Skip to content

Security: Jake-Network/radar

Security

docs/SECURITY.md

Security

Radar runs locally. It does not upload source, call an LLM or send telemetry.

What runs and what does not

Indexing, discovery, check, merge-check without --verify, and radar gate and radar watch without --run read repository files and Git objects as data. They do not run application code, package scripts, Git hooks or agent instructions.

land is static by default. --driver ... --allow-execution separately authorizes the named installed structural merge driver, with host permissions, inside a private candidate. Drivers are never selected from repository Git configuration, and merely installing one does not authorize execution. MCP exposes static landing/brief/overlap queries without driver, execution or apply flags.

Repository code runs only with gate --run, watch --run, merge-check --verify --allow-execution or test --allow-execution. Those commands run in a private copy of the selected commits or explicitly selected saved previews, with a timeout, an output limit and a reduced environment: PATH, a private HOME and temp directory, offline settings for package managers, and read access to existing dependency caches. Inherited secrets and language startup variables are removed. Build tools still read their own user configuration: Gradle reads gradle.properties in the shared GRADLE_USER_HOME, and Maven reads ~/.m2/settings.xml, so values kept there reach the build. A plan rule can add named variables (env), linked directories (link) and setup commands, so review those the same way you review the test command.

The private copy is not a sandbox. Test code runs as your user and can read your files, your credentials and the network. Review code before you run it, and use a disposable runner without secrets in CI.

watch --run is session authorization for repeated executions, including newly discovered worktrees when branch selection is automatic. The command prints that scope on stderr. Optional Observed Attribution reruns failed commands on branch subsets under the same authorization; it is separate from the full candidate Fast Gate and cannot strengthen its verdict. Watch is not exposed through MCP.

Watch limits sessions to 32 worktrees and selected branches, and dirty previews to eight selected snapshots of at most 32MiB changed contents each. Fast Gate and optional attribution have separate timeout budgets and share one worker slot.

watch --include-dirty separately opts into temporary snapshots of selected worktrees' staged, unstaged and nonignored untracked edits. Source indexes, refs and working files stay untouched. Ignored directories are excluded; unsupported entries and detected snapshot races block verification. Only relevant changed files are copied, but nonignored files can still contain secrets. See watch limitations.

Git and paths

Git is called with argument arrays, never through a shell. Radar's Git processes disable repository fsmonitor hooks and never enable a pager. The private candidate repository also runs without hooks and ignores user and system Git configuration. Radar does not merge, rebase, push, stash, clean or rewrite user branches in your repository. Explicit land --apply imports immutable objects and creates only a create-only refs/radar/integration/<id>; it does not fast-forward a user branch or modify its index/worktree. Combined candidates live in a temporary repository that is deleted afterward. Radar never fetches. When a shallow clone lacks the history it needs, the error tells you to run git fetch --unshallow.

Paths, refs, schemas and PR metadata are treated as untrusted. Paths must stay inside the repository, symlinks and submodules are refused in combined trees, remote $ref is not followed, and file sizes are capped.

Stored data

Passing gate --trailer records a repository-bound historical report under the existing private run-state directory; show does not promote it to current-head evidence. Digests detect edits but do not authenticate the writer. Syntax caches contain bounded source-derived parser facts and are disposable acceleration data; inventory, import resolution and required checks still run. Conflict briefs return explicitly requested committed excerpts and inferred test links, so their output may contain source text. Resource advisories compare database values privately and serialize no URL/credential values.

Evidence records store an output digest, never raw stdout or stderr. The last 4 KiB of output is shown to you (and included in --json) but not saved. Command arguments and the names of passed-through variables are saved, so keep secrets out of arguments.

Integrity hashes catch accidental changes to local records. They do not prove who made them. A review recorded with radar approve is a declaration with whatever name you pass, not an authenticated identity.

Agents

radar mcp serves one repository over stdio. It does not expose test, approve or any execution flag, so an agent cannot run repository code or record a review through it. It can write Radar state and new plan files, so it is not read-only. Native Claude/Codex Stop hooks run static saved-change previews; the legacy Claude script verifies an active plan. Neither executes repository tests. Hook trust and configuration belong to the agent runtime; skills are instructions, not a security boundary. See agent installation.

CI and installation

The GitHub Actions example uses pull_request with read-only permissions, pinned actions and no persisted checkout credentials. Static inspection runs by default. Running tests needs the repository owner to opt in.

Installers verify the archive checksum, version and member paths, extract only the binary, and leave an existing binary in place if anything fails. Checksums do not protect against a compromised publisher. Releases are not signed or notarized yet.

Reporting

Please report security issues privately to the maintainers before publishing details, and keep source and secrets out of public issues. There is no dedicated private reporting channel yet.

There aren't any published security advisories