@apifuse/provider-sdk 2.2.0-beta.4 → 2.2.0-beta.40

package/dist/server/self-test.js

@@ -1,5 +1,6 @@
 import { createHash, randomUUID } from "node:crypto";
 import { readFileSync } from "node:fs";
+import { TURN_KINDS } from "../auth-turn/index.js";
 import { Hono } from "hono";
 import { z } from "zod";
 import { resolveHealthCheckInputDateTokens } from "./self-test-input-tokens.js";
@@ -27,6 +28,36 @@ export const SelfTestRequestSchema = z.object({
     /** Credential material for requiresConnection cases; never persisted. */
     credentials: z.object({ inputs: z.record(z.string(), z.string()) }).optional(),
 });
+/**
+ * Skip reason reported when a declared auth flow does not complete in a single
+ * continue (OTP, retry loop). Cross-repo contract: the health-monitor maps
+ * this exact string to `self_test_incapable`; never vary it.
+ */
+export const SELF_TEST_AUTH_FLOW_MULTI_TURN_SKIP_REASON = "auth_flow_multi_turn";
+/**
+ * A `retry` turn after credential submission: the flow REJECTED the
+ * configured inputs (bad password, exchange failure). Distinct from the
+ * multi-turn gap so monitoring surfaces it as a real credential outage, and
+ * memoized like multi-turn so the probe does not re-submit rejected
+ * credentials every cycle (lockout safety).
+ */
+export const SELF_TEST_AUTH_FLOW_REJECTED_SKIP_REASON = "auth_flow_rejected";
+/**
+ * Known interactive turn kinds that justify the memoized multi-turn skip —
+ * they mean a human must participate (OTP, challenge, redirect, …).
+ * `retry` is deliberately excluded: after a credential submission it means
+ * rejection, not interaction (see SELF_TEST_AUTH_FLOW_REJECTED_SKIP_REASON).
+ * Kinds outside TURN_KINDS entirely are treated as flow errors.
+ */
+/**
+ * Post-submission /auth/continue statuses that mean the flow REJECTED the
+ * credentials (thrown AuthError -> 401, forbidden -> 403): memoized as
+ * `auth_flow_rejected`. Deliberately NOT 400 — the auth route maps generic
+ * ProviderErrors and Zod request errors there, which are often transient or
+ * fixable and must stay uncached retries (like 408/429/5xx).
+ */
+const AUTH_REJECTION_HTTP_STATUSES = new Set([401, 403]);
+const INTERACTIVE_TURN_KIND_SET = new Set(TURN_KINDS.filter((descriptor) => descriptor.rendering !== "terminal" && descriptor.kind !== "retry").map((descriptor) => descriptor.kind));
 function resolveSdkVersion() {
     try {
         const packageJsonUrl = new URL("../../package.json", import.meta.url);
@@ -101,8 +132,8 @@ export function isSelfTestReadOnlyOperation(operation) {
 }
 /** Binds the self-test executor to a tenant app's /v1 pipeline in-process. */
 export function createSelfTestInvoke(app) {
-    return async ({ operationId, input, connection, requestId }) => {
-        const response = await app.request(`/v1/${encodeURIComponent(operationId)}`, {
+    return async ({ operationId, input, connection, requestId, signal }) => {
+        const responsePromise = Promise.resolve(app.request(`/v1/${encodeURIComponent(operationId)}`, {
             method: "POST",
             headers: { "content-type": "application/json" },
             body: JSON.stringify({
@@ -110,7 +141,14 @@ export function createSelfTestInvoke(app) {
                 input: input ?? {},
                 ...(connection ? { connection } : {}),
             }),
-        });
+            ...(signal ? { signal } : {}),
+        }));
+        const response = await responsePromise;
+        if (signal?.aborted) {
+            await cancelSelfTestResponse(response);
+            signal.throwIfAborted();
+        }
+        // An abort after body consumption starts may let response.text() finish.
         const text = await response.text();
         let body = text;
         try {
@@ -130,25 +168,92 @@ export function createSelfTestInvoke(app) {
         return { status: response.status, data: body };
     };
 }
+async function cancelSelfTestResponse(response) {
+    await response.body?.cancel().catch(() => undefined);
+}
+function selfTestAbortReason(reason, sensitiveValues) {
+    const text = reason instanceof Error
+        ? reason.message || reason.name
+        : reason === undefined
+            ? "aborted"
+            : String(reason);
+    return redactSelfTestText(text, sensitiveValues);
+}
+function logSelfTestCancellation(provider, options, requestId, selected, signal, sensitiveValues) {
+    const event = {
+        level: "info",
+        event: "self_test_run_cancelled",
+        providerId: provider.id,
+        requestId,
+        operationId: selected.operationId,
+        caseName: selected.healthCase.name,
+        reason: selfTestAbortReason(signal.reason, sensitiveValues),
+    };
+    if (options.logger) {
+        options.logger(event);
+        return;
+    }
+    console.log(JSON.stringify(event));
+}
+/** Binds the self-test auth-flow driver to a tenant app's /auth pipeline in-process. */
+export function createSelfTestAuthFlowInvoke(app) {
+    return async ({ route, requestId, flowId, connectionId, externalRef, input, context, signal, }) => {
+        const response = await app.request(`/auth/${route}`, {
+            method: "POST",
+            headers: { "content-type": "application/json" },
+            body: JSON.stringify({
+                requestId,
+                flowId,
+                ...(connectionId ? { connectionId } : {}),
+                ...(externalRef ? { externalRef } : {}),
+                ...(input ? { input } : {}),
+                ...(context ? { context } : {}),
+            }),
+            ...(signal ? { signal } : {}),
+        });
+        const text = await response.text();
+        let body = text;
+        try {
+            body = text.length > 0 ? JSON.parse(text) : undefined;
+        }
+        catch {
+            // non-JSON transports keep the raw text as body
+        }
+        return { status: response.status, body };
+    };
+}
 class SelfTestCaseTimeoutError extends Error {
     constructor(timeoutMs) {
         super(`Self-test case timed out after ${timeoutMs}ms`);
         this.name = "SelfTestCaseTimeoutError";
     }
 }
-async function withCaseTimeout(run, timeoutMs) {
+async function withCaseTimeout(run, timeoutMs, controller) {
     let timer;
+    let onAbort;
     try {
         return await Promise.race([
             run(),
             new Promise((_, reject) => {
-                timer = setTimeout(() => reject(new SelfTestCaseTimeoutError(timeoutMs)), timeoutMs);
+                timer = setTimeout(() => {
+                    const error = new SelfTestCaseTimeoutError(timeoutMs);
+                    controller.abort(error);
+                    reject(error);
+                }, timeoutMs);
+            }),
+            new Promise((_, reject) => {
+                onAbort = () => reject(controller.signal.reason);
+                controller.signal.addEventListener("abort", onAbort, { once: true });
+                if (controller.signal.aborted)
+                    onAbort();
             }),
         ]);
     }
     finally {
         if (timer !== undefined)
             clearTimeout(timer);
+        if (onAbort)
+            controller.signal.removeEventListener("abort", onAbort);
     }
 }
 function objectProperty(value, key) {
@@ -171,6 +276,21 @@ function upstreamErrorMessage(body) {
     const message = objectProperty(error, "message");
     return typeof message === "string" ? message : undefined;
 }
+/**
+ * How long a memoized multi-turn flow outcome suppresses re-driving the auth
+ * flow. Generous on purpose: a multi-turn ceremony (OTP, device approval) is a
+ * provider property that changes on the timescale of releases, not probe
+ * cycles, and every re-drive is a REAL upstream login submission. The cache is
+ * in-process, so a pod restart also clears the entry.
+ */
+export const SELF_TEST_MULTI_TURN_RETRY_AFTER_MS = 24 * 60 * 60 * 1000;
+/**
+ * Age bound for POSITIVE cached credentials. Expiry modes that never produce
+ * a 401/403 (a 200 login page, an assertion failure) would otherwise replay
+ * the same stale session until pod restart — one re-login per day is the
+ * upstream-safe recovery for them.
+ */
+export const SELF_TEST_CREDENTIAL_MAX_AGE_MS = 24 * 60 * 60 * 1000;
 function resolveCaseTimeoutMs(execution, suite, healthCase) {
     const providerDefault = (execution.provider.healthProbe ?? execution.provider.healthMonitor)
         ?.defaultProbeTimeoutMs;
@@ -180,160 +300,646 @@ function resolveCaseTimeoutMs(execution, suite, healthCase) {
         providerDefault ??
         DEFAULT_CASE_TIMEOUT_MS);
 }
-function buildSelfTestConnection(execution, operationId, suite) {
+function credentialSessionCacheKey(providerId, inputs) {
+    const canonical = JSON.stringify(Object.keys(inputs)
+        .sort()
+        .map((key) => [key, inputs[key]]));
+    return `${providerId}:${createHash("sha256").update(canonical).digest("hex")}`;
+}
+function registerSensitiveValues(execution, values) {
+    for (const value of values) {
+        if (typeof value === "string" &&
+            value.length > 0 &&
+            !execution.sensitiveValues.includes(value)) {
+            execution.sensitiveValues.push(value);
+        }
+    }
+}
+function parseAuthFlowResponse(result) {
+    const errorEnvelope = objectProperty(result.body, "error");
+    if (result.status < 200 || result.status >= 300 || errorEnvelope !== undefined) {
+        const message = objectProperty(errorEnvelope, "message");
+        return {
+            ok: false,
+            code: "auth_flow_failed",
+            message: typeof message === "string"
+                ? message
+                : `Auth flow request failed with status ${result.status}`,
+            httpStatus: result.status,
+        };
+    }
+    const turnValue = objectProperty(result.body, "data");
+    const turnKind = objectProperty(turnValue, "kind");
+    if (typeof turnKind !== "string") {
+        return {
+            ok: false,
+            code: "auth_flow_failed",
+            message: "Auth flow returned an unrecognized turn.",
+            httpStatus: result.status,
+        };
+    }
+    const contextPatch = objectProperty(result.body, "contextPatch");
+    return {
+        ok: true,
+        turn: {
+            kind: turnKind,
+            data: objectProperty(turnValue, "data"),
+            expectedInput: objectProperty(turnValue, "expectedInput"),
+        },
+        ...(contextPatch && typeof contextPatch === "object" && !Array.isArray(contextPatch)
+            ? { contextPatch: contextPatch }
+            : {}),
+    };
+}
+function applyAuthFlowContextPatch(base, patch) {
+    if (!patch)
+        return base;
+    const next = { ...base };
+    for (const [key, value] of Object.entries(patch)) {
+        if (value === null) {
+            delete next[key];
+        }
+        else {
+            next[key] = value;
+        }
+    }
+    return next;
+}
+/**
+ * Extracts the completed credential from a complete turn's data payload — the
+ * same `data.credential` record the gateway persists as connection secrets in
+ * production (`persistCredential` → credential-service `UpdateCredential`).
+ */
+function completedCredentialFromTurn(turnData) {
+    const credential = objectProperty(turnData, "credential");
+    if (!credential || typeof credential !== "object" || Array.isArray(credential)) {
+        return undefined;
+    }
+    const secrets = {};
+    for (const [key, value] of Object.entries(credential)) {
+        if (typeof value === "string")
+            secrets[key] = value;
+    }
+    return Object.keys(secrets).length > 0 ? secrets : undefined;
+}
+/**
+ * Drives the provider's declared auth flow exactly like production does:
+ * `flow.start()` then a single `flow.continue(credentialInputs)`. Anything
+ * other than a complete turn is a visible multi-turn gap, never a fabricated
+ * probe failure.
+ */
+/**
+ * Fields an input-prompt turn actually requests. The canonical auth-turn
+ * shape carries the JSON schema DIRECTLY on `expectedInput` (`ctx.auth
+ * .nextForm`/`defineCredentialsAuth`, the committed fixtures); some providers
+ * nest it as `expectedInput.schema`. Both are honored. `null` when the turn
+ * declares no schema (legacy/loose flows keep full-input semantics).
+ */
+function turnRequestedFields(turn) {
+    const expectedInput = turn.expectedInput;
+    if (!expectedInput || typeof expectedInput !== "object" || Array.isArray(expectedInput)) {
+        return null;
+    }
+    const schemaOf = (candidate) => {
+        const properties = candidate && typeof candidate === "object"
+            ? candidate.properties
+            : undefined;
+        if (!properties || typeof properties !== "object" || Array.isArray(properties)) {
+            return null;
+        }
+        const requiredRaw = candidate && typeof candidate === "object"
+            ? candidate.required
+            : undefined;
+        const required = Array.isArray(requiredRaw)
+            ? requiredRaw.filter((field) => typeof field === "string")
+            : [];
+        return { properties: Object.keys(properties), required };
+    };
+    return (schemaOf(expectedInput) ?? schemaOf(expectedInput.schema));
+}
+async function materializeFlowCredential(execution, inputs, options = {}) {
+    const authFlow = execution.authFlow;
+    if (!authFlow) {
+        return {
+            kind: "flow_error",
+            code: "auth_flow_unavailable",
+            message: "Provider declares a credentials auth flow but the self-test host has no auth-flow driver.",
+        };
+    }
+    const flowId = `self-test-${randomUUID()}`;
+    // The login must ride the SAME proxy/connection affinity the probe will
+    // use (createAuthFlowContext keys affinity on connectionId) — otherwise
+    // IP/session-bound upstreams see the cookie arrive from a different
+    // session and reject it.
+    const started = parseAuthFlowResponse(await authFlow({
+        route: "start",
+        requestId: `${execution.requestId}-auth-start-${randomUUID()}`,
+        flowId,
+        ...(options.connectionId ? { connectionId: options.connectionId } : {}),
+        ...(options.externalRef ? { externalRef: options.externalRef } : {}),
+        ...(options.signal ? { signal: options.signal } : {}),
+    }));
+    if (!started.ok) {
+        return { kind: "flow_error", code: started.code, message: started.message };
+    }
+    let turn = started.turn;
+    const flowContext = applyAuthFlowContextPatch({}, started.contextPatch);
+    if (turn.kind === "abort") {
+        // Terminal turn: continuing after an abort would replay credentials into
+        // a flow that already refused to proceed. Not memoized (flow errors are
+        // never cached) — an abort can be transient upstream maintenance.
+        return {
+            kind: "flow_error",
+            code: "auth_flow_aborted",
+            message: "Auth flow aborted before requesting input.",
+        };
+    }
+    if (turn.kind !== "complete") {
+        // Validate the start turn BEFORE submitting credentials: an unknown
+        // kind may be a provider typo or a stage that must not receive the
+        // probe inputs. `retry` counts as an input prompt at this stage.
+        if (turn.kind !== "retry" && !INTERACTIVE_TURN_KIND_SET.has(turn.kind)) {
+            return {
+                kind: "flow_error",
+                code: "auth_flow_unexpected_turn",
+                message: `Auth flow start returned an unrecognized turn kind "${turn.kind}".`,
+            };
+        }
+        // Auto-continue ONLY into input prompts (form/retry). Other known
+        // interactive stages (redirect, poll, pending, challenge, message,
+        // multi_choice) are valid flows that are NOT asking for the credential
+        // inputs — posting the password there submits it to the wrong stage.
+        // They are a genuine headless gap: the memoized multi-turn skip.
+        if (turn.kind !== "form" && turn.kind !== "retry") {
+            return { kind: "skip", skipReason: SELF_TEST_AUTH_FLOW_MULTI_TURN_SKIP_REASON };
+        }
+        // Submit ONLY what the turn asks for: a first stage of a multi-step
+        // login may request a subset (or different fields entirely) — posting
+        // the full inputs would send secrets to the wrong stage. Only the
+        // schema's REQUIRED fields are mandatory (defineCredentialsAuth
+        // encodes optional fields by omitting them from `required`); a turn
+        // whose required fields we do not hold is a headless gap (multi-turn).
+        // A turn with no declared schema keeps full-input semantics.
+        const requestedFields = turnRequestedFields(turn);
+        let submitInputs = { ...inputs };
+        if (requestedFields !== null) {
+            if (requestedFields.required.some((field) => inputs[field] === undefined)) {
+                return { kind: "skip", skipReason: SELF_TEST_AUTH_FLOW_MULTI_TURN_SKIP_REASON };
+            }
+            submitInputs = Object.fromEntries(requestedFields.properties
+                .filter((field) => inputs[field] !== undefined)
+                .map((field) => [field, inputs[field]]));
+        }
+        // The case deadline may have fired while start() was still running.
+        // Never submit real credentials into a flow whose case already
+        // reported self_test_timeout — a late continue is a real upstream
+        // login/OTP attempt nobody is waiting for.
+        if (options.isAbandoned?.() === true) {
+            return {
+                kind: "flow_error",
+                code: "self_test_timeout",
+                message: "Case deadline passed before credential submission; flow abandoned.",
+            };
+        }
+        const continued = parseAuthFlowResponse(await authFlow({
+            route: "continue",
+            requestId: `${execution.requestId}-auth-continue-${randomUUID()}`,
+            flowId,
+            ...(options.connectionId ? { connectionId: options.connectionId } : {}),
+            ...(options.externalRef ? { externalRef: options.externalRef } : {}),
+            input: submitInputs,
+            ...(Object.keys(flowContext).length > 0 ? { context: flowContext } : {}),
+            ...(options.signal ? { signal: options.signal } : {}),
+        }));
+        if (!continued.ok) {
+            // Providers built with defineCredentialsAuth cannot return a retry
+            // turn — a rejected password THROWS and /auth/continue answers with
+            // an auth-shaped 401/403. That is a credential REJECTION (memoized,
+            // so the probe never hammers a locked-out login); every other
+            // status stays an uncached transient retry.
+            if (AUTH_REJECTION_HTTP_STATUSES.has(continued.httpStatus)) {
+                return { kind: "skip", skipReason: SELF_TEST_AUTH_FLOW_REJECTED_SKIP_REASON };
+            }
+            return { kind: "flow_error", code: continued.code, message: continued.message };
+        }
+        turn = continued.turn;
+    }
+    if (turn.kind === "abort") {
+        return {
+            kind: "flow_error",
+            code: "auth_flow_aborted",
+            message: "Auth flow aborted after credential submission.",
+        };
+    }
+    if (turn.kind === "retry") {
+        // A retry turn AFTER submission is a credential rejection, not an
+        // interactive gap — surfaced distinctly so monitoring can treat it as
+        // a real outage, and memoized by the caller (lockout safety).
+        return { kind: "skip", skipReason: SELF_TEST_AUTH_FLOW_REJECTED_SKIP_REASON };
+    }
+    if (turn.kind !== "complete") {
+        // Only KNOWN interactive kinds are a genuine "cannot complete headless"
+        // multi-turn gap (memoized by the caller). An unknown kind is ambiguous
+        // — it may encode a transient provider failure — so it reports as a
+        // flow error, which is never memoized, instead of freezing the signal.
+        if (!INTERACTIVE_TURN_KIND_SET.has(turn.kind)) {
+            return {
+                kind: "flow_error",
+                code: "auth_flow_unexpected_turn",
+                message: `Auth flow returned an unrecognized turn kind "${turn.kind}".`,
+            };
+        }
+        return { kind: "skip", skipReason: SELF_TEST_AUTH_FLOW_MULTI_TURN_SKIP_REASON };
+    }
+    const credential = completedCredentialFromTurn(turn.data);
+    if (!credential) {
+        return {
+            kind: "flow_error",
+            code: "auth_flow_invalid_credential",
+            message: "Auth flow completed without a string-valued credential payload.",
+        };
+    }
+    // Redaction contract: flow-issued secrets are registered BEFORE any probe
+    // output can be built from them.
+    registerSensitiveValues(execution, Object.values(credential));
+    return { credential };
+}
+async function resolveSelfTestConnection(execution, operationId, suite, options = {}) {
     if (!suite.requiresConnection)
-        return {};
+        return { kind: "connection" };
     const inputs = execution.credentials ?? {};
     const declaredFields = Object.keys((execution.provider.healthProbe ?? execution.provider.healthMonitor)?.credentialInputs ?? {});
     for (const field of declaredFields) {
         if (!inputs[field]) {
-            return { skipReason: `credential_missing:${field}` };
+            return { kind: "skip", skipReason: `credential_missing:${field}` };
         }
     }
     if (declaredFields.length === 0 && Object.keys(inputs).length === 0) {
-        return { skipReason: "credential_missing:credentials" };
+        return { kind: "skip", skipReason: "credential_missing:credentials" };
+    }
+    // The connection id seeds proxy/connection affinity in the provider
+    // context, so it must be STABLE per (provider, credentialInputs): a cached
+    // session replayed under a per-request id would ride a different proxy/IP
+    // each cycle and upstreams would treat the cookie as stale or suspicious.
+    // The id carries only a hash of the inputs, never the inputs themselves.
+    //
+    // Providers declaring `proxy.session.affinity: "operation"` pin the PROBE's
+    // proxy to `${providerId}/${operationId}` regardless of connection id — so
+    // the login must ride that exact key, and the session cache splits per
+    // operation (one shared cookie would otherwise hop between per-operation
+    // proxies).
+    const operationAffinity = typeof execution.provider.proxy === "object" &&
+        execution.provider.proxy?.session?.affinity === "operation";
+    const credentialKey = credentialSessionCacheKey(execution.provider.id, inputs);
+    const affinityKey = operationAffinity ? `${credentialKey}:${operationId}` : credentialKey;
+    // ONE id for the auth flow AND the probe connection: providers may bind
+    // the issued credential to FlowContext.connectionId and later compare it
+    // against ctx.request.connectionId. Operation-affinity providers use the
+    // probe's exact proxy key (providerId/operationId); everyone else uses the
+    // stable per-credential hash.
+    const connectionId = operationAffinity
+        ? `${execution.provider.id}/${operationId}`
+        : `self-test-${createHash("sha256").update(affinityKey).digest("hex").slice(0, 22)}`;
+    const buildConnection = (secrets) => ({
+        id: connectionId,
+        mode: "credentials",
+        secrets: { ...secrets },
+        metadata: { purpose: "provider-self-test", operationId },
+        externalRef: `${execution.provider.id}-${operationId}-self-test`,
+    });
+    const auth = execution.provider.auth;
+    if (auth?.mode !== "credentials" || !auth.flow) {
+        // Providers without a declared credentials flow keep raw-input semantics
+        // — and the pre-existing per-request connection id: there is no session
+        // to keep on one affinity, and a stable id would silently pin every
+        // cycle of a connection-affinity proxy to the same upstream session.
+        return {
+            kind: "connection",
+            connection: {
+                ...buildConnection(inputs),
+                id: `self-test-${execution.requestId}`,
+            },
+            credentialSource: "inputs",
+        };
+    }
+    const cacheKey = affinityKey;
+    const cached = execution.sessionCache.get(cacheKey);
+    // DR-7 upstream-account safety: a memoized multi-turn outcome
+    // short-circuits to the auth_flow_multi_turn skip WITHOUT re-driving
+    // flow.start()/flow.continue() — every re-drive is a real upstream login
+    // submission (OTP sends, lockout risk), and the probe scheduler would
+    // otherwise repeat it every cycle forever. Changed credentialInputs hash
+    // to a different key and re-attempt immediately; otherwise the entry
+    // expires after a generous TTL (or process restart) so a provider whose
+    // flow becomes single-turn again is eventually re-probed.
+    if (cached?.kind === "multi_turn" || cached?.kind === "rejected") {
+        if (Date.now() - cached.cachedAtMs < SELF_TEST_MULTI_TURN_RETRY_AFTER_MS) {
+            return {
+                kind: "skip",
+                skipReason: cached.kind === "rejected"
+                    ? SELF_TEST_AUTH_FLOW_REJECTED_SKIP_REASON
+                    : SELF_TEST_AUTH_FLOW_MULTI_TURN_SKIP_REASON,
+            };
+        }
+        execution.sessionCache.delete(cacheKey);
+    }
+    if (options.forceLogin !== true && cached?.kind === "credential") {
+        if (Date.now() - cached.cachedAtMs >= SELF_TEST_CREDENTIAL_MAX_AGE_MS) {
+            // Age-bounded: expiry modes that never 401 (login-page 200s,
+            // assertion failures) must not replay one stale session forever.
+            execution.sessionCache.delete(cacheKey);
+        }
+        else {
+            registerSensitiveValues(execution, Object.values(cached.credential));
+            return {
+                kind: "connection",
+                connection: buildConnection(cached.credential),
+                credentialSource: "cache",
+                cacheKey,
+            };
+        }
+    }
+    const materialized = await materializeFlowCredential(execution, inputs, {
+        ...(options.isAbandoned !== undefined ? { isAbandoned: options.isAbandoned } : {}),
+        connectionId,
+        externalRef: `${execution.provider.id}-${operationId}-self-test`,
+        ...(options.signal ? { signal: options.signal } : {}),
+    });
+    if (!("credential" in materialized)) {
+        // Only the multi-turn SKIP is negative-cached. Flow ERRORS
+        // (auth_flow_unavailable / auth_flow_failed / invalid credential
+        // payloads, or a thrown start/continue) are never memoized: they are
+        // typically transient upstream or host failures, so each cycle may
+        // retry — permanently caching an error would silently freeze the
+        // signal on a blip, while retrying a FAILED request is not a repeated
+        // successful login submission.
+        if (materialized.kind === "skip" && options.isAbandoned?.() !== true) {
+            if (materialized.skipReason === SELF_TEST_AUTH_FLOW_MULTI_TURN_SKIP_REASON) {
+                execution.sessionCache.set(cacheKey, { kind: "multi_turn", cachedAtMs: Date.now() });
+            }
+            else if (materialized.skipReason === SELF_TEST_AUTH_FLOW_REJECTED_SKIP_REASON) {
+                execution.sessionCache.set(cacheKey, { kind: "rejected", cachedAtMs: Date.now() });
+            }
+        }
+        return materialized;
     }
+    // Abort is cooperative: a flow may still complete after the case deadline
+    // if provider code ignores the signal. Never cache that late credential or
+    // let the next probe hide the timed-out login.
+    if (options.isAbandoned?.() === true) {
+        return {
+            kind: "flow_error",
+            code: "self_test_timeout",
+            message: "Auth flow completed after the case deadline; credential discarded.",
+        };
+    }
+    execution.sessionCache.set(cacheKey, {
+        kind: "credential",
+        credential: materialized.credential,
+        cachedAtMs: Date.now(),
+    });
     return {
-        connection: {
-            id: `self-test-${execution.requestId}`,
-            mode: "credentials",
-            secrets: { ...inputs },
-            metadata: { purpose: "provider-self-test", operationId },
-            externalRef: `${execution.provider.id}-${operationId}-self-test`,
-        },
+        kind: "connection",
+        connection: buildConnection(materialized.credential),
+        credentialSource: "flow",
+        cacheKey,
     };
 }
-async function executeSelfTestCase(execution, operationId, suite, healthCase) {
-    const { provider, invoke, sensitiveValues } = execution;
-    const redact = (text) => redactSelfTestText(text, sensitiveValues);
-    const startedAt = new Date().toISOString();
-    const startedAtMs = performance.now();
-    const finish = (partial) => ({
-        operationId,
-        caseName: healthCase.name,
-        startedAt,
-        finishedAt: new Date().toISOString(),
-        responseTimeMs: partial.responseTimeMs ?? Math.max(0, Math.round(performance.now() - startedAtMs)),
-        ...partial,
-    });
+/** A failed probe whose HTTP status is auth-shaped invalidates a cached session once. */
+function isAuthFailureCaseResult(result) {
+    return result.status === "failed" && (result.httpStatus === 401 || result.httpStatus === 403);
+}
+async function executeSelfTestCase(execution, operationId, suite, healthCase, caseController) {
+    const { provider, invoke } = execution;
+    // execution.sensitiveValues may grow while the case runs (flow-issued
+    // secrets); redact always reads the live array.
+    const redact = (text) => redactSelfTestText(text, execution.sensitiveValues);
     const defaultLabel = redact(healthCase.description ?? healthCase.name);
+    const timeoutMs = resolveCaseTimeoutMs(execution, suite, healthCase);
+    // One deadline for the WHOLE case: connection materialization (auth flow),
+    // the probe, and the one-shot auth retry all draw from the same budget —
+    // a 30s case must never take ~4×30s across its stages.
+    const caseDeadlineAtMs = performance.now() + timeoutMs;
+    const remainingCaseTimeoutMs = () => Math.max(1, Math.ceil(caseDeadlineAtMs - performance.now()));
+    const runWithCaseTimeout = (run, remainingMs) => withCaseTimeout(run, remainingMs, caseController);
+    const beginCase = () => {
+        const startedAt = new Date().toISOString();
+        const startedAtMs = performance.now();
+        return {
+            startedAtMs,
+            finish: (partial) => ({
+                operationId,
+                caseName: healthCase.name,
+                startedAt,
+                finishedAt: new Date().toISOString(),
+                responseTimeMs: partial.responseTimeMs ?? Math.max(0, Math.round(performance.now() - startedAtMs)),
+                ...partial,
+            }),
+        };
+    };
+    const caseScope = beginCase();
     if (healthCase.enabled && healthCase.enabled() === false) {
-        return finish({
+        return caseScope.finish({
             status: "skipped",
             label: defaultLabel,
             skipReason: "disabled",
         });
     }
-    const connectionResolution = buildSelfTestConnection(execution, operationId, suite);
-    if ("skipReason" in connectionResolution) {
-        return finish({
-            status: "skipped",
+    const resolveConnection = async (forceLogin) => {
+        // The deadline aborts the in-process flow request. Keep the abandonment
+        // guard as a backstop for provider code that ignores cancellation so a
+        // late completion still cannot write the session cache.
+        let abandoned = false;
+        try {
+            return await runWithCaseTimeout(() => resolveSelfTestConnection(execution, operationId, suite, {
+                forceLogin,
+                isAbandoned: () => abandoned,
+                signal: caseController.signal,
+            }), remainingCaseTimeoutMs());
+        }
+        catch (error) {
+            abandoned = true;
+            return {
+                kind: "flow_error",
+                code: error instanceof SelfTestCaseTimeoutError ? "self_test_timeout" : "auth_flow_failed",
+                message: error instanceof Error ? error.message : String(error),
+            };
+        }
+    };
+    const nonConnectionResult = (resolution) => {
+        if (resolution.kind === "skip") {
+            return caseScope.finish({
+                status: "skipped",
+                label: defaultLabel,
+                skipReason: resolution.skipReason,
+            });
+        }
+        return caseScope.finish({
+            status: "error",
             label: defaultLabel,
-            skipReason: connectionResolution.skipReason,
+            error: { code: resolution.code, message: redact(resolution.message) },
         });
-    }
-    const connection = connectionResolution.connection;
-    const timeoutMs = resolveCaseTimeoutMs(execution, suite, healthCase);
-    try {
-        return await withCaseTimeout(async () => {
-            const resolvedInput = resolveHealthCheckInputDateTokens(healthCase.input);
-            const preparedInput = healthCase.prepareInput
-                ? await healthCase.prepareInput({
-                    providerId: provider.id,
+    };
+    const runProbeAttempt = async (connection) => {
+        // gateway.execute keeps its contract — EVERY helper status is returned
+        // to the prepareInput hook (it may branch on 401 itself). The last
+        // auth-shaped helper status is only RECORDED: if the hook then throws,
+        // the case fails WITH that status so stale-session recovery triggers.
+        let prepareAuthStatus = null;
+        // Share the OUTER case scope: startedAt/responseTimeMs must cover the
+        // WHOLE case — auth-flow materialization included — not just the final
+        // operation attempt, or a slow login reads as a fast healthy case.
+        const { startedAtMs, finish } = caseScope;
+        try {
+            return await runWithCaseTimeout(async () => {
+                const resolvedInput = resolveHealthCheckInputDateTokens(healthCase.input);
+                const preparedInput = healthCase.prepareInput
+                    ? await healthCase.prepareInput({
+                        providerId: provider.id,
+                        operationId,
+                        input: resolvedInput,
+                        ...(connection ? { connectionId: connection.id } : {}),
+                        gateway: {
+                            execute: async (foreignProviderId, gatewayOperationId, gatewayInput) => {
+                                if (foreignProviderId !== provider.id) {
+                                    throw new Error(`Self-test prepareInput may only invoke provider "${provider.id}" operations (requested "${foreignProviderId}").`);
+                                }
+                                const startedGatewayMs = performance.now();
+                                const executed = await invoke({
+                                    operationId: gatewayOperationId,
+                                    input: gatewayInput,
+                                    connection,
+                                    requestId: `${execution.requestId}-prepare-${randomUUID()}`,
+                                    signal: caseController.signal,
+                                });
+                                if (executed.status === 401 || executed.status === 403) {
+                                    prepareAuthStatus = executed.status;
+                                }
+                                return {
+                                    status: executed.status,
+                                    duration: performance.now() - startedGatewayMs,
+                                    data: executed.data,
+                                    meta: executed.meta,
+                                };
+                            },
+                        },
+                    })
+                    : resolvedInput;
+                const executed = await invoke({
                     operationId,
-                    input: resolvedInput,
-                    ...(connection ? { connectionId: connection.id } : {}),
-                    gateway: {
-                        execute: async (foreignProviderId, gatewayOperationId, gatewayInput) => {
-                            if (foreignProviderId !== provider.id) {
-                                throw new Error(`Self-test prepareInput may only invoke provider "${provider.id}" operations (requested "${foreignProviderId}").`);
-                            }
-                            const startedGatewayMs = performance.now();
-                            const executed = await invoke({
-                                operationId: gatewayOperationId,
-                                input: gatewayInput,
-                                connection,
-                                requestId: `${execution.requestId}-prepare-${randomUUID()}`,
-                            });
-                            return {
-                                status: executed.status,
-                                duration: performance.now() - startedGatewayMs,
-                                data: executed.data,
-                                meta: executed.meta,
-                            };
+                    input: preparedInput,
+                    connection,
+                    requestId: `${execution.requestId}-${randomUUID()}`,
+                    signal: caseController.signal,
+                });
+                const durationMs = performance.now() - startedAtMs;
+                if (executed.status < 200 || executed.status >= 300) {
+                    return finish({
+                        status: "failed",
+                        label: defaultLabel,
+                        httpStatus: executed.status,
+                        error: {
+                            code: upstreamErrorCode(executed.data) ?? "operation_failed",
+                            message: redact(upstreamErrorMessage(executed.data) ??
+                                `Operation invocation failed with status ${executed.status}`),
                         },
-                    },
-                })
-                : resolvedInput;
-            const executed = await invoke({
-                operationId,
-                input: preparedInput,
-                connection,
-                requestId: `${execution.requestId}-${randomUUID()}`,
-            });
-            const durationMs = performance.now() - startedAtMs;
-            if (executed.status < 200 || executed.status >= 300) {
+                    });
+                }
+                const assertionContext = {
+                    status: executed.status,
+                    data: executed.data,
+                    durationMs,
+                    ...(executed.meta ? { meta: executed.meta } : {}),
+                };
+                let assertionResult;
+                try {
+                    assertionResult = await healthCase.assertions(assertionContext);
+                }
+                catch (assertionError) {
+                    return finish({
+                        status: "failed",
+                        label: defaultLabel,
+                        httpStatus: executed.status,
+                        assertion: {
+                            passed: false,
+                            message: redact(assertionError instanceof Error ? assertionError.message : String(assertionError)),
+                        },
+                    });
+                }
+                const statusValue = objectProperty(assertionResult, "status");
+                const overrideStatus = statusValue === "ok" || statusValue === "degraded" ? statusValue : undefined;
+                const labelValue = objectProperty(assertionResult, "label");
+                const overrideLabel = typeof labelValue === "string" ? redact(labelValue) : undefined;
                 return finish({
-                    status: "failed",
-                    label: defaultLabel,
+                    status: overrideStatus ?? "ok",
+                    label: overrideLabel ?? defaultLabel,
                     httpStatus: executed.status,
-                    error: {
-                        code: upstreamErrorCode(executed.data) ?? "operation_failed",
-                        message: redact(upstreamErrorMessage(executed.data) ??
-                            `Operation invocation failed with status ${executed.status}`),
-                    },
+                    assertion: { passed: true },
+                });
+            }, remainingCaseTimeoutMs());
+        }
+        catch (error) {
+            if (error instanceof SelfTestCaseTimeoutError) {
+                return finish({
+                    status: "error",
+                    label: defaultLabel,
+                    error: { code: "self_test_timeout", message: redact(error.message) },
                 });
             }
-            const assertionContext = {
-                status: executed.status,
-                data: executed.data,
-                durationMs,
-                ...(executed.meta ? { meta: executed.meta } : {}),
-            };
-            let assertionResult;
-            try {
-                assertionResult = await healthCase.assertions(assertionContext);
-            }
-            catch (assertionError) {
+            if (prepareAuthStatus !== null) {
                 return finish({
                     status: "failed",
                     label: defaultLabel,
-                    httpStatus: executed.status,
+                    httpStatus: prepareAuthStatus,
                     assertion: {
                         passed: false,
-                        message: redact(assertionError instanceof Error ? assertionError.message : String(assertionError)),
+                        message: redact(error instanceof Error ? error.message : String(error)),
                     },
                 });
             }
-            const statusValue = objectProperty(assertionResult, "status");
-            const overrideStatus = statusValue === "ok" || statusValue === "degraded" ? statusValue : undefined;
-            const labelValue = objectProperty(assertionResult, "label");
-            const overrideLabel = typeof labelValue === "string" ? redact(labelValue) : undefined;
-            return finish({
-                status: overrideStatus ?? "ok",
-                label: overrideLabel ?? defaultLabel,
-                httpStatus: executed.status,
-                assertion: { passed: true },
-            });
-        }, timeoutMs);
-    }
-    catch (error) {
-        if (error instanceof SelfTestCaseTimeoutError) {
             return finish({
                 status: "error",
                 label: defaultLabel,
-                error: { code: "self_test_timeout", message: redact(error.message) },
+                error: {
+                    code: "self_test_execution_error",
+                    message: redact(error instanceof Error ? error.message : String(error)),
+                },
             });
         }
-        return finish({
-            status: "error",
-            label: defaultLabel,
-            error: {
-                code: "self_test_execution_error",
-                message: redact(error instanceof Error ? error.message : String(error)),
-            },
-        });
+    };
+    const resolution = await resolveConnection(false);
+    if (resolution.kind !== "connection") {
+        return nonConnectionResult(resolution);
+    }
+    let result = await runProbeAttempt(resolution.connection);
+    // One-shot session recovery: a cached credential that fails the probe with
+    // an auth-shaped status is invalidated, the flow re-runs ONCE, and the
+    // probe retries once. Fresh (just-materialized) credentials never retry.
+    if (resolution.credentialSource === "cache" &&
+        resolution.cacheKey !== undefined &&
+        isAuthFailureCaseResult(result)) {
+        execution.sessionCache.delete(resolution.cacheKey);
+        const retryResolution = await resolveConnection(true);
+        if (retryResolution.kind !== "connection") {
+            return nonConnectionResult(retryResolution);
+        }
+        result = await runProbeAttempt(retryResolution.connection);
+        // The retry's fresh credential is subject to the same eviction rule
+        // as a first-attempt fresh credential (below).
+        if (retryResolution.cacheKey !== undefined && isAuthFailureCaseResult(result)) {
+            execution.sessionCache.delete(retryResolution.cacheKey);
+        }
+        return result;
+    }
+    // A FRESH credential the probe just rejected is known-bad: evict it so the
+    // next cycle logs in anew instead of replaying a guaranteed-stale session
+    // once before recovering. (No retry here — fresh credentials never retry.)
+    if (resolution.credentialSource === "flow" &&
+        resolution.cacheKey !== undefined &&
+        isAuthFailureCaseResult(result)) {
+        execution.sessionCache.delete(resolution.cacheKey);
     }
+    return result;
 }
 function selectCases(provider, request) {
     const singleCase = request.operationId !== undefined && request.caseName !== undefined;
@@ -420,6 +1026,10 @@ export function createSelfTestApp(provider, options) {
     const app = new Hono();
     const planDigest = computeSelfTestPlanDigest(provider);
     const requestBudgetMs = resolveRequestBudgetMs(options);
+    // In-process flow-credential session cache (providerId + credentialInputs
+    // hash → materialized credential). Lives as long as the app so consecutive
+    // probe cycles never log in to the upstream more than once per session.
+    const sessionCache = new Map();
     let busy = false;
     app.notFound((c) => c.json({ error: { code: "not_found", message: "Not found" } }, 404));
     app.get(SELF_TEST_HEALTHZ_PATH, (c) => c.json({ ok: true }));
@@ -486,6 +1096,7 @@ export function createSelfTestApp(provider, options) {
             const execution = {
                 provider,
                 invoke: options.invoke,
+                ...(options.authFlow ? { authFlow: options.authFlow } : {}),
                 requestId: request.requestId,
                 credentials: request.credentials?.inputs,
                 requestTimeoutMs: request.timeoutMs,
@@ -493,44 +1104,69 @@ export function createSelfTestApp(provider, options) {
                     env: options.env,
                     credentialInputs: request.credentials?.inputs,
                 }),
+                sessionCache,
             };
             const deadline = performance.now() + requestBudgetMs;
             const results = [];
+            const runSignal = c.req.raw.signal;
+            let interruptedCase;
             // Sequential execution (parallelism 1): self-tests run on serving pods
             // and must never compete with themselves for upstream quota.
             for (const selected of selection.cases) {
-                if (performance.now() >= deadline) {
-                    const now = new Date().toISOString();
-                    results.push({
-                        operationId: selected.operationId,
-                        caseName: selected.healthCase.name,
-                        status: "skipped",
-                        label: selected.healthCase.name,
-                        responseTimeMs: 0,
-                        skipReason: "budget_exhausted",
-                        startedAt: now,
-                        finishedAt: now,
-                    });
-                    continue;
+                const caseController = new AbortController();
+                const abortFromRun = () => caseController.abort(runSignal.reason);
+                runSignal.addEventListener("abort", abortFromRun, { once: true });
+                if (runSignal.aborted)
+                    abortFromRun();
+                try {
+                    if (runSignal.aborted) {
+                        interruptedCase = selected;
+                        break;
+                    }
+                    if (performance.now() >= deadline) {
+                        const now = new Date().toISOString();
+                        results.push({
+                            operationId: selected.operationId,
+                            caseName: selected.healthCase.name,
+                            status: "skipped",
+                            label: selected.healthCase.name,
+                            responseTimeMs: 0,
+                            skipReason: "budget_exhausted",
+                            startedAt: now,
+                            finishedAt: now,
+                        });
+                        continue;
+                    }
+                    if (!isSelfTestReadOnlyOperation(selected.operation)) {
+                        const now = new Date().toISOString();
+                        results.push({
+                            operationId: selected.operationId,
+                            caseName: selected.healthCase.name,
+                            status: "error",
+                            label: selected.healthCase.name,
+                            responseTimeMs: 0,
+                            error: {
+                                code: "operation_not_read_only",
+                                message: `Operation "${selected.operationId}" is not classified read-only; self-test refuses to execute it.`,
+                            },
+                            startedAt: now,
+                            finishedAt: now,
+                        });
+                        continue;
+                    }
+                    results.push(await executeSelfTestCase(execution, selected.operationId, selected.suite, selected.healthCase, caseController));
+                    if (runSignal.aborted) {
+                        interruptedCase = selected;
+                        break;
+                    }
                 }
-                if (!isSelfTestReadOnlyOperation(selected.operation)) {
-                    const now = new Date().toISOString();
-                    results.push({
-                        operationId: selected.operationId,
-                        caseName: selected.healthCase.name,
-                        status: "error",
-                        label: selected.healthCase.name,
-                        responseTimeMs: 0,
-                        error: {
-                            code: "operation_not_read_only",
-                            message: `Operation "${selected.operationId}" is not classified read-only; self-test refuses to execute it.`,
-                        },
-                        startedAt: now,
-                        finishedAt: now,
-                    });
-                    continue;
+                finally {
+                    runSignal.removeEventListener("abort", abortFromRun);
+                    caseController.abort();
                 }
-                results.push(await executeSelfTestCase(execution, selected.operationId, selected.suite, selected.healthCase));
+            }
+            if (interruptedCase) {
+                logSelfTestCancellation(provider, options, request.requestId, interruptedCase, runSignal, execution.sensitiveValues);
             }
             const singleCase = request.operationId !== undefined && request.caseName !== undefined;
             const response = {