feat(providers): OpenAI-compatible provider types as runtime data (providers.yaml)

Replace the five copy-paste OpenAI-compatible provider structs (moonshot,
moonshot_code, deepseek, zai, lm_studio) with one DeclaredProvider engine
driven by a providers.yaml catalog loaded at boot from the cwd — edit and
restart, no rebuild. The YAML carries identity, endpoints, per-model JSON
field mapping, id-glob enrichment rules and the reasoning knob (effort /
thinking request kinds); capability_flags keep vision and future input
modalities declarative. Anthropic, Ollama, OpenAI and OpenRouter stay
native (different wire protocols or bespoke parsing) and register
alongside; colliding declared ids are skipped. The shipped catalog is
validated by a unit test.
This commit is contained in:
2026-07-18 16:19:05 +01:00
parent 760dae06e5
commit 4fea04c57f
11 changed files with 1005 additions and 543 deletions
+1
View File
@@ -42,6 +42,7 @@ uuid = { version = "1", features = ["v4"] }
reqwest = { version = "0.13.4", default-features = false, features = ["rustls-no-provider", "charset", "http2", "system-proxy", "json", "multipart"] }
async-trait = "0.1"
serde_json = "1"
serde_yaml = "0.9"
indexmap = { version = "2", features = ["serde"] }
tracing = "0.1"
chrono = { version = "0.4", default-features = false, features = ["clock", "std"] }
@@ -0,0 +1,819 @@
//! Declarative OpenAI-compatible LLM providers, loaded at boot from a runtime
//! YAML file ([`PROVIDERS_FILE`], resolved against the process working
//! directory — the same convention as `config.yml`). Editing the file and
//! restarting picks up changes without a rebuild; a provider entry that fails
//! validation is logged and skipped, never fatal.
//!
//! One engine, [`DeclaredProvider`], implements [`ApiProvider`] for every
//! entry: the YAML carries identity, endpoints, per-model JSON field mapping,
//! id-prefix enrichment rules and the reasoning knob, while providers whose
//! behavior is not OpenAI-compatible (Anthropic, Ollama) or too bespoke
//! (OpenRouter's pricing/reasoning parsing, OpenAI's extra TTS/transcribe
//! services) stay native Rust and register alongside.
use std::collections::HashMap;
use std::path::Path;
use std::sync::Arc;
use anyhow::{anyhow, Context, Result};
use tracing::{info, warn};
use crate::chatbot::openai::OpenAiClient;
use crate::llm::providers::{extra_with_reasoning, RemoteLlmModelInfo};
use crate::llm::{LlmModelRecord, LlmProviderRecord};
use crate::provider::{
ApiProvider, BuiltLlmClient, ProviderField, ProviderUiMeta, ReasoningMode, ServiceType,
};
/// Runtime catalog of declarative providers, relative to the process cwd.
pub const PROVIDERS_FILE: &str = "providers.yaml";
const LLM_ONLY: &[ServiceType] = &[ServiceType::Llm];
// ── YAML spec ────────────────────────────────────────────────────────────────
#[derive(Debug, serde::Deserialize)]
struct ProvidersFile {
#[serde(default)]
providers: Vec<serde_yaml::Value>,
}
#[derive(Debug, serde::Deserialize)]
struct ProviderSpec {
id: String,
name: String,
/// Only `openai_compatible` exists today; the key is accepted (and
/// validated) so the file format can grow other kinds later.
kind: Option<String>,
base_url: String,
/// When true, the instance-level `base_url` stored in the DB overrides
/// `base_url` (local providers like LM Studio).
#[serde(default)]
base_url_overridable: bool,
#[serde(default)]
api_key: ApiKeySpec,
#[serde(default)]
prompt_cache: bool,
ui: UiSpec,
#[serde(default)]
fields: Vec<FieldSpec>,
models: Option<ModelsSpec>,
reasoning: Option<ReasoningSpec>,
}
#[derive(Debug, Default, serde::Deserialize)]
#[serde(rename_all = "snake_case")]
enum ApiKeySpec {
#[default]
Required,
Optional,
None,
}
#[derive(Debug, serde::Deserialize)]
struct UiSpec {
color: String,
icon: String,
description: Option<String>,
}
#[derive(Debug, serde::Deserialize)]
struct FieldSpec {
key: String,
label: String,
#[serde(default)]
required: bool,
#[serde(default)]
secret: bool,
}
#[derive(Debug, Default, serde::Deserialize)]
struct ModelsSpec {
/// GET path joined to `models_url`; the response must be the OpenAI
/// `{ "data": [...] }` envelope. Mutually exclusive with `static`.
endpoint: Option<String>,
/// Base for the listing endpoint; defaults to the resolved `base_url`.
models_url: Option<String>,
/// Defaults to `bearer` unless `api_key: none`.
auth: Option<AuthSpec>,
/// Static model-id catalog (provider exposes no listing endpoint).
#[serde(rename = "static")]
static_models: Option<Vec<String>>,
#[serde(default)]
map: MapSpec,
#[serde(default)]
base_capabilities: Vec<String>,
#[serde(default)]
defaults: DefaultsSpec,
/// First-matching rule wins.
#[serde(default)]
enrich: Vec<EnrichRule>,
}
#[derive(Debug, serde::Deserialize)]
#[serde(rename_all = "snake_case")]
enum AuthSpec {
Bearer,
None,
}
/// Per-model JSON field names → `RemoteLlmModelInfo` fields. Absent mappings
/// leave the corresponding field `None` (id defaults to `"id"`, name to id).
#[derive(Debug, Default, serde::Deserialize)]
struct MapSpec {
id: Option<String>,
name: Option<String>,
context_length: Option<String>,
max_completion_tokens: Option<String>,
knowledge_cutoff: Option<String>,
/// Boolean JSON field; when true it also adds the `vision` capability.
vision: Option<String>,
price_input_per_million: Option<String>,
price_output_per_million: Option<String>,
/// capability name → boolean JSON field that enables it.
#[serde(default)]
capability_flags: HashMap<String, String>,
}
#[derive(Debug, Default, serde::Deserialize)]
struct DefaultsSpec {
vision: Option<bool>,
}
#[derive(Debug, serde::Deserialize)]
struct EnrichRule {
#[serde(rename = "match")]
glob: String,
#[serde(default)]
mode: EnrichMode,
context_length: Option<u64>,
max_completion_tokens: Option<u64>,
vision: Option<bool>,
#[serde(default)]
add_capabilities: Vec<String>,
}
#[derive(Debug, Default, serde::Deserialize)]
#[serde(rename_all = "snake_case")]
enum EnrichMode {
/// Only fill fields the endpoint/static list left unset (endpoint wins).
#[default]
Fill,
/// Overwrite whatever the endpoint returned (rule wins).
Override,
}
#[derive(Debug, serde::Deserialize)]
struct ReasoningSpec {
request: ReasoningRequestSpec,
/// First-matching rule wins.
#[serde(default)]
modes: Vec<ReasoningModeRule>,
}
#[derive(Debug, serde::Deserialize)]
#[serde(tag = "kind", rename_all = "snake_case")]
enum ReasoningRequestSpec {
/// `{"reasoning_effort": value}`, with optional value remapping
/// (e.g. `disabled` → `none`).
Effort {
#[serde(default)]
remap: HashMap<String, String>,
},
/// `disabled`/`enabled` toggle via `{"thinking": {...}}`; any other value
/// also carries `reasoning_effort`.
Thinking,
}
#[derive(Debug, serde::Deserialize)]
struct ReasoningModeRule {
#[serde(default)]
when: WhenSpec,
values: Vec<String>,
default: Option<String>,
}
/// Selector for a reasoning mode: with both keys present the model matches on
/// either (OR); with neither, it matches every model.
#[derive(Debug, Default, serde::Deserialize)]
struct WhenSpec {
models: Option<Vec<String>>,
capability: Option<String>,
}
impl WhenSpec {
fn matches(&self, model_id: &str, capabilities: &[String]) -> bool {
let id_ok = self
.models
.as_ref()
.is_some_and(|gs| gs.iter().any(|g| glob_match(g, model_id)));
let cap_ok = self
.capability
.as_ref()
.is_some_and(|c| capabilities.iter().any(|x| x == c));
match (self.models.is_some(), self.capability.is_some()) {
(false, false) => true,
_ => id_ok || cap_ok,
}
}
}
// ── DeclaredProvider ─────────────────────────────────────────────────────────
/// [`ApiProvider`] strings are `&'static str`, but the spec is runtime data:
/// the handful of strings a provider exposes are leaked once at boot. The set
/// of declared providers is fixed for the process lifetime and tiny, so the
/// leak is bounded and intentional.
struct LeakedMeta {
id: &'static str,
name: &'static str,
description: Option<&'static str>,
color: &'static str,
icon: &'static str,
fields: &'static [ProviderField],
}
fn leak_str(s: &str) -> &'static str {
Box::leak(s.to_string().into_boxed_str())
}
pub struct DeclaredProvider {
spec: ProviderSpec,
meta: LeakedMeta,
/// Built on first use: with the `rustls-no-provider` reqwest feature a
/// `Client` can only be constructed after the process installs a crypto
/// provider (done by the shell at startup), so building one per provider
/// at registration time would panic outside the server binary (tests).
http: std::sync::OnceLock<reqwest::Client>,
}
impl DeclaredProvider {
fn new(spec: ProviderSpec) -> Self {
let fields: Vec<ProviderField> = spec
.fields
.iter()
.map(|f| ProviderField {
key: leak_str(&f.key),
label: leak_str(&f.label),
required: f.required,
secret: f.secret,
})
.collect();
let meta = LeakedMeta {
id: leak_str(&spec.id),
name: leak_str(&spec.name),
description: spec.ui.description.as_deref().map(leak_str),
color: leak_str(&spec.ui.color),
icon: leak_str(&spec.ui.icon),
fields: Box::leak(fields.into_boxed_slice()),
};
Self { spec, meta, http: std::sync::OnceLock::new() }
}
fn http(&self) -> &reqwest::Client {
self.http.get_or_init(reqwest::Client::new)
}
fn base_url(&self, record: &LlmProviderRecord) -> String {
if self.spec.base_url_overridable {
record
.base_url
.clone()
.unwrap_or_else(|| self.spec.base_url.clone())
} else {
self.spec.base_url.clone()
}
}
fn models_url(&self, models: &ModelsSpec, record: &LlmProviderRecord) -> String {
models
.models_url
.clone()
.unwrap_or_else(|| self.base_url(record))
}
/// Bearer key for the listing endpoint, or `None` when the entry is
/// unauthenticated. Errors when auth is required but the instance has no key.
fn auth_key(&self, models: &ModelsSpec, record: &LlmProviderRecord) -> Result<Option<String>> {
let bearer = match models.auth {
Some(AuthSpec::Bearer) => true,
Some(AuthSpec::None) => false,
None => !matches!(self.spec.api_key, ApiKeySpec::None),
};
if !bearer {
return Ok(None);
}
let key = record.api_key.clone().ok_or_else(|| {
anyhow!(
"provider '{}': api_key required for {} model listing",
record.name,
self.meta.name
)
})?;
Ok(Some(key))
}
fn blank_info(&self, id: &str, models: &ModelsSpec) -> RemoteLlmModelInfo {
RemoteLlmModelInfo {
id: id.to_string(),
name: id.to_string(),
context_length: None,
max_completion_tokens: None,
knowledge_cutoff: None,
capabilities: models.base_capabilities.clone(),
vision: models.defaults.vision,
price_input_per_million: None,
price_output_per_million: None,
reasoning: None,
}
}
fn map_model(&self, m: &serde_json::Value, models: &ModelsSpec) -> Option<RemoteLlmModelInfo> {
let map = &models.map;
let get = |f: &Option<String>| f.as_deref().map(|k| &m[k]);
let id = get(&map.id)
.or_else(|| Some(&m["id"]))
.and_then(|v| v.as_str())?
.to_string();
let name = get(&map.name)
.and_then(|v| v.as_str())
.unwrap_or(&id)
.to_string();
let mut vision = get(&map.vision).and_then(|v| v.as_bool());
if vision.is_none() {
vision = models.defaults.vision;
}
let mut capabilities = models.base_capabilities.clone();
let mut add_cap = |cap: &str| {
if !capabilities.iter().any(|c| c == cap) {
capabilities.push(cap.to_string());
}
};
if vision == Some(true) {
add_cap("vision");
}
for (cap, field) in &map.capability_flags {
if m[field].as_bool().unwrap_or(false) {
add_cap(cap);
}
}
Some(RemoteLlmModelInfo {
id,
name,
context_length: get(&map.context_length).and_then(|v| v.as_u64()),
max_completion_tokens: get(&map.max_completion_tokens).and_then(|v| v.as_u64()),
knowledge_cutoff: get(&map.knowledge_cutoff)
.and_then(|v| v.as_str())
.map(String::from),
capabilities,
vision,
price_input_per_million: get(&map.price_input_per_million).and_then(|v| v.as_f64()),
price_output_per_million: get(&map.price_output_per_million).and_then(|v| v.as_f64()),
reasoning: None,
})
}
async fn list_models(&self, record: &LlmProviderRecord) -> Result<Vec<RemoteLlmModelInfo>> {
let models = self.spec.models.as_ref().expect("checked by caller");
let mut list = if let Some(ids) = &models.static_models {
ids.iter().map(|id| self.blank_info(id, models)).collect()
} else {
let endpoint = models.endpoint.as_deref().expect("validated");
let base = self.models_url(models, record);
let url = format!("{}{}", base.trim_end_matches('/'), endpoint);
let key = self.auth_key(models, record)?;
let mut req = self.http().get(&url);
if let Some(k) = key {
req = req.bearer_auth(k);
}
let who = self.meta.name;
let resp: serde_json::Value = req
.send()
.await
.map_err(|e| anyhow!("{who} request failed: {e}"))?
.error_for_status()
.map_err(|e| anyhow!("{who} error response: {e}"))?
.json()
.await
.map_err(|e| anyhow!("{who} response parse failed: {e}"))?;
let raw = resp["data"]
.as_array()
.cloned()
.ok_or_else(|| anyhow!("unexpected {who} response shape"))?;
raw.iter().filter_map(|m| self.map_model(m, models)).collect()
};
for info in &mut list {
apply_enrich(&models.enrich, info);
}
Ok(list)
}
}
/// Applies the first matching enrich rule (later rules are not consulted).
fn apply_enrich(rules: &[EnrichRule], info: &mut RemoteLlmModelInfo) {
let Some(rule) = rules.iter().find(|r| glob_match(&r.glob, &info.id)) else {
return;
};
let set = |dst: &mut Option<u64>, v: Option<u64>| {
if let Some(v) = v {
match rule.mode {
EnrichMode::Fill => {
if dst.is_none() {
*dst = Some(v);
}
}
EnrichMode::Override => *dst = Some(v),
}
}
};
set(&mut info.context_length, rule.context_length);
set(&mut info.max_completion_tokens, rule.max_completion_tokens);
if let Some(v) = rule.vision {
match rule.mode {
EnrichMode::Fill => {
if info.vision.is_none() {
info.vision = Some(v);
}
}
EnrichMode::Override => info.vision = Some(v),
}
}
for cap in &rule.add_capabilities {
if !info.capabilities.iter().any(|c| c == cap) {
info.capabilities.push(cap.clone());
}
}
}
/// Case-insensitive glob where `*` is the only wildcard (any run, empty
/// included): `"k3*"` is a prefix match, `"*reasoner*"` a contains match, and
/// a pattern without `*` is an exact match.
fn glob_match(pattern: &str, text: &str) -> bool {
let p = pattern.to_lowercase();
let t = text.to_lowercase();
if !p.contains('*') {
return p == t;
}
let anchored_start = !p.starts_with('*');
let anchored_end = !p.ends_with('*');
let parts: Vec<&str> = p.split('*').filter(|s| !s.is_empty()).collect();
if parts.is_empty() {
return true;
}
let mut rest = t.as_str();
for (i, part) in parts.iter().enumerate() {
if i == 0 && anchored_start {
if !rest.starts_with(part) {
return false;
}
rest = &rest[part.len()..];
continue;
}
match rest.find(part) {
Some(pos) => rest = &rest[pos + part.len()..],
None => return false,
}
}
if anchored_end {
return t.ends_with(parts[parts.len() - 1]);
}
true
}
#[async_trait::async_trait]
impl ApiProvider for DeclaredProvider {
fn type_id(&self) -> &'static str {
self.meta.id
}
fn display_name(&self) -> &'static str {
self.meta.name
}
fn supported_types(&self) -> &'static [ServiceType] {
LLM_ONLY
}
async fn list_llm_models(
&self,
record: &LlmProviderRecord,
) -> Result<Option<Vec<RemoteLlmModelInfo>>> {
if self.spec.models.is_none() {
return Ok(None);
}
Ok(Some(self.list_models(record).await?))
}
fn reasoning_mode(&self, model_id: &str, capabilities: &[String]) -> Option<ReasoningMode> {
let spec = self.spec.reasoning.as_ref()?;
let rule = spec
.modes
.iter()
.find(|r| r.when.matches(model_id, capabilities))?;
Some(ReasoningMode::ValueSet {
values: rule.values.clone(),
default: rule.default.clone(),
})
}
fn reasoning_request(&self, value: &serde_json::Value) -> Option<serde_json::Value> {
let spec = self.spec.reasoning.as_ref()?;
let s = value.as_str()?;
match &spec.request {
ReasoningRequestSpec::Effort { remap } => {
let v = remap.get(s).map(String::as_str).unwrap_or(s);
Some(serde_json::json!({ "reasoning_effort": v }))
}
ReasoningRequestSpec::Thinking => match s {
"disabled" => Some(serde_json::json!({ "thinking": { "type": "disabled" } })),
"enabled" => Some(serde_json::json!({ "thinking": { "type": "enabled" } })),
effort => Some(serde_json::json!({
"thinking": { "type": "enabled" },
"reasoning_effort": effort,
})),
},
}
}
fn build_llm(
&self,
record: &LlmProviderRecord,
model: &LlmModelRecord,
) -> Option<Result<BuiltLlmClient>> {
Some((|| {
let key = match self.spec.api_key {
ApiKeySpec::Required => record.api_key.clone().with_context(|| {
format!(
"provider '{}': api_key required for {}",
record.name, self.meta.id
)
})?,
ApiKeySpec::Optional => record.api_key.clone().unwrap_or_default(),
ApiKeySpec::None => String::new(),
};
let extra = extra_with_reasoning(self, model);
let prompt_cache = self.spec.prompt_cache;
Ok(BuiltLlmClient {
client: Arc::new(OpenAiClient::new(self.base_url(record), key, extra, prompt_cache)),
prompt_cache,
})
})())
}
fn ui_meta(&self) -> ProviderUiMeta {
ProviderUiMeta {
type_id: self.meta.id,
display_name: self.meta.name,
description: self.meta.description,
color: self.meta.color,
icon: self.meta.icon,
lists_models: self.spec.models.is_some(),
fields: self.meta.fields,
}
}
}
// ── Loader ───────────────────────────────────────────────────────────────────
/// Loads every valid entry from `path`. A missing file, an unreadable file or
/// a malformed entry is logged and skipped — the native providers always
/// register regardless.
pub fn load(path: &Path) -> Vec<DeclaredProvider> {
let text = match std::fs::read_to_string(path) {
Ok(t) => t,
Err(e) if e.kind() == std::io::ErrorKind::NotFound => {
info!(path = %path.display(), "no declarative providers file — skipped");
return Vec::new();
}
Err(e) => {
warn!(path = %path.display(), error = %e, "cannot read declarative providers file — skipped");
return Vec::new();
}
};
let file: ProvidersFile = match serde_yaml::from_str(&text) {
Ok(f) => f,
Err(e) => {
warn!(path = %path.display(), error = %e, "invalid declarative providers file — skipped");
return Vec::new();
}
};
let mut seen = std::collections::HashSet::new();
let mut out = Vec::new();
for (i, entry) in file.providers.into_iter().enumerate() {
match parse_entry(entry) {
Ok(p) => {
if !seen.insert(p.meta.id) {
warn!(type_id = p.meta.id, "duplicate provider id in declarative file — skipped");
continue;
}
out.push(p);
}
Err(e) => {
warn!(path = %path.display(), entry = i, error = %e, "invalid provider entry — skipped");
}
}
}
info!(path = %path.display(), count = out.len(), "declarative providers loaded");
out
}
fn parse_entry(v: serde_yaml::Value) -> Result<DeclaredProvider> {
let spec: ProviderSpec = serde_yaml::from_value(v)?;
validate(&spec)?;
Ok(DeclaredProvider::new(spec))
}
fn validate(spec: &ProviderSpec) -> Result<()> {
if spec.id.trim().is_empty() {
return Err(anyhow!("provider id must not be empty"));
}
if let Some(kind) = &spec.kind
&& kind != "openai_compatible"
{
return Err(anyhow!(
"provider '{}': unsupported kind '{kind}' (only 'openai_compatible')",
spec.id
));
}
if spec.base_url.trim().is_empty() {
return Err(anyhow!("provider '{}': base_url must not be empty", spec.id));
}
if let Some(m) = &spec.models {
match (m.endpoint.is_some(), m.static_models.is_some()) {
(true, true) => {
return Err(anyhow!(
"provider '{}': models.endpoint and models.static are mutually exclusive",
spec.id
));
}
(false, false) => {
return Err(anyhow!(
"provider '{}': models needs either endpoint or static",
spec.id
));
}
_ => {}
}
}
Ok(())
}
// ── Tests ────────────────────────────────────────────────────────────────────
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn glob_prefix_contains_exact() {
assert!(glob_match("k3*", "K3-0718"));
assert!(glob_match("*coder*", "deepseek-Coder-v2"));
assert!(glob_match("*128k*", "glm-4-32b-0414-128k"));
assert!(glob_match("glm-5", "glm-5"));
assert!(!glob_match("glm-5", "glm-5-turbo"));
assert!(glob_match("a*b", "axbxb"));
assert!(!glob_match("a*b", "abx"));
assert!(glob_match("*b", "ab"));
assert!(!glob_match("k3*", "kimi-for-coding"));
}
fn provider(yaml: &str) -> DeclaredProvider {
parse_entry(serde_yaml::from_str(yaml).unwrap()).unwrap()
}
#[test]
fn thinking_request_mapping() {
let p = provider(
r#"
id: t
name: T
base_url: http://x
ui: { color: c, icon: i }
reasoning:
request: { kind: thinking }
modes:
- { values: [disabled, enabled], default: enabled }
"#,
);
let req = |v: &str| p.reasoning_request(&serde_json::json!(v)).unwrap();
assert_eq!(req("disabled"), serde_json::json!({ "thinking": { "type": "disabled" } }));
assert_eq!(req("enabled"), serde_json::json!({ "thinking": { "type": "enabled" } }));
assert_eq!(
req("high"),
serde_json::json!({ "thinking": { "type": "enabled" }, "reasoning_effort": "high" })
);
}
#[test]
fn effort_request_remap() {
let p = provider(
r#"
id: t
name: T
base_url: http://x
ui: { color: c, icon: i }
reasoning:
request: { kind: effort, remap: { disabled: none } }
modes:
- { values: [disabled, max], default: max }
"#,
);
let req = |v: &str| p.reasoning_request(&serde_json::json!(v)).unwrap();
assert_eq!(req("disabled"), serde_json::json!({ "reasoning_effort": "none" }));
assert_eq!(req("max"), serde_json::json!({ "reasoning_effort": "max" }));
}
#[test]
fn reasoning_mode_first_match_wins() {
let p = provider(
r#"
id: t
name: T
base_url: http://x
ui: { color: c, icon: i }
reasoning:
request: { kind: thinking }
modes:
- when: { models: ["glm-5.2*"] }
values: [disabled, max]
- when: { models: ["glm-5*"] }
values: [disabled, enabled]
"#,
);
let mode = |id: &str| p.reasoning_mode(id, &[]);
assert!(matches!(
mode("glm-5.2"),
Some(ReasoningMode::ValueSet { values, .. }) if values == ["disabled", "max"]
));
assert!(matches!(
mode("glm-5-turbo"),
Some(ReasoningMode::ValueSet { values, .. }) if values == ["disabled", "enabled"]
));
assert!(mode("glm-4").is_none());
}
#[test]
fn reasoning_mode_capability_or_models() {
let p = provider(
r#"
id: t
name: T
base_url: http://x
ui: { color: c, icon: i }
reasoning:
request: { kind: thinking }
modes:
- when: { capability: reasoning, models: ["*reasoner*"] }
values: [disabled, high]
"#,
);
assert!(p.reasoning_mode("anything", &["reasoning".to_string()]).is_some());
assert!(p.reasoning_mode("deepseek-reasoner", &[]).is_some());
assert!(p.reasoning_mode("deepseek-chat", &[]).is_none());
}
#[test]
fn enrich_first_match_fill_vs_override() {
let rules: Vec<EnrichRule> = serde_yaml::from_str(
r#"
- { match: "*coder*", context_length: 16384, mode: override }
- { match: "k3*", context_length: 1048576, vision: true }
"#,
)
.unwrap();
let mut info = RemoteLlmModelInfo {
id: "x-coder".into(),
name: "x".into(),
context_length: Some(999),
max_completion_tokens: None,
knowledge_cutoff: None,
capabilities: vec![],
vision: None,
price_input_per_million: None,
price_output_per_million: None,
reasoning: None,
};
apply_enrich(&rules, &mut info);
assert_eq!(info.context_length, Some(16384)); // override wins over endpoint
info.id = "k3-pro".into();
info.context_length = Some(999);
apply_enrich(&rules, &mut info);
assert_eq!(info.context_length, Some(999)); // fill keeps the endpoint value
assert_eq!(info.vision, Some(true));
}
/// The catalog shipped at the repository root must always parse: the file
/// is runtime data, but this test keeps a typo from reaching users.
#[test]
fn shipped_providers_yaml_is_valid() {
let text = include_str!("../../../../../providers.yaml");
let file: ProvidersFile = serde_yaml::from_str(text).unwrap();
assert!(!file.providers.is_empty());
let mut ids = std::collections::HashSet::new();
for entry in file.providers {
let p = parse_entry(entry).unwrap();
assert!(ids.insert(p.meta.id), "duplicate id {}", p.meta.id);
}
}
}
@@ -1,123 +0,0 @@
use anyhow::{Result, anyhow};
use crate::llm::{LlmModelRecord, LlmProviderRecord};
use crate::llm::providers::{RemoteLlmModelInfo, build_openai_llm, fetch_openai_models};
use crate::provider::{ApiProvider, BuiltLlmClient, ProviderField, ProviderUiMeta, ReasoningMode, ServiceType};
pub struct DeepSeekProvider {
http: reqwest::Client,
}
impl DeepSeekProvider {
pub fn new() -> Self {
Self { http: reqwest::Client::new() }
}
fn known_context_length(model_id: &str) -> Option<u64> {
let id = model_id.to_lowercase();
if id.contains("coder") { Some(16384) }
else if id.contains("reasoner") { Some(65536) }
else if id.starts_with("deepseek-v4") { Some(1_048_576) }
else if id.starts_with("deepseek-chat") || id.starts_with("deepseek-v3") { Some(65536) }
else { None }
}
fn known_max_output(model_id: &str) -> Option<u64> {
if model_id.to_lowercase().starts_with("deepseek-v4") { Some(393_216) } else { None }
}
fn known_capabilities(model_id: &str) -> Vec<String> {
let mut caps = vec!["function_calling".to_string()];
if model_id.to_lowercase().contains("reasoner") {
caps.push("reasoning".to_string());
}
caps
}
}
#[async_trait::async_trait]
impl ApiProvider for DeepSeekProvider {
fn type_id(&self) -> &'static str { "deepseek" }
fn display_name(&self) -> &'static str { "DeepSeek" }
fn supported_types(&self) -> &'static [ServiceType] {
&[ServiceType::Llm]
}
async fn list_llm_models(&self, record: &LlmProviderRecord) -> Result<Option<Vec<RemoteLlmModelInfo>>> {
let api_key = record.api_key.as_deref()
.ok_or_else(|| anyhow!("provider '{}': api_key required for deepseek model listing", record.name))?;
let raw = fetch_openai_models(&self.http, "https://api.deepseek.com", Some(api_key), "DeepSeek").await?;
let models = raw
.iter()
.filter_map(|m| {
let id = m["id"].as_str()?.to_string();
let name = id.clone();
let context_length = Self::known_context_length(&id).or_else(|| m["context_length"].as_u64());
let capabilities = Self::known_capabilities(&id);
let max_output = Self::known_max_output(&id);
Some(RemoteLlmModelInfo {
id, name, context_length,
max_completion_tokens: max_output,
knowledge_cutoff: None,
capabilities,
vision: None,
price_input_per_million: None,
price_output_per_million: None,
reasoning: None,
})
})
.collect();
Ok(Some(models))
}
fn reasoning_mode(&self, model_id: &str, capabilities: &[String]) -> Option<ReasoningMode> {
// Thinking mode (thinking.type) + graded reasoning_effort. "disabled"
// turns thinking off; effort levels low/medium map to high, xhigh to max.
let id = model_id.to_lowercase();
if capabilities.iter().any(|c| c == "reasoning")
|| id.contains("reasoner")
|| id.starts_with("deepseek-v4")
{
Some(ReasoningMode::ValueSet {
values: ["disabled", "low", "medium", "high", "xhigh", "max"]
.iter().map(|s| s.to_string()).collect(),
default: Some("high".to_string()),
})
} else {
None
}
}
fn reasoning_request(&self, value: &serde_json::Value) -> Option<serde_json::Value> {
// "disabled" → thinking off; "enabled" → thinking on (no effort);
// any effort level → thinking on + `reasoning_effort`.
match value.as_str()? {
"disabled" => Some(serde_json::json!({ "thinking": { "type": "disabled" } })),
"enabled" => Some(serde_json::json!({ "thinking": { "type": "enabled" } })),
effort => Some(serde_json::json!({
"thinking": { "type": "enabled" },
"reasoning_effort": effort,
})),
}
}
fn build_llm(&self, record: &LlmProviderRecord, model: &LlmModelRecord) -> Option<Result<BuiltLlmClient>> {
Some(build_openai_llm(self, "https://api.deepseek.com/v1", record, model, false))
}
fn ui_meta(&self) -> ProviderUiMeta {
ProviderUiMeta {
type_id: "deepseek",
display_name: "DeepSeek",
description: None,
color: "#0ea5e9",
icon: "bi-search",
lists_models: true,
fields: &[
ProviderField { key: "api_key", label: "API Key", required: true, secret: true },
],
}
}
}
@@ -1,77 +0,0 @@
use std::sync::Arc;
use anyhow::Result;
use crate::chatbot::lm_studio::LmStudioClient;
use crate::llm::{LlmModelRecord, LlmProviderRecord};
use crate::llm::providers::{RemoteLlmModelInfo, fetch_openai_models};
use crate::provider::{ApiProvider, BuiltLlmClient, ProviderField, ProviderUiMeta, ServiceType};
pub struct LmStudioProvider {
http: reqwest::Client,
}
impl LmStudioProvider {
pub fn new() -> Self {
Self { http: reqwest::Client::new() }
}
fn base_url(record: &LlmProviderRecord) -> String {
record.base_url.clone()
.unwrap_or_else(|| "http://localhost:1234/v1".to_string())
}
}
#[async_trait::async_trait]
impl ApiProvider for LmStudioProvider {
fn type_id(&self) -> &'static str { "lm_studio" }
fn display_name(&self) -> &'static str { "LM Studio" }
fn supported_types(&self) -> &'static [ServiceType] {
&[ServiceType::Llm]
}
async fn list_llm_models(&self, record: &LlmProviderRecord) -> Result<Option<Vec<RemoteLlmModelInfo>>> {
let raw = fetch_openai_models(&self.http, &Self::base_url(record), None, "LM Studio").await?;
let models = raw
.iter()
.filter_map(|m| {
let id = m["id"].as_str()?.to_string();
Some(RemoteLlmModelInfo {
name: id.clone(), id,
context_length: None,
max_completion_tokens: None,
knowledge_cutoff: None,
capabilities: vec![],
vision: None,
price_input_per_million: None,
price_output_per_million: None,
reasoning: None,
})
})
.collect();
Ok(Some(models))
}
fn build_llm(&self, record: &LlmProviderRecord, _model: &LlmModelRecord) -> Option<Result<BuiltLlmClient>> {
Some(Ok(BuiltLlmClient {
client: Arc::new(LmStudioClient::new(record.base_url.as_deref())),
prompt_cache: false,
}))
}
fn ui_meta(&self) -> ProviderUiMeta {
ProviderUiMeta {
type_id: "lm_studio",
display_name: "LM Studio",
description: Some("Local models via LM Studio"),
color: "#6b7280",
icon: "bi-window-stack",
lists_models: true,
fields: &[
ProviderField { key: "base_url", label: "Base URL", required: false, secret: false },
],
}
}
}
+1 -4
View File
@@ -1,11 +1,8 @@
pub mod anthropic;
pub mod deepseek;
pub mod lm_studio;
pub mod moonshot;
pub mod declared;
pub mod ollama;
pub mod openai;
pub mod openrouter;
pub mod zai;
// Re-export so existing code that uses `providers::ServiceType` / `providers::RemoteLlmModelInfo` keeps working.
pub use crate::provider::ServiceType;
@@ -1,197 +0,0 @@
use anyhow::{Result, anyhow};
use crate::llm::{LlmModelRecord, LlmProviderRecord};
use crate::llm::providers::{RemoteLlmModelInfo, build_openai_llm, fetch_openai_models};
use crate::provider::{ApiProvider, BuiltLlmClient, ProviderField, ProviderUiMeta, ReasoningMode, ServiceType};
/// Moonshot AI — pay-as-you-go platform.
///
/// Endpoint `https://api.moonshot.ai/v1/chat/completions` (OpenAI-compatible);
/// `OpenAiClient` appends `/chat/completions`, so the base URL is `.../v1`.
///
/// `GET /models` returns the full catalog including `context_length`,
/// `supports_image_in` and `supports_reasoning`, so the model list is entirely
/// endpoint-driven. Thinking models (kimi-k2-thinking…) always reason — the
/// platform exposes no request-level knob, hence no `ReasoningMode`.
pub struct MoonshotProvider {
http: reqwest::Client,
}
/// Moonshot AI — Kimi Code subscription.
///
/// Endpoint `https://api.kimi.com/coding/v1/chat/completions` (OpenAI-compatible).
/// The model catalog comes from `GET /models` like the platform; only the
/// metadata the endpoint omits (context size, vision) is filled from the
/// published docs. K3 is the only model with a reasoning knob: a graded
/// `reasoning_effort` (`low`/`high`/`max`, default `max`); the kimi-for-coding
/// series (K2.7 Code) always thinks and has no toggle.
pub struct MoonshotCodeProvider {
http: reqwest::Client,
}
impl MoonshotProvider {
pub fn new() -> Self {
Self { http: reqwest::Client::new() }
}
const BASE_URL: &'static str = "https://api.moonshot.ai/v1";
}
impl MoonshotCodeProvider {
pub fn new() -> Self {
Self { http: reqwest::Client::new() }
}
const BASE_URL: &'static str = "https://api.kimi.com/coding/v1";
/// Fills the metadata the Kimi Code `/models` endpoint may omit, from the
/// published docs: K3 → up to 1M context + native visual understanding;
/// the kimi-for-coding series → 256k. Values already present in the
/// endpoint response (e.g. a tier-specific context size) always win.
fn enrich(info: &mut RemoteLlmModelInfo) {
let id = info.id.to_lowercase();
if info.context_length.is_none() {
if id.starts_with("k3") {
info.context_length = Some(1_048_576);
} else if id.starts_with("kimi-for-coding") {
info.context_length = Some(262_144);
}
}
if info.vision.is_none() && id.starts_with("k3") {
info.vision = Some(true);
}
}
}
/// Shared model listing for both Moonshot APIs: same envelope, same per-model
/// fields (`context_length`, `supports_image_in`, `supports_reasoning` — all
/// optional, absent fields are left for the caller to enrich).
async fn list_models(
http: &reqwest::Client,
base_url: &str,
record: &LlmProviderRecord,
who: &str,
) -> Result<Vec<RemoteLlmModelInfo>> {
let api_key = record.api_key.as_deref()
.ok_or_else(|| anyhow!("provider '{}': api_key required for {who} model listing", record.name))?;
let raw = fetch_openai_models(http, base_url, Some(api_key), who).await?;
let models = raw
.iter()
.filter_map(|m| {
let id = m["id"].as_str()?.to_string();
let mut capabilities = vec!["function_calling".to_string()];
if m["supports_reasoning"].as_bool().unwrap_or(false) {
capabilities.push("reasoning".to_string());
}
if m["supports_image_in"].as_bool().unwrap_or(false) {
capabilities.push("vision".to_string());
}
Some(RemoteLlmModelInfo {
id,
name: m["id"].as_str()?.to_string(),
context_length: m["context_length"].as_u64(),
max_completion_tokens: None,
knowledge_cutoff: None,
capabilities,
vision: m["supports_image_in"].as_bool(),
price_input_per_million: None,
price_output_per_million: None,
reasoning: None,
})
})
.collect();
Ok(models)
}
#[async_trait::async_trait]
impl ApiProvider for MoonshotProvider {
fn type_id(&self) -> &'static str { "moonshot" }
fn display_name(&self) -> &'static str { "Moonshot AI pay-as-you-go" }
fn supported_types(&self) -> &'static [ServiceType] {
&[ServiceType::Llm]
}
async fn list_llm_models(&self, record: &LlmProviderRecord) -> Result<Option<Vec<RemoteLlmModelInfo>>> {
Ok(Some(list_models(&self.http, Self::BASE_URL, record, "Moonshot AI").await?))
}
fn build_llm(&self, record: &LlmProviderRecord, model: &LlmModelRecord) -> Option<Result<BuiltLlmClient>> {
Some(build_openai_llm(self, Self::BASE_URL, record, model, false))
}
fn ui_meta(&self) -> ProviderUiMeta {
ProviderUiMeta {
type_id: "moonshot",
display_name: "Moonshot AI pay-as-you-go",
description: Some("Kimi models on the Moonshot AI platform (OpenAI-compatible)"),
color: "#2563eb",
icon: "bi-moon-stars",
lists_models: true,
fields: &[
ProviderField { key: "api_key", label: "API Key", required: true, secret: true },
],
}
}
}
#[async_trait::async_trait]
impl ApiProvider for MoonshotCodeProvider {
fn type_id(&self) -> &'static str { "moonshot_code" }
fn display_name(&self) -> &'static str { "Moonshot AI Kimi Code" }
fn supported_types(&self) -> &'static [ServiceType] {
&[ServiceType::Llm]
}
async fn list_llm_models(&self, record: &LlmProviderRecord) -> Result<Option<Vec<RemoteLlmModelInfo>>> {
let mut models = list_models(&self.http, Self::BASE_URL, record, "Kimi Code").await?;
for m in &mut models {
Self::enrich(m);
}
Ok(Some(models))
}
fn reasoning_mode(&self, model_id: &str, _capabilities: &[String]) -> Option<ReasoningMode> {
let id = model_id.to_lowercase();
// K3 exposes a graded `reasoning_effort` (low/high/max; default max).
// "disabled" turns thinking off — the API then routes to K2.6.
// kimi-for-coding (K2.7 Code) always thinks and has no knob.
if id.starts_with("k3") {
Some(ReasoningMode::ValueSet {
values: ["disabled", "low", "high", "max"]
.iter().map(|s| s.to_string()).collect(),
default: Some("max".to_string()),
})
} else {
None
}
}
fn reasoning_request(&self, value: &serde_json::Value) -> Option<serde_json::Value> {
// The Kimi Code API accepts a flat `reasoning_effort`: "none" disables
// thinking, low/high/max select the effort (unknown values → HTTP 400).
match value.as_str()? {
"disabled" => Some(serde_json::json!({ "reasoning_effort": "none" })),
effort => Some(serde_json::json!({ "reasoning_effort": effort })),
}
}
fn build_llm(&self, record: &LlmProviderRecord, model: &LlmModelRecord) -> Option<Result<BuiltLlmClient>> {
Some(build_openai_llm(self, Self::BASE_URL, record, model, false))
}
fn ui_meta(&self) -> ProviderUiMeta {
ProviderUiMeta {
type_id: "moonshot_code",
display_name: "Moonshot AI Kimi Code",
description: Some("Kimi Code subscription models — k3 / kimi-for-coding (OpenAI-compatible)"),
color: "#000000",
icon: "bi-code-slash",
lists_models: true,
fields: &[
ProviderField { key: "api_key", label: "API Key", required: true, secret: true },
],
}
}
}
-135
View File
@@ -1,135 +0,0 @@
use anyhow::Result;
use crate::llm::{LlmModelRecord, LlmProviderRecord};
use crate::llm::providers::{RemoteLlmModelInfo, build_openai_llm};
use crate::provider::{ApiProvider, BuiltLlmClient, ProviderField, ProviderUiMeta, ReasoningMode, ServiceType};
/// Z.AI (Zhipu AI) — OpenAI-compatible GLM API.
///
/// Endpoint `https://api.z.ai/api/paas/v4/chat/completions`; `OpenAiClient`
/// appends `/chat/completions`, so the base URL is `.../paas/v4`.
///
/// Z.AI exposes no `GET /models` endpoint, so the model catalog is a curated
/// static list of the currently published GLM models.
pub struct ZaiProvider;
impl ZaiProvider {
pub fn new() -> Self {
Self
}
/// Base URL for the OpenAI-compatible chat endpoint (without `/chat/completions`).
const BASE_URL: &'static str = "https://api.z.ai/api/paas/v4";
/// Curated GLM catalog. Z.AI has no `GET /models` endpoint; this mirrors the
/// model menu published on the Z.AI console.
fn catalog() -> &'static [&'static str] {
&[
"glm-5.2",
"glm-5.1",
"glm-5",
"glm-5-turbo",
"glm-4.7",
"glm-4.6",
"glm-4.5",
"glm-4-32b-0414-128k",
]
}
fn known_context_length(model_id: &str) -> Option<u64> {
let id = model_id.to_lowercase();
if id.contains("128k") { Some(131_072) }
else if id.starts_with("glm-5") { Some(1_048_576) } // GLM-5.x: 1M context (per Z.AI)
else if id.starts_with("glm-4.7") { Some(200_000) }
else if id.starts_with("glm-4.6") { Some(200_000) }
else if id.starts_with("glm-4.5") { Some(131_072) }
else { None }
}
}
#[async_trait::async_trait]
impl ApiProvider for ZaiProvider {
fn type_id(&self) -> &'static str { "zai" }
fn display_name(&self) -> &'static str { "Z.AI" }
fn supported_types(&self) -> &'static [ServiceType] {
&[ServiceType::Llm]
}
async fn list_llm_models(&self, _record: &LlmProviderRecord) -> Result<Option<Vec<RemoteLlmModelInfo>>> {
let models = Self::catalog()
.iter()
.map(|id| RemoteLlmModelInfo {
id: id.to_string(),
name: id.to_string(),
context_length: Self::known_context_length(id),
max_completion_tokens: None,
knowledge_cutoff: None,
capabilities: vec!["function_calling".to_string()],
vision: Some(false),
price_input_per_million: None,
price_output_per_million: None,
reasoning: None,
})
.collect();
Ok(Some(models))
}
fn reasoning_mode(&self, model_id: &str, _capabilities: &[String]) -> Option<ReasoningMode> {
let id = model_id.to_lowercase();
// GLM-5.2 (and above) additionally expose a graded `reasoning_effort`
// on top of the thinking toggle, so offer the effort levels directly
// ("disabled" turns thinking off).
if id.starts_with("glm-5.2") {
Some(ReasoningMode::ValueSet {
values: ["disabled", "minimal", "low", "medium", "high", "xhigh", "max"]
.iter().map(|s| s.to_string()).collect(),
default: Some("max".to_string()),
})
// Deep-thinking toggle (thinking.type) is supported by the GLM-5.x
// series and GLM-4.5/4.6/4.7 (but not the older glm-4-32b).
} else if id.starts_with("glm-5")
|| id.starts_with("glm-4.7")
|| id.starts_with("glm-4.6")
|| id.starts_with("glm-4.5")
{
Some(ReasoningMode::ValueSet {
values: vec!["disabled".to_string(), "enabled".to_string()],
default: Some("enabled".to_string()),
})
} else {
None
}
}
fn reasoning_request(&self, value: &serde_json::Value) -> Option<serde_json::Value> {
// "disabled" → thinking off; "enabled" → thinking on (no effort);
// any effort level → thinking on + `reasoning_effort` (GLM-5.2+).
match value.as_str()? {
"disabled" => Some(serde_json::json!({ "thinking": { "type": "disabled" } })),
"enabled" => Some(serde_json::json!({ "thinking": { "type": "enabled" } })),
effort => Some(serde_json::json!({
"thinking": { "type": "enabled" },
"reasoning_effort": effort,
})),
}
}
fn build_llm(&self, record: &LlmProviderRecord, model: &LlmModelRecord) -> Option<Result<BuiltLlmClient>> {
Some(build_openai_llm(self, Self::BASE_URL, record, model, false))
}
fn ui_meta(&self) -> ProviderUiMeta {
ProviderUiMeta {
type_id: "zai",
display_name: "Z.AI",
description: Some("Zhipu AI GLM models (OpenAI-compatible)"),
color: "#4f46e5",
icon: "bi-stars",
lists_models: true,
fields: &[
ProviderField { key: "api_key", label: "API Key", required: true, secret: true },
],
}
}
}
+11 -5
View File
@@ -47,6 +47,7 @@ use crate::tts::TtsManager;
use tokio::sync::RwLock;
use core_api::plugin::Plugin;
use core_api::provider::ApiProvider;
use super::runtime::Runtime;
@@ -66,11 +67,16 @@ impl Models {
provider_registry.register_builtin(crate::llm::providers::anthropic::AnthropicProvider::new());
provider_registry.register_builtin(crate::llm::providers::openrouter::OpenRouterProvider::new());
provider_registry.register_builtin(crate::llm::providers::ollama::OllamaProvider::new());
provider_registry.register_builtin(crate::llm::providers::lm_studio::LmStudioProvider::new());
provider_registry.register_builtin(crate::llm::providers::deepseek::DeepSeekProvider::new());
provider_registry.register_builtin(crate::llm::providers::zai::ZaiProvider::new());
provider_registry.register_builtin(crate::llm::providers::moonshot::MoonshotProvider::new());
provider_registry.register_builtin(crate::llm::providers::moonshot::MoonshotCodeProvider::new());
// OpenAI-compatible providers are runtime data (providers.yaml), not code.
for p in crate::llm::providers::declared::load(std::path::Path::new(
crate::llm::providers::declared::PROVIDERS_FILE,
)) {
if provider_registry.contains(p.type_id()) {
warn!(type_id = p.type_id(), "declared provider id collides with a native provider — skipped");
continue;
}
provider_registry.register_builtin(p);
}
let provider_registry = Arc::new(provider_registry);
info!("provider registry ready ({} built-in providers)", provider_registry.all().len());