Add sandbox-git-write guard hook#10
Closed
technicalpickles wants to merge 5 commits into
Closed
Conversation
A PreToolUse/Bash hook (bin/claude-sandbox-guard) that denies sandboxed git-write and srb commands up front, telling Claude to retry with dangerouslyDisableSandbox=true instead of eating a cryptic EPERM and retrying. Both classes reliably fail under the command sandbox and are safe to run unsandboxed: - git writes hit .git/worktrees/.../index.lock and the hardcoded blocked paths (.claude/, .vscode/, .gitmodules) inside pt worktrees - sorbet (srb) wants to write its mdb / rubocop cache Read-only git (log/show/diff/status, worktree/stash/submodule list) is left sandboxed, so the common path is untouched. claudeconfig.sh now concatenates hooks.<event> arrays across base + active role. The previous deep merge replaced same-event arrays, which would silently drop a base-level PreToolUse hook whenever the active role also defined PreToolUse (work.jsonc does). Behaviour is locked down by a bats suite (test/): hand-written edge cases (env prefixes, git -C, compound chains, read-form list commands, already-unsandboxed calls) plus a data-driven sweep over 90 real commands pulled from the session corpus. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Capture the verified rationale in the hook header: git writes hit a hardcoded path deny (not overridable by allowWrite) and srb fails on LMDB's writable mmap (a syscall deny, not a path the cache dir is already writable inside cwd). Both ruled out allowlisting by testing, so unsandboxing is the only fix. Also makes the srb deny message accurate. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
sandbox-probe disproved the writable-mmap theory (mmap + flock both succeed sandboxed). srb still needs unsandboxing, but the mechanism is a non-path syscall deny, not mmap; suspect LMDB's POSIX named semaphores. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
srb's LMDB cache open fails sandboxed because LMDB on macOS uses SysV semaphores (MDB_USE_SYSV_SEM): the sandbox allows semget but denies semctl/semop with EPERM. Confirmed via pickletown projects/sandbox-probe (lmdb_replay.py replaying mdb_env_open). Replaces the earlier "named semaphores (sem_open)" suspicion in the comment and deny message; sem_open is not even linked into the sorbet binary. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ps and top are setuid-root; the sandbox denies exec of setuid/setgid binaries, so they EPERM sandboxed (bare `ps` is shimmed through rtk, which execs /bin/ps and hits the same deny). Both are read-only and safe unsandboxed, so they fit the existing deny-and-retry pattern. Matches bare, path-qualified, and rtk-wrapped forms; anchored so psql/topgrade/ pstree don't false-trigger. Only ps/top are listed (the tools agents actually run); other setuid binaries (su, crontab) are left to fail loudly rather than auto-unsandboxed. Mechanism confirmed via pickletown projects/sandbox-probe: the sysctl ps reads (KERN_PROC) is allowed sandboxed; the friction is purely setuid exec, which allowWrite can't touch. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Owner
Author
|
Closing without merging. This guard proactively denied sandboxed commands that reliably fail (git-write, srb, ps/top) and told Claude to retry unsandboxed. Auto-mode now handles that sandboxed -> unsandboxed decision in the harness itself, so the proactive deny is redundant in practice. The real value here was the investigation, and that's preserved: the sandbox deny-surface (setuid/setgid exec deny, LMDB SysV-semaphore deny, the path/LaunchServices/sysctl denies) is documented in pickletown projects/sandbox-probe plus a reference script (lmdb_replay.py) and memory notes. Its complement, the sandbox-first plugin, is being removed for the same reason (pickled-claude-plugins#90). Leaving the branch up in case the non-auto-mode edges (subagents, cron, teammates) make it worth revisiting. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
bin/claude-sandbox-guard, a PreToolUse/Bash hook that denies sandboxed git-write andsrbcommands and tells Claude to retry withdangerouslyDisableSandbox=true. Both reliably fail under the sandbox (git writes hit.git/worktrees/.../index.lockand the blocked.claude/.vscode/.gitmodulespaths in pt worktrees; sorbet wants its mdb/rubocop cache) and are safe to run unsandboxed. Read-only git stays sandboxed, so the common path is untouched.claudeconfig.shto concatenatehooks.<event>arrays across base + active role instead of deep-merging them. The old merge replaced same-event arrays, so a base-levelPreToolUsehook got silently dropped whenever the active role also definedPreToolUse(work.jsonc does).srbas the clear "always retry unsandboxed" patterns./tmpwrites were deliberately left out, since the fix there is$TMPDIR, not loosening the sandbox.Test plan
npm test(lint + bats). The bats suite covers hand-written edge cases (env prefixes,git -C, compound chains, read-formlistcommands, already-unsandboxed calls, non-Bash tools) plus a sweep over 90 real commands pulled from the corpus.Activating it
Dormant until
claudeconfig.shregenerates~/.claude/settings.json. Heads up first: the live settings hasbowerbird/rtk/plannotatorhooks that aren't tracked in any role, so a regen drops them. Pre-existing drift, worth sorting out separately.