formal-ai 0.299.0

//! Resolve a user turn into the concrete shell command the agentic loop should run.
//!
//! Split out of [`super::planner`] (issue #676): the agentic planner used to know
//! only the hardcoded `ls`, so `execute pwd` and every other seed shell token fell
//! through to the *unknown* fallback. The two data-driven strategies here — a named
//! command backed by `data/seed/terminal-commands.lino`, and a natural-language
//! directory-listing request — make the whole seed vocabulary reachable. Keeping the
//! matching prose (`PROSE_WORDS`, listing phrases) in one module also keeps the
//! planner file under the repository line budget.

use crate::seed::{self, ShellIntentArgument, ShellIntentVocabulary, TerminalCommandVocabulary};

const REPORT_ISSUE_ACTION: &str = "formal-ai:report-issue";

/// Resolve a user turn into the concrete shell command the agentic loop should run.
///
/// Two data-driven strategies, in order of specificity:
///
/// 1. **Named command** ([`named_shell_command`]): when the prompt pairs a run/execute
///    verb (or terminal/shell phrase) with a known shell token from
///    `data/seed/terminal-commands.lino` — e.g. *"execute pwd"*, *"run git status"*,
///    *«запусти ls»* — emit that command (with its flag/path/sub-command arguments).
///    This is what makes `pwd` (issue #676) and every other seed token reachable, not
///    just the hardcoded `ls` the fallback used to know.
///
/// 2. **Natural-language directory listing** ([`asks_for_directory_listing`]): when the
///    prompt asks, in prose, to see the files in the current place — e.g. *"give me a
///    list of files in current folder"*, *"what files are here?"* — resolve to `ls`.
///
/// The vocabulary lives in seed data, so a maintainer retunes coverage by editing a
/// `.lino` file rather than this function, upholding the project rule against hardcoded
/// natural language in the solver.
pub(super) fn shell_command_for_task(prompt: &str) -> Option<String> {
    let prompt = strip_balanced_outer_quotes(prompt.trim());
    let vocab = seed::terminal_command_vocabulary();
    let intent = intent_shell_command(prompt, &seed::shell_intent_vocabulary());
    let listing = asks_for_directory_listing(prompt);
    let web_search = crate::solver_handlers::web_search_query_for(prompt).is_some();
    let semantic = listing.then(|| String::from("ls")).or(intent);

    if let Some(command) = prefixed_shell_command(prompt, &vocab) {
        let first = command
            .split_whitespace()
            .next()
            .map(normalize_command_word);
        let names_known_command = first
            .as_ref()
            .is_some_and(|first| vocab.shell_tokens.iter().any(|token| token == first));
        let names_semantic_command = semantic.as_ref().is_some_and(|semantic| {
            semantic
                .split_whitespace()
                .next()
                .map(normalize_command_word)
                == first
        });
        if names_semantic_command {
            return semantic;
        }
        if names_known_command || (semantic.is_none() && !web_search) {
            return Some(command);
        }
    }

    semantic
        .or_else(|| named_shell_command(prompt, &vocab))
        .or_else(|| bare_shell_command(prompt, &vocab))
}

fn strip_balanced_outer_quotes(prompt: &str) -> &str {
    for quote in ['"', '\'', '`'] {
        if prompt.starts_with(quote) && prompt.ends_with(quote) && prompt.len() >= 2 {
            return &prompt[1..prompt.len() - 1];
        }
    }
    prompt
}

/// Pass an explicitly introduced command through byte-for-byte (apart from
/// surrounding whitespace). Explicit execution is an intent boundary: the
/// command need not appear in a maintained binary allowlist because the client
/// still owns its normal sandbox and permission decision.
fn prefixed_shell_command(prompt: &str, vocab: &TerminalCommandVocabulary) -> Option<String> {
    let prompt = prompt.trim();
    let lower = prompt.to_lowercase();
    if let Some(prefix) = vocab
        .passthrough_prefixes
        .iter()
        .filter(|prefix| prefix_boundary(&lower, prefix))
        .max_by_key(|prefix| prefix.chars().count())
    {
        let remainder = prompt.get(prefix.len()..)?.trim_start();
        let remainder = remainder
            .strip_prefix(':')
            .unwrap_or(remainder)
            .trim_start();
        return (!remainder.is_empty()).then(|| strip_balanced_outer_quotes(remainder).to_owned());
    }

    None
}

fn bare_shell_command(prompt: &str, vocab: &TerminalCommandVocabulary) -> Option<String> {
    let first = prompt.split_whitespace().next()?;
    let command = normalize_command_word(first);
    // A bare command token can also be an ordinary imperative (notably
    // `find information about ...`).  Explicit passthrough prefixes above are
    // unambiguous, but a seed-backed web-search request must remain available
    // to the dedicated search router instead of becoming a shell command.
    if crate::solver_handlers::web_search_query_for(prompt).is_some() {
        return None;
    }
    let is_known = vocab
        .bare_shell_tokens
        .iter()
        .any(|token| token == &command);
    let words = prompt.split_whitespace().collect::<Vec<_>>();
    (is_known && words[1..].iter().all(|word| !is_prose_word(word))).then(|| prompt.to_owned())
}

fn prefix_boundary(prompt: &str, prefix: &str) -> bool {
    prompt.starts_with(prefix)
        && prompt
            .get(prefix.len()..)
            .and_then(|rest| rest.chars().next())
            .is_some_and(|c| c.is_whitespace() || c == ':')
}

/// Recover the literal subject of a seed-backed source-code search request.
///
/// A client may advertise a dedicated grep/code-search tool instead of a shell.
/// The planner uses this semantic query before falling back to the `rg` lowering
/// returned by [`shell_command_for_task`].
pub(super) fn code_search_query_for_task(prompt: &str) -> Option<String> {
    let lower = prompt.to_lowercase();
    let vocab = seed::shell_intent_vocabulary();
    let intent_cues = vocab
        .intents
        .iter()
        .filter(|intent| intent.command == "rg")
        .flat_map(|intent| intent.cues.iter());
    let capability_registry = seed::agentic_tool_capabilities();
    let capability_cues = capability_registry
        .iter()
        .filter(|capability| capability.id == "grep")
        .flat_map(|capability| capability.cues.iter());
    let (_, cue) = intent_cues
        .chain(capability_cues)
        .filter(|cue| lower.contains(cue.as_str()))
        .map(|cue| (cue.chars().count(), cue))
        .max_by_key(|(length, _)| *length)?;
    local_search_query(prompt, cue, &vocab)
}

/// Resolve a semantic *intent* to its concrete command, backed by the seed
/// [`ShellIntentVocabulary`] (issue #680).
///
/// The two strategies above only fire when the prompt *names* the command (`run
/// pwd`) or asks, in prose, to list a directory. This third strategy handles the
/// common case where the user expresses an intent without naming the tool at all —
/// *"Print the current working directory"* → `pwd`, *"How much disk space is
/// free?"* → `df -h`, *"What is my username?"* → `whoami`. Each intent carries
/// multilingual cue phrases; the first intent whose cue is present in the prompt
/// wins (declaration order is most-specific-first), and its argument — if any — is
/// recovered from the prompt. An intent whose cue matches but whose required
/// argument is absent is skipped so the search continues rather than emitting an
/// argument-less command that would hang (`wc -l` on stdin).
fn intent_shell_command(prompt: &str, vocab: &ShellIntentVocabulary) -> Option<String> {
    let lower = prompt.to_lowercase();
    // Prefer the most specific matching cue across every intent. This prevents
    // a shorter generic cue (for example "current directory" → `pwd`) from
    // stealing a longer request ("list current directory" → `ls`).
    let (intent, cue) = vocab
        .intents
        .iter()
        .filter(|intent| intent.command != REPORT_ISSUE_ACTION)
        .flat_map(|intent| intent.cues.iter().map(move |cue| (intent, cue)))
        .filter(|(intent, cue)| {
            lower.contains(cue.as_str())
                && (intent.argument != ShellIntentArgument::SearchQuery
                    || vocab
                        .local_search_scopes
                        .iter()
                        .any(|scope| lower.contains(scope)))
        })
        .max_by_key(|(_, cue)| cue.chars().count())?;
    match intent.argument {
        ShellIntentArgument::None => resolve_shell_command(&intent.command, vocab),
        ShellIntentArgument::Path => {
            path_argument(prompt).map(|arg| format!("{} {arg}", intent.command))
        }
        ShellIntentArgument::NameLead => name_lead_argument(prompt, &vocab.name_leads)
            .map(|arg| format!("{} {arg}", intent.command)),
        ShellIntentArgument::OnePath => {
            path_arguments(prompt, cue, vocab, 1).map(|args| format!("{} {args}", intent.command))
        }
        ShellIntentArgument::TwoPaths => {
            path_arguments(prompt, cue, vocab, 2).map(|args| format!("{} {args}", intent.command))
        }
        ShellIntentArgument::Remainder => remainder_argument(prompt, &lower, cue)
            .map(|arg| format!("{} --fixed-strings -- '{arg}' .", intent.command)),
        ShellIntentArgument::SearchQuery => local_search_query(prompt, cue, vocab)
            .map(|arg| format!("{} --fixed-strings -- '{arg}' .", intent.command)),
    }
}

fn path_arguments(
    prompt: &str,
    cue: &str,
    vocab: &ShellIntentVocabulary,
    count: usize,
) -> Option<String> {
    let lower = prompt.to_lowercase();
    let after_cue = lower
        .find(cue)
        .and_then(|start| prompt.get(start + cue.len()..))
        .unwrap_or_default();
    let mut arguments = collect_path_arguments(after_cue, "", vocab, count);
    if arguments.len() < count {
        arguments = collect_path_arguments(prompt, cue, vocab, count);
    }
    (arguments.len() == count).then(|| arguments.join(" "))
}

fn collect_path_arguments(
    text: &str,
    cue: &str,
    vocab: &ShellIntentVocabulary,
    count: usize,
) -> Vec<String> {
    let cue = cue.to_lowercase();
    let mut arguments = Vec::new();
    for word in text.split_whitespace() {
        let candidate = word
            .trim_matches(|c: char| matches!(c, '`' | '"' | '\'' | ',' | ';' | ':' | '!' | '?'));
        let normalized = candidate.to_lowercase();
        if candidate.is_empty()
            || !is_safe_path(candidate)
            || cue.split_whitespace().any(|part| part == normalized)
            || vocab
                .argument_noise
                .iter()
                .any(|noise| noise == &normalized)
        {
            continue;
        }
        arguments.push(candidate.to_owned());
        if arguments.len() == count {
            break;
        }
    }
    arguments
}

fn local_search_query(prompt: &str, cue: &str, vocab: &ShellIntentVocabulary) -> Option<String> {
    let cue_words: Vec<String> = cue.split_whitespace().map(str::to_lowercase).collect();
    let scope_words: Vec<String> = vocab
        .local_search_scopes
        .iter()
        .flat_map(|scope| scope.split_whitespace().map(str::to_lowercase))
        .collect();
    let query = prompt
        .split_whitespace()
        .map(|word| {
            word.trim_matches(|c: char| !c.is_alphanumeric() && !matches!(c, '_' | '-' | '.' | ':'))
        })
        .filter(|word| !word.is_empty())
        .filter(|word| {
            let lower = word.to_lowercase();
            !cue_words.contains(&lower)
                && !scope_words.contains(&lower)
                && !vocab.argument_noise.contains(&lower)
        })
        .collect::<Vec<_>>()
        .join(" ");
    (!query.is_empty()
        && query.chars().all(|c| {
            c.is_alphanumeric() || c.is_whitespace() || matches!(c, '_' | '-' | '.' | ':')
        }))
    .then_some(query)
}

fn resolve_shell_command(command: &str, vocab: &ShellIntentVocabulary) -> Option<String> {
    let action = command.strip_prefix("formal-ai:workspace-");
    if let Some(action) = action {
        return workspace_command(action, vocab).map(str::to_owned);
    }
    Some(command.to_owned())
}

fn workspace_command<'a>(action: &str, vocab: &'a ShellIntentVocabulary) -> Option<&'a str> {
    vocab
        .workspace_commands
        .iter()
        .find(|commands| std::path::Path::new(&commands.marker).is_file())
        .and_then(|commands| match action {
            "test" => Some(commands.test.as_str()),
            "install" => Some(commands.install.as_str()),
            "build" => Some(commands.build.as_str()),
            _ => None,
        })
}

/// Recover a safe literal query following a matched semantic cue.
fn remainder_argument(prompt: &str, lower: &str, cue: &str) -> Option<String> {
    let start = lower.find(cue)? + cue.len();
    let remainder = prompt.get(start..)?.trim();
    (!remainder.is_empty()
        && remainder.chars().all(|c| {
            c.is_alphanumeric() || c.is_whitespace() || matches!(c, '_' | '-' | '.' | ':')
        }))
    .then(|| remainder.to_owned())
}

/// The first filename-looking token in the prompt: a token carrying an interior dot
/// (`Cargo.toml`, `src/lib.rs`) that is a safe relative path, not a URL or flag.
/// Used to fill a [`ShellIntentArgument::Path`] argument (`wc -l Cargo.toml`).
fn path_argument(prompt: &str) -> Option<String> {
    prompt
        .split(|c: char| c.is_whitespace())
        .map(|word| word.trim_matches(|c: char| matches!(c, '`' | '"' | '\'' | ',' | ';' | '?')))
        .map(|word| word.trim_end_matches(['.', '!']))
        .find(|token| {
            let interior_dot = token.trim_matches('.').contains('.');
            interior_dot && !token.contains("://") && is_safe_path(token)
        })
        .map(str::to_owned)
}

/// The name introduced by a name-lead cue (*called*/*named*/…): the token following
/// the first name-lead word. Used to fill a [`ShellIntentArgument::NameLead`]
/// argument (`mkdir build` from *"create a directory called build"*).
fn name_lead_argument(prompt: &str, name_leads: &[String]) -> Option<String> {
    let words: Vec<&str> = prompt
        .split(|c: char| c.is_whitespace())
        .filter(|w| !w.is_empty())
        .collect();
    let lead_index = words.iter().position(|word| {
        let normalized = word
            .trim_matches(|c: char| !c.is_alphanumeric())
            .to_lowercase();
        name_leads.iter().any(|lead| lead == &normalized)
    })?;
    let name = words
        .get(lead_index + 1)?
        .trim_matches(|c: char| matches!(c, '`' | '"' | '\'' | ',' | ';' | '.' | '!' | '?' | ':'));
    (!name.is_empty() && is_safe_path(name)).then(|| name.to_owned())
}

/// Whether a token is a safe relative path/name: no absolute or `..` escape, no
/// leading dash, only path-safe characters.
fn is_safe_path(token: &str) -> bool {
    !token.starts_with('/')
        && !token.starts_with('-')
        && !token.split('/').any(|part| part == ".." || part.is_empty())
        && token
            .chars()
            .all(|c| c.is_alphanumeric() || matches!(c, '/' | '.' | '_' | '-'))
}

/// Extract an explicit shell command named in the prompt, backed by the seed
/// [`TerminalCommandVocabulary`].
///
/// A prompt names a command when it carries *run context* — a run/execute verb, a
/// Chinese run verb, or a terminal/shell phrase — together with a recognized shell
/// token (`pwd`, `git`, `cargo`, …). Two shapes are handled:
///
/// * **Verb-adjacent** (`execute pwd`, `run git status`): the shell token immediately
///   follows a run verb, so the token *and its trailing arguments* are the command —
///   arguments run until a natural-language word ([`is_prose_word`]) ends them, so
///   `run git status` → `git status` but `run ls then stop` → `ls`.
/// * **Mentioned** (`Run the ls command to list files`): the token appears with run
///   context but is not directly after the verb, so only the single token is emitted
///   (`ls`) — the surrounding words are prose describing the request, not arguments.
fn named_shell_command(prompt: &str, vocab: &TerminalCommandVocabulary) -> Option<String> {
    let lower = prompt.to_ascii_lowercase();
    let has_phrase = vocab.terminal_phrases.iter().any(|p| lower.contains(p));
    let has_cjk_verb = vocab.cjk_run_verbs.iter().any(|v| lower.contains(v));

    // Word tokens of the original prompt, preserving case so command arguments
    // (paths, flags, filenames) survive intact.
    let words: Vec<&str> = prompt
        .split(|c: char| c.is_whitespace())
        .filter(|w| !w.is_empty())
        .collect();

    let is_run_verb = |word: &str| {
        let normalized = normalize_command_word(word);
        vocab.run_verbs.iter().any(|v| v == &normalized)
    };
    let is_shell_token = |word: &str| {
        let normalized = normalize_command_word(word);
        !normalized.is_empty() && vocab.shell_tokens.iter().any(|t| t == &normalized)
    };
    let has_verb = words.iter().any(|w| is_run_verb(w)) || has_cjk_verb;

    // Shape 1: a shell token directly after a run verb — the command plus its arguments.
    for (index, word) in words.iter().enumerate() {
        if index == 0 || !is_shell_token(word) {
            continue;
        }
        if is_run_verb(words[index - 1]) {
            return Some(collect_command(&words[index..]));
        }
    }

    // Shape 2: a shell token mentioned anywhere, given run context — the token alone.
    if has_verb || has_phrase {
        if let Some(word) = words.iter().find(|w| is_shell_token(w)) {
            return Some(normalize_command_word(word));
        }
    }

    None
}

/// Assemble a command from a token slice that starts at the command word: keep the
/// command and every following argument until a natural-language word ends it.
fn collect_command(words: &[&str]) -> String {
    let mut parts = vec![normalize_command_word(words[0])];
    for word in &words[1..] {
        if is_prose_word(word) {
            break;
        }
        let trimmed = word.trim_matches(|c: char| c == '`' || c == ',' || c == '.');
        if trimmed.is_empty() {
            break;
        }
        parts.push(trimmed.to_owned());
    }
    parts.join(" ")
}

/// Normalize a raw prompt word to a bare command token: lowercase, keeping only the
/// leading run of command-name characters (so ``` `pwd` ```, `pwd.` and `pwd,` all
/// normalize to `pwd`).
fn normalize_command_word(word: &str) -> String {
    word.trim_matches('`')
        .chars()
        .take_while(|c| c.is_ascii_alphanumeric() || *c == '_' || *c == '-')
        .collect::<String>()
        .to_ascii_lowercase()
}

/// Whether a word is natural-language prose rather than a command argument. Used to
/// stop argument collection at the boundary between a command and the sentence around
/// it (e.g. `git status` stops before `in the current directory`).
fn is_prose_word(word: &str) -> bool {
    const PROSE_WORDS: &[&str] = &[
        "command",
        "commands",
        "to",
        "in",
        "into",
        "on",
        "the",
        "a",
        "an",
        "and",
        "then",
        "please",
        "for",
        "of",
        "that",
        "which",
        "so",
        "this",
        "these",
        "those",
        "here",
        "there",
        "me",
        "us",
        "you",
        "it",
        "from",
        "at",
        "with",
        "will",
        "would",
        "can",
        "could",
        "should",
        "using",
        "via",
        "inside",
        "within",
        "output",
        "result",
        "results",
        "contents",
        "content",
        "directory",
        "folder",
        "folders",
        "file",
        "files",
        "currently",
        "again",
        "also",
        "just",
        "now",
    ];
    let normalized = word
        .trim_matches(|c: char| !c.is_ascii_alphanumeric())
        .to_ascii_lowercase();
    PROSE_WORDS.contains(&normalized.as_str())
}

/// Whether `prompt` asks, in natural language, to list the files in the current place.
///
/// Resolves prose directory-listing requests to `ls`: a listing phrase (or a
/// which-files question) scoped to the current/working/this directory or folder.
fn asks_for_directory_listing(prompt: &str) -> bool {
    let lower = prompt.to_ascii_lowercase();

    let asks_for_listing = contains_any(
        &lower,
        &[
            "list files",
            "list the files",
            "list all files",
            "list local files",
            "list of files",
            "list of all files",
            "list directory",
            "list the directory",
            "list the contents",
            "listing of files",
            "directory listing",
            "directory contents",
            "folder contents",
            "contents of this directory",
            "contents of the current directory",
            "contents of this folder",
            "contents of the current folder",
            "show files",
            "show the files",
            "show me the files",
            "see the files",
        ],
    );
    let asks_which_files = contains_any(
        &lower,
        &[
            "what files",
            "which files",
            "files are in",
            "files in the",
            "files in this",
            "files in current",
            "files in the current",
            "files exist",
            "files are here",
            "files are there",
        ],
    );
    let local_scope = contains_any(
        &lower,
        &[
            "here",
            "current directory",
            "working directory",
            "current working directory",
            "this directory",
            "the directory",
            "current folder",
            "this folder",
            "the folder",
            "local files",
        ],
    );

    (asks_for_listing || asks_which_files) && local_scope
}

fn contains_any(text: &str, phrases: &[&str]) -> bool {
    phrases.iter().any(|phrase| text.contains(phrase))
}