| Age | Commit message (Collapse) | Author |
|
Squash merge of 1 commit
|
|
security-guidance: bump version to 2.0.21
|
|
security-guidance: keep the interpreter check working through a .bat shim
|
|
security-guidance: skip other repositories whose settings change what git runs
|
|
Squash merge of 3 commits
|
|
Squash merge of 1 commit
|
|
Squash merge of 1 commit
|
|
The hooks run git in the session's repository and in those the session
moves to. A repository other than the session's own is now skipped when
its own config has settings that change what git runs. An index copy made
in another repository holds only the entries for files.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
pyenv-win's python names end in a batch file. In its released version
(3.1.1) cmd.exe read the ">=" in the check as a redirect: every hook left
an empty "(3" file in the project, and the interpreter was rejected. With
no other Python no hook ran.
The check now uses only characters that arrive unchanged. On Windows a
name that loses the argument counts as unable to read the hook script, so
the launcher goes on to its real interpreter.
See #7172.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
Windows can release a killed process's file lock a moment after the
process ended, so the wait that follows may see it held once more.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
Co-authored-by: mhegazy <mhegazy@anthropic.com>
# Conflicts:
# plugins/security-guidance/hooks/security_reminder_hook.py
|
|
security-guidance: keep a review's result and time on full metric rows
|
|
security-guidance: review text files renamed to code names
|
|
security-guidance: tell the commit reviewer its text is not read
|
|
security-guidance: review a commit-and-push again when the commit review gave no verdict
|
|
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
A python.org Python on macOS has no CA certificates until its "Install
Certificates" command is run, so every review from it failed to verify
the API's certificate. Where Python's default TLS context holds no
certificate at all, the plugin now adds /etc/ssl/cert.pem, the file
Apple's own Python uses. Verification is unchanged.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
The commit and push reviews captured, decoded and split the whole
`git show -p` / `git diff` before looking at the number of files. An
import of 5,800 files and 4.6 M lines took 6.4 GB, and the review was
then skipped for its file count.
Read git's output as it arrives and keep what the review uses: the
characters of each file the prompt takes, a count of the added lines
after them, and past the file cut-off only the file names.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
With findings, the critical and high counts could be cut from a full row. They now come with the result. commit_review_on and push_sweep_on, which are always true on a reviewed row, go last instead.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
security-guidance: fewer false edit warnings on ordinary JavaScript
|
|
security-guidance: edit hook no longer waits on App Installer's python alias
|
|
security-guidance: match git commit and push continued on the next line
|
|
security-guidance: custom rule globs can use braces
|
|
A text file whose name or folder says it holds secrets (api_keys.txt,
tokens.txt, secrets/prod.md) was read whole when it was renamed to a code
name. It now stays out of the review, and so does a file that recent commits
renamed from such a name.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
report a failure the push review did not
A message-only amend made while the first review was running counted as
reviewed at once. When that review then failed, a later push reviewed
nothing. The amend now waits for that review.
A failed commit review stayed quiet for the push review of its own command,
which can itself end quietly (a cap, no network, killed). It now waits for
the push review, and reports unless that one got a verdict or told of its
own failure.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
ANTHROPIC_DEFAULT_OPUS_MODEL is set
The reviewer then runs on the model that variable names, as it does with
SG_AGENTIC_MODEL, so the rule is left out there too.
The tests clear the variables the rule depends on in their shared fixture,
check that the stages ran, and no longer look models up.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
Squash merge of 1 commit
|
|
Four edit rules matched their word without looking at what stands next to it:
- eval: Playwright's and Puppeteer's page.$eval('selector', function) and
$$eval got the eval() warning. They run the function they are given. Left
out only when the selector is a plain string and the function is written
out in the call. AngularJS scope.$eval(expr) and every other $eval( still
warn.
- new Function: the whole word is needed, so new FunctionRegistry() is quiet.
- innerHTML and outerHTML: an assignment is needed, so == and === are quiet.
Nothing warns that did not warn before.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
git shows no content for a renamed file, so a text or document file renamed
to a code name (notes.txt to payload.py) was never read by the commit and
push reviews. They now read it whole, behind the files they already review
and within a small share of the prompt. Other renames are reviewed as before.
Part of #4852
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
`git \` with `commit` or `push` on the next line started the hook, but the
command patterns only allowed whitespace between the words, so nothing was
reviewed. They now also allow a line continued with a backslash (Bash) or a
backtick (PowerShell).
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
gave no verdict
When one command commits and pushes, the push review left the new commit to
the commit review and ended at once. A commit review that then failed, was
killed or skipped the commit left it pushed and unreviewed, and no later push
covers it.
The push review now waits for the commit review's process to end, and reviews
the pushed range itself when there is no verdict. Both choose the commit
through one helper, so they cannot pick differently.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
Only the first 20 keys of a hook's metrics are kept. Commit reviews with findings, and failed reviews that used tokens, often have more, so their time and sometimes their result were lost. The result, kind and time of a review now come first.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
security-guidance: bump version to 2.0.20
|
|
security-guidance: only run git in ordinary checkouts
|
|
On Windows without a packaged Python, python3 under WindowsApps is App
Installer's alias, which is not a Python. The launcher started it on every
hook. Starting it waits for Windows' package services, so while those are
stuck every hook in every session waits with them, and the edit hook had no
time limit of its own: each Edit or Write then took 600 seconds.
- The launcher tells that alias by where its link leads and does not start
it. Which interpreter runs the hook is unchanged. Aliases of a packaged
Python are still probed.
- The edit hook gets a 60 second limit.
Refs #6348, #3870, #6315
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
The reviewer returns its findings through a tool call, and nothing reads
what it writes beside it. With only the file-reading tools it spends much
of the review writing prose before that call.
The investigate stages now get a rule that says so, kept out of the shared
prompt in review_api and left out when thinking is off or the user names
the model, where the text is the reasoning.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
`*.{ts,tsx}` in paths or exclude_paths means either. The number of
combinations is bounded, per glob and over all rules. The README lists
MAX_COMMIT_REVIEWS_PER_HOUR.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
Git before 2.38 has no safe.bareRepository. Check before the prompt and
Stop hooks run git, following the same search git uses to find a
repository, and let a work tree kept elsewhere through core.worktree
through.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
Treat a folder as a repository to run git in only when it has its own
.git entry or the caller named the repository through GIT_DIR, and set
safe.bareRepository=explicit on the hook's git commands. Linked worktrees
and ordinary checkouts work as before.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
The commit and push reviews keep two small records (sg-reviewed-shas,
sg-reviewed-diffs) and one marker per commit under review
(sg-inflight-<sha>) in the git directory.
The plugin should only read and write files it made itself; the session
state file is already written in a similar way. All three are now opened
through one helper, which takes a regular file with no second hard link that
is no larger than a record can be. Anything else at that path is replaced by
a new file when the record is about to be written, and passed over when it
is only read. A folder is left alone. The open never waits, and reads from a
record are capped. Anything old at a marker's path that is not such a file,
other than a folder, is removed without being opened.
Three more fixes on the same lines:
- sg-reviewed-shas was locked with a blocking flock just before the findings
go out, so the findings waited for the lock, up to the hook's timeout. The
hook now waits half a second and then appends without the lock, as it does
where there is no flock.
- Bytes in that record that are not text ended the push review with a
traceback. They now cost only their own row.
- The rewritten record was still in the buffer when the lock was released, so
rows written by several sessions at once were lost. It is flushed first.
Cost: in a repository copied with hard links (cp -al) the records have two
names, so neither copy reads them. The first copy to write starts them
afresh, and commits reviewed before the copy may be reviewed once more.
Co-authored-by: mhegazy <mhegazy@anthropic.com>
|
|
Co-authored-by: Claude <noreply@anthropic.com>
|
|
security-guidance: take the review model from one place, default to the newest listed
|
|
security-guidance: review commits and pushes made through the PowerShell tool
|
|
security-guidance: keep the pip temp folder to a real folder of this user
|
|
security-guidance: review what a commit-and-push command sends out beyond its own commit
|
|
security-guidance: stop custom rule regexes that run too long
|
|
security-guidance: custom rule globs can be relative to the project
|
|
security-guidance: read yaml, torch and numpy load calls across their arguments
|
|
security-guidance: hold back a host's token by the host, not the output setting
|