lean-ctx-engine 3.12.2

pub(super) fn effective_allowlist() -> Vec<String> {
    filter_valid_entries(raw_allowlist_entries())
}

/// The same union `effective_allowlist()` builds, before `*`-placement
/// filtering — shared with the invalid-entry and upgrade-advisory doctor
/// checks, which need to see entries `effective_allowlist()` would drop.
fn raw_allowlist_entries() -> Vec<String> {
    use crate::core::config::shell_policy;
    // LEAN_CTX_SHELL_ALLOWLIST_OVERRIDE completely replaces the config (test
    // and debug builds only; see `config::shell_policy`).
    if let Some(ov) = shell_policy::env("LEAN_CTX_SHELL_ALLOWLIST_OVERRIDE") {
        return ov
            .split(',')
            .map(|s| s.trim().to_string())
            .filter(|s| !s.is_empty())
            .collect();
    }
    let cfg = shell_policy::config();
    let mut list = cfg.shell_allowlist.clone();
    // `shell_allowlist_extra` is purely additive (written by `lean-ctx allow <cmd>`),
    // so users can permit a command without nuking the built-in defaults. It only
    // matters in restricted mode — when the base list is empty all commands pass anyway.
    if !list.is_empty() {
        for entry in &cfg.shell_allowlist_extra {
            if !entry.is_empty() && !list.contains(entry) {
                list.push(entry.clone());
            }
        }
    }
    if let Some(env_val) = shell_policy::env("LEAN_CTX_SHELL_ALLOWLIST") {
        for entry in env_val
            .split(',')
            .map(|s| s.trim().to_string())
            .filter(|s| !s.is_empty())
        {
            if !list.contains(&entry) {
                list.push(entry);
            }
        }
    }
    list
}

/// Public accessor: entries the effective allowlist silently dropped for
/// invalid `*` placement (§4 of the design doc). Used by `lean-ctx doctor`
/// to surface a count, and by `--fix` to know which raw config lines to
/// comment out.
#[must_use]
pub fn invalid_allowlist_entries_pub() -> Vec<String> {
    raw_allowlist_entries()
        .into_iter()
        .filter(|e| !is_valid_allowlist_entry(e))
        .collect()
}

/// Public accessor: multi-word entries with no trailing `*` — their meaning
/// changed under subcommand scoping (now exact-match only, previously
/// "whole binary"). Used by the `lean-ctx doctor` upgrade advisory (§2 of
/// the design doc). Always empty when `shell_allowlist_subcommand_scoping =
/// false`, since those entries' meaning didn't change under the compat path.
#[must_use]
pub fn narrowed_allowlist_entries_pub() -> Vec<String> {
    if !crate::core::config::shell_policy::config().shell_allowlist_subcommand_scoping {
        return Vec::new();
    }
    raw_allowlist_entries()
        .into_iter()
        .filter(|e| is_valid_allowlist_entry(e))
        .filter(|e| {
            let words: Vec<&str> = e.split_whitespace().collect();
            words.len() > 1 && words.last() != Some(&"*")
        })
        .collect()
}

/// §2 "Shadowing warnings": is `a` made useless by a different entry already
/// in `entries`? Entry A is shadowed by B when B matches every command A
/// matches: B is a single word equal to A's first word, or B ends in `*` and
/// B's words before the `*` are a leading run of A's words (ignoring A's own
/// trailing `*`). Returns the first covering entry, in list order. Shared by
/// `lean-ctx allow` (write-time warning) and `lean-ctx doctor` (read-only
/// advisory) so the rule can't drift between the two.
#[must_use]
pub fn shadowing_entry(entries: &[String], a: &str) -> Option<String> {
    let a_words: Vec<&str> = a.split_whitespace().collect();
    let a_prefix: &[&str] = if a_words.last() == Some(&"*") {
        &a_words[..a_words.len() - 1]
    } else {
        &a_words
    };
    entries
        .iter()
        .find(|b| {
            if b.as_str() == a {
                return false;
            }
            let b_words: Vec<&str> = b.split_whitespace().collect();
            if b_words.len() == 1 {
                b_words.first() == a_words.first()
            } else if b_words.last() == Some(&"*") {
                let b_prefix = &b_words[..b_words.len() - 1];
                b_prefix.len() <= a_prefix.len() && a_prefix.starts_with(b_prefix)
            } else {
                false
            }
        })
        .cloned()
}

/// All shadowed entries in `entries`, paired with the entry that covers each
/// one — used by the `lean-ctx doctor` advisory (§2).
#[must_use]
pub fn shadowed_allowlist_entries_pub(entries: &[String]) -> Vec<(String, String)> {
    entries
        .iter()
        .filter_map(|a| shadowing_entry(entries, a).map(|b| (a.clone(), b)))
        .collect()
}

/// §4 of the design doc: an entry's only valid `*` placement is a single
/// trailing token. Bare `*`, an embedded `*`, or more than one `*` are
/// invalid.
pub fn is_valid_allowlist_entry(entry: &str) -> bool {
    let words: Vec<&str> = entry.split_whitespace().collect();
    match words.iter().filter(|w| **w == "*").count() {
        0 => true,
        1 => words.len() > 1 && words.last() == Some(&"*"),
        _ => false,
    }
}

/// Drops invalid entries rather than failing closed on the whole allowlist
/// (§4): removing one allow-entry can only make matching *more*
/// restrictive, never less, so this is safe by construction — mirrors the
/// `ShellSecurity::resolve()` precedent (an unknown value falls back to
/// enforce, never to open).
fn filter_valid_entries(entries: Vec<String>) -> Vec<String> {
    entries
        .into_iter()
        .filter(|entry| {
            let valid = is_valid_allowlist_entry(entry);
            if !valid {
                tracing::warn!(
                    "[shell_allowlist] dropping invalid allowlist entry '{entry}': `*` must be \
                     the single trailing token, not embedded, bare, or repeated"
                );
            }
            valid
        })
        .collect()
}

/// Could this token plausibly be an executable name?
///
/// Deliberately permissive — the point is to catch obvious scanner debris
/// (unbalanced parens, quotes, `$`, `&`, embedded spaces), not to validate
/// against the filesystem. A name that would need quoting to type is not a name
/// the user should be told to allowlist.
pub(super) fn is_plausible_command_name(base: &str) -> bool {
    if base.is_empty() {
        return false;
    }
    !base.chars().any(|c| {
        c.is_whitespace() || matches!(c, '(' | ')' | '"' | '\'' | '$' | '&' | '|' | ';' | '`')
    })
}

/// Builds the actionable, self-diagnosing message shown when a command's base binary
/// is not in the allowlist. Unlike a bare "not allowed" string, it tells the user
/// (1) the exact additive fix, (2) the real config path the MCP server reads, and
/// (3) — crucially — whether their `config.toml` silently failed to parse (in which
/// case lean-ctx is on defaults, which is the usual reason an allowlist edit "did
/// nothing"). That last signal is otherwise invisible over an MCP/stdio transport.
pub(super) fn allowlist_block_message(base: &str, tokens: &[String], entries: &[String]) -> String {
    let cfg_path = crate::core::config::shell_policy::config_path().map_or_else(
        || "~/.lean-ctx/config.toml".to_string(),
        |p| p.display().to_string(),
    );

    // #1814: a `$var` command word is a *correct* split, not a mis-split — the
    // variable genuinely is the command, and the allowlist cannot see what it
    // expands to at run time. Sending that caller to file a mis-split bug wastes
    // their time and ours, and "quote the fragment differently" cannot help:
    // no quoting makes a variable command allowable. Say what actually works.
    if let Some(name) = base.strip_prefix('$') {
        let hint = name.trim_start_matches('{').trim_end_matches('}');
        return format!(
            "[BLOCKED — DO NOT RETRY] '{base}' expands to the command name at run time, so \
             the allowlist cannot see which command would run.\n\
             Write the command literally instead (e.g. `echo hi`, `.venv/bin/python -c ...`) — \
             that is the form the allowlist can gate.\n\
             Do NOT run `lean-ctx allow {hint}`; it would allowlist the variable's name, not \
             the command.\n\
             Config in effect: {cfg_path}"
        );
    }

    // A base that cannot be a command name means the scanner mis-split the line,
    // not that the user needs to allow something. Printing
    // `lean-ctx allow print(urllib.parse.quote(sys.argv[1],safe=))` — a real
    // example from GH #1646 — invites the reader to copy a command that cannot
    // work and hides the actual fault. Say which it is.
    if !is_plausible_command_name(base) {
        return format!(
            "[BLOCKED] '{base}' is not in the shell allowlist — but it does not look like a \
             command name, which means lean-ctx split your command line wrongly rather than \
             finding a command to gate.\n\
             Do NOT run `lean-ctx allow` on it; that would allowlist a fragment.\n\
             Quote the fragment differently if you can, and please report the original command \
             at https://github.com/yvgude/lean-ctx/issues — a mis-split is a bug in lean-ctx.\n\
             Config in effect: {cfg_path}"
        );
    }

    // A nested shell (`bash -c '…'`, `sh script.sh`) is the most common block
    // agents retry: the command already runs in a shell, so the wrapper is
    // redundant — and allowlisting the shell would let any script through.
    if matches!(base, "sh" | "bash" | "zsh" | "dash" | "ksh") {
        return format!(
            "[BLOCKED — DO NOT RETRY] A nested '{base}' is not run: the command already \
             executes in a shell, and a nested shell would hide what it runs from the \
             allowlist.\n\
             Do this instead: drop the `{base} -c` wrapper and pass the inner command \
             directly (pipes, `&&`, loops and functions all work); for a script file, run \
             the commands it contains, or allow that one script deliberately.\n\
             Do NOT run `lean-ctx allow {base}`: it would admit every script, not this \
             command.\n\
             Config in effect: {cfg_path}"
        );
    }

    let mut msg = format!(
        "[BLOCKED — DO NOT RETRY] '{base}' is not in the shell allowlist. \
         This is a permanent restriction, not a transient error.\n\
         Fix (additive, keeps the defaults): run  lean-ctx allow {base}\n\
         Config in effect: {cfg_path}\n\
         Or disable the allowlist entirely: set  shell_allowlist = []\n\
         Or turn off all shell gating (you own the risk): set  shell_security = \"off\"  \
         in that file — compression still applies.\n\
         Do NOT reroute through ctx_execute(language=\"shell\"): both tools enforce the same \
         policy. Allow the command explicitly or change shell_security deliberately."
    );

    if let Some(detail) = closest_entry_detail(tokens, entries) {
        msg.push_str(&detail);
    }

    if crate::core::config::cloud_infra_commands().contains(&base) {
        msg.push_str(
            "\nNote: cloud/infra CLIs (terraform, kubectl, aws, …) are deliberately \
             excluded from the defaults — they mutate remote infrastructure with \
             ambient credentials. Opting in is a deliberate user decision.",
        );
    }

    // The notes below describe `Config::load`'s file; a release build's gate
    // reads only the file named above (`config::shell_policy`).
    let ambient = !crate::core::config::shell_policy::pinned();
    if ambient && let Some(parse_err) = crate::core::config::last_config_parse_error() {
        msg.push_str(&format!(
            "\n\n⚠ Your config.toml currently FAILS to parse, so lean-ctx is running on the \
             built-in defaults — this is almost certainly why editing the allowlist had no \
             effect. Fix the TOML error below, then retry:\n  {parse_err}\n  File: {cfg_path}"
        ));
    } else if ambient && let Some(missing) = crate::core::config::Config::missing_config_path() {
        // The resolved config doesn't exist → lean-ctx is on defaults. An edit
        // made to a config.toml in a different dir (XDG vs legacy ~/.lean-ctx) or
        // under a sandboxed/container HOME is never read — say so over MCP (#540).
        msg.push_str(&format!(
            "\n\n⚠ No config file exists at {} — lean-ctx is running on built-in defaults. \
             If you added the command to a config.toml in a DIFFERENT location (XDG \
             ~/.config/lean-ctx vs legacy ~/.lean-ctx, or your MCP client launches lean-ctx \
             in a sandbox/container with a different HOME), the runtime never reads it. \
             `lean-ctx doctor` prints the path actually in effect; pin it with \
             LEAN_CTX_CONFIG_DIR.",
            missing.display()
        ));
    }

    // A project-local `shell_allowlist`/`shell_allowlist_extra` is silently
    // withheld for an untrusted workspace; surface that here so the edit's
    // no-op reason isn't buried in an MCP-invisible stderr warning (#540).
    if let Some(notice) = crate::core::workspace_trust::untrusted_override_notice() {
        msg.push_str("\n\n⚠ ");
        msg.push_str(&notice);
    }

    msg
}

/// §6 of the design doc: names the closest matching allowlist entry and
/// where the command's tokens diverge from it, so an LLM agent can
/// self-correct without a human round-trip. Only fires when at least one
/// entry shares the base binary — otherwise "not in the shell allowlist"
/// already says enough. Longest common token prefix; ties resolve to the
/// first entry in list order (same convention as `levenshtein::closest()`).
fn closest_entry_detail(tokens: &[String], entries: &[String]) -> Option<String> {
    let base = tokens.first()?.as_str();
    let mut best: Option<(&str, usize)> = None;
    for entry in entries {
        let words: Vec<&str> = entry.split_whitespace().collect();
        if words.first() != Some(&base) {
            continue;
        }
        let shared = tokens
            .iter()
            .zip(words.iter())
            .take_while(|(t, w)| t.as_str() == **w)
            .count();
        if best.is_none_or(|(_, len)| shared > len) {
            best = Some((entry.as_str(), shared));
        }
    }
    let (entry, shared) = best?;
    let entry_words: Vec<&str> = entry.split_whitespace().collect();
    let diverges_at = entry_words.get(shared).copied();
    let got = tokens.get(shared).map(String::as_str);
    let mut detail = match (diverges_at, got) {
        (Some(expected), Some(got)) => {
            format!("\n(closest entry '{entry}' requires subcommand '{expected}', got '{got}')")
        }
        _ => format!("\n(closest entry '{entry}')"),
    };
    if got.is_some_and(|g| g.starts_with('-')) {
        detail.push_str(
            "\nGlobal options before the subcommand are not matched. Run \
             `cd <dir> && <cmd>` instead, or add an explicit allowlist entry.",
        );
    }
    Some(detail)
}
/// Public accessor: the fully-resolved allowlist actually enforced by the MCP tools
/// (base `shell_allowlist` + additive `shell_allowlist_extra` + env), deduplicated.
/// Empty means blocklist-only mode (all commands pass). Used by `lean-ctx allow`
/// and `lean-ctx doctor` to show users exactly what the runtime sees.
#[must_use]
pub fn effective_allowlist_pub() -> Vec<String> {
    effective_allowlist()
}

/// GH #1466: Check whether the user has explicitly allowed an interpreter for
/// inline/heredoc use in Claude Code's own permission system.
///
/// Reads `permissions.allow` from both global (`~/.claude/settings.json`) and
/// project-local (`.claude/settings.local.json`) settings. Entries like
/// `Bash(python3:*)` or `Bash(python3:)` grant the interpreter `python3`
/// inline-execution rights, bypassing the heredoc/eval-flag block.
///
/// Cached per-process to avoid re-reading settings on every shell invocation.
pub(super) fn claude_allows_interpreter_inline(interpreter: &str) -> bool {
    use std::sync::OnceLock;

    static CACHE: OnceLock<Vec<String>> = OnceLock::new();

    let allowed = CACHE.get_or_init(|| {
        let mut interpreters = Vec::new();
        collect_claude_bash_permissions(&mut interpreters);
        interpreters
    });

    allowed.iter().any(|a| a == interpreter)
}

/// Parse `Bash(<cmd>:...)` entries from Claude's `permissions.allow` arrays.
fn collect_claude_bash_permissions(out: &mut Vec<String>) {
    let paths = claude_settings_paths();
    for path in &paths {
        if let Ok(content) = std::fs::read_to_string(path) {
            if let Ok(json) = crate::core::jsonc::parse_jsonc(&content) {
                extract_bash_interpreters(&json, out);
            }
        }
    }
}

/// Returns candidate Claude settings paths (global + project-local).
fn claude_settings_paths() -> Vec<std::path::PathBuf> {
    let mut paths = Vec::with_capacity(2);

    if let Some(home) = crate::core::config::shell_policy::settings_home() {
        paths.push(home.join(".claude").join("settings.json"));
    }

    if let Ok(cwd) = std::env::current_dir() {
        paths.push(cwd.join(".claude").join("settings.local.json"));
    }

    paths
}

/// Extract interpreter names from `Bash(<interpreter>:...)` permission entries.
pub(super) fn extract_bash_interpreters(json: &serde_json::Value, out: &mut Vec<String>) {
    let allow = json
        .pointer("/permissions/allow")
        .and_then(|v| v.as_array());

    let Some(arr) = allow else { return };

    for entry in arr {
        let Some(s) = entry.as_str() else { continue };
        if let Some(inner) = parse_bash_permission(s) {
            if !inner.is_empty() && !out.contains(&inner) {
                out.push(inner);
            }
        }
    }
}

/// Parse a Claude permission entry like `Bash(python3:*)` → `Some("python3")`.
/// Accepted formats: `Bash(cmd:)`, `Bash(cmd:*)`, `Bash(cmd:<anything>)`.
pub(super) fn parse_bash_permission(entry: &str) -> Option<String> {
    let rest = entry.strip_prefix("Bash(")?;
    let rest = rest.strip_suffix(')')?;
    let colon_pos = rest.find(':')?;
    let cmd = &rest[..colon_pos];
    if cmd.is_empty() {
        return None;
    }
    let base = cmd.rsplit('/').next().unwrap_or(cmd);
    Some(base.to_string())
}

#[cfg(test)]
mod tests {
    use super::{allowlist_block_message, shadowing_entry};

    /// §2 of the design doc: pure-helper shadowing rule, all five agreed cases.
    #[test]
    fn shadowing_rule_matches_design_doc_cases() {
        let entries = |a: &str, b: &str| vec![a.to_string(), b.to_string()];

        assert_eq!(
            shadowing_entry(&entries("git status", "git"), "git status"),
            Some("git".to_string()),
            "a single-word entry shadows any multi-word entry sharing its first word"
        );
        assert_eq!(
            shadowing_entry(&entries("git status -s", "git status *"), "git status -s"),
            Some("git status *".to_string()),
            "a trailing-* entry shadows a longer exact entry sharing its prefix"
        );
        assert_eq!(
            shadowing_entry(&entries("git status", "git status *"), "git status"),
            Some("git status *".to_string()),
            "a trailing-* entry shadows the identical exact entry"
        );
        assert_eq!(
            shadowing_entry(&entries("git log", "git status *"), "git log"),
            None,
            "different subcommands must not shadow each other"
        );
        assert_eq!(
            shadowing_entry(&entries("git status *", "git status"), "git status *"),
            None,
            "an exact entry must not shadow a broader trailing-* entry"
        );
    }

    /// #1814: a `$var` command word is a *correct* split, not a mis-split. The
    /// old text asked the caller to file a bug for intended behaviour and
    /// suggested re-quoting, which cannot help — no quoting makes a variable
    /// command gateable.
    #[test]
    fn variable_command_word_explains_expansion_not_a_parser_bug() {
        for base in ["$X", "$PY", "${CMD}"] {
            let msg = allowlist_block_message(base, &[], &[]);
            assert!(
                msg.contains("expands to the command name at run time"),
                "{base}: must explain expansion, got: {msg}"
            );
            assert!(
                !msg.contains("mis-split"),
                "{base}: must not blame the parser, got: {msg}"
            );
            assert!(
                !msg.contains("github.com/yvgude/lean-ctx/issues"),
                "{base}: must not request a bug report for intended behaviour"
            );
        }
    }

    /// The genuine mis-split guidance must survive for tokens that really are
    /// not command names — that is what #1646 added it for.
    #[test]
    fn genuine_mis_splits_still_ask_for_a_report() {
        let msg = allowlist_block_message("print(urllib.parse.quote(sys.argv[1],safe=))", &[], &[]);
        assert!(msg.contains("mis-split"), "got: {msg}");
        assert!(msg.contains("github.com/yvgude/lean-ctx/issues"));
    }

    /// An ordinary unknown command keeps the actionable `lean-ctx allow` path.
    #[test]
    fn ordinary_command_still_offers_allow() {
        let msg = allowlist_block_message("ncat", &[], &[]);
        assert!(msg.contains("lean-ctx allow ncat"), "got: {msg}");
        assert!(!msg.contains("expands to the command name"));
    }

    /// A nested shell is redundant, and allowing it would admit every script:
    /// the message says to drop the wrapper and never to allow the shell.
    #[test]
    fn nested_shell_says_drop_the_wrapper_not_allow_it() {
        for shell in ["bash", "sh", "zsh"] {
            let msg = allowlist_block_message(shell, &[], &[]);
            assert!(
                msg.contains(&format!("drop the `{shell} -c` wrapper")),
                "got: {msg}"
            );
            assert!(
                msg.contains(&format!("Do NOT run `lean-ctx allow {shell}`")),
                "got: {msg}"
            );
            assert!(!msg.contains("Fix (additive"), "got: {msg}");
        }
    }
}