mesh-llm-client 0.68.0

/// Smart model router — classifies requests and picks the best model.
use serde_json::Value;

// ── Request categories ──────────────────────────────────────────────

#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
pub enum Category {
    Code,
    Reasoning,
    Chat,
    ToolCall,
    Creative,
    /// Factual lookup, summarization, knowledge retrieval
    Info,
    /// Image generation or analysis (future: multimodal models)
    Image,
}

/// How complex/heavy the request appears to be.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum Complexity {
    Quick,    // simple fact, short answer, casual
    Moderate, // normal conversation, standard code
    Deep,     // long reasoning, complex analysis, architecture
}

/// Full classification result.
#[derive(Debug, Clone, PartialEq)]
pub struct Classification {
    pub category: Category,
    pub complexity: Complexity,
    pub needs_tools: bool,
    pub has_media_inputs: bool,
}

#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
pub struct MediaRequirements {
    pub has_media: bool,
    pub needs_vision: bool,
    pub needs_audio: bool,
}

// ── Model profiles ──────────────────────────────────────────────────

/// Quality tier: higher = better quality, slower.
/// 1 = draft/tiny, 2 = good, 3 = strong, 4 = frontier
pub type Tier = u8;

pub struct ModelProfile {
    pub name: &'static str,
    pub strengths: &'static [Category],
    pub tier: Tier,
    /// Whether this model can handle tool-calling requests (function calling).
    /// Models without this set to true are filtered out when tools are present.
    pub tools: bool,
}

/// Static profiles for catalog models.
/// Order of strengths matters — first entry is primary strength.
pub static MODEL_PROFILES: &[ModelProfile] = &[
    // ── Tier 4: Frontier ────────────────────────────────────────
    ModelProfile {
        name: "Qwen3-235B-A22B-Q4_K_M",
        strengths: &[
            Category::Code,
            Category::Reasoning,
            Category::Chat,
            Category::Creative,
        ],
        tier: 4,
        tools: true,
    },
    ModelProfile {
        name: "Llama-3.1-405B-Instruct-Q2_K",
        strengths: &[Category::Chat, Category::Reasoning, Category::Code],
        tier: 4,
        tools: true,
    },
    ModelProfile {
        name: "MiniMax-M2.5-Q4_K_M",
        strengths: &[
            Category::Code,
            Category::Reasoning,
            Category::Chat,
            Category::Creative,
            Category::ToolCall,
        ],
        tier: 4,
        tools: true,
    },
    // ── Tier 3: Strong ──────────────────────────────────────────
    ModelProfile {
        name: "Qwen2.5-72B-Instruct-Q4_K_M",
        strengths: &[Category::Chat, Category::Reasoning, Category::Code],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "Llama-3.3-70B-Instruct-Q4_K_M",
        strengths: &[Category::Chat, Category::ToolCall, Category::Code],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "DeepSeek-R1-Distill-70B-Q4_K_M",
        strengths: &[Category::Reasoning],
        tier: 3,
        tools: false, // reasoning-only, no tool support
    },
    ModelProfile {
        name: "Mixtral-8x22B-Instruct-Q4_K_M",
        strengths: &[Category::Chat, Category::Code, Category::Reasoning],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "Qwen3-32B-Q4_K_M",
        strengths: &[Category::Reasoning, Category::Code, Category::Chat],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "Qwen2.5-Coder-32B-Instruct-Q4_K_M",
        strengths: &[Category::Code],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "DeepSeek-R1-Distill-Qwen-32B-Q4_K_M",
        strengths: &[Category::Reasoning],
        tier: 3,
        tools: false,
    },
    ModelProfile {
        name: "Qwen3-30B-A3B-Q4_K_M",
        strengths: &[Category::Chat, Category::Reasoning, Category::Code],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "Qwen3-Coder-30B-A3B-Instruct-Q4_K_M",
        strengths: &[Category::Code, Category::ToolCall],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "Qwen2.5-32B-Instruct-Q4_K_M",
        strengths: &[
            Category::Chat,
            Category::Reasoning,
            Category::Code,
            Category::ToolCall,
        ],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "Gemma-3-27B-it-Q4_K_M",
        strengths: &[Category::Reasoning, Category::Chat],
        tier: 3,
        tools: false, // unreliable tool calling
    },
    ModelProfile {
        name: "Qwen3.5-27B-Q4_K_M",
        strengths: &[Category::Code, Category::Reasoning, Category::Chat],
        tier: 3,
        tools: true,
    },
    ModelProfile {
        name: "Qwen3-Coder-Next-Q4_K_M",
        strengths: &[Category::Code, Category::ToolCall, Category::Reasoning],
        tier: 4,
        tools: true,
    },
    // ── Tier 2: Good ────────────────────────────────────────────
    ModelProfile {
        name: "Qwen3.5-9B-Q4_K_M",
        strengths: &[Category::Chat, Category::Code],
        tier: 2,
        tools: false,
    },
    ModelProfile {
        name: "Mistral-Small-3.1-24B-Instruct-Q4_K_M",
        strengths: &[Category::Chat, Category::ToolCall],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "Devstral-Small-2505-Q4_K_M",
        strengths: &[Category::Code, Category::ToolCall],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "GLM-4.7-Flash-Q4_K_M",
        strengths: &[Category::Chat, Category::ToolCall],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "GLM-4-32B-0414-Q4_K_M",
        strengths: &[Category::Chat, Category::ToolCall, Category::Code],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "Llama-4-Scout-Q4_K_M",
        strengths: &[Category::Chat, Category::ToolCall],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "Qwen3-14B-Q4_K_M",
        strengths: &[Category::Chat, Category::Reasoning],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "Qwen2.5-14B-Instruct-Q4_K_M",
        strengths: &[Category::Chat],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "Qwen2.5-Coder-14B-Instruct-Q4_K_M",
        strengths: &[Category::Code],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "DeepSeek-R1-Distill-Qwen-14B-Q4_K_M",
        strengths: &[Category::Reasoning],
        tier: 2,
        tools: false,
    },
    ModelProfile {
        name: "Gemma-3-12B-it-Q4_K_M",
        strengths: &[Category::Chat, Category::Reasoning],
        tier: 2,
        tools: false,
    },
    ModelProfile {
        name: "Qwen3-8B-Q4_K_M",
        strengths: &[Category::Chat, Category::Code],
        tier: 2,
        tools: true,
    },
    ModelProfile {
        name: "Hermes-2-Pro-Mistral-7B-Q4_K_M",
        strengths: &[Category::Chat],
        tier: 2,
        tools: false,
    },
    ModelProfile {
        name: "Qwen2.5-Coder-7B-Instruct-Q4_K_M",
        strengths: &[Category::Code],
        tier: 2,
        tools: true,
    },
    // ── Tier 1: Small / Draft ───────────────────────────────────
    ModelProfile {
        name: "Qwen3-4B-Q4_K_M",
        strengths: &[Category::Chat],
        tier: 1,
        tools: true,
    },
    ModelProfile {
        name: "Qwen2.5-3B-Instruct-Q4_K_M",
        strengths: &[Category::Chat],
        tier: 1,
        tools: true,
    },
    ModelProfile {
        name: "Llama-3.2-3B-Instruct-Q4_K_M",
        strengths: &[Category::Chat, Category::ToolCall],
        tier: 1,
        tools: true,
    },
];

pub fn profile_for(model_name: &str) -> Option<&'static ModelProfile> {
    // Direct match first
    if let Some(p) = MODEL_PROFILES.iter().find(|p| p.name == model_name) {
        return Some(p);
    }
    // Strip split GGUF suffix: "Model-00001-of-00004" → "Model"
    let clean = strip_split_suffix(model_name);
    if clean != model_name {
        return MODEL_PROFILES.iter().find(|p| p.name == clean);
    }
    None
}

/// Strip split GGUF suffix like "-00001-of-00004" from a model name.
pub fn strip_split_suffix(name: &str) -> &str {
    // Pattern: -NNNNN-of-NNNNN at the end
    if let Some(idx) = name.rfind("-of-") {
        // Check that what follows is digits and what precedes is -digits
        let after = &name[idx + 4..];
        if after.chars().all(|c| c.is_ascii_digit()) && !after.is_empty() {
            // Find the preceding -NNNNN
            if let Some(dash) = name[..idx].rfind('-') {
                let between = &name[dash + 1..idx];
                if between.chars().all(|c| c.is_ascii_digit()) && !between.is_empty() {
                    return &name[..dash];
                }
            }
        }
    }
    name
}

/// Owned version of strip_split_suffix for contexts that need a String.
pub fn strip_split_suffix_owned(name: &str) -> String {
    strip_split_suffix(name).to_string()
}

// ── Request classification ──────────────────────────────────────────

/// Classify a chat completion request body using heuristics.
/// No LLM call, just pattern matching on the request structure.
/// Classify a request body into category + complexity + needs_tools.
/// Tools presence is an attribute, not a category override — a code request
/// with tools is still Code (with needs_tools=true), not ToolCall.
pub fn classify(body: &Value) -> Classification {
    // Collect all text from messages for keyword analysis
    let text = collect_message_text(body);
    let lower = text.to_lowercase();
    let media = media_requirements(body);

    // Check if the request actually needs tool execution.
    // If the client sends a tools schema, this is an agentic session (Claude Code,
    // Goose, etc.) — always prefer the strongest tool-capable model regardless of
    // what the first message says.  Keyword matching on content is a secondary signal
    // but not required when tools are present.
    let has_tools_schema = body
        .get("tools")
        .and_then(|t| t.as_array())
        .map(|a| !a.is_empty())
        .unwrap_or(false);
    // Anthropic-style requests may include structured content blocks with
    // explicit tool_use/tool_result blocks — definitely tool-driven.
    let has_tool_blocks = body
        .get("messages")
        .and_then(|m| m.as_array())
        .map(|msgs| {
            msgs.iter().any(|msg| {
                msg.get("content")
                    .and_then(|c| c.as_array())
                    .map(|blocks| {
                        blocks.iter().any(|b| {
                            matches!(
                                b.get("type").and_then(|t| t.as_str()),
                                Some("tool_use") | Some("tool_result")
                            )
                        })
                    })
                    .unwrap_or(false)
            })
        })
        .unwrap_or(false);
    let needs_tools = has_tools_schema || has_tool_blocks;

    // Count last user message tokens (rough proxy for complexity)
    let last_user_len = last_user_message_len(body);

    // Code signals
    let code_signals = [
        "```",
        "def ",
        "fn ",
        "func ",
        "class ",
        "import ",
        "function",
        "const ",
        "let ",
        "var ",
        "return ",
        "write a program",
        "write code",
        "implement",
        "refactor",
        "debug",
        "fix the bug",
        "write a script",
        "code review",
        "pull request",
        "git ",
        "compile",
        "syntax",
        "python",
        "javascript",
        "typescript",
        " rust ",
        "golang",
        "java ",
        "c++",
        " ruby ",
        " swift ",
        "kotlin",
        "algorithm",
        "binary search",
        " sort ",
        "regex",
        " api ",
        " http ",
        " sql ",
        "database",
        " query ",
    ];
    let code_score: usize = code_signals.iter().filter(|s| lower.contains(*s)).count();

    // Reasoning signals
    let reasoning_signals = [
        "prove",
        "explain why",
        "step by step",
        "calculate",
        "solve",
        "derive",
        "what is the probability",
        "how many",
        "analyze",
        "compare and contrast",
        "evaluate",
        "mathematical",
        "theorem",
        "equation",
        "logic",
        "think carefully",
        "reason about",
    ];
    let reasoning_score: usize = reasoning_signals
        .iter()
        .filter(|s| lower.contains(*s))
        .count();

    // Creative signals
    let creative_signals = [
        "write a story",
        "write a poem",
        "creative",
        "imagine",
        "fiction",
        "narrative",
        "compose",
        "brainstorm",
        "write a song",
        "screenplay",
        "dialogue",
    ];
    let creative_score: usize = creative_signals
        .iter()
        .filter(|s| lower.contains(*s))
        .count();

    // Info/knowledge signals — factual lookup, summarization
    let info_signals = [
        "what is",
        "who is",
        "when did",
        "where is",
        "how does",
        "define ",
        "explain ",
        "summarize",
        "summary",
        "overview",
        "tell me about",
        "describe ",
        "what are the",
        "list the",
        "difference between",
        "compare ",
        "history of",
    ];
    let info_score: usize = info_signals.iter().filter(|s| lower.contains(*s)).count();

    // Image signals — generation or analysis (future)
    let image_signals = [
        "image",
        "picture",
        "photo",
        "draw",
        "generate an image",
        "visualize",
        "diagram",
        "screenshot",
        "describe this image",
    ];
    let image_score: usize = image_signals.iter().filter(|s| lower.contains(*s)).count();

    // Deep-thinking signals (want the biggest brain)
    let deep_signals = [
        "architect",
        "design a system",
        "trade-off",
        "tradeoff",
        "in depth",
        "comprehensive",
        "thorough",
        "detailed analysis",
        "long-term",
        "strategy",
        "plan for",
        "review this codebase",
        "rewrite",
        "from scratch",
    ];
    let deep_score: usize = deep_signals.iter().filter(|s| lower.contains(*s)).count();

    // System prompt hints
    let mut system_code = false;
    if let Some(messages) = body.get("messages").and_then(|m| m.as_array()) {
        for msg in messages {
            if msg.get("role").and_then(|r| r.as_str()) == Some("system")
                && let Some(content) = msg.get("content").and_then(|c| c.as_str())
            {
                let sys = content.to_lowercase();
                if sys.contains("developer") || sys.contains("coding") || sys.contains("programmer")
                {
                    system_code = true;
                }
            }
        }
    }

    // Pick category — tools don't override, content wins
    let category = if system_code
        || code_score >= 2
        || (code_score >= 1 && reasoning_score == 0 && creative_score == 0)
    {
        Category::Code
    } else if reasoning_score >= 2 {
        Category::Reasoning
    } else if creative_score >= 1 {
        Category::Creative
    } else if media.needs_vision || image_score >= 1 {
        Category::Image
    } else if needs_tools && code_score == 0 && reasoning_score == 0 && creative_score == 0 {
        // Only ToolCall if tools present AND no other signal dominates
        Category::ToolCall
    } else if info_score >= 2 && code_score == 0 {
        Category::Info
    } else {
        Category::Chat
    };

    // Complexity: Quick / Moderate / Deep
    let total_messages = body
        .get("messages")
        .and_then(|m| m.as_array())
        .map(|a| a.len())
        .unwrap_or(0);
    let complexity = if deep_score >= 1 || last_user_len > 500 || total_messages > 10 {
        Complexity::Deep
    } else if last_user_len < 60 && total_messages <= 2 && reasoning_score == 0 && deep_score == 0 {
        Complexity::Quick
    } else {
        Complexity::Moderate
    };

    Classification {
        category,
        complexity,
        needs_tools,
        has_media_inputs: media.has_media,
    }
}

pub fn media_requirements(body: &Value) -> MediaRequirements {
    let mut requirements = MediaRequirements::default();
    let Some(messages) = body.get("messages").and_then(|m| m.as_array()) else {
        return requirements;
    };

    for msg in messages {
        let Some(blocks) = msg.get("content").and_then(|c| c.as_array()) else {
            continue;
        };
        for block in blocks {
            let block_type = block
                .get("type")
                .and_then(|t| t.as_str())
                .unwrap_or_default();
            match block_type {
                "image_url" | "input_image" | "image" => {
                    requirements.has_media = true;
                    requirements.needs_vision = true;
                }
                "audio_url" | "input_audio" | "audio" => {
                    requirements.has_media = true;
                    requirements.needs_audio = true;
                }
                "file" | "input_file" => {
                    requirements.has_media = true;
                }
                _ => {
                    if block.get("image_url").is_some() || block.get("image").is_some() {
                        requirements.has_media = true;
                        requirements.needs_vision = true;
                    }
                    if block.get("audio_url").is_some() || block.get("audio").is_some() {
                        requirements.has_media = true;
                        requirements.needs_audio = true;
                    }
                }
            }
        }
    }

    requirements
}

/// Length of last user message in characters (rough complexity proxy).
fn last_user_message_len(body: &Value) -> usize {
    body.get("messages")
        .and_then(|m| m.as_array())
        .and_then(|msgs| {
            msgs.iter()
                .rev()
                .find(|m| m.get("role").and_then(|r| r.as_str()) == Some("user"))
        })
        .map(message_text)
        .map(|s| s.len())
        .unwrap_or(0)
}

fn collect_message_text(body: &Value) -> String {
    let mut text = String::new();
    if let Some(messages) = body.get("messages").and_then(|m| m.as_array()) {
        for msg in messages {
            let content = message_text(msg);
            if !content.is_empty() {
                text.push_str(&content);
                text.push('\n');
            }
        }
    }
    text
}

/// Extract message text for both OpenAI-style and Anthropic-style payloads.
fn message_text(msg: &Value) -> String {
    if let Some(s) = msg.get("content").and_then(|c| c.as_str()) {
        return s.to_string();
    }

    // Anthropic content blocks: [{"type":"text","text":"..."}, ...]
    if let Some(blocks) = msg.get("content").and_then(|c| c.as_array()) {
        let mut out = String::new();
        for b in blocks {
            if let Some(t) = b.get("text").and_then(|t| t.as_str()) {
                out.push_str(t);
                out.push('\n');
            }
        }
        return out;
    }

    String::new()
}

// ── Model selection ─────────────────────────────────────────────────

/// Pick the best model using full classification (category + complexity + tools).
pub fn pick_model_classified<'a>(
    classification: &Classification,
    available_models: &[(&'a str, f64)],
) -> Option<&'a str> {
    if available_models.is_empty() {
        return None;
    }

    // Filter for tool-capable models if tools are required
    let filtered: Vec<(&str, f64)> = if classification.needs_tools {
        available_models
            .iter()
            .filter(|(name, _)| profile_for(name).map(|p| p.tools).unwrap_or(false))
            .copied()
            .collect()
    } else {
        available_models.to_vec()
    };
    // Fall back to all models if no tool-capable model found
    let candidates = if filtered.is_empty() {
        available_models
    } else {
        &filtered
    };

    let category = classification.category;

    // Score each available model
    let mut scored: Vec<(&str, i32)> = candidates
        .iter()
        .map(|(name, tok_s)| {
            let profile = profile_for(name);
            let tier = profile.map(|p| p.tier).unwrap_or(1) as i32;

            // Task match is the primary signal.
            let has_match = profile
                .map(|p| p.strengths.contains(&category))
                .unwrap_or(false);

            let match_bonus = if has_match { 1000 } else { 0 };

            // Within matched models: primary > secondary > listed
            let position_bonus = profile
                .map(|p| {
                    p.strengths
                        .iter()
                        .position(|s| *s == category)
                        .map(|i| match i {
                            0 => 20,
                            1 => 10,
                            _ => 5,
                        })
                        .unwrap_or(0)
                })
                .unwrap_or(0);

            // Agentic vs chat scoring:
            // When tools are needed, strongly prefer the most capable model.
            // For chat, prefer the fastest model that matches.
            let tier_bonus = if classification.needs_tools {
                // Agentic: always prefer strongest. Tier dominates.
                // tier 1→20, tier 2→40, tier 3→60, tier 4→80
                tier * 20
            } else {
                // Chat/no-tools: always prefer bigger models, but less aggressively
                // than agentic. Small models are fallbacks, not first choice.
                match classification.complexity {
                    Complexity::Quick => tier * 5,     // tier 2→10, tier 3→15, tier 4→20
                    Complexity::Moderate => tier * 10, // tier 2→20, tier 3→30, tier 4→40
                    Complexity::Deep => tier * 15,     // tier 2→30, tier 3→45, tier 4→60
                }
            };

            // Speed bonus: higher for chat (speed matters), lower for agentic (quality matters)
            let speed_bonus = if classification.needs_tools {
                // Agentic: speed is a tiebreaker only
                (tok_s / 20.0).min(5.0) as i32
            } else {
                // Chat: speed matters more
                (tok_s / 5.0).min(20.0) as i32
            };

            let score = match_bonus + tier_bonus + position_bonus + speed_bonus;
            (*name, score)
        })
        .collect();

    scored.sort_by_key(|entry| std::cmp::Reverse(entry.1));

    // For non-agentic requests, spread load across top-scoring models.
    // Pick randomly among candidates within 15 points of the best score.
    // This avoids queueing all concurrent chat users on the same model
    // while keeping weak models as fallbacks, not equal contenders.
    if !classification.needs_tools && scored.len() > 1 {
        let best_score = scored[0].1;
        let top_tier: Vec<&(&str, i32)> = scored
            .iter()
            .filter(|(_, s)| best_score - s <= 15)
            .collect();
        if top_tier.len() > 1 {
            // Simple pseudo-random: use current time nanos to pick
            let nanos = std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap_or_default()
                .subsec_nanos() as usize;
            let pick = top_tier[nanos % top_tier.len()];
            return Some(pick.0);
        }
    }

    scored.first().map(|(name, _)| *name)
}

/// Legacy wrapper for tests that have category + tools but no complexity.
#[cfg(test)]
pub fn pick_model_with_tools<'a>(
    category: Category,
    available_models: &[(&'a str, f64)],
    tools_required: bool,
) -> Option<&'a str> {
    pick_model_classified(
        &Classification {
            category,
            complexity: Complexity::Moderate,
            needs_tools: tools_required,
            has_media_inputs: false,
        },
        available_models,
    )
}

// ── Tests ───────────────────────────────────────────────────────────

#[cfg(test)]
mod tests {
    use super::*;
    use serde_json::json;

    #[test]
    fn test_classify_tool_call() {
        // Content that implies tool use + tools schema = ToolCall
        let body = json!({
            "messages": [{"role": "user", "content": "Run the tests and check the output"}],
            "tools": [{"type": "function", "function": {"name": "bash"}}]
        });
        assert_eq!(classify(&body).category, Category::ToolCall);
    }

    #[test]
    fn test_classify_code() {
        let body = json!({
            "messages": [
                {"role": "user", "content": "Write a Python function to implement binary search and debug any issues"}
            ]
        });
        assert_eq!(classify(&body).category, Category::Code);
    }

    #[test]
    fn test_classify_reasoning() {
        let body = json!({
            "messages": [
                {"role": "user", "content": "Prove that the square root of 2 is irrational. Explain step by step."}
            ]
        });
        assert_eq!(classify(&body).category, Category::Reasoning);
    }

    #[test]
    fn test_classify_creative() {
        let body = json!({
            "messages": [
                {"role": "user", "content": "Write a story about a robot who learns to paint"}
            ]
        });
        assert_eq!(classify(&body).category, Category::Creative);
    }

    #[test]
    fn test_classify_chat_default() {
        let body = json!({
            "messages": [
                {"role": "user", "content": "What's the capital of France?"}
            ]
        });
        let cl = classify(&body);
        assert_eq!(cl.category, Category::Chat);
        assert_eq!(cl.complexity, Complexity::Quick); // short simple question
        assert!(!cl.needs_tools);
        assert!(!cl.has_media_inputs);
    }

    #[test]
    fn test_classify_deep_analysis() {
        let body = json!({
            "messages": [
                {"role": "user", "content": "Design a system architecture for a distributed database with strong consistency guarantees. Provide a detailed analysis of the trade-offs between CAP theorem constraints and explain how to handle network partitions in depth."}
            ]
        });
        let cl = classify(&body);
        assert_eq!(cl.complexity, Complexity::Deep);
    }

    #[test]
    fn test_classify_code_with_tools() {
        // Code request that happens to have tools — should be Code, not ToolCall
        let body = json!({
            "messages": [{"role": "user", "content": "Write a Python function to sort a list and debug it"}],
            "tools": [{"type": "function", "function": {"name": "bash"}}]
        });
        let cl = classify(&body);
        assert_eq!(cl.category, Category::Code);
        assert!(cl.needs_tools);
    }

    #[test]
    fn test_classify_tools_schema_always_needs_tools() {
        // Tools schema present = agentic session, always needs_tools
        // even if the message content is plain chat
        let body = json!({
            "messages": [{"role": "user", "content": "hello"}],
            "tools": [{"type": "function", "function": {"name": "bash"}}]
        });
        let cl = classify(&body);
        assert!(cl.needs_tools);
    }

    #[test]
    fn test_classify_tools_schema_with_tool_content() {
        // Tools in schema AND content implies tool use — needs tools
        let body = json!({
            "messages": [{"role": "user", "content": "Read the file and fix the bug"}],
            "tools": [{"type": "function", "function": {"name": "read"}}]
        });
        let cl = classify(&body);
        assert!(cl.needs_tools);
    }

    #[test]
    fn test_classify_anthropic_text_blocks_with_tools() {
        // Anthropic-style content blocks should still be parsed as text
        // and trigger needs_tools when tool-intent is present.
        let body = json!({
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": "List files in this directory and read README.md"}
                    ]
                }
            ],
            "tools": [{"name": "shell"}]
        });
        let cl = classify(&body);
        assert!(cl.needs_tools);
        assert!(matches!(cl.category, Category::Code | Category::ToolCall));
    }

    #[test]
    fn test_classify_anthropic_tool_use_block_sets_needs_tools() {
        // If an explicit tool_use/tool_result block is present, mark as needs_tools.
        let body = json!({
            "messages": [
                {
                    "role": "assistant",
                    "content": [
                        {"type": "tool_use", "id": "toolu_123", "name": "shell", "input": {"command": "ls"}}
                    ]
                }
            ]
        });
        let cl = classify(&body);
        assert!(cl.needs_tools);
    }

    #[test]
    fn test_anthropic_tool_request_prefers_stronger_tool_model() {
        // Reproduces Claude-like tool request shape and verifies needs_tools=true
        // pushes selection toward the stronger tool-capable model.
        let body = json!({
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": "List files in this directory and read README.md"}
                    ]
                }
            ],
            "tools": [{"name": "shell"}]
        });
        let cl = classify(&body);
        assert!(cl.needs_tools);

        let available = vec![("Qwen3-8B-Q4_K_M", 40.0), ("MiniMax-M2.5-Q4_K_M", 20.0)];
        let picked = pick_model_classified(&cl, &available);
        assert_eq!(picked, Some("MiniMax-M2.5-Q4_K_M"));
    }

    #[test]
    fn test_classify_system_prompt_code() {
        let body = json!({
            "messages": [
                {"role": "system", "content": "You are a senior developer and coding assistant."},
                {"role": "user", "content": "Help me with this."}
            ]
        });
        assert_eq!(classify(&body).category, Category::Code);
    }

    #[test]
    fn test_media_requirements_detect_audio_block() {
        let body = json!({
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": "Transcribe this clip"},
                        {"type": "audio_url", "audio_url": {"url": "mesh://blob/client-1/example"}}
                    ]
                }
            ]
        });
        let media = media_requirements(&body);
        assert!(media.has_media);
        assert!(media.needs_audio);
        assert!(!media.needs_vision);
        assert!(classify(&body).has_media_inputs);
    }

    #[test]
    fn test_media_requirements_detect_image_block() {
        let body = json!({
            "messages": [
                {
                    "role": "user",
                    "content": [
                        {"type": "image_url", "image_url": {"url": "data:image/png;base64,abc"}}
                    ]
                }
            ]
        });
        let media = media_requirements(&body);
        assert!(media.has_media);
        assert!(media.needs_vision);
        assert!(!media.needs_audio);
        assert!(classify(&body).has_media_inputs);
    }

    #[test]
    fn test_pick_model_primary_strength_wins() {
        // Qwen3-8B (tier 2, Chat primary) and 235B (tier 4, Chat 3rd) score within
        // 15 points at Moderate complexity, so either is a valid pick (load spread).
        let available = vec![("Qwen3-8B-Q4_K_M", 50.0), ("Qwen3-235B-A22B-Q4_K_M", 20.0)];
        let result = pick_model_classified(
            &Classification {
                category: Category::Chat,
                complexity: Complexity::Moderate,
                needs_tools: false,
                has_media_inputs: false,
            },
            &available,
        );
        assert!(result == Some("Qwen3-8B-Q4_K_M") || result == Some("Qwen3-235B-A22B-Q4_K_M"));
    }

    #[test]
    fn test_deep_complexity_prefers_bigger() {
        // Deep complexity amplifies tier bonus, but scores are within 15 points
        // so load spread makes either a valid pick.
        let available = vec![("Qwen3-8B-Q4_K_M", 50.0), ("Qwen3-235B-A22B-Q4_K_M", 20.0)];
        let result = pick_model_classified(
            &Classification {
                category: Category::Chat,
                complexity: Complexity::Deep,
                needs_tools: false,
                has_media_inputs: false,
            },
            &available,
        );
        assert!(result == Some("Qwen3-8B-Q4_K_M") || result == Some("Qwen3-235B-A22B-Q4_K_M"));
    }

    #[test]
    fn test_quick_complexity_prefers_smaller() {
        // Quick complexity: both score within 15 points (load spread applies).
        // Either is a valid pick — the key property is neither is excluded.
        let available = vec![
            ("Qwen3-8B-Q4_K_M", 50.0),
            ("Qwen2.5-72B-Instruct-Q4_K_M", 10.0),
        ];
        let result = pick_model_classified(
            &Classification {
                category: Category::Chat,
                complexity: Complexity::Quick,
                needs_tools: false,
                has_media_inputs: false,
            },
            &available,
        );
        assert!(result == Some("Qwen3-8B-Q4_K_M") || result == Some("Qwen2.5-72B-Instruct-Q4_K_M"));
    }

    #[test]
    fn test_pick_model_prefers_strength_match() {
        // Same tier, same speed — scores within 15 points, load spread applies.
        let available = vec![
            ("DeepSeek-R1-Distill-70B-Q4_K_M", 10.0), // tier 3, reasoning specialist
            ("Qwen2.5-72B-Instruct-Q4_K_M", 10.0),    // tier 3, chat primary
        ];
        let result = pick_model_classified(
            &Classification {
                category: Category::Reasoning,
                complexity: Complexity::Moderate,
                needs_tools: false,
                has_media_inputs: false,
            },
            &available,
        );
        assert!(
            result == Some("DeepSeek-R1-Distill-70B-Q4_K_M")
                || result == Some("Qwen2.5-72B-Instruct-Q4_K_M")
        );
    }

    #[test]
    fn test_pick_model_code_specialist() {
        // Same tier, same speed — scores within 15 points, load spread applies.
        let available = vec![
            ("Qwen2.5-Coder-32B-Instruct-Q4_K_M", 15.0),
            ("Qwen2.5-32B-Instruct-Q4_K_M", 15.0),
        ];
        let result = pick_model_classified(
            &Classification {
                category: Category::Code,
                complexity: Complexity::Moderate,
                needs_tools: false,
                has_media_inputs: false,
            },
            &available,
        );
        assert!(
            result == Some("Qwen2.5-Coder-32B-Instruct-Q4_K_M")
                || result == Some("Qwen2.5-32B-Instruct-Q4_K_M")
        );
    }

    #[test]
    fn test_pick_model_empty() {
        let available: Vec<(&str, f64)> = vec![];
        assert_eq!(
            pick_model_classified(
                &Classification {
                    category: Category::Chat,
                    complexity: Complexity::Moderate,
                    needs_tools: false,
                    has_media_inputs: false,
                },
                &available
            ),
            None
        );
    }

    #[test]
    fn test_pick_model_unknown_model_still_works() {
        let available = vec![("SomeUnknownModel", 30.0)];
        let result = pick_model_classified(
            &Classification {
                category: Category::Chat,
                complexity: Complexity::Moderate,
                needs_tools: false,
                has_media_inputs: false,
            },
            &available,
        );
        assert_eq!(result, Some("SomeUnknownModel"));
    }

    #[test]
    fn test_profile_lookup() {
        assert!(profile_for("Qwen3-235B-A22B-Q4_K_M").is_some());
        assert_eq!(profile_for("Qwen3-235B-A22B-Q4_K_M").unwrap().tier, 4);
        assert!(profile_for("nonexistent").is_none());
    }

    #[test]
    fn test_all_profiles_have_strengths() {
        for p in MODEL_PROFILES {
            assert!(!p.strengths.is_empty(), "{} has no strengths", p.name);
        }
    }

    #[test]
    fn test_classify_empty_tools_is_not_tool_call() {
        let body = json!({
            "messages": [{"role": "user", "content": "hello"}],
            "tools": []
        });
        assert_eq!(classify(&body).category, Category::Chat);
    }

    #[test]
    fn test_strip_split_suffix() {
        assert_eq!(
            strip_split_suffix("MiniMax-M2.5-Q4_K_M-00001-of-00004"),
            "MiniMax-M2.5-Q4_K_M"
        );
        assert_eq!(
            strip_split_suffix("Qwen3-Coder-Next-Q4_K_M-00001-of-00004"),
            "Qwen3-Coder-Next-Q4_K_M"
        );
        assert_eq!(
            strip_split_suffix("Hermes-2-Pro-Mistral-7B-Q4_K_M"),
            "Hermes-2-Pro-Mistral-7B-Q4_K_M"
        );
        assert_eq!(strip_split_suffix(""), "");
    }

    #[test]
    fn test_profile_for_split_gguf() {
        let p = profile_for("MiniMax-M2.5-Q4_K_M-00001-of-00004");
        assert!(p.is_some());
        assert_eq!(p.unwrap().name, "MiniMax-M2.5-Q4_K_M");
        assert_eq!(p.unwrap().tier, 4);
    }
}

#[test]
fn test_tools_filter_prefers_capable() {
    let available = vec![
        ("DeepSeek-R1-Distill-Qwen-32B-Q4_K_M", 10.0), // tools: false, Reasoning only
        ("Qwen2.5-32B-Instruct-Q4_K_M", 50.0),         // tools: true, Chat+Reasoning+Code
    ];
    // Without tools, Reasoning request: scores within 15 points, either valid
    let result = pick_model_with_tools(Category::Reasoning, &available, false);
    assert!(
        result == Some("DeepSeek-R1-Distill-Qwen-32B-Q4_K_M")
            || result == Some("Qwen2.5-32B-Instruct-Q4_K_M")
    );
    // With tools, Reasoning request: Qwen wins (DeepSeek filtered out — can't do tools)
    let result = pick_model_with_tools(Category::Reasoning, &available, true);
    assert_eq!(result, Some("Qwen2.5-32B-Instruct-Q4_K_M"));
}

#[test]
fn test_tools_filter_fallback_when_none_capable() {
    let available = vec![
        ("DeepSeek-R1-Distill-Qwen-32B-Q4_K_M", 10.0), // tools: false
    ];
    // With tools required but nothing capable: falls back to available
    let result = pick_model_with_tools(Category::Reasoning, &available, true);
    assert_eq!(result, Some("DeepSeek-R1-Distill-Qwen-32B-Q4_K_M"));
}

#[test]
fn test_agentic_prefers_strongest_model() {
    // Agentic (needs_tools=true): 32B (tier 3) should beat 7B (tier 2) even though 7B is faster
    let available = vec![
        ("Hermes-2-Pro-Mistral-7B-Q4_K_M", 87.0), // tier 2, tools: false
        ("Qwen2.5-Coder-7B-Instruct-Q4_K_M", 85.0), // tier 2, tools: true
        ("Qwen2.5-32B-Instruct-Q4_K_M", 18.0),    // tier 3, tools: true
    ];
    let cl = Classification {
        category: Category::Code,
        complexity: Complexity::Moderate,
        needs_tools: true,
        has_media_inputs: false,
    };
    let result = pick_model_classified(&cl, &available);
    // 32B should win: tier 3×20=60 beats Coder tier 2×20=40, despite lower speed
    assert_eq!(result, Some("Qwen2.5-32B-Instruct-Q4_K_M"));
}

#[test]
fn test_chat_prefers_fastest_model() {
    // Chat (needs_tools=false, Quick): scores within 15 points, load spread applies.
    let available = vec![
        ("Hermes-2-Pro-Mistral-7B-Q4_K_M", 87.0), // tier 2, fast
        ("Qwen2.5-32B-Instruct-Q4_K_M", 18.0),    // tier 3, slow
    ];
    let cl = Classification {
        category: Category::Chat,
        complexity: Complexity::Quick,
        needs_tools: false,
        has_media_inputs: false,
    };
    let result = pick_model_classified(&cl, &available);
    assert!(
        result == Some("Hermes-2-Pro-Mistral-7B-Q4_K_M")
            || result == Some("Qwen2.5-32B-Instruct-Q4_K_M")
    );
}

#[test]
fn test_agentic_deep_strongly_prefers_biggest() {
    // Deep agentic: tier 4 should massively beat tier 2
    let available = vec![
        ("Qwen2.5-Coder-7B-Instruct-Q4_K_M", 85.0), // tier 2
        ("MiniMax-M2.5-Q4_K_M", 21.0),              // tier 4
    ];
    let cl = Classification {
        category: Category::Code,
        complexity: Complexity::Deep,
        needs_tools: true,
        has_media_inputs: false,
    };
    let result = pick_model_classified(&cl, &available);
    assert_eq!(result, Some("MiniMax-M2.5-Q4_K_M"));
}