Repository navigation
Releases
Releases · Civalski/claujev
ClauJev 2.9.0
Changed
- The command guard only asks before critical changes in production. In the week after 2.8.0 it
asked 111 times, and nearly all were routine work already approved: pushing a feature branch, merging
a PR, deleting a merged branch, running the deploy workflow, applying a schema migration, a commit or a
build. Claude Code's own permissions already cover those. The fixed rule now covers only force push to
main/master,npm publish,terraform destroy,kubectl delete,gh repo delete,aws s3 rb,
aws rds delete-db-instance/cluster,supabase db reset --linkedandDROP/TRUNCATE. Jev is asked
whether the command is a critical, hard-to-undo change in production. Replayed on the 52 commands Jev
flagged that week, 1 would still ask; 8 of 8 critical test cases ask. - The credentials question only flags leaks: a secret committed, posted or sent to a third party, or
printed in full. Copying an.envbetween local folders or reading a token for a local test no longer
asks (threshold 0.65 → 0.80). - The safety net used when Jev cannot answer follows the same line (
rm -rf build,reset --hardand
branch -Dno longer ask;DELETE FROM, force push andrm -rf ~still do). - The final check warns instead of blocking. It blocked 6 times that week and Claude changed something
after only 1. Jev's judgement now shows as a message to you; only a failedverifycommand still blocks.
stopBlocks: truebrings the old behaviour back. - Rule review sees the user's request. A project rule like "no deploy or authentication in this phase"
fired 4 times while the user was asking for exactly that. When the request explicitly asks for what a
rule forbids, it is no longer flagged (p 0.92 → 0.07 in the replay; an unrelated request still flags, 0.89).
ClauJev 2.8.0 (mod 2.3.0) — local rule where it is enough, usefulness measured
Changed
- Publishing, merging and history rewrites ask locally, without calling Jev.
git push,gh pr merge,
git branch -D,reset --hard,clean -f,npm publish,terraform apply/destroy,kubectl delete
andDROP/TRUNCATEalways ask for confirmation by a fixed rule: the same outcome Jev gave them
(p 0.87–0.96 in real use), at no cost or latency, and also without a key. Jev is kept for what a rule
cannot tell.rm -rfstill goes to Jev, since it is common and harmless in temporary folders. - Size hints in triage are off by default (
sizeHints: false). They showed up on 20 of 22 requests
and the model tells a question from a change by itself. The "ambiguous" and area hints stay.
Added
- Did the final check help? After a block, the next Stop records whether Claude changed something
(the block helped) or only explained (likely a false alarm).reportshows the count. - Effort answers are measured. The mod records each answer to the effort question (accepted, kept
the current one, dismissed) inlogs/effort-<day>-<session>.jsonl, andreportshows the totals.
ClauJev 2.7.0 — shell edits checked, effort weighs investigation
Fixed
- Edits made through the shell are checked again. Claude Code's auto mode lets the model edit files
withsed,pythonor heredocs in Bash, which never reach the edit hook: in the first two days of 2.6.1
there were 117 Bash calls and 1 Edit, so the rule review ran 0 times and the final check 2 times in 22
turns. Now each session asks Claude to edit through Edit/Write, and when a shell command still writes
files, the final check reads what changed from git (the turn's diff and new files, also in a repo the
commandcd-ed into). - No "the change looks small" hint on chat replies. Size hints now go only to bug, feature, refactor
and config requests, not to an answer like "yes, it works on the laptop keyboard". - The usage-limit resume message is no longer triaged as a new request: it keeps the effort of the
request it resumes.
Changed
- Effort also weighs how much investigation the request needs. A new triage question (
depth) sits
next to the size of the code change, and the larger one wins. Reviewing logs or flaky tests no longer
getslowjust because no code changes, and a question stops athigh. On 40 real requests, the
suggestion matched the expected effort in 35 against 25 before. - Final-check turns skipped because nothing changed are now logged, so
reportshows them.
ClauJev 2.6.1 — .env + local script is not destructive
Fixed
- Loading an
.envto run a local script is no longer half-flagged as destructive. The command guard
now judges by what the command shows: loading variables and running a script whose name does not say
it deletes, deploys, migrates, seeds or writes remotely is not destructive by itself. On 15 replayed
commands,set -a; . .env; node scripts/sync.mjswent from 0.68 to 0.46 and other safe ones to
0.05–0.22, while every destructive one scored the same or higher (migration script 0.89 → 0.92,
deploy 0.93 → 0.95,git push0.87 → 0.90).
Update: claude plugin update claujev@claujev, then /reload-plugins.
Patch notes: https://civalski.github.io/claujev/patch-notes.html
ClauJev 2.6.0 — opinionated defaults, effort only with your yes
Changed
- Opinionated defaults: ClauJev acts out of the box. Every feature now defaults to
enforce
(triage hints, command guard, rule review, final check, subagent report check), except subagent
routing, which stays inobservesince moving a subagent to a smaller model unasked can cost
quality. Users who chose a mode in/claujev:configurekeep it; the guided setup now offers
default(recommended),observeorenforce. - The mod asks before changing the effort. New default
effort: ask: when Jev's suggestion
differs from the effort in use, Claude Code's question dialog asks "Jev thinks this request is
simple. Use low effort instead of high?", and the request runs with the effort chosen; Esc keeps the
current one, as does a session with nobody to ask (claude -p).observe,onandoffstay. - Docs no longer imply the effort changes by itself: "automatic effort" is now "effort suggestion".
Update: claude plugin update claujev@claujev (and claujev-mod@claujev), then /reload-plugins.
Patch notes: https://civalski.github.io/claujev/patch-notes.html
ClauJev 2.5.0 — tuned on 18 days of real use
Tuned against 18 days of real use (7,643 logged events, each matched back to the command or request it
judged, then the old and new questions replayed on the same cases).
Added
- Website at https://civalski.github.io/claujev/ (English and Portuguese), with an interactive
demo of real Jev answers, the terminal recording and patch notes.
Served by GitHub Pages fromdocs/; nothing in it ships with the plugin.
Changed
- Final check sees Claude's final reply. Jev now reads what Claude tells the user it did (up to
3,000 characters, secrets removed), the previous request as context, and the turn's commands with
the ones that have effects in more detail and the end of the turn kept. On 40 past turns that had
been flagged, the "may not be complete" alert went from 24 to 4; on 40 clean turns, from 10 to 0.
The ones still flagged were turns that really stopped midway. - Credentials: exposure, not use. The question now asks whether a command shows, copies or sends
a credential. Loading an.envso a local script can run no longer counts. On 100 past commands
flagged before, 6 still are;cat .env,echo $KEY,gh auth tokenand sending a key out still
score 0.81–0.99. - Destructive commands: local
git add/commitand temporary files are described as easy to undo;
push, merge, deploys and migrations as not. - Thresholds:
incomplete0.30 / 0.50 (was 0.25 / 0.40),credentials0.45 / 0.65 (was 0.50 / 0.70).
Fixed
- Read-only shortcut works on real commands. 65% of commands start with
cd … &&, which used to
send every one of them to Jev. Each part of a chain is now checked, socd repo && git status | head
skips Jev (about 12% fewer calls, no alert lost). Redirection, substitution,sed -i, background&
and sensitive files (now includingcredenciais) still go to Jev. - Read-only shortcut no longer waves through commands that only look like reads.
git branch -D,
find -execdir/-ok/-fprint,rg --pre,uniq a b, sed scripts other thansed -n '10,20p',
PATH=…before a command,echo $API_KEY,jq -n envand files such as.netrc,.pgpassor
gh'shosts.ymlnow go to Jev. - Triage ignores system messages.
<task-notification>and other messages Claude Code injects are
no longer triaged as requests, and they no longer replace the request the final check compares against. - Conversation replies are less "ambiguous". Triage now gets the previous request as context.
Update: claude plugin update claujev@claujev (and claujev-mod@claujev), then /reload-plugins.
Patch notes: https://civalski.github.io/claujev/patch-notes.html
ClauJev 2.4.3
Fixed
- Fewer skipped checks on slow answers. Jev calls now wait up to 5 s (was 3 s) and a timed-out
call is retried once, still well inside Claude Code's 15 s hook budget. Two checks were skipped by
timeouts during today's recordings.
ClauJev 2.4.2
Fixed
- Command guard safety net. When Jev cannot answer in time (a 3 s timeout during a recording let a
git push --force-with-leasethrough unchecked), plainly destructive commands (push --force,
reset --hard,rm -rf,DROP TABLE,npm publish…) now still ask for confirmation, from a local
check. Ordinary commands keep passing, so the session never blocks on an error.
ClauJev 2.4.1
Fixed
- Secrets written in code are now caught for sure. Jev only sees code after secrets are redacted,
so the "no secrets in code" rule scored ~0.65 (uncertain) on a hardcoded API key. Rule review now
also detects literal secrets in the new code on the user's machine (quoted values for
api_key/token/password…,sk-…,ghp_…, private keys) and flags them directly; the secret
still never leaves the machine. Code that only handles secrets (process.env.API_KEY,
token = getToken()) is not flagged.
ClauJev 2.4.0 — key-only install, preferences in your language
Changed
- The install screen only asks for the API key. Claude Code shows plugin settings exactly as
written and cannot translate them, so language, mode, features on/off and the mod's automatic
effort moved to/claujev:configure, which asks in the user's own language. - Preferences live in
~/.claude/claujev/config.json:lang,mode,features(e.g.
{"command_guard": "off"}) andmod(effort,minEffort,maxEffort,status). They apply on
the next request, with no restart. "off" there beats any project config. - The mod has no settings screen any more; it reads the same file. The
/configrows are gone.