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.
The previous note said a fresh clone gets the skill back from `make
plugin` without naming the command, and I picked the wrong one when the
files actually went missing. There are two, and only one restores the
skill:
- `graphify install --platform claude` copies SKILL.md, references/
and .graphify_version. Touches nothing else. This is the recovery
command, verified: the skill came back at 0.9.61 and the four
guarded configs were byte-identical afterwards.
- `graphify claude install` writes the CLAUDE.md section and the
.claude/settings.json PreToolUse hooks, rewrites both of those
guarded configs (EVAL-020, reproduced today), and does NOT copy the
skill.
LRN-154 records why the files vanished in the first place. `git rm
--cached` keeps the working file, but `gitflow finish` checks out the
target branch, where it is still tracked, so git restores it and the
merge then deletes it from disk. .gitignore does not protect it; it only
stops a re-add. The file survives the commit and dies at the merge, which
reads as unrelated.
graphify: `graphify claude install` (install-plugins.sh STEP graphify)
writes SKILL.md, references/ and .graphify_version straight into the repo,
because ~/.claude/skills is a symlink to skills/. Every `pipx upgrade
graphifyy` therefore dirtied the tree and cost a `chore(graphify): sync
vendored skill X -> Y` commit. Now gitignored and untracked; a fresh clone
gets them back from `make plugin`. test-prompts.json is hand-written for
darwin and stays tracked. The accepted trade-off, documented in CLAUDE.md,
is that an upstream release can change the skill's prompt with no diff to
review.
settings.local.json (gitignored, so not in this commit) went from 14.6 KB
to 6.2 KB. It was a near-complete shadow copy of the global settings at a
higher precedence tier, which hid its own drift until the global moved.
Two entries were actively defeating BDR-090, merged an hour earlier:
- local `deny` still carried rsync / kill -9 / killall / pkill, the four
rules deliberately moved out of global deny. deny wins across sources,
so autoMode.soft_deny was a dead letter in this repo.
- local `allow` carried `sed *`, `cp *` and `python3 -`. An allow rule
short-circuits the classifier, punching a hole through the same
soft_deny rules.
deny and ask are dropped whole (102 and 27 of their entries duplicated the
global; ask gates nothing under defaultMode auto). allow went 185 -> 98:
81 duplicates plus six policy conflicts, the three above and
Read(//home/bchanot/**), WebSearch, and a leftover command-injection test
payload that had been allowlisted verbatim. Every non-permissions key was
a verbatim copy of the global, including a hooks block whose only original
entry pointed at hooks/config-protection.sh, a script that exists nowhere.
BDR-090 records why the ask tier was abandoned rather than repopulated,
the three alternatives rejected, and the deliberate caveat that the
guardrail hard_deny bars removing a deny entry but not adding one.
LRN-153 records the two traps the block carries: every autoMode list is
a full replacement without "$defaults", and a user-scope block reaches
every project on the machine.
TODO also logs F1-F3, found but not fixed: .claude/settings.local.json
is a 14.6 KB shadow copy of the global settings at higher precedence,
including a PreToolUse hook whose script does not exist.
`permissions.ask` gates nothing under `defaultMode: auto` (LRN-146,
verified live), so the ten rules that left the static tiers had no cover
left: rsync / kill -9 / killall / pkill out of deny, and python3 -c /
python -c / xargs / sed / cp / mv out of ask.
autoMode.soft_deny (7 rules) takes over what an explicit instruction
should be able to clear: writes outside the working directory,
rsync --delete, SIGKILL and kill-by-name, in-place edits spanning more
than one file, directory moves, and inline interpreters or xargs that
delete or write outside the cwd. Intent clears a soft block for the
current turn only, stated as a rule since no setting expresses it.
autoMode.hard_deny (3 rules) takes the classes no command pattern can
express: secret exfiltration, production deployment, and disarming the
guardrails. Adding a restriction stays allowed, removing one does not.
permissions.deny gains ten .env reader rules (sed awk cut tr sort uniq
diff od xxd strings). Six of those tools sat in permissions.allow, so
reading a .env through them triggered nothing.
autoMode.environment named another project, its FTP deploy target and its
customer data, inside the file link.sh:21 symlinks to
~/.claude/settings.json, where it reached every repo and contradicted
this one's Gitea remote. Rewritten machine-generic; the project facts
moved to that project's gitignored .claude/settings.local.json. All three
lists now open with "$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.
Extract gstack_bump_playwright_if_unsupported from install-plugins.sh into
lib/gstack-playwright.sh and call it from update-all.sh too. A submodule
update no longer leaves the OS-support bump unapplied until the next
`make plugin` — that was BDR-029's open caveat.
The update helper never touches the submodule working tree: on failure it
prints git's own message and points at `make plugin`, and returns non-zero
so the existing `else warn` arm still handles it.
Add a read-only `Playwright browsers` section to doctor.sh: cache size,
which registered install requires each revision, and counts of unreferenced
directories and broken links. No pruning is written — Playwright's own
`install` already unions the required set across every registered install,
and all three installs here are live (rev 1228 for gstack + gsd-pi 1.61,
rev 1243 for gsd-pi 1.63).
Carries two latent-bug fixes from the moved code: the ostag capture exited
1 on every non-Ubuntu host and aborted the caller under inherited errexit,
and the bun calls had no timeout.
Ending a turn right after spawning a subagent fired the bell and a toast
saying the response was finished, while the work continued. The Stop
payload carries background_tasks, so skip the signal when it is not
empty; the next turn end signals once the work is really done.
Interaction requests still signal during background work. A missing
field still signals, so an older client loses nothing. Also drop the
message suffix when it merely restates the label.
Only turn end and the moments needing the user should signal. Any other
event reaching the hook, such as agent_completed or auth_success, now
exits without emitting, so a subagent finishing rings nothing even if the
Notification matcher is ignored.
The toast body showed Claude's own message when present and the raw
notification_type otherwise, so permission_prompt and idle_prompt reached
the user as snake_case. Map every event the matcher covers to a readable
label, and keep Claude's message as a suffix when it adds detail.
Notification matcher covers input-needed events only; end of turn had no
signal but idle_prompt, ~60s late. Wire notify-attention.sh on Stop too,
branching on hook_event_name for the message. Signal only: returns
terminalSequence + suppressOutput, never blocks (guard vs BDR-083).
Header documents both client-side prerequisites found in BLK-020.
Notification hook (permission_prompt|idle_prompt|agent_needs_input|
elicitation_*) returns BEL x2 + OSC 777 via the terminalSequence JSON
field (hooks have no controlling TTY). Client side over Remote-SSH:
VS Code accessibility.signals.terminalBell sound:on for the beep,
wenbopan.vscode-terminal-osc-notifier extension for the Windows toast.
Two or more project paths dispatch one general-purpose runner per repo,
all in a single message, instead of processing repos one by one. The
runner inherits the session model — no pin, it carries tour's reflection
(fix decisions, convergence) — and every agent inside keeps its defined
tier (security-auditor sonnet, Phase B opus, doc-syncer sonnet). A dead
or mute runner becomes an explicit RUNNER FAILED summary row; the gated
capitalize offer stays in the main loop, never in a runner.
Bounded LRN-083 derogation recorded in BDR-084: the per-project fix loop
moves into its runner, but nothing a runner decides touches shared state
— independent repos, per-repo chore branches, branches left unmerged for
human review exactly as inline. Mechanics proven before building: nested
probe, 3 sub-agent windows all overlapping, 9.1s vs ~18s sequential.
Census §12: 6 locks, flip-tested. Single-project path unchanged.
The include is authoritative, but feat/bugfix/ship-feature/init-project
each restate the verify loop inline — an orchestrator following the
restatement alone would have skipped the floor. Each now carries the
GATE 0 bullet ahead of GATE 1 (4 new structure locks, flip-tested).
The contract-interview weight table stops promising a hotfix oracle
nothing executes: hotfix runs no floor, the hotfixer runs the suite
itself. CHANGELOG extended with the wiring + the RED result.
BDR-083 records what was taken from unlazy and, more usefully, what was
refused and why. LRN-141: an external skill's machinery encodes its threat
model, not yours — take the invariants, refuse the machinery. LRN-142:
structure locks are fixed-string, so reflowing a doctrine paragraph reds
them; fix the doc, not the lock.
GATE 1 is an LLM dispatch and the verifier's mandatory PROOF: line is a line
the verifier writes — nothing structurally stops it being produced without
anything being executed. Nothing deterministic sat between the executor and
that dispatch.
An acceptance criterion can now carry an oracle: indented CHECK: (command),
EXPECT: (success-only marker), EVIDENCE: (slot). lib/gates.sh runs them
fail-closed — MET requires exit 0 AND the marker, so a nonzero process never
passes on its error text carrying the token — and writes the outcome back
into the contract, so the fresh verifier reads evidence as fact rather than
trusting the executor's report.
GATE 0 runs that floor before any verifier is dispatched; a red build sends
the executor back for free, on its own iteration budget. ABANDON: <id>
<reason> turns an impossible criterion into a visible handoff that blocks
CONFORME and routes to the human gate, via the new ABANDONED(n) verdict —
a distinct token because it routes distinctly, never a dev loop. feater and
bugfixer gain a four-pass completion discipline, scoped so a pass can never
widen the contract.
The runner's parse fails closed on partial oracles, duplicate ids,
unindented attributes and runnable criteria with no EVIDENCE: line, and
executes nothing at all when the ledger is malformed. status never executes
and never writes; run always re-executes, since trusting written evidence is
the failure being closed.
Adapted from the unlazy skill (Leonxlnx/unlazy, MIT). Its Stop hook,
approval store, .unlazy/ tree, depth-tree arithmetic and Node checker were
deliberately refused — BDR-083 records each reason.
64 assertions in lib/tests/gates.test.sh, non-execution proved by sentinel
with its own positive control asserted first.
Bastien Chanot