sausage/agent/src/llm.rs

//! Inference through two API shapes: OpenAI Responses and Anthropic Messages.
//!
//! The conversation is stored in a neutral format, so the agent can switch
//! between the two shapes.

use serde::{Deserialize, Serialize};
use serde_json::{Value, json};
use wstd::http::{Body, Client, Request};

#[derive(Deserialize, Clone)]
pub struct Config {
    /// `openai-responses` or `anthropic-messages`.
    pub api: String,
    pub base_url: String,
    pub model: String,
    pub token: String,
    /// Optional reasoning effort for OpenAI models only (wins over `thinking` there).
    #[serde(default)]
    pub reasoning_effort: Option<String>,
    #[serde(default)]
    pub max_tokens: Option<u32>,
    /// Thinking effort for both APIs: "low", "medium", "high", "xhigh", "max" (`LEVELS`), or "off"/unset for none.
    /// Anthropic: adaptive thinking + output_config.effort; OpenAI: reasoning.effort. A level the model rejects is
    /// replaced by the nearest accepted one (logged), never by "off".
    #[serde(default)]
    pub thinking: Option<String>,
}

/// Thinking levels, lowest first.
pub const LEVELS: [&str; 5] = ["low", "medium", "high", "xhigh", "max"];

/// The level nearest to `want` (lower on a tie) among those a 400 error lists in quotes (else all levels), not
/// yet `tried`. None: nothing left, so the request goes out without an effort field (model default).
fn fallback_effort(want: &str, error: &str, tried: &[String]) -> Option<String> {
    let listed: Vec<&str> = error.split('\'').skip(1).step_by(2).filter(|w| LEVELS.contains(w)).collect();
    let pool = if listed.iter().any(|l| *l != want) { listed } else { LEVELS.to_vec() };
    let at = LEVELS.iter().position(|l| *l == want).unwrap_or(1) as i32;
    pool.into_iter()
        .filter(|l| *l != want && !tried.iter().any(|t| t == l))
        .min_by_key(|l| {
            let i = LEVELS.iter().position(|x| x == l).unwrap_or(1) as i32;
            ((i - at).abs(), i > at)
        })
        .map(str::to_string)
}

#[cfg(test)]
mod tests {
    use super::fallback_effort;

    #[test]
    fn nearest_accepted_level() {
        let anthropic = "HTTP 400: output_config.effort: Input should be 'low', 'medium', 'high' or 'max'";
        assert_eq!(fallback_effort("xhigh", anthropic, &["xhigh".into()]).as_deref(), Some("high"));
        let openai = "HTTP 400: Invalid value: 'ultra'. Supported values are: 'none', 'minimal', 'low', 'medium', 'high', 'xhigh', and 'max'.";
        assert_eq!(fallback_effort("ultra", openai, &["ultra".into()]).as_deref(), Some("medium"));
        // No list in the error: step to the neighbour, never to off.
        assert_eq!(fallback_effort("max", "HTTP 400: effort max not supported", &["max".into()]).as_deref(), Some("xhigh"));
    }
}

#[derive(Serialize, Deserialize, Clone, Debug)]
#[serde(tag = "kind", rename_all = "snake_case")]
pub enum Entry {
    User { text: String },
    Assistant {
        text: String,
        #[serde(default, skip_serializing_if = "Option::is_none")]
        phase: Option<String>,
    },
    ToolCall { id: String, name: String, arguments: String },
    ToolResult { id: String, output: String },
    /// A provider item that only this API shape understands (OpenAI reasoning, Anthropic thinking).
    Opaque { api: String, item: Value },
}

pub struct Tool {
    pub name: &'static str,
    pub description: &'static str,
    pub parameters: Value,
}

pub async fn complete(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]) -> Result<Vec<Entry>, String> {
    let want = match cfg.api.as_str() {
        "openai-responses" => cfg.reasoning_effort.clone().or_else(|| cfg.thinking.clone()),
        _ => cfg.thinking.clone(),
    }
    .filter(|t| t != "off");
    let mut effort = want.clone();
    let mut tried = Vec::new();
    loop {
        let result = match cfg.api.as_str() {
            "openai-responses" => openai(cfg, system, tools, history, effort.as_deref()).await,
            "anthropic-messages" => anthropic(cfg, system, tools, history, want.is_some(), effort.as_deref()).await,
            other => return Err(format!("unknown API shape {other:?}")),
        };
        match (result, effort.clone()) {
            // The model rejected the effort level: retry with the nearest accepted one (thinking stays on).
            (Err(e), Some(current)) if e.starts_with("HTTP 400") && e.contains("effort") => {
                tried.push(current.clone());
                effort = fallback_effort(want.as_deref().unwrap_or(""), &e, &tried);
                eprintln!("agent: {} rejected thinking effort {current:?}; retrying with {effort:?}: {e}", cfg.model);
            }
            (result, _) => return result,
        }
    }
}

async fn post(url: &str, headers: &[(&str, String)], body: &Value) -> Result<Value, String> {
    let mut req = Request::post(url).header("content-type", "application/json");
    for (k, v) in headers {
        req = req.header(*k, v.as_str());
    }
    let req = req.body(Body::from(body.to_string())).map_err(|e| format!("request: {e}"))?;
    let client = Client::new();
    let resp = client.send(req).await.map_err(|e| format!("HTTP request failed: {e:?}"))?;
    let status = resp.status();
    let mut body = resp.into_body();
    let text = body.str_contents().await.map_err(|e| format!("reading the response failed: {e:?}"))?.to_string();
    if !status.is_success() {
        return Err(format!("HTTP {status}: {text}"));
    }
    serde_json::from_str(&text).map_err(|e| format!("invalid JSON in response: {e}"))
}

async fn openai(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry], effort: Option<&str>) -> Result<Vec<Entry>, String> {
    let mut input = Vec::new();
    for e in history {
        match e {
            Entry::User { text } => input.push(json!({"role": "user", "content": text})),
            Entry::Assistant { text, phase } => {
                let mut item = json!({"role": "assistant", "content": text});
                if let Some(phase) = phase {
                    item["phase"] = json!(phase);
                }
                input.push(item);
            }
            Entry::ToolCall { id, name, arguments } => {
                input.push(json!({"type": "function_call", "call_id": id, "name": name, "arguments": arguments}))
            }
            Entry::ToolResult { id, output } => {
                input.push(json!({"type": "function_call_output", "call_id": id, "output": output}))
            }
            Entry::Opaque { api, item } if api == "openai-responses" => input.push(item.clone()),
            Entry::Opaque { .. } => {}
        }
    }
    let tools: Vec<Value> = tools
        .iter()
        .map(|t| json!({"type": "function", "name": t.name, "description": t.description, "parameters": t.parameters}))
        .collect();
    let mut body = json!({
        "model": cfg.model,
        "instructions": system,
        "input": input,
        "tools": tools,
        "store": false,
        "include": ["reasoning.encrypted_content"],
    });
    if let Some(max_tokens) = cfg.max_tokens {
        body["max_output_tokens"] = json!(max_tokens);
    }
    if let Some(effort) = effort {
        body["reasoning"] = json!({"effort": effort});
    }
    let url = format!("{}/v1/responses", cfg.base_url);
    let resp = post(&url, &[("authorization", format!("Bearer {}", cfg.token))], &body).await?;
    let mut out = Vec::new();
    for item in resp["output"].as_array().cloned().unwrap_or_default() {
        match item["type"].as_str() {
            Some("message") => {
                let text: String = item["content"]
                    .as_array()
                    .map(|parts| {
                        parts
                            .iter()
                            .filter(|p| p["type"] == "output_text")
                            .filter_map(|p| p["text"].as_str())
                            .collect()
                    })
                    .unwrap_or_default();
                let phase = item["phase"].as_str().map(str::to_string);
                if !text.is_empty() || phase.is_some() {
                    out.push(Entry::Assistant { text, phase });
                }
            }
            Some("function_call") => out.push(Entry::ToolCall {
                id: item["call_id"].as_str().unwrap_or_default().to_string(),
                name: item["name"].as_str().unwrap_or_default().to_string(),
                arguments: item["arguments"].as_str().unwrap_or("{}").to_string(),
            }),
            Some("reasoning") => out.push(Entry::Opaque { api: "openai-responses".into(), item }),
            _ => {}
        }
    }
    if resp["status"] == "incomplete" && out.is_empty() {
        return Err(format!("incomplete response: {}", resp["incomplete_details"]));
    }
    Ok(out)
}

async fn anthropic(
    cfg: &Config,
    system: &str,
    tools: &[Tool],
    history: &[Entry],
    thinking: bool,
    effort: Option<&str>,
) -> Result<Vec<Entry>, String> {
    // Merge consecutive blocks of the same role into one message.
    let mut messages: Vec<(String, Vec<Value>)> = Vec::new();
    let mut push = |role: &str, block: Value| match messages.last_mut() {
        Some((r, blocks)) if r == role => blocks.push(block),
        _ => messages.push((role.to_string(), vec![block])),
    };
    for e in history {
        match e {
            Entry::User { text } => push("user", json!({"type": "text", "text": text})),
            Entry::Assistant { text, .. } => {
                push("assistant", json!({"type": "text", "text": text}))
            }
            Entry::ToolCall { id, name, arguments } => {
                let input: Value = serde_json::from_str(arguments).unwrap_or(json!({}));
                push("assistant", json!({"type": "tool_use", "id": id, "name": name, "input": input}))
            }
            Entry::ToolResult { id, output } => {
                push("user", json!({"type": "tool_result", "tool_use_id": id, "content": output}))
            }
            // Thinking blocks go back unchanged (with their signature): required next to tool use.
            Entry::Opaque { api, item } if api == "anthropic-messages" && thinking => push("assistant", item.clone()),
            Entry::Opaque { .. } => {}
        }
    }
    let messages: Vec<Value> = messages.into_iter().map(|(role, content)| json!({"role": role, "content": content})).collect();
    let tools: Vec<Value> = tools
        .iter()
        .map(|t| json!({"name": t.name, "description": t.description, "input_schema": t.parameters}))
        .collect();
    let mut body = json!({
        "model": cfg.model,
        "system": system,
        "messages": messages,
        "tools": tools,
        "max_tokens": cfg.max_tokens.unwrap_or(32000),
    });
    if thinking {
        body["thinking"] = json!({"type": "adaptive"});
        if let Some(effort) = effort {
            body["output_config"] = json!({"effort": effort});
        }
    }
    let url = format!("{}/v1/messages", cfg.base_url);
    let resp = post(
        &url,
        &[("x-api-key", cfg.token.clone()), ("anthropic-version", "2023-06-01".to_string())],
        &body,
    )
    .await?;
    let mut out = Vec::new();
    for block in resp["content"].as_array().cloned().unwrap_or_default() {
        match block["type"].as_str() {
            Some("text") => {
                let text = block["text"].as_str().unwrap_or_default().to_string();
                if !text.is_empty() {
                    out.push(Entry::Assistant { text, phase: None });
                }
            }
            Some("tool_use") => out.push(Entry::ToolCall {
                id: block["id"].as_str().unwrap_or_default().to_string(),
                name: block["name"].as_str().unwrap_or_default().to_string(),
                arguments: block["input"].to_string(),
            }),
            Some("thinking") | Some("redacted_thinking") => {
                out.push(Entry::Opaque { api: "anthropic-messages".into(), item: block })
            }
            _ => {}
        }
    }
    Ok(out)
}