claude/sigit-cloud-agent-expansion-reox0a
claude/github-issue-40-0au5ua
claude/linked-issue-acp-tui-u4vvpa
claude/minor-release-cut-vkugow
claude/sigit-acp-local-chat-cx6380
claude/sigit-cloud-agent-expansion-reox0a
claude/tool-permission-system
claude/zen-feynman-0u78dk
development
feature/acp-cancel-request
feature/acp-session-list-cursor
feature/acp-v1-conformance
feature/agent-tools-multiedit-glob-todos-remember
feature/cargo-workspace
feature/hooks
feature/load-local-model-explicitly
feature/sigit-agent-profile
feature/sigit-code-cloud
feature/xcode-mcp-server
fix/acp-replay-loaded-session-history
fix/acp-replay-reasoning-and-attachments
fix/dsml-doubled-bar-markers
main
release/v1.5.10
release/v1.5.6
release/v1.6.1
| 1 | //! Tool permission policy: which agent tools may run, and when to ask. |
| 2 | //! |
| 3 | //! Every tool call funnels through one decision point before execution |
| 4 | //! (`decision_for`). Tools are classified by risk: *read-only* tools (reading |
| 5 | //! files, searching, listing, fetching a web page) always run, while *mutating* |
| 6 | //! tools (writing files, deleting, shell commands, MCP tools) are governed by |
| 7 | //! policy. The policy layers, first match wins: |
| 8 | //! |
| 9 | //! 1. **Plan mode** — a per-session switch that denies every mutating tool with |
| 10 | //! a message telling the model to present a plan instead. Toggled via |
| 11 | //! `/plan on|off` (TUI and ACP). |
| 12 | //! 2. **Session grants** — "always allow this session", recorded when the user |
| 13 | //! picks that option in an approval prompt. For `run_command` the grant is |
| 14 | //! scoped to the command's first two whitespace-separated tokens (approving |
| 15 | //! `git push origin main` records `run_command(git push)`); other tools |
| 16 | //! record the bare tool name. |
| 17 | //! 3. **Rule lists** — `[permissions.rules]` in `settings.toml`: ordered |
| 18 | //! `deny` and `allow` lists of rules shaped `tool_name` or |
| 19 | //! `tool_name(argument_pattern)`, e.g. `run_command(git status)`, |
| 20 | //! `run_command(cargo *)`, `edit_file(src/*)`. The pattern matches the |
| 21 | //! command string for `run_command` and the path argument for the |
| 22 | //! file-mutating tools; for tools with no obvious argument (MCP tools, |
| 23 | //! unknown tools) only a bare `tool_name` rule matches. `deny` is checked |
| 24 | //! before `allow`, so a deny always beats an allow matching the same call. |
| 25 | //! 4. **Per-tool override** — `[permissions.tools]` in `settings.toml`, e.g. |
| 26 | //! `run_command = "ask"`, `edit_file = "allow"`, `delete_file = "deny"`. |
| 27 | //! 5. **Default mode** — `[permissions] default = "ask"|"allow"|"deny"` in |
| 28 | //! `settings.toml`; `ask` on a fresh install. |
| 29 | //! |
| 30 | //! Pattern matching (rules and session grants share it): `*` is a glob-style |
| 31 | //! wildcard (the `glob` tool's translator). A pattern ending in `*` matches |
| 32 | //! everything from the wildcard on — `run_command(cargo *)` covers |
| 33 | //! `cargo build src/main.rs`. A pattern without a trailing `*` must be the |
| 34 | //! whole argument or end at a whitespace boundary: `run_command(git status)` |
| 35 | //! matches `git status` and `git status --short` but not `git status-x`. |
| 36 | //! |
| 37 | //! The `SIGIT_PERMISSIONS` env var (`allow`/`ask`/`deny`) overrides the stored |
| 38 | //! default without writing the file — the escape hatch for ACP clients that |
| 39 | //! cannot answer `session/request_permission` and for CI/headless runs. |
| 40 | //! |
| 41 | //! Tools discovered from MCP servers (`mcp__*`) and any unknown tool name are |
| 42 | //! treated as mutating: external tools can have arbitrary side effects, so the |
| 43 | //! safe assumption is to gate them. The exceptions are the query tools of the |
| 44 | //! two baked-in servers — the official sigit.si server and the smbCloud CLI — |
| 45 | //! which are read-only (see [`classify`]). |
| 46 | //! |
| 47 | //! Session state (grants + plan mode) lives in a process-global keyed by |
| 48 | //! session id — the same pattern as `mcp.rs`'s server cache — so the ACP |
| 49 | //! multi-session surface and the single-session TUI share one implementation. |
| 50 | |
| 51 | use std::collections::{HashMap, HashSet}; |
| 52 | use std::sync::{Mutex, OnceLock}; |
| 53 | |
| 54 | use regex::Regex; |
| 55 | |
| 56 | use crate::settings::{self, PermissionMode}; |
| 57 | use crate::tools::glob_to_regex; |
| 58 | |
| 59 | /// Session key used by the interactive TUI, which only ever has one session. |
| 60 | /// The TUI (`chat.rs`) is `#[cfg(unix)]`, so this is its only consumer and is |
| 61 | /// dead on non-Unix targets — the rest of the module is used on all platforms. |
| 62 | #[cfg_attr(not(unix), allow(dead_code))] |
| 63 | pub const TUI_SESSION: &str = "tui"; |
| 64 | |
| 65 | /// How risky a tool is to run without the user's sign-off. |
| 66 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 67 | pub enum ToolRisk { |
| 68 | /// Observes state without changing it; always allowed to run. |
| 69 | ReadOnly, |
| 70 | /// Changes files, runs commands, or has unknown side effects; governed by |
| 71 | /// the permission policy. |
| 72 | Mutating, |
| 73 | } |
| 74 | |
| 75 | /// The outcome of the policy check for one tool call. |
| 76 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 77 | pub enum Decision { |
| 78 | /// Run the tool without asking. |
| 79 | Allow, |
| 80 | /// Ask the user before running (surface-specific: ACP permission request |
| 81 | /// or TUI approval prompt). |
| 82 | Ask, |
| 83 | /// Do not run the tool; the string is returned to the model as the tool |
| 84 | /// result so it can adapt instead of retrying blindly. |
| 85 | Deny(String), |
| 86 | } |
| 87 | |
| 88 | /// Classify a tool by name. Unknown names and MCP tools are mutating: the |
| 89 | /// conservative default for anything whose side effects we can't see. |
| 90 | /// `task` is read-only because the subagent it launches is restricted to the |
| 91 | /// read-only toolset (see `SUBAGENT_TOOL_NAMES` in `tools.rs`), so delegated |
| 92 | /// research stays available in plan mode. The native `web_search` tool is |
| 93 | /// read-only for the same reason `read_website` is — it only ever returns |
| 94 | /// data, even though it forwards to an `mcp__*` tool under the hood (see |
| 95 | /// `tools.rs`'s `web_search` module doc for why it isn't classified by that |
| 96 | /// `mcp__` prefix). |
| 97 | /// |
| 98 | /// Two MCP exceptions, both baked-in first-party servers. The *official* |
| 99 | /// sigit.si server's (`mcp__sigit__*`) query tools — names starting `list_` |
| 100 | /// or `get_`, plus `search_code` and `web_search` — are read-only and never |
| 101 | /// prompt (the TUI's Repo tab depends on this). The smbCloud CLI server's |
| 102 | /// (`mcp__smbcloud__*`) read tools are exempted by an explicit allow-list |
| 103 | /// rather than a name shape, since its names don't follow the `list_`/`get_` |
| 104 | /// convention — a tool the list doesn't know (e.g. from a newer smb) fails |
| 105 | /// safe to mutating. Every other `mcp__*` tool stays mutating. |
| 106 | pub fn classify(tool_name: &str) -> ToolRisk { |
| 107 | match tool_name { |
| 108 | "read_file" | "list_directory" | "search_files" | "glob" | "read_website" |
| 109 | | "write_todos" | "skill" | "task" | "command_output" | "web_search" => ToolRisk::ReadOnly, |
| 110 | _ => { |
| 111 | if let Some(bare) = crate::mcp::official_tool_suffix(tool_name) |
| 112 | && (bare.starts_with("list_") |
| 113 | || bare.starts_with("get_") |
| 114 | || bare == "search_code" |
| 115 | || bare == "web_search") |
| 116 | { |
| 117 | return ToolRisk::ReadOnly; |
| 118 | } |
| 119 | if let Some(bare) = crate::mcp::smbcloud_tool_suffix(tool_name) |
| 120 | && matches!(bare, "me" | "deployments" | "project_list" | "project_show") |
| 121 | { |
| 122 | return ToolRisk::ReadOnly; |
| 123 | } |
| 124 | ToolRisk::Mutating |
| 125 | } |
| 126 | } |
| 127 | } |
| 128 | |
| 129 | /// Per-session permission state. |
| 130 | #[derive(Default)] |
| 131 | struct SessionPerms { |
| 132 | /// Tools the user chose "always allow this session" for. |
| 133 | always_allow: HashSet<String>, |
| 134 | /// When set, every mutating tool is denied with a plan-mode message. |
| 135 | plan_mode: bool, |
| 136 | } |
| 137 | |
| 138 | fn sessions() -> &'static Mutex<HashMap<String, SessionPerms>> { |
| 139 | static SESSIONS: OnceLock<Mutex<HashMap<String, SessionPerms>>> = OnceLock::new(); |
| 140 | SESSIONS.get_or_init(|| Mutex::new(HashMap::new())) |
| 141 | } |
| 142 | |
| 143 | fn with_session<T>(session: &str, f: impl FnOnce(&mut SessionPerms) -> T) -> T { |
| 144 | let mut map = sessions() |
| 145 | .lock() |
| 146 | .unwrap_or_else(|poisoned| poisoned.into_inner()); |
| 147 | f(map.entry(session.to_string()).or_default()) |
| 148 | } |
| 149 | |
| 150 | /// The message returned to the model when a mutating tool is blocked by plan |
| 151 | /// mode. Instructive rather than terse so the model changes course in one turn. |
| 152 | fn plan_mode_denial(tool_name: &str) -> String { |
| 153 | format!( |
| 154 | "Plan mode is active: `{tool_name}` was not executed because it modifies state. \ |
| 155 | Present a concise plan of the changes you intend to make and ask the user to \ |
| 156 | approve it (they can run /plan off to enable execution). Read-only tools \ |
| 157 | (read_file, search_files, glob, list_directory) remain available for research." |
| 158 | ) |
| 159 | } |
| 160 | |
| 161 | /// The message returned to the model when the user (or policy) denies a tool. |
| 162 | pub fn user_denial(tool_name: &str) -> String { |
| 163 | format!( |
| 164 | "The user denied permission to run `{tool_name}`. Do not retry the same call. \ |
| 165 | Explain what you wanted to do and ask the user how to proceed, or continue \ |
| 166 | with an approach that does not need this tool." |
| 167 | ) |
| 168 | } |
| 169 | |
| 170 | /// Render a tool call's arguments for an approval prompt. The person deciding |
| 171 | /// must be able to see what they are approving, so the cap is generous and any |
| 172 | /// cut is marked with how much is hidden — silently truncating could hide the |
| 173 | /// tail of a command from the user who is about to allow it. |
| 174 | pub fn approval_preview(arguments: &str) -> String { |
| 175 | const MAX_CHARS: usize = 500; |
| 176 | let total = arguments.chars().count(); |
| 177 | if total <= MAX_CHARS { |
| 178 | return arguments.to_string(); |
| 179 | } |
| 180 | let shown: String = arguments.chars().take(MAX_CHARS).collect(); |
| 181 | format!("{shown}… [+{} more chars]", total - MAX_CHARS) |
| 182 | } |
| 183 | |
| 184 | // ── Rule patterns ──────────────────────────────────────────────────────────── |
| 185 | // A rule is `tool_name` or `tool_name(argument_pattern)`; rules from |
| 186 | // `[permissions.rules]` and session grants share this matcher. |
| 187 | |
| 188 | /// Split a rule into tool name and optional argument pattern: |
| 189 | /// `run_command(git *)` → `("run_command", Some("git *"))`; |
| 190 | /// `edit_file` → `("edit_file", None)`. |
| 191 | fn parse_rule(rule: &str) -> (&str, Option<&str>) { |
| 192 | let rule = rule.trim(); |
| 193 | if let Some(open) = rule.find('(') |
| 194 | && let Some(inner) = rule[open + 1..].strip_suffix(')') |
| 195 | { |
| 196 | return (rule[..open].trim(), Some(inner)); |
| 197 | } |
| 198 | (rule, None) |
| 199 | } |
| 200 | |
| 201 | /// Compile a rule's argument pattern. Reuses the `glob` tool's translator |
| 202 | /// (`*`, `**`, `?`, `{a,b}`), then swaps its end anchor for rule semantics: |
| 203 | /// a pattern ending in `*` matches everything from the wildcard on (prefix |
| 204 | /// semantics — `cargo *` covers `cargo build src/main.rs`), while any other |
| 205 | /// pattern must be the whole argument or end at a whitespace boundary |
| 206 | /// (`git status` matches `git status --short` but not `git status-x`). |
| 207 | fn pattern_regex(pattern: &str) -> Option<Regex> { |
| 208 | let anchored = glob_to_regex(pattern); |
| 209 | let body = anchored.strip_suffix('$').unwrap_or(&anchored); |
| 210 | let source = if pattern.ends_with('*') { |
| 211 | body.to_string() |
| 212 | } else { |
| 213 | format!("{body}(?:$|\\s)") |
| 214 | }; |
| 215 | Regex::new(&source).ok() |
| 216 | } |
| 217 | |
| 218 | /// Whether one rule covers one tool call. A bare `tool_name` rule matches any |
| 219 | /// call of that tool; a pattern rule additionally needs the call's matchable |
| 220 | /// argument to fit the pattern — so a pattern rule never matches a tool that |
| 221 | /// has no matchable argument (an unreadable rule must not widen access). |
| 222 | fn rule_matches(rule: &str, tool_name: &str, argument: Option<&str>) -> bool { |
| 223 | let (rule_tool, pattern) = parse_rule(rule); |
| 224 | if rule_tool != tool_name { |
| 225 | return false; |
| 226 | } |
| 227 | match (pattern, argument) { |
| 228 | (None, _) => true, |
| 229 | (Some(pattern), Some(argument)) => { |
| 230 | pattern_regex(pattern).is_some_and(|re| re.is_match(argument)) |
| 231 | } |
| 232 | (Some(_), None) => false, |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | /// The argument a rule pattern is matched against, extracted from the tool |
| 237 | /// call's raw JSON arguments: the command string for `run_command`, the path |
| 238 | /// for the file-mutating tools. Read-only tools never reach the matcher, and |
| 239 | /// other tools (MCP, unknown) have no obvious single argument, so they return |
| 240 | /// `None` and are governed only by bare `tool_name` rules. |
| 241 | fn matchable_argument(tool_name: &str, arguments: &str) -> Option<String> { |
| 242 | let key = match tool_name { |
| 243 | "run_command" => "command", |
| 244 | "edit_file" | "create_file" | "multi_edit" | "delete_file" | "create_directory" => "path", |
| 245 | _ => return None, |
| 246 | }; |
| 247 | let value: serde_json::Value = serde_json::from_str(arguments).ok()?; |
| 248 | Some(value.get(key)?.as_str()?.to_string()) |
| 249 | } |
| 250 | |
| 251 | /// Policy check for one tool call. `arguments` is the call's raw JSON argument |
| 252 | /// string, consulted by rule patterns and granular session grants. See the |
| 253 | /// module docs for the layering. |
| 254 | pub fn decision_for(session: &str, tool_name: &str, arguments: &str) -> Decision { |
| 255 | let read_only = classify(tool_name) == ToolRisk::ReadOnly; |
| 256 | let argument = matchable_argument(tool_name, arguments); |
| 257 | let rules = settings::permission_rules(); |
| 258 | |
| 259 | // An explicit deny applies to every tool, read-only ones included. This is |
| 260 | // what lets a user still switch off a first-party query tool that |
| 261 | // `classify` now treats as read-only (`web_search`, `list_issues`, …); |
| 262 | // without it, the reclassification would silently stop honoring a deny that |
| 263 | // was enforced when those names counted as mutating. |
| 264 | if let Some(rule) = rules |
| 265 | .deny |
| 266 | .iter() |
| 267 | .find(|rule| rule_matches(rule, tool_name, argument.as_deref())) |
| 268 | { |
| 269 | return Decision::Deny(format!( |
| 270 | "`{tool_name}` is denied by the permission rule `{rule}` in settings.toml. \ |
| 271 | Do not retry it; work without this tool or ask the user to change \ |
| 272 | the policy." |
| 273 | )); |
| 274 | } |
| 275 | |
| 276 | if read_only { |
| 277 | // Read-only tools never prompt and are never plan-mode blocked, but an |
| 278 | // explicit per-tool `deny` still turns one off. Crucially we consult |
| 279 | // the *explicit* override, not `permission_mode_for`, so the default |
| 280 | // `ask`/`deny` mode never starts prompting on every read_file. |
| 281 | if settings::permission_mode_explicit(tool_name) == Some(PermissionMode::Deny) { |
| 282 | return Decision::Deny(format!( |
| 283 | "`{tool_name}` is denied by the permission policy in settings.toml. \ |
| 284 | Do not retry it; work without this tool or ask the user to change \ |
| 285 | the policy." |
| 286 | )); |
| 287 | } |
| 288 | return Decision::Allow; |
| 289 | } |
| 290 | |
| 291 | let (plan_mode, granted) = with_session(session, |s| { |
| 292 | ( |
| 293 | s.plan_mode, |
| 294 | s.always_allow |
| 295 | .iter() |
| 296 | .any(|grant| rule_matches(grant, tool_name, argument.as_deref())), |
| 297 | ) |
| 298 | }); |
| 299 | |
| 300 | if plan_mode { |
| 301 | return Decision::Deny(plan_mode_denial(tool_name)); |
| 302 | } |
| 303 | if granted { |
| 304 | return Decision::Allow; |
| 305 | } |
| 306 | |
| 307 | if rules |
| 308 | .allow |
| 309 | .iter() |
| 310 | .any(|rule| rule_matches(rule, tool_name, argument.as_deref())) |
| 311 | { |
| 312 | return Decision::Allow; |
| 313 | } |
| 314 | |
| 315 | match settings::permission_mode_for(tool_name) { |
| 316 | PermissionMode::Allow => Decision::Allow, |
| 317 | PermissionMode::Ask => Decision::Ask, |
| 318 | PermissionMode::Deny => Decision::Deny(format!( |
| 319 | "`{tool_name}` is denied by the permission policy in settings.toml. \ |
| 320 | Do not retry it; work without this tool or ask the user to change \ |
| 321 | the policy." |
| 322 | )), |
| 323 | } |
| 324 | } |
| 325 | |
| 326 | /// Record an "always allow this session" grant for a tool call. For |
| 327 | /// `run_command` the grant is scoped to the command's first two |
| 328 | /// whitespace-separated tokens (approving `git push origin main` records |
| 329 | /// `run_command(git push)`, covering the `git push …` family only); other |
| 330 | /// tools record the bare tool name, matching any call. |
| 331 | pub fn grant_for_session(session: &str, tool_name: &str, arguments: &str) { |
| 332 | let grant = session_grant_rule(tool_name, arguments); |
| 333 | with_session(session, |s| { |
| 334 | s.always_allow.insert(grant); |
| 335 | }); |
| 336 | } |
| 337 | |
| 338 | /// The rule string recorded for one approved call (see [`grant_for_session`]). |
| 339 | /// Falls back to the bare tool name when the command is absent or empty, which |
| 340 | /// grants the whole tool — exactly what the pre-granular behavior was. |
| 341 | fn session_grant_rule(tool_name: &str, arguments: &str) -> String { |
| 342 | if tool_name == "run_command" |
| 343 | && let Some(command) = matchable_argument(tool_name, arguments) |
| 344 | { |
| 345 | let prefix: Vec<&str> = command.split_whitespace().take(2).collect(); |
| 346 | if !prefix.is_empty() { |
| 347 | return format!("{tool_name}({})", prefix.join(" ")); |
| 348 | } |
| 349 | } |
| 350 | tool_name.to_string() |
| 351 | } |
| 352 | |
| 353 | /// Toggle plan mode for a session. Returns the new state. |
| 354 | pub fn set_plan_mode(session: &str, enabled: bool) -> bool { |
| 355 | with_session(session, |s| { |
| 356 | s.plan_mode = enabled; |
| 357 | s.plan_mode |
| 358 | }) |
| 359 | } |
| 360 | |
| 361 | /// Whether plan mode is active for a session. |
| 362 | pub fn plan_mode(session: &str) -> bool { |
| 363 | with_session(session, |s| s.plan_mode) |
| 364 | } |
| 365 | |
| 366 | /// Drop all recorded state for a session (fresh session, /clear, or session |
| 367 | /// teardown) so grants never outlive the conversation they were given in. |
| 368 | pub fn reset_session(session: &str) { |
| 369 | let mut map = sessions() |
| 370 | .lock() |
| 371 | .unwrap_or_else(|poisoned| poisoned.into_inner()); |
| 372 | map.remove(session); |
| 373 | } |
| 374 | |
| 375 | /// Drop the recorded state for *every* session. Called at ACP session |
| 376 | /// boundaries (new/load/fork): the agent drives one shared engine, so only one |
| 377 | /// conversation is live at a time and grants must never cross a boundary. This |
| 378 | /// also keeps the map from accumulating entries for session ids that will |
| 379 | /// never be used again. |
| 380 | pub fn reset_all() { |
| 381 | let mut map = sessions() |
| 382 | .lock() |
| 383 | .unwrap_or_else(|poisoned| poisoned.into_inner()); |
| 384 | map.clear(); |
| 385 | } |
| 386 | |
| 387 | /// Status summary for `/permissions` and `/status`: the default mode, plan |
| 388 | /// mode, the granular session grants, and the active rule lists. |
| 389 | pub fn describe(session: &str) -> String { |
| 390 | let plan = if plan_mode(session) { "on" } else { "off" }; |
| 391 | let default = settings::permission_default(); |
| 392 | let granted = with_session(session, |s| { |
| 393 | let mut names: Vec<&str> = s.always_allow.iter().map(String::as_str).collect(); |
| 394 | names.sort_unstable(); |
| 395 | names.join(", ") |
| 396 | }); |
| 397 | let granted = if granted.is_empty() { |
| 398 | "none".to_string() |
| 399 | } else { |
| 400 | granted |
| 401 | }; |
| 402 | let rules = settings::permission_rules(); |
| 403 | let render = |list: &[String]| { |
| 404 | if list.is_empty() { |
| 405 | "none".to_string() |
| 406 | } else { |
| 407 | list.join(", ") |
| 408 | } |
| 409 | }; |
| 410 | format!( |
| 411 | "permissions: default={default} | plan mode: {plan} | session grants: {granted}\n\ |
| 412 | rules: deny: {} | allow: {}\n\ |
| 413 | read-only tools always run; configure [permissions] in settings.toml", |
| 414 | render(&rules.deny), |
| 415 | render(&rules.allow), |
| 416 | ) |
| 417 | } |
| 418 | |
| 419 | #[cfg(test)] |
| 420 | mod tests { |
| 421 | use super::*; |
| 422 | |
| 423 | /// `decision_for` reads settings (env + file), and the settings test |
| 424 | /// mutates `SIGIT_CONFIG_DIR`/`SIGIT_PERMISSIONS` under this lock — hold it |
| 425 | /// here too so parallel test runs don't race, and point the config dir at |
| 426 | /// an empty sandbox so a developer's real settings.toml can't skew results. |
| 427 | fn env_guard() -> std::sync::MutexGuard<'static, ()> { |
| 428 | let guard = crate::ENV_TEST_LOCK |
| 429 | .lock() |
| 430 | .unwrap_or_else(|poisoned| poisoned.into_inner()); |
| 431 | let dir = std::env::temp_dir().join(format!("sigit_perm_tests_{}", std::process::id())); |
| 432 | // Start from an empty sandbox: a settings.toml written by an earlier |
| 433 | // test in this process (e.g. one storing rule lists) must not leak |
| 434 | // into the next. |
| 435 | let _ = std::fs::remove_dir_all(&dir); |
| 436 | // SAFETY: process-global env mutation, serialized by ENV_TEST_LOCK; the |
| 437 | // other env-touching tests re-set these before reading. |
| 438 | unsafe { std::env::set_var("SIGIT_CONFIG_DIR", &dir) }; |
| 439 | unsafe { std::env::remove_var("SIGIT_PERMISSIONS") }; |
| 440 | guard |
| 441 | } |
| 442 | |
| 443 | /// Persist rule lists into the sandboxed settings.toml. |
| 444 | fn store_rules(allow: &[&str], deny: &[&str]) { |
| 445 | let mut settings = settings::load(); |
| 446 | settings.permissions.rules.allow = allow.iter().map(|s| s.to_string()).collect(); |
| 447 | settings.permissions.rules.deny = deny.iter().map(|s| s.to_string()).collect(); |
| 448 | settings::store(&settings).unwrap(); |
| 449 | } |
| 450 | |
| 451 | /// `decision_for` on a `run_command` call with the given command string. |
| 452 | fn run_command_decision(session: &str, command: &str) -> Decision { |
| 453 | let args = serde_json::json!({ "command": command }).to_string(); |
| 454 | decision_for(session, "run_command", &args) |
| 455 | } |
| 456 | |
| 457 | #[test] |
| 458 | fn read_only_tools_always_allowed() { |
| 459 | let _guard = env_guard(); |
| 460 | for tool in [ |
| 461 | "read_file", |
| 462 | "list_directory", |
| 463 | "search_files", |
| 464 | "glob", |
| 465 | "read_website", |
| 466 | "write_todos", |
| 467 | "skill", |
| 468 | "task", |
| 469 | "command_output", |
| 470 | "web_search", |
| 471 | ] { |
| 472 | assert_eq!(classify(tool), ToolRisk::ReadOnly, "{tool}"); |
| 473 | assert_eq!(decision_for("t-ro", tool, "{}"), Decision::Allow, "{tool}"); |
| 474 | } |
| 475 | } |
| 476 | |
| 477 | #[test] |
| 478 | fn mutating_and_unknown_tools_are_gated() { |
| 479 | for tool in [ |
| 480 | "edit_file", |
| 481 | "multi_edit", |
| 482 | "create_file", |
| 483 | "create_directory", |
| 484 | "delete_file", |
| 485 | "run_command", |
| 486 | "kill_command", |
| 487 | "remember", |
| 488 | "mcp__server__anything", |
| 489 | "totally_unknown_tool", |
| 490 | ] { |
| 491 | assert_eq!(classify(tool), ToolRisk::Mutating, "{tool}"); |
| 492 | } |
| 493 | } |
| 494 | |
| 495 | #[test] |
| 496 | fn official_mcp_query_tools_are_read_only() { |
| 497 | let _guard = env_guard(); |
| 498 | for tool in [ |
| 499 | "mcp__sigit__list_issues", |
| 500 | "mcp__sigit__list_pull_requests", |
| 501 | "mcp__sigit__get_issue", |
| 502 | "mcp__sigit__get_pull_request", |
| 503 | "mcp__sigit__list_repositories", |
| 504 | "mcp__sigit__get_file_contents", |
| 505 | "mcp__sigit__search_code", |
| 506 | "mcp__sigit__web_search", |
| 507 | ] { |
| 508 | assert_eq!(classify(tool), ToolRisk::ReadOnly, "{tool}"); |
| 509 | // Read-only means it never prompts, whatever the policy layers say. |
| 510 | assert_eq!( |
| 511 | decision_for("t-official", tool, "{}"), |
| 512 | Decision::Allow, |
| 513 | "{tool}" |
| 514 | ); |
| 515 | } |
| 516 | } |
| 517 | |
| 518 | #[test] |
| 519 | fn smbcloud_mcp_read_tools_are_read_only() { |
| 520 | let _guard = env_guard(); |
| 521 | for tool in [ |
| 522 | "mcp__smbcloud__me", |
| 523 | "mcp__smbcloud__deployments", |
| 524 | "mcp__smbcloud__project_list", |
| 525 | "mcp__smbcloud__project_show", |
| 526 | ] { |
| 527 | assert_eq!(classify(tool), ToolRisk::ReadOnly, "{tool}"); |
| 528 | assert_eq!( |
| 529 | decision_for("t-smbcloud", tool, "{}"), |
| 530 | Decision::Allow, |
| 531 | "{tool}" |
| 532 | ); |
| 533 | } |
| 534 | } |
| 535 | |
| 536 | #[test] |
| 537 | fn smbcloud_mcp_mutating_and_unknown_tools_stay_gated() { |
| 538 | for tool in [ |
| 539 | // smbcloud server, but not on the read allow-list |
| 540 | "mcp__smbcloud__project_create", |
| 541 | "mcp__smbcloud__project_update", |
| 542 | "mcp__smbcloud__project_delete", |
| 543 | "mcp__smbcloud__", |
| 544 | // allow-listed names on other servers get no exemption |
| 545 | "mcp__other__project_list", |
| 546 | "mcp__other__me", |
| 547 | // `smbcloud` must match the whole server name |
| 548 | "mcp__smbcloudx__project_list", |
| 549 | ] { |
| 550 | assert_eq!(classify(tool), ToolRisk::Mutating, "{tool}"); |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | #[test] |
| 555 | fn read_only_tool_still_honors_an_explicit_deny() { |
| 556 | let _guard = env_guard(); |
| 557 | // A deny rule listing a read-only first-party query tool is enforced, |
| 558 | // even though the tool never prompts otherwise. |
| 559 | store_rules(&[], &["mcp__sigit__web_search"]); |
| 560 | assert!(matches!( |
| 561 | decision_for("t", "mcp__sigit__web_search", "{}"), |
| 562 | Decision::Deny(_) |
| 563 | )); |
| 564 | // A tool with no deny rule is unaffected. |
| 565 | assert_eq!( |
| 566 | decision_for("t", "mcp__sigit__list_issues", "{}"), |
| 567 | Decision::Allow |
| 568 | ); |
| 569 | |
| 570 | // An explicit per-tool `deny` mode also turns a read-only tool off, |
| 571 | // while the default mode never makes read-only tools prompt. |
| 572 | let mut settings = settings::load(); |
| 573 | settings.permissions.rules.deny.clear(); |
| 574 | settings.permissions.default = PermissionMode::Ask; |
| 575 | settings |
| 576 | .permissions |
| 577 | .tools |
| 578 | .insert("mcp__sigit__web_search".to_string(), PermissionMode::Deny); |
| 579 | settings::store(&settings).unwrap(); |
| 580 | assert!(matches!( |
| 581 | decision_for("t", "mcp__sigit__web_search", "{}"), |
| 582 | Decision::Deny(_) |
| 583 | )); |
| 584 | assert_eq!(decision_for("t", "read_file", "{}"), Decision::Allow); |
| 585 | } |
| 586 | |
| 587 | #[test] |
| 588 | fn official_mcp_mutating_and_foreign_servers_stay_gated() { |
| 589 | for tool in [ |
| 590 | // official server, but not a query tool |
| 591 | "mcp__sigit__create_issue", |
| 592 | "mcp__sigit__merge_pull_request", |
| 593 | "mcp__sigit__delete_repository", |
| 594 | "mcp__sigit__", |
| 595 | // query-shaped names on other servers get no exemption |
| 596 | "mcp__other__list_issues", |
| 597 | "mcp__other__get_issue", |
| 598 | "mcp__github__search_code", |
| 599 | // `sigit` must match the whole server name |
| 600 | "mcp__sigitx__list_issues", |
| 601 | ] { |
| 602 | assert_eq!(classify(tool), ToolRisk::Mutating, "{tool}"); |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | #[test] |
| 607 | fn plan_mode_denies_mutating_and_spares_read_only() { |
| 608 | let _guard = env_guard(); |
| 609 | let session = "t-plan"; |
| 610 | reset_session(session); |
| 611 | set_plan_mode(session, true); |
| 612 | assert!(matches!( |
| 613 | run_command_decision(session, "ls"), |
| 614 | Decision::Deny(_) |
| 615 | )); |
| 616 | assert_eq!(decision_for(session, "read_file", "{}"), Decision::Allow); |
| 617 | set_plan_mode(session, false); |
| 618 | reset_session(session); |
| 619 | } |
| 620 | |
| 621 | #[test] |
| 622 | fn session_grant_short_circuits_ask() { |
| 623 | let _guard = env_guard(); |
| 624 | let session = "t-grant"; |
| 625 | reset_session(session); |
| 626 | let args = r#"{"path":"src/a.rs","old_text":"a","new_text":"b"}"#; |
| 627 | grant_for_session(session, "edit_file", args); |
| 628 | assert_eq!(decision_for(session, "edit_file", args), Decision::Allow); |
| 629 | // Non-run_command grants record the bare tool name: any path is covered. |
| 630 | assert_eq!( |
| 631 | decision_for(session, "edit_file", r#"{"path":"docs/other.md"}"#), |
| 632 | Decision::Allow |
| 633 | ); |
| 634 | // Other tools are unaffected by the grant. |
| 635 | assert_ne!( |
| 636 | decision_for(session, "delete_file", r#"{"path":"src/a.rs"}"#), |
| 637 | Decision::Allow |
| 638 | ); |
| 639 | reset_session(session); |
| 640 | assert_ne!(decision_for(session, "edit_file", args), Decision::Allow); |
| 641 | } |
| 642 | |
| 643 | #[test] |
| 644 | fn approval_preview_shows_short_arguments_in_full() { |
| 645 | let args = r#"{"command":"cargo test"}"#; |
| 646 | assert_eq!(approval_preview(args), args); |
| 647 | } |
| 648 | |
| 649 | #[test] |
| 650 | fn approval_preview_marks_truncation_explicitly() { |
| 651 | let args = format!(r#"{{"command":"echo {}; rm -rf /"}}"#, "x".repeat(600)); |
| 652 | let preview = approval_preview(&args); |
| 653 | assert!(preview.chars().count() < args.chars().count()); |
| 654 | assert!( |
| 655 | preview.contains("more chars]"), |
| 656 | "hidden content must be flagged, got: {preview}" |
| 657 | ); |
| 658 | } |
| 659 | |
| 660 | #[test] |
| 661 | fn plan_mode_outranks_session_grant() { |
| 662 | let _guard = env_guard(); |
| 663 | let session = "t-rank"; |
| 664 | reset_session(session); |
| 665 | let args = r#"{"path":"src/a.rs"}"#; |
| 666 | grant_for_session(session, "edit_file", args); |
| 667 | set_plan_mode(session, true); |
| 668 | assert!(matches!( |
| 669 | decision_for(session, "edit_file", args), |
| 670 | Decision::Deny(_) |
| 671 | )); |
| 672 | reset_session(session); |
| 673 | } |
| 674 | |
| 675 | #[test] |
| 676 | fn rules_gate_run_command_by_argument() { |
| 677 | let _guard = env_guard(); |
| 678 | let session = "t-rules"; |
| 679 | reset_session(session); |
| 680 | store_rules(&["run_command(git *)"], &["run_command(git push*)"]); |
| 681 | |
| 682 | assert_eq!(run_command_decision(session, "git status"), Decision::Allow); |
| 683 | assert!(matches!( |
| 684 | run_command_decision(session, "git push"), |
| 685 | Decision::Deny(_) |
| 686 | )); |
| 687 | assert!(matches!( |
| 688 | run_command_decision(session, "git push --force"), |
| 689 | Decision::Deny(_) |
| 690 | )); |
| 691 | assert_eq!( |
| 692 | run_command_decision(session, "cargo test"), |
| 693 | Decision::Ask, |
| 694 | "an unmatched command falls through to the default mode" |
| 695 | ); |
| 696 | reset_session(session); |
| 697 | } |
| 698 | |
| 699 | #[test] |
| 700 | fn deny_rule_beats_matching_allow_rule() { |
| 701 | let _guard = env_guard(); |
| 702 | let session = "t-deny-wins"; |
| 703 | reset_session(session); |
| 704 | store_rules(&["run_command(git *)"], &["run_command(git *)"]); |
| 705 | match run_command_decision(session, "git status") { |
| 706 | Decision::Deny(reason) => assert!( |
| 707 | reason.contains("run_command(git *)"), |
| 708 | "the denial must name the rule, got: {reason}" |
| 709 | ), |
| 710 | other => panic!("expected a deny, got {other:?}"), |
| 711 | } |
| 712 | reset_session(session); |
| 713 | } |
| 714 | |
| 715 | #[test] |
| 716 | fn rule_pattern_wildcard_and_prefix_edges() { |
| 717 | // Whole-token prefix: a pattern without a trailing `*` matches at a |
| 718 | // whitespace boundary or the end, never mid-token. |
| 719 | let rule = "run_command(git status)"; |
| 720 | assert!(rule_matches(rule, "run_command", Some("git status"))); |
| 721 | assert!(rule_matches( |
| 722 | rule, |
| 723 | "run_command", |
| 724 | Some("git status --short") |
| 725 | )); |
| 726 | assert!(!rule_matches(rule, "run_command", Some("git status-x"))); |
| 727 | assert!(!rule_matches(rule, "run_command", Some("git statu"))); |
| 728 | assert!(!rule_matches(rule, "run_command", Some("xgit status"))); |
| 729 | |
| 730 | // Trailing `*`: everything from the wildcard on matches. |
| 731 | let rule = "run_command(git push*)"; |
| 732 | assert!(rule_matches(rule, "run_command", Some("git push"))); |
| 733 | assert!(rule_matches(rule, "run_command", Some("git pushx"))); |
| 734 | assert!(rule_matches(rule, "run_command", Some("git push --force"))); |
| 735 | assert!(!rule_matches(rule, "run_command", Some("git pus"))); |
| 736 | let rule = "run_command(cargo *)"; |
| 737 | assert!(rule_matches( |
| 738 | rule, |
| 739 | "run_command", |
| 740 | Some("cargo test --locked") |
| 741 | )); |
| 742 | assert!( |
| 743 | rule_matches(rule, "run_command", Some("cargo build --bin src/x")), |
| 744 | "a trailing `*` also covers arguments containing `/`" |
| 745 | ); |
| 746 | assert!(!rule_matches(rule, "run_command", Some("cargo"))); |
| 747 | |
| 748 | // A rule only applies to its own tool. |
| 749 | assert!(!rule_matches(rule, "delete_file", Some("cargo test"))); |
| 750 | // Bare tool rules match any call, including argument-less tools. |
| 751 | assert!(rule_matches("run_command", "run_command", Some("anything"))); |
| 752 | assert!(rule_matches("mcp__srv__tool", "mcp__srv__tool", None)); |
| 753 | // A pattern rule never matches a tool without a matchable argument. |
| 754 | assert!(!rule_matches("mcp__srv__tool(x)", "mcp__srv__tool", None)); |
| 755 | } |
| 756 | |
| 757 | #[test] |
| 758 | fn run_command_session_grant_is_scoped_to_command_prefix() { |
| 759 | let _guard = env_guard(); |
| 760 | let session = "t-grant-scope"; |
| 761 | reset_session(session); |
| 762 | grant_for_session( |
| 763 | session, |
| 764 | "run_command", |
| 765 | r#"{"command":"git push origin main"}"#, |
| 766 | ); |
| 767 | // The grant is `run_command(git push)`: the `git push …` family only. |
| 768 | assert_eq!(run_command_decision(session, "git push"), Decision::Allow); |
| 769 | assert_eq!( |
| 770 | run_command_decision(session, "git push --force-with-lease"), |
| 771 | Decision::Allow |
| 772 | ); |
| 773 | assert_eq!(run_command_decision(session, "git pull"), Decision::Ask); |
| 774 | assert_eq!(run_command_decision(session, "git pushx"), Decision::Ask); |
| 775 | assert_eq!(run_command_decision(session, "rm -rf /"), Decision::Ask); |
| 776 | |
| 777 | // A single-token command grants that token's family. |
| 778 | grant_for_session(session, "run_command", r#"{"command":"ls"}"#); |
| 779 | assert_eq!(run_command_decision(session, "ls -la"), Decision::Allow); |
| 780 | assert_eq!(run_command_decision(session, "lsof"), Decision::Ask); |
| 781 | reset_session(session); |
| 782 | } |
| 783 | |
| 784 | #[test] |
| 785 | fn file_tool_rules_match_on_path() { |
| 786 | let _guard = env_guard(); |
| 787 | let session = "t-file-rules"; |
| 788 | reset_session(session); |
| 789 | store_rules(&["edit_file(src/*)"], &["delete_file(src/*)"]); |
| 790 | assert_eq!( |
| 791 | decision_for( |
| 792 | session, |
| 793 | "edit_file", |
| 794 | r#"{"path":"src/main.rs","old_text":"a","new_text":"b"}"# |
| 795 | ), |
| 796 | Decision::Allow |
| 797 | ); |
| 798 | assert_eq!( |
| 799 | decision_for(session, "edit_file", r#"{"path":"docs/readme.md"}"#), |
| 800 | Decision::Ask, |
| 801 | "a path outside the rule falls through to the default mode" |
| 802 | ); |
| 803 | assert!(matches!( |
| 804 | decision_for(session, "delete_file", r#"{"path":"src/main.rs"}"#), |
| 805 | Decision::Deny(_) |
| 806 | )); |
| 807 | assert_eq!( |
| 808 | decision_for(session, "delete_file", r#"{"path":"docs/readme.md"}"#), |
| 809 | Decision::Ask |
| 810 | ); |
| 811 | reset_session(session); |
| 812 | } |
| 813 | } |