brazen 0.0.10

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
161
162
163
164
165
166
167
168

//! The canonical-content projections of a Responses request (§3.2/§3.3): `system` →
//! the top-level `instructions` string, and `messages[]` → the typed `input[]` array
//! (`message` / `reasoning` / `function_call` / `function_call_output` items, with the
//! image and document parts each role admits). This is the whole of encode that reads
//! `Content` — `super::encode` calls [`instructions`] and [`input_value`], and the
//! text-only slot rejection (`slot_err`) lives here since only these projections use
//! it. The tool-argument string encoding is the shared `protocol::json::to_json_string`.

use serde_json::{json, Value};

use crate::canonical::{
    CanonicalError, CanonicalRequest, Content, DocumentSource, ErrorKind, ImageSource, Message,
    Role,
};
use crate::protocol::json::to_json_string;

/// A text-only wire slot rejected non-text content (§3.2/§3.3).
fn slot_err(slot: &str) -> CanonicalError {
    CanonicalError {
        kind: ErrorKind::ParseInput,
        message: format!("{slot} accepts only text content"),
        provider_detail: None,
        retry_after_seconds: None,
    }
}

/// `system` → the top-level `instructions` string (§3.2): text-only, `None` when
/// empty. `Role::System` messages stay distinct in `input[]` (§3.3).
pub(super) fn instructions(req: &CanonicalRequest) -> Result<Option<String>, CanonicalError> {
    let Some(system) = req.system.as_ref().filter(|s| !s.is_empty()) else {
        return Ok(None);
    };
    let mut text = String::new();
    for c in system {
        match c {
            Content::Text(t) => text.push_str(t),
            _ => return Err(slot_err("instructions")),
        }
    }
    Ok(Some(text))
}

/// Project `messages[]` to the typed `input[]` (§3.3): each message yields a
/// `message` item for its text/image parts plus standalone `function_call` /
/// `function_call_output` items for tool use/results.
pub(super) fn input_value(req: &CanonicalRequest) -> Result<Value, CanonicalError> {
    let mut items = Vec::new();
    for m in &req.messages {
        message_items(m, &mut items)?;
    }
    Ok(Value::Array(items))
}

/// One `Message` → its `input[]` items (§3.3). A `Role::Tool` carries only
/// `function_call_output` items; other roles fold text/image into one `message`
/// item, hoisting `ToolUse` to standalone `function_call` items (kept in order).
fn message_items(m: &Message, items: &mut Vec<Value>) -> Result<(), CanonicalError> {
    let (role, text_type) = match m.role {
        Role::User => ("user", "input_text"),
        Role::System => ("system", "input_text"),
        Role::Assistant => ("assistant", "output_text"),
        Role::Tool => {
            for c in &m.content {
                items.push(function_call_output(c)?);
            }
            return Ok(());
        }
    };
    let mut content = Vec::new();
    let mut reasonings = Vec::new();
    let mut calls = Vec::new();
    for c in &m.content {
        match c {
            Content::Text(t) => content.push(json!({ "type": text_type, "text": t })),
            Content::Image { source } if role == "user" => content.push(input_image(source)),
            Content::Document { source } if role == "user" => content.push(input_file(source)),
            Content::ToolUse {
                id, name, input, ..
            } => calls.push(json!({
                "type": "function_call", "call_id": id, "name": name,
                "arguments": to_json_string(input),
            })),
            // A reasoning item replays ONLY when its encrypted_content is present (the
            // stateless store:false path) — a bare summary cannot be replayed (§3.3, bl-61a9).
            Content::Thinking {
                text,
                id,
                encrypted_content: Some(enc),
                ..
            } => reasonings.push(reasoning_item(text, id.as_deref(), enc)),
            Content::Thinking { .. } | Content::RedactedThinking { .. } => {} // dropped (§3.3)
            _ => return Err(slot_err(role)),
        }
    }
    items.extend(reasonings); // reasoning precedes the message/function_call it reasoned about
    if !content.is_empty() {
        items.push(json!({ "type": "message", "role": role, "content": content }));
    }
    items.extend(calls);
    Ok(())
}

/// A `reasoning` input item for stateless (`store:false`) replay (§3.3, bl-61a9): the
/// `encrypted_content` blob IS the reasoning state; the `id` (`rs_…`) is echoed when
/// present; the `summary` carries `text` when non-empty, else `[]`.
fn reasoning_item(text: &str, id: Option<&str>, encrypted_content: &str) -> Value {
    let summary = if text.is_empty() {
        json!([])
    } else {
        json!([{ "type": "summary_text", "text": text }])
    };
    let mut item = json!({
        "type": "reasoning", "summary": summary, "encrypted_content": encrypted_content,
    });
    if let Some(id) = id {
        item["id"] = json!(id);
    }
    item
}

/// `ToolResult` → a `function_call_output` item (§3.3): text-only `output`, keyed by
/// `call_id`. `is_error` surfaces textually (prefix); non-`Text` content rejects.
fn function_call_output(c: &Content) -> Result<Value, CanonicalError> {
    let Content::ToolResult {
        tool_use_id,
        content,
        is_error,
    } = c
    else {
        return Err(slot_err("tool"));
    };
    let mut text = String::new();
    for part in content {
        match part {
            Content::Text(t) => text.push_str(t),
            _ => return Err(slot_err("tool_result")),
        }
    }
    if *is_error {
        text = format!("[error] {text}");
    }
    Ok(json!({ "type": "function_call_output", "call_id": tool_use_id, "output": text }))
}

/// `Image` source → a Responses `input_image` part (§3.3): base64 embeds as a
/// data-URI (round-trips, as Chat Completions); a URL passes through.
fn input_image(source: &ImageSource) -> Value {
    let url = match source {
        ImageSource::Base64 { media_type, data } => format!("data:{media_type};base64,{data}"),
        ImageSource::Url { url } => url.clone(),
    };
    json!({ "type": "input_image", "image_url": url })
}

/// `Document` source → a Responses `input_file` part (§3.3): base64 embeds as a data-URI
/// in `file_data` (with a `filename` synthesized from the media type, required for
/// `file_data`); a URL passes through as `file_url` — Responses fetches web URLs, so BOTH
/// sources express here (unlike Chat, which rejects the URL, §6 CR-C6).
fn input_file(source: &DocumentSource) -> Value {
    match source {
        DocumentSource::Base64 { media_type, data } => json!({
            "type": "input_file",
            "filename": format!("document.{}", media_type.rsplit('/').next().unwrap_or("bin")),
            "file_data": format!("data:{media_type};base64,{data}"),
        }),
        DocumentSource::Url { url } => json!({ "type": "input_file", "file_url": url }),
    }
}