akribes-sdk 0.22.6

//! Integration tests for the bench sub-clients (`BenchClient` +
//! `BenchRunsClient`).
//!
//! Each test asserts the *request shape the SDK puts on the wire* (method,
//! path, query, body JSON) against what the corresponding `akribes-server`
//! bench handler parses, plus correct deserialisation of the response body.
//! The wire shapes mirror `crates/akribes-server/src/handlers/bench.rs` and
//! `crates/akribes-server/src/models.rs`.

use akribes_sdk::AkribesError;
use akribes_sdk::models::{
    CreateBenchCaseRequest, CreateOrUpdateBenchRequest, PatchBenchCaseRequest, PromoteCaseEdits,
    PromoteExecutionRequest, TriggerBenchRunRequest,
};
use akribes_sdk::{AkribesClient, BenchClient, BenchRunsClient};
use mockito::{Matcher, Server};

fn make_client(server: &Server) -> AkribesClient {
    AkribesClient::builder(server.url())
        .project_id(7)
        .name("bench-test")
        .id("bench-id")
        .build()
}

fn bench(server: &Server) -> BenchClient {
    make_client(server).project(7).bench()
}

fn runs(server: &Server) -> BenchRunsClient {
    make_client(server).bench_runs()
}

// ── Project-wide summary ──────────────────────────────────────────────────────

#[tokio::test]
async fn list_project_summaries_maps_to_projects_benches_route() {
    let mut server = Server::new_async().await;
    // GET /projects/{id}/benches — the project id comes from the scope (7),
    // NOT a query param. Mirror the server's `list_project_benches` row shape
    // (crates/akribes-server/src/handlers/bench.rs / models.rs
    // ProjectBenchSummary): full row with a populated latest-run block.
    let _m = server
        .mock("GET", "/projects/7/benches")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"[{"bench_id":3,"script_id":11,"script_name":"summarise",
                "judge_script_id":12,"judge_script_name":"judge_quality",
                "judge_channel":"production","case_count":18,
                "latest_run_id":42,"latest_run_status":"completed",
                "latest_run_channel":"production","latest_run_workflow_version_id":99,
                "latest_run_at":"2026-01-02T00:00:00Z","latest_run_mean_score":0.91,
                "latest_run_cost_usd":0.42,"updated_at":"2026-01-02T00:00:00Z"}]"#,
        )
        .create_async()
        .await;
    let rows = bench(&server).list_project_summaries().await.unwrap();
    assert_eq!(rows.len(), 1);
    let r = &rows[0];
    assert_eq!(r.bench_id, 3);
    assert_eq!(r.script_name, "summarise");
    assert_eq!(r.judge_script_id, Some(12));
    assert_eq!(r.judge_script_name.as_deref(), Some("judge_quality"));
    assert_eq!(r.case_count, 18);
    assert_eq!(r.latest_run_id, Some(42));
    assert_eq!(r.latest_run_status.as_deref(), Some("completed"));
    assert_eq!(r.latest_run_mean_score, Some(0.91));
    assert_eq!(r.latest_run_cost_usd, Some(0.42));
}

#[tokio::test]
async fn list_project_summaries_tolerates_null_latest_run_block() {
    let mut server = Server::new_async().await;
    // A bench that has never been run leaves the entire latest-run block
    // NULL on the server (LEFT JOIN LATERAL miss). Every `latest_run_*`
    // field plus judge fields are `#[serde(default)]` Options, so the row
    // must still deserialise.
    let _m = server
        .mock("GET", "/projects/7/benches")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"[{"bench_id":4,"script_id":13,"script_name":"never_run",
                "judge_script_id":null,"judge_script_name":null,
                "judge_channel":"draft","case_count":0,
                "latest_run_id":null,"latest_run_status":null,
                "latest_run_channel":null,"latest_run_workflow_version_id":null,
                "latest_run_at":null,"latest_run_mean_score":null,
                "latest_run_cost_usd":null,"updated_at":"2026-01-02T00:00:00Z"}]"#,
        )
        .create_async()
        .await;
    let rows = bench(&server).list_project_summaries().await.unwrap();
    assert_eq!(rows.len(), 1);
    assert_eq!(rows[0].script_name, "never_run");
    assert_eq!(rows[0].case_count, 0);
    assert!(rows[0].latest_run_id.is_none());
    assert!(rows[0].judge_script_id.is_none());
    assert_eq!(rows[0].judge_channel, "draft");
}

#[tokio::test]
async fn list_project_summaries_404_is_empty() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/projects/7/benches")
        .with_status(404)
        .create_async()
        .await;
    assert!(
        bench(&server)
            .list_project_summaries()
            .await
            .unwrap()
            .is_empty()
    );
}

// ── Bench config CRUD ────────────────────────────────────────────────────────

#[tokio::test]
async fn get_bench_hit() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/projects/7/scripts/summarise/bench")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"id":3,"script_id":11,"judge_script_id":12,"judge_channel":"production",
                "config":{"k":"v"},"created_at":"2026-01-01T00:00:00Z",
                "updated_at":"2026-01-02T00:00:00Z"}"#,
        )
        .create_async()
        .await;

    let got = bench(&server).get("summarise").await.unwrap();
    let b = got.expect("bench present");
    assert_eq!(b.id, 3);
    assert_eq!(b.script_id, 11);
    assert_eq!(b.judge_script_id, Some(12));
    assert_eq!(b.judge_channel, "production");
    assert_eq!(b.config["k"], "v");
}

#[tokio::test]
async fn get_bench_404_is_none() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/projects/7/scripts/nope/bench")
        .with_status(404)
        .create_async()
        .await;

    assert!(bench(&server).get("nope").await.unwrap().is_none());
}

#[tokio::test]
async fn create_or_update_bench_sends_only_set_fields() {
    let mut server = Server::new_async().await;
    // The SDK omits `None` fields via skip_serializing_if. A request that
    // only sets `judge_script_id` must NOT carry `judge_channel`/`config`
    // keys — the server defaults them, and an explicit `null` would change
    // behaviour. Assert exact body.
    let _m = server
        .mock("POST", "/projects/7/scripts/summarise/bench")
        .match_body(Matcher::Json(serde_json::json!({"judge_script_id": 12})))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"id":3,"script_id":11,"judge_script_id":12,"judge_channel":"draft",
                "config":{},"created_at":"2026-01-01T00:00:00Z",
                "updated_at":"2026-01-01T00:00:00Z"}"#,
        )
        .create_async()
        .await;

    let req = CreateOrUpdateBenchRequest {
        judge_script_id: Some(12),
        judge_channel: None,
        config: None,
    };
    let b = bench(&server)
        .create_or_update("summarise", &req)
        .await
        .unwrap();
    assert_eq!(b.judge_channel, "draft");
}

#[tokio::test]
async fn delete_bench_true_on_200() {
    let mut server = Server::new_async().await;
    // delete_bench returns 200 with `{"deleted":true}`, NOT 204. The SDK's
    // generic delete returns `status != 404` → true.
    let _m = server
        .mock("DELETE", "/projects/7/scripts/summarise/bench")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(r#"{"deleted":true}"#)
        .create_async()
        .await;
    assert!(bench(&server).delete("summarise").await.unwrap());
}

#[tokio::test]
async fn delete_bench_false_on_404() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("DELETE", "/projects/7/scripts/gone/bench")
        .with_status(404)
        .create_async()
        .await;
    assert!(!bench(&server).delete("gone").await.unwrap());
}

#[tokio::test]
async fn get_signature_maps_to_signature_route_not_bench() {
    let mut server = Server::new_async().await;
    // Regression guard: the signature endpoint is
    // `/projects/{id}/scripts/{name}/signature`, NOT under `/bench`.
    let _m = server
        .mock("GET", "/projects/7/scripts/summarise/signature")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(r#"{"inputs":[{"name":"q","type":"String"}]}"#)
        .create_async()
        .await;
    let sig = bench(&server).get_signature("summarise").await.unwrap();
    assert_eq!(sig["inputs"][0]["name"], "q");
}

#[tokio::test]
async fn get_signature_404_yields_empty_object() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/projects/7/scripts/x/signature")
        .with_status(404)
        .create_async()
        .await;
    let sig = bench(&server).get_signature("x").await.unwrap();
    assert_eq!(sig, serde_json::json!({}));
}

#[tokio::test]
async fn contract_preview_sends_judge_and_channel_query() {
    let mut server = Server::new_async().await;
    // judge is required; channel only present when Some.
    let _m = server
        .mock(
            "GET",
            "/projects/7/scripts/summarise/bench/contract-preview",
        )
        .match_query(Matcher::AllOf(vec![
            Matcher::UrlEncoded("judge".into(), "12".into()),
            Matcher::UrlEncoded("channel".into(), "production".into()),
        ]))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(r#"{"breaks":[]}"#)
        .create_async()
        .await;
    let prev = bench(&server)
        .contract_preview("summarise", 12, Some("production"))
        .await
        .unwrap();
    assert_eq!(prev["breaks"], serde_json::json!([]));
}

#[tokio::test]
async fn contract_preview_omits_channel_when_none() {
    let mut server = Server::new_async().await;
    // Exact full query: only `judge`, no `channel` key at all.
    let _m = server
        .mock(
            "GET",
            "/projects/7/scripts/summarise/bench/contract-preview",
        )
        .match_query(Matcher::Exact("judge=12".into()))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(r#"{"breaks":[]}"#)
        .create_async()
        .await;
    bench(&server)
        .contract_preview("summarise", 12, None)
        .await
        .unwrap();
}

// ── Cases ────────────────────────────────────────────────────────────────────

#[tokio::test]
async fn list_cases_404_is_empty() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/projects/7/scripts/summarise/bench/cases")
        .with_status(404)
        .create_async()
        .await;
    assert!(
        bench(&server)
            .list_cases("summarise")
            .await
            .unwrap()
            .is_empty()
    );
}

#[tokio::test]
async fn list_cases_deserialises_rows() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/projects/7/scripts/summarise/bench/cases")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"[{"id":"case_a","project_id":7,"script_name":"summarise","bench_id":3,
                "kind":"case","frozen":true,"case_name":"happy",
                "inputs":{"q":"hi"},"expected_output":{"a":1},
                "created_at":"2026-01-01T00:00:00Z"}]"#,
        )
        .create_async()
        .await;
    let cases = bench(&server).list_cases("summarise").await.unwrap();
    assert_eq!(cases.len(), 1);
    assert_eq!(cases[0].id, "case_a");
    assert_eq!(cases[0].case_name.as_deref(), Some("happy"));
    assert!(cases[0].frozen);
    assert_eq!(cases[0].inputs.as_ref().unwrap()["q"], "hi");
}

#[tokio::test]
async fn create_case_sends_inputs_and_optional_fields() {
    let mut server = Server::new_async().await;
    // `inputs` is required and always serialised; expected_output present,
    // ground_truth/name omitted when None.
    let _m = server
        .mock("POST", "/projects/7/scripts/summarise/bench/cases")
        .match_body(Matcher::Json(serde_json::json!({
            "inputs": {"q": "hi"},
            "expected_output": {"a": 1},
        })))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"id":"case_new","project_id":7,"script_name":"summarise","kind":"case",
                "frozen":true,"created_at":"2026-01-01T00:00:00Z"}"#,
        )
        .create_async()
        .await;
    let req = CreateBenchCaseRequest {
        inputs: serde_json::json!({"q": "hi"}),
        expected_output: Some(serde_json::json!({"a": 1})),
        ground_truth: None,
        name: None,
    };
    let case = bench(&server).create_case("summarise", &req).await.unwrap();
    assert_eq!(case.id, "case_new");
}

#[tokio::test]
async fn case_contract_drift_404_yields_empty_report() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock(
            "GET",
            "/projects/7/scripts/summarise/bench/cases/contract-drift",
        )
        .with_status(404)
        .create_async()
        .await;
    let report = bench(&server)
        .case_contract_drift("summarise")
        .await
        .unwrap();
    assert!(report.drifted.is_empty());
    assert_eq!(report.summary, "");
}

#[tokio::test]
async fn case_contract_drift_parses_drifted_rows() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock(
            "GET",
            "/projects/7/scripts/summarise/bench/cases/contract-drift",
        )
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"drifted":[{"case_id":"case_a","label":"happy",
                "what_broke":"outputs.score removed"}],
                "script_version_id":99,"published_at":"2026-01-01T00:00:00Z",
                "published_by":"alice","summary":"1 case drifted"}"#,
        )
        .create_async()
        .await;
    let report = bench(&server)
        .case_contract_drift("summarise")
        .await
        .unwrap();
    assert_eq!(report.drifted.len(), 1);
    assert_eq!(report.drifted[0].case_id, "case_a");
    assert_eq!(report.drifted[0].what_broke, "outputs.score removed");
    assert_eq!(report.script_version_id, Some(99));
    assert_eq!(report.summary, "1 case drifted");
}

// ── Runs (project-scoped) ────────────────────────────────────────────────────

#[tokio::test]
async fn list_runs_sends_limit_offset_query() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/projects/7/scripts/summarise/bench/runs")
        .match_query(Matcher::AllOf(vec![
            Matcher::UrlEncoded("limit".into(), "5".into()),
            Matcher::UrlEncoded("offset".into(), "10".into()),
        ]))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"[{"id":1,"bench_id":3,"channel":"production","workflow_version_id":99,
                "judge_version_id":100,"status":"completed","triggered_at":"2026-01-01T00:00:00Z",
                "mean_headline_score":0.91,"ok_cases":7}]"#,
        )
        .create_async()
        .await;
    let runs = bench(&server)
        .list_runs("summarise", Some(5), Some(10))
        .await
        .unwrap();
    assert_eq!(runs.len(), 1);
    assert_eq!(runs[0].id, 1);
    assert_eq!(runs[0].mean_headline_score, Some(0.91));
    assert_eq!(runs[0].ok_cases, Some(7));
}

#[tokio::test]
async fn list_runs_omits_query_when_no_pagination() {
    let mut server = Server::new_async().await;
    // Both None → no query string at all (the helper returns the bare base
    // URL). A stray `?limit=&offset=` would change the server's parse.
    let _m = server
        .mock("GET", "/projects/7/scripts/summarise/bench/runs")
        .match_query(Matcher::Missing)
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body("[]")
        .create_async()
        .await;
    let runs = bench(&server)
        .list_runs("summarise", None, None)
        .await
        .unwrap();
    assert!(runs.is_empty());
}

#[tokio::test]
async fn trigger_run_sends_channel_required_and_case_ids() {
    let mut server = Server::new_async().await;
    // channel is non-optional; case_ids present, notes omitted when None.
    let _m = server
        .mock("POST", "/projects/7/scripts/summarise/bench/runs")
        .match_body(Matcher::Json(serde_json::json!({
            "channel": "production",
            "case_ids": ["case_a", "case_b"],
        })))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"id":42,"bench_id":3,"channel":"production","workflow_version_id":99,
                "judge_version_id":100,"status":"pending","triggered_at":"2026-01-01T00:00:00Z"}"#,
        )
        .create_async()
        .await;
    let req = TriggerBenchRunRequest {
        channel: "production".into(),
        notes: None,
        case_ids: Some(vec!["case_a".into(), "case_b".into()]),
    };
    let run = bench(&server).trigger_run("summarise", &req).await.unwrap();
    assert_eq!(run.id, 42);
    assert_eq!(run.status, "pending");
}

// ── Run-scoped (cross-project) ───────────────────────────────────────────────

#[tokio::test]
async fn run_get_404_is_none() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/bench-runs/42")
        .with_status(404)
        .create_async()
        .await;
    assert!(runs(&server).get(42).await.unwrap().is_none());
}

#[tokio::test]
async fn run_delete_succeeds_on_204() {
    let mut server = Server::new_async().await;
    // The server emits 204 No Content; the SDK discards the bool → Ok(()).
    let _m = server
        .mock("DELETE", "/bench-runs/42")
        .with_status(204)
        .create_async()
        .await;
    runs(&server).delete(42).await.unwrap();
}

#[tokio::test]
async fn list_results_deserialises() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/bench-runs/42/results")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"[{"id":1,"bench_run_id":42,"case_id":"case_a","headline_score":0.8,
                "status":"ok","created_at":"2026-01-01T00:00:00Z"}]"#,
        )
        .create_async()
        .await;
    let results = runs(&server).list_results(42).await.unwrap();
    assert_eq!(results.len(), 1);
    assert_eq!(results[0].case_id, "case_a");
    assert_eq!(results[0].headline_score, Some(0.8));
    assert_eq!(results[0].status, "ok");
}

#[tokio::test]
async fn cancel_posts_empty_body_and_returns_run() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("POST", "/bench-runs/42/cancel")
        .match_body(Matcher::Json(serde_json::json!({})))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"id":42,"bench_id":3,"channel":"production","workflow_version_id":99,
                "judge_version_id":100,"status":"canceled","triggered_at":"2026-01-01T00:00:00Z"}"#,
        )
        .create_async()
        .await;
    let run = runs(&server).cancel(42).await.unwrap();
    assert_eq!(run.status, "canceled");
}

#[tokio::test]
async fn tag_session_patches_with_mcp_session_id() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("PATCH", "/bench-runs/42/tag-session")
        .match_body(Matcher::Json(serde_json::json!({
            "mcp_session_id": "sess_abc"
        })))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(r#"{"tagged":true,"run_id":42,"mcp_session_id":"sess_abc"}"#)
        .create_async()
        .await;
    let resp = runs(&server).tag_session(42, "sess_abc").await.unwrap();
    assert!(resp.tagged);
    assert_eq!(resp.run_id, 42);
    assert_eq!(resp.mcp_session_id, "sess_abc");
}

#[tokio::test]
async fn promote_execution_posts_edits_overlay() {
    let mut server = Server::new_async().await;
    // promote-to-case lives on /executions/{id}, NOT /bench-runs.
    let _m = server
        .mock("POST", "/executions/exec_123/promote-to-case")
        .match_body(Matcher::Json(serde_json::json!({
            "edits": {"expected_output": {"a": 2}},
            "name": "T_1",
        })))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"id":"case_p","project_id":7,"script_name":"summarise","kind":"case",
                "frozen":true,"case_name":"T_1","created_at":"2026-01-01T00:00:00Z"}"#,
        )
        .create_async()
        .await;
    let req = PromoteExecutionRequest {
        edits: Some(PromoteCaseEdits {
            inputs: None,
            expected_output: Some(serde_json::json!({"a": 2})),
            ground_truth: None,
        }),
        name: Some("T_1".into()),
    };
    let case = runs(&server)
        .promote_execution("exec_123", &req)
        .await
        .unwrap();
    assert_eq!(case.id, "case_p");
    assert_eq!(case.case_name.as_deref(), Some("T_1"));
}

#[tokio::test]
async fn promote_execution_url_encodes_exec_id() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("POST", "/executions/exec%2F123/promote-to-case")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"id":"c","project_id":7,"script_name":"s","kind":"case","frozen":true,
                "created_at":"2026-01-01T00:00:00Z"}"#,
        )
        .create_async()
        .await;
    runs(&server)
        .promote_execution("exec/123", &PromoteExecutionRequest::default())
        .await
        .unwrap();
}

#[tokio::test]
async fn compare_parses_aggregate_and_per_case() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/bench-runs/10/compare/20")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"run_a_id":10,"run_b_id":20,
                "aggregate":{"mean_score_delta":0.05,"cost_delta_usd":-0.1,
                    "n_regressed":1,"n_improved":3,"n_unchanged":2},
                "per_case":[{"case_id":"case_a","case_label":"happy","score_a":0.8,
                    "score_b":0.9,"delta":0.1,"flag":"improved"}]}"#,
        )
        .create_async()
        .await;
    let report = runs(&server).compare(10, 20).await.unwrap();
    assert_eq!(report.run_a_id, 10);
    assert_eq!(report.aggregate.n_improved, 3);
    assert_eq!(report.per_case.len(), 1);
    assert_eq!(report.per_case[0].flag, "improved");
    assert_eq!(report.per_case[0].delta, Some(0.1));
}

#[tokio::test]
async fn compare_404_surfaces_http_status() {
    let mut server = Server::new_async().await;
    // compare uses get_opt then maps None→HttpStatus 404 (not a silent
    // empty report) because there's no meaningful empty comparison.
    let _m = server
        .mock("GET", "/bench-runs/10/compare/20")
        .with_status(404)
        .create_async()
        .await;
    let err = runs(&server).compare(10, 20).await.unwrap_err();
    match err {
        AkribesError::HttpStatus { status, .. } => assert_eq!(status, 404),
        other => panic!("expected HttpStatus 404, got {other:?}"),
    }
}

#[tokio::test]
async fn patch_case_sends_sparse_update() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("PATCH", "/cases/case_a")
        .match_body(Matcher::Json(serde_json::json!({"name": "renamed"})))
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(
            r#"{"id":"case_a","project_id":7,"script_name":"s","kind":"case","frozen":true,
                "case_name":"renamed","created_at":"2026-01-01T00:00:00Z"}"#,
        )
        .create_async()
        .await;
    let req = PatchBenchCaseRequest {
        name: Some("renamed".into()),
        ..Default::default()
    };
    let case = runs(&server).patch_case("case_a", &req).await.unwrap();
    assert_eq!(case.case_name.as_deref(), Some("renamed"));
}

#[tokio::test]
async fn delete_case_succeeds() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("DELETE", "/cases/case_a")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(r#"{"deleted":true}"#)
        .create_async()
        .await;
    runs(&server).delete_case("case_a").await.unwrap();
}

#[tokio::test]
async fn get_case_404_yields_null_value() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/executions/case_a")
        .with_status(404)
        .create_async()
        .await;
    let v = runs(&server).get_case("case_a").await.unwrap();
    assert_eq!(v, serde_json::Value::Null);
}

#[tokio::test]
async fn bench_by_id_404_is_none() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/benches/5")
        .with_status(404)
        .create_async()
        .await;
    assert!(runs(&server).bench_by_id(5).await.unwrap().is_none());
}

#[tokio::test]
async fn bench_by_id_returns_value() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/benches/5")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(r#"{"id":5,"project_id":7,"script_name":"summarise"}"#)
        .create_async()
        .await;
    let v = runs(&server).bench_by_id(5).await.unwrap().unwrap();
    assert_eq!(v["script_name"], "summarise");
}

#[tokio::test]
async fn mcp_session_cost_returns_value() {
    let mut server = Server::new_async().await;
    let _m = server
        .mock("GET", "/mcp-sessions/sess_abc/cost")
        .with_status(200)
        .with_header("content-type", "application/json")
        .with_body(r#"{"session_id":"sess_abc","total_cost_usd":1.25,"breakdown":[]}"#)
        .create_async()
        .await;
    let v = runs(&server).mcp_session_cost("sess_abc").await.unwrap();
    assert_eq!(v["total_cost_usd"], 1.25);
}