car-inference 0.47.0

//! Curated OpenRouter model definitions and credential resolution.
//!
//! The built-in table produces exactly eleven reviewed personal models plus
//! the eleven opaque Parslee-managed aliases. Personal availability follows the
//! resolved credential on every registry snapshot; CAR deliberately does not
//! project OpenRouter's full catalog. Managed aliases remain bounded and never
//! disclose upstream provider ids.

use car_secrets::{SecretRef, SecretStore};
#[cfg(test)]
use std::cell::Cell;
#[cfg(test)]
use std::sync::MutexGuard;
use std::sync::{Mutex, OnceLock};

use crate::schema::{
    ApiProtocol, CostModel, GenerateParam, ModelCapability, ModelSchema, ModelSource,
    PerformanceEnvelope, ProprietaryAuth, ProprietaryProtocol, TrustTier,
};

pub const API_KEY_ENV: &str = "OPENROUTER_API_KEY";
pub const OAUTH_KEYCHAIN_KEY: &str = car_secrets::OPENROUTER_OAUTH_KEY;
pub const DEFAULT_API_BASE: &str = "https://openrouter.ai/api";

fn oauth_keychain_service() -> String {
    // Test-double verification can isolate its credential without touching the
    // operator's real OAuth connection. Production never sets this variable.
    std::env::var("CAR_OPENROUTER_OAUTH_KEYCHAIN_SERVICE")
        .unwrap_or_else(|_| car_secrets::DEFAULT_SERVICE.to_string())
}

fn oauth_secret_ref() -> SecretRef {
    SecretRef::new(oauth_keychain_service(), OAUTH_KEYCHAIN_KEY)
}

/// The OAuth credential is daemon-private. Generic secret RPC/FFI surfaces
/// must fail closed for this exact `(service, key)` pair; only this module's
/// request lease and the daemon-owned OAuth manager may access it.
pub fn is_reserved_oauth_secret(service: Option<&str>, key: &str) -> bool {
    service.unwrap_or(car_secrets::DEFAULT_SERVICE) == oauth_keychain_service()
        && key == OAUTH_KEYCHAIN_KEY
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum CredentialSource {
    Env,
    Pasted,
    Oauth,
}

impl CredentialSource {
    pub fn as_str(self) -> &'static str {
        match self {
            Self::Env => "env",
            Self::Pasted => "pasted",
            Self::Oauth => "oauth",
        }
    }
}

/// Resolve only existence/source metadata. Registry listing and routing use
/// this path so they never fetch the credential value from the OS keychain.
pub fn credential_source() -> Option<CredentialSource> {
    #[cfg(test)]
    CREDENTIAL_SOURCE_CALLS.with(|calls| calls.set(calls.get() + 1));
    #[cfg(test)]
    if let Some(value) = test_credential_override()
        .lock()
        .unwrap_or_else(|p| p.into_inner())
        .clone()
    {
        return value.map(|_| CredentialSource::Env);
    }
    if environment_key_exists() {
        return Some(CredentialSource::Env);
    }
    if pasted_key_exists() {
        return Some(CredentialSource::Pasted);
    }
    oauth_key_exists().then_some(CredentialSource::Oauth)
}

#[cfg(test)]
thread_local! {
    // Per-thread instrumentation keeps the "one credential probe per routing
    // decision" assertion isolated from unrelated router tests running in
    // parallel in the same process.
    static CREDENTIAL_SOURCE_CALLS: Cell<usize> = const { Cell::new(0) };
}

#[cfg(test)]
pub(crate) fn reset_credential_source_call_count() {
    CREDENTIAL_SOURCE_CALLS.with(|calls| calls.set(0));
}

#[cfg(test)]
pub(crate) fn credential_source_call_count() -> usize {
    CREDENTIAL_SOURCE_CALLS.with(Cell::get)
}

pub fn environment_key_exists() -> bool {
    std::env::var(API_KEY_ENV).is_ok_and(|value| !value.trim().is_empty())
}

/// Resolve on every request: environment, separately pasted key, then OAuth.
pub fn resolve_credential() -> Option<(String, CredentialSource)> {
    #[cfg(test)]
    if let Some(value) = test_credential_override()
        .lock()
        .unwrap_or_else(|p| p.into_inner())
        .clone()
    {
        return value.map(|key| (key, CredentialSource::Env));
    }
    if let Ok(value) = std::env::var(API_KEY_ENV) {
        if !value.trim().is_empty() {
            return Some((value, CredentialSource::Env));
        }
    }
    let store = SecretStore::new();
    if store.is_available() {
        if let Ok(value) = store.get(&SecretRef::with_default_service(API_KEY_ENV)) {
            if !value.trim().is_empty() {
                return Some((value, CredentialSource::Pasted));
            }
        }
        if let Ok(value) = store.get(&oauth_secret_ref()) {
            if !value.trim().is_empty() {
                return Some((value, CredentialSource::Oauth));
            }
        }
    }
    None
}

#[cfg(test)]
fn test_credential_override() -> &'static Mutex<Option<Option<String>>> {
    static OVERRIDE: OnceLock<Mutex<Option<Option<String>>>> = OnceLock::new();
    OVERRIDE.get_or_init(|| Mutex::new(None))
}

#[cfg(test)]
fn test_credential_serial_lock() -> &'static Mutex<()> {
    static LOCK: OnceLock<Mutex<()>> = OnceLock::new();
    LOCK.get_or_init(|| Mutex::new(()))
}

/// Serializes tests that exercise the process-wide credential resolver. The
/// guard also clears the override on unwind so a failed test cannot poison the
/// availability seen by the rest of the suite.
#[cfg(test)]
pub(crate) struct TestCredentialScope {
    _guard: MutexGuard<'static, ()>,
}

#[cfg(test)]
impl Drop for TestCredentialScope {
    fn drop(&mut self) {
        clear_test_credential();
    }
}

#[cfg(test)]
pub(crate) fn test_credential_scope() -> TestCredentialScope {
    TestCredentialScope {
        _guard: test_credential_serial_lock()
            .lock()
            .unwrap_or_else(|p| p.into_inner()),
    }
}

#[cfg(test)]
pub(crate) fn test_environment_scope() -> tokio::sync::MutexGuard<'static, ()> {
    test_environment_lock().blocking_lock()
}

#[cfg(test)]
pub(crate) async fn test_environment_scope_async() -> tokio::sync::MutexGuard<'static, ()> {
    test_environment_lock().lock().await
}

#[cfg(test)]
fn test_environment_lock() -> &'static tokio::sync::Mutex<()> {
    static LOCK: OnceLock<tokio::sync::Mutex<()>> = OnceLock::new();
    LOCK.get_or_init(|| tokio::sync::Mutex::new(()))
}

/// Deterministic credential seam for provider-wire tests. `Some(key)` forces a
/// key, `None` forces no credential; [`clear_test_credential`] restores normal
/// env/keychain resolution.
#[cfg(test)]
pub(crate) fn set_test_credential(value: Option<&str>) {
    *test_credential_override()
        .lock()
        .unwrap_or_else(|p| p.into_inner()) = Some(value.map(str::to_string));
}

#[cfg(test)]
pub(crate) fn clear_test_credential() {
    *test_credential_override()
        .lock()
        .unwrap_or_else(|p| p.into_inner()) = None;
}

pub fn pasted_key_exists() -> bool {
    SecretStore::new()
        .status(&SecretRef::with_default_service(API_KEY_ENV))
        .map(|status| status.exists)
        .unwrap_or(false)
}

pub fn oauth_key_exists() -> bool {
    SecretStore::new()
        .status(&oauth_secret_ref())
        .map(|status| status.exists)
        .unwrap_or(false)
}

pub fn store_oauth_credential(value: &str) -> Result<(), car_secrets::SecretError> {
    SecretStore::new().put(&oauth_secret_ref(), value)
}

pub fn delete_oauth_credential() -> Result<(), car_secrets::SecretError> {
    match SecretStore::new().delete(&oauth_secret_ref()) {
        Ok(()) | Err(car_secrets::SecretError::NotFound { .. }) => Ok(()),
        Err(error) => Err(error),
    }
}

#[derive(Clone, Copy)]
struct CuratedModel {
    upstream_id: &'static str,
    managed_alias: &'static str,
    family: &'static str,
    /// Date of the `https://openrouter.ai/api/v1/models` snapshot this row's
    /// context window, output ceiling and prices were read from. Per-row
    /// rather than table-wide so a later addition doesn't have to claim the
    /// original snapshot's date. Published as `ModelSchema::version`.
    snapshot: &'static str,
    context_length: usize,
    max_output_tokens: usize,
    input_per_mtok: f64,
    output_per_mtok: f64,
    cache_read_input_per_mtok: Option<f64>,
    cache_write_input_per_mtok: Option<f64>,
    high_context_pricing: Option<CuratedPricingTier>,
    capabilities: &'static [ModelCapability],
    tags: &'static [&'static str],
    supported_params: &'static [GenerateParam],
}

#[derive(Clone, Copy)]
struct CuratedPricingTier {
    min_prompt_tokens: usize,
    input_per_mtok: f64,
    output_per_mtok: f64,
    cache_read_input_per_mtok: Option<f64>,
    cache_write_input_per_mtok: Option<f64>,
}

use ModelCapability as C;
const CODE_REASON_VISION: &[ModelCapability] = &[
    C::Generate,
    C::Code,
    C::Reasoning,
    C::ToolUse,
    C::Summarize,
    C::Vision,
];
const CODE_REASON: &[ModelCapability] =
    &[C::Generate, C::Code, C::Reasoning, C::ToolUse, C::Summarize];
const CODE_REASON_MULTI_TOOL: &[ModelCapability] = &[
    C::Generate,
    C::Code,
    C::Reasoning,
    C::ToolUse,
    C::MultiToolCall,
    C::Summarize,
];
const CODE: &[ModelCapability] = &[C::Generate, C::Code, C::ToolUse, C::Summarize];

use GenerateParam as P;
// Keep these bounded to fields `OpenRouterHandler` actually places on the
// Chat Completions request. The reviewed upstream snapshot may advertise
// additional controls, but CAR must not claim them until `ApiRequest` and the
// adapter carry them end to end.
const PARAMS_OPENAI_REASONING: &[GenerateParam] =
    &[P::MaxTokens, P::ResponseFormat, P::ExtendedThinking];
const PARAMS_ANTHROPIC_REASONING: &[GenerateParam] = &[
    P::Temperature,
    P::MaxTokens,
    P::ResponseFormat,
    P::ExtendedThinking,
];
const PARAMS_GEMINI_REASONING: &[GenerateParam] = &[
    P::Temperature,
    P::MaxTokens,
    P::ResponseFormat,
    P::ExtendedThinking,
];
const PARAMS_STANDARD_REASONING: &[GenerateParam] = &[
    P::Temperature,
    P::MaxTokens,
    P::ResponseFormat,
    P::ExtendedThinking,
];
const PARAMS_QWEN_REASONING: &[GenerateParam] = &[
    P::Temperature,
    P::MaxTokens,
    P::ResponseFormat,
    P::ExtendedThinking,
];
const PARAMS_STANDARD: &[GenerateParam] = &[P::Temperature, P::MaxTokens, P::ResponseFormat];

/// Rows read from the official `https://openrouter.ai/api/v1/models` catalog.
/// Each row carries the date of the snapshot it was read from in `snapshot`,
/// so a later addition doesn't restate the whole table's provenance.
/// Prices are USD/MTok; high-context overrides use the endpoint's inclusive
/// `min_prompt_tokens` thresholds.
const CURATED: &[CuratedModel] = &[
    CuratedModel {
        upstream_id: "openai/gpt-5.4",
        managed_alias: "frontier-general",
        family: "gpt-5.4",
        snapshot: "2026-07-22",
        context_length: 1_050_000,
        max_output_tokens: 128_000,
        input_per_mtok: 2.5,
        output_per_mtok: 15.0,
        cache_read_input_per_mtok: Some(0.25),
        cache_write_input_per_mtok: None,
        high_context_pricing: Some(CuratedPricingTier {
            min_prompt_tokens: 272_000,
            input_per_mtok: 5.0,
            output_per_mtok: 22.5,
            cache_read_input_per_mtok: Some(0.5),
            cache_write_input_per_mtok: None,
        }),
        capabilities: CODE_REASON_VISION,
        tags: &["frontier"],
        supported_params: PARAMS_OPENAI_REASONING,
    },
    CuratedModel {
        upstream_id: "anthropic/claude-opus-4.6",
        managed_alias: "frontier-deep",
        family: "claude-4.6",
        snapshot: "2026-07-22",
        context_length: 1_000_000,
        max_output_tokens: 128_000,
        input_per_mtok: 5.0,
        output_per_mtok: 25.0,
        cache_read_input_per_mtok: Some(0.5),
        cache_write_input_per_mtok: Some(6.25),
        high_context_pricing: None,
        capabilities: CODE_REASON_VISION,
        tags: &["frontier"],
        supported_params: PARAMS_ANTHROPIC_REASONING,
    },
    CuratedModel {
        upstream_id: "google/gemini-3.1-pro-preview",
        managed_alias: "frontier-multimodal",
        family: "gemini-3.1",
        snapshot: "2026-07-22",
        context_length: 1_048_576,
        max_output_tokens: 65_536,
        input_per_mtok: 2.0,
        output_per_mtok: 12.0,
        cache_read_input_per_mtok: Some(0.2),
        cache_write_input_per_mtok: Some(0.375),
        high_context_pricing: Some(CuratedPricingTier {
            min_prompt_tokens: 200_000,
            input_per_mtok: 4.0,
            output_per_mtok: 18.0,
            cache_read_input_per_mtok: Some(0.4),
            cache_write_input_per_mtok: Some(0.375),
        }),
        capabilities: CODE_REASON_VISION,
        tags: &["frontier", "preview"],
        supported_params: PARAMS_GEMINI_REASONING,
    },
    CuratedModel {
        upstream_id: "anthropic/claude-sonnet-4.6",
        managed_alias: "balanced-general",
        family: "claude-4.6",
        snapshot: "2026-07-22",
        context_length: 1_000_000,
        max_output_tokens: 128_000,
        input_per_mtok: 3.0,
        output_per_mtok: 15.0,
        cache_read_input_per_mtok: Some(0.3),
        cache_write_input_per_mtok: Some(3.75),
        high_context_pricing: None,
        capabilities: CODE_REASON_VISION,
        tags: &["balanced"],
        supported_params: PARAMS_ANTHROPIC_REASONING,
    },
    CuratedModel {
        upstream_id: "moonshotai/kimi-k2.5",
        managed_alias: "open-multimodal",
        family: "kimi-k2.5",
        snapshot: "2026-07-22",
        context_length: 262_144,
        max_output_tokens: 262_144,
        input_per_mtok: 0.57,
        output_per_mtok: 2.85,
        cache_read_input_per_mtok: Some(0.095),
        cache_write_input_per_mtok: None,
        high_context_pricing: None,
        capabilities: CODE_REASON_VISION,
        tags: &["cheap", "open-weight"],
        supported_params: PARAMS_STANDARD_REASONING,
    },
    CuratedModel {
        upstream_id: "qwen/qwen3.5-plus-02-15",
        managed_alias: "open-long-context",
        family: "qwen3.5",
        snapshot: "2026-07-22",
        context_length: 1_000_000,
        max_output_tokens: 65_536,
        input_per_mtok: 0.26,
        output_per_mtok: 1.56,
        cache_read_input_per_mtok: None,
        cache_write_input_per_mtok: None,
        high_context_pricing: Some(CuratedPricingTier {
            min_prompt_tokens: 256_000,
            input_per_mtok: 0.325,
            output_per_mtok: 1.95,
            cache_read_input_per_mtok: None,
            cache_write_input_per_mtok: None,
        }),
        capabilities: CODE_REASON_VISION,
        tags: &["cheap", "open-weight"],
        supported_params: PARAMS_QWEN_REASONING,
    },
    CuratedModel {
        upstream_id: "deepseek/deepseek-v3.2",
        managed_alias: "open-reasoning",
        family: "deepseek-v3.2",
        snapshot: "2026-07-22",
        context_length: 163_840,
        max_output_tokens: 65_536,
        input_per_mtok: 0.269,
        output_per_mtok: 0.4,
        cache_read_input_per_mtok: Some(0.1345),
        cache_write_input_per_mtok: None,
        high_context_pricing: None,
        capabilities: CODE_REASON,
        tags: &["cheap", "open-weight"],
        supported_params: PARAMS_STANDARD_REASONING,
    },
    CuratedModel {
        upstream_id: "minimax/minimax-m2.5",
        managed_alias: "open-fast",
        family: "minimax-m2.5",
        snapshot: "2026-07-22",
        context_length: 204_800,
        max_output_tokens: 196_608,
        input_per_mtok: 0.15,
        output_per_mtok: 0.9,
        cache_read_input_per_mtok: Some(0.05),
        cache_write_input_per_mtok: None,
        high_context_pricing: None,
        capabilities: CODE_REASON_MULTI_TOOL,
        tags: &["cheap", "open-weight"],
        supported_params: PARAMS_STANDARD_REASONING,
    },
    CuratedModel {
        upstream_id: "openai/gpt-5.3-codex",
        managed_alias: "coding-frontier",
        family: "gpt-5.3-codex",
        snapshot: "2026-07-22",
        context_length: 400_000,
        max_output_tokens: 128_000,
        input_per_mtok: 1.75,
        output_per_mtok: 14.0,
        cache_read_input_per_mtok: Some(0.175),
        cache_write_input_per_mtok: None,
        high_context_pricing: None,
        capabilities: CODE_REASON_VISION,
        tags: &["code"],
        supported_params: PARAMS_OPENAI_REASONING,
    },
    CuratedModel {
        upstream_id: "qwen/qwen3-coder-next",
        managed_alias: "coding-efficient",
        family: "qwen3-coder",
        snapshot: "2026-07-22",
        context_length: 262_144,
        max_output_tokens: 262_144,
        input_per_mtok: 0.11,
        output_per_mtok: 0.8,
        cache_read_input_per_mtok: Some(0.07),
        cache_write_input_per_mtok: None,
        high_context_pricing: None,
        capabilities: CODE,
        tags: &["code", "cheap", "open-weight"],
        supported_params: PARAMS_STANDARD,
    },
    // Appended from the live `https://openrouter.ai/api/v1/models` catalog on
    // 2026-07-31. OpenRouter publishes USD *per token* as strings; these are
    // those numbers × 1e6: prompt `0.000005` → 5.0 USD/MTok, completion
    // `0.000025` → 25.0, `input_cache_read` `0.0000005` → 0.5,
    // `input_cache_write` `0.00000625` → 6.25. `context_length` 1_000_000 and
    // `top_provider.max_completion_tokens` 128_000 come from the same row. The
    // `-fast` variant (2× price) is deliberately not curated.
    //
    // Note this row is a price/capability/context clone of `claude-opus-4.6`,
    // so the two tie exactly on every adaptive-routing score term. The tie
    // resolves deterministically but on an incidental key, and NOT the way an
    // id-ascending catalog order suggests: `UnifiedRegistry::list` sorts ids
    // ascending, `AdaptiveRouter`'s scoring sort is stable-descending so the
    // tied pair keeps that order, and the Quality lane picks with
    // `Iterator::max_by`, which returns the LAST of equal maxima — so the
    // primary is `4.8`, while the fallback chain, built from the same ordering,
    // still leads with `4.6`. Deterministic, but decided by string ordering
    // rather than merit. Preferring one on purpose needs an explicit signal (a
    // benchmark score, a `deprecated` flag, a tag), not a rename.
    CuratedModel {
        upstream_id: "anthropic/claude-opus-4.8",
        managed_alias: "frontier-deep-next",
        family: "claude-4.8",
        snapshot: "2026-07-31",
        context_length: 1_000_000,
        max_output_tokens: 128_000,
        input_per_mtok: 5.0,
        output_per_mtok: 25.0,
        cache_read_input_per_mtok: Some(0.5),
        cache_write_input_per_mtok: Some(6.25),
        high_context_pricing: None,
        capabilities: CODE_REASON_VISION,
        tags: &["frontier"],
        supported_params: PARAMS_ANTHROPIC_REASONING,
    },
];

fn schema(model: CuratedModel, gateway: bool, parslee_api_base: &str) -> ModelSchema {
    let id = if gateway {
        format!("parslee/openrouter/{}", model.managed_alias)
    } else {
        format!("openrouter/{}", model.upstream_id)
    };
    // Personal OpenRouter receives its raw upstream id. The Parslee gateway
    // receives the full allow-listed gateway id.
    let name = if gateway {
        id.clone()
    } else {
        model.upstream_id.to_string()
    };
    let mut tags = vec!["builtin".to_string(), "openrouter".to_string()];
    tags.extend(model.tags.iter().map(|tag| (*tag).to_string()));
    if gateway {
        tags.extend(["parslee".to_string(), "managed".to_string()]);
    } else {
        tags.extend([
            "personal-key".to_string(),
            "openrouter-baseline".to_string(),
        ]);
    }
    let source = if gateway {
        ModelSource::Proprietary {
            provider: "parslee".to_string(),
            endpoint: parslee_api_base.to_string(),
            auth: ProprietaryAuth::OAuth2Pkce {
                authority: parslee_api_base.to_string(),
                client_id: "parslee-car".to_string(),
                scopes: vec!["inference:invoke".to_string(), "models:list".to_string()],
            },
            protocol: ProprietaryProtocol {
                chat_path: "/api/v1/orgs/{orgId}/inference/responses".to_string(),
                content_type: "application/json".to_string(),
                streaming: true,
                extra_headers: Default::default(),
            },
        }
    } else {
        ModelSource::RemoteApi {
            endpoint: DEFAULT_API_BASE.to_string(),
            api_key_env: API_KEY_ENV.to_string(),
            api_key_envs: Vec::new(),
            api_version: None,
            protocol: ApiProtocol::OpenRouter,
        }
    };
    ModelSchema {
        id,
        name,
        provider: if gateway { "parslee" } else { "openrouter" }.to_string(),
        family: model.family.to_string(),
        version: model.snapshot.to_string(),
        capabilities: model.capabilities.to_vec(),
        context_length: model.context_length,
        max_output_tokens: Some(model.max_output_tokens),
        param_count: String::new(),
        quantization: None,
        performance: PerformanceEnvelope::default(),
        cost: CostModel {
            input_per_mtok: Some(model.input_per_mtok),
            output_per_mtok: Some(model.output_per_mtok),
            cache_read_input_per_mtok: model.cache_read_input_per_mtok,
            cache_write_input_per_mtok: model.cache_write_input_per_mtok,
            pricing_tiers: model
                .high_context_pricing
                .into_iter()
                .map(|tier| crate::schema::TokenPricingTier {
                    min_prompt_tokens: tier.min_prompt_tokens,
                    prices: crate::schema::TokenPrices {
                        input_per_mtok: Some(tier.input_per_mtok),
                        output_per_mtok: Some(tier.output_per_mtok),
                        cache_read_input_per_mtok: tier.cache_read_input_per_mtok,
                        cache_write_input_per_mtok: tier.cache_write_input_per_mtok,
                    },
                })
                .collect(),
            size_mb: None,
            ram_mb: None,
        },
        source,
        tags,
        supported_params: if gateway {
            // The managed CAR → Parslee transport currently forwards only the
            // output-token limit. Keep the advertised contract bounded to the
            // request body we actually send.
            vec![P::MaxTokens]
        } else {
            model.supported_params.to_vec()
        },
        public_benchmarks: Vec::new(),
        trust_tier: TrustTier::Curated,
        deprecated: false,
        available: false,
        // Remote OpenRouter/Parslee models never require a local weight
        // download. Live credential availability remains a separate gate.
        weights_ready: true,
    }
}

/// Return the canonical selector for an allowlisted model using the exact
/// Parslee-managed OpenRouter transport.
///
/// The selector is the reviewed alias ID, never the mutable display/name
/// field. This helper intentionally excludes tags and trust tier: callers use
/// it both to bind outbound requests safely and as one part of the stricter
/// adaptive-trust predicate below.
pub fn canonical_managed_gateway_selector(model: &ModelSchema) -> Option<&str> {
    let canonical_id = is_curated_managed_gateway_alias(&model.id);

    (canonical_id
        && model.provider == "parslee"
        && matches!(
            &model.source,
            ModelSource::Proprietary {
                provider,
                endpoint,
                auth:
                    ProprietaryAuth::OAuth2Pkce {
                        authority,
                        client_id,
                        scopes,
                    },
                protocol,
            } if provider == "parslee"
                && !endpoint.is_empty()
                && endpoint.trim_end_matches('/') == authority.trim_end_matches('/')
                && client_id == "parslee-car"
                && scopes.as_slice() == ["inference:invoke", "models:list"]
                && protocol.chat_path == "/api/v1/orgs/{orgId}/inference/responses"
                && protocol.content_type == "application/json"
                && protocol.streaming
                && protocol.extra_headers.is_empty()
        ))
    .then_some(model.id.as_str())
}

/// Whether `id` is one of the fixed, reviewed Parslee-managed OpenRouter
/// aliases. This reads only the compiled curated catalog: it deliberately
/// does not build or refresh the runtime model registry.
pub fn is_curated_managed_gateway_alias(id: &str) -> bool {
    id.strip_prefix("parslee/openrouter/")
        .is_some_and(|alias| CURATED.iter().any(|entry| entry.managed_alias == alias))
}

/// How long a "this gateway has no OpenRouter upstream" observation suppresses
/// the managed aliases before CAR optimistically advertises them again.
///
/// Bounded rather than permanent because the condition is a *provisioning*
/// state, not a fact about the models: an operator can configure OpenRouter on
/// the Parslee environment at any moment, and CAR has no signal when that
/// happens (`parslee.capabilities` discovers product entitlements — studio,
/// aie, odi, crm — and nothing that enumerates inference upstreams). Fifteen
/// minutes costs at most one failed request per quarter-hour to rediscover,
/// which is the whole price of being wrong in the recoverable direction.
const GATEWAY_UNCONFIGURED_TTL: std::time::Duration = std::time::Duration::from_secs(15 * 60);

fn gateway_unconfigured_at() -> &'static Mutex<Option<std::time::Instant>> {
    static AT: OnceLock<Mutex<Option<std::time::Instant>>> = OnceLock::new();
    AT.get_or_init(|| Mutex::new(None))
}

/// Record that the managed gateway answered "OpenRouter is not configured on
/// this environment".
///
/// Called from the dispatch loop when the gateway returns that condition. It is
/// environment-scoped — the gateway has no OpenRouter configuration *at all*,
/// so it is identical for every `parslee/openrouter/*` alias and says nothing
/// about any individual model (Parslee-ai/car#786).
pub fn note_gateway_unconfigured() {
    let mut guard = match gateway_unconfigured_at().lock() {
        Ok(guard) => guard,
        Err(poisoned) => poisoned.into_inner(),
    };
    *guard = Some(std::time::Instant::now());
}

/// Whether a recent observation says the managed OpenRouter namespace cannot be
/// served. Drives `available` for those aliases so the catalog stops advertising
/// models that are certain to fail.
///
/// This is CAR *learning* rather than deriving. Availability for these rows
/// derives from `ProprietaryAuth::OAuth2Pkce`, which resolves to exactly "is a
/// Parslee session signed in" — true regardless of what the gateway can serve.
/// There is no discovery endpoint to ask instead, so the only truthful signal
/// available is the gateway's own answer to a real request.
pub fn gateway_unconfigured() -> bool {
    let guard = match gateway_unconfigured_at().lock() {
        Ok(guard) => guard,
        Err(poisoned) => poisoned.into_inner(),
    };
    guard.is_some_and(|at| at.elapsed() < GATEWAY_UNCONFIGURED_TTL)
}

/// Forget the observation. Called when the Parslee session goes away, because
/// what was learned was learned about *that* environment and the next sign-in
/// may be to one that has OpenRouter configured.
///
/// Note this catches sign-out, not an org switch within a live session — CAR
/// observes auth presence here, not identity. An org switch therefore keeps the
/// suppression until the TTL expires, which fails in the safe direction (a
/// usable namespace under-advertised for at most one TTL, rather than an
/// unusable one advertised).
pub fn clear_gateway_unconfigured() {
    let mut guard = match gateway_unconfigured_at().lock() {
        Ok(guard) => guard,
        Err(poisoned) => poisoned.into_inner(),
    };
    *guard = None;
}

/// Whether a schema has the exact built-in Parslee-managed OpenRouter
/// identity, metadata, and transport contract.
///
/// Provider names and free-form tags are caller-controlled, so adaptive
/// bootstrap must not treat those strings alone as project curation. This
/// shape is deliberately tied to the reviewed alias allowlist and Parslee's
/// proprietary OAuth/Responses gateway. User-controlled ingestion boundaries
/// independently demote imported schemas to `Community`.
pub(crate) fn is_managed_gateway_schema(model: &ModelSchema) -> bool {
    let Some(canonical_id) = canonical_managed_gateway_selector(model) else {
        return false;
    };
    let has_tag = |expected: &str| model.tags.iter().any(|tag| tag == expected);

    model.name == canonical_id
        && ["builtin", "openrouter", "parslee", "managed"]
            .into_iter()
            .all(has_tag)
}

/// How many curated OpenRouter models the built-in table declares — one
/// personal row and one managed alias each. Tests assert against this rather
/// than a hardcoded literal so adding a reviewed model is a one-line change
/// instead of a hunt through four unrelated count assertions.
pub fn curated_model_count() -> usize {
    CURATED.len()
}

pub fn curated_schemas() -> Vec<ModelSchema> {
    // Use the same explicit override -> persisted connection -> production
    // default resolution as Parslee sign-in. The registry is descriptive;
    // the remote request seam resolves again so a connection switched after
    // registry construction takes effect without restarting CAR.
    let parslee_api_base = car_auth::api_base(None);
    CURATED
        .iter()
        .flat_map(|model| {
            [
                schema(*model, false, &parslee_api_base),
                schema(*model, true, &parslee_api_base),
            ]
        })
        .collect()
}

/// Built-in registry rows: eleven reviewed personal models plus eleven managed
/// aliases. The personal rows are always discoverable; registry availability
/// follows the current credential without adding unreviewed model IDs.
pub fn builtin_schemas() -> Vec<ModelSchema> {
    curated_schemas()
}

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

    #[test]
    fn managed_aliases_follow_the_configured_parslee_api_base() {
        let _environment = test_environment_scope();
        unsafe {
            std::env::set_var(
                car_auth::PARSLEE_API_BASE_KEY,
                "https://staging-api.parslee.ai/",
            );
        }

        let managed: Vec<_> = curated_schemas()
            .into_iter()
            .filter(|schema| schema.provider == "parslee")
            .collect();
        assert_eq!(managed.len(), CURATED.len());
        for schema in managed {
            let ModelSource::Proprietary { endpoint, auth, .. } = schema.source else {
                panic!("managed alias must use the Parslee proprietary transport");
            };
            assert_eq!(endpoint, "https://staging-api.parslee.ai");
            let ProprietaryAuth::OAuth2Pkce { authority, .. } = auth else {
                panic!("managed alias must use the Parslee OAuth connection");
            };
            assert_eq!(authority, "https://staging-api.parslee.ai");
        }

        unsafe {
            std::env::remove_var(car_auth::PARSLEE_API_BASE_KEY);
        }
    }

    #[test]
    fn managed_alias_schema_keeps_the_production_default() {
        let schema = schema(CURATED[0], true, car_auth::DEFAULT_API_BASE);
        let ModelSource::Proprietary { endpoint, auth, .. } = schema.source else {
            panic!("managed alias must use the Parslee proprietary transport");
        };
        assert_eq!(endpoint, car_auth::DEFAULT_API_BASE);
        let ProprietaryAuth::OAuth2Pkce { authority, .. } = auth else {
            panic!("managed alias must use the Parslee OAuth connection");
        };
        assert_eq!(authority, car_auth::DEFAULT_API_BASE);
    }

    #[test]
    fn managed_aliases_match_the_gateway_contract() {
        let managed: Vec<_> = curated_schemas()
            .into_iter()
            .filter(|schema| schema.provider == "parslee")
            .collect();
        assert!(
            managed.iter().all(is_managed_gateway_schema),
            "every reviewed alias must satisfy the managed gateway predicate"
        );
        let ids: Vec<_> = managed.into_iter().map(|schema| schema.id).collect();
        assert_eq!(
            ids,
            [
                "parslee/openrouter/frontier-general",
                "parslee/openrouter/frontier-deep",
                "parslee/openrouter/frontier-multimodal",
                "parslee/openrouter/balanced-general",
                "parslee/openrouter/open-multimodal",
                "parslee/openrouter/open-long-context",
                "parslee/openrouter/open-reasoning",
                "parslee/openrouter/open-fast",
                "parslee/openrouter/coding-frontier",
                "parslee/openrouter/coding-efficient",
                "parslee/openrouter/frontier-deep-next",
            ]
        );
    }

    #[test]
    fn managed_gateway_predicate_rejects_allowlisted_id_with_wrong_selector_name() {
        let mut schema = curated_schemas()
            .into_iter()
            .find(|schema| schema.id == "parslee/openrouter/frontier-general")
            .expect("managed frontier alias");
        schema.name = "attacker-controlled-upstream-selector".into();

        assert!(
            !is_managed_gateway_schema(&schema),
            "an allowlisted id cannot make a different outbound selector trusted"
        );
    }

    #[test]
    fn managed_aliases_deserialize_the_gateway_models_fixture_without_upstream_ids() {
        let fixture: serde_json::Value = serde_json::from_str(include_str!(
            "../tests/fixtures/parslee-openrouter-models.json"
        ))
        .expect("gateway fixture must be valid JSON");
        assert_eq!(fixture["object"], "list");
        let rows = fixture["data"].as_array().expect("data array");
        let ids: Vec<_> = rows
            .iter()
            .map(|row| row["id"].as_str().expect("opaque id"))
            .collect();
        let projected: Vec<_> = curated_schemas()
            .into_iter()
            .filter(|schema| schema.provider == "parslee")
            .map(|schema| schema.id)
            .collect();
        assert_eq!(ids, projected);
        for row in rows {
            assert_eq!(row["object"], "model");
            assert_eq!(row["owned_by"], "parslee");
            assert_eq!(row["provider"], "openrouter");
            assert!(row.get("upstream_model").is_none());
            assert!(!row.to_string().contains("openai/"));
            assert!(!row.to_string().contains("anthropic/"));
            assert!(row["capabilities"].is_array());
            assert!(row["tags"].is_array());
        }
    }

    #[test]
    fn managed_aliases_advertise_only_parameters_the_parslee_transport_forwards() {
        let managed: Vec<_> = builtin_schemas()
            .into_iter()
            .filter(|schema| schema.provider == "parslee")
            .collect();
        assert_eq!(managed.len(), CURATED.len());
        for schema in managed {
            assert_eq!(
                schema.supported_params,
                vec![P::MaxTokens],
                "{} advertises a request control the CAR-to-Parslee transport drops",
                schema.id
            );
        }
    }

    #[test]
    fn projections_are_bounded_distinct_and_price_identical() {
        let schemas = curated_schemas();
        assert_eq!(schemas.len(), CURATED.len() * 2);
        for model in CURATED {
            let personal = schemas
                .iter()
                .find(|s| s.id == format!("openrouter/{}", model.upstream_id))
                .unwrap();
            let gateway = schemas
                .iter()
                .find(|s| s.id == format!("parslee/openrouter/{}", model.managed_alias))
                .unwrap();
            assert_eq!(personal.cost.input_per_mtok, gateway.cost.input_per_mtok);
            assert_eq!(personal.cost.output_per_mtok, gateway.cost.output_per_mtok);
            assert_eq!(
                personal.cost.cache_read_input_per_mtok,
                gateway.cost.cache_read_input_per_mtok
            );
            assert_eq!(
                personal.cost.cache_write_input_per_mtok,
                gateway.cost.cache_write_input_per_mtok
            );
            assert_eq!(personal.cost.pricing_tiers, gateway.cost.pricing_tiers);
            assert_eq!(personal.capabilities, gateway.capabilities);
        }
    }
    #[test]
    fn only_calibrated_models_are_frontier() {
        let frontier: Vec<_> = curated_schemas()
            .into_iter()
            .filter(|s| s.provider == "openrouter" && s.tags.iter().any(|t| t == "frontier"))
            .collect();
        assert_eq!(frontier.len(), 4);
        assert!(frontier
            .iter()
            .all(|s| !s.tags.iter().any(|t| t == "cheap")));
    }

    #[test]
    fn personal_registry_is_exactly_the_reviewed_rows() {
        let personal: Vec<_> = curated_schemas()
            .into_iter()
            .filter(|schema| schema.provider == "openrouter")
            .collect();
        let ids: Vec<_> = personal.iter().map(|schema| schema.id.as_str()).collect();
        assert_eq!(
            ids,
            [
                "openrouter/openai/gpt-5.4",
                "openrouter/anthropic/claude-opus-4.6",
                "openrouter/google/gemini-3.1-pro-preview",
                "openrouter/anthropic/claude-sonnet-4.6",
                "openrouter/moonshotai/kimi-k2.5",
                "openrouter/qwen/qwen3.5-plus-02-15",
                "openrouter/deepseek/deepseek-v3.2",
                "openrouter/minimax/minimax-m2.5",
                "openrouter/openai/gpt-5.3-codex",
                "openrouter/qwen/qwen3-coder-next",
                "openrouter/anthropic/claude-opus-4.8",
            ]
        );
        assert!(personal.iter().all(|schema| {
            schema.trust_tier == TrustTier::Curated
                && schema.tags.iter().any(|tag| tag == "openrouter")
                && schema.tags.iter().any(|tag| tag == "personal-key")
                && !schema.tags.iter().any(|tag| tag == "dynamic")
        }));
    }

    #[test]
    fn personal_capabilities_and_params_match_reviewed_chat_transport_contract() {
        let handler = crate::protocol::handler_for(ApiProtocol::OpenRouter);
        assert!(!handler.supports_audio());
        assert!(!handler.supports_video());
        let expected = [
            ("openrouter/openai/gpt-5.4", CODE_REASON_VISION),
            ("openrouter/anthropic/claude-opus-4.6", CODE_REASON_VISION),
            (
                "openrouter/google/gemini-3.1-pro-preview",
                CODE_REASON_VISION,
            ),
            ("openrouter/anthropic/claude-sonnet-4.6", CODE_REASON_VISION),
            ("openrouter/moonshotai/kimi-k2.5", CODE_REASON_VISION),
            ("openrouter/qwen/qwen3.5-plus-02-15", CODE_REASON_VISION),
            ("openrouter/deepseek/deepseek-v3.2", CODE_REASON),
            ("openrouter/minimax/minimax-m2.5", CODE_REASON_MULTI_TOOL),
            ("openrouter/openai/gpt-5.3-codex", CODE_REASON_VISION),
            ("openrouter/qwen/qwen3-coder-next", CODE),
            ("openrouter/anthropic/claude-opus-4.8", CODE_REASON_VISION),
        ];
        let personal: Vec<_> = curated_schemas()
            .into_iter()
            .filter(|schema| schema.provider == "openrouter")
            .collect();
        for (id, capabilities) in expected {
            let schema = personal.iter().find(|schema| schema.id == id).unwrap();
            assert_eq!(
                schema.capabilities, capabilities,
                "{id} capability metadata drifted from the reviewed model contract"
            );
            assert!(!schema.has_capability(C::VideoUnderstanding));
            assert!(!schema.has_capability(C::AudioUnderstanding));
            assert!(!schema.has_capability(C::ImageGeneration));
            assert!(!schema.has_capability(C::VideoGeneration));
        }

        let transport_params = [
            P::Temperature,
            P::MaxTokens,
            P::ResponseFormat,
            P::ExtendedThinking,
        ];
        for schema in personal {
            assert!(
                schema
                    .supported_params
                    .iter()
                    .all(|parameter| transport_params.contains(parameter)),
                "{} advertises a request parameter CAR's OpenRouter Chat adapter does not send",
                schema.id
            );
        }
    }

    #[test]
    fn current_catalog_prices_include_tiers_and_per_model_cache_rates() {
        let schemas = curated_schemas();
        let gpt = schemas
            .iter()
            .find(|schema| schema.id == "openrouter/openai/gpt-5.4")
            .unwrap();
        assert_eq!(gpt.cost.prices_for(271_999).input_per_mtok, Some(2.5));
        assert_eq!(gpt.cost.prices_for(272_000).input_per_mtok, Some(5.0));
        assert_eq!(gpt.cost.prices_for(272_000).output_per_mtok, Some(22.5));
        assert_eq!(
            gpt.cost.prices_for(272_000).cache_read_input_per_mtok,
            Some(0.5)
        );

        let gemini = schemas
            .iter()
            .find(|schema| schema.id == "openrouter/google/gemini-3.1-pro-preview")
            .unwrap();
        assert_eq!(
            gemini.cost.prices_for(199_999).cache_write_input_per_mtok,
            Some(0.375)
        );
        assert_eq!(gemini.cost.prices_for(200_000).input_per_mtok, Some(4.0));
        assert_eq!(gemini.cost.prices_for(200_000).output_per_mtok, Some(18.0));
        assert_eq!(
            gemini.cost.prices_for(200_000).cache_read_input_per_mtok,
            Some(0.4)
        );

        let kimi = schemas
            .iter()
            .find(|schema| schema.id == "openrouter/moonshotai/kimi-k2.5")
            .unwrap();
        assert_eq!(kimi.cost.cache_read_input_per_mtok, Some(0.095));
        assert_eq!(kimi.cost.cache_write_input_per_mtok, None);
    }

    /// The one place `estimated_usd_bounded`'s substitution ranges are
    /// load-bearing rather than merely descriptive.
    ///
    /// A missing cache-WRITE rate is flagged in both directions, so it stays
    /// honest at any ratio; a missing OUTPUT rate refuses, so refusing is never
    /// wrong. But a missing cache-READ rate claims a real `≤` ceiling, and that
    /// ceiling rests on cache reads being cheaper than uncached input. The
    /// argument is sound — a cache read is a substitute good for sending the
    /// tokens fresh, so a provider pricing it above input would be selling
    /// something whose rational use is "don't use it" — but an argument in a
    /// docstring is not a guard. This makes the assumption fail loudly the day
    /// a provider violates it, instead of silently issuing a ceiling that is
    /// not one.
    #[test]
    fn no_curated_cache_read_rate_exceeds_its_input_rate() {
        for schema in curated_schemas() {
            // Check the base sheet and every declared tier: a tier can
            // override either component independently.
            let thresholds = std::iter::once(0).chain(
                schema
                    .cost
                    .pricing_tiers
                    .iter()
                    .map(|tier| tier.min_prompt_tokens),
            );
            for threshold in thresholds {
                let prices = schema.cost.prices_for(threshold);
                let (Some(input), Some(cache_read)) =
                    (prices.input_per_mtok, prices.cache_read_input_per_mtok)
                else {
                    continue;
                };
                assert!(
                    cache_read <= input,
                    "{} prices a cache read at {cache_read}/MTok above its {input}/MTok input \
                     rate at prompt threshold {threshold}. The `≤` ceiling that \
                     `estimated_usd_bounded` puts on a MISSING cache-read rate assumes this \
                     never happens — if a provider now surcharges cached reads, that \
                     substitution needs both direction flags the way cache write does.",
                    schema.id
                );
            }
        }
    }

    /// Live `https://openrouter.ai/api/v1/models` row for
    /// `anthropic/claude-opus-4.8`, read 2026-07-31: `context_length`
    /// 1000000, `top_provider.max_completion_tokens` 128000, and USD/token
    /// pricing strings `prompt: "0.000005"`, `completion: "0.000025"`,
    /// `input_cache_read: "0.0000005"`, `input_cache_write: "0.00000625"` —
    /// each × 1e6 for this catalog's USD/MTok unit.
    #[test]
    fn claude_opus_4_8_matches_the_live_openrouter_catalog_row() {
        let schemas = curated_schemas();
        let personal = schemas
            .iter()
            .find(|schema| schema.id == "openrouter/anthropic/claude-opus-4.8")
            .expect("built-in curated row, no runtime registration");
        assert_eq!(personal.provider, "openrouter");
        assert_eq!(personal.context_length, 1_000_000);
        assert_eq!(personal.max_output_tokens, Some(128_000));
        assert_eq!(personal.cost.input_per_mtok, Some(5.0));
        assert_eq!(personal.cost.output_per_mtok, Some(25.0));
        assert_eq!(personal.cost.cache_read_input_per_mtok, Some(0.5));
        assert_eq!(personal.cost.cache_write_input_per_mtok, Some(6.25));
        assert!(personal.cost.pricing_tiers.is_empty());
        assert_eq!(personal.version, "2026-07-31");

        // The alias fronts it without naming it.
        let alias = schemas
            .iter()
            .find(|schema| schema.id == "parslee/openrouter/frontier-deep-next")
            .expect("matching managed alias row");
        assert!(is_managed_gateway_schema(alias));
        assert!(!serde_json::to_string(alias).unwrap().contains("opus-4.8"));
    }

    /// Frozen expectations for the ten rows that predate `claude-opus-4.8`,
    /// so a future catalog edit that silently reprices or resizes one fails
    /// here rather than in a customer's invoice.
    ///
    /// This cements the **2026-07-22 snapshot**, deliberately. Live OpenRouter
    /// has already drifted from it in at least one place (`qwen3-coder-next`
    /// publishes 0.12 input today against the 0.11 frozen here). That is not a
    /// bug to fix in passing: each row declares the date it was read on, and
    /// re-snapshotting prices is a separate, deliberate act that updates the
    /// numbers and the `snapshot` date together. If you are here because this
    /// test failed after a re-snapshot, update both.
    ///
    /// Capabilities are asserted by
    /// `personal_capabilities_and_params_match_reviewed_chat_transport_contract`,
    /// which is why this one does not claim them in its name.
    #[test]
    fn the_ten_preexisting_rows_keep_their_ids_prices_and_sizes() {
        // id, input, output, cache read, cache write, context, max output.
        let expected: &[(&str, f64, f64, Option<f64>, Option<f64>, usize, usize)] = &[
            (
                "openrouter/openai/gpt-5.4",
                2.5,
                15.0,
                Some(0.25),
                None,
                1_050_000,
                128_000,
            ),
            (
                "openrouter/anthropic/claude-opus-4.6",
                5.0,
                25.0,
                Some(0.5),
                Some(6.25),
                1_000_000,
                128_000,
            ),
            (
                "openrouter/google/gemini-3.1-pro-preview",
                2.0,
                12.0,
                Some(0.2),
                Some(0.375),
                1_048_576,
                65_536,
            ),
            (
                "openrouter/anthropic/claude-sonnet-4.6",
                3.0,
                15.0,
                Some(0.3),
                Some(3.75),
                1_000_000,
                128_000,
            ),
            (
                "openrouter/moonshotai/kimi-k2.5",
                0.57,
                2.85,
                Some(0.095),
                None,
                262_144,
                262_144,
            ),
            (
                "openrouter/qwen/qwen3.5-plus-02-15",
                0.26,
                1.56,
                None,
                None,
                1_000_000,
                65_536,
            ),
            (
                "openrouter/deepseek/deepseek-v3.2",
                0.269,
                0.4,
                Some(0.1345),
                None,
                163_840,
                65_536,
            ),
            (
                "openrouter/minimax/minimax-m2.5",
                0.15,
                0.9,
                Some(0.05),
                None,
                204_800,
                196_608,
            ),
            (
                "openrouter/openai/gpt-5.3-codex",
                1.75,
                14.0,
                Some(0.175),
                None,
                400_000,
                128_000,
            ),
            (
                "openrouter/qwen/qwen3-coder-next",
                0.11,
                0.8,
                Some(0.07),
                None,
                262_144,
                262_144,
            ),
        ];
        let schemas = curated_schemas();
        assert_eq!(
            expected.len() + 1,
            curated_model_count(),
            "this list must cover every curated row except the one added after it"
        );
        for (id, input, output, cache_read, cache_write, context, max_output) in expected {
            let schema = schemas
                .iter()
                .find(|schema| &schema.id == id)
                .unwrap_or_else(|| panic!("{id} must not be removed or renamed"));
            assert_eq!(schema.context_length, *context, "{id} context window");
            assert_eq!(
                schema.max_output_tokens,
                Some(*max_output),
                "{id} max output tokens"
            );
            assert_eq!(schema.cost.input_per_mtok, Some(*input), "{id} input price");
            assert_eq!(
                schema.cost.output_per_mtok,
                Some(*output),
                "{id} output price"
            );
            assert_eq!(
                schema.cost.cache_read_input_per_mtok, *cache_read,
                "{id} cache-read price"
            );
            assert_eq!(
                schema.cost.cache_write_input_per_mtok, *cache_write,
                "{id} cache-write price"
            );
            assert_eq!(schema.version, "2026-07-22", "{id} snapshot date");
        }

        // The one high-context tier nothing else pinned. gpt-5.4's and
        // gemini's tiers are asserted by
        // `current_catalog_prices_include_tiers_and_per_model_cache_rates`;
        // qwen3.5-plus's was unguarded until now.
        let qwen = schemas
            .iter()
            .find(|schema| schema.id == "openrouter/qwen/qwen3.5-plus-02-15")
            .unwrap();
        assert_eq!(qwen.cost.prices_for(255_999).input_per_mtok, Some(0.26));
        assert_eq!(qwen.cost.prices_for(255_999).output_per_mtok, Some(1.56));
        assert_eq!(qwen.cost.prices_for(256_000).input_per_mtok, Some(0.325));
        assert_eq!(qwen.cost.prices_for(256_000).output_per_mtok, Some(1.95));
        // It publishes no cache rates at either tier — the case that makes the
        // stats column report a bounded estimate rather than a fabricated $0.
        assert_eq!(
            qwen.cost.prices_for(256_000).cache_read_input_per_mtok,
            None
        );
        assert_eq!(
            qwen.cost.prices_for(256_000).cache_write_input_per_mtok,
            None
        );
    }
}