brazen 0.0.4

A stateless, swiss-army-knife adapter for every LLM provider and protocol.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160

//! Top-level body decode (openai-chat-mapping §2.1, inverted): each known
//! `chat/completions` key lands on its typed canonical field — the single source of
//! truth the egress encoders re-project — and every unknown key rides `extra`
//! verbatim. The `messages[]` fold lives in [`super::messages`].

use serde_json::Value;

use super::{arr_of, bad, err, messages, obj_of, opt_bool, opt_str, str_of};
use crate::canonical::{CanonicalRequest, OutputFormat, ReasoningEffort, Tool, ToolChoice};
use crate::ingress::IngressError;

/// OpenAI `chat/completions` request bytes → `CanonicalRequest` (ingress.md §2).
pub(crate) fn decode_request(bytes: &[u8]) -> Result<CanonicalRequest, IngressError> {
    let top: Value =
        serde_json::from_slice(bytes).map_err(|e| err(format!("request body is not JSON: {e}")))?;
    let Value::Object(body) = top else {
        return Err(bad("request body", "a JSON object"));
    };
    let mut req = CanonicalRequest::default();
    for (k, v) in body {
        match k.as_str() {
            "model" => req.model = str_of(Some(&v), "model")?.to_owned(),
            "messages" => messages::fold(&v, &mut req)?,
            "tools" => req.tools = tools(&v)?,
            "tool_choice" => req.tool_choice = tool_choice(&v)?,
            "parallel_tool_calls" => {
                req.parallel_tool_calls = Some(v.as_bool().ok_or_else(|| bad(&k, "a boolean"))?);
            }
            // Both wire spellings are the ONE canonical fact; the egress encoder
            // re-picks the key from the `reasoning` signal (openai-chat-mapping §2.7).
            "max_tokens" | "max_completion_tokens" => req.max_tokens = Some(u32_of(&v, &k)?),
            "temperature" => req.temperature = Some(f32_of(&v, &k)?),
            "top_p" => req.top_p = Some(f32_of(&v, &k)?),
            "reasoning_effort" => req.reasoning = Some(effort(&v)?),
            "stop" => req.stop = stop(&v)?,
            "stream" => req.stream = Some(v.as_bool().ok_or_else(|| bad(&k, "a boolean"))?),
            "response_format" => req.output = output(&v)?,
            // The long-tail valve (arch §3.1): unknown top-level keys — including the
            // client's `stream_options`, kept for the response encoder's shape
            // decision — forward verbatim, never rejected (ingress.md §2).
            _ => {
                req.extra.insert(k, v);
            }
        }
    }
    Ok(req)
}

/// A `u32` wire field (the token bound); floats and negatives are shapeless here.
fn u32_of(v: &Value, path: &str) -> Result<u32, IngressError> {
    v.as_u64()
        .and_then(|n| u32::try_from(n).ok())
        .ok_or_else(|| bad(path, "an unsigned 32-bit integer"))
}

/// An `f32` wire field (the sampling knobs).
fn f32_of(v: &Value, path: &str) -> Result<f32, IngressError> {
    v.as_f64()
        .map(|f| f as f32)
        .ok_or_else(|| bad(path, "a number"))
}

/// `reasoning_effort` → the closed canonical effort (providers §6). An effort the
/// canonical model cannot represent has no projection → rung 4.
fn effort(v: &Value) -> Result<ReasoningEffort, IngressError> {
    str_of(Some(v), "reasoning_effort")?
        .parse()
        .map_err(|()| bad("reasoning_effort", "one of \"low\" | \"medium\" | \"high\""))
}

/// `stop`: the wire allows a bare string or an array; canonically always the array
/// (the string-vs-list distinction dies at decode, arch §3.1).
fn stop(v: &Value) -> Result<Vec<String>, IngressError> {
    match v {
        Value::String(s) => Ok(vec![s.clone()]),
        Value::Array(a) => a
            .iter()
            .enumerate()
            .map(|(i, s)| Ok(str_of(Some(s), &format!("stop[{i}]"))?.to_owned()))
            .collect(),
        _ => Err(bad("stop", "a string or an array of strings")),
    }
}

/// `tools[]` → `Tool::Custom` (openai-chat-mapping §2.5, inverted), including the
/// lifted `function.strict` knob. Only `"function"` tools have a canonical
/// projection; a missing `type` is tolerated as the unambiguous default.
fn tools(v: &Value) -> Result<Vec<Tool>, IngressError> {
    let mut out = Vec::new();
    for (i, t) in arr_of(Some(v), "tools")?.iter().enumerate() {
        let path = format!("tools[{i}]");
        let obj = obj_of(Some(t), &path)?;
        if let Some(ty) = obj.get("type") {
            if str_of(Some(ty), &format!("{path}.type"))? != "function" {
                return Err(err(format!(
                    "`{path}.type` has no canonical projection: only \"function\" tools are representable"
                )));
            }
        }
        let fp = format!("{path}.function");
        let f = obj_of(obj.get("function"), &fp)?;
        out.push(Tool::Custom {
            name: str_of(f.get("name"), &format!("{fp}.name"))?.to_owned(),
            description: opt_str(f.get("description"), &format!("{fp}.description"))?,
            input_schema: f.get("parameters").cloned().unwrap_or(Value::Null),
            strict: opt_bool(f.get("strict"), &format!("{fp}.strict"))?,
        });
    }
    Ok(out)
}

/// `tool_choice` spellings, inverted (§2.6): `"auto"` → `Auto` (also the default
/// when the key is absent), `"required"` → `Any`, `"none"` → `None`, and the named
/// function object → `Tool{name}`.
fn tool_choice(v: &Value) -> Result<ToolChoice, IngressError> {
    match v {
        Value::String(s) => match s.as_str() {
            "auto" => Ok(ToolChoice::Auto),
            "required" => Ok(ToolChoice::Any),
            "none" => Ok(ToolChoice::None),
            other => Err(err(format!(
                "`tool_choice` \"{other}\" has no canonical projection"
            ))),
        },
        Value::Object(o) => {
            if str_of(o.get("type"), "tool_choice.type")? != "function" {
                return Err(err(
                    "`tool_choice.type` has no canonical projection: only \"function\" is representable",
                ));
            }
            let f = obj_of(o.get("function"), "tool_choice.function")?;
            Ok(ToolChoice::Tool {
                name: str_of(f.get("name"), "tool_choice.function.name")?.to_owned(),
            })
        }
        _ => Err(bad("tool_choice", "a string or an object")),
    }
}

/// `response_format` → the portable `output` knob (§2.5.1, inverted): `"text"` is
/// the wire's spelling of the default → `None`; `json_object` → JSON mode;
/// `json_schema` unwraps chat's nested `{name, schema, strict}` object.
fn output(v: &Value) -> Result<Option<OutputFormat>, IngressError> {
    let o = obj_of(Some(v), "response_format")?;
    match str_of(o.get("type"), "response_format.type")? {
        "text" => Ok(None),
        "json_object" => Ok(Some(OutputFormat::Json)),
        "json_schema" => {
            let js = obj_of(o.get("json_schema"), "response_format.json_schema")?;
            Ok(Some(OutputFormat::JsonSchema {
                name: opt_str(js.get("name"), "response_format.json_schema.name")?,
                schema: js.get("schema").cloned().unwrap_or(Value::Null),
                strict: opt_bool(js.get("strict"), "response_format.json_schema.strict")?,
            }))
        }
        other => Err(err(format!(
            "`response_format.type` \"{other}\" has no canonical projection"
        ))),
    }
}