mockforge_bench/conformance/self_test.rs
1//! Positive + per-category negative request driver against a live server.
2//!
3//! Issue #79 round 13 (4) — Srikanth's (e) ask: a way to test both
4//! positive and negative compliance scenarios separately, where the
5//! positive cases should pass and the negative cases should be
6//! rejected.
7//!
8//! This module sits *alongside* the existing conformance executor
9//! (which drives k6 / native checks on a single positive call per
10//! operation). The self-test driver synthesises per-category
11//! deliberately-bad requests and asserts that the server actually
12//! rejects them with a 4xx — useful when verifying that
13//! `validate_request_with_all` is wired correctly for the user's spec
14//! (the exact gap that round-13 (3) fixed).
15//!
16//! Scope of the initial MVP: covers the highest-signal negatives —
17//! empty body when one is required, missing required query/header
18//! params, and wrong-type path params. Doesn't try to mutate every
19//! field of a JSON-Schema-validated body; that's a follow-up.
20
21use super::spec_driven::{AnnotatedOperation, ApiKeyLocation, SecuritySchemeInfo};
22use reqwest::{Client, Method};
23use std::collections::BTreeMap;
24use std::net::IpAddr;
25use std::sync::atomic::{AtomicUsize, Ordering};
26use std::sync::{Arc, Mutex};
27use std::time::Duration;
28
29/// Round 23 (c-iii) — per-direction body cap when capturing
30/// request/response payloads to `conformance-self-test-requests.jsonl`.
31/// 16 KiB keeps a 1000-case run under ~32 MB even if every payload
32/// fills the cap, while still preserving enough of a typical JSON body
33/// (or a stack-trace error response) to debug from.
34const CAPTURE_BODY_CAP_BYTES: usize = 16 * 1024;
35
36/// Round 17.2 — cap on schema-driven negatives per operation. A spec
37/// with 100 properties per body could produce hundreds of mutations
38/// for a single operation; combined with thousands of operations
39/// that's a runaway test matrix. 12 covers the highest-signal
40/// mutations (type mismatch + required-removed + a few constraint
41/// breaks) without exploding wall time on large specs.
42const SCHEMA_MUTATION_CAP: usize = 12;
43
44/// Round 25 (k) — content-type swap probes. For operations declaring a
45/// JSON request body, each entry below produces one probe that lies
46/// about Content-Type while keeping the JSON payload. A spec-compliant
47/// server should respond 415 (or 400). Order matches the order
48/// Srikanth listed in his round-23 reply: XML, YAML, multipart, and
49/// the URL-encoded variant he added in round 24.
50const CONTENT_TYPE_SWAP_VARIANTS: &[(&str, &str)] = &[
51 ("application/xml", "request-body:content-type-mismatch:xml"),
52 ("application/yaml", "request-body:content-type-mismatch:yaml"),
53 ("multipart/form-data", "request-body:content-type-mismatch:multipart"),
54 (
55 "application/x-www-form-urlencoded",
56 "request-body:content-type-mismatch:urlencoded",
57 ),
58];
59
60/// Round 27 (k variant b) — embedded content payloads. Content-Type
61/// stays `application/json` and the envelope IS valid JSON; we just
62/// stuff a non-JSON snippet into a string field's value. The test
63/// surfaces servers that try to parse string field contents (e.g.
64/// XML-EE expanders, YAML loaders, urlencoded parsers) and crash on
65/// the payload — a 5xx here is the finding. Label, payload pairs:
66const EMBEDDED_CONTENT_VARIANTS: &[(&str, &str)] = &[
67 ("request-body:embedded-content:xml", "<root><cmd>execute()</cmd></root>"),
68 ("request-body:embedded-content:yaml", "key: value\n- item1\n- item2"),
69 (
70 "request-body:embedded-content:multipart",
71 "--boundary\r\nContent-Disposition: form-data; name=\"x\"\r\n\r\nval\r\n--boundary--",
72 ),
73 ("request-body:embedded-content:urlencoded", "a=1&b=2&c=hello%20world"),
74];
75
76/// Configuration for a self-test run.
77#[derive(Debug, Clone)]
78pub struct SelfTestConfig {
79 pub target_url: String,
80 pub skip_tls_verify: bool,
81 pub timeout: Duration,
82 /// Optional extra headers to attach to every request (e.g. auth).
83 pub extra_headers: Vec<(String, String)>,
84 /// Delay between requests to avoid hammering the server.
85 pub delay_between_requests: Duration,
86 /// Round 18.1 — base path to prepend to every spec path. When the
87 /// spec declares `/users` and the deployed API is served under
88 /// `/api`, `--base-path /api` should make the self-test hit
89 /// `https://target/api/users` instead of `https://target/users`.
90 /// Pre-fix this was ignored entirely and every operation 404'd
91 /// (Srikanth's vCenter run on 0.3.152: 1275 positives, 1275 4xx).
92 pub base_path: Option<String>,
93 /// Round 18.5 — local source IPs to bind outgoing requests to.
94 /// Each IP must already be assigned to an interface on the host.
95 /// Operations round-robin through the resulting client pool.
96 pub source_ips: Vec<IpAddr>,
97 /// Round 18.5 — fake source IPs to advertise via forwarded-IP
98 /// headers (used to exercise GEODB lookup at the destination).
99 /// Rotated per operation.
100 pub geo_source_ips: Vec<IpAddr>,
101 /// Which forwarded-IP header(s) to populate when `geo_source_ips`
102 /// is non-empty. Empty → no-op; default below sets the standard
103 /// three-header set.
104 pub geo_source_headers: Vec<String>,
105 /// Round 23 (c-iii) — when `Some`, every probe captures method, URL,
106 /// request headers/body and response status/headers/body into this
107 /// sink. Caller drains it after `run_self_test` and writes
108 /// `conformance-self-test-requests.jsonl`. None → no capture (zero
109 /// extra allocations on the hot path).
110 pub capture: Option<Arc<Mutex<Vec<CaseCapture>>>>,
111 /// Round 25 — when true, validate every probe's response body
112 /// against the spec's response schema for the actual status
113 /// returned (closes round 21.3 / Srikanth's a2 / a3 ask). The
114 /// validation result lands in `CaseCapture::response_schema_error`
115 /// (None → matched, or no schema for that status). Default false:
116 /// JSON-Schema validation of large response bodies adds wall-clock
117 /// time and the user has to opt in.
118 pub validate_response_schemas: bool,
119 /// Round 33 (#823) — human-readable label for the OpenAPI spec
120 /// this run is exercising. Stamped on every `CaseCapture` so the
121 /// per-endpoint summary can attribute rows back to a spec in
122 /// multi-spec / multi-target benches. `None` when the bench didn't
123 /// track a spec path.
124 pub spec_label: Option<String>,
125 /// Round 47 (#79) — Srikanth on 0.3.191: "I did not see network
126 /// logs in the mockforge bench and conformance traffic if used
127 /// the [self-test] command". The r46 wire-level event sink only
128 /// existed on the native conformance executor; this matches it on
129 /// the self-test side. When `Some`, every `reqwest::Error` from
130 /// `send().await` is classified and pushed to this sink; caller
131 /// drains it into `conformance-network-events.json` next to the
132 /// JSONL capture. None → no extra allocations on the hot path.
133 pub network_events: Option<Arc<Mutex<Vec<NetworkEvent>>>>,
134}
135
136/// Round 47 (#79) — wire-level network event captured by the self-test
137/// driver. Same shape as the native executor's `NetworkEvent` so
138/// downstream tooling can consume one file across executor variants.
139#[derive(Debug, Clone, serde::Serialize)]
140pub struct NetworkEvent {
141 pub timestamp: chrono::DateTime<chrono::Utc>,
142 pub check: String,
143 pub method: String,
144 pub url: String,
145 pub kind: String,
146 pub message: String,
147}
148
149/// Round 23 (c-iii) — one captured request/response pair, one per
150/// probe (positive or negative). Serialised as a JSON line in
151/// `conformance-self-test-requests.jsonl`. Headers are kept as
152/// `BTreeMap` for stable ordering. Bodies are truncated to
153/// `CAPTURE_BODY_CAP_BYTES`; `*_truncated` flags whether more was
154/// dropped.
155#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
156pub struct CaseCapture {
157 pub label: String,
158 pub method: String,
159 pub url: String,
160 pub request_headers: BTreeMap<String, String>,
161 pub request_body: Option<String>,
162 pub request_body_truncated: bool,
163 pub response_status: u16,
164 pub response_headers: BTreeMap<String, String>,
165 pub response_body: Option<String>,
166 pub response_body_truncated: bool,
167 pub error: Option<String>,
168 /// Round 25 — when `validate_response_schemas` is on and the spec
169 /// declares a schema for `response_status`, this carries the
170 /// validation message (or None when the body matched, or no schema
171 /// was declared for that status). Serialised verbatim in the JSONL
172 /// and rendered in the HTML viewer.
173 #[serde(default, skip_serializing_if = "Option::is_none")]
174 pub response_schema_error: Option<String>,
175 /// Round 28 — Srikanth's "Is it possible to put expected response
176 /// code status in both jsonl and jsonl report" ask. Human-readable
177 /// expected status range: `"2xx-3xx"` for positive probes,
178 /// `"4xx"` for negatives. Lets users `jq` for misses
179 /// (`.response_status as $s | .expected_status_range == "4xx"
180 /// and ($s < 400 or $s >= 500)`) and powers the HTML viewer's
181 /// "show mismatches only" filter.
182 #[serde(default)]
183 pub expected_status_range: String,
184 /// Round 33 (#823) — the spec's path template (e.g.
185 /// `/users/{id}`) before path-param substitution. Lets the
186 /// per-endpoint summary collapse `/users/X` and `/users/Y` into
187 /// one row. Empty string when the call site predates this field
188 /// (older `CaseCapture` payloads on disk also deserialise OK).
189 #[serde(default)]
190 pub path_template: String,
191 /// Round 33 (#823) — basename (or fallback to full path) of the
192 /// OpenAPI spec file this probe came from. Lets multi-spec runs
193 /// attribute rows back to the spec they came from. `None` when
194 /// the bench didn't track a spec path.
195 #[serde(default, skip_serializing_if = "Option::is_none")]
196 pub spec_label: Option<String>,
197 /// Round 36 (#876) — mockforge version that ran the probe.
198 /// Stamped from `CARGO_PKG_VERSION` at compile time. Also sent
199 /// as the `X-Mockforge-Client-Version` request header so a
200 /// matching `ServerConformanceViolation.client_mockforge_version`
201 /// can be cross-correlated. Empty string when the capture
202 /// pre-dates this field.
203 #[serde(default)]
204 pub mockforge_version: String,
205 /// Round 36 (#876) — wall-clock moment the bench driver sent the
206 /// request, as RFC3339 / ISO-8601. Also sent as the
207 /// `X-Mockforge-Client-Sent-At` request header so the server-side
208 /// `ServerConformanceViolation.client_sent_at` carries the same
209 /// value. Empty string when the capture pre-dates this field.
210 #[serde(default)]
211 pub client_sent_at: String,
212}
213
214impl Default for SelfTestConfig {
215 fn default() -> Self {
216 Self {
217 target_url: "http://localhost:3000".into(),
218 skip_tls_verify: false,
219 timeout: Duration::from_secs(15),
220 extra_headers: Vec::new(),
221 delay_between_requests: Duration::from_millis(0),
222 base_path: None,
223 source_ips: Vec::new(),
224 geo_source_ips: Vec::new(),
225 geo_source_headers: default_geo_source_headers(),
226 capture: None,
227 validate_response_schemas: false,
228 spec_label: None,
229 network_events: None,
230 }
231 }
232}
233
234/// Truncate `body` to `CAPTURE_BODY_CAP_BYTES` on a UTF-8 boundary,
235/// returning the trimmed string and whether truncation occurred. Used
236/// for both request and response bodies in the capture sink.
237fn truncate_body_for_capture(body: &str) -> (String, bool) {
238 if body.len() <= CAPTURE_BODY_CAP_BYTES {
239 return (body.to_string(), false);
240 }
241 let mut end = CAPTURE_BODY_CAP_BYTES;
242 while end > 0 && !body.is_char_boundary(end) {
243 end -= 1;
244 }
245 (body[..end].to_string(), true)
246}
247
248/// Default forwarded-IP header set. Covers the three conventions a
249/// real GEODB front-end is likely to read in this order of
250/// preference: Cloudflare (`CF-Connecting-IP`), Akamai/CloudFront
251/// (`True-Client-IP`), then the de-facto standard
252/// `X-Forwarded-For`. Override via `--geo-source-header` to test a
253/// specific stack.
254pub fn default_geo_source_headers() -> Vec<String> {
255 vec![
256 "X-Forwarded-For".to_string(),
257 "True-Client-IP".to_string(),
258 "CF-Connecting-IP".to_string(),
259 ]
260}
261
262/// Outcome of a single test case (positive or negative).
263#[derive(Debug, Clone, serde::Serialize)]
264pub struct CaseOutcome {
265 pub label: String,
266 pub expected_4xx: bool,
267 pub actual_status: u16,
268 /// True when the response status matches expectation
269 /// (positive → 2xx-3xx, negative → 4xx).
270 pub passed: bool,
271}
272
273/// All cases run against one annotated operation.
274#[derive(Debug, Clone, serde::Serialize)]
275pub struct OperationResult {
276 pub method: String,
277 pub path: String,
278 pub positive: Option<CaseOutcome>,
279 pub negatives: Vec<CaseOutcome>,
280}
281
282/// Summary report rolled up across all operations.
283#[derive(Debug, Default, Clone, serde::Serialize)]
284pub struct SelfTestReport {
285 pub positive_pass: usize,
286 pub positive_fail: usize,
287 /// Per category: count of negative cases the server correctly
288 /// rejected with a 4xx (we caught the spec violation).
289 pub negative_caught: BTreeMap<String, usize>,
290 /// Per category: count of negative cases that should have been
291 /// rejected but came back with a non-4xx (validator gap).
292 pub negative_missed: BTreeMap<String, usize>,
293 pub operations: Vec<OperationResult>,
294}
295
296impl SelfTestReport {
297 /// All-pass means every positive case got 2xx-3xx and every
298 /// negative case got 4xx.
299 pub fn all_passed(&self) -> bool {
300 self.positive_fail == 0 && self.negative_missed.values().sum::<usize>() == 0
301 }
302
303 /// Round 18.1 — detect the "self-test target is misconfigured"
304 /// case where every positive failed with the *same* status code.
305 /// The classic example: `--base-path /api` was forgotten so every
306 /// request hits a path the server doesn't know and returns 404.
307 /// Pre-warning, the user saw all-green negative buckets (because
308 /// "missing route" 404s look like "validator rejected") and no
309 /// indication that the run was meaningless. Returns Some(status)
310 /// when ≥10 positives all failed with the same status, else None.
311 pub fn detect_target_misconfiguration(&self) -> Option<u16> {
312 if self.positive_pass > 0 || self.positive_fail < 10 {
313 return None;
314 }
315 let mut seen: Option<u16> = None;
316 for op in &self.operations {
317 let Some(p) = &op.positive else {
318 continue;
319 };
320 if p.passed {
321 return None;
322 }
323 match seen {
324 None => seen = Some(p.actual_status),
325 Some(s) if s != p.actual_status => return None,
326 _ => {}
327 }
328 }
329 seen
330 }
331
332 /// Round 47 (#79) — fold a second iteration of the self-test into
333 /// this report so multi-iteration runs aggregate counters across
334 /// passes. Per-category caught / missed counters sum; positive
335 /// counters sum; the `operations` vec records every probe outcome
336 /// so the iteration-N misconfiguration detector still works. Used
337 /// by command.rs's `--conformance-self-test-iterations` /
338 /// `--conformance-self-test-duration` loop.
339 pub fn merge_iteration(&mut self, other: SelfTestReport) {
340 self.positive_pass = self.positive_pass.saturating_add(other.positive_pass);
341 self.positive_fail = self.positive_fail.saturating_add(other.positive_fail);
342 for (k, v) in other.negative_caught {
343 let slot = self.negative_caught.entry(k).or_insert(0);
344 *slot = slot.saturating_add(v);
345 }
346 for (k, v) in other.negative_missed {
347 let slot = self.negative_missed.entry(k).or_insert(0);
348 *slot = slot.saturating_add(v);
349 }
350 self.operations.extend(other.operations);
351 }
352
353 /// Human-readable summary string. One line for positives, one per
354 /// category for negatives. Designed to slot into existing
355 /// `TerminalReporter` output.
356 pub fn render_summary(&self) -> String {
357 let mut out = String::new();
358 out.push_str(&format!(
359 "Positives: {} pass / {} fail\n",
360 self.positive_pass, self.positive_fail
361 ));
362 let mut keys: Vec<&String> =
363 self.negative_caught.keys().chain(self.negative_missed.keys()).collect();
364 keys.sort();
365 keys.dedup();
366 for cat in keys {
367 let caught = self.negative_caught.get(cat).copied().unwrap_or(0);
368 let missed = self.negative_missed.get(cat).copied().unwrap_or(0);
369 let mark = if missed == 0 { "✓" } else { "⚠" };
370 out.push_str(&format!(
371 "Negatives [{}]: {} caught / {} missed {}\n",
372 cat, caught, missed, mark
373 ));
374 }
375 out
376 }
377}
378
379/// Execute the self-test plan against `config.target_url` for every
380/// `AnnotatedOperation`. Returns the aggregated report; callers
381/// decide how to display it (e.g. via `render_summary` or by writing
382/// the JSON serialisation to disk).
383pub async fn run_self_test(
384 operations: &[AnnotatedOperation],
385 config: &SelfTestConfig,
386) -> Result<SelfTestReport, reqwest::Error> {
387 // Round 18.5 — build a client pool when `source_ips` is set,
388 // one reqwest::Client per IP, each bound to its local address.
389 // Operations round-robin through the pool. Empty pool → single
390 // default client (the pre-18.5 behaviour).
391 let clients = build_client_pool(config)?;
392 let client_cursor = AtomicUsize::new(0);
393 let geo_cursor = AtomicUsize::new(0);
394
395 let mut report = SelfTestReport::default();
396 for op in operations {
397 let client_idx = client_cursor.fetch_add(1, Ordering::Relaxed) % clients.len();
398 let client = &clients[client_idx];
399 let geo_ip = if config.geo_source_ips.is_empty() {
400 None
401 } else {
402 let idx = geo_cursor.fetch_add(1, Ordering::Relaxed) % config.geo_source_ips.len();
403 Some(config.geo_source_ips[idx])
404 };
405 let result = test_operation(client, config, op, geo_ip).await;
406 if let Some(p) = &result.positive {
407 if p.passed {
408 report.positive_pass += 1;
409 } else {
410 report.positive_fail += 1;
411 }
412 }
413 for neg in &result.negatives {
414 let cat = neg.label.split(':').next().unwrap_or("other").to_string();
415 if neg.passed {
416 *report.negative_caught.entry(cat).or_insert(0) += 1;
417 } else {
418 *report.negative_missed.entry(cat).or_insert(0) += 1;
419 }
420 }
421 report.operations.push(result);
422 if !config.delay_between_requests.is_zero() {
423 tokio::time::sleep(config.delay_between_requests).await;
424 }
425 }
426 Ok(report)
427}
428
429/// Round 18.5 — append GEODB forwarded-IP headers to the
430/// operation's declared headers. Returns the original vec untouched
431/// when `geo_ip` is None or `geo_headers` is empty.
432///
433/// If the operation already declares one of the geo headers (rare
434/// but legal), we keep the operation's value — the caller's spec
435/// wins.
436fn effective_op_headers(
437 base: &[(String, String)],
438 geo_ip: Option<IpAddr>,
439 geo_headers: &[String],
440) -> Vec<(String, String)> {
441 let mut out = base.to_vec();
442 let Some(ip) = geo_ip else {
443 return out;
444 };
445 let value = ip.to_string();
446 for h in geo_headers {
447 // Case-insensitive duplicate check: don't override the
448 // spec's own declared value for the header.
449 if out.iter().any(|(k, _)| k.eq_ignore_ascii_case(h)) {
450 continue;
451 }
452 out.push((h.clone(), value.clone()));
453 }
454 out
455}
456
457/// Round 18.5 — build a pool of reqwest clients, one per declared
458/// source IP. Empty `source_ips` → a single default client.
459///
460/// The OS must already have each `source_ip` assigned to an
461/// interface; reqwest's `.local_address()` issues a `bind()` syscall
462/// at connect time, so an IP the kernel doesn't recognise surfaces
463/// as `EADDRNOTAVAIL` at request time, not at builder time.
464fn build_client_pool(config: &SelfTestConfig) -> Result<Vec<Client>, reqwest::Error> {
465 let make = |bind: Option<IpAddr>| -> Result<Client, reqwest::Error> {
466 let mut builder = Client::builder().timeout(config.timeout);
467 if config.skip_tls_verify {
468 builder = builder.danger_accept_invalid_certs(true);
469 }
470 if let Some(addr) = bind {
471 builder = builder.local_address(addr);
472 }
473 builder.build()
474 };
475 if config.source_ips.is_empty() {
476 Ok(vec![make(None)?])
477 } else {
478 config.source_ips.iter().map(|ip| make(Some(*ip))).collect()
479 }
480}
481
482async fn test_operation(
483 client: &Client,
484 config: &SelfTestConfig,
485 op: &AnnotatedOperation,
486 geo_ip: Option<IpAddr>,
487) -> OperationResult {
488 // Round 25 — track the sink length BEFORE we run any probes for
489 // this operation, so that after the probes finish we can mutate
490 // exactly the entries that belong to this op (the capture sink is
491 // shared but `run_self_test` iterates operations sequentially).
492 // Used by the response-schema validation pass below.
493 let sink_start = config.capture.as_ref().and_then(|s| s.lock().ok().map(|g| g.len()));
494
495 let url = build_url_with_base(
496 &config.target_url,
497 config.base_path.as_deref(),
498 &op.path,
499 &op.path_params,
500 );
501 let method = Method::from_bytes(op.method.to_uppercase().as_bytes()).unwrap_or(Method::GET);
502
503 // Round 34 (#828) — stamp every `CaseCapture` with the spec
504 // template PREFIXED by `--base-path`, so the per-endpoint
505 // summary's `path` column matches what the user sees in URLs
506 // and logs. Srikanth searched for `/api/appliance/access/...`
507 // and didn't find it because round 33 stored just `/appliance/
508 // access/...`. Same normalization as `build_url_with_base`:
509 // leading `/` auto-added, trailing `/` stripped, empty
510 // base_path → no prefix at all.
511 let path_template = {
512 let prefix = match config.base_path.as_deref() {
513 Some(bp) if !bp.is_empty() => {
514 let trimmed = bp.trim_end_matches('/');
515 if trimmed.starts_with('/') {
516 trimmed.to_string()
517 } else {
518 format!("/{}", trimmed)
519 }
520 }
521 _ => String::new(),
522 };
523 let path = if op.path.starts_with('/') {
524 op.path.clone()
525 } else {
526 format!("/{}", op.path)
527 };
528 format!("{prefix}{path}")
529 };
530
531 // Round 18.5 — pre-compute the operation's effective headers
532 // with the geo source IP baked in. Doing it once here keeps the
533 // per-case `send_case` calls below unchanged. When `geo_ip` is
534 // None the result equals `op.header_params`.
535 let op_headers = effective_op_headers(&op.header_params, geo_ip, &config.geo_source_headers);
536
537 // ── Positive case ────────────────────────────────────────────
538 let positive = send_case(
539 client,
540 config,
541 method.clone(),
542 &url,
543 "positive",
544 ExpectedOutcome::Success,
545 op.sample_body.as_deref(),
546 op.query_params.clone(),
547 op_headers.clone(),
548 &path_template,
549 )
550 .await;
551
552 // ── Negative cases ───────────────────────────────────────────
553 let mut negatives = Vec::new();
554
555 // (a) empty body when one is required.
556 //
557 // Round 16 — drop the `sample_body.is_some()` precondition. Operations
558 // whose body annotator couldn't synthesize a sample previously got
559 // zero negatives (so the self-test reported "all passing" even on
560 // POST /resource with a required body). The spec saying the operation
561 // *has* a request body is enough — an empty object is a valid
562 // negative regardless of whether we have a positive sample.
563 if op.request_body_content_type.is_some() {
564 negatives.push(
565 send_case(
566 client,
567 config,
568 method.clone(),
569 &url,
570 "request-body:empty",
571 ExpectedOutcome::ClientError,
572 Some("{}"),
573 op.query_params.clone(),
574 op_headers.clone(),
575 &path_template,
576 )
577 .await,
578 );
579
580 // (b) wrong-shaped body (array instead of object) — exercises
581 // top-level type validation independently of which fields are
582 // required.
583 negatives.push(
584 send_case(
585 client,
586 config,
587 method.clone(),
588 &url,
589 "request-body:wrong-type",
590 ExpectedOutcome::ClientError,
591 Some("[]"),
592 op.query_params.clone(),
593 op_headers.clone(),
594 &path_template,
595 )
596 .await,
597 );
598
599 // Round 25 (k) — content-type swap probes.
600 //
601 // For operations declaring `application/json` request bodies, send
602 // the SAME json payload (or a synthesised one) under four other
603 // content types: `application/xml`, `application/yaml`,
604 // `multipart/form-data`, `application/x-www-form-urlencoded`.
605 // The spec says the endpoint accepts only JSON, so a strict server
606 // should respond 415 Unsupported Media Type (or 400 if it tries
607 // to parse and fails). A 2xx means the server is accepting
608 // payloads outside its declared content negotiation, which is the
609 // failure mode behind a lot of "we crashed on a malformed XML
610 // upload" incidents.
611 //
612 // Variant (a) of Srikanth's round-23 g ask: lie about the
613 // Content-Type header. The body shape is honest JSON; only the
614 // header is swapped. Variant (b) (JSON envelope with embedded
615 // non-JSON field values) is deferred to round 26 because it
616 // requires a schema-aware field walker.
617 if op
618 .request_body_content_type
619 .as_deref()
620 .map(|ct| ct.contains("json"))
621 .unwrap_or(false)
622 {
623 let payload = op.sample_body.as_deref().unwrap_or("{}");
624 for (ct, label) in CONTENT_TYPE_SWAP_VARIANTS {
625 negatives.push(
626 send_case_with_extra(
627 client,
628 config,
629 method.clone(),
630 &url,
631 label,
632 ExpectedOutcome::ClientError,
633 Some(payload),
634 op.query_params.clone(),
635 // Strip any Content-Type already on the operation
636 // headers (the spec's positive value) so the
637 // probe's value is the only one the server sees.
638 op_headers
639 .iter()
640 .filter(|(k, _)| !k.eq_ignore_ascii_case("content-type"))
641 .cloned()
642 .collect(),
643 // The wrong Content-Type rides on `extra_headers`
644 // so it lands AFTER `send_case_with_extra`'s
645 // unconditional `application/json` insertion in
646 // request-body mode. Actually `send_case_with_extra`
647 // only sets Content-Type when a body is present
648 // AND there's no manual override; passing the
649 // override here wins because reqwest preserves
650 // the last-set header value.
651 vec![("Content-Type".to_string(), (*ct).to_string())],
652 &path_template,
653 )
654 .await,
655 );
656 }
657
658 // Round 27 (k variant b) — embedded non-JSON content
659 // inside a valid JSON envelope. Content-Type stays
660 // application/json (honest) and the body parses as JSON;
661 // only the string-valued payload changes. We expect 2xx-3xx
662 // because the envelope is spec-shape, so the probe surfaces
663 // servers that crash (5xx) trying to parse the embedded
664 // snippet as XML/YAML/etc. A 4xx is also a finding because
665 // it usually means the server's pattern/format validator
666 // tripped on the payload contents, but the user can decide
667 // from the JSONL whether that's a bug or correct narrow-
668 // string-field behaviour.
669 for (label, snippet) in EMBEDDED_CONTENT_VARIANTS {
670 let payload = op.sample_body.as_deref().unwrap_or("{}");
671 // Round 34 (#829) — skip the probe entirely when the
672 // positive sample has no string leaf we can mutate.
673 // The previous round-27 fallback `{"data": <snippet>}`
674 // produced a body that doesn't match the spec's actual
675 // schema for endpoints like vCenter's `consolecli` PUT
676 // (which wants `{enabled: bool}`), so the server
677 // correctly 400'd and the bench misreported the
678 // mismatch as an expectation failure.
679 let Some(body) = embed_payload_in_first_string_field(payload, snippet) else {
680 continue;
681 };
682 negatives.push(
683 send_case(
684 client,
685 config,
686 method.clone(),
687 &url,
688 label,
689 // expected_4xx=false: any non-2xx is a probe
690 // failure. 5xx in particular is "server panicked
691 // on the embedded content".
692 ExpectedOutcome::NotServerError,
693 Some(&body),
694 op.query_params.clone(),
695 op_headers.clone(),
696 &path_template,
697 )
698 .await,
699 );
700 }
701 }
702
703 // Round 17.2 — schema-aware negatives.
704 //
705 // When both a positive sample AND the resolved body schema are
706 // available, mutate the sample per-field (type mismatch,
707 // min/max bounds, pattern, enum out-of-range, required-field
708 // removal) and assert each is rejected with 4xx. Capped at
709 // SCHEMA_MUTATION_CAP per operation so a 100-property body
710 // doesn't explode the test matrix.
711 if let (Some(sample_str), Some(schema)) =
712 (op.sample_body.as_deref(), op.request_body_schema.as_ref())
713 {
714 if let Ok(sample) = serde_json::from_str::<serde_json::Value>(sample_str) {
715 let mutations = super::schema_mutator::mutate_body(&sample, schema);
716 for m in mutations.into_iter().take(SCHEMA_MUTATION_CAP) {
717 let body_str = serde_json::to_string(&m.body).unwrap_or_default();
718 negatives.push(
719 send_case(
720 client,
721 config,
722 method.clone(),
723 &url,
724 &m.label,
725 ExpectedOutcome::ClientError,
726 Some(&body_str),
727 op.query_params.clone(),
728 // Round 24 (f) — was `op.header_params`, which
729 // skipped the geo-IP header. Use `op_headers`
730 // so the geo IP rides with the negative probe
731 // too (positive vs negative coverage must be
732 // symmetric, otherwise a GEODB front-end sees
733 // the rotating IP only on positives).
734 op_headers.clone(),
735 &path_template,
736 )
737 .await,
738 );
739 }
740 }
741 }
742 }
743
744 // Round 17.2 — URI-length probe. Spec-agnostic but schema-aware in
745 // spirit: most servers cap URIs at 8 KB or so. Append a 9 KB query
746 // string to the URL and expect 414 URI Too Long (or 400). Skipped
747 // for operations that already have a heavy positive query.
748 {
749 let pad = "p=".to_string() + &"x".repeat(9_000);
750 let bad_url = if url.contains('?') {
751 format!("{url}&{pad}")
752 } else {
753 format!("{url}?{pad}")
754 };
755 negatives.push(
756 send_case(
757 client,
758 config,
759 method.clone(),
760 &bad_url,
761 "parameters:uri-too-long",
762 ExpectedOutcome::ClientError,
763 op.sample_body.as_deref(),
764 op.query_params.clone(),
765 // Round 24 (f) — see schema-mutation note above. Use
766 // `op_headers` (carries geo IP) instead of bare
767 // `op.header_params`.
768 op_headers.clone(),
769 &path_template,
770 )
771 .await,
772 );
773 }
774
775 // (e) Round 16 — path-param type probe. Send the first path
776 // parameter as a literal `"self-test-invalid-id"`: a string that
777 // contains hyphens, won't parse as an integer, won't parse as a
778 // UUID, and won't match any typical regex pattern. Operations
779 // whose spec types the param as `integer` or `string` with a
780 // `format`/`pattern` will catch this (caught: server returned
781 // 4xx); operations whose spec lets path params be free-form
782 // strings will let it through (missed: server returned 2xx).
783 // Either outcome is informative: a category that's all "missed"
784 // tells the user their spec is loose on path-param types, which
785 // is itself worth knowing. Addresses Srikanth's "always all
786 // passing" report — operations with a path param now produce at
787 // least one probe instead of zero.
788 if !op.path_params.is_empty() {
789 let mut url_with_placeholder = op.path.clone();
790 if let Some((first_name, _)) = op.path_params.first() {
791 // Substitute every other path-param with its sample so the
792 // route shape stays intact and only the first param is bad.
793 for (name, value) in op.path_params.iter().skip(1) {
794 if !value.is_empty() {
795 url_with_placeholder =
796 url_with_placeholder.replace(&format!("{{{name}}}"), value);
797 }
798 }
799 // Substitute the first param with a guaranteed-invalid
800 // sentinel that's unlikely to match any reasonable schema:
801 // contains characters disallowed in numeric IDs *and* UUIDs.
802 url_with_placeholder =
803 url_with_placeholder.replace(&format!("{{{first_name}}}"), "self-test-invalid-id");
804 // Round 18.1 — honour `base_path` here too, otherwise the
805 // probe URL differs from the positive case and the
806 // resulting 404 is misattributed to "bad path param".
807 let bad_url = build_url_with_base(
808 &config.target_url,
809 config.base_path.as_deref(),
810 &url_with_placeholder,
811 &[],
812 );
813 negatives.push(
814 send_case(
815 client,
816 config,
817 method.clone(),
818 &bad_url,
819 "parameters:bad-path-param",
820 ExpectedOutcome::ClientError,
821 op.sample_body.as_deref(),
822 op.query_params.clone(),
823 op_headers.clone(),
824 &path_template,
825 )
826 .await,
827 );
828 }
829 }
830
831 // (c) drop the first required query param
832 if !op.query_params.is_empty() {
833 let mut q = op.query_params.clone();
834 q.remove(0);
835 negatives.push(
836 send_case(
837 client,
838 config,
839 method.clone(),
840 &url,
841 "parameters:missing-query",
842 ExpectedOutcome::ClientError,
843 op.sample_body.as_deref(),
844 q,
845 op_headers.clone(),
846 &path_template,
847 )
848 .await,
849 );
850 }
851
852 // (s) Round 17.3 — security probes.
853 //
854 // Operations whose spec declares a security requirement get a
855 // dedicated set of negatives. The point isn't to test whether the
856 // server's *real* auth works (the positive case already does that
857 // via `extra_headers`) — it's to check whether deliberately-bad
858 // credentials are still rejected, which is exactly the failure
859 // mode that lets an attacker through a half-wired validator.
860 //
861 // Each probe replaces or omits the relevant auth credential and
862 // expects 401 / 403. A 2xx here is a hard finding: "spec says
863 // this endpoint is protected, server let unauthenticated /
864 // wrong-credential traffic through".
865 //
866 // Bounded: at most one probe per declared scheme kind, so an
867 // operation with 3 security requirements doesn't 4× the request
868 // volume. Skips entirely when `op.security_schemes` is empty.
869 for probe in build_security_probes(&op.security_schemes) {
870 // Strip any pre-existing Authorization or known API-key
871 // header from extra_headers + header_params so the probe
872 // value is the *only* credential the server sees.
873 let stripped_extra = strip_auth(&config.extra_headers, &op.security_schemes);
874 let stripped_headers = strip_auth(&op.header_params, &op.security_schemes);
875 let stripped_query = strip_auth_query(&op.query_params, &op.security_schemes);
876 let mut req_headers = stripped_headers;
877 for (k, v) in &probe.headers {
878 req_headers.push((k.clone(), v.clone()));
879 }
880 // Round 24 (f) — security probes build req_headers from
881 // `op.header_params` directly (we need the stripped-auth
882 // variant), so the geo-IP header doesn't ride along
883 // automatically. Append it here so a GEODB / WAF in front
884 // of the auth layer still sees the rotating source IP.
885 if let Some(ip) = geo_ip {
886 let ip_str = ip.to_string();
887 for h in &config.geo_source_headers {
888 let already = req_headers.iter().any(|(k, _)| k.eq_ignore_ascii_case(h));
889 if !already {
890 req_headers.push((h.clone(), ip_str.clone()));
891 }
892 }
893 }
894 let mut req_query = stripped_query;
895 for (k, v) in &probe.query {
896 req_query.push((k.clone(), v.clone()));
897 }
898 negatives.push(
899 send_case_with_extra(
900 client,
901 config,
902 method.clone(),
903 &url,
904 &probe.label,
905 ExpectedOutcome::ClientError,
906 op.sample_body.as_deref(),
907 req_query,
908 req_headers,
909 stripped_extra,
910 &path_template,
911 )
912 .await,
913 );
914 }
915
916 // (d) drop the first required header
917 if !op.header_params.is_empty() {
918 // Round 24 (f) — start from `op_headers` (so the geo IP rides
919 // along) and only strip the first OPERATION-declared header.
920 // Slicing past `op.header_params.len()` would otherwise risk
921 // dropping the geo header itself; `op_headers` is built as
922 // `op.header_params ++ geo` so index 0 is always operational.
923 let mut h = op_headers.clone();
924 if !h.is_empty() {
925 h.remove(0);
926 }
927 negatives.push(
928 send_case(
929 client,
930 config,
931 method.clone(),
932 &url,
933 "parameters:missing-header",
934 ExpectedOutcome::ClientError,
935 op.sample_body.as_deref(),
936 op.query_params.clone(),
937 h,
938 &path_template,
939 )
940 .await,
941 );
942 }
943
944 // (w) Round 17.5 — OWASP/WAF unification.
945 //
946 // Pull one canonical payload per OWASP category from the existing
947 // `SecurityPayloads` library and emit an injection probe per
948 // category. Targets in priority order: (1) substitute the first
949 // query param's value, (2) substitute the first string field of
950 // the positive JSON body, (3) skip if neither is available.
951 //
952 // Label format `owasp:<category>`, so the existing
953 // `negative_caught` / `negative_missed` rollup groups all OWASP
954 // findings under one `owasp` bucket. Expected 4xx (server should
955 // reject malicious input). A 5xx is a hard finding (server
956 // crashed on the payload); a 2xx is a soft finding (input passed
957 // through unfiltered — may or may not be a real vuln).
958 //
959 // Bounded: at most one probe per category (7 categories total).
960 // Skips the operation entirely if no injection target is
961 // available — open GET endpoints with no params get zero OWASP
962 // probes, no false signal.
963 for probe in build_owasp_probes(op) {
964 negatives.push(
965 send_case(
966 client,
967 config,
968 method.clone(),
969 &url,
970 &probe.label,
971 ExpectedOutcome::ClientError,
972 probe.body.as_deref(),
973 probe.query,
974 // Round 24 (f) — OWASP injection probes must also
975 // carry the geo IP, otherwise a WAF / GEODB rule
976 // tuned to a specific source IP would silently let
977 // them through.
978 op_headers.clone(),
979 &path_template,
980 )
981 .await,
982 );
983 }
984
985 // Round 25 — response-body shape validation pass. For each capture
986 // this op pushed onto the sink, look up the spec's schema for the
987 // actual response status and validate. Result lands in
988 // `response_schema_error` (Some(message) on failure, None on
989 // pass or no-schema-for-this-status). Runs only when the user
990 // opted in AND capture is on (we need the body).
991 if config.validate_response_schemas {
992 if let (Some(sink), Some(start)) = (config.capture.as_ref(), sink_start) {
993 if !op.response_schemas.is_empty() {
994 if let Ok(mut guard) = sink.lock() {
995 let end = guard.len();
996 for i in start..end {
997 let Some(entry) = guard.get_mut(i) else {
998 continue;
999 };
1000 let Some(body) = entry.response_body.as_deref() else {
1001 continue;
1002 };
1003 let Some(schema) = op.response_schemas.get(&entry.response_status) else {
1004 continue;
1005 };
1006 entry.response_schema_error = validate_body_against_schema(body, schema);
1007 }
1008 }
1009 }
1010 }
1011 }
1012
1013 OperationResult {
1014 method: op.method.clone(),
1015 path: op.path.clone(),
1016 positive: Some(positive),
1017 negatives,
1018 }
1019}
1020
1021/// Round 25 — validate a JSON body string against an OpenAPI response
1022/// schema (already converted to a `serde_json::Value`). Returns
1023/// `Some(message)` describing the first violation, or `None` on a
1024/// clean pass / non-JSON body / schema-build failure (in which case
1025/// the absence of an error means "we didn't have anything to compare
1026/// against", not "passed"; the caller-side semantics treat absence as
1027/// success because that's what the user sees as silence).
1028/// Round 27 (k variant b) — return a JSON body string identical to
1029/// `sample` except that the first string-valued leaf has been
1030/// replaced with `snippet`. Walks objects depth-first and stops at
1031/// the first string. Returns `None` when `sample` is not parseable
1032/// JSON or has no string field anywhere; the caller skips emitting
1033/// a probe in that case (Round 34 #829: Srikanth on 0.3.178 found
1034/// that the previous `{"data": <snippet>}` fallback envelope didn't
1035/// match real-API schemas like vCenter's `{enabled: bool}` and the
1036/// server correctly 400'd, which the bench then misreported as a
1037/// `2xx-3xx` expectation miss).
1038fn embed_payload_in_first_string_field(sample: &str, snippet: &str) -> Option<String> {
1039 let mut parsed: serde_json::Value = serde_json::from_str(sample).ok()?;
1040 if !replace_first_string(&mut parsed, snippet) {
1041 return None;
1042 }
1043 serde_json::to_string(&parsed).ok()
1044}
1045
1046/// Helper for `embed_payload_in_first_string_field`: recursively
1047/// walk the value and replace the FIRST string leaf encountered.
1048/// Returns true when a replacement happened. Honors document order
1049/// for objects (BTreeMap-backed `serde_json::Map` iterates in
1050/// insertion order) so the choice of which field to mutate is
1051/// stable across runs.
1052fn replace_first_string(v: &mut serde_json::Value, snippet: &str) -> bool {
1053 match v {
1054 serde_json::Value::String(s) => {
1055 *s = snippet.to_string();
1056 true
1057 }
1058 serde_json::Value::Object(map) => {
1059 for (_k, child) in map.iter_mut() {
1060 if replace_first_string(child, snippet) {
1061 return true;
1062 }
1063 }
1064 false
1065 }
1066 serde_json::Value::Array(arr) => {
1067 for child in arr.iter_mut() {
1068 if replace_first_string(child, snippet) {
1069 return true;
1070 }
1071 }
1072 false
1073 }
1074 _ => false,
1075 }
1076}
1077
1078fn validate_body_against_schema(body: &str, schema: &serde_json::Value) -> Option<String> {
1079 let parsed: serde_json::Value = serde_json::from_str(body).ok()?;
1080 let validator = jsonschema::validator_for(schema).ok()?;
1081 let mut errors = validator.iter_errors(&parsed);
1082 let first = errors.next()?;
1083 // Round 28 — Srikanth on 0.3.170 wanted the message to show the
1084 // actual expected schema alongside the kind label so it reads as
1085 // "expected schema {...} but got <kind>". We emit a compact JSON
1086 // serialisation of the schema as a suffix; the kind label still
1087 // names what went wrong in plain English for quick scanning.
1088 // Round 26 — Srikanth on 0.3.169: the prior `format!("{:?}", first.kind)
1089 // .split('(').next()` produced "Type { kind: Single" (broken Rust
1090 // syntax, mismatched braces). Switch to the human-readable mapping
1091 // already used in executor.rs: handle the common kinds (Type,
1092 // Required, AdditionalProperties, Enum, MinLength, MaxLength,
1093 // Minimum, Maximum, Pattern) explicitly; fall back to the
1094 // jsonschema crate's Display impl on the error (which produces
1095 // something like "{...} is not of type \"string\"") for the long
1096 // tail. Combined with `at <instance-path>` for the field location.
1097 let path = first.instance_path.to_string();
1098 let path = if path.is_empty() { "/" } else { path.as_str() };
1099 // Round 31 — Srikanth on 0.3.174 hit the vCenter case where the
1100 // error is "required field missing: comment" but the printed
1101 // schema was the WHOLE parent object schema (with descriptions of
1102 // every property), not just the missing field's sub-schema. The
1103 // jsonschema crate emits `Required` errors with
1104 // `instance_path == /` (the parent), so the round-30 sub-schema
1105 // walker had no extra info to focus the suffix. Carry the missing
1106 // property name out of the kind match so we can descend one more
1107 // step into `properties[property]` for the printed schema.
1108 let mut required_property: Option<String> = None;
1109 let kind_msg: String = match &first.kind {
1110 jsonschema::error::ValidationErrorKind::Type { kind } => {
1111 // `kind` is `TypeKind::Single(JsonType)` or
1112 // `TypeKind::Multiple(JsonTypeSet)`. `JsonType` has its
1113 // own `Display` impl ("string", "object", etc.).
1114 match kind {
1115 jsonschema::error::TypeKind::Single(t) => format!("expected type {t}"),
1116 jsonschema::error::TypeKind::Multiple(_) => "expected one of multiple types".into(),
1117 }
1118 }
1119 jsonschema::error::ValidationErrorKind::Required { property } => {
1120 // `property.to_string()` returns the Display of the JSON
1121 // value, which for a string is `"name"` (with quotes).
1122 // Strip them for the lookup; keep them in the human message.
1123 let raw = property.to_string();
1124 let unquoted = raw
1125 .strip_prefix('"')
1126 .and_then(|s| s.strip_suffix('"'))
1127 .unwrap_or(&raw)
1128 .to_string();
1129 required_property = Some(unquoted);
1130 format!("required field missing: {property}")
1131 }
1132 jsonschema::error::ValidationErrorKind::AdditionalProperties { unexpected } => {
1133 format!("unexpected additional properties: {unexpected:?}")
1134 }
1135 jsonschema::error::ValidationErrorKind::Enum { options } => {
1136 format!("value not in allowed enum: {options}")
1137 }
1138 jsonschema::error::ValidationErrorKind::MinLength { limit } => {
1139 format!("string shorter than min length ({limit})")
1140 }
1141 jsonschema::error::ValidationErrorKind::MaxLength { limit } => {
1142 format!("string longer than max length ({limit})")
1143 }
1144 jsonschema::error::ValidationErrorKind::Minimum { limit } => {
1145 format!("value below minimum ({limit})")
1146 }
1147 jsonschema::error::ValidationErrorKind::Maximum { limit } => {
1148 format!("value above maximum ({limit})")
1149 }
1150 jsonschema::error::ValidationErrorKind::Pattern { pattern } => {
1151 format!("value did not match pattern {pattern}")
1152 }
1153 // Long tail: lean on jsonschema's Display impl, which is the
1154 // built-in human-readable error message ("X is not of type Y").
1155 // Strip trailing newlines so the JSONL line stays one line.
1156 _ => first.to_string().trim().to_string(),
1157 };
1158 // Round 30 — Srikanth on 0.3.173 asked how a deeper nested mismatch
1159 // reads. The prior output printed the WHOLE top-level schema even for
1160 // a single-field mismatch, which buried the actual constraint that
1161 // failed. Walk the instance pointer through the schema's properties
1162 // chain and print the most specific sub-schema we can find. Falls
1163 // back to the full schema for paths the walker can't resolve
1164 // (additionalProperties, oneOf, allOf, $ref un-resolved, etc.).
1165 let mut focused_schema = sub_schema_at_pointer(schema, path).unwrap_or_else(|| schema.clone());
1166 // Round 31 — for Required errors, descend one more step into
1167 // `properties[<missing>]` so the printed schema is the missing
1168 // field's own constraint, not the whole parent.
1169 if let Some(prop_name) = required_property.as_ref() {
1170 if let Some(prop_schema) =
1171 focused_schema.get("properties").and_then(|p| p.get(prop_name.as_str()))
1172 {
1173 focused_schema = prop_schema.clone();
1174 }
1175 }
1176 // Round 34 (#827) — Srikanth on 0.3.178 hit the vCenter
1177 // `enabled: boolean` case where the schema's multi-paragraph
1178 // `description` (and other prose fields) ate the 300-char budget
1179 // before the actually-useful `type` keyword could appear. Strip
1180 // the noise-fields recursively before serializing so the type
1181 // signal survives truncation; constraint keywords (`type`,
1182 // `properties`, `required`, `format`, `items`, etc.) stay.
1183 let focused_schema = strip_schema_noise(&focused_schema);
1184 let schema_str = serde_json::to_string(&focused_schema).unwrap_or_else(|_| "<schema>".into());
1185 let schema_str = if schema_str.len() > 300 {
1186 format!("{}...", &schema_str[..300])
1187 } else {
1188 schema_str
1189 };
1190 // Round 29 — Srikanth on 0.3.172 was confused by `at /:` thinking
1191 // it referenced the URL path; it's actually a JSON pointer into
1192 // the RESPONSE BODY. Reword so that's unambiguous: explicit
1193 // "response body" prefix and a human label for the root case.
1194 let location = if path == "/" {
1195 "response body root".to_string()
1196 } else {
1197 format!("response body at {path}")
1198 };
1199 Some(format!("{location}: {kind_msg}; expected schema {schema_str}"))
1200}
1201
1202/// Round 34 (#827) — drop the human-readable / documentation-only
1203/// fields from a JSON Schema before printing it inside a
1204/// `response_schema_error` message. The validator only cares about
1205/// constraint keywords (`type`, `required`, `properties`, `items`,
1206/// `format`, `enum`, `min*`/`max*`, `pattern`, `oneOf`/`anyOf`/
1207/// `allOf`/`not`); the prose fields can be paragraphs long for real-
1208/// world specs (vCenter's `enabled: bool` field has a multi-paragraph
1209/// description) and were eating the 300-char truncation budget before
1210/// the actually-useful type info could appear. Stripped fields:
1211/// `description`, `example`, `examples`, `summary`, `title`,
1212/// `externalDocs`, `xml`, `discriminator.description`.
1213fn strip_schema_noise(schema: &serde_json::Value) -> serde_json::Value {
1214 const NOISE_KEYS: &[&str] = &[
1215 "description",
1216 "example",
1217 "examples",
1218 "summary",
1219 "title",
1220 "externalDocs",
1221 "xml",
1222 ];
1223 match schema {
1224 serde_json::Value::Object(map) => {
1225 let mut out = serde_json::Map::with_capacity(map.len());
1226 for (k, v) in map {
1227 if NOISE_KEYS.contains(&k.as_str()) {
1228 continue;
1229 }
1230 out.insert(k.clone(), strip_schema_noise(v));
1231 }
1232 serde_json::Value::Object(out)
1233 }
1234 serde_json::Value::Array(items) => {
1235 serde_json::Value::Array(items.iter().map(strip_schema_noise).collect())
1236 }
1237 other => other.clone(),
1238 }
1239}
1240
1241/// Round 30 — walk a JSON-Pointer-style instance path through a JSON
1242/// Schema and return the sub-schema describing the value at that
1243/// position. For path `/name/age` on
1244/// `{"properties":{"name":{"properties":{"age":{"type":"integer"}}}}}`
1245/// returns `{"type":"integer"}`. Returns `None` for paths the walker
1246/// can't follow (array indices into `items` with no per-index schema,
1247/// `additionalProperties`, `oneOf`/`allOf`, unresolved `$ref`); callers
1248/// should fall back to the full schema in that case.
1249fn sub_schema_at_pointer(schema: &serde_json::Value, pointer: &str) -> Option<serde_json::Value> {
1250 if pointer.is_empty() || pointer == "/" {
1251 return Some(schema.clone());
1252 }
1253 let mut current = schema;
1254 for seg in pointer.trim_start_matches('/').split('/') {
1255 let unescaped = seg.replace("~1", "/").replace("~0", "~");
1256 if let Some(props) = current.get("properties") {
1257 if let Some(sub) = props.get(&unescaped) {
1258 current = sub;
1259 continue;
1260 }
1261 }
1262 if let Some(items) = current.get("items") {
1263 if items.is_object() {
1264 current = items;
1265 continue;
1266 }
1267 }
1268 return None;
1269 }
1270 Some(current.clone())
1271}
1272
1273/// Round 17.5 — one OWASP injection probe to send.
1274#[derive(Debug, Clone)]
1275struct OwaspProbe {
1276 label: String,
1277 body: Option<String>,
1278 query: Vec<(String, String)>,
1279}
1280
1281/// Build one OWASP probe per `SecurityCategory` for `op`. Targets the
1282/// first query param if any, else the first string field of the
1283/// positive JSON body. Returns empty if neither target is available.
1284fn build_owasp_probes(op: &AnnotatedOperation) -> Vec<OwaspProbe> {
1285 use crate::security_payloads::{SecurityCategory, SecurityPayloads};
1286
1287 let categories = [
1288 SecurityCategory::SqlInjection,
1289 SecurityCategory::Xss,
1290 SecurityCategory::CommandInjection,
1291 SecurityCategory::PathTraversal,
1292 SecurityCategory::Ssti,
1293 SecurityCategory::LdapInjection,
1294 SecurityCategory::Xxe,
1295 ];
1296
1297 // Pick an injection target ONCE per operation; reuse it across
1298 // categories. (A single op gets up to 7 probes — one per category
1299 // — all attacking the same field.)
1300 let injection_target = pick_injection_target(op);
1301 let Some(target) = injection_target else {
1302 return Vec::new();
1303 };
1304
1305 let mut probes = Vec::new();
1306 for cat in categories {
1307 // Take the *first* payload from each category. The
1308 // collection's first entry is the canonical low-risk
1309 // representative; later entries include time-based / blind
1310 // probes that aren't useful as a one-shot rejection test.
1311 let Some(payload) = SecurityPayloads::get_by_category(cat).into_iter().next() else {
1312 continue;
1313 };
1314 let mut query = op.query_params.clone();
1315 let mut body = op.sample_body.clone();
1316 match &target {
1317 InjectionTarget::Query(idx) => {
1318 if let Some(slot) = query.get_mut(*idx) {
1319 slot.1 = payload.payload.clone();
1320 }
1321 }
1322 InjectionTarget::BodyStringField(field) => {
1323 body = inject_into_body_field(body.as_deref(), field, &payload.payload);
1324 }
1325 }
1326 probes.push(OwaspProbe {
1327 label: format!("owasp:{}", cat),
1328 body,
1329 query,
1330 });
1331 }
1332 probes
1333}
1334
1335#[derive(Debug, Clone)]
1336enum InjectionTarget {
1337 Query(usize),
1338 BodyStringField(String),
1339}
1340
1341fn pick_injection_target(op: &AnnotatedOperation) -> Option<InjectionTarget> {
1342 if !op.query_params.is_empty() {
1343 return Some(InjectionTarget::Query(0));
1344 }
1345 let sample = op.sample_body.as_deref()?;
1346 let parsed: serde_json::Value = serde_json::from_str(sample).ok()?;
1347 let obj = parsed.as_object()?;
1348 for (k, v) in obj {
1349 if v.is_string() {
1350 return Some(InjectionTarget::BodyStringField(k.clone()));
1351 }
1352 }
1353 None
1354}
1355
1356/// Replace the value of `field` in a JSON-object body with `payload`.
1357/// Returns the mutated body as a JSON string. Returns `None` if the
1358/// body doesn't parse as a JSON object.
1359fn inject_into_body_field(body: Option<&str>, field: &str, payload: &str) -> Option<String> {
1360 let raw = body?;
1361 let mut parsed: serde_json::Value = serde_json::from_str(raw).ok()?;
1362 let obj = parsed.as_object_mut()?;
1363 obj.insert(field.to_string(), serde_json::json!(payload));
1364 serde_json::to_string(&parsed).ok()
1365}
1366
1367#[allow(clippy::too_many_arguments)]
1368/// Round 17.3 — one synthesised bad credential to send.
1369#[derive(Debug, Clone)]
1370struct SecurityProbe {
1371 /// Self-test label, e.g. `security:bad-bearer`.
1372 label: String,
1373 /// Headers to attach to the probe request.
1374 headers: Vec<(String, String)>,
1375 /// Query parameters to attach (API key in query case).
1376 query: Vec<(String, String)>,
1377}
1378
1379/// For each declared security scheme, produce one bad-credential
1380/// probe plus a single "no auth at all" probe that exercises the
1381/// missing-credential code path. Deduplicates by scheme kind so an
1382/// operation declaring `[bearer, bearer]` only yields one Bearer
1383/// probe.
1384fn build_security_probes(schemes: &[SecuritySchemeInfo]) -> Vec<SecurityProbe> {
1385 if schemes.is_empty() {
1386 return Vec::new();
1387 }
1388 let mut probes: Vec<SecurityProbe> = Vec::new();
1389 let mut seen_bearer = false;
1390 let mut seen_basic = false;
1391 // `(loc_tag, name)` — ApiKeyLocation doesn't implement Ord, so
1392 // we tag it with a short discriminant string for dedup.
1393 let mut seen_apikey: std::collections::BTreeSet<(&'static str, String)> = Default::default();
1394 for s in schemes {
1395 match s {
1396 SecuritySchemeInfo::Bearer if !seen_bearer => {
1397 seen_bearer = true;
1398 probes.push(SecurityProbe {
1399 label: "security:bad-bearer".into(),
1400 headers: vec![(
1401 "Authorization".into(),
1402 "Bearer self-test-invalid-token".into(),
1403 )],
1404 query: Vec::new(),
1405 });
1406 }
1407 SecuritySchemeInfo::Basic if !seen_basic => {
1408 seen_basic = true;
1409 // base64("self-test:invalid") — valid base64, wrong creds.
1410 probes.push(SecurityProbe {
1411 label: "security:bad-basic".into(),
1412 headers: vec![(
1413 "Authorization".into(),
1414 "Basic c2VsZi10ZXN0OmludmFsaWQ=".into(),
1415 )],
1416 query: Vec::new(),
1417 });
1418 }
1419 SecuritySchemeInfo::ApiKey { location, name } => {
1420 let loc_tag = match location {
1421 ApiKeyLocation::Header => "header",
1422 ApiKeyLocation::Query => "query",
1423 ApiKeyLocation::Cookie => "cookie",
1424 };
1425 if seen_apikey.contains(&(loc_tag, name.clone())) {
1426 continue;
1427 }
1428 seen_apikey.insert((loc_tag, name.clone()));
1429 let label = format!("security:bad-apikey:{}", name);
1430 let bad = "self-test-invalid-key".to_string();
1431 match location {
1432 ApiKeyLocation::Header => probes.push(SecurityProbe {
1433 label,
1434 headers: vec![(name.clone(), bad)],
1435 query: Vec::new(),
1436 }),
1437 ApiKeyLocation::Query => probes.push(SecurityProbe {
1438 label,
1439 headers: Vec::new(),
1440 query: vec![(name.clone(), bad)],
1441 }),
1442 ApiKeyLocation::Cookie => probes.push(SecurityProbe {
1443 label,
1444 headers: vec![("Cookie".into(), format!("{}={}", name, bad))],
1445 query: Vec::new(),
1446 }),
1447 }
1448 }
1449 _ => {}
1450 }
1451 }
1452 // Always add a "no auth at all" probe when *any* security scheme
1453 // is declared — useful even if all schemes failed to resolve to a
1454 // testable kind, because it surfaces validators that aren't
1455 // checking auth presence at all.
1456 probes.push(SecurityProbe {
1457 label: "security:no-auth".into(),
1458 headers: Vec::new(),
1459 query: Vec::new(),
1460 });
1461 probes
1462}
1463
1464/// Remove Authorization and any API-key headers declared by the
1465/// operation's security schemes from `headers`, so a security probe
1466/// can supply its own credential (or none) cleanly.
1467fn strip_auth(
1468 headers: &[(String, String)],
1469 schemes: &[SecuritySchemeInfo],
1470) -> Vec<(String, String)> {
1471 let mut apikey_headers: std::collections::BTreeSet<String> = Default::default();
1472 for s in schemes {
1473 if let SecuritySchemeInfo::ApiKey {
1474 location: ApiKeyLocation::Header,
1475 name,
1476 } = s
1477 {
1478 apikey_headers.insert(name.to_lowercase());
1479 }
1480 if let SecuritySchemeInfo::ApiKey {
1481 location: ApiKeyLocation::Cookie,
1482 ..
1483 } = s
1484 {
1485 apikey_headers.insert("cookie".into());
1486 }
1487 }
1488 headers
1489 .iter()
1490 .filter(|(k, _)| {
1491 let lk = k.to_lowercase();
1492 lk != "authorization" && !apikey_headers.contains(&lk)
1493 })
1494 .cloned()
1495 .collect()
1496}
1497
1498/// Remove API-key query parameters declared by the operation's
1499/// security schemes from `query`, so a probe can supply its own.
1500fn strip_auth_query(
1501 query: &[(String, String)],
1502 schemes: &[SecuritySchemeInfo],
1503) -> Vec<(String, String)> {
1504 let mut apikey_query: std::collections::BTreeSet<String> = Default::default();
1505 for s in schemes {
1506 if let SecuritySchemeInfo::ApiKey {
1507 location: ApiKeyLocation::Query,
1508 name,
1509 } = s
1510 {
1511 apikey_query.insert(name.clone());
1512 }
1513 }
1514 query.iter().filter(|(k, _)| !apikey_query.contains(k)).cloned().collect()
1515}
1516
1517/// Round 35 (#859) — Srikanth on 0.3.179: embedded-content variant-b
1518/// probes were flagging well-behaved 4xx responses as mismatches when
1519/// in reality only a 5xx (server CRASHED trying to parse the embedded
1520/// XML/YAML/multipart/urlencoded payload) is the bug the probe was
1521/// designed to find. Tristate replaces the older `expected_4xx: bool`
1522/// so variant-b probes can opt into "anything but 5xx is fine".
1523#[derive(Debug, Clone, Copy, PartialEq, Eq)]
1524pub(crate) enum ExpectedOutcome {
1525 /// Positive probe: spec-compliant request, expect 2xx or 3xx.
1526 Success,
1527 /// Negative probe: invalid request, expect 4xx.
1528 ClientError,
1529 /// Embedded-content variant-b probe: spec-shape envelope with a
1530 /// non-JSON payload embedded in the first string field. Any
1531 /// response that isn't a 5xx is fine; the probe is here to catch
1532 /// server crashes on the embedded payload.
1533 NotServerError,
1534}
1535
1536impl ExpectedOutcome {
1537 /// Whether `actual_status` counts as a pass for this outcome.
1538 fn passes(self, actual_status: u16) -> bool {
1539 match self {
1540 ExpectedOutcome::Success => (200..400).contains(&actual_status),
1541 ExpectedOutcome::ClientError => (400..500).contains(&actual_status),
1542 ExpectedOutcome::NotServerError => {
1543 actual_status >= 200 && !(500..600).contains(&actual_status)
1544 }
1545 }
1546 }
1547
1548 /// Human-readable hint persisted in the JSONL capture + HTML
1549 /// viewer's "show mismatches only" filter; also what users `jq`
1550 /// against.
1551 fn as_str(self) -> &'static str {
1552 match self {
1553 ExpectedOutcome::Success => "2xx-3xx",
1554 ExpectedOutcome::ClientError => "4xx",
1555 ExpectedOutcome::NotServerError => "2xx-4xx",
1556 }
1557 }
1558}
1559
1560/// Variant of `send_case` that takes an explicit `extra_headers`
1561/// (rather than reading them from `config`). Used by security probes
1562/// to substitute or strip the configured Authorization header.
1563#[allow(clippy::too_many_arguments)]
1564async fn send_case_with_extra(
1565 client: &Client,
1566 config: &SelfTestConfig,
1567 method: Method,
1568 url: &str,
1569 label: &str,
1570 expected: ExpectedOutcome,
1571 body: Option<&str>,
1572 query: Vec<(String, String)>,
1573 headers: Vec<(String, String)>,
1574 extra_headers: Vec<(String, String)>,
1575 // Round 33 (#823) — spec path template (e.g. `/users/{id}`)
1576 // for the operation this probe belongs to. Stamped on the
1577 // capture so the per-endpoint summary can group by template.
1578 path_template: &str,
1579) -> CaseOutcome {
1580 let mut req = client.request(method.clone(), url);
1581 let mut capture_headers: BTreeMap<String, String> = BTreeMap::new();
1582 for (k, v) in &query {
1583 req = req.query(&[(k.as_str(), v.as_str())]);
1584 }
1585 // Round 36 (#876) — stamp the client side first so the same
1586 // `client_sent_at` string flows into both the request headers
1587 // (so the server-side `ServerConformanceViolation` records it
1588 // verbatim) and the on-disk `CaseCapture` JSONL line. Don't
1589 // re-call `Utc::now()` after `req.send()` — that would record
1590 // a different timestamp than the server sees.
1591 let mockforge_version = env!("CARGO_PKG_VERSION").to_string();
1592 let client_sent_at = chrono::Utc::now().to_rfc3339();
1593 // Round 28 — reqwest's `.header(k, v)` APPENDS rather than replaces
1594 // (.headers().insert() would replace but isn't on the builder).
1595 // The previous round-25 fix relied on "last-write-wins" semantics
1596 // that don't exist; for content-type-swap probes the request went
1597 // out with BOTH `Content-Type: application/json` AND `Content-Type:
1598 // application/xml`, and axum's `Json<>` extractor picked the JSON
1599 // one and accepted, so the server-side validator never saw the
1600 // mismatch. Build a `HeaderMap` ourselves so the override
1601 // replaces the body-block default exactly once.
1602 let mut final_headers: reqwest::header::HeaderMap = reqwest::header::HeaderMap::new();
1603 if let Some(_b) = body {
1604 if let Ok(v) = reqwest::header::HeaderValue::from_str("application/json") {
1605 final_headers.insert(reqwest::header::CONTENT_TYPE, v);
1606 }
1607 capture_headers.insert("Content-Type".to_string(), "application/json".to_string());
1608 }
1609 for (k, v) in &headers {
1610 if let (Ok(hn), Ok(hv)) = (
1611 reqwest::header::HeaderName::from_bytes(k.as_bytes()),
1612 reqwest::header::HeaderValue::from_str(v),
1613 ) {
1614 final_headers.insert(hn, hv);
1615 }
1616 capture_headers.insert(k.clone(), v.clone());
1617 }
1618 for (k, v) in &extra_headers {
1619 if let (Ok(hn), Ok(hv)) = (
1620 reqwest::header::HeaderName::from_bytes(k.as_bytes()),
1621 reqwest::header::HeaderValue::from_str(v),
1622 ) {
1623 final_headers.insert(hn, hv);
1624 }
1625 capture_headers.insert(k.clone(), v.clone());
1626 }
1627 // Round 36 (#876) — outbound client stamps. Inserted last so
1628 // they can't be clobbered by user-supplied extra-headers, and
1629 // recorded in `capture_headers` so the JSONL line shows the
1630 // exact bytes that went on the wire.
1631 {
1632 let v_header = mockforge_foundation::conformance_violations::CLIENT_VERSION_HEADER;
1633 let s_header = mockforge_foundation::conformance_violations::CLIENT_SENT_AT_HEADER;
1634 if let (Ok(hn), Ok(hv)) = (
1635 reqwest::header::HeaderName::from_bytes(v_header.as_bytes()),
1636 reqwest::header::HeaderValue::from_str(&mockforge_version),
1637 ) {
1638 final_headers.insert(hn, hv);
1639 }
1640 if let (Ok(hn), Ok(hv)) = (
1641 reqwest::header::HeaderName::from_bytes(s_header.as_bytes()),
1642 reqwest::header::HeaderValue::from_str(&client_sent_at),
1643 ) {
1644 final_headers.insert(hn, hv);
1645 }
1646 capture_headers.insert(v_header.to_string(), mockforge_version.clone());
1647 capture_headers.insert(s_header.to_string(), client_sent_at.clone());
1648 }
1649 if let Some(b) = body {
1650 req = req.body(b.to_string());
1651 }
1652 req = req.headers(final_headers);
1653 let (actual_status, response_capture) = match req.send().await {
1654 Ok(resp) => {
1655 let status = resp.status().as_u16();
1656 if let Some(sink) = &config.capture {
1657 let resp_headers: BTreeMap<String, String> = resp
1658 .headers()
1659 .iter()
1660 .map(|(k, v)| (k.as_str().to_string(), v.to_str().unwrap_or("").to_string()))
1661 .collect();
1662 let text = resp.text().await.unwrap_or_default();
1663 let (rb, truncated) = truncate_body_for_capture(&text);
1664 (status, Some((Some((rb, truncated)), resp_headers, None, sink.clone())))
1665 } else {
1666 (status, None)
1667 }
1668 }
1669 Err(e) => {
1670 let err_str = e.to_string();
1671 // Round 47 (#79) — classify + push to the wire-level
1672 // network-events sink (when present) so the user has a
1673 // grep-able log of connect/timeout/tls failures during
1674 // self-test, matching the r46 native-executor behaviour.
1675 if let Some(sink) = &config.network_events {
1676 let kind = if e.is_connect() {
1677 "connect"
1678 } else if e.is_timeout() {
1679 "timeout"
1680 } else if e.is_request() {
1681 "request"
1682 } else if e.is_body() {
1683 "body"
1684 } else if e.is_decode() {
1685 "decode"
1686 } else if err_str.to_ascii_lowercase().contains("tls") {
1687 "tls"
1688 } else {
1689 "other"
1690 };
1691 if let Ok(mut guard) = sink.lock() {
1692 guard.push(NetworkEvent {
1693 timestamp: chrono::Utc::now(),
1694 check: label.to_string(),
1695 method: method.to_string(),
1696 url: build_query_url(url, &query),
1697 kind: kind.to_string(),
1698 message: err_str.clone(),
1699 });
1700 }
1701 }
1702 if let Some(sink) = &config.capture {
1703 (0, Some((None, BTreeMap::new(), Some(err_str), sink.clone())))
1704 } else {
1705 (0, None)
1706 }
1707 }
1708 };
1709 let passed = expected.passes(actual_status);
1710 if let Some((resp_body, resp_headers, error, sink)) = response_capture {
1711 let (request_body, request_body_truncated) = match body {
1712 Some(b) => {
1713 let (rb, t) = truncate_body_for_capture(b);
1714 (Some(rb), t)
1715 }
1716 None => (None, false),
1717 };
1718 let (response_body, response_body_truncated) = match resp_body {
1719 Some((rb, t)) => (Some(rb), t),
1720 None => (None, false),
1721 };
1722 let entry = CaseCapture {
1723 label: label.to_string(),
1724 method: method.to_string(),
1725 url: build_query_url(url, &query),
1726 request_headers: capture_headers,
1727 request_body,
1728 request_body_truncated,
1729 response_status: actual_status,
1730 response_headers: resp_headers,
1731 response_body,
1732 response_body_truncated,
1733 error,
1734 // Filled in by the per-operation validation pass after
1735 // every probe finishes; the capture itself is unaware of
1736 // the schema map.
1737 response_schema_error: None,
1738 // Round 28 — derive the expected range from the probe's
1739 // outcome shape so the JSONL line and HTML viewer can
1740 // filter mismatches without re-deriving on the read side.
1741 // Round 35 (#859) — add a third value `"2xx-4xx"` for
1742 // embedded-content variant-b probes whose only failure
1743 // mode is a 5xx server crash.
1744 expected_status_range: expected.as_str().to_string(),
1745 // Round 33 (#823) — path_template carries the spec's
1746 // pre-substitution path so the per-endpoint summary can
1747 // collapse `/users/X` and `/users/Y` into one row.
1748 // spec_label is constant per run, read from the config.
1749 path_template: path_template.to_string(),
1750 spec_label: config.spec_label.clone(),
1751 // Round 36 (#876) — same values that went on the wire as
1752 // request headers, so a server-side
1753 // `ServerConformanceViolation` recorded with
1754 // `client_mockforge_version` + `client_sent_at` matches
1755 // the JSONL line byte-for-byte.
1756 mockforge_version: mockforge_version.clone(),
1757 client_sent_at: client_sent_at.clone(),
1758 };
1759 if let Ok(mut guard) = sink.lock() {
1760 guard.push(entry);
1761 }
1762 }
1763 // Round 35 (#859) — keep the `expected_4xx` field on `CaseOutcome`
1764 // semantically tied to "negative probe expecting 400-class", so
1765 // downstream code in `report_html.rs` doesn't have to learn about
1766 // the new tristate. `NotServerError` reports as `expected_4xx:
1767 // false` (it's a positive probe in spirit) and instead carries
1768 // its outcome through the per-capture `expected_status_range`.
1769 let expected_4xx = matches!(expected, ExpectedOutcome::ClientError);
1770 CaseOutcome {
1771 label: label.to_string(),
1772 expected_4xx,
1773 actual_status,
1774 passed,
1775 }
1776}
1777
1778// HTTP request shape needs all of these: client, config (for capture
1779// sink + extra headers), method, url, label (probe id), expected_4xx
1780// (pass/fail decision), body, query, headers. A struct wrapper would
1781// just move the arity from positional to field access without making
1782// the call sites clearer.
1783#[allow(clippy::too_many_arguments)]
1784async fn send_case(
1785 client: &Client,
1786 config: &SelfTestConfig,
1787 method: Method,
1788 url: &str,
1789 label: &str,
1790 expected: ExpectedOutcome,
1791 body: Option<&str>,
1792 query: Vec<(String, String)>,
1793 headers: Vec<(String, String)>,
1794 path_template: &str,
1795) -> CaseOutcome {
1796 // Forwarding to `send_case_with_extra` keeps the capture logic in
1797 // one place so request/response tracing can't drift between the
1798 // two entrypoints.
1799 send_case_with_extra(
1800 client,
1801 config,
1802 method,
1803 url,
1804 label,
1805 expected,
1806 body,
1807 query,
1808 headers,
1809 config.extra_headers.clone(),
1810 path_template,
1811 )
1812 .await
1813}
1814
1815/// Round 23 (c-iii) — rebuild the query-stringified URL for capture so
1816/// the JSONL trace shows the URL that actually went over the wire
1817/// (reqwest applies `.query(..)` after the request URL string is
1818/// rendered, so capturing the raw `url` argument alone loses the
1819/// query params).
1820fn build_query_url(base: &str, query: &[(String, String)]) -> String {
1821 if query.is_empty() {
1822 return base.to_string();
1823 }
1824 let qs: String = query
1825 .iter()
1826 .map(|(k, v)| format!("{}={}", urlencoding::encode(k), urlencoding::encode(v)))
1827 .collect::<Vec<_>>()
1828 .join("&");
1829 if base.contains('?') {
1830 format!("{base}&{qs}")
1831 } else {
1832 format!("{base}?{qs}")
1833 }
1834}
1835
1836/// Substitute `{param}` placeholders in the spec path with their
1837/// sample values from `path_params`, then prepend `target_url`. Empty
1838/// values are kept as `{param}` so an upstream router still matches
1839/// the template — useful when `path_params` is empty and we want to
1840/// hit the same route the spec defines.
1841///
1842/// All current call sites went through `build_url_with_base` after
1843/// round 18.1, so this no-base-path helper is unused; keep it as the
1844/// documented shim for future external callers (one-arg simplification).
1845#[allow(dead_code)]
1846fn build_url(target: &str, path_template: &str, path_params: &[(String, String)]) -> String {
1847 build_url_with_base(target, None, path_template, path_params)
1848}
1849
1850/// Round 18.1 — variant of `build_url` that takes a `base_path`
1851/// (e.g. `Some("/api")`). When set, prepends it to the spec path so a
1852/// spec declaring `/users` against a target served behind `/api`
1853/// resolves to `<target>/api/users`. `base_path` is normalised: leading
1854/// `/` is auto-added, trailing `/` is stripped.
1855fn build_url_with_base(
1856 target: &str,
1857 base_path: Option<&str>,
1858 path_template: &str,
1859 path_params: &[(String, String)],
1860) -> String {
1861 let mut url = path_template.to_string();
1862 for (name, value) in path_params {
1863 let placeholder = format!("{{{}}}", name);
1864 if !value.is_empty() {
1865 url = url.replace(&placeholder, value);
1866 }
1867 }
1868 let target = target.trim_end_matches('/');
1869 let prefix = match base_path {
1870 Some(bp) if !bp.is_empty() => {
1871 let trimmed = bp.trim_end_matches('/');
1872 if trimmed.starts_with('/') {
1873 trimmed.to_string()
1874 } else {
1875 format!("/{}", trimmed)
1876 }
1877 }
1878 _ => String::new(),
1879 };
1880 let path = if url.starts_with('/') {
1881 url
1882 } else {
1883 format!("/{url}")
1884 };
1885 format!("{target}{prefix}{path}")
1886}
1887
1888#[cfg(test)]
1889mod tests {
1890 use super::*;
1891
1892 fn op(
1893 method: &str,
1894 path: &str,
1895 body: Option<&str>,
1896 query: Vec<(&str, &str)>,
1897 headers: Vec<(&str, &str)>,
1898 path_params: Vec<(&str, &str)>,
1899 ) -> AnnotatedOperation {
1900 AnnotatedOperation {
1901 method: method.into(),
1902 path: path.into(),
1903 features: Vec::new(),
1904 request_body_content_type: body.map(|_| "application/json".into()),
1905 sample_body: body.map(|s| s.to_string()),
1906 query_params: query.into_iter().map(|(a, b)| (a.into(), b.into())).collect(),
1907 header_params: headers.into_iter().map(|(a, b)| (a.into(), b.into())).collect(),
1908 path_params: path_params.into_iter().map(|(a, b)| (a.into(), b.into())).collect(),
1909 response_schema: None,
1910 response_schemas: std::collections::BTreeMap::new(),
1911 request_body_schema: None,
1912 security_schemes: Vec::new(),
1913 }
1914 }
1915
1916 /// Round 36 (#876) — older JSONL lines (written before the stamp
1917 /// fields existed) must still deserialise without error and
1918 /// default to empty strings. Prevents a back-compat regression
1919 /// the next time we extend `CaseCapture`.
1920 #[test]
1921 fn case_capture_back_compat_when_stamp_fields_missing() {
1922 let pre_r36 = serde_json::json!({
1923 "label": "positive",
1924 "method": "GET",
1925 "url": "http://api/users",
1926 "request_headers": {},
1927 "request_body_truncated": false,
1928 "response_status": 200,
1929 "response_headers": {},
1930 "response_body_truncated": false,
1931 });
1932 let capture: CaseCapture =
1933 serde_json::from_value(pre_r36).expect("pre-r36 payload must deserialise");
1934 assert!(capture.mockforge_version.is_empty(), "default to empty");
1935 assert!(capture.client_sent_at.is_empty(), "default to empty");
1936 }
1937
1938 /// Round 36 (#876) — when the bench stamps fields itself (the
1939 /// happy path), they round-trip through serde unchanged. Pins
1940 /// the on-wire shape so tooling that grep's `mockforge_version`
1941 /// out of the JSONL stays valid.
1942 #[test]
1943 fn case_capture_stamps_round_trip_through_serde() {
1944 let stamped = CaseCapture {
1945 label: "positive".into(),
1946 method: "GET".into(),
1947 url: "http://api/users".into(),
1948 request_headers: BTreeMap::new(),
1949 request_body: None,
1950 request_body_truncated: false,
1951 response_status: 200,
1952 response_headers: BTreeMap::new(),
1953 response_body: None,
1954 response_body_truncated: false,
1955 error: None,
1956 response_schema_error: None,
1957 expected_status_range: "2xx-3xx".into(),
1958 path_template: "/users".into(),
1959 spec_label: None,
1960 mockforge_version: "0.3.183".into(),
1961 client_sent_at: "2026-06-17T12:34:56+00:00".into(),
1962 };
1963 let json = serde_json::to_string(&stamped).unwrap();
1964 assert!(json.contains("\"mockforge_version\":\"0.3.183\""));
1965 assert!(json.contains("\"client_sent_at\":\"2026-06-17T12:34:56+00:00\""));
1966 let back: CaseCapture = serde_json::from_str(&json).unwrap();
1967 assert_eq!(back.mockforge_version, "0.3.183");
1968 assert_eq!(back.client_sent_at, "2026-06-17T12:34:56+00:00");
1969 }
1970
1971 #[test]
1972 fn build_url_substitutes_path_params() {
1973 let url = build_url(
1974 "https://api.test/",
1975 "/users/{id}/posts/{pid}",
1976 &[("id".into(), "42".into()), ("pid".into(), "7".into())],
1977 );
1978 assert_eq!(url, "https://api.test/users/42/posts/7");
1979 }
1980
1981 /// Round 18.1 — a run where every positive 404s should be flagged
1982 /// as a likely target misconfiguration, not silently treated as a
1983 /// successful conformance run.
1984 #[test]
1985 fn detect_target_misconfiguration_when_all_positives_share_status() {
1986 let mut report = SelfTestReport {
1987 positive_pass: 0,
1988 positive_fail: 50,
1989 ..Default::default()
1990 };
1991 for i in 0..50 {
1992 report.operations.push(OperationResult {
1993 method: "GET".into(),
1994 path: format!("/r/{i}"),
1995 positive: Some(CaseOutcome {
1996 label: "positive".into(),
1997 expected_4xx: false,
1998 actual_status: 404,
1999 passed: false,
2000 }),
2001 negatives: Vec::new(),
2002 });
2003 }
2004 assert_eq!(report.detect_target_misconfiguration(), Some(404));
2005 }
2006
2007 #[test]
2008 fn detect_target_misconfiguration_returns_none_when_some_pass() {
2009 let mut report = SelfTestReport {
2010 positive_pass: 5,
2011 positive_fail: 50,
2012 ..Default::default()
2013 };
2014 for i in 0..55 {
2015 report.operations.push(OperationResult {
2016 method: "GET".into(),
2017 path: format!("/r/{i}"),
2018 positive: Some(CaseOutcome {
2019 label: "positive".into(),
2020 expected_4xx: false,
2021 actual_status: if i < 5 { 200 } else { 404 },
2022 passed: i < 5,
2023 }),
2024 negatives: Vec::new(),
2025 });
2026 }
2027 assert_eq!(report.detect_target_misconfiguration(), None);
2028 }
2029
2030 /// Round 18.1 — `--base-path /api` should prepend `/api` to
2031 /// every spec path. Pre-fix, the self-test ignored base_path and
2032 /// 404'd every positive when the deployed API was behind a path
2033 /// prefix.
2034 #[test]
2035 fn build_url_applies_base_path_when_present() {
2036 let url = build_url_with_base(
2037 "https://api.example.com",
2038 Some("/api"),
2039 "/users/{id}",
2040 &[("id".into(), "42".into())],
2041 );
2042 assert_eq!(url, "https://api.example.com/api/users/42");
2043 }
2044
2045 /// Round 18.1 — base_path is normalised: missing leading slash
2046 /// gets one added, trailing slash is stripped, empty string is
2047 /// the same as None.
2048 #[test]
2049 fn build_url_normalises_base_path() {
2050 let no_slash = build_url_with_base("https://t", Some("api"), "/x", &[]);
2051 assert_eq!(no_slash, "https://t/api/x");
2052 let trailing = build_url_with_base("https://t", Some("/api/"), "/x", &[]);
2053 assert_eq!(trailing, "https://t/api/x");
2054 let empty = build_url_with_base("https://t", Some(""), "/x", &[]);
2055 assert_eq!(empty, "https://t/x");
2056 let none = build_url_with_base("https://t", None, "/x", &[]);
2057 assert_eq!(none, "https://t/x");
2058 }
2059
2060 #[test]
2061 fn build_url_keeps_placeholders_when_no_sample() {
2062 let url = build_url("https://api.test", "/users/{id}", &[]);
2063 assert_eq!(url, "https://api.test/users/{id}");
2064 }
2065
2066 #[test]
2067 fn report_summary_calls_out_misses() {
2068 let r = SelfTestReport {
2069 positive_pass: 3,
2070 positive_fail: 0,
2071 negative_caught: BTreeMap::from([("request-body".into(), 2)]),
2072 negative_missed: BTreeMap::from([("request-body".into(), 1)]),
2073 operations: Vec::new(),
2074 };
2075 let summary = r.render_summary();
2076 assert!(summary.contains("Positives: 3 pass / 0 fail"));
2077 assert!(summary.contains("Negatives [request-body]: 2 caught / 1 missed"));
2078 assert!(summary.contains("⚠"));
2079 assert!(!r.all_passed());
2080 }
2081
2082 #[test]
2083 fn report_all_passed_when_no_miss() {
2084 let r = SelfTestReport {
2085 positive_pass: 5,
2086 positive_fail: 0,
2087 negative_caught: BTreeMap::from([("parameters".into(), 3)]),
2088 negative_missed: BTreeMap::new(),
2089 operations: Vec::new(),
2090 };
2091 assert!(r.all_passed());
2092 assert!(r.render_summary().contains("✓"));
2093 }
2094
2095 #[tokio::test]
2096 async fn run_self_test_against_unreachable_target_marks_all_failed() {
2097 // Use an obviously-dead port so we exercise the timeout/error
2098 // path without needing a live server in tests.
2099 let cfg = SelfTestConfig {
2100 target_url: "http://127.0.0.1:1".into(),
2101 timeout: Duration::from_millis(200),
2102 ..Default::default()
2103 };
2104 let ops = vec![op(
2105 "POST",
2106 "/users",
2107 Some("{\"name\":\"a\"}"),
2108 vec![],
2109 vec![],
2110 vec![],
2111 )];
2112 let report = run_self_test(&ops, &cfg).await.expect("client builds");
2113 // All cases hit the connect-error path → actual_status=0.
2114 // Positive expects 2xx-3xx → 0 is fail. Negatives expect 4xx
2115 // → 0 is also fail (we missed catching).
2116 assert_eq!(report.positive_fail, 1);
2117 assert!(report.negative_missed.values().sum::<usize>() >= 1);
2118 assert!(!report.all_passed());
2119 }
2120
2121 /// Round 17.2 — operations with both a positive sample AND a
2122 /// resolved request-body schema produce schema-driven negatives
2123 /// in addition to the spec-agnostic empty/wrong-type ones. The
2124 /// labels carry the field path so a per-category report can tell
2125 /// you exactly which field caught.
2126 #[tokio::test]
2127 async fn schema_driven_negatives_fire_when_schema_present() {
2128 use openapiv3::{ObjectType, ReferenceOr, Schema, SchemaData, SchemaKind, Type};
2129 let cfg = SelfTestConfig {
2130 target_url: "http://127.0.0.1:1".into(),
2131 timeout: Duration::from_millis(200),
2132 ..Default::default()
2133 };
2134 // Build an operation whose schema has a required `name` string
2135 // and an `age` integer. The mutator should produce, at
2136 // minimum: required-removed:name, required-removed:age,
2137 // type-mismatch:name, type-mismatch:age, integer-as-float:age,
2138 // plus the root-level type-mismatch.
2139 let mut obj = ObjectType::default();
2140 obj.properties.insert(
2141 "name".to_string(),
2142 ReferenceOr::Item(Box::new(Schema {
2143 schema_data: SchemaData::default(),
2144 schema_kind: SchemaKind::Type(Type::String(Default::default())),
2145 })),
2146 );
2147 obj.properties.insert(
2148 "age".to_string(),
2149 ReferenceOr::Item(Box::new(Schema {
2150 schema_data: SchemaData::default(),
2151 schema_kind: SchemaKind::Type(Type::Integer(Default::default())),
2152 })),
2153 );
2154 obj.required = vec!["name".into(), "age".into()];
2155 let schema = Schema {
2156 schema_data: SchemaData::default(),
2157 schema_kind: SchemaKind::Type(Type::Object(obj)),
2158 };
2159
2160 let mut o =
2161 op("POST", "/users", Some(r#"{"name":"Ada","age":30}"#), vec![], vec![], vec![]);
2162 o.request_body_schema = Some(schema);
2163 let report = run_self_test(&[o], &cfg).await.expect("client builds");
2164 // Bucket labels from the operation result.
2165 let labels: std::collections::BTreeSet<String> = report
2166 .operations
2167 .iter()
2168 .flat_map(|op| op.negatives.iter().map(|n| n.label.clone()))
2169 .collect();
2170 assert!(
2171 labels.iter().any(|l| l.starts_with("request-body:type-mismatch:")),
2172 "missing type-mismatch negative; got {labels:?}"
2173 );
2174 assert!(
2175 labels.iter().any(|l| l.starts_with("request-body:required-removed:")),
2176 "missing required-removed negative; got {labels:?}"
2177 );
2178 assert!(
2179 labels.iter().any(|l| l == "parameters:uri-too-long"),
2180 "missing URI-length negative; got {labels:?}"
2181 );
2182 }
2183
2184 /// Round 16 — operations with a body OR a path-param now produce
2185 /// negatives even without a sample body. Previously a POST whose
2186 /// body annotator failed produced *zero* negatives, so the self-test
2187 /// always reported "all passing" for that endpoint.
2188 #[tokio::test]
2189 async fn no_sample_body_still_produces_request_body_negatives() {
2190 let cfg = SelfTestConfig {
2191 target_url: "http://127.0.0.1:1".into(),
2192 timeout: Duration::from_millis(200),
2193 ..Default::default()
2194 };
2195 // POST with a body content type but no sample (annotator gap).
2196 let ops = vec![op("POST", "/x", None, vec![], vec![], vec![])];
2197 // No sample_body but request_body_content_type set:
2198 let mut ops_fixed = ops;
2199 ops_fixed[0].request_body_content_type = Some("application/json".into());
2200 let report = run_self_test(&ops_fixed, &cfg).await.expect("client builds");
2201 // Both request-body negatives (empty + wrong-type) should fire,
2202 // landing in `negative_missed` because the unreachable target
2203 // returns no 4xx. The point: count > 0.
2204 assert!(
2205 report.negative_missed.values().sum::<usize>() >= 2,
2206 "expected ≥2 request-body negatives, got {:?}",
2207 report.negative_missed
2208 );
2209 }
2210
2211 /// Round 16 — operations with a path-param now get a probe even
2212 /// when there's no body / required query / required header.
2213 /// Previously `/teams/{team-id}` with no other required fields
2214 /// produced zero negatives → always "all passing".
2215 #[tokio::test]
2216 async fn path_param_only_endpoint_produces_a_probe() {
2217 let cfg = SelfTestConfig {
2218 target_url: "http://127.0.0.1:1".into(),
2219 timeout: Duration::from_millis(200),
2220 ..Default::default()
2221 };
2222 let ops = vec![op(
2223 "GET",
2224 "/teams/{team-id}",
2225 None,
2226 vec![],
2227 vec![],
2228 vec![("team-id", "1")],
2229 )];
2230 let report = run_self_test(&ops, &cfg).await.expect("client builds");
2231 let total: usize = report.negative_caught.values().sum::<usize>()
2232 + report.negative_missed.values().sum::<usize>();
2233 assert!(total >= 1, "expected ≥1 path-param probe, got {:?}", report);
2234 }
2235
2236 /// Round 18.5 — when `geo_ip` is set, every default forwarded-
2237 /// IP header gets the IP appended (X-Forwarded-For,
2238 /// True-Client-IP, CF-Connecting-IP).
2239 #[test]
2240 fn effective_op_headers_appends_geo_ip_to_default_headers() {
2241 let ip: IpAddr = "203.0.113.42".parse().unwrap();
2242 let headers = effective_op_headers(
2243 &[("Accept".into(), "application/json".into())],
2244 Some(ip),
2245 &default_geo_source_headers(),
2246 );
2247 let names: Vec<&str> = headers.iter().map(|(k, _)| k.as_str()).collect();
2248 assert!(names.contains(&"Accept"));
2249 assert!(names.contains(&"X-Forwarded-For"));
2250 assert!(names.contains(&"True-Client-IP"));
2251 assert!(names.contains(&"CF-Connecting-IP"));
2252 // Every geo header carries the same IP value.
2253 let geo_values: Vec<&str> =
2254 headers.iter().filter(|(k, _)| k != "Accept").map(|(_, v)| v.as_str()).collect();
2255 for v in geo_values {
2256 assert_eq!(v, "203.0.113.42");
2257 }
2258 }
2259
2260 /// Round 18.5 — operations that already declare a forwarded-IP
2261 /// header (rare but legal — some specs hard-code one) keep their
2262 /// declared value; we don't clobber the spec.
2263 #[test]
2264 fn effective_op_headers_respects_spec_declared_header() {
2265 let ip: IpAddr = "203.0.113.99".parse().unwrap();
2266 let headers = effective_op_headers(
2267 &[("x-forwarded-for".into(), "10.0.0.1".into())],
2268 Some(ip),
2269 &["X-Forwarded-For".to_string()],
2270 );
2271 // The spec's lower-case value wins; we shouldn't add a
2272 // second X-Forwarded-For row that overrides it.
2273 let xff: Vec<&str> = headers
2274 .iter()
2275 .filter(|(k, _)| k.eq_ignore_ascii_case("x-forwarded-for"))
2276 .map(|(_, v)| v.as_str())
2277 .collect();
2278 assert_eq!(xff, vec!["10.0.0.1"]);
2279 }
2280
2281 /// Round 18.5 — None geo_ip and/or empty header list is a no-op.
2282 #[test]
2283 fn effective_op_headers_is_a_noop_without_geo_ip() {
2284 let base = vec![("Accept".into(), "json".into())];
2285 let h1 = effective_op_headers(&base, None, &default_geo_source_headers());
2286 assert_eq!(h1, base);
2287 let ip: IpAddr = "10.0.0.1".parse().unwrap();
2288 let h2 = effective_op_headers(&base, Some(ip), &[]);
2289 assert_eq!(h2, base);
2290 }
2291
2292 /// Round 18.5 — empty `source_ips` builds a single default
2293 /// client; a non-empty list builds N clients each attempting to
2294 /// bind. We can't reliably test the actual bind on CI (no
2295 /// loopback aliases), but a loopback IP is always bind-able.
2296 #[test]
2297 fn build_client_pool_one_per_source_ip() {
2298 let mut cfg = SelfTestConfig {
2299 target_url: "http://127.0.0.1:1".into(),
2300 timeout: Duration::from_millis(200),
2301 ..Default::default()
2302 };
2303 // Empty → one default client.
2304 assert_eq!(build_client_pool(&cfg).expect("default builds").len(), 1);
2305 // Non-empty → one per IP. Loopback bind is portable.
2306 cfg.source_ips = vec!["127.0.0.1".parse().unwrap()];
2307 assert_eq!(build_client_pool(&cfg).expect("bind loopback").len(), 1);
2308 }
2309
2310 /// Round 18.5 — geo IPs round-robin across operations. Hits an
2311 /// unreachable target so we can inspect the case outcomes; the
2312 /// point is to confirm `op_headers` carried the geo IP through
2313 /// (CaseOutcome doesn't surface headers directly, so we just
2314 /// verify the run completes without panicking and the result
2315 /// shape is correct when source_ips is non-empty too).
2316 #[tokio::test]
2317 async fn run_self_test_with_geo_source_completes() {
2318 let cfg = SelfTestConfig {
2319 target_url: "http://127.0.0.1:1".into(),
2320 timeout: Duration::from_millis(200),
2321 geo_source_ips: vec![
2322 "203.0.113.1".parse().unwrap(),
2323 "203.0.113.2".parse().unwrap(),
2324 ],
2325 ..Default::default()
2326 };
2327 let ops = vec![
2328 op("GET", "/a", None, vec![], vec![], vec![]),
2329 op("GET", "/b", None, vec![], vec![], vec![]),
2330 op("GET", "/c", None, vec![], vec![], vec![]),
2331 ];
2332 let report = run_self_test(&ops, &cfg).await.expect("client builds");
2333 assert_eq!(report.operations.len(), 3);
2334 }
2335
2336 /// Round 24 (f) — Srikanth saw the geo header on positive probes
2337 /// only; the four negative-probe call sites were passing
2338 /// `op.header_params` directly instead of `op_headers`, so the
2339 /// geo IP got dropped. This test runs a self-test that includes
2340 /// negative probes (uri-too-long, missing-query, etc.) under
2341 /// `--conformance-self-test-capture`, then asserts that EVERY
2342 /// captured probe (positive AND negative) carries one of the
2343 /// configured forwarded-IP headers.
2344 #[tokio::test]
2345 async fn geo_headers_present_on_every_probe_with_capture() {
2346 let sink: Arc<Mutex<Vec<CaseCapture>>> = Arc::new(Mutex::new(Vec::new()));
2347 let cfg = SelfTestConfig {
2348 target_url: "http://127.0.0.1:1".into(),
2349 timeout: Duration::from_millis(50),
2350 geo_source_ips: vec!["203.0.113.5".parse().unwrap()],
2351 capture: Some(sink.clone()),
2352 ..Default::default()
2353 };
2354 // An operation rich enough to trip several negative-probe
2355 // branches: header param (→ missing-header), query param
2356 // (→ missing-query), and a sample body (→ schema mutations
2357 // wouldn't fire without a schema, but uri-too-long always
2358 // does).
2359 let ops = vec![op(
2360 "GET",
2361 "/items",
2362 Some("{}"),
2363 vec![("id", "1")],
2364 vec![("X-Trace", "x")],
2365 vec![],
2366 )];
2367 let _ = run_self_test(&ops, &cfg).await.expect("client builds");
2368 let captures = sink.lock().unwrap();
2369 assert!(!captures.is_empty(), "self-test should record probes");
2370 // For every captured probe, at least one of the default geo
2371 // headers must be present and equal to the configured IP.
2372 let geo_headers: std::collections::HashSet<&str> =
2373 ["X-Forwarded-For", "True-Client-IP", "CF-Connecting-IP"].into_iter().collect();
2374 for c in captures.iter() {
2375 let has_geo = c
2376 .request_headers
2377 .iter()
2378 .any(|(k, v)| geo_headers.contains(k.as_str()) && v == "203.0.113.5");
2379 assert!(
2380 has_geo,
2381 "probe `{}` is missing the geo IP header; got headers: {:?}",
2382 c.label, c.request_headers
2383 );
2384 }
2385 }
2386
2387 /// Round 25 (k) — operations with a JSON request body now get four
2388 /// content-type-swap probes (xml / yaml / multipart / urlencoded).
2389 /// Verify they:
2390 /// 1. fire only when the operation declares a JSON body
2391 /// 2. carry the wrong Content-Type the probe is testing for
2392 /// 3. don't fire on body-less operations
2393 #[tokio::test]
2394 async fn content_type_swap_probes_fire_for_json_bodies() {
2395 let sink: Arc<Mutex<Vec<CaseCapture>>> = Arc::new(Mutex::new(Vec::new()));
2396 let cfg = SelfTestConfig {
2397 target_url: "http://127.0.0.1:1".into(),
2398 timeout: Duration::from_millis(50),
2399 capture: Some(sink.clone()),
2400 ..Default::default()
2401 };
2402 let ops = vec![
2403 op("POST", "/users", Some("{\"name\":\"a\"}"), vec![], vec![], vec![]),
2404 op("GET", "/ping", None, vec![], vec![], vec![]),
2405 ];
2406 let _ = run_self_test(&ops, &cfg).await.expect("client builds");
2407 let captures = sink.lock().unwrap();
2408
2409 let swap_labels: Vec<&str> = captures
2410 .iter()
2411 .filter(|c| c.label.starts_with("request-body:content-type-mismatch:"))
2412 .map(|c| c.label.as_str())
2413 .collect();
2414 assert_eq!(
2415 swap_labels.len(),
2416 4,
2417 "expected 4 content-type-swap probes (one per variant), got: {swap_labels:?}"
2418 );
2419 let expected_labels = [
2420 "request-body:content-type-mismatch:xml",
2421 "request-body:content-type-mismatch:yaml",
2422 "request-body:content-type-mismatch:multipart",
2423 "request-body:content-type-mismatch:urlencoded",
2424 ];
2425 for want in expected_labels {
2426 assert!(swap_labels.contains(&want), "missing swap probe `{want}`");
2427 }
2428
2429 // Each swap probe must carry the wrong Content-Type it's
2430 // testing for — that's the whole point.
2431 for c in captures.iter() {
2432 let Some(suffix) = c.label.strip_prefix("request-body:content-type-mismatch:") else {
2433 continue;
2434 };
2435 let want_ct = match suffix {
2436 "xml" => "application/xml",
2437 "yaml" => "application/yaml",
2438 "multipart" => "multipart/form-data",
2439 "urlencoded" => "application/x-www-form-urlencoded",
2440 _ => continue,
2441 };
2442 let got_ct = c
2443 .request_headers
2444 .iter()
2445 .find(|(k, _)| k.eq_ignore_ascii_case("content-type"))
2446 .map(|(_, v)| v.as_str())
2447 .unwrap_or("");
2448 assert_eq!(got_ct, want_ct, "swap probe `{}` sent wrong CT", c.label);
2449 }
2450
2451 // The body-less operation must NOT produce content-type-swap
2452 // probes (no body → no content type to lie about).
2453 let body_less_swaps = captures
2454 .iter()
2455 .filter(|c| {
2456 c.label.starts_with("request-body:content-type-mismatch:")
2457 && c.url.ends_with("/ping")
2458 })
2459 .count();
2460 assert_eq!(
2461 body_less_swaps, 0,
2462 "GET /ping has no request body; should not produce content-type-swap probes"
2463 );
2464 }
2465
2466 /// Round 27 (k variant b) — Srikanth's round-23 follow-up on (k):
2467 /// JSON envelope with embedded non-JSON field values. For each
2468 /// JSON-body operation, four extra probes fire that send valid
2469 /// JSON with an XML/YAML/multipart/urlencoded snippet stuffed
2470 /// into a string field. Content-Type stays `application/json`;
2471 /// expected is 2xx-3xx (the body parses); a 5xx flags a server
2472 /// that crashed on the embedded content.
2473 #[tokio::test]
2474 async fn embedded_content_probes_fire_with_honest_content_type() {
2475 let sink: Arc<Mutex<Vec<CaseCapture>>> = Arc::new(Mutex::new(Vec::new()));
2476 let cfg = SelfTestConfig {
2477 target_url: "http://127.0.0.1:1".into(),
2478 timeout: Duration::from_millis(50),
2479 capture: Some(sink.clone()),
2480 ..Default::default()
2481 };
2482 let ops = vec![op(
2483 "POST",
2484 "/users",
2485 Some("{\"name\":\"alice\",\"age\":30}"),
2486 vec![],
2487 vec![],
2488 vec![],
2489 )];
2490 let _ = run_self_test(&ops, &cfg).await.expect("client builds");
2491 let captures = sink.lock().unwrap();
2492 let embedded: Vec<&CaseCapture> = captures
2493 .iter()
2494 .filter(|c| c.label.starts_with("request-body:embedded-content:"))
2495 .collect();
2496 assert_eq!(
2497 embedded.len(),
2498 4,
2499 "expected 4 embedded-content probes, got: {:?}",
2500 embedded.iter().map(|c| &c.label).collect::<Vec<_>>()
2501 );
2502 // Every embedded probe must carry the honest application/json
2503 // Content-Type (NOT lie like the variant-a content-type-swap
2504 // probes do) and a request body that still parses as JSON.
2505 for c in &embedded {
2506 let ct = c
2507 .request_headers
2508 .iter()
2509 .find(|(k, _)| k.eq_ignore_ascii_case("content-type"))
2510 .map(|(_, v)| v.as_str())
2511 .unwrap_or("");
2512 assert!(
2513 ct.contains("application/json"),
2514 "embedded probe `{}` should keep Content-Type honest, got {ct}",
2515 c.label
2516 );
2517 let body = c.request_body.as_deref().unwrap_or("");
2518 assert!(
2519 serde_json::from_str::<serde_json::Value>(body).is_ok(),
2520 "embedded probe `{}` body should still be valid JSON, got: {body}",
2521 c.label
2522 );
2523 }
2524 }
2525
2526 /// `embed_payload_in_first_string_field` walks objects depth-first
2527 /// and replaces only the FIRST string-valued leaf, leaving the
2528 /// surrounding structure intact.
2529 #[test]
2530 fn embed_payload_replaces_first_string_only() {
2531 let sample = r#"{"name":"alice","age":30,"tags":["admin","user"]}"#;
2532 let mutated = embed_payload_in_first_string_field(sample, "<x/>")
2533 .expect("string field present so probe constructed");
2534 let v: serde_json::Value = serde_json::from_str(&mutated).unwrap();
2535 assert_eq!(v["name"], serde_json::json!("<x/>"));
2536 // age stays an integer (not stringified by the mutation).
2537 assert_eq!(v["age"], serde_json::json!(30));
2538 // tags array's strings stay untouched (we only replace the
2539 // first encountered string leaf, depth-first).
2540 assert_eq!(v["tags"][0], serde_json::json!("admin"));
2541 assert_eq!(v["tags"][1], serde_json::json!("user"));
2542 }
2543
2544 /// Round 34 (#829) — Srikanth on 0.3.178: when the positive
2545 /// sample has NO string field, the previous `{"data": <snippet>}`
2546 /// fallback produced an envelope that doesn't match real-API
2547 /// schemas (e.g. vCenter's `consolecli` PUT wants
2548 /// `{enabled: bool}`), so the server correctly 400'd and the
2549 /// bench misreported the 2xx-3xx expectation. Now we return None
2550 /// and the caller skips the probe.
2551 #[test]
2552 fn embed_payload_returns_none_when_no_string_field() {
2553 let no_strings = r#"{"a":1,"b":[2,3]}"#;
2554 assert!(embed_payload_in_first_string_field(no_strings, "<x><y></y></x>").is_none());
2555 // The exact vCenter-style case Srikanth hit.
2556 let bool_only = r#"{"enabled":true}"#;
2557 assert!(embed_payload_in_first_string_field(bool_only, "<x/>").is_none());
2558 }
2559
2560 #[test]
2561 fn embed_payload_returns_none_for_invalid_json_sample() {
2562 assert!(embed_payload_in_first_string_field("garbage", "a=1&b=2").is_none());
2563 }
2564
2565 /// Round 35 (#859) — Srikanth on 0.3.179 saw variant-b probes flag
2566 /// every 4xx as a mismatch when the spec field had a `pattern` /
2567 /// `format` validator that correctly rejected the embedded
2568 /// payload. The probe was only ever meant to catch 5xx (server
2569 /// crashed parsing the embedded content); 4xx is the well-behaved
2570 /// outcome. Tristate `ExpectedOutcome::NotServerError` lets a
2571 /// variant-b probe pass on 2xx-4xx and fail only on 5xx.
2572 #[test]
2573 fn expected_outcome_pass_rules() {
2574 // Success (positive): 2xx-3xx pass, 4xx + 5xx fail.
2575 assert!(ExpectedOutcome::Success.passes(200));
2576 assert!(ExpectedOutcome::Success.passes(201));
2577 assert!(ExpectedOutcome::Success.passes(204));
2578 assert!(ExpectedOutcome::Success.passes(301));
2579 assert!(!ExpectedOutcome::Success.passes(400));
2580 assert!(!ExpectedOutcome::Success.passes(415));
2581 assert!(!ExpectedOutcome::Success.passes(500));
2582 assert!(!ExpectedOutcome::Success.passes(0));
2583
2584 // ClientError (negative): only 4xx pass.
2585 assert!(!ExpectedOutcome::ClientError.passes(200));
2586 assert!(ExpectedOutcome::ClientError.passes(400));
2587 assert!(ExpectedOutcome::ClientError.passes(404));
2588 assert!(ExpectedOutcome::ClientError.passes(422));
2589 assert!(!ExpectedOutcome::ClientError.passes(500));
2590
2591 // NotServerError (variant-b): 2xx-4xx pass, 5xx fails.
2592 assert!(ExpectedOutcome::NotServerError.passes(200));
2593 assert!(ExpectedOutcome::NotServerError.passes(204));
2594 assert!(ExpectedOutcome::NotServerError.passes(400), "Srikanth's vCenter consolecli case: 400 from a pattern validator should NOT be a probe failure");
2595 assert!(ExpectedOutcome::NotServerError.passes(415));
2596 assert!(ExpectedOutcome::NotServerError.passes(422));
2597 assert!(
2598 !ExpectedOutcome::NotServerError.passes(500),
2599 "Server CRASH on embedded content is the only real failure"
2600 );
2601 assert!(!ExpectedOutcome::NotServerError.passes(502));
2602 assert!(!ExpectedOutcome::NotServerError.passes(503));
2603 // status 0 (network error / probe never reached the server) does not pass either
2604 assert!(!ExpectedOutcome::NotServerError.passes(0));
2605 }
2606
2607 /// Round 35 (#859) — the per-capture `expected_status_range`
2608 /// string is what the HTML viewer's "show mismatches only"
2609 /// filter and Srikanth's `jq` pipelines key off, so the new
2610 /// tristate must surface a third distinct value.
2611 #[test]
2612 fn expected_outcome_string_labels() {
2613 assert_eq!(ExpectedOutcome::Success.as_str(), "2xx-3xx");
2614 assert_eq!(ExpectedOutcome::ClientError.as_str(), "4xx");
2615 assert_eq!(ExpectedOutcome::NotServerError.as_str(), "2xx-4xx");
2616 }
2617
2618 /// Round 26 — Srikanth saw `at /: Type { kind: Single` in his
2619 /// 0.3.169 capture for the vCenter `infraprofile/configs` 202
2620 /// response (spec promised `type: string`, server returned a
2621 /// JSON object). The output was a broken-syntax debug string.
2622 /// This test reproduces his exact spec+body and asserts the
2623 /// message is readable.
2624 #[test]
2625 fn response_schema_error_message_is_readable() {
2626 let schema = serde_json::json!({"type": "string"});
2627 let body = r#"{"data":{},"id":"generated_id","status":"created"}"#;
2628 let err = validate_body_against_schema(body, &schema).expect("type-mismatch fires");
2629 // The message must NOT contain Rust debug syntax leftovers
2630 // ("Type { kind:", trailing "{" or "(" tokens). It SHOULD say
2631 // what type was expected.
2632 assert!(!err.contains("Type { kind"), "stale debug output: {err}");
2633 assert!(!err.contains("{ kind:"), "stale debug output: {err}");
2634 assert!(err.contains("string"), "should name expected type: {err}");
2635 // Round 29 — Srikanth on 0.3.172 was confused by `at /:`,
2636 // thinking it pointed to the URL path. The new format
2637 // explicitly says "response body root" for the root case
2638 // (and "response body at /<pointer>" for nested fields).
2639 assert!(
2640 err.contains("response body root"),
2641 "should label root explicitly so reader knows it's not the URL: {err}"
2642 );
2643 // Round 28 — Srikanth wanted the expected schema embedded
2644 // in the message so it reads as 'expected schema {"type":"string"}'.
2645 assert!(
2646 err.contains("expected schema") && err.contains("\"type\":\"string\""),
2647 "should include expected schema JSON: {err}"
2648 );
2649 }
2650
2651 /// Round 29 — for non-root paths the format reads
2652 /// "response body at /name: ...". Catches the case where the
2653 /// root rewording accidentally dropped the JSON-pointer for
2654 /// nested fields.
2655 #[test]
2656 fn response_schema_error_uses_response_body_prefix_for_nested_paths() {
2657 let schema = serde_json::json!({
2658 "type": "object",
2659 "required": ["name"],
2660 "properties": {"name": {"type": "string"}}
2661 });
2662 let body = r#"{"name": 123}"#;
2663 let err = validate_body_against_schema(body, &schema).expect("type-mismatch fires");
2664 assert!(
2665 err.contains("response body at /name"),
2666 "nested path should read 'response body at /name': {err}"
2667 );
2668 assert!(!err.contains("response body root"), "wrong label for nested: {err}");
2669 // Round 30 — the "expected schema" suffix should be the
2670 // sub-schema at /name, not the entire object schema. Reader
2671 // shouldn't have to scan a 300-char object to find the
2672 // constraint that failed.
2673 assert!(
2674 err.contains(r#"expected schema {"type":"string"}"#),
2675 "should show only the /name sub-schema, not the full object: {err}"
2676 );
2677 }
2678
2679 /// Round 30 — Srikanth asked how a deeper nested mismatch reads.
2680 /// Schema: `name.type` should be a string; body has it as a number.
2681 /// JSON pointer is `/name/type`.
2682 #[test]
2683 fn response_schema_error_uses_response_body_prefix_for_deep_nested_paths() {
2684 let schema = serde_json::json!({
2685 "type": "object",
2686 "properties": {
2687 "name": {
2688 "type": "object",
2689 "properties": {"type": {"type": "string"}}
2690 }
2691 }
2692 });
2693 let body = r#"{"name": {"type": 123}}"#;
2694 let err = validate_body_against_schema(body, &schema).expect("type-mismatch fires");
2695 assert!(
2696 err.contains("response body at /name/type"),
2697 "deep nested path should read 'response body at /name/type': {err}"
2698 );
2699 // Round 30 — for deep paths the sub-schema is the leaf
2700 // {"type":"string"}, not the wrapping object schemas.
2701 assert!(
2702 err.contains(r#"expected schema {"type":"string"}"#),
2703 "should show only the /name/type leaf sub-schema: {err}"
2704 );
2705 }
2706
2707 /// Round 30 — when the instance pointer can't be resolved through
2708 /// the schema's `properties` chain (e.g. additionalProperties hit),
2709 /// `sub_schema_at_pointer` returns None and the message falls back
2710 /// to the full schema. Verifies the fallback path is wired.
2711 #[test]
2712 fn sub_schema_at_pointer_falls_back_for_unresolvable_paths() {
2713 let schema = serde_json::json!({"type":"object","additionalProperties":true});
2714 // Walker can't resolve /unknown, so we get the full schema back.
2715 assert_eq!(
2716 sub_schema_at_pointer(&schema, "/unknown"),
2717 None,
2718 "unresolvable path should return None to trigger fallback"
2719 );
2720 // Root path returns the whole schema.
2721 assert_eq!(sub_schema_at_pointer(&schema, "/"), Some(schema.clone()));
2722 assert_eq!(sub_schema_at_pointer(&schema, ""), Some(schema));
2723 }
2724
2725 #[test]
2726 fn response_schema_error_required_field_is_readable() {
2727 let schema = serde_json::json!({
2728 "type": "object",
2729 "required": ["id"],
2730 "properties": {"id": {"type": "integer"}}
2731 });
2732 let body = r#"{"other": 1}"#;
2733 let err = validate_body_against_schema(body, &schema).expect("required-missing fires");
2734 assert!(err.contains("required field missing"), "{err}");
2735 assert!(err.contains("id"), "{err}");
2736 }
2737
2738 /// Round 31 — Srikanth's vCenter case on 0.3.174: the
2739 /// `Appliance.Recovery.Backup.SystemName.Archive.Info` schema has
2740 /// a multi-paragraph description and ~6 required fields, of which
2741 /// `comment` was missing in the response. Before this fix the
2742 /// printed schema was the WHOLE parent object schema (parent's
2743 /// description bleeding in, all sibling property schemas dumped)
2744 /// truncated to 300 chars; after the fix it's the missing field's
2745 /// own schema. Verifies (a) parent description is gone and
2746 /// (b) sibling property names don't appear in the message.
2747 #[test]
2748 fn response_schema_error_required_focuses_on_missing_field_only() {
2749 let schema = serde_json::json!({
2750 "description": "The Appliance.Recovery.Backup.SystemName.Archive.Info schema represents backup archive information.\n\nThis schema was added in vSphere API 6.7.",
2751 "type": "object",
2752 "required": ["comment", "location", "parts", "system_name", "timestamp", "version"],
2753 "properties": {
2754 "comment": {
2755 "type": "string",
2756 "description": "Custom comment added by the user for this backup."
2757 },
2758 "location": {"type": "string", "description": "Backup location URL."},
2759 "parts": {"type": "array", "items": {"type": "string"}},
2760 "system_name": {"type": "string"},
2761 "timestamp": {"type": "string", "format": "date-time"},
2762 "version": {"type": "string"}
2763 }
2764 });
2765 let body = r#"{"location":"x","parts":[],"system_name":"y","timestamp":"z","version":"v"}"#;
2766 let err = validate_body_against_schema(body, &schema).expect("required-missing fires");
2767 assert!(err.contains("required field missing: \"comment\""), "{err}");
2768 // Parent's description should not appear; only the `comment`
2769 // field's own description (if any) may.
2770 assert!(
2771 !err.contains("Appliance.Recovery.Backup"),
2772 "parent description should not bleed into focused schema: {err}"
2773 );
2774 // No sibling property names should appear in the focused schema
2775 // suffix.
2776 for sibling in ["location", "parts", "system_name", "timestamp", "version"] {
2777 assert!(
2778 !err.contains(&format!("\"{sibling}\"")),
2779 "sibling field {sibling} should not appear in focused schema: {err}"
2780 );
2781 }
2782 }
2783
2784 #[test]
2785 fn response_schema_error_none_on_match() {
2786 let schema = serde_json::json!({"type": "string"});
2787 assert_eq!(validate_body_against_schema("\"hello\"", &schema), None);
2788 }
2789
2790 /// Round 34 (#827) — Srikanth on 0.3.178 hit the vCenter
2791 /// `consolecli` PUT where the `enabled: boolean` property has a
2792 /// multi-paragraph description. The schema printout truncated
2793 /// mid-description, hiding `type: boolean` past the 300-char cap.
2794 /// Stripping `description` (and friends) before serializing must
2795 /// keep the type info visible.
2796 #[test]
2797 fn response_schema_error_strips_description_so_type_survives_truncation() {
2798 // Schema crafted so without stripping, `description` would
2799 // push `type` past the 300-char truncation cap. The
2800 // description we use here is intentionally close to the
2801 // vCenter-spec wording Srikanth quoted.
2802 let big_desc = "In the result of the #get and #list operations this property indicates whether proxying is enabled for a particular protocol. In the input to the test and set operations this property specifies whether proxying should be enabled for a particular protocol. This property was added in vSphere API 6.7. Defaults to enabled if both this field and the value field are unset.";
2803 let schema = serde_json::json!({
2804 "type": "object",
2805 "required": ["enabled"],
2806 "properties": {
2807 "enabled": {
2808 "type": "boolean",
2809 "description": big_desc,
2810 "example": true,
2811 }
2812 }
2813 });
2814 let body = r#"{}"#;
2815 let err = validate_body_against_schema(body, &schema).expect("required-missing fires");
2816 assert!(err.contains("required field missing: \"enabled\""), "{err}");
2817 assert!(
2818 err.contains(r#""type":"boolean""#),
2819 "the `type: boolean` keyword must survive truncation: {err}"
2820 );
2821 // Description should NOT appear (we stripped it) so the
2822 // suffix is type-focused, not prose.
2823 assert!(
2824 !err.contains("proxying is enabled"),
2825 "description should be stripped from the printed schema: {err}"
2826 );
2827 assert!(
2828 !err.contains("\"example\""),
2829 "`example` field should be stripped from the printed schema: {err}"
2830 );
2831 }
2832
2833 /// Round 34 (#827) — strip_schema_noise should keep all
2834 /// constraint keywords intact; only the prose noise goes.
2835 #[test]
2836 fn strip_schema_noise_preserves_constraint_keywords() {
2837 let schema = serde_json::json!({
2838 "type": "object",
2839 "required": ["a", "b"],
2840 "description": "should be stripped",
2841 "title": "should be stripped",
2842 "example": {"a": 1, "b": 2},
2843 "properties": {
2844 "a": {"type": "string", "format": "uri", "minLength": 1, "description": "drop"},
2845 "b": {"type": "integer", "minimum": 0, "maximum": 100, "summary": "drop"},
2846 },
2847 });
2848 let stripped = strip_schema_noise(&schema);
2849 let s = serde_json::to_string(&stripped).unwrap();
2850 // Constraint keywords survive.
2851 for keep in [
2852 "\"type\"",
2853 "\"required\"",
2854 "\"properties\"",
2855 "\"format\"",
2856 "\"minLength\"",
2857 "\"minimum\"",
2858 "\"maximum\"",
2859 ] {
2860 assert!(s.contains(keep), "should keep {keep}: {s}");
2861 }
2862 // Noise fields are gone.
2863 for drop in ["description", "title", "example", "summary"] {
2864 assert!(!s.contains(&format!("\"{drop}\"")), "should strip {drop}: {s}");
2865 }
2866 }
2867
2868 #[test]
2869 fn json_serialises_report() {
2870 let r = SelfTestReport {
2871 positive_pass: 1,
2872 positive_fail: 0,
2873 negative_caught: BTreeMap::new(),
2874 negative_missed: BTreeMap::new(),
2875 operations: vec![OperationResult {
2876 method: "GET".into(),
2877 path: "/x".into(),
2878 positive: Some(CaseOutcome {
2879 label: "positive".into(),
2880 expected_4xx: false,
2881 actual_status: 200,
2882 passed: true,
2883 }),
2884 negatives: Vec::new(),
2885 }],
2886 };
2887 let json = serde_json::to_value(&r).expect("serialises");
2888 assert_eq!(json["positive_pass"], serde_json::json!(1));
2889 assert_eq!(json["operations"][0]["positive"]["actual_status"], serde_json::json!(200));
2890 }
2891}