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 from a
281    /// provider's public pricing API, filling in any pricing fields left
282    /// unset. `"openrouter"` always uses the `OpenRouter` models API;
283    /// `"auto"` does so only when the endpoint URL host is `openrouter.ai`.
284    /// Explicit `pricing` values win over fetched ones.
285    #[serde(default)]
286    pub pricing_source: Option<String>,
287    /// Task types this endpoint serves by default
288    /// (e.g. `["targeting", "assertion"]`).
289    #[serde(default)]
290    pub default_for: Vec<String>,
291    /// Command to launch an MCP server subprocess (stdio transport).
292    #[serde(default)]
293    pub command: Option<String>,
294    /// Arguments for the MCP server command.
295    #[serde(default)]
296    pub args: Vec<String>,
297    /// Whether this LLM endpoint accepts image parts (vision) in addition
298    /// to text. `assert` steps with `screenshot = true` require a vision
299    /// endpoint.
300    #[serde(default)]
301    pub vision: bool,
302    /// How often a single chat completion against this endpoint is retried
303    /// on transient failures (HTTP 429/5xx, empty 200 bodies, network
304    /// errors) before the fallback chain is tried. Default: 3 (override
305    /// globally with `HARNESS_LLM_CALL_ATTEMPTS`).
306    #[serde(default)]
307    pub max_attempts: Option<u32>,
308    /// Ordered names of other endpoints to try when this endpoint exhausts
309    /// its attempts. Only LLM endpoints are eligible. Useful for pairing a
310    /// cheap primary model with a more powerful/expensive fallback.
311    #[serde(default)]
312    pub fallbacks: Vec<String>,
313    /// LLM provider protocol: `openai` (default, OpenAI-compatible chat
314    /// completions), `azure` (`Azure` `OpenAI`), or `bedrock` (AWS Bedrock
315    /// Converse API; requires the `aws` cargo feature).
316    #[serde(default)]
317    pub provider: Provider,
318    /// `Azure` `OpenAI` deployment name (`provider = "azure"`). Defaults to
319    /// `model` when unset.
320    #[serde(default)]
321    pub deployment: Option<String>,
322    /// `Azure` `OpenAI` API version (`provider = "azure"`). Defaults to
323    /// `2024-10-21`.
324    #[serde(default)]
325    pub api_version: Option<String>,
326    /// Authentication configuration for LLM endpoints (API key, token
327    /// command, Entra ID client credentials / managed identity).
328    #[serde(default)]
329    pub auth: AuthConfig,
330    /// Extra HTTP headers produced by running a command per call, keyed by
331    /// header name. The command's stdout (first line) becomes the header
332    /// value. Provider-agnostic — applies to every LLM provider.
333    #[serde(default, deserialize_with = "deserialize_headers")]
334    pub header_commands: HashMap<String, String>,
335    /// AWS credential settings (`provider = "bedrock"`).
336    #[serde(default)]
337    pub aws: AwsConfig,
338    /// Send provider-side prompt-cache markers from this endpoint
339    /// (default `true`). Only providers that require explicit markers are
340    /// affected: AWS Bedrock gets a `cachePoint` block, and Anthropic-style
341    /// OpenAI-compatible models (model name contains `claude`/`anthropic`,
342    /// e.g. via `OpenRouter`) get a `cache_control: ephemeral` block on the
343    /// system message. `OpenAI`, `Azure`, Groq, xAI and `DeepSeek` cache
344    /// automatically and need no markers. Set to `false` to disable.
345    #[serde(default)]
346    pub cache: Option<bool>,
347}
348
349/// Type discriminator for endpoint configuration.
350#[derive(Debug, Deserialize, Clone, PartialEq, Eq, Default)]
351#[serde(rename_all = "lowercase")]
352pub enum EndpointType {
353    /// OpenAI-compatible LLM API.
354    #[default]
355    Llm,
356    /// Model Context Protocol server.
357    Mcp,
358    /// Agent-to-Agent protocol agent.
359    A2a,
360}
361
362/// LLM provider protocol used by an LLM endpoint.
363#[derive(Debug, Deserialize, Clone, Copy, PartialEq, Eq, Default)]
364#[serde(rename_all = "lowercase")]
365pub enum Provider {
366    /// OpenAI-compatible Chat Completions API
367    /// (`POST <url>/v1/chat/completions`).
368    #[default]
369    Openai,
370    /// `Azure` `OpenAI`
371    /// (`POST <url>/openai/deployments/<deployment>/chat/completions`).
372    Azure,
373    /// AWS Bedrock Converse API (`POST https://bedrock-runtime.<region>
374    /// .amazonaws.com/model/<model>/converse`). Requires the `aws` cargo
375    /// feature; uses the standard AWS credential chain unless overridden.
376    Bedrock,
377}
378
379/// Authentication mode for an LLM endpoint.
380#[derive(Debug, Deserialize, Clone, Copy, PartialEq, Eq, Default)]
381#[serde(rename_all = "kebab-case")]
382pub enum AuthMode {
383    /// Static API key. Sent as `Authorization: Bearer <key>` by default;
384    /// set [`AuthConfig::api_key_header`] (or use the `azure` provider)
385    /// to send it in a different header.
386    #[default]
387    ApiKey,
388    /// Execute a command and use its stdout (first line) as the bearer
389    /// token. Uses the `token_command` field. Provider-agnostic escape
390    /// hatch — e.g. `az account get-access-token …`.
391    TokenCommand,
392    /// Entra ID (`Azure` AD) OAuth 2.0 client-credentials grant: exchanges
393    /// `client_id` + `client_secret` in `tenant_id` for a bearer token.
394    EntraClientCredentials,
395    /// Entra ID managed identity: fetches a bearer token from the IMDS
396    /// endpoint (works on `Azure` VMs / App Service / ACI with a
397    /// system-assigned identity; zero credentials in the config).
398    EntraManagedIdentity,
399}
400
401/// Authentication settings for an LLM endpoint.
402#[derive(Debug, Deserialize, Clone, Default)]
403pub struct AuthConfig {
404    /// Which authentication mode to use. Default: `api-key`.
405    #[serde(default)]
406    pub mode: AuthMode,
407    /// Header name that receives the API key instead of
408    /// `Authorization: Bearer` (`mode = "api-key"`). For example the `Azure`
409    /// `OpenAI` `api-key` header.
410    #[serde(default)]
411    pub api_key_header: Option<String>,
412    /// Command whose stdout (first line) is used as the bearer token
413    /// (`mode = "token-command"`).
414    #[serde(default)]
415    pub token_command: Option<String>,
416    /// Entra tenant id (`mode = "entra-client-credentials"`,
417    /// `"entra-managed-identity"`).
418    #[serde(default)]
419    pub tenant_id: Option<String>,
420    /// Entra client id (`mode = "entra-client-credentials"`).
421    #[serde(default)]
422    pub client_id: Option<String>,
423    /// Entra client secret (`mode = "entra-client-credentials"`).
424    #[serde(default)]
425    pub client_secret: Option<String>,
426    /// Entra token scope. Defaults to
427    /// `https://cognitiveservices.azure.com/.default`.
428    #[serde(default)]
429    pub scope: Option<String>,
430    /// Override for the Entra token endpoint
431    /// (`mode = "entra-client-credentials"`), default
432    /// `https://login.microsoftonline.com`. Also used as the IMDS identity
433    /// endpoint base for `"entra-managed-identity"` (default
434    /// `http://169.254.169.254`). Mainly useful for tests and proxies.
435    #[serde(default)]
436    pub token_url: Option<String>,
437    /// Seconds to reuse a fetched (non-API-key) token before refetching.
438    /// For Entra modes the server-issued expiry is used when available;
439    /// this caps the reuse window. Default 300.
440    #[serde(default)]
441    pub cache_ttl_secs: Option<u64>,
442}
443
444/// AWS credential and region settings for a `bedrock` provider endpoint.
445///
446/// When no explicit access key is given, the standard AWS credential chain
447/// is used (env vars, shared config `~/.aws/config` + `~/.aws/credentials`,
448/// SSO, ECS/IMDS) — the same behavior as the AWS CLI. `profile` selects a
449/// named profile from the shared config, and `region` overrides the chain
450/// default.
451#[derive(Debug, Deserialize, Clone, Default)]
452pub struct AwsConfig {
453    /// Named profile from `~/.aws/config` / `~/.aws/credentials` to use
454    /// (default: the AWS CLI default selection via `AWS_PROFILE` or
455    /// `default`).
456    #[serde(default)]
457    pub profile: Option<String>,
458    /// AWS region (default: `AWS_REGION` env or the profile's region;
459    /// required if neither is set).
460    #[serde(default)]
461    pub region: Option<String>,
462    /// Explicit access key id (bypasses the credential chain).
463    #[serde(default)]
464    pub access_key_id: Option<String>,
465    /// Explicit secret access key (with `access_key_id`).
466    #[serde(default)]
467    pub secret_access_key: Option<String>,
468    /// Optional session token for explicit temporary credentials.
469    #[serde(default)]
470    pub session_token: Option<String>,
471}
472
473/// Pricing configuration for an endpoint.
474#[derive(Debug, Deserialize, Clone, Default)]
475pub struct PricingConfig {
476    /// Cost per 1M input tokens (USD).
477    #[serde(default)]
478    pub input_per_1m_tokens: f64,
479    /// Cost per 1M output tokens (USD).
480    #[serde(default)]
481    pub output_per_1m_tokens: f64,
482    /// Flat cost per call (USD), used for MCP/agent endpoints.
483    #[serde(default)]
484    pub per_call: f64,
485    /// Cost per 1M cached-input (prompt cache read) tokens (USD). When
486    /// unset, `input_per_1m_tokens * cache_read_multiplier` is used.
487    #[serde(default)]
488    pub cached_input_per_1m_tokens: Option<f64>,
489    /// Cost per 1M cache-write (cache creation) input tokens (USD). When
490    /// unset, `input_per_1m_tokens * cache_write_multiplier` is used.
491    #[serde(default)]
492    pub cache_write_per_1m_tokens: Option<f64>,
493    /// Multiplier applied to `input_per_1m_tokens` for cache reads when
494    /// `cached_input_per_1m_tokens` is unset. Default: 0.1.
495    #[serde(default)]
496    pub cache_read_multiplier: Option<f64>,
497    /// Multiplier applied to `input_per_1m_tokens` for cache writes when
498    /// `cache_write_per_1m_tokens` is unset. Default: 1.25.
499    #[serde(default)]
500    pub cache_write_multiplier: Option<f64>,
501    /// Bill cache reads/writes at their cache rates (default `true`). Set
502    /// to `false` to bill every prompt token at the flat input price.
503    #[serde(default)]
504    pub cache_pricing: Option<bool>,
505}
506
507/// Budget limits for test execution.
508#[derive(Debug, Deserialize, Clone, Default)]
509pub struct BudgetsConfig {
510    /// Global budget across all tests in the scenario.
511    #[serde(default)]
512    pub global: Option<BudgetDef>,
513    /// Default per-test budget. Individual tests can override.
514    #[serde(default)]
515    pub per_test_default: Option<BudgetDef>,
516}
517
518/// A budget definition with limits and enforcement mode.
519#[derive(Debug, Deserialize, Clone)]
520pub struct BudgetDef {
521    /// Maximum cost in USD.
522    #[serde(default)]
523    pub max_cost: Option<f64>,
524    /// Maximum total tokens (input + output).
525    #[serde(default)]
526    pub max_tokens: Option<u64>,
527    /// Maximum number of calls (LLM, MCP, agent combined).
528    #[serde(default)]
529    pub max_calls: Option<u64>,
530    /// Enforcement mode: `hard` (abort) or `soft` (warn and continue).
531    #[serde(default)]
532    pub enforcement: Option<BudgetEnforcement>,
533}
534
535/// Budget enforcement strategy.
536#[derive(Debug, Deserialize, Clone, PartialEq, Eq)]
537#[serde(rename_all = "lowercase")]
538pub enum BudgetEnforcement {
539    /// Abort the test or run when budget is exceeded.
540    Hard,
541    /// Log a warning but continue execution.
542    Soft,
543}
544
545/// MCP server exposure configuration.
546#[derive(Debug, Deserialize, Clone)]
547pub struct McpServerConfig {
548    /// Whether to enable the embedded MCP server.
549    #[serde(default)]
550    pub enabled: bool,
551    /// Port to listen on.
552    #[serde(default = "default_mcp_port")]
553    pub port: u16,
554}
555
556const fn default_mcp_port() -> u16 {
557    3000
558}
559
560/// A2A agent server exposure configuration.
561#[derive(Debug, Deserialize, Clone)]
562pub struct A2aServerConfig {
563    /// Whether to enable the embedded A2A agent server.
564    #[serde(default)]
565    pub enabled: bool,
566    /// Port to listen on.
567    #[serde(default = "default_a2a_port")]
568    pub port: u16,
569}
570
571const fn default_a2a_port() -> u16 {
572    3100
573}
574
575fn deserialize_headers<'de, D>(deserializer: D) -> Result<HashMap<String, String>, D::Error>
576where
577    D: serde::Deserializer<'de>,
578{
579    let raw: Option<serde_json::Value> = Option::deserialize(deserializer)?;
580    let Some(json) = raw else {
581        return Ok(HashMap::new());
582    };
583    let serde_json::Value::Object(obj) = json else {
584        return Ok(HashMap::new());
585    };
586    Ok(obj
587        .into_iter()
588        .filter_map(|(k, v)| v.as_str().map(|s| (k, s.to_owned())))
589        .collect())
590}
591
592const fn default_auto_navigate() -> bool {
593    true
594}
595
596fn default_layout_ignore_classes() -> Vec<String> {
597    vec![
598        "cdk-visually-hidden".to_owned(),
599        "cdk-describedby-message-container".to_owned(),
600        "cdk-overlay-container".to_owned(),
601    ]
602}
603
604const fn default_temperature() -> f64 {
605    0.0
606}
607
608fn deserialize_model_params<'de, D>(deserializer: D) -> Result<HashMap<String, Value>, D::Error>
609where
610    D: serde::Deserializer<'de>,
611{
612    #[derive(Deserialize)]
613    #[serde(untagged)]
614    enum Raw {
615        Map(HashMap<String, Value>),
616        Table(HashMap<String, Value>),
617    }
618    let raw: Option<Raw> = Option::deserialize(deserializer)?;
619    Ok(match raw {
620        Some(Raw::Map(m) | Raw::Table(m)) => m,
621        None => HashMap::new(),
622    })
623}
624
625fn deserialize_screenshot_max_height<'de, D>(
626    deserializer: D,
627) -> Result<Option<ScreenshotHeight>, D::Error>
628where
629    D: serde::Deserializer<'de>,
630{
631    let raw: Option<serde_json::Value> = Option::deserialize(deserializer)?;
632    match raw {
633        None => Ok(None),
634        Some(value) => match value.as_u64() {
635            // Integer: absolute pixel count.
636            Some(px) if px <= u64::from(MAX_SCREENSHOT_HEIGHT) => {
637                #[allow(clippy::cast_possible_truncation, clippy::cast_sign_loss)]
638                Ok(Some(ScreenshotHeight::Pixels(px as u32)))
639            }
640            // String: viewport multiple like "2x" or "1.5x".
641            None => match value.as_str() {
642                Some(s) => {
643                    let lower = s.trim().to_lowercase();
644                    if lower.ends_with('x') {
645                        let num = lower.strip_suffix("x");
646                        if let Some(num) = num {
647                            if let Ok(m) = num.parse::<f64>() {
648                                if m > 0.0 {
649                                    return Ok(Some(ScreenshotHeight::ViewportTimes(m)));
650                                }
651                            }
652                        }
653                    }
654                    Err(serde::de::Error::custom(format_args!(
655                        "screenshot_max_height must be a pixel count (e.g. 2880) or a viewport multiple like \"2x\", found {s:?}"
656                    )))
657                }
658                None => Err(serde::de::Error::custom(format_args!(
659                    "screenshot_max_height must be a pixel count (e.g. 2880) or a viewport multiple like \"2x\", found {value:?}"
660                ))),
661            },
662            Some(_) => Err(serde::de::Error::custom(format_args!(
663                "screenshot_max_height value too large (max {MAX_SCREENSHOT_HEIGHT} px)"
664            ))),
665        },
666    }
667}
668
669/// Reusable assertion definition referenced by name from `assert` steps.
670///
671/// Definitions can either reference a built-in preset via `preset`, supply a
672/// custom LLM `prompt`, or define a **custom preset** by providing both
673/// `system` and `user_template`. Custom presets support the same template
674/// variables as built-in presets: `{url}`, `{title}`, `{content}`,
675/// `{expected_text}`, `{description}`.
676#[derive(Debug, Deserialize, Clone)]
677pub struct AssertDefinition {
678    /// Unique name used to reference this definition from steps.
679    pub name: String,
680    /// Predefined assertion preset name
681    /// (e.g. `no_error_on_page`, `text_visible`).
682    #[serde(default)]
683    pub preset: Option<String>,
684    /// Custom LLM prompt for assertion evaluation.
685    #[serde(default)]
686    pub prompt: Option<String>,
687    /// System prompt for a custom preset.
688    #[serde(default)]
689    pub system: Option<String>,
690    /// User template (with `{placeholders}`) for a custom preset.
691    #[serde(default)]
692    pub user_template: Option<String>,
693    /// Text that the `text_visible` preset checks for, or the
694    /// `{expected_text}` placeholder value for custom presets.
695    #[serde(default)]
696    pub assert_text: Option<String>,
697    /// Agent endpoint to call for this assertion.
698    #[serde(default)]
699    pub agent: Option<String>,
700    /// Agent task template for this assertion.
701    #[serde(default)]
702    pub task_template: Option<String>,
703}
704
705/// A group of steps that form a single test scenario.
706#[derive(Debug, Deserialize, Clone)]
707pub struct TestGroup {
708    /// Human-readable test name.
709    pub name: String,
710    /// Override the global `start_url` for this test.
711    #[serde(default)]
712    pub start_url: Option<String>,
713    /// Override the global `auto_navigate` for this test.
714    #[serde(default)]
715    pub auto_navigate: Option<bool>,
716    /// Override the global `base_url` for this test.
717    #[serde(default)]
718    pub base_url: Option<String>,
719    /// Override the global `timeout_secs` for this test.
720    #[serde(default)]
721    pub timeout_secs: Option<u64>,
722    /// Override the global `browser_headless` for this test.
723    #[serde(default)]
724    pub browser_headless: Option<bool>,
725    /// Override the global viewport width for this test (applied via CDP
726    /// `Emulation.setDeviceMetricsOverride` before the test runs).
727    #[serde(default)]
728    pub viewport_width: Option<u32>,
729    /// Override the global viewport height for this test.
730    #[serde(default)]
731    pub viewport_height: Option<u32>,
732    /// Per-test budget override.
733    #[serde(default)]
734    pub budget: Option<BudgetDef>,
735    /// Endpoint to use for all steps in this test (can be overridden
736    /// per-step).
737    #[serde(default)]
738    pub endpoint: Option<String>,
739    /// Ordered steps to execute.
740    #[serde(default)]
741    pub steps: Vec<TestStep>,
742}
743
744/// A single step in a test. The `kind` field determines which variant is
745/// deserialized and which field constraints apply.
746#[derive(Debug, Deserialize, Clone)]
747#[serde(tag = "kind")]
748pub enum TestStep {
749    /// Navigate the browser to a URL.
750    #[serde(rename = "navigate")]
751    Navigate {
752        /// URL to navigate to (absolute, or relative to the test's
753        /// base URL).
754        url: String,
755        /// Milliseconds to wait after navigation completes.
756        #[serde(default)]
757        wait_after_ms: Option<u64>,
758    },
759
760    /// Click an element described in natural language.
761    #[serde(rename = "click")]
762    Click {
763        /// Natural language description of the element. The LLM resolves
764        /// this to a CSS selector at runtime.
765        target: String,
766        /// Explicit CSS selector override (bypasses LLM resolution).
767        #[serde(default)]
768        selector: Option<String>,
769        /// Milliseconds to wait after the click.
770        #[serde(default)]
771        wait_after_ms: Option<u64>,
772        /// Endpoint to use for LLM element targeting.
773        #[serde(default)]
774        endpoint: Option<String>,
775        /// Idempotent: when the target element is absent the step is
776        /// reported skipped instead of failed (the action was already
777        /// done / not applicable).
778        #[serde(default)]
779        idempotent: bool,
780    },
781
782    /// Type text into an input element.
783    #[serde(rename = "type")]
784    Type {
785        /// Natural language description of the target input element.
786        target: String,
787        /// Text to type into the element.
788        text: String,
789        /// Explicit CSS selector override (bypasses LLM resolution).
790        #[serde(default)]
791        selector: Option<String>,
792        /// Milliseconds to wait after typing.
793        #[serde(default)]
794        wait_after_ms: Option<u64>,
795        /// Endpoint to use for LLM element targeting.
796        #[serde(default)]
797        endpoint: Option<String>,
798        /// Idempotent: when the target element is absent the step is
799        /// reported skipped instead of failed (the action was already
800        /// done / not applicable).
801        #[serde(default)]
802        idempotent: bool,
803    },
804
805    /// Wait for an element to appear on the page.
806    #[serde(rename = "wait")]
807    Wait {
808        /// Natural language description of the element to wait for.
809        target: String,
810        /// Explicit CSS selector override (bypasses LLM resolution).
811        #[serde(default)]
812        selector: Option<String>,
813        /// Wait until the page's visible text contains this substring
814        /// (alternative to `selector`; either or both may be set — both are
815        /// required to hold when both are set).
816        #[serde(default)]
817        text: Option<String>,
818        /// Maximum milliseconds to wait (default: 10000).
819        #[serde(default)]
820        timeout_ms: Option<u64>,
821        /// Endpoint to use for LLM element targeting.
822        #[serde(default)]
823        endpoint: Option<String>,
824        /// Idempotent: when the condition never becomes true within the
825        /// timeout the step is reported skipped instead of failed (the
826        /// condition was not applicable, e.g. already-authenticated
827        /// pages in a viewport matrix).
828        #[serde(default)]
829        idempotent: bool,
830    },
831
832    /// Evaluate an assertion against the current page content.
833    #[serde(rename = "assert")]
834    Assert {
835        /// Reference to a named `[[definitions]]` entry.
836        #[serde(default)]
837        definition: Option<String>,
838        /// Inline predefined assertion preset (e.g. `no_error_on_page`).
839        #[serde(default)]
840        preset: Option<String>,
841        /// Inline custom LLM prompt for assertion evaluation.
842        #[serde(default)]
843        prompt: Option<String>,
844        /// Text that the `text_visible` preset checks for.
845        #[serde(default)]
846        assert_text: Option<String>,
847        /// Endpoint to use for this assertion's LLM call.
848        #[serde(default)]
849        endpoint: Option<String>,
850        /// Attach a screenshot of the current viewport to the assertion so
851        /// the LLM can evaluate visuals (overlaps, clipping, layout).
852        /// Requires the resolved endpoint to declare `vision = true`.
853        #[serde(default)]
854        screenshot: bool,
855    },
856
857    /// Take a screenshot of the current page.
858    #[serde(rename = "screenshot")]
859    Screenshot {
860        /// File path to save the screenshot (default: `screenshot.png`).
861        #[serde(default)]
862        path: Option<String>,
863    },
864
865    /// Call an A2A agent with a task.
866    #[serde(rename = "agent")]
867    Agent {
868        /// Name of the agent endpoint to call.
869        agent: String,
870        /// Task description / prompt for the agent.
871        task: String,
872        /// Optional definition name with a task template.
873        #[serde(default)]
874        definition: Option<String>,
875    },
876
877    /// Call an MCP server tool.
878    #[serde(rename = "mcp")]
879    Mcp {
880        /// Name of the MCP server endpoint.
881        server: String,
882        /// Tool name to invoke on the server.
883        tool: String,
884        /// Tool arguments as JSON.
885        #[serde(default)]
886        args: Option<serde_json::Value>,
887    },
888}
llm_browser_testkit/scenario.rs

llm_browser_testkit/
scenario.rs
