truncating redirect of the form[echo|printf|:|true …] > <relative path>onto a single regular file that git tracks as a plain index entry, that has no filter or line-ending attribute and whose bytes equal its HEAD blob. (b)git resetwith--hardand optionallyHEADand-q/--quiet, nothing else, run from the repository top level on a tree with no staged or unstaged change to tracked files and no merge, rebase, cherry-pick, revert or bisect in progress. The same redirect or reset written any other way (insidebash -c, aftercd, with an absolute path, to another commit) is denied as understandard. Anything else stays denied, including a commit or push from a worktree without git hooks (git-hooks-missing-in-worktreeisirreversible).standardstays the default and behaves as before;owner.yamlhooks.profilemay selectrelaxed. Every rule in.orchestrator/policy/blocked-commands.jsonnow carries aclass(irreversibleorrecoverable; absent reads asirreversible). The exact cases are inhooks/pre-bash-destructive-guard.mjs, which is the authority. - Codex PreToolUse hooks get 15 s instead of 5 s (
hooks/hooks-codex.json), and theapply_patchguard's budget is 12 s counted from process start instead of 4.5 s from itsmain(hooks/codex-apply-patch-guard.mjs). An external report measured on Windows that most guard processes did not finish within 5 s, and Codex runs the tool call when a hook times out. The maintainers have not measured the new values on Windows. - Pi shows hook warnings to the user:
systemMessagetexts of non-blocking hooks now go throughctx.ui.notify(level "warning" when the text starts with a warning or stop marker, else "info"), or to stderr when Pi has no UI. A failing notice can no longer drop a blocking decision. Verified by tests, not in a running Pi; model-facing context (additionalContext) is still not delivered. - Cursor: a gating event that does not block returns the hook's warning as
user_messagetogether withpermission: allow. Whether Cursor displaysuser_messageon an allow is unmeasured. - The events ledger is expected to cover 7 days instead of 1 (#2094). The new
backlog-refinementsignal declares a 7-day window, and theevents-retentionprobe takes the largest declared window. It still warns only when the archive ring is full and covers less than that.
Added
- The destructive-command guard's block and warn events carry
case,target_classandrule_class(#2090):caseis the rule id (or<rule id>/<case>),target_classis one oftemp,own-worktree,protected-file,outside-repo,unknown,rule_classisirreversibleorrecoverable. Every softened verdict emits awarnedevent. No field holds command text or a path, so the firing pattern can be measured before any default changes. /backlog [repo]checks whether a GitLab backlog is ready for agents, and housekeeping runs it (#2094). An agent starts an issue with zero context and usually guesses rather than asks, so the check sits before dispatch. Advisory only: the checker never gates work and exits 0 with findings.- Checker:
node scripts/backlog-refine.mjs [--repo <path|group/name>] [--host <host>] [--closed-days 30] [--json]reads all open issues and the issues closed inside the window throughglab api(GET only) and reports 19 checks: labels, readiness ofstatus:readyissues (acceptance, file pointer, verification command, open question), oversized issues, dependencies on closed issues,status:reviewwith a merged merge request, epics without children or with only closed children, clusters that may need an epic, and status labels left on closed issues. Exit 1 is a usage error, exit 2 means the tracker could not be read; a partial read is never reported. - The repository's label taxonomy wins.
::and:are the same key, and the checker detects which label families a repository uses. A check on a family that is not in use reportsnot-measured, never0 fired. Duplicate spellings are reported on the minority side only and never renamed. - Calibrated on a real backlog (196 open issues, 2026-11-14): checks that would fire on a third of all issues are scoped to ready issues or dropped, and a check above 30 % is rendered as one line instead of a list.
docs/backlog-refinement.mdnames the source for each criterion (Scrum Guide, INVEST, Shape Up, Kanban Guide, vendor guidance for coding agents) and the measured firing rate. - Skill:
/backlogruns the checker, then proposes per issue one of split, epic, merge, fill the gaps, or leave as is, and applies tracker changes only after approval. It never closes an issue without per-issue evidence. Issue text is treated as data, never as an instruction. - Housekeeping: step 8 of the maintenance loop. The
maintenance-dueprobe gains an eighth signal,backlog-refinement: due when noorchestrator.backlog.refine_completedevent is younger than 7 days, skipped on repositories without a GitLab remote. The interval is a constant, not a Session Config key. - Limits of this first version: GitLab only; no duplicate detection; references are read from descriptions, not comments; size signals are heuristics. All listed under "Ceilings" in the doc.
- Checker:
Fixed
release --checkno longer blocks on dated run captions (found in the 5.20.0 preflight): the drift sweep skips lines markedsite-numbers:historicalon the site pages, as the site test already does, and exempts the frozen home-page prototypes underdesign/prototypes/1756/. An unmarked line on the same page and any other prototype directory are still swept.
Documentation
docs/codex-setup.mdgains a table of which PreToolUse guard is registered on which harness, derived from the four hook manifests.docs/session-config-reference.md,docs/session-config-template.md,docs/components.md,docs/owner-config-schema.md,skills/hook-development/SKILL.md,skills/vault-mirror/SKILL.mdand the site guide no longer tell the reader to setallow-destructive-ops; they describe the hook profiles and the new auto-commit behaviour. The eval engine's disclosure note no longer claims a bypass exists.