mermaid_cli/providers/model/
anthropic.rs

1//! Anthropic provider — wraps `models::adapters::anthropic::AnthropicAdapter`.
2//!
3//! Same pattern as `ollama.rs`: the adapter handles the wire format
4//! (cache_control blocks, extended-thinking signature round-trip);
5//! this wrapper plumbs `ChatRequest` / `StreamContext` into it.
6//!
7//! Anthropic is the one provider that emits a `provider_continuation`
8//! that MUST round-trip on the next request. The adapter's
9//! `ModelResponse.provider_continuation` already carries it; we forward
10//! that onto the `FinalResponse` so the reducer can commit it via
11//! `ChatMessage::with_provider_continuation`.
12
13use async_trait::async_trait;
14
15use crate::domain::ChatRequest;
16use crate::models::adapters::anthropic::AnthropicAdapter;
17use crate::models::{Model, ModelConfig, ModelError, Result};
18
19use super::super::capabilities::Capabilities;
20use super::super::ctx::{FinalResponse, StreamContext, StreamEvent};
21use super::{ContextSizing, ModelProvider, resolve_limits_cached};
22
23/// Anthropic's Messages-API root, and the env var its key lives in.
24///
25/// One definition, next to the provider, like `meta`'s. `ProviderFactory`
26/// builds the endpoint from these and `providers::discovery` reports it from
27/// the same constants — the literal used to be spelled out separately in both.
28pub const DEFAULT_BASE_URL: &str = "https://api.anthropic.com/v1";
29pub const DEFAULT_API_KEY_ENV: &str = "ANTHROPIC_API_KEY";
30
31/// Anthropic adapter fronted by `ModelProvider`.
32pub struct AnthropicProvider {
33    adapter: AnthropicAdapter,
34    capabilities: Capabilities,
35}
36
37impl AnthropicProvider {
38    pub fn new(api_key: String, model_name: String, base_url: String) -> Result<Self> {
39        let adapter = AnthropicAdapter::new(api_key, model_name, base_url)?;
40        let capabilities =
41            Capabilities::from_legacy(adapter.capabilities()).with_provider_continuation();
42        Ok(Self {
43            adapter,
44            capabilities,
45        })
46    }
47}
48
49#[async_trait]
50impl ModelProvider for AnthropicProvider {
51    fn capabilities(&self) -> &Capabilities {
52        &self.capabilities
53    }
54
55    /// Live limit discovery via Anthropic's Models API (`GET /v1/models/
56    /// {id}` → `max_input_tokens` window + `max_tokens` output ceiling).
57    /// Cache-first via `provider_probes` (TTL-bounded), one live fetch on a
58    /// miss; a fetch failure resolves all-`None` (adapter floors apply).
59    async fn resolve_context_window(&self, request: &ChatRequest) -> ContextSizing {
60        let _ = request;
61        let model = Model::name(&self.adapter).to_string();
62        let limits =
63            resolve_limits_cached("anthropic", &model, || self.adapter.fetch_model_limits()).await;
64        let window = limits.as_ref().and_then(|l| l.max_context_tokens);
65        ContextSizing {
66            model_max: window,
67            effective: window,
68            source: None,
69            max_output: limits.as_ref().and_then(|l| l.max_output_tokens),
70        }
71    }
72
73    async fn chat(&self, request: ChatRequest, ctx: StreamContext) -> Result<FinalResponse> {
74        let config = build_model_config(&request);
75        // F2: ordered relay — see stream_bridge docs.
76        let (relay_tx, relay_handle) = super::stream_bridge::ordered_relay(ctx.sink.clone());
77        let callback = super::stream_bridge::forward_callback(relay_tx.clone());
78        let chat_fut = self
79            .adapter
80            .chat(&request.messages, &config, Some(callback));
81
82        let response = tokio::select! {
83            biased;
84            _ = ctx.token.cancelled() => {
85                return Err(ModelError::Cancelled);
86            },
87            r = chat_fut => r?,
88        };
89
90        let usage = response.usage.clone();
91        let provider_continuation = response.provider_continuation.clone();
92        let stop_reason = response.stop_reason.clone();
93        // Terminal Done through the ordered relay, then drain (see stream_bridge).
94        let _ = relay_tx.send(StreamEvent::Done {
95            usage: usage.clone(),
96            provider_continuation: provider_continuation.clone(),
97            stop_reason: stop_reason.clone(),
98        });
99        drop(relay_tx);
100        crate::utils::join_logged(relay_handle.take(), "stream_relay").await;
101
102        Ok(FinalResponse {
103            usage,
104            provider_continuation,
105            tool_calls: response.tool_calls.unwrap_or_default(),
106            stop_reason,
107        })
108    }
109}
110
111fn build_model_config(request: &ChatRequest) -> ModelConfig {
112    ModelConfig {
113        model: request.model_id.clone(),
114        temperature: request.temperature,
115        max_tokens: request.max_tokens,
116        reasoning: request.reasoning,
117        system_prompt: Some(request.system_prompt.clone()),
118        dynamic_system_suffix: request.instructions.clone(),
119        tools: request.tools.iter().map(|t| t.to_openai_json()).collect(),
120        resolved_context_window: request.resolved_context_window,
121        resolved_max_output: request.resolved_max_output,
122        // The adapter maps this to `output_config.format` (native
123        // structured output); client-side validation stays the final gate.
124        output_schema: request.output_schema.clone(),
125        ..Default::default()
126    }
127}
128
129#[cfg(test)]
130mod tests {
131    use super::*;
132
133    #[test]
134    fn build_model_config_maps_fields() {
135        let req = ChatRequest {
136            model_id: "anthropic/claude-opus-4-7".to_string(),
137            messages: vec![],
138            system_prompt: "sys".to_string(),
139            instructions: Some("MERMAID.md content".to_string()),
140            reasoning: crate::models::ReasoningLevel::XHigh,
141            temperature: 0.7,
142            max_tokens: 8192,
143            tools: vec![],
144
145            ollama_num_ctx: None,
146            ollama_allow_ram_offload: None,
147            resolved_context_window: None,
148            resolved_max_output: None,
149            output_schema: None,
150            suppress_auto_compact: false,
151            suppressed_builtin_tools: Vec::new(),
152        };
153        let cfg = build_model_config(&req);
154        assert_eq!(cfg.reasoning, crate::models::ReasoningLevel::XHigh);
155        assert_eq!(cfg.max_tokens, 8192);
156        assert_eq!(
157            cfg.dynamic_system_suffix.as_deref(),
158            Some("MERMAID.md content")
159        );
160    }
161}
mermaid_cli/providers/model/anthropic.rs

mermaid_cli/providers/model/
anthropic.rs
