aidaemon 0.11.1

use crate::llm_markers::INTENT_GATE_MARKER;

#[cfg(test)]
fn is_pseudo_tool_line(line: &str) -> bool {
    let lower = line.trim().to_ascii_lowercase();
    lower.starts_with("[tool_use:")
        || lower.starts_with("[tool_call:")
        || lower.starts_with("[function_call:")
        || lower.starts_with("[functioncall:")
}

#[cfg(test)]
fn is_tool_name_like(name: &str) -> bool {
    if name.is_empty() {
        return false;
    }
    let lower = name.to_ascii_lowercase();
    matches!(
        lower.as_str(),
        "terminal"
            | "browser"
            | "web_search"
            | "web_fetch"
            | "system_info"
            | "remember_fact"
            | "manage_config"
            | "send_file"
            | "spawn_agent"
            | "cli_agent"
            | "manage_cli_agents"
            | "health_probe"
            | "manage_skills"
            | "use_skill"
            | "skill_resources"
            | "manage_people"
            | "manage_api"
            | "http_request"
            | "manage_http_auth"
            | "manage_oauth"
            | "read_channel_history"
    ) || lower.starts_with("mcp__")
        || lower.contains("__")
}

#[cfg(test)]
fn parse_name_field(line: &str) -> Option<String> {
    let trimmed = line.trim();
    let (key, value) = trimmed.split_once(':')?;
    if !key.trim().eq_ignore_ascii_case("name") {
        return None;
    }
    let name = value.trim();
    if name.is_empty() || name.contains(' ') {
        return None;
    }
    Some(name.to_string())
}

pub(super) fn looks_like_deferred_action_response(text: &str) -> bool {
    let lower = text.trim().to_ascii_lowercase();

    // Pattern-based detection: catch "I'll [verb]", "I will [verb]", "Let me [verb]",
    // "Shall I [verb]", "Would you like me to [verb]" where verb needs tools.
    // This is dynamic — any new action verb the LLM uses is automatically caught.
    if has_action_promise(&lower) {
        return true;
    }

    // Structural format markers — substring match appropriate for these patterns
    lower.contains("[consultation]")
        || lower.contains(&INTENT_GATE_MARKER.to_ascii_lowercase())
        || lower.contains("[tool_use:")
        || lower.contains("[tool_call:")
}

/// Detect first-person past-tense side-effect claims like "I have deleted…",
/// "I've removed…", "I created…". Used by the completion phase to catch
/// fabricated action claims: a reply asserting a completed side effect in a
/// task that made zero tool calls cannot be truthful.
pub(super) fn claims_completed_side_effect(text: &str) -> bool {
    // Same normalization as has_action_promise: unify Unicode apostrophes
    // so "I’ve" matches "I've".
    let normalized = text
        .trim()
        .to_ascii_lowercase()
        .replace(['\u{2018}', '\u{2019}', '`', '\u{02BC}'], "'");

    // Past-tense verbs that imply a tool-mediated side effect. Knowledge
    // verbs ("explained", "summarized") are intentionally absent — those
    // can be fulfilled without tools.
    const PAST_ACTION_VERBS: &[&str] = &[
        "deleted",
        "removed",
        "created",
        "wrote",
        "written",
        "updated",
        "edited",
        "modified",
        "replaced",
        "moved",
        "renamed",
        "executed",
        "ran",
        "installed",
        "saved",
        "stored",
        "copied",
        "combined",
        "merged",
        "deployed",
        "restarted",
        "stopped",
        "killed",
        "downloaded",
        "cloned",
        "pushed",
        "pulled",
        "committed",
        "cleaned",
        "cleared",
        "built",
        "generated",
        "appended",
        "uploaded",
        "sent",
        "added",
    ];
    // Adverbs allowed between the subject and the verb:
    // "I have successfully executed", "I've just removed".
    const FILLER_ADVERBS: &[&str] = &["now", "just", "successfully", "already", "also"];

    let words: Vec<String> = normalized
        .split_whitespace()
        .map(|w| {
            w.trim_matches(|c: char| c.is_ascii_punctuation() && c != '\'')
                .to_lowercase()
        })
        .filter(|w| !w.is_empty())
        .collect();

    let is_action_verb = |w: &str| PAST_ACTION_VERBS.contains(&w);
    let verb_after = |start: usize| -> bool {
        let mut idx = start;
        while words
            .get(idx)
            .is_some_and(|w| FILLER_ADVERBS.contains(&w.as_str()))
        {
            idx += 1;
        }
        words.get(idx).is_some_and(|w| is_action_verb(w))
    };

    for i in 0..words.len() {
        let claim = if words[i] == "i've" {
            // "I've removed …"
            verb_after(i + 1)
        } else if words[i] == "i" && words.get(i + 1).is_some_and(|w| w == "have") {
            // "I have deleted …"
            verb_after(i + 2)
        } else if words[i] == "i" {
            // Simple past: "I deleted …"
            verb_after(i + 1)
        } else {
            false
        };
        if claim {
            return true;
        }
    }
    false
}

/// Detect claims that a delegated agent has already been started. Without a
/// corresponding tool call, these statements fabricate asynchronous work that
/// will never produce a result.
pub(super) fn claims_delegation_started(text: &str) -> bool {
    let normalized = text
        .trim()
        .to_ascii_lowercase()
        .replace(['\u{2018}', '\u{2019}', '`', '\u{02BC}'], "'");

    let mentions_delegated_agent = [
        "specialist agent",
        "specialized review agent",
        "review agent",
        "research agent",
        "sub-agent",
        "subagent",
        "background agent",
    ]
    .iter()
    .any(|phrase| normalized.contains(phrase));
    if !mentions_delegated_agent {
        return false;
    }

    [
        "i've initiated",
        "i have initiated",
        "i initiated",
        "i've started",
        "i have started",
        "i started",
        "i've launched",
        "i have launched",
        "i launched",
        "i've spawned",
        "i have spawned",
        "i spawned",
        "i've delegated",
        "i have delegated",
        "i delegated",
        "is now running",
        "is running in the background",
        "has been started",
    ]
    .iter()
    .any(|phrase| normalized.contains(phrase))
}

/// Detect action-promise patterns like "I'll create", "I will run", "Let me check".
/// Returns true when the verb following the prefix is NOT a knowledge-only verb
/// (e.g., "explain", "describe", "summarize"), meaning the LLM needs tools to fulfill it.
pub(super) fn has_action_promise(text: &str) -> bool {
    // Normalize common Unicode apostrophes so contractions like "I’ll"
    // are treated the same as "I'll".
    let normalized = text.replace(['\u{2018}', '\u{2019}', '`', '\u{02BC}'], "'");

    // Verbs the LLM can fulfill without tools — pure knowledge/explanation verbs
    const KNOWLEDGE_ONLY_VERBS: &[&str] = &[
        "explain",
        "describe",
        "summarize",
        "clarify",
        "elaborate",
        "outline",
        "note",
        "mention",
        "address",
        "highlight",
        "tell",
        "share",
        "say",
        "answer",
        "provide",
        "be",
        "give",
        "offer",
        "know",
        "rephrase",
        "restate",
        // Memory/recall verbs — can be answered from conversation context or stored facts
        "recall",
        "confirm",
        "remember",
        "think",
        "point",
        "help",
    ];

    let words: Vec<String> = normalized
        .split_whitespace()
        .map(|w| {
            w.trim_matches(|c: char| c.is_ascii_punctuation() && c != '\'')
                .to_lowercase()
        })
        .filter(|w| !w.is_empty())
        .collect();

    for i in 0..words.len() {
        // Determine the index of the verb after the action-promise prefix
        let verb_idx = if words[i] == "i'll" {
            // "I'll [verb]"
            Some(i + 1)
        } else if words[i] == "i" && words.get(i + 1).is_some_and(|w| w == "will") {
            // "I will [verb]"
            Some(i + 2)
        } else if words[i] == "let" && words.get(i + 1).is_some_and(|w| w == "me") {
            // "Let me [verb]"
            Some(i + 2)
        } else if words[i] == "shall" && words.get(i + 1).is_some_and(|w| w == "i") {
            // "Shall I [verb]"
            Some(i + 2)
        } else if words[i] == "would"
            && words.get(i + 1).is_some_and(|w| w == "you")
            && words.get(i + 2).is_some_and(|w| w == "like")
            && words.get(i + 3).is_some_and(|w| w == "me")
            && words.get(i + 4).is_some_and(|w| w == "to")
        {
            // "Would you like me to [verb]"
            Some(i + 5)
        } else {
            None
        };

        if let Some(vi) = verb_idx {
            if let Some(verb) = words.get(vi) {
                if !KNOWLEDGE_ONLY_VERBS.contains(&verb.as_str()) {
                    return true;
                }
            }
        }
    }

    false
}

/// Check whether a model response contains substantive text content rather
/// than just deferred-action phrases.  Used to decide whether to accept a
/// text-only response after repeated deferred-no-tool retries: if the model
/// finally produced real content (greeting, explanation, joke, etc.) we should
/// let it through instead of stalling further.
///
/// The heuristic:
/// 1. The text must be at least `min_len` characters (default 50).
/// 2. After stripping lines that are purely deferred-action phrases, there must
///    still be substantive content left.
pub(super) fn is_substantive_text_response(text: &str, min_len: usize) -> bool {
    let trimmed = text.trim();
    if trimmed.len() < min_len {
        return false;
    }

    // Strip lines that are purely deferred-action phrases.
    // If most of the text survives, the response is substantive.
    let substantive_lines: Vec<&str> = trimmed
        .lines()
        .filter(|line| {
            let l = line.trim();
            if l.is_empty() {
                return false;
            }
            // Keep lines that do NOT look like pure deferral text
            !has_action_promise(&l.to_ascii_lowercase())
        })
        .collect();

    let substantive_text: String = substantive_lines.join(" ");
    let substantive_len = substantive_text.trim().len();

    // Must have at least min_len chars of non-deferred content
    substantive_len >= min_len
}

/// Heuristic: does the user's message look like a multi-part request that
/// warrants a detailed response?  We check for numbered lists, explanation
/// keywords, and conjunction-heavy compound tasks.
pub(super) fn looks_like_multi_part_request(text: &str) -> bool {
    let lower = text.to_ascii_lowercase();

    // Count numbered/lettered items: "1)", "2)", "a)", "b)", "step 1", etc.
    let numbered_items = {
        let re = regex::Regex::new(r"(?:^|\s)(?:\d+[.)]\s|[a-e][.)]\s|step\s+\d)").unwrap();
        re.find_iter(&lower).count()
    };
    if numbered_items >= 2 {
        return true;
    }

    // Explanation keywords: user explicitly wants reasoning
    let explanation_words = [
        "explain why",
        "explain how",
        "tell me why",
        "describe how",
        "show me",
        "what did you",
        "summarize what",
        "thorough review",
        "find all",
        "list all",
        "review it",
        "review the",
        "audit",
    ];
    let has_explanation_request = explanation_words.iter().any(|w| lower.contains(w));

    // Compound task indicators
    let compound_signals = [
        "also ",
        "then ",
        "after that",
        "additionally",
        "finally ",
        "and then",
        "before ",
        "as well",
    ];
    let compound_count = compound_signals
        .iter()
        .filter(|s| lower.contains(*s))
        .count();

    // Multi-part if explanation requested, or ≥2 compound signals
    has_explanation_request || compound_count >= 2
}

/// Remove leaked text-only control markers and pseudo tool-call text.
#[cfg(test)]
pub(super) fn sanitize_response_analysis(analysis: &str) -> String {
    let lines: Vec<&str> = analysis.lines().collect();
    let has_pseudo_tool_block = lines.iter().any(|line| is_pseudo_tool_line(line));

    let mut cleaned: Vec<String> = Vec::with_capacity(lines.len());
    let mut i = 0usize;
    while i < lines.len() {
        let line = lines[i];
        let trimmed = line.trim();
        let lower = trimmed.to_ascii_lowercase();

        if lower == "arguments:" {
            let mut j = i + 1;
            let mut block_has_tool_signature = false;
            while j < lines.len() {
                let next = lines[j].trim();
                if next.is_empty() {
                    break;
                }
                if let Some(name) = parse_name_field(next) {
                    if is_tool_name_like(&name) {
                        block_has_tool_signature = true;
                    }
                }
                let next_lower = next.to_ascii_lowercase();
                if next_lower.starts_with("cmd:")
                    || next_lower.starts_with("command:")
                    || next_lower.starts_with("args:")
                    || next_lower.starts_with("arguments:")
                {
                    block_has_tool_signature = true;
                }
                j += 1;
            }

            if block_has_tool_signature {
                i = j;
                continue;
            }
        }

        if is_pseudo_tool_line(line) {
            i += 1;
            continue;
        }

        let replaced = line.replace(crate::llm_markers::TEXT_ONLY_RESPONSE_MARKER, "");
        let trimmed_replaced = replaced.trim();
        let lower_replaced = trimmed_replaced.to_ascii_lowercase();

        if lower_replaced == "[consultation]" {
            i += 1;
            continue;
        }

        if lower_replaced.starts_with(&INTENT_GATE_MARKER.to_ascii_lowercase()) {
            i += 1;
            continue;
        }

        // Some models echo the text-only control instructions verbatim.
        // Strip the control header and nearby instruction lines so they don't
        // pollute the injected warm-start context for iteration 2.
        if lower_replaced.starts_with("[important:")
            && (lower_replaced.contains("consultation")
                || (lower_replaced.contains("you are being consulted")
                    && lower_replaced.contains("respond with text only")))
        {
            i += 1;
            continue;
        }
        if lower_replaced.contains("text only")
            && (lower_replaced.contains("no tools")
                || lower_replaced.contains("no function calls")
                || lower_replaced.contains("tool_use")
                || lower_replaced.contains("functioncall"))
        {
            i += 1;
            continue;
        }
        if lower_replaced.starts_with("end your response with")
            || lower_replaced.starts_with("end with one line")
            || lower_replaced == "guidelines:"
            || lower_replaced.starts_with("- complexity:")
            || lower_replaced.starts_with("- only include schedule")
            || lower_replaced.starts_with("- domains is optional")
        {
            i += 1;
            continue;
        }

        if has_pseudo_tool_block
            && (lower_replaced.starts_with("cmd:")
                || lower_replaced.starts_with("command:")
                || lower_replaced.starts_with("args:")
                || lower_replaced.starts_with("arguments:")
                || parse_name_field(trimmed_replaced)
                    .as_deref()
                    .is_some_and(is_tool_name_like))
        {
            i += 1;
            continue;
        }

        if trimmed_replaced.is_empty() {
            if cleaned.last().is_some_and(|prev| prev.is_empty()) {
                i += 1;
                continue;
            }
            cleaned.push(String::new());
        } else {
            cleaned.push(replaced.trim_end().to_string());
        }
        i += 1;
    }

    cleaned.join("\n").trim().to_string()
}