llm_browser_testkit/
scenario.rs

1//! Scenario types for human-readable browser test case definitions.
2//!
3//! Scenarios are written in [TOML](https://toml.io) and describe groups of
4//! browser interaction tests with reusable assertion definitions and
5//! configurable test-level overrides.
6//!
7//! # Structure
8//!
9//! ```toml
10//! [config]                         # Global defaults
11//! start_url = "/dashboard"
12//!
13//! [[definitions]]                  # Reusable assertion definitions
14//! name = "no_errors"
15//! preset = "no_error_on_page"
16//!
17//! [[test]]                         # Test group
18//! name = "Dashboard Smoke"
19//! start_url = "/dashboard"         # Override global start_url (optional)
20//!
21//! [[test.steps]]                   # Ordered steps — the `kind` field
22//! kind = "navigate"                # determines which other fields apply
23//! url = "/dashboard"
24//!
25//! [[test.steps]]
26//! kind = "click"
27//! target = "the Login button"      # Natural language — LLM resolves to selector
28//!
29//! [[test.steps]]
30//! kind = "assert"
31//! definition = "no_errors"
32//! ```
33//!
34//! ## Step Kinds
35//!
36//! | `kind`        | Required fields  | Optional fields                          |
37//! |---------------|-----------------|------------------------------------------|
38//! | `navigate`    | `url`           | `wait_after_ms`                          |
39//! | `click`       | `target`        | `selector`, `wait_after_ms`              |
40//! | `type`        | `target`, `text`| `selector`, `wait_after_ms`              |
41//! | `wait`        | `target`        | `selector`, `timeout_ms`                 |
42//! | `assert`      | *one of below*  | —                                        |
43//! | `screenshot`  | —               | `path`                                   |
44//! | `agent`       | `agent`, `task` | —                                        |
45//! | `mcp`         | `server`, `tool`| `args`                                   |
46//!
47//! Assert steps require one of: `definition` (references a named
48//! `[[definitions]]` entry), `preset` (built-in preset name), or `prompt`
49//! (custom LLM evaluation prompt).
50
51use std::collections::HashMap;
52
53use serde::Deserialize;
54use serde_json::Value;
55
56/// Top-level scenario file, deserialized from TOML.
57#[derive(Debug, Deserialize)]
58pub struct Scenario {
59    /// Global configuration (overridable per test).
60    #[serde(default)]
61    pub config: ScenarioConfig,
62
63    /// Reusable assertion definitions referenced by name in `assert` steps.
64    #[serde(default)]
65    pub definitions: Vec<AssertDefinition>,
66
67    /// Ordered test groups to execute.
68    #[serde(default)]
69    pub test: Vec<TestGroup>,
70}
71
72/// Global scenario configuration with per-test overridable fields.
73#[derive(Debug, Deserialize, Default, Clone)]
74pub struct ScenarioConfig {
75    /// Base URL for relative navigation.
76    #[serde(default)]
77    pub base_url: Option<String>,
78    /// LLM server base URL (deprecated; prefer `[config.endpoints]`).
79    #[serde(default)]
80    pub llm_url: Option<String>,
81    /// LLM model name (deprecated; prefer `[config.endpoints]`).
82    #[serde(default)]
83    pub llm_model: Option<String>,
84    /// LLM API key (Bearer token).
85    #[serde(default)]
86    pub llm_api_key: Option<String>,
87    /// Custom HTTP headers as JSON key-value pairs.
88    #[serde(default, deserialize_with = "deserialize_headers")]
89    pub llm_headers: HashMap<String, String>,
90    /// Run browser in headless mode.
91    #[serde(default)]
92    pub browser_headless: Option<bool>,
93    /// HTTP / browser action timeout in seconds.
94    #[serde(default)]
95    pub timeout_secs: Option<u64>,
96    /// Browser viewport width.
97    #[serde(default)]
98    pub viewport_width: Option<u32>,
99    /// Browser viewport height.
100    #[serde(default)]
101    pub viewport_height: Option<u32>,
102    /// Default URL every test auto-navigates to before running its steps.
103    #[serde(default)]
104    pub start_url: Option<String>,
105    /// Whether to auto-navigate to `start_url` before test steps.
106    ///
107    /// Disable when a test starts with click-based navigation.
108    #[serde(default = "default_auto_navigate")]
109    pub auto_navigate: bool,
110    /// LLM temperature (0.0–1.0). Lower = more deterministic.
111    #[serde(default = "default_temperature")]
112    pub temperature: f64,
113    /// Enable thinking/reasoning tokens. `None` means the provider default
114    /// is used (no `thinking` key is sent). Set to `true`/`false` to
115    /// explicitly enable or disable.
116    #[serde(default)]
117    pub thinking: Option<bool>,
118    /// Provider-specific model parameters merged into the chat completion
119    /// request body (e.g. `effort = "high"` for Anthropic).
120    #[serde(default, deserialize_with = "deserialize_model_params")]
121    pub model_params: HashMap<String, Value>,
122    /// Named endpoints (LLM, MCP, A2A agents) with pricing.
123    #[serde(default)]
124    pub endpoints: HashMap<String, EndpointConfig>,
125    /// Global and per-test budgets for cost/token/call limits.
126    #[serde(default)]
127    pub budgets: BudgetsConfig,
128    /// MCP server exposure configuration.
129    #[serde(default)]
130    pub mcp_server: Option<McpServerConfig>,
131    /// A2A agent server exposure configuration.
132    #[serde(default)]
133    pub a2a_server: Option<A2aServerConfig>,
134    /// Whether to continue running the remaining steps of a test after a
135    /// step fails. Default `false` = fail fast: the first failed step ends
136    /// the test and the rest are reported as skipped. Set to `true` to run
137    /// every step (more diagnostics, more LLM cost on broken apps).
138    #[serde(default)]
139    pub continue_on_failure: bool,
140    /// Longest edge (px) of screenshots attached to `screenshot = true`
141    /// assert steps, before they are JPEG-encoded and sent to the vision
142    /// endpoint. Downscaling keeps vision token cost/quality sane.
143    /// Default: 1400.
144    #[serde(default)]
145    pub screenshot_max_dimension: Option<u32>,
146    /// Height cap (px) of page coverage for screenshots attached to
147    /// `screenshot = true` assert steps. The full scrollable page is
148    /// captured, then split into viewport-tall tiles (each sent as its own
149    /// image part, ordered from the top) covering at most this many pixels
150    /// — below-the-fold content stays visible to the vision model at 1:1
151    /// detail while token cost stays bounded. Accepts an absolute pixel
152    /// count (`2880`) or a viewport multiple (`"20x"` = twenty times the
153    /// currently applied viewport height, which follows viewport-matrix
154    /// and per-test overrides). A value below the viewport height is
155    /// raised to it, so the visible viewport is always fully included
156    /// (`0` / `"0x"` therefore means "viewport only", the pre-full-page
157    /// behavior). Default: `"20x"`.
158    #[serde(default, deserialize_with = "deserialize_screenshot_max_height")]
159    pub screenshot_max_height: Option<ScreenshotHeight>,
160    /// Directory for failure artifacts (screenshots, page snapshots).
161    /// Defaults to `artifacts`.
162    #[serde(default)]
163    pub artifacts_dir: Option<String>,
164    /// Optional viewport matrix: when set, every test in the scenario is
165    /// expanded into one variant per viewport (e.g. mobile/tablet/desktop).
166    /// Each variant overrides the test's viewport and gets a ` — <name>`
167    /// suffix on the test name. Per-test budgets apply per variant.
168    #[serde(default)]
169    pub viewport_matrix: Option<ViewportMatrix>,
170    /// Class-name prefixes the `layout_no_issues` scan skips: elements
171    /// whose class matches any prefix are ignored by the fixed-element,
172    /// text-clipped, and overlap checks. Defaults cover the Angular CDK
173    /// screen-reader helpers (`cdk-visually-hidden`,
174    /// `cdk-describedby-message-container`, `cdk-overlay-container`),
175    /// which are intentionally 1x1 / off-screen.
176    #[serde(default = "default_layout_ignore_classes")]
177    pub layout_ignore_classes: Vec<String>,
178    /// Concurrency group for parallel runs across scenario files.
179    /// When several scenario files are run together (`--parallel > 1`),
180    /// files that declare the **same** `concurrency_group` are never
181    /// executed at the same time — use this for files that touch the same
182    /// shared backend state and would interfere if run concurrently. A file
183    /// with no group gets its own implicit group, so distinct files run in
184    /// parallel by default. Only honored when files are passed to the
185    /// runner as a batch; has no effect on the steps within a single file,
186    /// which always run sequentially.
187    #[serde(default)]
188    pub concurrency_group: Option<String>,
189    /// Default for per-endpoint `cache` across all endpoints (default
190    /// `true`). Set to `false` to disable provider-side prompt-cache
191    /// markers globally.
192    #[serde(default)]
193    pub cache: Option<bool>,
194    /// Default for per-endpoint `pricing.cache_pricing` across all
195    /// endpoints (default `true`). Set to `false` to bill all prompt
196    /// tokens at the flat input price instead of cache rates.
197    #[serde(default)]
198    pub cache_pricing: Option<bool>,
199}
200
201/// A list of named viewports a scenario is expanded across.
202#[derive(Debug, Deserialize, Clone, Default)]
203pub struct ViewportMatrix {
204    /// The viewport variants (`{name, width, height}`).
205    #[serde(default)]
206    pub viewports: Vec<ViewportDef>,
207}
208
209/// Height cap for screenshots attached to `screenshot = true` assert
210/// steps: either an absolute pixel count or a multiple of the currently
211/// applied viewport height.
212#[derive(Debug, Clone, PartialEq)]
213pub enum ScreenshotHeight {
214    /// Absolute height in pixels.
215    Pixels(u32),
216    /// Multiple of the active viewport height (e.g. `2x` = twice the
217    /// current viewport's pixel height).
218    ViewportTimes(f64),
219}
220
221/// Ceiling (px) applied when resolving screenshot height specs, so an
222/// absurdly large multiplier can never overflow or produce an unusable
223/// capture.
224const MAX_SCREENSHOT_HEIGHT: u32 = 4_000_000;
225
226impl ScreenshotHeight {
227    /// Resolves this height spec to a pixel value for the given viewport
228    /// height. Multipliers are rounded and clamped to
229    /// [`MAX_SCREENSHOT_HEIGHT`].
230    #[must_use]
231    pub fn to_px(&self, viewport_height: u32) -> u32 {
232        match self {
233            Self::Pixels(px) => *px,
234            Self::ViewportTimes(mult) => {
235                let scaled = f64::from(viewport_height) * mult;
236                #[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
237                let px = scaled
238                    .round()
239                    .max(1.0)
240                    .min(f64::from(MAX_SCREENSHOT_HEIGHT)) as u32;
241                px
242            }
243        }
244    }
245}
246
247/// One named viewport size in a matrix.
248#[derive(Debug, Deserialize, Clone)]
249pub struct ViewportDef {
250    /// Human-readable variant name (appended to test names, e.g.
251    /// `— mobile`).
252    pub name: String,
253    /// Browser viewport width in pixels.
254    pub width: u32,
255    /// Browser viewport height in pixels.
256    pub height: u32,
257}
258
259/// A named endpoint definition with pricing.
260#[derive(Debug, Deserialize, Clone, Default)]
261pub struct EndpointConfig {
262    /// Endpoint type: `llm`, `mcp`, or `a2a`.
263    #[serde(rename = "type")]
264    pub endpoint_type: EndpointType,
265    /// Base URL for the endpoint.
266    #[serde(default)]
267    pub url: Option<String>,
268    /// Model name (LLM endpoints only).
269    #[serde(default)]
270    pub model: Option<String>,
271    /// API key / bearer token.
272    #[serde(default)]
273    pub api_key: Option<String>,
274    /// Custom HTTP headers as JSON key-value pairs.
275    #[serde(default, deserialize_with = "deserialize_headers")]
276    pub headers: HashMap<String, String>,
277    /// Pricing configuration.
278    #[serde(default)]
279    pub pricing: Option<PricingConfig>,
280    /// Automatically fetch exact per-token pricing for this endpoint at
281    /// startup from a provider's public pricing API, filling in any pricing
282    /// fields left unset (explicit `pricing` values win). Defaults to
283    /// `"auto"`: Bedrock endpoints use the AWS Price List and `openrouter.ai`
284    /// URLs use the `OpenRouter` models API; other providers are a no-op.
285    /// Force a source with `"bedrock"` / `"openrouter"`, or disable the
286    /// lookup with `"off"` / `"none"` / `"disabled"`.
287    #[serde(default)]
288    pub pricing_source: Option<String>,
289    /// Task types this endpoint serves by default
290    /// (e.g. `["targeting", "assertion"]`).
291    #[serde(default)]
292    pub default_for: Vec<String>,
293    /// Command to launch an MCP server subprocess (stdio transport).
294    #[serde(default)]
295    pub command: Option<String>,
296    /// Arguments for the MCP server command.
297    #[serde(default)]
298    pub args: Vec<String>,
299    /// Whether this LLM endpoint accepts image parts (vision) in addition
300    /// to text. `assert` steps with `screenshot = true` require a vision
301    /// endpoint.
302    #[serde(default)]
303    pub vision: bool,
304    /// How often a single chat completion against this endpoint is retried
305    /// on transient failures (HTTP 429/5xx, empty 200 bodies, network
306    /// errors) before the fallback chain is tried. Default: 3 (override
307    /// globally with `HARNESS_LLM_CALL_ATTEMPTS`).
308    #[serde(default)]
309    pub max_attempts: Option<u32>,
310    /// Ordered names of other endpoints to try when this endpoint exhausts
311    /// its attempts. Only LLM endpoints are eligible. Useful for pairing a
312    /// cheap primary model with a more powerful/expensive fallback.
313    #[serde(default)]
314    pub fallbacks: Vec<String>,
315    /// LLM provider protocol: `openai` (default, OpenAI-compatible chat
316    /// completions), `azure` (`Azure` `OpenAI`), or `bedrock` (AWS Bedrock
317    /// Converse API; requires the `aws` cargo feature).
318    #[serde(default)]
319    pub provider: Provider,
320    /// `Azure` `OpenAI` deployment name (`provider = "azure"`). Defaults to
321    /// `model` when unset.
322    #[serde(default)]
323    pub deployment: Option<String>,
324    /// `Azure` `OpenAI` API version (`provider = "azure"`). Defaults to
325    /// `2024-10-21`.
326    #[serde(default)]
327    pub api_version: Option<String>,
328    /// Authentication configuration for LLM endpoints (API key, token
329    /// command, Entra ID client credentials / managed identity).
330    #[serde(default)]
331    pub auth: AuthConfig,
332    /// Extra HTTP headers produced by running a command per call, keyed by
333    /// header name. The command's stdout (first line) becomes the header
334    /// value. Provider-agnostic — applies to every LLM provider.
335    #[serde(default, deserialize_with = "deserialize_headers")]
336    pub header_commands: HashMap<String, String>,
337    /// AWS credential settings (`provider = "bedrock"`).
338    #[serde(default)]
339    pub aws: AwsConfig,
340    /// Send provider-side prompt-cache markers from this endpoint
341    /// (default `true`). Only providers that require explicit markers are
342    /// affected: AWS Bedrock gets a `cachePoint` block, and Anthropic-style
343    /// OpenAI-compatible models (model name contains `claude`/`anthropic`,
344    /// e.g. via `OpenRouter`) get a `cache_control: ephemeral` block on the
345    /// system message. `OpenAI`, `Azure`, Groq, xAI and `DeepSeek` cache
346    /// automatically and need no markers. Set to `false` to disable.
347    #[serde(default)]
348    pub cache: Option<bool>,
349}
350
351/// Type discriminator for endpoint configuration.
352#[derive(Debug, Deserialize, Clone, PartialEq, Eq, Default)]
353#[serde(rename_all = "lowercase")]
354pub enum EndpointType {
355    /// OpenAI-compatible LLM API.
356    #[default]
357    Llm,
358    /// Model Context Protocol server.
359    Mcp,
360    /// Agent-to-Agent protocol agent.
361    A2a,
362}
363
364/// LLM provider protocol used by an LLM endpoint.
365#[derive(Debug, Deserialize, Clone, Copy, PartialEq, Eq, Default)]
366#[serde(rename_all = "lowercase")]
367pub enum Provider {
368    /// OpenAI-compatible Chat Completions API
369    /// (`POST <url>/v1/chat/completions`).
370    #[default]
371    Openai,
372    /// `Azure` `OpenAI`
373    /// (`POST <url>/openai/deployments/<deployment>/chat/completions`).
374    Azure,
375    /// AWS Bedrock Converse API (`POST https://bedrock-runtime.<region>
376    /// .amazonaws.com/model/<model>/converse`). Requires the `aws` cargo
377    /// feature; uses the standard AWS credential chain unless overridden.
378    Bedrock,
379}
380
381/// Authentication mode for an LLM endpoint.
382#[derive(Debug, Deserialize, Clone, Copy, PartialEq, Eq, Default)]
383#[serde(rename_all = "kebab-case")]
384pub enum AuthMode {
385    /// Static API key. Sent as `Authorization: Bearer <key>` by default;
386    /// set [`AuthConfig::api_key_header`] (or use the `azure` provider)
387    /// to send it in a different header.
388    #[default]
389    ApiKey,
390    /// Execute a command and use its stdout (first line) as the bearer
391    /// token. Uses the `token_command` field. Provider-agnostic escape
392    /// hatch — e.g. `az account get-access-token …`.
393    TokenCommand,
394    /// Entra ID (`Azure` AD) OAuth 2.0 client-credentials grant: exchanges
395    /// `client_id` + `client_secret` in `tenant_id` for a bearer token.
396    EntraClientCredentials,
397    /// Entra ID managed identity: fetches a bearer token from the IMDS
398    /// endpoint (works on `Azure` VMs / App Service / ACI with a
399    /// system-assigned identity; zero credentials in the config).
400    EntraManagedIdentity,
401}
402
403/// Authentication settings for an LLM endpoint.
404#[derive(Debug, Deserialize, Clone, Default)]
405pub struct AuthConfig {
406    /// Which authentication mode to use. Default: `api-key`.
407    #[serde(default)]
408    pub mode: AuthMode,
409    /// Header name that receives the API key instead of
410    /// `Authorization: Bearer` (`mode = "api-key"`). For example the `Azure`
411    /// `OpenAI` `api-key` header.
412    #[serde(default)]
413    pub api_key_header: Option<String>,
414    /// Command whose stdout (first line) is used as the bearer token
415    /// (`mode = "token-command"`).
416    #[serde(default)]
417    pub token_command: Option<String>,
418    /// Entra tenant id (`mode = "entra-client-credentials"`,
419    /// `"entra-managed-identity"`).
420    #[serde(default)]
421    pub tenant_id: Option<String>,
422    /// Entra client id (`mode = "entra-client-credentials"`).
423    #[serde(default)]
424    pub client_id: Option<String>,
425    /// Entra client secret (`mode = "entra-client-credentials"`).
426    #[serde(default)]
427    pub client_secret: Option<String>,
428    /// Entra token scope. Defaults to
429    /// `https://cognitiveservices.azure.com/.default`.
430    #[serde(default)]
431    pub scope: Option<String>,
432    /// Override for the Entra token endpoint
433    /// (`mode = "entra-client-credentials"`), default
434    /// `https://login.microsoftonline.com`. Also used as the IMDS identity
435    /// endpoint base for `"entra-managed-identity"` (default
436    /// `http://169.254.169.254`). Mainly useful for tests and proxies.
437    #[serde(default)]
438    pub token_url: Option<String>,
439    /// Seconds to reuse a fetched (non-API-key) token before refetching.
440    /// For Entra modes the server-issued expiry is used when available;
441    /// this caps the reuse window. Default 300.
442    #[serde(default)]
443    pub cache_ttl_secs: Option<u64>,
444}
445
446/// AWS credential and region settings for a `bedrock` provider endpoint.
447///
448/// When no explicit access key is given, the standard AWS credential chain
449/// is used (env vars, shared config `~/.aws/config` + `~/.aws/credentials`,
450/// SSO, ECS/IMDS) — the same behavior as the AWS CLI. `profile` selects a
451/// named profile from the shared config, and `region` overrides the chain
452/// default.
453#[derive(Debug, Deserialize, Clone, Default)]
454pub struct AwsConfig {
455    /// Named profile from `~/.aws/config` / `~/.aws/credentials` to use
456    /// (default: the AWS CLI default selection via `AWS_PROFILE` or
457    /// `default`).
458    #[serde(default)]
459    pub profile: Option<String>,
460    /// AWS region (default: `AWS_REGION` env or the profile's region;
461    /// required if neither is set).
462    #[serde(default)]
463    pub region: Option<String>,
464    /// Explicit access key id (bypasses the credential chain).
465    #[serde(default)]
466    pub access_key_id: Option<String>,
467    /// Explicit secret access key (with `access_key_id`).
468    #[serde(default)]
469    pub secret_access_key: Option<String>,
470    /// Optional session token for explicit temporary credentials.
471    #[serde(default)]
472    pub session_token: Option<String>,
473}
474
475/// Pricing configuration for an endpoint.
476#[derive(Debug, Deserialize, Clone, Default)]
477pub struct PricingConfig {
478    /// Cost per 1M input tokens (USD).
479    #[serde(default)]
480    pub input_per_1m_tokens: f64,
481    /// Cost per 1M output tokens (USD).
482    #[serde(default)]
483    pub output_per_1m_tokens: f64,
484    /// Flat cost per call (USD), used for MCP/agent endpoints.
485    #[serde(default)]
486    pub per_call: f64,
487    /// Cost per 1M cached-input (prompt cache read) tokens (USD). When
488    /// unset, `input_per_1m_tokens * cache_read_multiplier` is used.
489    #[serde(default)]
490    pub cached_input_per_1m_tokens: Option<f64>,
491    /// Cost per 1M cache-write (cache creation) input tokens (USD). When
492    /// unset, `input_per_1m_tokens * cache_write_multiplier` is used.
493    #[serde(default)]
494    pub cache_write_per_1m_tokens: Option<f64>,
495    /// Multiplier applied to `input_per_1m_tokens` for cache reads when
496    /// `cached_input_per_1m_tokens` is unset. Default: 0.1.
497    #[serde(default)]
498    pub cache_read_multiplier: Option<f64>,
499    /// Multiplier applied to `input_per_1m_tokens` for cache writes when
500    /// `cache_write_per_1m_tokens` is unset. Default: 1.25.
501    #[serde(default)]
502    pub cache_write_multiplier: Option<f64>,
503    /// Bill cache reads/writes at their cache rates (default `true`). Set
504    /// to `false` to bill every prompt token at the flat input price.
505    #[serde(default)]
506    pub cache_pricing: Option<bool>,
507}
508
509/// Budget limits for test execution.
510#[derive(Debug, Deserialize, Clone, Default)]
511pub struct BudgetsConfig {
512    /// Global budget across all tests in the scenario.
513    #[serde(default)]
514    pub global: Option<BudgetDef>,
515    /// Default per-test budget. Individual tests can override.
516    #[serde(default)]
517    pub per_test_default: Option<BudgetDef>,
518}
519
520/// A budget definition with limits and enforcement mode.
521#[derive(Debug, Deserialize, Clone)]
522pub struct BudgetDef {
523    /// Maximum cost in USD.
524    #[serde(default)]
525    pub max_cost: Option<f64>,
526    /// Maximum total tokens (input + output).
527    #[serde(default)]
528    pub max_tokens: Option<u64>,
529    /// Maximum number of calls (LLM, MCP, agent combined).
530    #[serde(default)]
531    pub max_calls: Option<u64>,
532    /// Enforcement mode: `hard` (abort) or `soft` (warn and continue).
533    #[serde(default)]
534    pub enforcement: Option<BudgetEnforcement>,
535}
536
537/// Budget enforcement strategy.
538#[derive(Debug, Deserialize, Clone, PartialEq, Eq)]
539#[serde(rename_all = "lowercase")]
540pub enum BudgetEnforcement {
541    /// Abort the test or run when budget is exceeded.
542    Hard,
543    /// Log a warning but continue execution.
544    Soft,
545}
546
547/// MCP server exposure configuration.
548#[derive(Debug, Deserialize, Clone)]
549pub struct McpServerConfig {
550    /// Whether to enable the embedded MCP server.
551    #[serde(default)]
552    pub enabled: bool,
553    /// Port to listen on.
554    #[serde(default = "default_mcp_port")]
555    pub port: u16,
556}
557
558const fn default_mcp_port() -> u16 {
559    3000
560}
561
562/// A2A agent server exposure configuration.
563#[derive(Debug, Deserialize, Clone)]
564pub struct A2aServerConfig {
565    /// Whether to enable the embedded A2A agent server.
566    #[serde(default)]
567    pub enabled: bool,
568    /// Port to listen on.
569    #[serde(default = "default_a2a_port")]
570    pub port: u16,
571}
572
573const fn default_a2a_port() -> u16 {
574    3100
575}
576
577fn deserialize_headers<'de, D>(deserializer: D) -> Result<HashMap<String, String>, D::Error>
578where
579    D: serde::Deserializer<'de>,
580{
581    let raw: Option<serde_json::Value> = Option::deserialize(deserializer)?;
582    let Some(json) = raw else {
583        return Ok(HashMap::new());
584    };
585    let serde_json::Value::Object(obj) = json else {
586        return Ok(HashMap::new());
587    };
588    Ok(obj
589        .into_iter()
590        .filter_map(|(k, v)| v.as_str().map(|s| (k, s.to_owned())))
591        .collect())
592}
593
594const fn default_auto_navigate() -> bool {
595    true
596}
597
598fn default_layout_ignore_classes() -> Vec<String> {
599    vec![
600        "cdk-visually-hidden".to_owned(),
601        "cdk-describedby-message-container".to_owned(),
602        "cdk-overlay-container".to_owned(),
603    ]
604}
605
606const fn default_temperature() -> f64 {
607    0.0
608}
609
610fn deserialize_model_params<'de, D>(deserializer: D) -> Result<HashMap<String, Value>, D::Error>
611where
612    D: serde::Deserializer<'de>,
613{
614    #[derive(Deserialize)]
615    #[serde(untagged)]
616    enum Raw {
617        Map(HashMap<String, Value>),
618        Table(HashMap<String, Value>),
619    }
620    let raw: Option<Raw> = Option::deserialize(deserializer)?;
621    Ok(match raw {
622        Some(Raw::Map(m) | Raw::Table(m)) => m,
623        None => HashMap::new(),
624    })
625}
626
627fn deserialize_screenshot_max_height<'de, D>(
628    deserializer: D,
629) -> Result<Option<ScreenshotHeight>, D::Error>
630where
631    D: serde::Deserializer<'de>,
632{
633    let raw: Option<serde_json::Value> = Option::deserialize(deserializer)?;
634    match raw {
635        None => Ok(None),
636        Some(value) => match value.as_u64() {
637            // Integer: absolute pixel count.
638            Some(px) if px <= u64::from(MAX_SCREENSHOT_HEIGHT) => {
639                #[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
640                Ok(Some(ScreenshotHeight::Pixels(px as u32)))
641            }
642            // String: viewport multiple like "2x" or "1.5x".
643            None => match value.as_str() {
644                Some(s) => {
645                    let lower = s.trim().to_lowercase();
646                    if lower.ends_with('x') {
647                        let num = lower.strip_suffix("x");
648                        if let Some(num) = num {
649                            if let Ok(m) = num.parse::<f64>() {
650                                if m > 0.0 {
651                                    return Ok(Some(ScreenshotHeight::ViewportTimes(m)));
652                                }
653                            }
654                        }
655                    }
656                    Err(serde::de::Error::custom(format_args!(
657                        "screenshot_max_height must be a pixel count (e.g. 2880) or a viewport multiple like \"2x\", found {s:?}"
658                    )))
659                }
660                None => Err(serde::de::Error::custom(format_args!(
661                    "screenshot_max_height must be a pixel count (e.g. 2880) or a viewport multiple like \"2x\", found {value:?}"
662                ))),
663            },
664            Some(_) => Err(serde::de::Error::custom(format_args!(
665                "screenshot_max_height value too large (max {MAX_SCREENSHOT_HEIGHT} px)"
666            ))),
667        },
668    }
669}
670
671/// Reusable assertion definition referenced by name from `assert` steps.
672///
673/// Definitions can either reference a built-in preset via `preset`, supply a
674/// custom LLM `prompt`, or define a **custom preset** by providing both
675/// `system` and `user_template`. Custom presets support the same template
676/// variables as built-in presets: `{url}`, `{title}`, `{content}`,
677/// `{expected_text}`, `{description}`.
678#[derive(Debug, Deserialize, Clone)]
679pub struct AssertDefinition {
680    /// Unique name used to reference this definition from steps.
681    pub name: String,
682    /// Predefined assertion preset name
683    /// (e.g. `no_error_on_page`, `text_visible`).
684    #[serde(default)]
685    pub preset: Option<String>,
686    /// Custom LLM prompt for assertion evaluation.
687    #[serde(default)]
688    pub prompt: Option<String>,
689    /// System prompt for a custom preset.
690    #[serde(default)]
691    pub system: Option<String>,
692    /// User template (with `{placeholders}`) for a custom preset.
693    #[serde(default)]
694    pub user_template: Option<String>,
695    /// Text that the `text_visible` preset checks for, or the
696    /// `{expected_text}` placeholder value for custom presets.
697    #[serde(default)]
698    pub assert_text: Option<String>,
699    /// Agent endpoint to call for this assertion.
700    #[serde(default)]
701    pub agent: Option<String>,
702    /// Agent task template for this assertion.
703    #[serde(default)]
704    pub task_template: Option<String>,
705}
706
707/// A group of steps that form a single test scenario.
708#[derive(Debug, Deserialize, Clone)]
709pub struct TestGroup {
710    /// Human-readable test name.
711    pub name: String,
712    /// Override the global `start_url` for this test.
713    #[serde(default)]
714    pub start_url: Option<String>,
715    /// Override the global `auto_navigate` for this test.
716    #[serde(default)]
717    pub auto_navigate: Option<bool>,
718    /// Override the global `base_url` for this test.
719    #[serde(default)]
720    pub base_url: Option<String>,
721    /// Override the global `timeout_secs` for this test.
722    #[serde(default)]
723    pub timeout_secs: Option<u64>,
724    /// Override the global `browser_headless` for this test.
725    #[serde(default)]
726    pub browser_headless: Option<bool>,
727    /// Override the global viewport width for this test (applied via CDP
728    /// `Emulation.setDeviceMetricsOverride` before the test runs).
729    #[serde(default)]
730    pub viewport_width: Option<u32>,
731    /// Override the global viewport height for this test.
732    #[serde(default)]
733    pub viewport_height: Option<u32>,
734    /// Per-test budget override.
735    #[serde(default)]
736    pub budget: Option<BudgetDef>,
737    /// Endpoint to use for all steps in this test (can be overridden
738    /// per-step).
739    #[serde(default)]
740    pub endpoint: Option<String>,
741    /// Ordered steps to execute.
742    #[serde(default)]
743    pub steps: Vec<TestStep>,
744}
745
746/// A single step in a test. The `kind` field determines which variant is
747/// deserialized and which field constraints apply.
748#[derive(Debug, Deserialize, Clone)]
749#[serde(tag = "kind")]
750pub enum TestStep {
751    /// Navigate the browser to a URL.
752    #[serde(rename = "navigate")]
753    Navigate {
754        /// URL to navigate to (absolute, or relative to the test's
755        /// base URL).
756        url: String,
757        /// Milliseconds to wait after navigation completes.
758        #[serde(default)]
759        wait_after_ms: Option<u64>,
760    },
761
762    /// Click an element described in natural language.
763    #[serde(rename = "click")]
764    Click {
765        /// Natural language description of the element. The LLM resolves
766        /// this to a CSS selector at runtime.
767        target: String,
768        /// Explicit CSS selector override (bypasses LLM resolution).
769        #[serde(default)]
770        selector: Option<String>,
771        /// Milliseconds to wait after the click.
772        #[serde(default)]
773        wait_after_ms: Option<u64>,
774        /// Endpoint to use for LLM element targeting.
775        #[serde(default)]
776        endpoint: Option<String>,
777        /// Idempotent: when the target element is absent the step is
778        /// reported skipped instead of failed (the action was already
779        /// done / not applicable).
780        #[serde(default)]
781        idempotent: bool,
782    },
783
784    /// Type text into an input element.
785    #[serde(rename = "type")]
786    Type {
787        /// Natural language description of the target input element.
788        target: String,
789        /// Text to type into the element.
790        text: String,
791        /// Explicit CSS selector override (bypasses LLM resolution).
792        #[serde(default)]
793        selector: Option<String>,
794        /// Milliseconds to wait after typing.
795        #[serde(default)]
796        wait_after_ms: Option<u64>,
797        /// Endpoint to use for LLM element targeting.
798        #[serde(default)]
799        endpoint: Option<String>,
800        /// Idempotent: when the target element is absent the step is
801        /// reported skipped instead of failed (the action was already
802        /// done / not applicable).
803        #[serde(default)]
804        idempotent: bool,
805    },
806
807    /// Wait for an element to appear on the page.
808    #[serde(rename = "wait")]
809    Wait {
810        /// Natural language description of the element to wait for.
811        target: String,
812        /// Explicit CSS selector override (bypasses LLM resolution).
813        #[serde(default)]
814        selector: Option<String>,
815        /// Wait until the page's visible text contains this substring
816        /// (alternative to `selector`; either or both may be set — both are
817        /// required to hold when both are set).
818        #[serde(default)]
819        text: Option<String>,
820        /// Maximum milliseconds to wait (default: 10000).
821        #[serde(default)]
822        timeout_ms: Option<u64>,
823        /// Endpoint to use for LLM element targeting.
824        #[serde(default)]
825        endpoint: Option<String>,
826        /// Idempotent: when the condition never becomes true within the
827        /// timeout the step is reported skipped instead of failed (the
828        /// condition was not applicable, e.g. already-authenticated
829        /// pages in a viewport matrix).
830        #[serde(default)]
831        idempotent: bool,
832    },
833
834    /// Evaluate an assertion against the current page content.
835    #[serde(rename = "assert")]
836    Assert {
837        /// Reference to a named `[[definitions]]` entry.
838        #[serde(default)]
839        definition: Option<String>,
840        /// Inline predefined assertion preset (e.g. `no_error_on_page`).
841        #[serde(default)]
842        preset: Option<String>,
843        /// Inline custom LLM prompt for assertion evaluation.
844        #[serde(default)]
845        prompt: Option<String>,
846        /// Text that the `text_visible` preset checks for.
847        #[serde(default)]
848        assert_text: Option<String>,
849        /// Endpoint to use for this assertion's LLM call.
850        #[serde(default)]
851        endpoint: Option<String>,
852        /// Attach a screenshot of the current viewport to the assertion so
853        /// the LLM can evaluate visuals (overlaps, clipping, layout).
854        /// Requires the resolved endpoint to declare `vision = true`.
855        #[serde(default)]
856        screenshot: bool,
857    },
858
859    /// Take a screenshot of the current page.
860    #[serde(rename = "screenshot")]
861    Screenshot {
862        /// File path to save the screenshot (default: `screenshot.png`).
863        #[serde(default)]
864        path: Option<String>,
865    },
866
867    /// Call an A2A agent with a task.
868    #[serde(rename = "agent")]
869    Agent {
870        /// Name of the agent endpoint to call.
871        agent: String,
872        /// Task description / prompt for the agent.
873        task: String,
874        /// Optional definition name with a task template.
875        #[serde(default)]
876        definition: Option<String>,
877    },
878
879    /// Call an MCP server tool.
880    #[serde(rename = "mcp")]
881    Mcp {
882        /// Name of the MCP server endpoint.
883        server: String,
884        /// Tool name to invoke on the server.
885        tool: String,
886        /// Tool arguments as JSON.
887        #[serde(default)]
888        args: Option<serde_json::Value>,
889    },
890}
llm_browser_testkit/scenario.rs

llm_browser_testkit/
scenario.rs
