Layer C of the plan written after the 2026-09-21 wipe (BDR-095): a reviewer
sub-agent traced `lftp mirror --delete` against a local file:// tree, the
prose tiers named neither lftp nor a local trace, the brief had authorized
it, and four days of commits had never left the machine.
- gitflow: `start` pushes the branch with its upstream, merge targets are
pushed after each merge, and `init`/`install-hook` write post-commit and
post-merge hooks that push every commit as it lands (warn, never block;
GITFLOW_NO_PUSH=1 for throwaway repos). T18 + T19 (installed == emitted).
- hooks/unpushed-guard.sh on SessionStart and Stop: branch ahead of its
upstream, no upstream, or no origin. Non-blocking systemMessage.
- settings.json: static deny for transfer and mirror tools, rsync --delete,
xargs rm, pipe-to-shell, chmod/chown -R, sudo/doas/pkexec, disk tools,
chattr, docker volume drops/prune/--privileged/socket/-v /:, git history
destruction, --no-verify and core.hooksPath; new hard_deny "destructive
tool against a local path, brief carries no user authority"; soft_deny
reworded + discarding uncommitted work; environment records the incident.
- CLAUDE.global.md "Destructive tools & data loss"; the four report-only
agents trace by reading, never by running, whatever the brief says.
- lib/tests/guard-bash.test.sh: executable spec of the PreToolUse guard
(214 cases). The hook itself is not shipped (BLK-022); the spec skips.
Every checklist command is emitted on exactly one line, however long; a
legacy backslash continuation in the runbook is joined at instantiation,
and bootstrap / learn patches write runbook lines the same way. After the
checklist the hand-back carries a Post-deploy tests block derived from the
delta diff: by-hand checks tied to delta files plus Suggestions for gaps.
Cold-resume re-display and re-hand-back regenerate both.
RED/GREEN on a scratch runbook: 4/4 baseline runs reproduced the
continuation verbatim and printed no tests; 4/4 runs on the edited skill
joined it and printed the block in the recipe's shape.
`permissions.ask` gates nothing under `defaultMode: auto` (LRN-146,
verified live), so the ten rules that left the static tiers had no cover
left: rsync / kill -9 / killall / pkill out of deny, and python3 -c /
python -c / xargs / sed / cp / mv out of ask.
autoMode.soft_deny (7 rules) takes over what an explicit instruction
should be able to clear: writes outside the working directory,
rsync --delete, SIGKILL and kill-by-name, in-place edits spanning more
than one file, directory moves, and inline interpreters or xargs that
delete or write outside the cwd. Intent clears a soft block for the
current turn only, stated as a rule since no setting expresses it.
autoMode.hard_deny (3 rules) takes the classes no command pattern can
express: secret exfiltration, production deployment, and disarming the
guardrails. Adding a restriction stays allowed, removing one does not.
permissions.deny gains ten .env reader rules (sed awk cut tr sort uniq
diff od xxd strings). Six of those tools sat in permissions.allow, so
reading a .env through them triggered nothing.
autoMode.environment named another project, its FTP deploy target and its
customer data, inside the file link.sh:21 symlinks to
~/.claude/settings.json, where it reached every repo and contradicted
this one's Gitea remote. Rewritten machine-generic; the project facts
moved to that project's gitignored .claude/settings.local.json. All three
lists now open with "$defaults", which the original omitted, so the
built-in classifier entries are inherited rather than replaced.
doctor.sh check_automode backstops both defects. SETTINGS.md documents
the block and a tier-choice table. README no longer claims the ask tier
makes every mcp__magic__* call require a live confirmation.
Startup emitted 15 warnings: "Write(**/.env) is not matched by file
permission checks — only Edit(path) rules are."
Write(path) rules never matched. The 5 secret-file write bans were dead
config — .env, secrets/**, *.pem, *.key were freely writable. Converting
to Edit() makes them enforced: permissions.md:242 "Edit rules apply to all
built-in tools that edit files", and :244 prescribes exactly this ("add an
Edit deny rule for paths no tool may change").
- settings.json: Write(...) -> Edit(...) on the 5 patterns.
- Mirror the 9 secret patterns Read denied but Edit did not: *.p12, *.pfx,
id_rsa*, id_ed25519*, .ssh/**, credentials, credentials.json,
.aws/credentials, .azure/**. Read/Edit parity now 14/14. Claude could
previously overwrite an SSH private key or ~/.aws/credentials.
- New read-allowed/write-denied class: lockfiles (*.lock,
package-lock.json, pnpm-lock.yaml, go.sum) + node_modules/**. Reading
aids diagnosis; hand-editing is always wrong — the package manager
regenerates them via Bash, which Edit deny does not block.
- templates/settings/SETTINGS.md taught the broken Write() pattern; fixed
at the source so /onboard stops propagating it.
Rule syntax has no negation and deny beats allow, so deny globs cannot
carry exceptions — see the .env.example conflict noted in the follow-up.
Live failure (run 2): the checklist printed above AskUserQuestion never
reached the user. Fix is structural: the checklist is never written to a
file (throwaway — PENDING.json + live runbook regenerate it in any
session) and every hand-back/re-display ends the turn with the full
checklist as the FINAL text, no tool call after it. Cold resume without a
report regenerates + re-displays. Artifacts 5 -> 4 files; bootstrap
gitignore step drops NEXT.sh; mistakes/red-flags updated (no tool call
after the print, no file 'for reference').
First-real-run UX feedback (EVAL-016): one command per line as typed in an
interactive session (ssh opens the box, following lines run on it, local
steps flagged), never folded ssh compounds; the hand-back prints the full
checklist in the conversation (and every re-hand-back reprints it). Step
defined as a block (header + command lines to next blank line), @delta
governs the block. Template restyled to match.
Move tasks/ into .claude/tasks/ and create .claude/memory/ with 5
persistent registries (decisions/learnings/blockers/journal/evals) plus
.claude/audits/ for audit reports. Also seed templates/memory/ for
onboarder to copy into user projects.
- tasks/TODO.md, tasks/rtk-upstream-issue.md → .claude/tasks/
- tasks/LESSONS.md deleted (empty template; superseded by learnings.md)
- .claude/memory/: 5 append-only registries with YAML schema + seed entries
- templates/memory/: 5 empty-index templates for new projects
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
bastien