fix(server): improve self hosted usability (#15510)

fix #15505 fix #15502 fix #15496 fix #15491 #### PR Dependency Tree * **PR #15510** 👈 This tree was auto-generated by [Charcoal](https://github.com/danerwilliams/charcoal)  ## Summary by CodeRabbit - **New Features** - Added configurable delays for invitations, invite links, and document publishing by newly created accounts. - Added workspace action checks that explain blocked actions and retry timing. - BYOK setup now verifies model capabilities and saves only validated options. - **Bug Fixes** - Improved BYOK probing for chat, structured responses, tool calls, embeddings, reranking, and image generation. - Preserved probe request order and strengthened response validation. - Authentication configuration changes now reload correctly.
2026-08-24 13:17:10 +08:00 · 2026-08-22 04:47:21 +08:00
parent 591f874dad
commit dec1a01449
22 changed files with 683 additions and 279 deletions
@@ -34,6 +34,7 @@ export declare class BackendRuntime {
  claimInviteAbuseAction(actionId: string, workerId: string): Promise<boolean>
  claimRetryableInviteAbuseActions(workerId: string, limit: number): Promise<Array<RuntimeInviteAbuseClaimedAction>>
  markInviteAbuseAction(actionId: string, workerId: string, status: string, error?: string | undefined | null): Promise<boolean>
+  evaluateWorkspaceActionV1(actorUserId: string, workspaceId: string): Promise<RuntimeWorkspaceActionDecision>
  assertWorkspaceInviteQuotaV1(input: RuntimeWorkspaceInviteQuotaInput): Promise<RuntimeWorkspaceInviteQuotaDecision>
  commitWorkspaceInviteQuotaV1(reservationId: string, usage: RuntimeWorkspaceInviteQuotaUsage): Promise<boolean>
  releaseWorkspaceInviteQuotaV1(reservationId: string): Promise<boolean>
@@ -1496,6 +1497,12 @@ export interface RuntimeVerificationTokenRecord {
  expiresAtMs: number
 }

+export interface RuntimeWorkspaceActionDecision {
+  allowed: boolean
+  retryAfterSeconds?: number
+  reason?: string
+}
+
 export interface RuntimeWorkspaceArtifact {
  id: string
  workspaceId: string
@@ -6,10 +6,10 @@ use llm_adapter::{
    AttachmentKind, AttachmentSource, ModelFeature, ModelInput, ModelOutput, ModelRequirements, declared_model_matches,
  },
  core::{
-    CoreContent, CoreMessage, CoreRequest, CoreRole, CoreToolDefinition, EmbeddingRequest, ImageOptions,
-    ImageProviderOptions, ImageRequest, RerankCandidate, RerankRequest, StructuredRequest,
+    CoreContent, CoreMessage, CoreRequest, CoreRole, CoreToolChoice, CoreToolDefinition, EmbeddingRequest,
+    ImageOptions, ImageProviderOptions, ImageRequest, RerankCandidate, RerankRequest, StructuredRequest,
  },
-  router::{ExecutablePreparedRoute, ExecutableRequest, dispatch_prepared_route},
+  router::{ExecutablePreparedRoute, ExecutableRequest, ExecutableResponse, dispatch_prepared_route},
  target::{
    BackendCredential, BackendEndpoint, BackendOperation, BackendTargetInput, EgressPolicy, compile_backend_target,
  },
@@ -30,9 +30,11 @@ pub(super) async fn execute_probe(
  checks: Vec<ByokProbeCheckInput>,
 ) -> RuntimeResult<ByokProbeResultOutput> {
  let tested_at_ms = chrono::Utc::now().timestamp_millis();
-  let mut requested = HashSet::new();
+  let mut requested = Vec::new();
+  let mut requested_set = HashSet::new();
  for check in checks {
-    if !requested.insert((check.model_id.clone(), check.operation.clone())) {
+    let key = (check.model_id.clone(), check.operation.clone());
+    if !requested_set.insert(key.clone()) {
      return Err(RuntimeError::invalid_input("duplicate BYOK probe check"));
    }
    if !matches!(
@@ -41,6 +43,7 @@ pub(super) async fn execute_probe(
    ) {
      return Err(RuntimeError::invalid_input("unknown BYOK probe operation"));
    }
+    requested.push(key);
  }

  let mut models = Vec::new();
@@ -160,7 +163,49 @@ fn dispatch_check(
    Err(_) => return failed(checked_at, "invalid_probe_request"),
  };
  match dispatch_prepared_route(&DefaultHttpClient::default(), &route) {
-    Ok(_) => verified(checked_at),
+    Ok(ExecutableResponse::Chat(response)) => {
+      let valid = if operation == "tool_calling" {
+        response
+          .message
+          .content
+          .iter()
+          .any(|content| matches!(content, CoreContent::ToolCall { name, .. } if name == "byok_probe"))
+      } else {
+        response
+          .message
+          .content
+          .iter()
+          .any(|content| matches!(content, CoreContent::Text { text } if !text.trim().is_empty()))
+      };
+      if valid {
+        verified(checked_at)
+      } else {
+        failed(checked_at, "invalid_response")
+      }
+    }
+    Ok(ExecutableResponse::Structured(response)) => {
+      let valid = response.output_json.as_ref().is_some_and(|output| {
+        let ExecutableRequest::Structured(request) = &route.request else {
+          return false;
+        };
+        llm_adapter::schema::validate_json_schema(&request.schema, output).is_ok()
+      });
+      if valid {
+        verified(checked_at)
+      } else {
+        failed(checked_at, "invalid_response")
+      }
+    }
+    Ok(ExecutableResponse::Embedding(response)) if operation == "embedding" && !response.embeddings.is_empty() => {
+      verified(checked_at)
+    }
+    Ok(ExecutableResponse::Rerank(response)) if operation == "rerank" && !response.scores.is_empty() => {
+      verified(checked_at)
+    }
+    Ok(ExecutableResponse::Image(response)) if operation == "image" && !response.images.is_empty() => {
+      verified(checked_at)
+    }
+    Ok(_) => failed(checked_at, "invalid_response"),
    Err(error) => failed(checked_at, backend_error_kind(&error)),
  }
 }
@@ -169,7 +214,11 @@ fn probe_request_for_operation(operation: &str) -> ExecutableRequest {
  let message = CoreMessage {
    role: CoreRole::User,
    content: vec![CoreContent::Text {
-      text: "Reply with OK.".to_string(),
+      text: match operation {
+        "tool_calling" => "Call the byok_probe tool.".to_string(),
+        "structured" => "Return exactly {\"ok\":true}.".to_string(),
+        _ => "Reply with OK.".to_string(),
+      },
    }],
  };
  match operation {
@@ -177,8 +226,8 @@ fn probe_request_for_operation(operation: &str) -> ExecutableRequest {
      model: String::new(),
      messages: vec![message],
      stream: false,
-      max_tokens: Some(8),
-      temperature: Some(0.0),
+      max_tokens: Some(64),
+      temperature: None,
      tools: if operation == "tool_calling" {
        vec![CoreToolDefinition {
          name: "byok_probe".to_string(),
@@ -188,7 +237,9 @@ fn probe_request_for_operation(operation: &str) -> ExecutableRequest {
      } else {
        vec![]
      },
-      tool_choice: None,
+      tool_choice: (operation == "tool_calling").then_some(CoreToolChoice::Specific {
+        name: "byok_probe".to_string(),
+      }),
      include: None,
      reasoning: None,
      response_schema: None,
@@ -202,8 +253,8 @@ fn probe_request_for_operation(operation: &str) -> ExecutableRequest {
        "required": ["ok"],
        "additionalProperties": false
      }),
-      max_tokens: Some(16),
-      temperature: Some(0.0),
+      max_tokens: Some(128),
+      temperature: None,
      reasoning: None,
      strict: Some(true),
      response_mime_type: Some("application/json".to_string()),
@@ -426,7 +477,48 @@ mod tests {
        let mut stream = stream.unwrap();
        let request = read_request(&mut stream);
        let responses = request.starts_with("POST /v1/responses ");
-        let body = if responses {
+        let embedding = request.starts_with("POST /v1/embeddings ");
+        let image = request.starts_with("POST /v1/images/generations ");
+        let rerank = request.contains("\"logprobs\":true");
+        let tool_calling = request.contains("byok_probe");
+        let body = if embedding {
+          json!({
+            "model": "smoke-model",
+            "data": [{ "embedding": [0.1], "index": 0 }],
+            "usage": { "prompt_tokens": 1, "total_tokens": 1 }
+          })
+        } else if image {
+          json!({
+            "created": 0,
+            "data": [{ "url": "https://example.com/smoke.png" }]
+          })
+        } else if rerank {
+          json!({
+            "model": "smoke-model",
+            "choices": [{
+              "logprobs": { "content": [{
+                "top_logprobs": [
+                  { "token": "Yes", "logprob": 0.0 },
+                  { "token": "No", "logprob": -1.0 }
+                ]
+              }] }
+            }]
+          })
+        } else if responses && tool_calling {
+          json!({
+            "id": "resp_smoke",
+            "model": "smoke-model",
+            "status": "completed",
+            "output": [{
+              "type": "function_call",
+              "id": "fc_smoke",
+              "call_id": "call_smoke",
+              "name": "byok_probe",
+              "arguments": "{}"
+            }],
+            "usage": { "input_tokens": 1, "output_tokens": 1, "total_tokens": 2 }
+          })
+        } else if responses {
          json!({
            "id": "resp_smoke",
            "model": "smoke-model",
@@ -439,6 +531,25 @@ mod tests {
            }],
            "usage": { "input_tokens": 1, "output_tokens": 1, "total_tokens": 2 }
          })
+        } else if tool_calling {
+          json!({
+            "id": "chat_smoke",
+            "model": "smoke-model",
+            "choices": [{
+              "index": 0,
+              "message": {
+                "role": "assistant",
+                "content": null,
+                "tool_calls": [{
+                  "id": "call_smoke",
+                  "type": "function",
+                  "function": { "name": "byok_probe", "arguments": "{}" }
+                }]
+              },
+              "finish_reason": "tool_calls"
+            }],
+            "usage": { "prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2 }
+          })
        } else {
          json!({
            "id": "chat_smoke",
@@ -489,7 +600,7 @@ mod tests {

  #[test]
  fn openai_compatible_probe_smoke_uses_the_selected_dialect() {
-    let operations = ["chat", "structured", "tool_calling"];
+    let operations = ["chat", "structured", "tool_calling", "embedding", "rerank", "image"];
    let (endpoint, requests, server) = serve_openai_compatible(operations.len() * 2);

    for dialect in [OpenAiDialect::Responses, OpenAiDialect::ChatCompletions] {
@@ -520,19 +631,39 @@ mod tests {
        .iter()
        .filter(|request| request.starts_with("POST /v1/responses "))
        .count(),
-      operations.len()
+      3
    );
    assert_eq!(
      requests
        .iter()
        .filter(|request| request.starts_with("POST /v1/chat/completions "))
        .count(),
-      operations.len()
+      5
+    );
+    assert_eq!(
+      requests
+        .iter()
+        .filter(|request| request.starts_with("POST /v1/embeddings "))
+        .count(),
+      2
+    );
+    assert_eq!(
+      requests
+        .iter()
+        .filter(|request| request.starts_with("POST /v1/images/generations "))
+        .count(),
+      2
    );
    assert!(requests.iter().all(|request| !request.contains("/models")));
    assert_eq!(
      requests.iter().filter(|request| request.contains("byok_probe")).count(),
      2
    );
+    assert!(
+      requests
+        .iter()
+        .filter(|request| !request.contains("\"logprobs\":true"))
+        .all(|request| !request.contains("\"temperature\""))
+    );
  }
 }
@@ -5,7 +5,7 @@ use super::{
  BackendRuntime, RuntimeError, RuntimeInviteAbuseClaimedAction, RuntimeResult, napi_error, workspace_subject_key,
 };

-async fn invite_abuse_user_quarantined_or_banned(pool: &PgPool, user_id: &str) -> RuntimeResult<bool> {
+pub(super) async fn invite_abuse_user_quarantined_or_banned(pool: &PgPool, user_id: &str) -> RuntimeResult<bool> {
  let row: Option<i32> = sqlx::query_scalar(
    r#"
    SELECT 1
@@ -22,7 +22,7 @@ async fn invite_abuse_user_quarantined_or_banned(pool: &PgPool, user_id: &str) -
  Ok(row.is_some())
 }

-async fn invite_abuse_workspace_quarantined(pool: &PgPool, workspace_id: &str) -> RuntimeResult<bool> {
+pub(super) async fn invite_abuse_workspace_quarantined(pool: &PgPool, workspace_id: &str) -> RuntimeResult<bool> {
  let row: Option<i32> = sqlx::query_scalar(
    r#"
    SELECT 1
@@ -6,6 +6,7 @@ mod reservation;
 mod workspace_invite;
 mod workspace_invite_policy;

+use invite_abuse_actions::{invite_abuse_user_quarantined_or_banned, invite_abuse_workspace_quarantined};
 use mail_delivery::{build_mail_scopes, decision_from_violation as mail_decision_from_violation, mail_class};
 use napi::Result;
 use reservation::{
@@ -15,8 +16,8 @@ use reservation::{
 use sha2::{Digest, Sha256};
 use workspace_invite_policy::{
  ActorFacts, InviteAbuseDecision, InviteActivityFacts, QuotaFacts, WorkspaceFacts, build_invite_scopes,
-  evaluate_projection, high_confidence_invite_abuse, invite_commit_usage_for_scope, source_cohort_subject_key,
-  subject_hash, sum_domains,
+  evaluate_projection, high_confidence_invite_abuse, invite_commit_usage_for_scope, new_account_action_retry_after,
+  source_cohort_subject_key, subject_hash, sum_domains,
 };

 #[cfg(test)]
@@ -26,7 +27,8 @@ pub(super) use super::{
  types::{
    RuntimeInviteAbuseActionRequired, RuntimeInviteAbuseClaimedAction, RuntimeMailDeliveryQuotaDecision,
    RuntimeMailDeliveryQuotaInput, RuntimeQuotaSourceInput, RuntimeQuotaTargetDomainInput,
-    RuntimeWorkspaceInviteQuotaDecision, RuntimeWorkspaceInviteQuotaInput, RuntimeWorkspaceInviteQuotaUsage,
+    RuntimeWorkspaceActionDecision, RuntimeWorkspaceInviteQuotaDecision, RuntimeWorkspaceInviteQuotaInput,
+    RuntimeWorkspaceInviteQuotaUsage,
  },
 };

@@ -5,19 +5,32 @@ use sqlx::{PgPool, Row};

 use super::{
  ActorFacts, BackendRuntime, InviteAbuseDecision, InviteActivityFacts, InviteQuotaConfig, QuotaFacts, QuotaViolation,
-  RuntimeError, RuntimeInviteAbuseActionRequired, RuntimeResult, RuntimeWorkspaceInviteQuotaDecision,
-  RuntimeWorkspaceInviteQuotaInput, RuntimeWorkspaceInviteQuotaUsage, WorkspaceFacts, build_invite_scopes,
-  commit_reservation, evaluate_projection, high_confidence_invite_abuse, invite_commit_usage_for_scope, napi_error,
-  normalize_domain, release_reservation, reserve_scopes, short_hash, source_cohort_subject_key, source_prefix,
-  subject_hash, sum_domains, workspace_subject_key,
+  RuntimeError, RuntimeInviteAbuseActionRequired, RuntimeResult, RuntimeWorkspaceActionDecision,
+  RuntimeWorkspaceInviteQuotaDecision, RuntimeWorkspaceInviteQuotaInput, RuntimeWorkspaceInviteQuotaUsage,
+  WorkspaceFacts, build_invite_scopes, commit_reservation, evaluate_projection, high_confidence_invite_abuse,
+  invite_abuse_user_quarantined_or_banned, invite_abuse_workspace_quarantined, invite_commit_usage_for_scope,
+  napi_error, new_account_action_retry_after, normalize_domain, release_reservation, reserve_scopes, short_hash,
+  source_cohort_subject_key, source_prefix, subject_hash, sum_domains, workspace_subject_key,
 };

 async fn load_actor(pool: &PgPool, user_id: &str) -> RuntimeResult<ActorFacts> {
  let row = sqlx::query(
    r#"
-    SELECT email, created_at, registered, email_verified IS NOT NULL AS email_verified, disabled
+    SELECT
+      users.email,
+      users.created_at,
+      users.registered,
+      users.email_verified IS NOT NULL AS email_verified,
+      users.disabled,
+      CASE
+        WHEN quota.known
+          AND NOT quota.stale
+          AND (quota.stale_after IS NULL OR quota.stale_after > clock_timestamp())
+        THEN quota.plan
+      END AS quota_plan
    FROM users
-    WHERE id = $1
+    LEFT JOIN effective_user_quota_states quota ON quota.user_id = users.id
+    WHERE users.id = $1
    "#,
  )
  .bind(user_id)
@@ -32,6 +45,7 @@ async fn load_actor(pool: &PgPool, user_id: &str) -> RuntimeResult<ActorFacts> {
    registered: row.get("registered"),
    email_verified: row.get("email_verified"),
    disabled: row.get("disabled"),
+    quota_plan: row.get("quota_plan"),
  })
 }

@@ -285,6 +299,49 @@ fn decision_from_violation(violation: QuotaViolation, reason: &str) -> RuntimeWo

 #[napi_derive::napi]
 impl BackendRuntime {
+  #[napi]
+  pub async fn evaluate_workspace_action_v1(
+    &self,
+    actor_user_id: String,
+    workspace_id: String,
+  ) -> Result<RuntimeWorkspaceActionDecision> {
+    let runtime_config = self.config()?;
+    let pool = self.pool().await?;
+    if invite_abuse_user_quarantined_or_banned(&pool, &actor_user_id).await? {
+      return Ok(RuntimeWorkspaceActionDecision {
+        allowed: false,
+        retry_after_seconds: None,
+        reason: Some("abuse_subject".to_string()),
+      });
+    }
+    if invite_abuse_workspace_quarantined(&pool, &workspace_id).await? {
+      return Ok(RuntimeWorkspaceActionDecision {
+        allowed: false,
+        retry_after_seconds: None,
+        reason: Some("abuse_workspace".to_string()),
+      });
+    }
+    let now: DateTime<Utc> = sqlx::query_scalar("SELECT clock_timestamp()")
+      .fetch_one(&pool)
+      .await
+      .map_err(|err| RuntimeError::database("failed to read database clock", err))?;
+    let actor = load_actor(&pool, &actor_user_id).await?;
+    let quota = load_quota(&pool, &workspace_id).await?;
+    let current_quota = quota.as_ref().filter(|quota| evaluate_projection(quota, now).is_none());
+    let retry_after_seconds = new_account_action_retry_after(
+      runtime_config.deployment,
+      &runtime_config.invite_quota,
+      &actor,
+      current_quota,
+      now,
+    );
+    Ok(RuntimeWorkspaceActionDecision {
+      allowed: retry_after_seconds.is_none(),
+      retry_after_seconds,
+      reason: retry_after_seconds.map(|_| "new_account_action_delay".to_string()),
+    })
+  }
+
  #[napi]
  pub async fn assert_workspace_invite_quota_v1(
    &self,
@@ -388,6 +445,22 @@ impl BackendRuntime {
        action_required: None,
      });
    }
+    if let Some(retry_after_seconds) =
+      new_account_action_retry_after(runtime_config.deployment, config, &actor, Some(&quota), now)
+    {
+      return Ok(RuntimeWorkspaceInviteQuotaDecision {
+        allowed: false,
+        reservation_id: None,
+        retry_after_seconds: Some(retry_after_seconds),
+        reason: Some("new_account_action_delay".to_string()),
+        scope_key: None,
+        window_seconds: None,
+        limit: None,
+        current: None,
+        requested: Some(input.target_count),
+        action_required: None,
+      });
+    }
    if let Some(abuse_decision) = high_confidence_invite_abuse(&input, &actor, config) {
      let reason = abuse_decision.reason;
      let scope_key = match abuse_decision.subject_kind {
@@ -10,6 +10,7 @@ use super::{
  InviteQuotaConfig, RuntimeQuotaTargetDomainInput, RuntimeWorkspaceInviteQuotaInput, ScopeLimit, bucket_seconds,
  high_risk_domain, napi_error, normalize_domain, scope, short_hash, source_prefix, workspace_subject_key,
 };
+use crate::llm::Deployment;

 #[derive(Clone, Debug)]
 pub(super) struct ActorFacts {
@@ -18,6 +19,7 @@ pub(super) struct ActorFacts {
  pub(super) registered: bool,
  pub(super) email_verified: bool,
  pub(super) disabled: bool,
+  pub(super) quota_plan: Option<String>,
 }

 #[derive(Clone, Debug)]
@@ -108,9 +110,6 @@ fn base_invite_limits(
  let mut per_day = 15;
  let mut per_week = 30;

-  if account_age < Duration::hours(24) {
-    return (0, 0, 0, 0);
-  }
  if !actor.email_verified {
    single = 1;
    per_hour = 1;
@@ -185,6 +184,31 @@ pub(super) fn evaluate_projection(quota: &QuotaFacts, now: DateTime<Utc>) -> Opt
  None
 }

+pub(super) fn new_account_action_retry_after(
+  deployment: Deployment,
+  config: &InviteQuotaConfig,
+  actor: &ActorFacts,
+  workspace_quota: Option<&QuotaFacts>,
+  now: DateTime<Utc>,
+) -> Option<i32> {
+  if deployment == Deployment::SelfHosted || config.new_account_action_delay_seconds <= 0 {
+    return None;
+  }
+  if actor
+    .quota_plan
+    .as_deref()
+    .is_some_and(|plan| matches!(plan, "pro" | "lifetime_pro" | "ai"))
+    || workspace_quota.map(|quota| quota.plan.as_str()).is_some_and(|plan| {
+      matches!(plan, "pro" | "lifetime_pro" | "ai") || plan.contains("team") && !plan.contains("trial")
+    })
+  {
+    return None;
+  }
+
+  let remaining = config.new_account_action_delay_seconds - (now - actor.created_at).num_seconds();
+  (remaining > 0).then(|| remaining.min(i64::from(i32::MAX)) as i32)
+}
+
 pub(super) fn build_invite_scopes(
  input: &RuntimeWorkspaceInviteQuotaInput,
  actor: &ActorFacts,
@@ -442,6 +466,7 @@ mod tests {
      registered: true,
      email_verified: true,
      disabled: false,
+      quota_plan: None,
    }
  }

@@ -467,7 +492,7 @@ mod tests {
  }

  #[test]
-  fn seat_based_weekly_limit_binds_paid_team_and_high_risk_domain() {
+  fn invite_scopes_apply_plan_ceiling_domain_risk_and_graduated_limits() {
    let now = Utc.with_ymd_and_hms(2026, 7, 6, 0, 0, 0).single().unwrap();
    let input = RuntimeWorkspaceInviteQuotaInput {
      actor_user_id: "u1".to_string(),
@@ -500,6 +525,67 @@ mod tests {
      .find(|scope| scope.scope_key == "invite:quota_subject_domain:workspace:w1:qq.com")
      .unwrap();
    assert_eq!(high_risk.limit, 5);
+
+    let fresh_actor_scopes = build_invite_scopes(
+      &input,
+      &user(now - Duration::hours(1)),
+      &workspace(now - Duration::hours(1)),
+      &quota("paid_team", 10),
+      &InviteActivityFacts::default(),
+      &invite_config(),
+      now,
+    )
+    .unwrap();
+    assert_eq!(fresh_actor_scopes[0].limit, 3);
+
+    let mut fresh_actor = user(now - Duration::hours(1));
+    assert_eq!(
+      new_account_action_retry_after(
+        Deployment::Cloud,
+        &invite_config(),
+        &fresh_actor,
+        Some(&quota("free", 3)),
+        now,
+      ),
+      Some(23 * 60 * 60)
+    );
+    assert_eq!(
+      new_account_action_retry_after(
+        Deployment::SelfHosted,
+        &invite_config(),
+        &fresh_actor,
+        Some(&quota("free", 3)),
+        now,
+      ),
+      None
+    );
+    assert_eq!(
+      new_account_action_retry_after(
+        Deployment::Cloud,
+        &invite_config(),
+        &fresh_actor,
+        Some(&quota("paid_team", 10)),
+        now,
+      ),
+      None
+    );
+    let mut no_delay = invite_config();
+    no_delay.new_account_action_delay_seconds = 0;
+    assert_eq!(
+      new_account_action_retry_after(Deployment::Cloud, &no_delay, &fresh_actor, Some(&quota("free", 3)), now,),
+      None
+    );
+    fresh_actor.quota_plan = Some("pro".to_string());
+    assert_eq!(
+      new_account_action_retry_after(
+        Deployment::Cloud,
+        &invite_config(),
+        &fresh_actor,
+        Some(&quota("free", 3)),
+        now,
+      ),
+      None
+    );
  }

  #[test]
@@ -319,6 +319,7 @@ impl TryFrom<CopilotManagedProfileConfigFile> for CopilotManagedProfileConfig {

 #[derive(Clone, Debug)]
 pub(crate) struct InviteQuotaConfig {
+  pub(crate) new_account_action_delay_seconds: i64,
  pub(crate) high_risk_target_domains: Vec<String>,
  pub(crate) subject_hash_salt: String,
  pub(crate) mail_class_mapping: BTreeMap<String, String>,
@@ -327,6 +328,7 @@ pub(crate) struct InviteQuotaConfig {
 impl Default for InviteQuotaConfig {
  fn default() -> Self {
    Self {
+      new_account_action_delay_seconds: 24 * 60 * 60,
      high_risk_target_domains: [
        "qq.com",
        "proton.me",
@@ -479,12 +481,19 @@ fn deployment_from_env() -> Deployment {

 #[derive(Default, Deserialize)]
 struct AppConfigFile {
+  auth: Option<AuthConfigFile>,
  db: Option<DbConfigFile>,
  crypto: Option<CryptoConfigFile>,
  copilot: Option<CopilotRuntimeConfigFile>,
  indexer: Option<SearchRuntimeConfigFile>,
 }

+#[derive(Default, Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct AuthConfigFile {
+  new_account_action_delay: Option<i64>,
+}
+
 #[derive(Default, Deserialize)]
 #[serde(rename_all = "camelCase", default)]
 struct SearchRuntimeConfigFile {
@@ -542,7 +551,11 @@ impl AppConfigFile {
  }

  fn invite_quota_config(&self) -> InviteQuotaConfig {
-    InviteQuotaConfig::default()
+    let mut config = InviteQuotaConfig::default();
+    if let Some(delay) = self.auth.as_ref().and_then(|auth| auth.new_account_action_delay) {
+      config.new_account_action_delay_seconds = delay.max(0);
+    }
+    config
  }
 }

@@ -985,14 +998,16 @@ mod tests {
  }

  #[test]
-  fn invite_quota_policy_is_internal_not_app_configurable() {
+  fn invite_abuse_policy_is_internal_while_action_delay_is_configurable() {
    let app_config = app_config_from_flat_overrides([
+      ("auth.newAccountActionDelay", serde_json::json!(123)),
      ("auth.untrustedPolicyOverride", serde_json::json!("runtime-salt-v2")),
      ("auth.untrustedDomainList", serde_json::json!(["Example.COM."])),
    ])
    .unwrap();

    let config = app_config.invite_quota_config();
+    assert_eq!(config.new_account_action_delay_seconds, 123);
    assert!(!config.high_risk_target_domains.contains(&"example.com".to_string()));
    assert_ne!(config.subject_hash_salt, "runtime-salt-v2");
    assert_eq!(
@@ -263,6 +263,13 @@ pub struct RuntimeWorkspaceInviteQuotaUsage {
  pub target_domains: Vec<RuntimeQuotaTargetDomainInput>,
 }

+#[napi_derive::napi(object)]
+pub struct RuntimeWorkspaceActionDecision {
+  pub allowed: bool,
+  pub retry_after_seconds: Option<i32>,
+  pub reason: Option<String>,
+}
+
 #[napi_derive::napi(object)]
 pub struct RuntimeInviteAbuseActionRequired {
  pub action: String,
@@ -18,7 +18,7 @@ export interface AuthConfig {
  allowSignupForOauth: boolean;
  requireEmailDomainVerification: boolean;
  requireEmailVerification: boolean;
-  newAccountShareActionDelay: number;
+  newAccountActionDelay: number;
  trustedCloudflareHeaders: boolean;
  signInRateLimit: ConfigItem<{
    ttl: number;
@@ -56,8 +56,8 @@ defineModuleConfig('auth', {
    desc: 'Whether require email verification before accessing restricted resources(not implemented).',
    default: true,
  },
-  newAccountShareActionDelay: {
-    desc: 'Minimum account age in seconds before new accounts can invite members or create share links.',
+  newAccountActionDelay: {
+    desc: 'Minimum account age in seconds before new accounts can invite members, create invite links, or publish documents. Set to 0 to disable.',
    default: 24 * 60 * 60,
    shape: z.number().int().min(0),
  },
@@ -121,6 +121,12 @@ export type RuntimeWorkspaceInviteQuotaUsage = {
  targetDomains: RuntimeQuotaTargetDomainInput[];
 };

+export type RuntimeWorkspaceActionDecision = {
+  allowed: boolean;
+  retryAfterSeconds?: number;
+  reason?: string;
+};
+
 export type RuntimeInviteAbuseAction =
  | 'ban_actor'
  | 'quarantine_actor'
@@ -213,6 +219,10 @@ export type RuntimeMailDeliveryQuotaDecision = {
 };

 type RuntimeQuotaMethods = RuntimeInstance & {
+  evaluateWorkspaceActionV1(
+    actorUserId: string,
+    workspaceId: string
+  ): Promise<RuntimeWorkspaceActionDecision>;
  assertWorkspaceInviteQuotaV1(
    input: RuntimeWorkspaceInviteQuotaInput
  ): Promise<NativeRuntimeWorkspaceInviteQuotaDecision>;
@@ -328,6 +338,7 @@ export class BackendRuntimeProvider
      !updates.copilot &&
      !updates.crypto &&
      !updates.db &&
+      !updates.auth &&
      !updates.indexer &&
      !updates.storages
    ) {
@@ -537,6 +548,12 @@ export class BackendRuntimeProvider
    );
  }

+  async evaluateWorkspaceActionV1(actorUserId: string, workspaceId: string) {
+    return await this.measured('evaluateWorkspaceActionV1', rt =>
+      this.quotaRuntime(rt).evaluateWorkspaceActionV1(actorUserId, workspaceId)
+    );
+  }
+
  async commitWorkspaceInviteQuotaV1(
    reservationId: string,
    usage: RuntimeWorkspaceInviteQuotaUsage
@@ -15,6 +15,9 @@ import { Mockers } from '../../../__tests__/mocks';
 import { Config } from '../../../base';
 import { ActionForbidden, TooManyRequest } from '../../../base/error';
 import { Models, WorkspaceRole } from '../../../models';
+import { BackendRuntimeProvider } from '../../backend-runtime';
+import { EntitlementService } from '../../entitlement';
+import { QuotaService } from '../../quota';
 import {
  getAbuseRequestSource,
  InviteAbuseDispositionService,
@@ -23,6 +26,7 @@ import {

 let app: TestingApp;
 const quota = {
+  assertWorkspaceActionAllowed: Sinon.stub(),
  assertWorkspaceInviteQuota: Sinon.stub(),
  commitWorkspaceInviteQuota: Sinon.stub(),
  releaseWorkspaceInviteQuota: Sinon.stub(),
@@ -41,6 +45,7 @@ test.before(async () => {
 });

 test.beforeEach(() => {
+  quota.assertWorkspaceActionAllowed.reset();
  quota.assertWorkspaceInviteQuota.reset();
  quota.commitWorkspaceInviteQuota.reset();
  quota.releaseWorkspaceInviteQuota.reset();
@@ -345,8 +350,8 @@ test('workspace quarantine blocks invite link creation', async t => {
      updated_at = now()
  `;

-  const previousDelay = config.auth.newAccountShareActionDelay;
-  config.auth.newAccountShareActionDelay = 0;
+  const previousDelay = config.auth.newAccountActionDelay;
+  config.auth.newAccountActionDelay = 0;
  try {
    await app.login(owner);
    await t.throwsAsync(
@@ -359,32 +364,58 @@ test('workspace quarantine blocks invite link creation', async t => {
      })
    );
  } finally {
-    config.auth.newAccountShareActionDelay = previousDelay;
+    config.auth.newAccountActionDelay = previousDelay;
  }
 });

-test('domain workspace name blocks invite link creation', async t => {
-  const config = app.get(Config);
+test('workspace action admission applies exemption before content policy', async t => {
+  const db = app.get(PrismaClient);
+  const inviteQuota = new InviteQuotaAssertService(
+    app.get(Config),
+    app.get(QuotaService),
+    app.get(BackendRuntimeProvider),
+    app.get(InviteAbuseDispositionService)
+  );
  const owner = await app.create(Mockers.User);
+  await db.user.update({
+    where: { id: owner.id },
+    data: { createdAt: new Date() },
+  });
  const workspace = await app.create(Mockers.Workspace, {
    owner,
    name: 'Join example.com',
  });

-  const previousDelay = config.auth.newAccountShareActionDelay;
-  config.auth.newAccountShareActionDelay = 0;
-  try {
-    await app.login(owner);
-    await t.throwsAsync(
-      app.gql({
-        query: createInviteLinkMutation,
-        variables: {
-          workspaceId: workspace.id,
-          expireTime: WorkspaceInviteLinkExpireTime.OneDay,
-        },
-      })
-    );
-  } finally {
-    config.auth.newAccountShareActionDelay = previousDelay;
-  }
+  await t.throwsAsync(
+    inviteQuota.assertWorkspaceActionAllowed({
+      actorUserId: owner.id,
+      workspaceId: workspace.id,
+      action: 'inviteMember',
+    }),
+    { instanceOf: ActionForbidden }
+  );
+
+  await app.get(EntitlementService).upsertAdminGrant({
+    targetType: 'user',
+    targetId: owner.id,
+    plan: 'pro',
+  });
+  await t.notThrowsAsync(
+    inviteQuota.assertWorkspaceActionAllowed({
+      actorUserId: owner.id,
+      workspaceId: workspace.id,
+      action: 'inviteMember',
+    })
+  );
+
+  await app.login(owner);
+  await t.throwsAsync(
+    app.gql({
+      query: createInviteLinkMutation,
+      variables: {
+        workspaceId: workspace.id,
+        expireTime: WorkspaceInviteLinkExpireTime.OneDay,
+      },
+    })
+  );
 });
@@ -39,14 +39,6 @@ declare global {
  }
 }

-export function canUserExecuteLimitedActions(
-  user: { createdAt: Date },
-  minimumAccountAgeMs: number
-) {
-  if (minimumAccountAgeMs <= 0) return true;
-  return Date.now() - user.createdAt.getTime() >= minimumAccountAgeMs;
-}
-
 function parseAsn(value: string | undefined) {
  if (!value) {
    return;
@@ -237,6 +229,28 @@ export class InviteQuotaAssertService {
    private readonly disposition: InviteAbuseDispositionService
  ) {}

+  async assertWorkspaceActionAllowed(input: {
+    actorUserId: string;
+    workspaceId: string;
+    action: 'inviteMember' | 'createInviteLink' | 'publishDoc';
+    docId?: string;
+  }) {
+    const decision = await this.runtime.evaluateWorkspaceActionV1(
+      input.actorUserId,
+      input.workspaceId
+    );
+    if (decision.allowed) return;
+
+    this.logger.warn('Workspace action rejected', {
+      ...input,
+      reason: decision.reason,
+      retryAfter: decision.retryAfterSeconds,
+    });
+    throw new ActionForbidden(
+      'This feature is temporarily unavailable for you.'
+    );
+  }
+
  async assertWorkspaceInviteQuota(input: {
    actorUserId: string;
    workspaceId: string;
@@ -384,7 +398,11 @@ export class InviteQuotaAssertService {
  private mapDecision(
    decision: RuntimeWorkspaceInviteQuotaDecision
  ): UserFriendlyError {
-    if (decision.reason === 'abuse_subject' || decision.actionRequired) {
+    if (
+      decision.reason === 'abuse_subject' ||
+      decision.reason === 'new_account_action_delay' ||
+      decision.actionRequired
+    ) {
      return new ActionForbidden('This feature is temporarily unavailable.');
    }
    return new TooManyRequest();
@@ -15,9 +15,7 @@ import { Prisma, PrismaClient } from '@prisma/client';
 import { SafeIntResolver } from 'graphql-scalars';

 import {
-  ActionForbidden,
  Cache,
-  Config,
  DocActionDenied,
  DocDefaultRoleCanNotBeOwner,
  DocNotFound,
@@ -46,7 +44,7 @@ import {
  PermissionAccess,
 } from '../../permission';
 import { PublicUserType, WorkspaceUserType } from '../../user';
-import { canUserExecuteLimitedActions } from '../abuse';
+import { InviteQuotaAssertService } from '../abuse';
 import { DocGrantsService } from '../doc-grants';
 import { WorkspaceType } from '../types';
 import { TimeBucket, TimeWindow } from './analytics-types';
@@ -302,51 +300,10 @@ export class WorkspaceDocResolver {
    private readonly models: Models,
    private readonly cache: Cache,
    private readonly event: EventBus,
-    private readonly config: Config,
-    private readonly runtime: BackendRuntimeProvider
+    private readonly runtime: BackendRuntimeProvider,
+    private readonly inviteQuota: InviteQuotaAssertService
  ) {}

-  private async assertCanShare(
-    userId: string,
-    context: { workspaceId: string; docId: string; action: 'publishDoc' }
-  ) {
-    if (await this.runtime.isInviteAbuseUserQuarantinedOrBanned(userId)) {
-      this.logger.warn('Share action blocked for quarantined actor', {
-        userId,
-        ...context,
-      });
-      throw new ActionForbidden(
-        'This feature is temporarily unavailable for you.'
-      );
-    }
-    if (
-      await this.runtime.isInviteAbuseWorkspaceQuarantined(context.workspaceId)
-    ) {
-      this.logger.warn('Share action blocked for quarantined workspace', {
-        userId,
-        ...context,
-      });
-      throw new ActionForbidden(
-        'This feature is temporarily unavailable for you.'
-      );
-    }
-    const user = await this.models.user.get(userId);
-    const newAccountAgeMs = this.config.auth.newAccountShareActionDelay * 1000;
-    if (!user || !canUserExecuteLimitedActions(user, newAccountAgeMs)) {
-      this.logger.warn('Share action blocked for new account', {
-        userId,
-        email: user?.email,
-        createdAt: user?.createdAt,
-        accountAgeMs: user ? Date.now() - user.createdAt.getTime() : null,
-        minimumAccountAgeMs: newAccountAgeMs,
-        ...context,
-      });
-      throw new ActionForbidden(
-        'This feature is temporarily unavailable for you.'
-      );
-    }
-  }
-
  @ResolveField(() => WorkspaceDocMeta, {
    description: 'Cloud page metadata of workspace',
    complexity: 2,
@@ -475,7 +432,8 @@ export class WorkspaceDocResolver {
    }

    await this.ac.user(user.id).doc(workspaceId, docId).assert('Doc.Publish');
-    await this.assertCanShare(user.id, {
+    await this.inviteQuota.assertWorkspaceActionAllowed({
+      actorUserId: user.id,
      workspaceId,
      docId,
      action: 'publishDoc',
@@ -1,4 +1,3 @@
-import { Logger } from '@nestjs/common';
 import {
  Args,
  Context,
@@ -39,7 +38,6 @@ import {
 import type { GraphqlContext } from '../../../base/graphql';
 import { Models, type WorkspaceUserCompat } from '../../../models';
 import { CurrentUser, Public } from '../../auth';
-import { BackendRuntimeProvider } from '../../backend-runtime';
 import { containsUrlOrDomain } from '../../content-policy';
 import {
  PermissionAccess,
@@ -49,11 +47,7 @@ import {
 import { QuotaService } from '../../quota';
 import { UserType } from '../../user';
 import { validators } from '../../utils/validators';
-import {
-  canUserExecuteLimitedActions,
-  getAbuseRequestSource,
-  InviteQuotaAssertService,
-} from '../abuse';
+import { getAbuseRequestSource, InviteQuotaAssertService } from '../abuse';
 import { WorkspaceService } from '../service';
 import {
  InvitationType,
@@ -92,8 +86,6 @@ function aggregateTargetDomains(candidates: InviteCandidate[]) {
 */
@Resolver(() => WorkspaceType)
 export class WorkspaceMemberResolver {
-  private readonly logger = new Logger(WorkspaceMemberResolver.name);
-
  constructor(
    private readonly cache: Cache,
    private readonly event: EventBus,
@@ -105,55 +97,9 @@ export class WorkspaceMemberResolver {
    private readonly workspaceService: WorkspaceService,
    private readonly quota: QuotaService,
    private readonly config: Config,
-    private readonly inviteQuota: InviteQuotaAssertService,
-    private readonly runtime: BackendRuntimeProvider
+    private readonly inviteQuota: InviteQuotaAssertService
  ) {}

-  private async assertCanInviteOrShare(
-    userId: string,
-    context: {
-      workspaceId: string;
-      action: 'createInviteLink';
-    }
-  ) {
-    if (await this.runtime.isInviteAbuseUserQuarantinedOrBanned(userId)) {
-      this.logger.warn('Share action blocked for quarantined actor', {
-        userId,
-        ...context,
-      });
-      throw new ActionForbidden(
-        'This feature is temporarily unavailable for you.'
-      );
-    }
-    if (
-      await this.runtime.isInviteAbuseWorkspaceQuarantined(context.workspaceId)
-    ) {
-      this.logger.warn('Share action blocked for quarantined workspace', {
-        userId,
-        ...context,
-      });
-      throw new ActionForbidden(
-        'This feature is temporarily unavailable for you.'
-      );
-    }
-    // Member invites are owned by native quota; this guard stays for invite links until share/link actions migrate.
-    const user = await this.models.user.get(userId);
-    const newAccountAgeMs = this.config.auth.newAccountShareActionDelay * 1000;
-    if (!user || !canUserExecuteLimitedActions(user, newAccountAgeMs)) {
-      this.logger.warn('Share action blocked for new account', {
-        userId,
-        email: user?.email,
-        createdAt: user?.createdAt,
-        accountAgeMs: user ? Date.now() - user.createdAt.getTime() : null,
-        minimumAccountAgeMs: newAccountAgeMs,
-        ...context,
-      });
-      throw new ActionForbidden(
-        'This feature is temporarily unavailable for you.'
-      );
-    }
-  }
-
  private async assertWorkspaceNameCanInvite(workspaceId: string) {
    const workspace = await this.workspaceService.getWorkspaceInfo(workspaceId);
    if (containsUrlOrDomain(workspace.name)) {
@@ -287,6 +233,12 @@ export class WorkspaceMemberResolver {
      return results;
    }

+    await this.inviteQuota.assertWorkspaceActionAllowed({
+      actorUserId: me.id,
+      workspaceId,
+      action: 'inviteMember',
+    });
+
    // lock to prevent concurrent invite
    const lockFlag = `invite:${workspaceId}`;
    await using lock = await this.mutex.acquire(lockFlag);
@@ -452,7 +404,8 @@ export class WorkspaceMemberResolver {
      .user(user.id)
      .workspace(workspaceId)
      .assert('Workspace.Users.Manage');
-    await this.assertCanInviteOrShare(user.id, {
+    await this.inviteQuota.assertWorkspaceActionAllowed({
+      actorUserId: user.id,
      workspaceId,
      action: 'createInviteLink',
    });
@@ -87,9 +87,9 @@
      "type": "Boolean",
      "desc": "Whether require email verification before accessing restricted resources(not implemented)."
    },
-    "newAccountShareActionDelay": {
+    "newAccountActionDelay": {
      "type": "Number",
-      "desc": "Minimum account age in seconds before new accounts can invite members or create share links."
+      "desc": "Minimum account age in seconds before new accounts can invite members, create invite links, or publish documents. Set to 0 to disable."
    },
    "trustedCloudflareHeaders": {
      "type": "Boolean",
@@ -58,9 +58,9 @@ export const KNOWN_CONFIG_GROUPS = [
      'allowSignup',
      'allowSignupForOauth',
      {
-        key: 'newAccountShareActionDelay',
+        key: 'newAccountActionDelay',
        type: 'Number',
-        desc: 'Minimum account age in seconds before new accounts can invite members or create share links.',
+        desc: 'Minimum account age in seconds before accounts can invite members, create invite links, or publish documents. Set to 0 to disable.',
      },
      // nested json object
      {
@@ -27,6 +27,7 @@ import {
  type ModelDeclaration,
  modelUseCases,
  probeChecks,
+  retainVerifiedCapabilities,
 } from './model-utils';
 import type { ByokDefinition, ByokKey, ByokSettings, GqlFn } from './types';
 import { ByokStorage } from './types';
@@ -138,7 +139,7 @@ export const AddKeyModal = ({

  const invalidateTest = () => setTestStatus(null);
  const runProbe = useCallback(async () => {
-    if (!gql) return false;
+    if (!gql) return { passed: false, definition };
    const canReuseServerCredential =
      editingKey?.storage === ByokStorage.server && !apiKey;
    const checks = probeChecks(models, includeImageProbe);
@@ -159,21 +160,19 @@ export const AddKeyModal = ({
      },
    });
    const probe = result.probeWorkspaceByokDraft;
-    const verifiedChecks = new Set(
-      probe.models.flatMap(model =>
-        model.checks
-          .filter(check => check.status.kind === 'verified')
-          .map(check => `${model.modelId}\0${check.operation}`)
-      )
+    const nextModels = retainVerifiedCapabilities(models, probe.models);
+    const nextDefinition = { ...definition, models: nextModels };
+    const hasVerifiedCheck = probe.models.some(model =>
+      model.checks.some(check => check.status.kind === 'verified')
    );
    const passed =
      checks.length > 0 &&
      probe.connection.kind === 'verified' &&
-      checks.every(check =>
-        verifiedChecks.has(`${check.modelId}\0${check.operation}`)
-      );
+      hasVerifiedCheck &&
+      nextModels.some(model => model.enabled && model.capabilities.length > 0);
+    if (passed) setModels(nextModels);
    setTestStatus(passed ? 'passed' : 'failed');
-    return passed;
+    return { passed, definition: nextDefinition };
  }, [
    apiKey,
    definition,
@@ -185,112 +184,118 @@ export const AddKeyModal = ({
    workspaceId,
  ]);

-  const persist = useCallback(async () => {
-    if (!gql) return;
-    if (storage === ByokStorage.local) {
-      const saved = await upsertLocalKey(workspaceId, {
-        id:
-          editingKey?.storage === ByokStorage.local
-            ? editingKey.id
-            : crypto.randomUUID(),
-        provider,
-        name,
-        description,
-        credential: apiKey,
-        definition,
-        sortOrder:
-          editingKey?.storage === ByokStorage.local
-            ? editingKey.sortOrder
-            : localKeys.length,
-        enabled: profileEnabled,
-      });
-      if (!saved) {
-        notify.error({
-          title: byokT(t, 'notify.local-save-failed.title'),
-          message: byokT(t, 'notify.local-save-failed.message'),
+  const persist = useCallback(
+    async (persistedDefinition = definition) => {
+      if (!gql) return;
+      if (storage === ByokStorage.local) {
+        const saved = await upsertLocalKey(workspaceId, {
+          id:
+            editingKey?.storage === ByokStorage.local
+              ? editingKey.id
+              : crypto.randomUUID(),
+          provider,
+          name,
+          description,
+          credential: apiKey,
+          definition: persistedDefinition,
+          sortOrder:
+            editingKey?.storage === ByokStorage.local
+              ? editingKey.sortOrder
+              : localKeys.length,
+          enabled: profileEnabled,
        });
-        return;
-      }
-      setLocalKeys(await readLocalKeys(workspaceId));
-    } else if (editingKey?.storage === ByokStorage.server) {
-      if (editingKey.revision === undefined) {
-        notify.error({
-          title: byokT(t, 'notify.reload-required.title'),
-          message: byokT(t, 'notify.reload-required.message'),
+        if (!saved) {
+          notify.error({
+            title: byokT(t, 'notify.local-save-failed.title'),
+            message: byokT(t, 'notify.local-save-failed.message'),
+          });
+          return;
+        }
+        setLocalKeys(await readLocalKeys(workspaceId));
+      } else if (editingKey?.storage === ByokStorage.server) {
+        if (editingKey.revision === undefined) {
+          notify.error({
+            title: byokT(t, 'notify.reload-required.title'),
+            message: byokT(t, 'notify.reload-required.message'),
+          });
+          return;
+        }
+        await gql({
+          query: replaceWorkspaceByokProfileMutation,
+          variables: {
+            input: {
+              workspaceId,
+              profileId: editingKey.id,
+              expectedRevision: editingKey.revision,
+              name,
+              description: description || null,
+              credential: apiKey || null,
+              definition: persistedDefinition,
+              enabled: profileEnabled,
+            },
+          },
        });
-        return;
+        await onSaved();
+      } else {
+        await gql({
+          query: createWorkspaceByokProfileMutation,
+          variables: {
+            input: {
+              workspaceId,
+              provider,
+              name,
+              description: description || null,
+              credential: apiKey,
+              definition: persistedDefinition,
+              enabled: profileEnabled,
+            },
+          },
+        });
+        await onSaved();
      }
-      await gql({
-        query: replaceWorkspaceByokProfileMutation,
-        variables: {
-          input: {
-            workspaceId,
-            profileId: editingKey.id,
-            expectedRevision: editingKey.revision,
-            name,
-            description: description || null,
-            credential: apiKey || null,
-            definition,
-            enabled: profileEnabled,
-          },
-        },
-      });
-      await onSaved();
-    } else {
-      await gql({
-        query: createWorkspaceByokProfileMutation,
-        variables: {
-          input: {
-            workspaceId,
-            provider,
-            name,
-            description: description || null,
-            credential: apiKey,
-            definition,
-            enabled: profileEnabled,
-          },
-        },
-      });
-      await onSaved();
-    }
-    onOpenChange(false);
-  }, [
-    apiKey,
-    definition,
-    description,
-    editingKey,
-    gql,
-    localKeys.length,
-    name,
-    onOpenChange,
-    onSaved,
-    provider,
-    profileEnabled,
-    setLocalKeys,
-    storage,
-    t,
-    workspaceId,
-  ]);
+      onOpenChange(false);
+    },
+    [
+      apiKey,
+      definition,
+      description,
+      editingKey,
+      gql,
+      localKeys.length,
+      name,
+      onOpenChange,
+      onSaved,
+      provider,
+      profileEnabled,
+      setLocalKeys,
+      storage,
+      t,
+      workspaceId,
+    ]
+  );

  const connect = useCallback(async () => {
    if (busyRef.current) return;
    busyRef.current = true;
    setBusy(true);
    try {
-      const passed = testStatus === 'passed' || (await runProbe());
-      if (!passed) {
+      const probe =
+        testStatus === 'passed'
+          ? { passed: true, definition }
+          : await runProbe();
+      if (!probe.passed) {
        notify.error({
          title: byokT(t, 'notify.test-failed.title'),
          message: byokT(t, 'notify.operation-failed.message'),
        });
        return;
      }
-      await persist();
+      await persist(probe.definition);
    } finally {
      busyRef.current = false;
      setBusy(false);
    }
-  }, [persist, runProbe, t, testStatus]);
+  }, [definition, persist, runProbe, t, testStatus]);

  const testConnection = useCallback(async () => {
    if (busyRef.current) return;
@@ -309,7 +314,10 @@ export const AddKeyModal = ({
    !!name.trim() &&
    hasCredential &&
    models.length > 0 &&
-    models.every(model => model.modelId.trim() && model.capabilities.length) &&
+    models.every(
+      model =>
+        model.modelId.trim() && (!model.enabled || model.capabilities.length)
+    ) &&
    new Set(models.map(model => model.modelId.trim())).size === models.length &&
    (!customEndpoint || (!!endpoint.trim() && dialect !== null));

@@ -11,6 +11,7 @@ import {
  capabilitiesForUseCases,
  type ModelDeclaration,
  modelUseCases,
+  retainVerifiedCapabilities,
 } from './model-utils';

 describe('BYOK model capabilities', () => {
@@ -60,4 +61,53 @@ describe('BYOK model capabilities', () => {
      capabilitiesForUseCases(model, ['chat', 'actions', 'vision'])
    ).toEqual([capability]);
  });
+
+  test('drops capabilities that imply failed uses while keeping independent uses', () => {
+    const embeddingCapability = {
+      input: [ByokModelInput.text],
+      output: [ByokModelOutput.embedding],
+      features: [],
+      attachmentKinds: [],
+      attachmentSources: [],
+    };
+    const model: ModelDeclaration = {
+      modelId: 'multimodal-tools',
+      enabled: true,
+      capabilities: [
+        {
+          input: [ByokModelInput.text, ByokModelInput.image],
+          output: [ByokModelOutput.text],
+          features: [ByokModelFeature.tool_calling],
+          attachmentKinds: [ByokAttachmentKind.image],
+          attachmentSources: [
+            ByokAttachmentSource.url,
+            ByokAttachmentSource.data,
+            ByokAttachmentSource.bytes,
+            ByokAttachmentSource.file_handle,
+          ],
+        },
+        embeddingCapability,
+      ],
+    };
+
+    const retained = retainVerifiedCapabilities(
+      [model],
+      [
+        {
+          modelId: model.modelId,
+          checks: [
+            { operation: 'chat', status: { kind: 'failed' } },
+            { operation: 'tool_calling', status: { kind: 'verified' } },
+            { operation: 'embedding', status: { kind: 'verified' } },
+          ],
+        },
+      ]
+    )[0];
+
+    expect(retained).toEqual({
+      ...model,
+      capabilities: [embeddingCapability],
+    });
+    expect(modelUseCases(retained)).toEqual(['embedding']);
+  });
 });
@@ -163,6 +163,55 @@ export function probeChecks(models: ModelDeclaration[], includeImage: boolean) {
    );
 }

+export function retainVerifiedCapabilities(
+  models: ModelDeclaration[],
+  probeModels: Array<{
+    modelId: string;
+    checks: Array<{ operation: string; status: { kind: string } }>;
+  }>
+) {
+  return models.map(model => {
+    const probe = probeModels.find(item => item.modelId === model.modelId);
+    if (!probe) return model;
+    const failedOperations = new Set(
+      probe.checks
+        .filter(check => check.status.kind !== 'verified')
+        .map(check => check.operation)
+    );
+    const selectedUseCases = modelUseCases(model);
+    const verifiedUseCases = selectedUseCases.filter(useCase => {
+      const operation = useCase === 'actions' ? 'tool_calling' : useCase;
+      if (failedOperations.has(operation)) return false;
+
+      const represented = modelUseCases({
+        ...model,
+        capabilities: [capabilityForUseCase(useCase)],
+      });
+      return represented.every(
+        representedUseCase =>
+          !failedOperations.has(
+            representedUseCase === 'actions'
+              ? 'tool_calling'
+              : representedUseCase
+          )
+      );
+    });
+    const rebuiltCapabilities =
+      selectedUseCases.length > 0 &&
+      verifiedUseCases.length === selectedUseCases.length
+        ? model.capabilities
+        : verifiedUseCases.map(capabilityForUseCase);
+    const capabilities = rebuiltCapabilities.length
+      ? rebuiltCapabilities
+      : model.capabilities;
+    return {
+      ...model,
+      enabled: rebuiltCapabilities.length > 0 && model.enabled,
+      capabilities,
+    };
+  });
+}
+
 export function catalogModels(settings: ByokSettings, provider: ByokProvider) {
  return (
    settings.catalog.providers.find(item => item.provider === provider)