diff --git a/.github/workflows/security-review.yml b/.github/workflows/security-review.yml index 8926ca5..1bd6ee3 100644 --- a/.github/workflows/security-review.yml +++ b/.github/workflows/security-review.yml @@ -14,15 +14,18 @@ name: Security-Review # Claude Code reviews the head of the pull request and gives its report as its # last answer. A second job, which the agent never sees, posts the report as a # review (changes requested with findings, approved without, and for a pull -# request from outside never more than a comment) and sets the commit status +# request from outside or of a bot never more than a comment) and sets the +# commit status # PR-Security-Review on that head; pull_request_target # attaches the checks of its jobs to the base branch, so a status on the head # is the one result both triggers can give. # # Two modes, each on its own trigger; the run of the other trigger ends after # its first job: -# - A pull request from a branch of the repository itself, opened by a member, -# is reviewed on pull_request with a shell. The shell runs in the sandbox of +# - A pull request from a branch of the repository itself, opened by a member +# or by the coding agent of CodeRabbit, which works on tasks a maintainer +# starts and which the organisation trusts like a member, is reviewed on +# pull_request with a shell. The shell runs in the sandbox of # Claude Code: a process namespace of its own, without the variables that # request an identity token, without the token file, without sudo and # Docker, writing only in the checkout and reaching nothing but the npm @@ -33,12 +36,23 @@ name: Security-Review # inside that sandbox, for the length of one command; none runs beside it. # The organisation variable SECURITY_REVIEW_DATABASE, set to "aus", # switches that off for every repository without a pull request. -# - Anything else (a fork, an author outside the organisation, a bot) is -# reviewed by reading alone on pull_request_target, whose definition comes -# from main: no shell, nothing of the pull request is installed or run. Such -# a job runs in the environment security-review-extern, which each +# - Anything else (a fork, an author outside the organisation, any other bot, +# Dependabot among them) is reviewed by reading alone on +# pull_request_target, whose definition comes from main: no shell, nothing +# of the pull request is installed or run. Such a job runs in the +# environment security-review-extern, which each # repository protects with a maintainer as required reviewer, so that no -# one outside spends the budget of the review. +# one outside spends the budget of the review. Its agent has Read, Grep and +# Glob and no other tool: --tools sets the tools Claude Code offers at all, +# while --allowedTools only lets some run without asking and left Bash, +# among others, on offer, where Claude Code runs a command it judges +# read-only without asking. Before the review a run without sign-in checks +# which tools Claude Code offers, and one more stops the job. So it has no +# bubblewrap and needs none: the agent starts no subprocess that the +# variable CLAUDE_CODE_SUBPROCESS_ENV_SCRUB would guard (no Bash, no hooks, +# no MCP), and the variable is set for the review with a shell only. Set +# without bubblewrap it stopped Claude Code before its first answer, and +# from the start until 09.11.2026 no review from outside got a report. # In both, the agent reads only inside the checkout and writes nothing: the # report is its answer, not a file something in the pull request could # overwrite. Settings, hooks, skills, MCP servers and CLAUDE.md of the @@ -74,7 +88,7 @@ on: permissions: {} env: - CLAUDE_CODE_VERSION: 2.1.296 + CLAUDE_CODE_VERSION: 2.1.321 POSTGRES_VERSION: 18 # The signing key of apt.postgresql.org. POSTGRES_KEY: B97B0AFCAA1A47F044F244A07FCC7D46ACCC4CF8 @@ -96,13 +110,37 @@ jobs: env: ASSOCIATION: ${{ github.event.pull_request.author_association }} AUTHOR_TYPE: ${{ github.event.pull_request.user.type }} + AUTHOR_LOGIN: ${{ github.event.pull_request.user.login }} + AUTHOR_ID: ${{ github.event.pull_request.user.id }} HEAD_REPO: ${{ github.event.pull_request.head.repo.full_name }} + # The bot is trusted by its login and its account id together: the + # login of an app cannot be taken by a user, and the id stays the same + # if an app is ever renamed. Its branch must be one of the repository + # itself, as a member's must. Every command of CodeRabbit that changes + # code, commits or opens a pull request (unit tests, docstrings, fix + # comments, fix-ci, autofix, merge conflicts, local commit, simplify, + # recipes, configuration) needs write access to the repository of the + # one who asks, by its reference of review commands; the coding agent + # is started from its dashboard by members. A pull request of it with + # code from a fork is one a maintainer asked for, as if they had pushed + # that code to a branch themselves. That this rests on the check of + # write access at CodeRabbit, which this workflow cannot test, the + # maintainer accepted on 10.11.2026. Dependabot stays outside: whether + # a run it starts gets the variables of the sign-in is not settled, and + # what it brings in is packages of others, which the review with a + # shell would install. run: | trusted=false - if [ "$HEAD_REPO" = "$GITHUB_REPOSITORY" ] && [ "$AUTHOR_TYPE" = "User" ]; then - case "$ASSOCIATION" in - OWNER|MEMBER|COLLABORATOR) trusted=true ;; - esac + if [ "$HEAD_REPO" = "$GITHUB_REPOSITORY" ]; then + if [ "$AUTHOR_TYPE" = "User" ]; then + case "$ASSOCIATION" in + OWNER|MEMBER|COLLABORATOR) trusted=true ;; + esac + elif [ "$AUTHOR_TYPE" = "Bot" ]; then + case "$AUTHOR_LOGIN:$AUTHOR_ID" in + 'coderabbitai[bot]:136622811') trusted=true ;; + esac + fi fi handles=false if [ "$trusted" = "true" ] && [ "$GITHUB_EVENT_NAME" = "pull_request" ]; then @@ -114,6 +152,8 @@ jobs: echo "handles=$handles" >> "$GITHUB_OUTPUT" if [ "$handles" = "false" ]; then echo "Diesen PR prüft der Lauf des anderen Auslösers." + elif [ "$trusted" = "true" ] && [ "$AUTHOR_TYPE" = "Bot" ]; then + echo "Review mit Shell in der Sandbox: PR aus dem Repository, Autor $AUTHOR_LOGIN." elif [ "$trusted" = "true" ]; then echo "Review mit Shell in der Sandbox: PR aus dem Repository, Autor $ASSOCIATION." else @@ -432,8 +472,10 @@ jobs: # Allowing them opens the kernel to every process of the job through # user namespaces again: a flaw there leads to root, and root to the # token in the memory of the runner. That risk is carried knowingly, for - # pull requests of members only; one from outside gets no shell, and - # this step does not run for it. + # pull requests of members and of the coding agent of CodeRabbit only, + # each from a branch of the repository, which only those with write + # access push to; one from outside gets no shell, and this step does not + # run for it. - name: Sandbox vorbereiten if: ${{ env.TRUSTED == 'true' }} run: | @@ -681,6 +723,7 @@ jobs: - name: Auftrag zusammenstellen env: + AUTHOR_TYPE: ${{ github.event.pull_request.user.type }} PR_NUMBER: ${{ github.event.pull_request.number }} SHA: ${{ github.event.pull_request.head.sha }} BASE_REF: ${{ github.event.pull_request.base.ref }} @@ -860,7 +903,15 @@ jobs: 'des PRs. Was sich nur mit Datenbank belegen ließe, bleibt `open_proof_gap`. ' ) mode = ( - 'Review mit Shell: Der PR kommt von einem Mitglied aus einem Branch des Repositorys. ' + 'Review mit Shell: Der PR kommt ' + + ( + 'vom Coding-Agent von CodeRabbit aus einem Branch des Repositorys. Seinen Code hat ' + 'ein Modell geschrieben, auch nach Text, den andere verfasst haben können; prüfe ' + 'ihn so streng wie den eines Menschen. ' + if os.environ.get('AUTHOR_TYPE') == 'Bot' + else 'von einem Mitglied aus einem Branch des Repositorys. ' + ) + + 'Deine Shell läuft in einer Sandbox: ohne Netz außer der npm-Registry, ohne sudo, ' 'schreibend nur im Arbeitsverzeichnis. Abhängigkeiten sind nicht installiert, und im ' 'Arbeitsverzeichnis selbst lässt sich nichts installieren: `package.json`, die ' @@ -911,8 +962,10 @@ jobs: # read the proxy variables by itself: NODE_USE_ENV_PROXY makes Node do # it, and means nothing outside the sandbox, where none is set. # --safe-mode and --setting-sources user - # load nothing of the checkout; dontAsk turns down every tool not - # allowed instead of asking. + # load nothing of the checkout; --tools offers the allowed tools and no + # other, and dontAsk turns down whatever would ask. With the scrub on, + # Claude Code sets the mode back to default; nothing asks there either, + # since every tool on offer is allowed. - name: Review mit Claude Code run: | token_file="$RUNNER_TEMP/anthropic-identity-token" @@ -963,11 +1016,53 @@ jobs: settings='{"permissions": {"blockReadsOutsideWorkingDirectories": true}}' fi + # The scrub needs bubblewrap, which only the review with a shell installs + # ("Sandbox vorbereiten"). Without a shell, hooks or MCP servers the review by + # reading alone starts nothing it would guard; set there anyway, it ended + # Claude Code before its first answer. + if [ "$TRUSTED" = "true" ]; then + scrub=1 + denied=(WebFetch WebSearch Write Edit NotebookEdit) + else + scrub=0 + denied=(Bash WebFetch WebSearch Write Edit NotebookEdit) + fi + + # The tools Claude Code offers at all (--tools), not only those it runs + # without asking (--allowedTools): without it Bash, Task, Workflow + # and more were on offer in both modes, and Claude Code runs a Bash + # command it judges read-only without asking, also in the review by + # reading alone, where nothing sandboxes it. Checked before the + # review: a run without sign-in names the tools in its first event + # and ends at the sign-in, and one more than allowed stops the job. + offered="$(env -u ACTIONS_ID_TOKEN_REQUEST_TOKEN -u ACTIONS_ID_TOKEN_REQUEST_URL \ + -u ANTHROPIC_FEDERATION_RULE_ID -u ANTHROPIC_ORGANIZATION_ID \ + -u ANTHROPIC_SERVICE_ACCOUNT_ID -u ANTHROPIC_WORKSPACE_ID \ + CLAUDE_CONFIG_DIR="$RUNNER_TEMP/claude-preflight" \ + CLAUDE_CODE_SUBPROCESS_ENV_SCRUB="$scrub" \ + claude -p --safe-mode --setting-sources user --strict-mcp-config --no-session-persistence \ + --settings "$settings" \ + --model "$REVIEW_MODEL" --permission-mode dontAsk \ + --output-format stream-json --verbose \ + --tools "${allowed[@]}" \ + --allowedTools "${allowed[@]}" \ + --disallowedTools "${denied[@]}" \ + <<< "Werkzeuge" 2> /dev/null \ + | jq -c 'select(.type == "system" and .subtype == "init") | .tools' | head -n 1 || true)" + extra="$(jq -n -r --argjson offered "${offered:-null}" --args ' + if $offered == null then "kein Start" else ($offered - $ARGS.positional | join(" ")) end' \ + "${allowed[@]}")" + if [ -n "$extra" ]; then + echo "Claude Code bietet mehr Werkzeuge an als erlaubt: $extra. Das Review läuft so nicht." + exit 1 + fi + echo "Werkzeuge: $offered" + # Arguments are NAME=VALUE pairs for the environment of this run. run_review() { env -u ACTIONS_ID_TOKEN_REQUEST_TOKEN -u ACTIONS_ID_TOKEN_REQUEST_URL \ ANTHROPIC_IDENTITY_TOKEN_FILE="$token_file" \ - CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 \ + CLAUDE_CODE_SUBPROCESS_ENV_SCRUB="$scrub" \ COREPACK_HOME="$cache/corepack" \ NODE_USE_ENV_PROXY=1 \ PATH="$GITHUB_WORKSPACE/pr/.security-review/bin:$PATH" \ @@ -979,8 +1074,9 @@ jobs: --settings "$settings" \ --model "$REVIEW_MODEL" --permission-mode dontAsk \ --output-format stream-json --verbose \ + --tools "${allowed[@]}" \ --allowedTools "${allowed[@]}" \ - --disallowedTools WebFetch WebSearch Write Edit NotebookEdit \ + --disallowedTools "${denied[@]}" \ < "$RUNNER_TEMP/prompt.md" > "$RUNNER_TEMP/claude.jsonl" } # The run ended without an answer because the account has no credit @@ -1178,6 +1274,7 @@ jobs: # artifact comes from the job that read the pull request and only # says what its report found. TRUSTED: ${{ needs.mode.outputs.trusted }} + AUTHOR_TYPE: ${{ github.event.pull_request.user.type }} run: | run_url="$GITHUB_SERVER_URL/$GITHUB_REPOSITORY/actions/runs/$GITHUB_RUN_ID" result="$(cat result/result.txt 2> /dev/null || echo missing)" @@ -1214,10 +1311,12 @@ jobs: # found something or is unclear, approved where it found nothing. # A pull request from outside is never approved by this review: a # model judged text that can be written to steer it, and an approval - # is what such text would aim for. It gets a comment instead. + # is what such text would aim for. It gets a comment instead. So does + # one of a bot: its code is what a model wrote, after text that others + # can have written, and its approval stays with a person. if [ "$result" != "clean" ]; then event=REQUEST_CHANGES - elif [ "$TRUSTED" = "true" ]; then + elif [ "$TRUSTED" = "true" ] && [ "$AUTHOR_TYPE" = "User" ]; then event=APPROVE else event=COMMENT diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index 961edef..07fe5df 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -67,7 +67,7 @@ Das gilt für jede Datei in jedem Repository der Organisation, und die CI prüft 3. Pull Request eröffnen, die Vorlage ausfüllen und auf das zugehörige Issue verweisen. Soll der Pull Request ein Issue schließen, gehört unter den deutschen Text eine englische Zeile `Closes #`: GitHub erkennt nur die englischen Schlüsselwörter, ein „Schließt #12“ bewirkt nichts. 4. Auf die Prüfung warten. `main` ist geschützt, direkte Pushes gibt es nicht. -Jeder Pull Request bekommt dazu ein automatisches Security-Review, das als Kommentar am Pull Request und als Status `PR-Security-Review` erscheint. Geprüft wird jeweils der neueste Commit. Kommt der Pull Request aus einem Fork oder von außerhalb der Organisation, liest das Review den Code nur und führt nichts davon aus, und es startet erst, wenn ein Maintainer es freigibt. Das Review ist ein Hinweis, keine Freigabe: gemergt wird nach dem Blick eines Menschen. +Jeder Pull Request bekommt dazu ein automatisches Security-Review, das als Kommentar am Pull Request und als Status `PR-Security-Review` erscheint. Geprüft wird jeweils der neueste Commit. Kommt der Pull Request aus einem Fork oder von außerhalb der Organisation, liest das Review den Code nur und führt nichts davon aus, und es startet erst, wenn ein Maintainer es freigibt. Pull Requests des Coding-Agents von CodeRabbit aus einem Branch des Repositorys prüft es wie die eines Mitglieds, ohne Freigabe. Ein Approve gibt es ihnen nie: Findet das Review etwas oder bleibt es unklar, bittet es um Änderungen, sonst bleibt es bei einem Kommentar. Das Review ist ein Hinweis, keine Freigabe: gemergt wird nach dem Blick eines Menschen. Große Änderungen bitte vorher als Issue oder Discussion abstimmen. Ein fertiger Pull Request, dessen Richtung nicht passt, ist für beide Seiten ärgerlich.