diff --git a/changelog.d/features/11104-operator-error-rules.md b/changelog.d/features/11104-operator-error-rules.md
new file mode 100644
index 0000000000..f31e78c01f
--- /dev/null
+++ b/changelog.d/features/11104-operator-error-rules.md
@@ -0,0 +1 @@
+- **feat(providers):** let operators declare per-provider error rules through `settings.providerErrorRules` instead of patching the catalog — an operator-supplied rule for a provider is consulted before the built-in `providerRuleRegistry`, receives the raw error text, and has its declared scope/cooldown/reason actually honored end to end, for any provider (declaring the rule is the opt-in — no extra allowlist entry needed). Matches are plain case-insensitive substrings (never RegExp) and bounded to 50 rules to keep the hot path safe ([#11104](https://github.com/diegosouzapw/OmniRoute/pull/11104))
diff --git a/changelog.d/fixes/11060-perplexity-filter.md b/changelog.d/fixes/11060-perplexity-filter.md
new file mode 100644
index 0000000000..c221d3ccab
--- /dev/null
+++ b/changelog.d/fixes/11060-perplexity-filter.md
@@ -0,0 +1 @@
+- fix(providers): filter Perplexity model import to the Sonar family so Agent-API catalog ids stop surfacing as routable chat models (#11060)
diff --git a/changelog.d/fixes/11095-termux-onnx.md b/changelog.d/fixes/11095-termux-onnx.md
new file mode 100644
index 0000000000..8c26803710
--- /dev/null
+++ b/changelog.d/fixes/11095-termux-onnx.md
@@ -0,0 +1 @@
+- fix(install): make the ONNX dependency chain optional so Termux/Android installs succeed again (#11095)
diff --git a/changelog.d/fixes/11101-reject-silent-validation.md b/changelog.d/fixes/11101-reject-silent-validation.md
new file mode 100644
index 0000000000..04b2a67d5a
--- /dev/null
+++ b/changelog.d/fixes/11101-reject-silent-validation.md
@@ -0,0 +1 @@
+- **fix(providers):** Reject silent validation degradation on provider connection patch — unknown `rateLimitOverrides` keys (e.g. a typo'd `tpm`) and empty/non-numeric values now return `400` with the rejected key list instead of being silently dropped ([#11101](https://github.com/diegosouzapw/OmniRoute/pull/11101))
diff --git a/changelog.d/fixes/11102-combo-suggestion-count.md b/changelog.d/fixes/11102-combo-suggestion-count.md
new file mode 100644
index 0000000000..3cbf6f11d3
--- /dev/null
+++ b/changelog.d/fixes/11102-combo-suggestion-count.md
@@ -0,0 +1 @@
+- **Autopilot suggestion counter:** the combo health autopilot summary now reports `suggestionCount` (the real number of suggested actions across all issues) instead of conflating it with link counts, while keeping `actionableCount` as a deprecated alias for backward compatibility. The `run_combo_test` action now links to the dashboard with the combo id (`/dashboard/combos?test=<comboId>`) rather than the read-only API route, so operators can actually trigger a test from the UI ([#11102](https://github.com/diegosouzapw/OmniRoute/pull/11102)).
diff --git a/changelog.d/fixes/11103-persist-config-audit-log.md b/changelog.d/fixes/11103-persist-config-audit-log.md
new file mode 100644
index 0000000000..aeb53b1781
--- /dev/null
+++ b/changelog.d/fixes/11103-persist-config-audit-log.md
@@ -0,0 +1 @@
+- **Config audit persistence:** persist the configuration audit trail to SQLite (`config_audit_log`) instead of an in-memory buffer capped at 1000 volatile entries, and bound its growth with `cleanupConfigAudit()` driven by the `retention.configAudit` setting (default 30 days), wired into `runAutoCleanup` ([#11103](https://github.com/diegosouzapw/OmniRoute/pull/11103)).
diff --git a/changelog.d/fixes/11109-stream-recovery-toolcall.md b/changelog.d/fixes/11109-stream-recovery-toolcall.md
new file mode 100644
index 0000000000..04a43382e4
--- /dev/null
+++ b/changelog.d/fixes/11109-stream-recovery-toolcall.md
@@ -0,0 +1 @@
+- fix(sse): resume mid-stream recovery after a _completed_ tool call — `finish_reason: "tool_calls"` is now tracked per-call instead of as a general terminal marker, so truncation of trailing prose after a fully-delivered tool call is recoverable while in-flight calls stay blocked ([#11109](https://github.com/diegosouzapw/OmniRoute/pull/11109))
diff --git a/changelog.d/fixes/11116-reasoning-effort-capability-discovery.md b/changelog.d/fixes/11116-reasoning-effort-capability-discovery.md
new file mode 100644
index 0000000000..fbc4dfe694
--- /dev/null
+++ b/changelog.d/fixes/11116-reasoning-effort-capability-discovery.md
@@ -0,0 +1 @@
+- **fix(providers):** `reasoning_effort` now learns the accepted values from a provider's own 400/422 response and clamps to the highest one instead of forwarding an unsupported `xhigh`/`max` (or a hardcoded `"high"` fallback) — fixes custom OpenAI-compatible connections and registered providers with no reasoning metadata ([#11116](https://github.com/diegosouzapw/OmniRoute/pull/11116)) — thanks @maxmad64bis
diff --git a/changelog.d/fixes/11144-responses-parallel-tool-calls-index.md b/changelog.d/fixes/11144-responses-parallel-tool-calls-index.md
new file mode 100644
index 0000000000..35ba19b279
--- /dev/null
+++ b/changelog.d/fixes/11144-responses-parallel-tool-calls-index.md
@@ -0,0 +1 @@
+- **fix(sse):** parallel `function_call` items in a Responses API stream (e.g. several tool calls dispatched in the same turn) now each get a stable, distinct `index`/`id` when translated to Chat Completions streaming deltas, instead of colliding on index 0 and tripping strict stream parsers with `Expected 'id' to be a string.` ([#11144](https://github.com/diegosouzapw/OmniRoute/pull/11144))
diff --git a/changelog.d/maintenance/vi-harimport-parity.md b/changelog.d/maintenance/vi-harimport-parity.md
new file mode 100644
index 0000000000..b08b8dc92f
--- /dev/null
+++ b/changelog.d/maintenance/vi-harimport-parity.md
@@ -0,0 +1 @@
+- fix(i18n): translate the 14 `providers.harImport*` keys into Vietnamese (parity gap left by #11069)
diff --git a/docs/architecture/RESILIENCE_GUIDE.md b/docs/architecture/RESILIENCE_GUIDE.md
index 030f65dd41..0048761e60 100644
--- a/docs/architecture/RESILIENCE_GUIDE.md
+++ b/docs/architecture/RESILIENCE_GUIDE.md
@@ -448,14 +448,14 @@ classification rules pick the fallback `reason` and lock `scope`
 Classification rules only see full error **text** (needed to match body
 markers like `额度不足`) for providers listed in the `FULL_TEXT_RULE_PROVIDERS`
 allowlist in `providerErrorRules.ts` — currently only `"agentrouter"`. For
-every other provider, `checkFallbackError` hands `getProviderErrorRuleMatch`
-only the structured error (`{code, type}`), which is enough for
-header/status/code-based rules but blind to body-text markers. The helper
-`resolveRuleMatchBody()` performs this selection: full error text for
-allowlisted providers, the structured error otherwise. Adding a provider to
-`FULL_TEXT_RULE_PROVIDERS` is an explicit per-provider opt-in — it exists so
-that the default path for every provider not on the list stays
-byte-for-byte unchanged.
+every other **built-in catalog** provider, `checkFallbackError` hands
+`getProviderErrorRuleMatch` only the structured error (`{code, type}`), which
+is enough for header/status/code-based rules but blind to body-text markers.
+The helper `resolveRuleMatchBody()` performs this selection: full error text
+for allowlisted providers, the structured error otherwise. Adding a
+**built-in** provider to `FULL_TEXT_RULE_PROVIDERS` is an explicit per-provider
+opt-in — it exists so that the default path for every provider not on the
+list stays byte-for-byte unchanged.
 
 A rule's `scope` (`model` / `provider` / `connection`) is a separate opt-in
 from `FULL_TEXT_RULE_PROVIDERS`: `checkFallbackError` only surfaces it as
@@ -466,6 +466,31 @@ honorsRuleLockScope()` — today only `"agentrouter"`). See "Restated quota
 errors" above for what a `scope: "connection"` match actually does once a
 provider is on that allowlist.
 
+**#11104 — operator-declared rules bypass both allowlists.** An operator can
+declare a per-provider rule at runtime via `settings.providerErrorRules`
+(`open-sse/config/providerErrorRules.ts::setOperatorProviderErrorRules`)
+without editing this file. Gating an operator rule behind
+`FULL_TEXT_RULE_PROVIDERS`/`HONORS_RULE_LOCK_SCOPE_PROVIDERS` — allowlists
+meant to protect the **default** behavior of built-in catalog rules — would
+make the settings mechanism inert for every provider except the ones already
+listed there, since declaring the rule is already the operator's explicit
+opt-in. `resolveRuleMatchBody()` and `honorsRuleLockScope()` both check
+`hasOperatorRuleForProvider()` first: a provider with an operator rule gets
+the raw error text and has its declared `scope` honored, regardless of
+whether it also appears in either allowlist.
+
+**Known gap — `providerRuleRegistry` is never consulted for HTTP 400.**
+`checkFallbackError`'s `BAD_REQUEST` branch classifies status 400 entirely
+through its own pattern arrays (`MODEL_ACCESS_DENIED_PATTERNS`,
+`CONTEXT_OVERFLOW_PATTERNS`, etc. in `accountFallback.ts`) and returns before
+the `configuredRule`/`getProviderErrorRuleMatch` branch above it is reached.
+A built-in catalog rule (or an operator rule) with `status: 400` is
+syntactically valid but will never fire. No existing rule targets 400 today,
+so nothing in production is affected — but a future 400 rule needs this
+branch touched first, which is a larger change than adding a rule (it
+reclassifies 400 for every provider already relying on the pattern-array
+behavior) and is out of scope for a single-provider rule addition.
+
 ### Adding a new quota-misstating gateway
 
 1. Register one rule array in `statusRestatementRegistry`
diff --git a/open-sse/config/constants.ts b/open-sse/config/constants.ts
index c45cf542a6..f1fa055c93 100644
--- a/open-sse/config/constants.ts
+++ b/open-sse/config/constants.ts
@@ -171,6 +171,7 @@ export const HTTP_STATUS = {
   FORBIDDEN: 403,
   NOT_FOUND: 404,
   NOT_ACCEPTABLE: 406,
+  UNPROCESSABLE_ENTITY: 422,
   REQUEST_TIMEOUT: 408,
   GONE: 410,
   RATE_LIMITED: 429,
@@ -263,11 +264,17 @@ export const PROVIDER_PROFILES = {
     circuitBreakerReset: envInt("OMNIROUTE_CIRCUIT_BREAKER_API_KEY_RESET_MS", 30000),
     // Provider-level circuit breaker (entire provider cooldown after repeated failures)
     providerFailureThreshold: envInt("OMNIROUTE_PROVIDER_BREAKER_API_KEY_FAILURE_THRESHOLD", 15), // Scaled for 500+ connections (was 5)
-    providerFailureWindowMs: envInt("OMNIROUTE_PROVIDER_BREAKER_API_KEY_FAILURE_WINDOW_MS", 1800000), // 30min window (was 20min)
+    providerFailureWindowMs: envInt(
+      "OMNIROUTE_PROVIDER_BREAKER_API_KEY_FAILURE_WINDOW_MS",
+      1800000
+    ), // 30min window (was 20min)
     providerCooldownMs: envInt("OMNIROUTE_PROVIDER_BREAKER_API_KEY_COOLDOWN_MS", 600000), // 10min cooldown when threshold reached
     degradationThreshold: envInt("OMNIROUTE_PROVIDER_BREAKER_API_KEY_DEGRADATION_THRESHOLD", 7),
     maxBackoffMultiplier: envInt("OMNIROUTE_PROVIDER_BREAKER_API_KEY_MAX_BACKOFF_MULTIPLIER", 4),
-    backoffEscalationCount: envInt("OMNIROUTE_PROVIDER_BREAKER_API_KEY_BACKOFF_ESCALATION_COUNT", 3),
+    backoffEscalationCount: envInt(
+      "OMNIROUTE_PROVIDER_BREAKER_API_KEY_BACKOFF_ESCALATION_COUNT",
+      3
+    ),
   },
   // Local providers (localhost inference backends like Ollama, LM Studio, oMLX).
   // Not yet wired into getProviderProfile() — will be used when local provider_nodes
diff --git a/open-sse/config/providerErrorRules.ts b/open-sse/config/providerErrorRules.ts
index c910e3fb4d..6f00d2c5f1 100644
--- a/open-sse/config/providerErrorRules.ts
+++ b/open-sse/config/providerErrorRules.ts
@@ -30,21 +30,63 @@ export type ProviderErrorRule = {
 export type ProviderErrorRuleMatch = {
   reason: ConfiguredErrorReason;
   /**
-   * Intended lock scope. #10334: this field is CONSUMED end-to-end only for
-   * providers in `HONORS_RULE_LOCK_SCOPE_PROVIDERS` (agentrouter-exclusive
-   * today, gated by `honorsRuleLockScope()`) — for those, `checkFallbackError`
-   * surfaces it as `ruleScope` on its return value for the persistence layer
-   * to honor instead of re-deriving scope from `hasPerModelQuota()`. For
-   * every other provider it remains INFORMATIONAL: `getProviderErrorRuleMatch`
-   * callers still read only `reason`/`cooldownMs`, and the actual lock scope
-   * is decided independently by each call site. Widening the allowlist is
-   * tracked as a follow-up — see `docs/architecture/RESILIENCE_GUIDE.md` §7.
+   * Intended lock scope. #10334: for a BUILT-IN catalog rule, this field is
+   * CONSUMED end-to-end only for providers in `HONORS_RULE_LOCK_SCOPE_PROVIDERS`
+   * (agentrouter-exclusive today, gated by `honorsRuleLockScope()`) — for those,
+   * `checkFallbackError` surfaces it as `ruleScope` on its return value for the
+   * persistence layer to honor instead of re-deriving scope from
+   * `hasPerModelQuota()`. For every other built-in-rule provider it remains
+   * INFORMATIONAL. #11104: an OPERATOR-declared rule (`OperatorProviderErrorRule`)
+   * is exempt from this allowlist — `honorsRuleLockScope()` always returns true
+   * when the provider has one, since the operator already opted in by declaring
+   * the rule. Widening `HONORS_RULE_LOCK_SCOPE_PROVIDERS` itself (for a new
+   * built-in catalog rule) is tracked as a follow-up — see
+   * `docs/architecture/RESILIENCE_GUIDE.md` §7.
    */
   scope: "model" | "provider" | "connection";
   /** Optional explicit cooldown; falls back to the existing per-reason defaults. */
   cooldownMs?: number;
 };
 
+/**
+ * Operator-declared per-provider error rule (settings-driven).
+ *
+ * Mirrors the catalog `ProviderErrorRule` but is data-only so an operator can
+ * add a scope/cooldown/reason override for a provider without editing this
+ * file. `match` is a plain case-insensitive SUBSTRING of the error body — never
+ * a RegExp — so an operator-supplied pattern can never introduce a ReDoS on the
+ * error-classification hot path. Bounded to <= 50 rules total by the settings
+ * schema. An operator rule is consulted BEFORE the built-in `providerRuleRegistry`
+ * and wins on the first status+substring match for a provider.
+ */
+export type OperatorProviderErrorRule = {
+  status: number;
+  match: string;
+  scope: "model" | "provider" | "connection";
+  reason?: ConfiguredErrorReason;
+  cooldownMs?: number;
+};
+
+let operatorProviderErrorRules: Record<string, OperatorProviderErrorRule[]> = {};
+
+/**
+ * Inject operator-declared rules. Called from the runtime-settings applier
+ * (`applyRuntimeSettings`) once at boot and on every settings update, with the
+ * value validated by the settings schema. Pass `undefined`/empty/null to clear.
+ * Provider keys are lowercased so lookups are case-insensitive.
+ */
+export function setOperatorProviderErrorRules(
+  rules: Record<string, OperatorProviderErrorRule[]> | undefined | null
+): void {
+  operatorProviderErrorRules = {};
+  if (!rules) return;
+  for (const [provider, list] of Object.entries(rules)) {
+    if (Array.isArray(list) && list.length > 0) {
+      operatorProviderErrorRules[provider.toLowerCase()] = list;
+    }
+  }
+}
+
 // ─── Opencode ───────────────────────────────────────────────────────────────────
 // Opencode Go uses an account-wide quota. The body usually says "rate limit
 // reached" but the presence of `x-ratelimit-remaining-requests: 0` is the
@@ -272,11 +314,21 @@ export const providerRuleRegistry = new Map<string, ProviderErrorRule[]>([
  * FULL_TEXT_RULE_PROVIDERS: that set controls what body a rule matches against
  * (input), this one controls whether the matched scope changes caller behavior
  * (output). A provider could need one without the other.
+ *
+ * Providers with an operator-declared rule (`setOperatorProviderErrorRules`)
+ * are honored too, without being added here: the allowlist exists to gate
+ * BUILT-IN catalog rules, which change default behavior for every operator
+ * running that provider — an operator rule is already an explicit, per-operator
+ * opt-in, so gating it a second time behind this list would make the settings
+ * mechanism (#11104) silently inert for every provider except the ones listed
+ * below. See `hasOperatorRuleForProvider`.
  */
 const HONORS_RULE_LOCK_SCOPE_PROVIDERS = new Set(["agentrouter"]);
 
 export function honorsRuleLockScope(provider: string | null | undefined): boolean {
-  return !!provider && HONORS_RULE_LOCK_SCOPE_PROVIDERS.has(provider.toLowerCase());
+  if (!provider) return false;
+  const key = provider.toLowerCase();
+  return HONORS_RULE_LOCK_SCOPE_PROVIDERS.has(key) || hasOperatorRuleForProvider(key);
 }
 
 /**
@@ -310,28 +362,51 @@ export function egressBucketedLockProviders(): string[] {
 }
 
 /**
- * Providers whose rules match on the FULL upstream error text.
- * checkFallbackError's rule lookup normally passes only the structured
+ * Providers whose BUILT-IN catalog rules match on the FULL upstream error
+ * text. checkFallbackError's rule lookup normally passes only the structured
  * error ({code, type} — message stripped by the combo callers), which is
  * enough for header/status/code rules but blind to body-text markers like
  * agentrouter's "额度不足". Providers in this set get the raw error text as
  * the match body instead. EXCLUSIVE allowlist by owner decision (2026-08-13):
  * adding a provider here is an explicit opt-in — the default path for every
  * other provider must remain byte-for-byte unchanged.
+ *
+ * Operator-declared rules bypass this allowlist entirely (see
+ * `hasOperatorRuleForProvider`): the operator's `match` is a literal substring
+ * of the error body by construction, so a rule that never sees body text could
+ * never match anything, defeating the point of declaring it.
  */
 const FULL_TEXT_RULE_PROVIDERS = new Set(["agentrouter"]);
 
+/**
+ * True when an operator has declared at least one rule for this provider via
+ * `settings.providerErrorRules` (injected through `setOperatorProviderErrorRules`).
+ * Presence of the rule IS the opt-in — no separate allowlist to maintain, and
+ * no widening decision needed as new operators configure new providers.
+ */
+export function hasOperatorRuleForProvider(provider: string | null | undefined): boolean {
+  if (!provider) return false;
+  const rules = operatorProviderErrorRules[provider.toLowerCase()];
+  return !!rules && rules.length > 0;
+}
+
 /**
  * Resolve the body handed to getProviderErrorRuleMatch inside
- * checkFallbackError: full error text for FULL_TEXT_RULE_PROVIDERS,
- * the structured error for everyone else.
+ * checkFallbackError: full error text for FULL_TEXT_RULE_PROVIDERS or any
+ * provider with an operator-declared rule, the structured error for everyone
+ * else.
  */
 export function resolveRuleMatchBody(
   provider: string | null | undefined,
   structuredError: unknown,
   errorText: string | null | undefined
 ): unknown {
-  if (provider && FULL_TEXT_RULE_PROVIDERS.has(provider.toLowerCase()) && errorText) {
+  if (
+    provider &&
+    (FULL_TEXT_RULE_PROVIDERS.has(provider.toLowerCase()) ||
+      hasOperatorRuleForProvider(provider)) &&
+    errorText
+  ) {
     return errorText;
   }
   return structuredError ?? null;
@@ -346,10 +421,32 @@ export function getProviderErrorRuleMatch(
   provider: string | null | undefined,
   status: number,
   headers: Headers | Record<string, string> | null | undefined,
-  body?: unknown
+  body?: unknown,
+  operatorRules?: Record<string, OperatorProviderErrorRule[]>
 ): ProviderErrorRuleMatch | null {
   if (!provider) return null;
-  const rules = providerRuleRegistry.get(provider.toLowerCase());
+  const key = provider.toLowerCase();
+
+  // Operator-declared rules win first: an operator can override any catalog
+  // rule for a provider without editing this file. `operatorRules` is the
+  // injected source (tests / direct callers); when omitted we fall back to the
+  // settings-backed cache populated by `setOperatorProviderErrorRules`.
+  const opRules = (operatorRules ?? operatorProviderErrorRules)?.[key];
+  if (opRules && opRules.length > 0) {
+    const text = typeof body === "string" ? body : JSON.stringify(body ?? "");
+    const lowered = text.toLowerCase();
+    for (const r of opRules) {
+      if (r.status === status && lowered.includes(r.match.toLowerCase())) {
+        return {
+          reason: r.reason ?? "quota_exhausted",
+          scope: r.scope,
+          cooldownMs: r.cooldownMs,
+        };
+      }
+    }
+  }
+
+  const rules = providerRuleRegistry.get(key);
   if (!rules) return null;
   // Normalize headers: accept either a `Headers` object (from `fetch()`) or
   // a plain record. Provider rules access headers via plain object indexing.
diff --git a/open-sse/config/providerRegistry.ts b/open-sse/config/providerRegistry.ts
index bb3b6e0f9f..94bb036c95 100644
--- a/open-sse/config/providerRegistry.ts
+++ b/open-sse/config/providerRegistry.ts
@@ -10,6 +10,7 @@ export {
 } from "./providers/registry/alibaba/index.ts";
 export { REGISTRY } from "./providers/index.ts";
 import { REGISTRY } from "./providers/index.ts";
+import { isPrivateHost } from "@/shared/network/outboundUrlGuard";
 import {
   RegistryModel,
   REASONING_UNSUPPORTED,
@@ -132,11 +133,8 @@ export function isLocalProvider(baseUrl?: string | null): boolean {
   try {
     const url = new URL(baseUrl);
     const hostname = url.hostname;
-    // Strictly matching 172.16.0.0/12 (Docker/local) and explicitly blocking ::1 per SSRF hardening
-    return (
-      LOCAL_HOSTNAMES.has(hostname) ||
-      /^172\.(1[6-9]|2[0-9]|3[0-1])\.\d{1,3}\.\d{1,3}$/.test(hostname)
-    );
+    if (!hostname) return false;
+    return LOCAL_HOSTNAMES.has(hostname) || isPrivateHost(hostname);
   } catch {
     return false;
   }
diff --git a/open-sse/config/providers/registry/cline/index.ts b/open-sse/config/providers/registry/cline/index.ts
index 21e8c600ed..aecf811f19 100644
--- a/open-sse/config/providers/registry/cline/index.ts
+++ b/open-sse/config/providers/registry/cline/index.ts
@@ -27,7 +27,7 @@ export const clineProvider: RegistryEntry = {
   // the official free bucket and text-output models advertised as zero-cost.
   models: [
     {
-      id: "zai/glm-5.2",
+      id: "z-ai/glm-5.2",
       name: "GLM 5.2",
       toolCalling: true,
       supportsReasoning: true,
diff --git a/open-sse/config/providers/registry/opencode/go/index.ts b/open-sse/config/providers/registry/opencode/go/index.ts
index 14304426c5..abebd92c0f 100644
--- a/open-sse/config/providers/registry/opencode/go/index.ts
+++ b/open-sse/config/providers/registry/opencode/go/index.ts
@@ -14,8 +14,6 @@ export const opencode_goProvider: RegistryEntry = {
   authPrefix: "Bearer",
   defaultContextLength: 200000,
   models: [
-    ...OPENCODE_ZEN_GO_SHARED_MODELS,
-
     // Port from decolua/9router 8efacc11: align with official Go endpoints —
     // glm-5.2 is now advertised and Kimi chat traffic must route through
     // `kimi-k2.7-code` (the live API rejects the plain `kimi-k2.7` alias for
@@ -26,6 +24,10 @@ export const opencode_goProvider: RegistryEntry = {
     { id: "glm-5.2", name: "GLM-5.2", supportsReasoning: true },
     { id: "glm-5.2-high", name: "GLM-5.2 (high effort)", supportsReasoning: true },
     { id: "glm-5.2-max", name: "GLM-5.2 (max effort)", supportsReasoning: true },
+
+    ...OPENCODE_ZEN_GO_SHARED_MODELS,
+    // models[0] (glm-5.2) is the dashboard default (LlmChatCard/ProviderTestSlideOver take models[0]).
+
     { id: "glm-5.1", name: "GLM-5.1" },
     { id: "glm-5", name: "GLM-5" },
     // kimi-k2.7-code declared identically on opencode-zen — see OPENCODE_ZEN_GO_SHARED_MODELS.
diff --git a/open-sse/config/providers/registry/opencode/zen/index.ts b/open-sse/config/providers/registry/opencode/zen/index.ts
index 76b3811e39..9fdca2fc0a 100644
--- a/open-sse/config/providers/registry/opencode/zen/index.ts
+++ b/open-sse/config/providers/registry/opencode/zen/index.ts
@@ -16,8 +16,6 @@ export const opencode_zenProvider: RegistryEntry = {
   // from the live API response so new models work without a code deploy.
   passthroughModels: true,
   models: [
-    ...OPENCODE_ZEN_GO_SHARED_MODELS,
-
     // ── Chat / Coding ──────────────────────────────────────────
     // #2900: big-pickle's upstream runs DeepSeek thinking mode — declare the
     // interleaved reasoning_content contract so follow-up/tool-use turns replay
@@ -28,6 +26,10 @@ export const opencode_zenProvider: RegistryEntry = {
       supportsReasoning: true,
       interleavedField: "reasoning_content",
     },
+
+    ...OPENCODE_ZEN_GO_SHARED_MODELS,
+    // models[0] (big-pickle) is the dashboard default; SHARED spread kept after it.
+
     { id: "gpt-5.6-sol", name: "GPT 5.6 Sol" },
     { id: "gpt-5.6-terra", name: "GPT 5.6 Terra" },
     { id: "gpt-5.6-luna", name: "GPT 5.6 Luna" },
@@ -67,6 +69,8 @@ export const opencode_zenProvider: RegistryEntry = {
       supportsReasoning: true,
       targetFormat: "openai-responses",
     },
+    // Explicit wire-format overlay of the base opencode provider's muse-spark entry
+    // (targetFormat: openai-responses). Keep in sync with base on catalog syncs.
     {
       id: "muse-spark-1.2-contributor-free",
       name: "Muse Spark 1.2 Contributor Free",
diff --git a/open-sse/executors/base.ts b/open-sse/executors/base.ts
index 66e10b67e9..1c13442aff 100644
--- a/open-sse/executors/base.ts
+++ b/open-sse/executors/base.ts
@@ -20,6 +20,10 @@ import {
   recordLearnedThinkingCap,
   parseThinkingBudgetMax,
 } from "../services/learnedThinkingCaps.ts";
+import {
+  recordLearnedReasoningEffort,
+  parseReasoningEffortEnum,
+} from "../services/learnedReasoningEffortCaps.ts";
 import {
   getParamFilterConfig,
   addParamToBlocklist,
@@ -826,6 +830,9 @@ export class BaseExecutor {
     // loop. The learned cap is also recorded process-wide via
     // recordLearnedThinkingCap so future requests skip the 400 entirely.
     let thinkingBudgetClampedMax: number | null = null;
+    // Set by the reasoning_effort 4xx clamp-and-retry below — guards the same
+    // "fires at most once per URL" invariant as thinkingBudgetClampedMax above.
+    let reasoningEffortClamped = false;
 
     for (let urlIndex = 0; urlIndex < fallbackCount; urlIndex++) {
       const requestCredentials = withForcedResponsesUpstream(
@@ -1529,6 +1536,49 @@ export class BaseExecutor {
           }
         }
 
+        // Reasoning-effort enum 4xx clamp-and-retry (any provider/model without a
+        // declared reasoning_effort capability — custom OpenAI-compatible
+        // connections, or a registered provider the registry hasn't caught up
+        // with). Mirrors the thinking_budget clamp-and-retry above: parse the
+        // upstream-advertised accepted values, record them process-wide (so
+        // FUTURE requests clamp proactively via sanitizeReasoningEffortForProvider
+        // → getLearnedReasoningEffort), clamp the live transformedBody by
+        // re-running the sanitizer, and retry the same URL once.
+        if (
+          (response.status === HTTP_STATUS.BAD_REQUEST ||
+            response.status === HTTP_STATUS.UNPROCESSABLE_ENTITY) &&
+          !reasoningEffortClamped &&
+          transformedBody &&
+          typeof transformedBody === "object"
+        ) {
+          const errText = await response
+            .clone()
+            .text()
+            .catch(() => "");
+          const acceptedValues = parseReasoningEffortEnum(errText);
+          if (acceptedValues) {
+            reasoningEffortClamped = true;
+            const learned = recordLearnedReasoningEffort(this.provider, model, acceptedValues);
+            if (learned) {
+              transformedBody = sanitizeReasoningEffortForProvider(
+                transformedBody,
+                this.provider,
+                model,
+                log
+              );
+              let retryBody = JSON.stringify(transformedBody);
+              if (usesClaudeCodeProtocol || this.provider === "claude") {
+                retryBody = await signRequestBody(retryBody);
+              }
+              log?.info?.(
+                "REASONING_SANITIZE",
+                `Upstream ${response.status} rejected reasoning_effort on ${url} — clamped to ${learned} and retrying (learned for ${this.provider}/${model})`
+              );
+              response = await fetchWithStartTimeout(url, { ...fetchOptions, body: retryBody });
+            }
+          }
+        }
+
         // Generic reactive 400 field-downgrade; each field is stripped at most once.
         if (
           response.status === HTTP_STATUS.BAD_REQUEST &&
diff --git a/open-sse/executors/base/reasoningEffort.ts b/open-sse/executors/base/reasoningEffort.ts
index fa416dbc1a..8dd99904fd 100644
--- a/open-sse/executors/base/reasoningEffort.ts
+++ b/open-sse/executors/base/reasoningEffort.ts
@@ -8,6 +8,10 @@ import {
   getProviderModel,
   getProviderModels,
 } from "../../config/providerModels.ts";
+import {
+  getLearnedReasoningEffort,
+  REASONING_EFFORT_ORDER,
+} from "../../services/learnedReasoningEffortCaps.ts";
 
 /**
  * Sanitize reasoning_effort for providers that don't accept all values.
@@ -338,10 +342,24 @@ export function sanitizeReasoningEffortForProvider(
 
   const supportsXHigh = supportsXHighEffort(provider, modelStr);
   const supportsMax = supportsMaxEffortForProvider(provider, modelStr);
+  // Highest value we've actually seen this provider+model accept in a real
+  // upstream 4xx (learnedReasoningEffortCaps.ts) — takes priority over the
+  // static registry (which defaults to "supports everything" when there's no
+  // entry, e.g. custom OpenAI-compatible connections) and over the hardcoded
+  // "high" fallback below (which isn't always valid either).
+  const learnedCap = getLearnedReasoningEffort(provider, modelStr);
+  const learnedRank = learnedCap ? REASONING_EFFORT_ORDER.indexOf(learnedCap) : -1;
 
   // ── xhigh handling ──────────────────────────────────────────────────────
   // xhigh is OmniRoute-internal. Map it to the best effort the model accepts.
   if (effortStr === "xhigh") {
+    if (learnedCap && learnedRank < REASONING_EFFORT_ORDER.indexOf("xhigh")) {
+      log?.info?.(
+        "REASONING_SANITIZE",
+        `${provider}/${modelStr}: clamped reasoning_effort xhigh → ${learnedCap} (learned)`
+      );
+      return writeEffortValue(b, learnedCap, c);
+    }
     if (supportsXHigh) return body; // model accepts xhigh natively
     if (supportsMax) {
       log?.info?.(
@@ -366,6 +384,13 @@ export function sanitizeReasoningEffortForProvider(
   // upstream, and if it 400s the user gets a clear signal. This prevents
   // new models from being unusable for weeks until they're whitelisted (#8057).
   if (effortStr === "max") {
+    if (learnedCap && learnedRank < REASONING_EFFORT_ORDER.indexOf("max")) {
+      log?.info?.(
+        "REASONING_SANITIZE",
+        `${provider}/${modelStr}: clamped reasoning_effort max → ${learnedCap} (learned)`
+      );
+      return writeEffortValue(b, learnedCap, c);
+    }
     if (supportsMax) return body; // explicitly known to accept max
 
     // A model that explicitly advertises its accepted tiers is safe to normalize.
diff --git a/open-sse/services/accountFallback.ts b/open-sse/services/accountFallback.ts
index d63c58e94c..9e1f5f2ee3 100644
--- a/open-sse/services/accountFallback.ts
+++ b/open-sse/services/accountFallback.ts
@@ -21,7 +21,7 @@ import {
   honorsRuleLockScope,
 } from "../config/providerErrorRules.ts";
 import * as rot from "./rotationConfig.ts";
-import { getPassthroughProviders, getProviderCategory } from "../config/providerRegistry.ts";
+import { getPassthroughProviders, getProviderCategory, isLocalProvider } from "../config/providerRegistry.ts";
 import {
   DEFAULT_RESILIENCE_SETTINGS,
   resolveResilienceSettings,
@@ -37,7 +37,7 @@ import {
   type FailureKind,
 } from "../../src/shared/utils/classify429";
 import { recordProviderSuccess as resetCooldownFailureCount } from "./providerCooldownTracker.ts";
-import { resolveProviderId } from "../../src/shared/constants/providers";
+import { resolveProviderId, isLocalProvider as isLocalProviderId, isSelfHostedChatProvider } from "../../src/shared/constants/providers";
 import { resolveUseUpstream429BreakerHints } from "../../src/shared/utils/providerHints";
 import { getCodexModelScope } from "../config/codexQuotaScopes.ts";
 import { getQuotaScopedModelForProvider } from "./antigravityQuotaFamily.ts";
@@ -791,12 +791,14 @@ export function hasPerModelQuota(
     return connectionPassthroughModels;
   }
   if (!provider) return false;
-  if (getCanonicalLockProvider(provider) === "antigravity") return true;
-  if (getCanonicalLockProvider(provider) === "codex") return true;
-  if (provider === "gemini" || provider === "github") return true;
-  if (provider === "antigravity" || provider === "agy") return true;
-  if (getPassthroughProviders().has(provider)) return true;
-  if (isCompatibleProvider(provider)) return true;
+  const canonicalId = resolveProviderId(provider);
+  if (getCanonicalLockProvider(canonicalId) === "antigravity") return true;
+  if (getCanonicalLockProvider(canonicalId) === "codex") return true;
+  if (canonicalId === "gemini" || canonicalId === "github") return true;
+  if (canonicalId === "antigravity" || canonicalId === "agy") return true;
+  if (getPassthroughProviders().has(canonicalId)) return true;
+  if (isCompatibleProvider(canonicalId)) return true;
+  if (isLocalProviderId(canonicalId) || isSelfHostedChatProvider(canonicalId)) return true;
   return false;
 }
 
diff --git a/open-sse/services/learnedReasoningEffortCaps.ts b/open-sse/services/learnedReasoningEffortCaps.ts
new file mode 100644
index 0000000000..b0125d8683
--- /dev/null
+++ b/open-sse/services/learnedReasoningEffortCaps.ts
@@ -0,0 +1,126 @@
+/**
+ * Learned Reasoning-Effort Caps — reactive capability memory for providers/models
+ * OmniRoute has no static registry entry for (custom OpenAI-compatible connections,
+ * or any registered provider whose registry entry carries no reasoning metadata).
+ *
+ * Same shape as `learnedThinkingCaps.ts` (thinking_budget), generalized from a
+ * numeric budget to an ordinal reasoning_effort scale: on a 4xx whose body
+ * enumerates the accepted values, `base.ts`'s executor calls
+ * `recordLearnedReasoningEffort`, which stores the highest recognized value in a
+ * module-level Map keyed "provider:model" (lowercased). Subsequent requests for
+ * the same provider+model read the cap via `getLearnedReasoningEffort` (consulted
+ * by `sanitizeReasoningEffortForProvider` in `executors/base/reasoningEffort.ts`)
+ * so the 4xx→retry round-trip is paid at most once per process per provider+model.
+ *
+ * In-memory only (same operator-accepted tradeoff as the thinking-budget cache):
+ * restart resets, the first request after a restart may re-learn at the cost of
+ * one upstream 4xx.
+ */
+
+export const REASONING_EFFORT_ORDER: readonly string[] = [
+  "none",
+  "minimal",
+  "low",
+  "medium",
+  "high",
+  "xhigh",
+  "max",
+];
+
+// key: `${provider}:${model}` lowercased → highest value known to be accepted.
+const learnedCaps = new Map<string, string>();
+
+function buildKey(provider: string | null | undefined, model: string | null | undefined): string {
+  const p = typeof provider === "string" ? provider.trim().toLowerCase() : "";
+  const m = typeof model === "string" ? model.trim().toLowerCase() : "";
+  if (!p || !m) return "";
+  return `${p}:${m}`;
+}
+
+function rankOf(value: string): number {
+  return REASONING_EFFORT_ORDER.indexOf(value);
+}
+
+/**
+ * Return the learned cap for provider+model, or null when nothing has been
+ * learned yet (no upstream 4xx recorded). Keyed case-insensitively.
+ */
+export function getLearnedReasoningEffort(
+  provider: string | null | undefined,
+  model: string | null | undefined
+): string | null {
+  const key = buildKey(provider, model);
+  if (!key) return null;
+  return learnedCaps.get(key) ?? null;
+}
+
+/**
+ * Record that `acceptedValues` is the enum the upstream advertised for
+ * provider+model, and store the highest recognized value as the learned cap.
+ * Returns the stored value, or null when `acceptedValues` contained no token
+ * from `REASONING_EFFORT_ORDER` (nothing usable to learn) or the key is unusable.
+ *
+ * Always monotonically decreases: if a cap already stored ranks lower than the
+ * newly computed highest, the stored (lower) value wins and is returned
+ * unchanged. This keeps a later, laxer-looking response (or a race between
+ * concurrent requests) from ratcheting the cap back up.
+ */
+export function recordLearnedReasoningEffort(
+  provider: string | null | undefined,
+  model: string | null | undefined,
+  acceptedValues: string[]
+): string | null {
+  const key = buildKey(provider, model);
+  if (!key) return null;
+
+  let best: string | null = null;
+  let bestRank = -1;
+  for (const raw of acceptedValues) {
+    const rank = rankOf(raw);
+    if (rank > bestRank) {
+      bestRank = rank;
+      best = raw;
+    }
+  }
+  if (best === null) return null;
+
+  const existing = learnedCaps.get(key);
+  if (existing !== undefined && rankOf(existing) <= bestRank) {
+    return existing; // already learned an equal-or-lower cap; keep it
+  }
+  learnedCaps.set(key, best);
+  return best;
+}
+
+// Matches both prose shapes observed: OVH's `@ai-sdk/openai-compatible`
+// deserializer ("expected one of `a`, `b`") and a generic vendor prose form
+// ("Supported types are a, b, and c").
+const LIST_INTRO = /(?:expected one of|supported (?:types|values) are)[:\s]*([^.]+)/i;
+
+/**
+ * Extract the upstream-advertised accepted reasoning_effort values from a 4xx
+ * error body. Returns only tokens present in REASONING_EFFORT_ORDER (unknown
+ * tokens are dropped defensively) in the order they appeared, or null when the
+ * text names no recognized enum member.
+ */
+export function parseReasoningEffortEnum(errText: unknown): string[] | null {
+  if (typeof errText !== "string" || !errText) return null;
+  const match = LIST_INTRO.exec(errText);
+  if (!match) return null;
+  const tokens = match[1]
+    .split(/,|\band\b|&/i)
+    .map((t) =>
+      t
+        .replace(/`/g, "")
+        .replace(/\([^)]*\)/g, "")
+        .trim()
+        .toLowerCase()
+    )
+    .filter((t) => t.length > 0 && REASONING_EFFORT_ORDER.includes(t));
+  return tokens.length > 0 ? tokens : null;
+}
+
+/** Test-only: clear the learned-cap Map between tests. */
+export function __test_resetLearnedReasoningEffortCaps(): void {
+  learnedCaps.clear();
+}
diff --git a/open-sse/services/reasoningInputPolicy.ts b/open-sse/services/reasoningInputPolicy.ts
index 71a283a706..94f27ddcdc 100644
--- a/open-sse/services/reasoningInputPolicy.ts
+++ b/open-sse/services/reasoningInputPolicy.ts
@@ -1,5 +1,6 @@
 import { REGISTRY } from "../config/providerRegistry.ts";
 import type { ReasoningTransport } from "../config/providerRegistry.ts";
+import { isValidResponsesItemId } from "./responsesItemId.ts";
 
 type JsonRecord = Record<string, unknown>;
 
@@ -279,13 +280,27 @@ function sanitizeResponsesInput(
       if (!hasPlaintext && !hasOpaque && (!hasDisplaySummary(next) || stripOrphanedSummaries)) {
         continue;
       }
-      if (!hasOpaque && typeof next.id === "string") delete next.id;
+      // `id` is only worth keeping on an opaque item with a valid string value —
+      // non-opaque items don't replay their id, and a malformed value (e.g. `null`,
+      // observed on opencode/zen) must not survive either way (#11108).
+      if (!hasOpaque || !isValidResponsesItemId(next.id)) delete next.id;
+      // Some upstreams (e.g. opencode/zen) omit `summary` entirely on opaque
+      // reasoning items instead of sending an empty array. Replaying that shape
+      // verbatim trips strict Responses-API validators that require the field
+      // to be present on every `input[]` item of type `reasoning` (#11108).
+      // Plaintext-only items intentionally have no `summary` key and must stay
+      // untouched.
+      if (hasOpaque && next.summary === undefined) next.summary = [];
       filtered.push(next);
       continue;
     }
 
     const cloned = { ...record };
-    if (typeof cloned.id === "string") delete cloned.id;
+    // Strip `id` whenever present, valid or not: these items don't need a
+    // replayed server id, and a malformed one (e.g. `null`, same opencode/zen
+    // omission pattern as the reasoning branch above) must not survive either
+    // (#11108).
+    if (cloned.id !== undefined) delete cloned.id;
     filtered.push(cloned);
   }
   return filtered;
diff --git a/open-sse/services/responsesInputSanitizer.ts b/open-sse/services/responsesInputSanitizer.ts
index 5d81b787a1..94cd99f934 100644
--- a/open-sse/services/responsesInputSanitizer.ts
+++ b/open-sse/services/responsesInputSanitizer.ts
@@ -1,3 +1,5 @@
+import { isValidResponsesItemId } from "./responsesItemId.ts";
+
 type JsonRecord = Record<string, unknown>;
 type SanitizeResponsesInputOptions = {
   dropInternalAssistantMessages?: boolean;
@@ -40,7 +42,12 @@ function sanitizeFunctionName(name: string): string {
 }
 
 function sanitizeInputItemId(record: JsonRecord): JsonRecord {
-  if (typeof record.id !== "string") return record;
+  if (record.id === undefined) return record;
+  if (!isValidResponsesItemId(record.id)) {
+    const next = { ...record };
+    delete next.id;
+    return next;
+  }
 
   const type = typeof record.type === "string" ? record.type : "";
   const expectedPrefix = SERVER_ITEM_ID_PREFIX_BY_TYPE[type];
diff --git a/open-sse/services/responsesItemId.ts b/open-sse/services/responsesItemId.ts
new file mode 100644
index 0000000000..a57ac92e42
--- /dev/null
+++ b/open-sse/services/responsesItemId.ts
@@ -0,0 +1,7 @@
+// Shared by reasoningInputPolicy.ts and responsesInputSanitizer.ts: both strip a
+// Responses-API `input[]` item's `id` field when it isn't a valid string before
+// replay, so a malformed value (e.g. `null`, observed on opencode/zen) never
+// survives to trip a strict upstream with "Expected 'id' to be a string." (#11108).
+export function isValidResponsesItemId(id: unknown): id is string {
+  return typeof id === "string";
+}
diff --git a/open-sse/services/streamRecovery.ts b/open-sse/services/streamRecovery.ts
index a0a397fd87..37a4867f13 100644
--- a/open-sse/services/streamRecovery.ts
+++ b/open-sse/services/streamRecovery.ts
@@ -185,7 +185,21 @@ export interface OpenAiSseScan {
   text: string;
   /** True if any `choices[].delta.tool_calls` appeared — NEVER continue those. */
   sawToolCall: boolean;
-  /** True if a terminal marker (`[DONE]` or a non-null `finish_reason`) appeared. */
+  /**
+   * True only when `tool_calls` appeared in this scan AND its own
+   * `finish_reason: "tool_calls"` has NOT also appeared in the same scan — i.e. the
+   * call is still being streamed (arguments may be mid-flight). Once
+   * `finish_reason: "tool_calls"` closes it, the call is complete, not in flight: the
+   * client has the full arguments and a truncation past this point only drops
+   * trailing prose, which continuation can safely recover.
+   */
+  sawToolCallInFlight: boolean;
+  /**
+   * True if a terminal marker for the OVERALL stream appeared: `[DONE]`, or a
+   * `finish_reason` other than `"tool_calls"`. A `finish_reason: "tool_calls"` ends
+   * that one choice but is not terminal for continuation purposes — the model turn
+   * (and the client-visible SSE) is still eligible to be resumed past it.
+   */
   terminal: boolean;
   /** True if at least one OpenAI-shaped `choices[].delta` was parsed (format gate). */
   parsedOpenAi: boolean;
@@ -199,10 +213,11 @@ export interface OpenAiSseScan {
 export function scanOpenAiSseText(sse: string): OpenAiSseScan {
   let text = "";
   let sawToolCall = false;
+  let toolCallFinished = false;
   let terminal = false;
   let parsedOpenAi = false;
   if (typeof sse !== "string" || sse.length === 0) {
-    return { text, sawToolCall, terminal, parsedOpenAi };
+    return { text, sawToolCall, sawToolCallInFlight: false, terminal, parsedOpenAi };
   }
   for (const line of sse.split("\n")) {
     const trimmed = line.trimStart();
@@ -231,10 +246,17 @@ export function scanOpenAiSseText(sse: string): OpenAiSseScan {
         if (Array.isArray(toolCalls) && toolCalls.length > 0) sawToolCall = true;
       }
       const finishReason = (choice as { finish_reason?: unknown })?.finish_reason;
-      if (finishReason != null) terminal = true;
+      if (finishReason === "tool_calls") {
+        // Ends this one choice, but the overall stream/turn stays continuable —
+        // never counts as the general terminal marker (see OpenAiSseScan.terminal).
+        toolCallFinished = true;
+      } else if (finishReason != null) {
+        terminal = true;
+      }
     }
   }
-  return { text, sawToolCall, terminal, parsedOpenAi };
+  const sawToolCallInFlight = sawToolCall && !toolCallFinished;
+  return { text, sawToolCall, sawToolCallInFlight, terminal, parsedOpenAi };
 }
 
 export interface ContinuableBody {
@@ -369,7 +391,7 @@ export function createRecoverableStream(
   let emittedTail = ""; // raw SSE not yet scanned (awaiting an event boundary)
   let emittedText = ""; // assistant text already delivered to the client
   let emittedTerminal = false;
-  let emittedToolCall = false;
+  let emittedToolCallInFlight = false;
   let emittedParsedOpenAi = false;
 
   // Enqueue to the client and, when continuation is enabled, fold the chunk into the
@@ -388,7 +410,7 @@ export function createRecoverableStream(
     const scan = scanOpenAiSseText(complete);
     emittedText += scan.text;
     if (scan.terminal) emittedTerminal = true;
-    if (scan.sawToolCall) emittedToolCall = true;
+    if (scan.sawToolCallInFlight) emittedToolCallInFlight = true;
     if (scan.parsedOpenAi) emittedParsedOpenAi = true;
   };
 
@@ -402,7 +424,7 @@ export function createRecoverableStream(
     continueEnabled &&
     continuations < maxContinuations &&
     emittedParsedOpenAi &&
-    !emittedToolCall &&
+    !emittedToolCallInFlight &&
     !emittedTerminal &&
     emittedText.length > 0;
 
diff --git a/open-sse/translator/request/openai-responses/toResponses.ts b/open-sse/translator/request/openai-responses/toResponses.ts
index bee5efab1a..988835edea 100644
--- a/open-sse/translator/request/openai-responses/toResponses.ts
+++ b/open-sse/translator/request/openai-responses/toResponses.ts
@@ -201,6 +201,14 @@ export function openaiToOpenAIResponsesRequest(
         input.push({
           type: "reasoning",
           content: [{ type: "reasoning_text", text: reasoning }],
+          // Strict Responses-API upstreams (e.g. opencode/zen) require `summary`
+          // on every `input[]` item of type "reasoning", plaintext or opaque —
+          // omitting it rejects the request with `input[N] missing required
+          // field summary`. This item is always freshly built from a chat
+          // client's plaintext reasoning, so there is no source summary to
+          // preserve; default to an empty array like the replay sanitizer does
+          // for opaque items in reasoningInputPolicy.ts (#11108).
+          summary: [],
         });
       }
 
diff --git a/open-sse/translator/response/openai-responses.ts b/open-sse/translator/response/openai-responses.ts
index 0c59fac4fb..01ad55d72f 100644
--- a/open-sse/translator/response/openai-responses.ts
+++ b/open-sse/translator/response/openai-responses.ts
@@ -866,21 +866,25 @@ export function openaiResponsesToOpenAIResponse(chunk, state) {
 
 function openaiResponsesToOpenAIResponseStream(chunk, state) {
   if (!chunk) {
-    if (
-      state.currentToolCallNeedsNormalization &&
-      state.currentToolCallArgsBuffer &&
-      state.currentToolCallName
-    ) {
-      const toolSchema = state.toolSchemas?.get(state.currentToolCallName);
-      const argsToEmit = stripEmptyOptionalToolArgs(
-        state.currentToolCallArgsBuffer,
-        state.currentToolCallName,
-        toolSchema
-      );
-      const argsStr =
-        typeof argsToEmit === "string" ? argsToEmit : JSON.stringify(argsToEmit ?? {});
-      state.currentToolCallArgsBuffer = "";
-      state.currentToolCallNeedsNormalization = false;
+    // Iterate every still-open call needing schema-aware normalization, not just a
+    // single one — multiple parallel calls can each be pending here if the stream
+    // ends before their output_item.done arrives.
+    const pendingNormalized: Array<{ index: number; argsStr: string }> = [];
+    if (state.toolCallByCallId instanceof Map) {
+      for (const entry of state.toolCallByCallId.values()) {
+        if (entry.needsNormalization && entry.argsBuffer) {
+          const toolSchema = state.toolSchemas?.get(entry.name);
+          const argsToEmit = stripEmptyOptionalToolArgs(entry.argsBuffer, entry.name, toolSchema);
+          pendingNormalized.push({
+            index: entry.index,
+            argsStr: typeof argsToEmit === "string" ? argsToEmit : JSON.stringify(argsToEmit ?? {}),
+          });
+          entry.argsBuffer = "";
+          entry.needsNormalization = false;
+        }
+      }
+    }
+    if (pendingNormalized.length > 0) {
       state.finishReasonSent = true;
       state.finishReason = "tool_calls";
       const common = {
@@ -889,24 +893,21 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
         created: state.created,
         model: state.model || "gpt-4",
       };
-      return [
-        {
-          ...common,
-          choices: [
-            {
-              index: 0,
-              delta: {
-                tool_calls: [{ index: state.toolCallIndex, function: { arguments: argsStr } }],
-              },
-              finish_reason: null,
-            },
-          ],
-        },
-        {
-          ...common,
-          choices: [{ index: 0, delta: {}, finish_reason: "tool_calls" }],
-        },
-      ];
+      const chunks: Record<string, unknown>[] = pendingNormalized.map(({ index, argsStr }) => ({
+        ...common,
+        choices: [
+          {
+            index: 0,
+            delta: { tool_calls: [{ index, function: { arguments: argsStr } }] },
+            finish_reason: null,
+          },
+        ],
+      }));
+      chunks.push({
+        ...common,
+        choices: [{ index: 0, delta: {}, finish_reason: "tool_calls" }],
+      });
+      return chunks;
     }
     // Flush: send final chunk with finish_reason
     if (!state.finishReasonSent && state.started) {
@@ -952,7 +953,23 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
     state.chatId = `chatcmpl-${Date.now()}`;
     state.created = Math.floor(Date.now() / 1000);
     state.toolCallIndex = 0;
+    // Kept for computeFinishReason (synthesizeCompletedToolCalls.ts) compatibility —
+    // that snapshot path mutates it directly and expects it to exist. In a turn with
+    // multiple parallel calls this only ever reflects the LAST one opened/closed, so
+    // it must never be used to identify a specific call — only as the "is at least
+    // one tool call in flight this turn" signal computeFinishReason needs, which
+    // toolCallIndex > 0 already covers on its own once any call has been added.
     state.currentToolCallId = null;
+    // Per-call state keyed by call_id (replaces the old singular
+    // currentToolCallId/ArgsBuffer/Name/NeedsNormalization/Deferred fields, which
+    // assumed only one function_call could ever be in flight at a time).
+    state.toolCallByCallId = new Map();
+    // response.function_call_arguments.delta carries `item_id`/`output_index`, not
+    // `call_id` — resolve either one back to the call_id key used by
+    // toolCallByCallId (two independent reverse maps, since some upstreams omit
+    // item_id on delta events but still send output_index).
+    state.toolCallItemToCallId = new Map();
+    state.toolCallOutputIndexToCallId = new Map();
   }
 
   // Text content delta
@@ -983,22 +1000,48 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
   // Function call started
   if (eventType === "response.output_item.added" && data.item?.type === "function_call") {
     const item = data.item;
-    state.currentToolCallId = item.call_id || fallbackToolCallId();
-    state.currentToolCallArgsBuffer = ""; // reset per-call arg buffer
-    state.currentToolCallDeferred = false;
+    const callId = item.call_id || fallbackToolCallId();
+    // Kept for computeFinishReason (synthesizeCompletedToolCalls.ts) compatibility.
+    state.currentToolCallId = callId;
+
+    const toolName = normalizeToolName(item.name);
+    // Assign this call's index NOW, at .added, not at .done — two calls opened before
+    // either closes (a genuine parallel dispatch) must never share an index. Deferred
+    // (still-nameless) calls are the one exception: they don't claim an index until
+    // .done resolves a real name, so a call that never gets one never burns a slot
+    // another call could have used.
+    let index: number | null = null;
+    if (toolName) {
+      index = state.toolCallIndex ?? 0;
+      state.toolCallIndex = index + 1;
+    }
+
+    if (!(state.toolCallByCallId instanceof Map)) state.toolCallByCallId = new Map();
+    state.toolCallByCallId.set(callId, {
+      index,
+      name: toolName,
+      argsBuffer: "",
+      deferred: !toolName,
+      needsNormalization: toolName === "Agent",
+    });
+    if (!(state.toolCallItemToCallId instanceof Map)) state.toolCallItemToCallId = new Map();
+    if (item.id) state.toolCallItemToCallId.set(item.id, callId);
+    // `output_index` is a top-level field on every Responses API streamed event
+    // (response.output_item.added/.done AND function_call_arguments.delta alike) —
+    // an identifier independent of item_id, for upstreams that omit item_id on delta
+    // events.
+    if (!(state.toolCallOutputIndexToCallId instanceof Map)) {
+      state.toolCallOutputIndexToCallId = new Map();
+    }
+    if (data.output_index != null) state.toolCallOutputIndexToCallId.set(data.output_index, callId);
 
     // Track this call_id so response.completed doesn't synthesize a duplicate
     if (!state.toolCallIdsSeen) state.toolCallIdsSeen = new Set();
-    if (state.currentToolCallId) state.toolCallIdsSeen.add(state.currentToolCallId);
+    state.toolCallIdsSeen.add(callId);
 
-    const toolName = normalizeToolName(item.name);
-    state.currentToolName = toolName; // track for schema lookup at done time
-    state.currentToolCallName = toolName;
-    state.currentToolCallNeedsNormalization = toolName === "Agent";
     if (!toolName) {
       // Some Responses providers briefly emit placeholder/empty tool names.
       // Defer emission until output_item.done in case the final name is populated there.
-      state.currentToolCallDeferred = true;
       return null;
     }
 
@@ -1013,8 +1056,8 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
           delta: {
             tool_calls: [
               {
-                index: state.toolCallIndex,
-                id: state.currentToolCallId,
+                index,
+                id: callId,
                 type: "function",
                 function: {
                   name: toolName,
@@ -1037,11 +1080,26 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
     const argsDelta = data.delta || "";
     if (!argsDelta) return null;
 
-    state.currentToolCallArgsBuffer = (state.currentToolCallArgsBuffer || "") + argsDelta;
-    if (state.currentToolCallDeferred || state.currentToolCallNeedsNormalization) return null;
+    // Resolve which in-flight call this delta belongs to. Try item_id first (the
+    // field the Responses API documents for this event), then output_index (also a
+    // top-level field on this event, and independent of item_id — covers upstreams
+    // that omit item_id on delta events but still send output_index). Only once both
+    // identifying fields are absent/unresolved do we fall back to guessing (the
+    // single open call, or the most recently opened one as a last resort).
+    const map = state.toolCallByCallId instanceof Map ? state.toolCallByCallId : null;
+    let callId = data.item_id ? state.toolCallItemToCallId?.get(data.item_id) : undefined;
+    if (!callId && data.output_index != null) {
+      callId = state.toolCallOutputIndexToCallId?.get(data.output_index);
+    }
+    if (!callId && map) {
+      callId = map.size === 1 ? [...map.keys()][0] : state.currentToolCallId;
+    }
+    const entry = callId ? map?.get(callId) : undefined;
+    if (!entry) return null;
 
     // #9168: buffer arguments until output_item.done for schema-aware null normalization
     // Previously emitted raw null values for optional enum fields (e.g. isolation: null).
+    entry.argsBuffer = (entry.argsBuffer || "") + argsDelta;
     return null;
   }
 
@@ -1061,13 +1119,30 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
   // carry the complete arguments only in output_item.done (no preceding delta events).
   if (eventType === "response.output_item.done" && data.item?.type === "function_call") {
     const item = data.item;
-    const buffered = state.currentToolCallArgsBuffer || "";
-    const currentIndex = state.toolCallIndex; // capture before increment
-    const callId = item.call_id || state.currentToolCallId || fallbackToolCallId();
+    const map = state.toolCallByCallId instanceof Map ? state.toolCallByCallId : null;
+    let callId = item.call_id;
+    if (!callId && item.id) callId = state.toolCallItemToCallId?.get(item.id);
+    if (!callId) callId = state.currentToolCallId || fallbackToolCallId();
+    const trackedEntry = callId ? map?.get(callId) : undefined;
+    // Some upstreams (e.g. Codex) send the complete payload only in output_item.done,
+    // with no preceding output_item.added at all — there is no tracked entry to read an
+    // index from.
+    const entry = trackedEntry || { index: null, argsBuffer: "", deferred: false };
+
+    const buffered = entry.argsBuffer || "";
     const toolName = normalizeToolName(item.name);
+
+    // Claim (and advance) this call's index now if it wasn't assigned at .added — either
+    // a deferred call whose name has just now resolved, or a Codex-style done-only
+    // payload that never had an .added at all. A deferred call whose name is STILL empty
+    // never claims an index (nothing was ever emitted for it either way).
+    if (entry.index == null && toolName) {
+      entry.index = state.toolCallIndex ?? 0;
+      state.toolCallIndex = entry.index + 1;
+    }
+    const currentIndex = entry.index;
     const toolSchema = state.toolSchemas?.get(toolName);
     const shouldNormalizeArguments = toolName === "Agent";
-    state.currentToolCallNeedsNormalization = shouldNormalizeArguments;
 
     if (toolName && state.toolCalls instanceof Map) {
       const completedArguments =
@@ -1077,6 +1152,9 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
         toolName,
         toolSchema
       );
+      // Keyed by index, not insertion order — readers that need call order for
+      // parallel calls closed out of order should sort by this key rather than
+      // relying on Map iteration order.
       state.toolCalls.set(currentIndex, {
         id: callId,
         index: currentIndex,
@@ -1095,17 +1173,17 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
     if (!state.toolCallIdsSeen) state.toolCallIdsSeen = new Set();
     if (callId) state.toolCallIdsSeen.add(callId);
 
-    if (state.currentToolCallDeferred) {
-      state.currentToolCallDeferred = false;
-      state.currentToolCallArgsBuffer = "";
-      state.currentToolCallId = null;
+    // This call is fully closed — remove it from the in-flight map (bounds the map
+    // to genuinely in-flight calls, and keeps the single-open-call fallback in the
+    // function_call_arguments.delta handler correct for whichever call opens next).
+    if (map && callId) map.delete(callId);
+    if (state.currentToolCallId === callId) state.currentToolCallId = null;
 
+    if (entry.deferred) {
       if (!toolName) {
         return null;
       }
 
-      state.toolCallIndex++;
-
       const terminalArguments =
         typeof item.arguments === "string"
           ? item.arguments.length > 0
@@ -1148,12 +1226,7 @@ function openaiResponsesToOpenAIResponseStream(chunk, state) {
       };
     }
 
-    state.toolCallIndex++;
-    state.currentToolCallArgsBuffer = ""; // reset for next tool call
-    state.currentToolCallId = null;
-    const needsNormalization = state.currentToolCallNeedsNormalization === true;
-    state.currentToolCallNeedsNormalization = false;
-    state.currentToolCallName = "";
+    const needsNormalization = shouldNormalizeArguments;
 
     // Nullable omission sentinels must be normalized before any argument bytes reach the client.
     // Other tool calls retain immediate argument streaming.
diff --git a/open-sse/utils/streamPayloadCollector.ts b/open-sse/utils/streamPayloadCollector.ts
index ea1b2e658e..31b9e818f4 100644
--- a/open-sse/utils/streamPayloadCollector.ts
+++ b/open-sse/utils/streamPayloadCollector.ts
@@ -337,6 +337,7 @@ function createOpenAIReducer(fallbackModel?: string | null): SummaryReducer {
       // same-name tool_calls) into its own separate tool_calls entries.
       const finalToolCalls: ToolCall[] = [];
       let nextIndex = 0;
+      // Normalize tool_call indexes to contiguous 0-based (OpenAI contract).
       for (const tc of mergedToolCalls) {
         const splitArgs = splitConcatenatedToolCallArguments(tc.function.arguments);
         if (!splitArgs) {
diff --git a/package-lock.json b/package-lock.json
index 57f810910e..bf77bab602 100644
--- a/package-lock.json
+++ b/package-lock.json
@@ -18,7 +18,6 @@
         "@dnd-kit/core": "^6.3.1",
         "@dnd-kit/sortable": "^10.0.0",
         "@dnd-kit/utilities": "^3.2.2",
-        "@huggingface/transformers": "^4.2.0",
         "@lobehub/icons": "^5.16.0",
         "@modelcontextprotocol/sdk": "^1.29.0",
         "@monaco-editor/react": "^4.7.0",
@@ -61,7 +60,6 @@
         "next-themes": "^0.4.6",
         "node-machine-id": "^1.1.12",
         "omniglyph": "^1.4.0",
-        "onnxruntime-node": "1.24.3",
         "open": "^11.0.1",
         "ora": "^9.4.1",
         "parse5": "^8.0.1",
@@ -156,9 +154,11 @@
       },
       "optionalDependencies": {
         "@atjsh/llmlingua-2": "3.0.0",
+        "@huggingface/transformers": "^4.2.0",
         "better-sqlite3": "^13.0.2",
         "js-tiktoken": "^1.0.20",
         "keytar": "^7.9.0",
+        "onnxruntime-node": "1.24.3",
         "sqlite-vec": "^0.1.9",
         "tls-client-node": "^0.2.0",
         "wreq-js": "^3.0.0"
@@ -4510,6 +4510,7 @@
       "resolved": "https://registry.npmjs.org/@huggingface/jinja/-/jinja-0.5.9.tgz",
       "integrity": "sha512-uWTG+l3VJRsl7EXxYizuL3P+cCPoc3cRqbWWRcQN0FhejRfbdq0RNhCmbY/YDtnTcz9icdLYuLDjsnz4d8JMuw==",
       "license": "MIT",
+      "optional": true,
       "engines": {
         "node": ">=18"
       }
@@ -4518,13 +4519,15 @@
       "version": "0.1.3",
       "resolved": "https://registry.npmjs.org/@huggingface/tokenizers/-/tokenizers-0.1.3.tgz",
       "integrity": "sha512-8rF/RRT10u+kn7YuUbUg0OF30K8rjTc78aHpxT+qJ1uWSqxT1MHi8+9ltwYfkFYJzT/oS+qw3JVfHtNMGAdqyA==",
-      "license": "Apache-2.0"
+      "license": "Apache-2.0",
+      "optional": true
     },
     "node_modules/@huggingface/transformers": {
       "version": "4.2.0",
       "resolved": "https://registry.npmjs.org/@huggingface/transformers/-/transformers-4.2.0.tgz",
       "integrity": "sha512-8BRCoBMH0XsWaEIamuR0LrJGAfftgHAfb2Vrffy0VKlSAE/MnUJ5/h/zTfEP3fDIft+nk7TqB8xXEyABGitBjQ==",
       "license": "Apache-2.0",
+      "optional": true,
       "dependencies": {
         "@huggingface/jinja": "^0.5.6",
         "@huggingface/tokenizers": "^0.1.3",
@@ -9483,30 +9486,35 @@
       "version": "1.1.2",
       "resolved": "https://registry.npmjs.org/@protobufjs/aspromise/-/aspromise-1.1.2.tgz",
       "integrity": "sha512-j+gKExEuLmKwvz3OgROXtrJ2UG2x8Ch2YZUxahh+s1F2HZ+wAceUNLkvy6zKCPVRkU++ZWQrdxsUeQXmcg4uoQ==",
+      "devOptional": true,
       "license": "BSD-3-Clause"
     },
     "node_modules/@protobufjs/base64": {
       "version": "1.1.2",
       "resolved": "https://registry.npmjs.org/@protobufjs/base64/-/base64-1.1.2.tgz",
       "integrity": "sha512-AZkcAA5vnN/v4PDqKyMR5lx7hZttPDgClv83E//FMNhR2TMcLUhfRUBHCmSl0oi9zMgDDqRUJkSxO3wm85+XLg==",
+      "devOptional": true,
       "license": "BSD-3-Clause"
     },
     "node_modules/@protobufjs/codegen": {
       "version": "2.0.5",
       "resolved": "https://registry.npmjs.org/@protobufjs/codegen/-/codegen-2.0.5.tgz",
       "integrity": "sha512-zgXFLzW3Ap33e6d0Wlj4MGIm6Ce8O89n/apUaGNB/jx+hw+ruWEp7EwGUshdLKVRCxZW12fp9r40E1mQrf/34g==",
+      "devOptional": true,
       "license": "BSD-3-Clause"
     },
     "node_modules/@protobufjs/eventemitter": {
       "version": "1.1.1",
       "resolved": "https://registry.npmjs.org/@protobufjs/eventemitter/-/eventemitter-1.1.1.tgz",
       "integrity": "sha512-vW1GmwMZNnL+gMRaovlh9yZX74kc+TTU3FObkkurpMaRtBfLP3ldjS9KQWlwZgraRE0+dheEEoAxdzcJQ8eXZg==",
+      "devOptional": true,
       "license": "BSD-3-Clause"
     },
     "node_modules/@protobufjs/fetch": {
       "version": "1.1.1",
       "resolved": "https://registry.npmjs.org/@protobufjs/fetch/-/fetch-1.1.1.tgz",
       "integrity": "sha512-GpptLrs57adMSuHi3VNj0mAF8dwh36LMaYF6XyJ6JMWlVsc+t42tm1HSEDmOs3A8fC9yyeisgLhsTVQokOZ0zw==",
+      "devOptional": true,
       "license": "BSD-3-Clause",
       "dependencies": {
         "@protobufjs/aspromise": "^1.1.1"
@@ -9516,24 +9524,28 @@
       "version": "1.0.2",
       "resolved": "https://registry.npmjs.org/@protobufjs/float/-/float-1.0.2.tgz",
       "integrity": "sha512-Ddb+kVXlXst9d+R9PfTIxh1EdNkgoRe5tOX6t01f1lYWOvJnSPDBlG241QLzcyPdoNTsblLUdujGSE4RzrTZGQ==",
+      "devOptional": true,
       "license": "BSD-3-Clause"
     },
     "node_modules/@protobufjs/path": {
       "version": "1.1.2",
       "resolved": "https://registry.npmjs.org/@protobufjs/path/-/path-1.1.2.tgz",
       "integrity": "sha512-6JOcJ5Tm08dOHAbdR3GrvP+yUUfkjG5ePsHYczMFLq3ZmMkAD98cDgcT2iA1lJ9NVwFd4tH/iSSoe44YWkltEA==",
+      "devOptional": true,
       "license": "BSD-3-Clause"
     },
     "node_modules/@protobufjs/pool": {
       "version": "1.1.0",
       "resolved": "https://registry.npmjs.org/@protobufjs/pool/-/pool-1.1.0.tgz",
       "integrity": "sha512-0kELaGSIDBKvcgS4zkjz1PeddatrjYcmMWOlAuAPwAeccUrPHdUqo/J6LiymHHEiJT5NrF1UVwxY14f+fy4WQw==",
+      "devOptional": true,
       "license": "BSD-3-Clause"
     },
     "node_modules/@protobufjs/utf8": {
       "version": "1.1.1",
       "resolved": "https://registry.npmjs.org/@protobufjs/utf8/-/utf8-1.1.1.tgz",
       "integrity": "sha512-oOAWABowe8EAbMyWKM0tYDKi8Yaox52D+HWZhAIJqQXbqe0xI/GV7FhLWqlEKreMkfDjshR5FKgi3mnle0h6Eg==",
+      "devOptional": true,
       "license": "BSD-3-Clause"
     },
     "node_modules/@radix-ui/number": {
@@ -12737,6 +12749,7 @@
       "version": "26.2.0",
       "resolved": "https://registry.npmjs.org/@types/node/-/node-26.2.0.tgz",
       "integrity": "sha512-5IviulTZeRNp2vAJ514cc/HUlY5nZ9fCbq9DMyC52BrhFZACo3nI0R7qBxhQmo/d27NFe96ur/b7Wwxklda+kg==",
+      "devOptional": true,
       "license": "MIT",
       "dependencies": {
         "undici-types": "~8.3.0"
@@ -13998,6 +14011,7 @@
       "resolved": "https://registry.npmjs.org/adm-zip/-/adm-zip-0.6.0.tgz",
       "integrity": "sha512-XleryMhbuksdKtofnWZ9Sk+4CUTbms4Mb/EU32SZwToAyZ5RgVos/ki8n+yr0LWHOGKuakbXTuuYNHLQjhddgg==",
       "license": "MIT",
+      "optional": true,
       "engines": {
         "node": ">=14.0"
       }
@@ -14971,7 +14985,8 @@
       "resolved": "https://registry.npmjs.org/boolean/-/boolean-3.2.0.tgz",
       "integrity": "sha512-d0II/GO9uf9lfUHH2BQsjxzRJZBdsjgsBiW4BvhWk/3qoKwQFjIDVN19PfX8F2D/r9PCMTtLWjYVCFrpeYUzsw==",
       "deprecated": "Package no longer supported. Contact Support at https://www.npmjs.com/support for more info.",
-      "license": "MIT"
+      "license": "MIT",
+      "optional": true
     },
     "node_modules/bottleneck": {
       "version": "2.19.5",
@@ -17935,6 +17950,7 @@
       "version": "1.1.4",
       "resolved": "https://registry.npmjs.org/define-data-property/-/define-data-property-1.1.4.tgz",
       "integrity": "sha512-rBMvIzlpA8v6E+SJZoo++HAYqsLrkg7MSfIinMPFhmkorw7X+dOXVJQs+QT69zGkzMyfDnIMN2Wid1+NbL3T+A==",
+      "devOptional": true,
       "license": "MIT",
       "dependencies": {
         "es-define-property": "^1.0.0",
@@ -17964,6 +17980,7 @@
       "version": "1.2.1",
       "resolved": "https://registry.npmjs.org/define-properties/-/define-properties-1.2.1.tgz",
       "integrity": "sha512-8QmQKqEASLd5nx0U1B1okLElbUuuttJ/AnYmRXbbbGDWh6uS208EjD4Xqq/I9wK7u0v6O08XhTWnt5XtEbR6Dg==",
+      "devOptional": true,
       "license": "MIT",
       "dependencies": {
         "define-data-property": "^1.0.1",
@@ -18064,7 +18081,8 @@
       "version": "2.1.0",
       "resolved": "https://registry.npmjs.org/detect-node/-/detect-node-2.1.0.tgz",
       "integrity": "sha512-T0NIuQpnTvFDATNuHN5roPwSBG83rFsuO+MXXH9/3N1eFbn4wcPjttvjMLEPWJ0RGUYgQE7cGgS3tNxbqCGM7g==",
-      "license": "MIT"
+      "license": "MIT",
+      "optional": true
     },
     "node_modules/detect-node-es": {
       "version": "1.1.0",
@@ -18908,7 +18926,8 @@
       "version": "4.1.1",
       "resolved": "https://registry.npmjs.org/es6-error/-/es6-error-4.1.1.tgz",
       "integrity": "sha512-Um/+FxMr9CISWh0bi5Zv0iOD+4cFh5qLeks1qhAopKVAJw3drgKbKySikp7wGhDL0HPeaja0P5ULZrxLkniUVg==",
-      "license": "MIT"
+      "license": "MIT",
+      "optional": true
     },
     "node_modules/es6-promisify": {
       "version": "7.0.0",
@@ -20400,7 +20419,8 @@
       "version": "25.9.23",
       "resolved": "https://registry.npmjs.org/flatbuffers/-/flatbuffers-25.9.23.tgz",
       "integrity": "sha512-MI1qs7Lo4Syw0EOzUl0xjs2lsoeqFku44KpngfIduHBYvzm8h2+7K8YMQh1JtVVVrUvhLpNwqVi4DERegUJhPQ==",
-      "license": "Apache-2.0"
+      "license": "Apache-2.0",
+      "optional": true
     },
     "node_modules/flatted": {
       "version": "3.4.2",
@@ -21226,6 +21246,7 @@
       "resolved": "https://registry.npmjs.org/global-agent/-/global-agent-3.0.0.tgz",
       "integrity": "sha512-PT6XReJ+D07JvGoxQMkT6qji/jVNfX/h364XHZOWeRzy64sSFr+xJ5OX7LI3b4MPQzdL4H8Y8M0xzPpsVMwA8Q==",
       "license": "BSD-3-Clause",
+      "optional": true,
       "dependencies": {
         "boolean": "^3.0.1",
         "es6-error": "^4.1.1",
@@ -21243,6 +21264,7 @@
       "resolved": "https://registry.npmjs.org/semver/-/semver-7.8.5.tgz",
       "integrity": "sha512-Y7/KDsb8LjooZpwaqGyulO6DQlksgCncchHGk+sZIY4SBvUocMBEFH5Ur1fI4dV+Jvl0w6cjvucaIi40puRioA==",
       "license": "ISC",
+      "optional": true,
       "bin": {
         "semver": "bin/semver.js"
       },
@@ -21291,6 +21313,7 @@
       "version": "1.0.4",
       "resolved": "https://registry.npmjs.org/globalthis/-/globalthis-1.0.4.tgz",
       "integrity": "sha512-DpLKbNU4WylpxJykQujfCcwYWiV/Jhm50Goo0wrVILAv5jOr9d+H+UR3PhSCD2rCCEIg0uc+G+muBTwD54JhDQ==",
+      "devOptional": true,
       "license": "MIT",
       "dependencies": {
         "define-properties": "^1.2.1",
@@ -21645,7 +21668,8 @@
       "version": "1.0.9",
       "resolved": "https://registry.npmjs.org/guid-typescript/-/guid-typescript-1.0.9.tgz",
       "integrity": "sha512-Y8T4vYhEfwJOTbouREvG+3XDsjr8E3kIr7uf+JZ0BYloFsttiHU0WfvANVsR7TxNUJa/WpCnw/Ino/p+DeBhBQ==",
-      "license": "ISC"
+      "license": "ISC",
+      "optional": true
     },
     "node_modules/hachure-fill": {
       "version": "0.5.2",
@@ -21679,6 +21703,7 @@
       "version": "1.0.2",
       "resolved": "https://registry.npmjs.org/has-property-descriptors/-/has-property-descriptors-1.0.2.tgz",
       "integrity": "sha512-55JNKuIW+vq4Ke1BjOTjM2YctQIvCT7GFzHwmfZPGo5wnrgkid0YQtnAleFSqumZm4az3n2BS+erby5ipJdgrg==",
+      "devOptional": true,
       "license": "MIT",
       "dependencies": {
         "es-define-property": "^1.0.0"
@@ -24790,7 +24815,8 @@
       "version": "5.0.1",
       "resolved": "https://registry.npmjs.org/json-stringify-safe/-/json-stringify-safe-5.0.1.tgz",
       "integrity": "sha512-ZClg6AaYvamvYEE82d3Iyd3vSSIjQ+odgjaTzRuO3s7toCdFKczob2i0zCh7JE8kWn17yvAWhUVxvqGwUalsRA==",
-      "license": "ISC"
+      "license": "ISC",
+      "optional": true
     },
     "node_modules/json5": {
       "version": "2.2.3",
@@ -26654,6 +26680,7 @@
       "resolved": "https://registry.npmjs.org/matcher/-/matcher-3.0.0.tgz",
       "integrity": "sha512-OkeDaAZ/bQCxeFAozM55PKcKU0yJMPGifLwV4Qgjitu+5MoAfSQN4lsLJeXZ1b8w0x+/Emda6MZgXS1jvsapng==",
       "license": "MIT",
+      "optional": true,
       "dependencies": {
         "escape-string-regexp": "^4.0.0"
       },
@@ -29425,6 +29452,7 @@
       "version": "1.1.1",
       "resolved": "https://registry.npmjs.org/object-keys/-/object-keys-1.1.1.tgz",
       "integrity": "sha512-NuAESUOUMrlIXOfHKzD6bpPu3tYt3xvjNdRIQ+FeT0lNb4K8WR70CaDxhuNguS2XG+GjkyMwOzsN5ZktImfhLA==",
+      "devOptional": true,
       "license": "MIT",
       "engines": {
         "node": ">= 0.4"
@@ -29632,7 +29660,8 @@
       "version": "1.24.3",
       "resolved": "https://registry.npmjs.org/onnxruntime-common/-/onnxruntime-common-1.24.3.tgz",
       "integrity": "sha512-GeuPZO6U/LBJXvwdaqHbuUmoXiEdeCjWi/EG7Y1HNnDwJYuk6WUbNXpF6luSUY8yASul3cmUlLGrCCL1ZgVXqA==",
-      "license": "MIT"
+      "license": "MIT",
+      "optional": true
     },
     "node_modules/onnxruntime-node": {
       "version": "1.24.3",
@@ -29640,6 +29669,7 @@
       "integrity": "sha512-JH7+czbc8ALA819vlTgcV+Q214/+VjGeBHDjX81+ZCD0PCVCIFGFNtT0V4sXG/1JXypKPgScQcB3ij/hk3YnTg==",
       "hasInstallScript": true,
       "license": "MIT",
+      "optional": true,
       "os": [
         "win32",
         "darwin",
@@ -29656,6 +29686,7 @@
       "resolved": "https://registry.npmjs.org/onnxruntime-web/-/onnxruntime-web-1.26.0-dev.20260416-b7804b056c.tgz",
       "integrity": "sha512-MD6Ss4GSpQBo6zqoJzyT9LRbKYs7x/JVN23FT24EcEvlqF4VuzPOeH6X38orZPKHQDbprn7K+SBpu0/mj2CQiw==",
       "license": "MIT",
+      "optional": true,
       "dependencies": {
         "flatbuffers": "^25.1.24",
         "guid-typescript": "^1.0.9",
@@ -29669,13 +29700,15 @@
       "version": "5.3.2",
       "resolved": "https://registry.npmjs.org/long/-/long-5.3.2.tgz",
       "integrity": "sha512-mNAgZ1GmyNhD7AuqnTG3/VQ26o760+ZYBPKjPvugO8+nLbYfX6TVpJPseBvopbdY+qpZ/lKUnmEc1LeZYS3QAA==",
-      "license": "Apache-2.0"
+      "license": "Apache-2.0",
+      "optional": true
     },
     "node_modules/onnxruntime-web/node_modules/onnxruntime-common": {
       "version": "1.24.0-dev.20251116-b39e144322",
       "resolved": "https://registry.npmjs.org/onnxruntime-common/-/onnxruntime-common-1.24.0-dev.20251116-b39e144322.tgz",
       "integrity": "sha512-BOoomdHYmNRL5r4iQ4bMvsl2t0/hzVQ3OM3PHD0gxeXu1PmggqBv3puZicEUVOA3AtHHYmqZtjMj9FOfGrATTw==",
-      "license": "MIT"
+      "license": "MIT",
+      "optional": true
     },
     "node_modules/open": {
       "version": "11.0.1",
@@ -30906,7 +30939,8 @@
       "version": "1.3.6",
       "resolved": "https://registry.npmjs.org/platform/-/platform-1.3.6.tgz",
       "integrity": "sha512-fnWVljUchTro6RiCFvCXBbNhJc2NijN7oIQxbwsyL0buWJPG85v81ehlHI9fXrJsMNgTofEoWIQeClKpgxFLrg==",
-      "license": "MIT"
+      "license": "MIT",
+      "optional": true
     },
     "node_modules/playwright": {
       "version": "1.62.1",
@@ -31861,6 +31895,7 @@
       "version": "7.6.5",
       "resolved": "https://registry.npmjs.org/protobufjs/-/protobufjs-7.6.5.tgz",
       "integrity": "sha512-/FPD0nUc9jH6rfFjji9IBqOz4pcSE3CsT1m7Ep6Mdb0LxSUMj8hgl6GomOvZzpNpAqqGaXA0P3VSrZLFzIhQrw==",
+      "devOptional": true,
       "hasInstallScript": true,
       "license": "BSD-3-Clause",
       "dependencies": {
@@ -31884,6 +31919,7 @@
       "version": "5.3.2",
       "resolved": "https://registry.npmjs.org/long/-/long-5.3.2.tgz",
       "integrity": "sha512-mNAgZ1GmyNhD7AuqnTG3/VQ26o760+ZYBPKjPvugO8+nLbYfX6TVpJPseBvopbdY+qpZ/lKUnmEc1LeZYS3QAA==",
+      "devOptional": true,
       "license": "Apache-2.0"
     },
     "node_modules/proxy-addr": {
@@ -33280,6 +33316,7 @@
       "resolved": "https://registry.npmjs.org/roarr/-/roarr-2.15.4.tgz",
       "integrity": "sha512-CHhPh+UNHD2GTXNYhPWLnU8ONHdI+5DI+4EYIAOaiD63rHeYlZvyh8P+in5999TTSFgUYuKUAjzRI4mdh/p+2A==",
       "license": "BSD-3-Clause",
+      "optional": true,
       "dependencies": {
         "boolean": "^3.0.1",
         "detect-node": "^2.0.4",
@@ -33655,7 +33692,8 @@
       "version": "1.0.0",
       "resolved": "https://registry.npmjs.org/semver-compare/-/semver-compare-1.0.0.tgz",
       "integrity": "sha512-YM3/ITh2MJ5MtzaM429anh+x2jiLVjqILF4m4oyQB18W7Ggea7BfqdH/wGMK7dDiMghv/6WG7znWMwUDzJiXow==",
-      "license": "MIT"
+      "license": "MIT",
+      "optional": true
     },
     "node_modules/send": {
       "version": "1.2.1",
@@ -33688,6 +33726,7 @@
       "resolved": "https://registry.npmjs.org/serialize-error/-/serialize-error-7.0.1.tgz",
       "integrity": "sha512-8I8TjW5KMOKsZQTvoxjuSIa7foAwPWGOts+6o7sgjz41/qMD9VQHEDxi6PBvK2l0MXUmqZyNpUK+T2tQaaElvw==",
       "license": "MIT",
+      "optional": true,
       "dependencies": {
         "type-fest": "^0.13.1"
       },
@@ -33703,6 +33742,7 @@
       "resolved": "https://registry.npmjs.org/type-fest/-/type-fest-0.13.1.tgz",
       "integrity": "sha512-34R7HTnG0XIJcBSn5XhDd7nNFPRcXYRZrBB2O2jdKqYODldSzBAqzsWoZYYvduky73toYS/ESqxPvkDf/F0XMg==",
       "license": "(MIT OR CC0-1.0)",
+      "optional": true,
       "engines": {
         "node": ">=10"
       },
@@ -34474,7 +34514,8 @@
       "version": "1.1.3",
       "resolved": "https://registry.npmjs.org/sprintf-js/-/sprintf-js-1.1.3.tgz",
       "integrity": "sha512-Oo+0REFV59/rz3gfJNKQiBlwfHaSESl1pcGyABQsnnIfWOFt6JNj5gCog2U6MLZ//IGYD+nA8nI+mTShREReaA==",
-      "license": "BSD-3-Clause"
+      "license": "BSD-3-Clause",
+      "optional": true
     },
     "node_modules/sql.js": {
       "version": "1.14.2",
@@ -36187,6 +36228,7 @@
       "version": "8.3.0",
       "resolved": "https://registry.npmjs.org/undici-types/-/undici-types-8.3.0.tgz",
       "integrity": "sha512-j375ScV60dom+YkPFIfTLcOiPxkN/buHz5GobjLhixFuANaNs3C9l4GmrWqejgXWJ7BbJcFYpTEUkS1Ge8bpZQ==",
+      "devOptional": true,
       "license": "MIT"
     },
     "node_modules/unicode-emoji-modifier-base": {
diff --git a/package.json b/package.json
index bf2740f233..55cc574f9e 100644
--- a/package.json
+++ b/package.json
@@ -265,7 +265,6 @@
     "@dnd-kit/core": "^6.3.1",
     "@dnd-kit/sortable": "^10.0.0",
     "@dnd-kit/utilities": "^3.2.2",
-    "@huggingface/transformers": "^4.2.0",
     "@lobehub/icons": "^5.16.0",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@monaco-editor/react": "^4.7.0",
@@ -308,7 +307,6 @@
     "next-themes": "^0.4.6",
     "node-machine-id": "^1.1.12",
     "omniglyph": "^1.4.0",
-    "onnxruntime-node": "1.24.3",
     "open": "^11.0.1",
     "ora": "^9.4.1",
     "parse5": "^8.0.1",
@@ -343,9 +341,11 @@
   },
   "optionalDependencies": {
     "@atjsh/llmlingua-2": "3.0.0",
+    "@huggingface/transformers": "^4.2.0",
     "better-sqlite3": "^13.0.2",
     "js-tiktoken": "^1.0.20",
     "keytar": "^7.9.0",
+    "onnxruntime-node": "1.24.3",
     "sqlite-vec": "^0.1.9",
     "tls-client-node": "^0.2.0",
     "wreq-js": "^3.0.0"
diff --git a/skills/omni-webhooks/SKILL.md b/skills/omni-webhooks/SKILL.md
index 251b5f5817..c60df46eca 100644
--- a/skills/omni-webhooks/SKILL.md
+++ b/skills/omni-webhooks/SKILL.md
@@ -1,12 +1,12 @@
 ---
 name: omni-webhooks
-description: Register, list, test, and remove webhook endpoints. Configure event subscriptions (request.completed, provider.error, budget.exceeded, etc.) and manage delivery retries.
+description: Register, list, test, and remove webhook endpoints. Configure event subscriptions (request.completed, request.failed, quota.exceeded, etc.) and manage delivery retries.
 ---
 <!-- generated by src/lib/agentSkills/generator.ts; manual edits will be overwritten -->
 
 ## Overview
 
-Register, list, test, and remove webhook endpoints. Configure event subscriptions (request.completed, provider.error, budget.exceeded, etc.) and manage delivery retries.
+Register, list, test, and remove webhook endpoints. Configure event subscriptions (request.completed, request.failed, quota.exceeded, etc.) and manage delivery retries.
 
 ## Authentication
 
diff --git a/src/app/(dashboard)/dashboard/analytics/ComboHealthTab.tsx b/src/app/(dashboard)/dashboard/analytics/ComboHealthTab.tsx
index 7970c40a8e..8d8c5c5573 100644
--- a/src/app/(dashboard)/dashboard/analytics/ComboHealthTab.tsx
+++ b/src/app/(dashboard)/dashboard/analytics/ComboHealthTab.tsx
@@ -291,7 +291,9 @@ function ComboAutopilotPanel({ report }: { report: ComboAutopilotReport }) {
             icon="monitor_heart"
             label={t("comboHealthIssues")}
             value={report.summary.issueCount.toLocaleString()}
-            subValue={t("comboHealthActionable", { count: report.summary.actionableCount })}
+            subValue={t("comboHealthActionable", {
+              count: report.summary.suggestionCount ?? report.summary.actionableCount ?? 0,
+            })}
           />
           <MetricBlock
             icon="error"
diff --git a/src/app/api/providers/[id]/models/discovery/providerModelsConfig.ts b/src/app/api/providers/[id]/models/discovery/providerModelsConfig.ts
index 62988798d7..0939b59b55 100644
--- a/src/app/api/providers/[id]/models/discovery/providerModelsConfig.ts
+++ b/src/app/api/providers/[id]/models/discovery/providerModelsConfig.ts
@@ -87,6 +87,22 @@ export function parseAlibabaModelStudioModelsForConnection(
 export function parseQwenCloudTextModels(data: any): any[] {
   return parseCuratedDashscopeModels(data, QWEN_CLOUD_TEXT_MODELS, QWEN_CLOUD_TEXT_MODEL_IDS);
 }
+
+// Perplexity's /v1/models lists the Agent API catalog (vendor-prefixed ids like
+// "anthropic/claude-fable-5"), but chat requests always go to the classic
+// /chat/completions endpoint, which only accepts the Sonar family. Filter
+// discovery to Sonar-family ids so agent-style ids never surface as routable
+// chat models (#11060). Bounded pattern — no ReDoS-prone quantifiers.
+export function parsePerplexitySonarModels(data: any): any[] {
+  const models = Array.isArray(data?.data)
+    ? data.data
+    : Array.isArray(data?.models)
+      ? data.models
+      : [];
+  return models.filter(
+    (model: any) => typeof model?.id === "string" && /^sonar(-|$)/.test(model.id)
+  );
+}
 type ProviderModelsHeaderContext = {
   authType?: string;
   providerSpecificData?: unknown;
@@ -659,6 +675,17 @@ export const PROVIDER_MODELS_CONFIG: Record<string, ProviderModelsConfigEntry> =
     headers: { Accept: "application/json" },
     parseResponse: parseClinepassRecommendedModels,
   },
+  // Perplexity's /v1/models lists the Agent API catalog (vendor-prefixed agent
+  // ids), but chat only accepts the Sonar family on /chat/completions. Import
+  // must keep Sonar-family ids only (#11060).
+  perplexity: {
+    url: "https://api.perplexity.ai/v1/models",
+    method: "GET",
+    headers: { "Content-Type": "application/json" },
+    authHeader: "Authorization",
+    authPrefix: "Bearer ",
+    parseResponse: parsePerplexitySonarModels,
+  },
   cohere: {
     url: "https://api.cohere.com/v2/models",
     method: "GET",
diff --git a/src/app/api/providers/[id]/route.ts b/src/app/api/providers/[id]/route.ts
index 562dad0744..4f588571b8 100644
--- a/src/app/api/providers/[id]/route.ts
+++ b/src/app/api/providers/[id]/route.ts
@@ -121,7 +121,17 @@ export async function PUT(request: Request, { params }: { params: Promise<{ id:
     const { id } = await params;
     const validation = validateBody(updateProviderConnectionSchema, rawBody);
     if (isValidationFailure(validation)) {
-      return NextResponse.json({ error: validation.error }, { status: 400 });
+      // never drop an operator's intent silently. Surface the rejected
+      // keys (field paths and unrecognized-key names) alongside the existing
+      // error envelope so clients and the UI can tell exactly what was refused.
+      const rejected = [
+        ...validation.error.details.map((d) => d.field).filter(Boolean),
+        ...validation.error.details.flatMap((d) => d.keys ?? []),
+      ];
+      return NextResponse.json(
+        { error: { ...validation.error, rejected } },
+        { status: 400 }
+      );
     }
     const body = validation.data;
     const {
diff --git a/src/domain/configAudit.ts b/src/domain/configAudit.ts
index c2701b2a21..873d3b046a 100644
--- a/src/domain/configAudit.ts
+++ b/src/domain/configAudit.ts
@@ -13,6 +13,8 @@
  * - Optional human notes
  */
 
+import { getDbInstance } from "../lib/db/core";
+
 /** Types of configuration entities that can be audited */
 export type AuditTarget = "provider" | "combo" | "policy" | "connection" | "settings";
 
@@ -72,10 +74,8 @@ export interface ConfigSnapshot {
   data: Record<string, unknown>;
 }
 
-// ── In-memory store ──────────────────────────────────────────────────────────
-// In production, persist to SQLite alongside other domain state.
+// ── SQLite-backed store ───────────────────────────────────────────────────────
 
-let auditLog: ConfigAuditEntry[] = [];
 let idCounter = 0;
 
 function generateId(): string {
@@ -85,6 +85,40 @@ function generateId(): string {
   return `audit-${ts}-${seq}`;
 }
 
+function db() {
+  return getDbInstance();
+}
+
+interface ConfigAuditRow {
+  id: string;
+  timestamp: string;
+  action: string;
+  target: string;
+  target_id: string;
+  target_name: string;
+  before_json: string | null;
+  after_json: string | null;
+  diff_json: string;
+  source: string;
+  note: string | null;
+}
+
+function rowToEntry(row: ConfigAuditRow): ConfigAuditEntry {
+  return {
+    id: row.id,
+    timestamp: row.timestamp,
+    action: row.action as AuditAction,
+    target: row.target as AuditTarget,
+    targetId: row.target_id,
+    targetName: row.target_name,
+    before: row.before_json === null ? null : (JSON.parse(row.before_json) as Record<string, unknown> | null),
+    after: row.after_json === null ? null : (JSON.parse(row.after_json) as Record<string, unknown> | null),
+    source: row.source as AuditSource,
+    diff: JSON.parse(row.diff_json) as ConfigDiff,
+    note: row.note,
+  };
+}
+
 /**
  * Compute a structured diff between two configuration states.
  */
@@ -159,12 +193,24 @@ export function recordChange(
     note: note ?? null,
   };
 
-  auditLog.push(entry);
-
-  // Keep log bounded (max 1000 entries in memory)
-  if (auditLog.length > 1000) {
-    auditLog = auditLog.slice(-1000);
-  }
+  db().prepare(
+    `INSERT INTO config_audit_log
+       (id, timestamp, action, target, target_id, target_name, before_json, after_json, diff_json, source, note)
+     VALUES
+       (@id, @timestamp, @action, @target, @targetId, @targetName, @beforeJson, @afterJson, @diffJson, @source, @note)`
+  ).run({
+    id: entry.id,
+    timestamp: entry.timestamp,
+    action: entry.action,
+    target: entry.target,
+    targetId: entry.targetId,
+    targetName: entry.targetName,
+    beforeJson: before === null ? null : JSON.stringify(before),
+    afterJson: after === null ? null : JSON.stringify(after),
+    diffJson: JSON.stringify(entry.diff),
+    source: entry.source,
+    note: entry.note,
+  });
 
   return entry;
 }
@@ -181,42 +227,57 @@ export function getAuditLog(options?: {
   limit?: number;
   offset?: number;
 }): { entries: ConfigAuditEntry[]; total: number } {
-  let filtered = auditLog;
+  const where: string[] = [];
+  const params: Record<string, unknown> = {};
 
   if (options?.target) {
-    filtered = filtered.filter((e) => e.target === options.target);
+    where.push("target = @target");
+    params.target = options.target;
   }
   if (options?.targetId) {
-    filtered = filtered.filter((e) => e.targetId === options.targetId);
+    where.push("target_id = @targetId");
+    params.targetId = options.targetId;
   }
   if (options?.action) {
-    filtered = filtered.filter((e) => e.action === options.action);
+    where.push("action = @action");
+    params.action = options.action;
   }
   if (options?.source) {
-    filtered = filtered.filter((e) => e.source === options.source);
+    where.push("source = @source");
+    params.source = options.source;
   }
   if (options?.since) {
-    filtered = filtered.filter((e) => e.timestamp >= options.since!);
+    where.push("timestamp >= @since");
+    params.since = options.since;
   }
 
-  const total = filtered.length;
+  const whereSql = where.length > 0 ? `WHERE ${where.join(" AND ")}` : "";
 
-  // Sort newest first
-  filtered = [...filtered].sort((a, b) => b.timestamp.localeCompare(a.timestamp));
+  const totalRow = db()
+    .prepare(`SELECT COUNT(*) AS c FROM config_audit_log ${whereSql}`)
+    .get(params) as { c: number };
+  const total = totalRow.c;
 
-  // Paginate
   const offset = options?.offset ?? 0;
   const limit = options?.limit ?? 50;
-  filtered = filtered.slice(offset, offset + limit);
 
-  return { entries: filtered, total };
+  const rows = db()
+    .prepare(
+      `SELECT * FROM config_audit_log ${whereSql} ORDER BY datetime(timestamp) DESC, id DESC LIMIT @limit OFFSET @offset`
+    )
+    .all({ ...params, limit, offset }) as ConfigAuditRow[];
+
+  return { entries: rows.map(rowToEntry), total };
 }
 
 /**
  * Get a specific audit entry by ID.
  */
 export function getAuditEntry(id: string): ConfigAuditEntry | null {
-  return auditLog.find((e) => e.id === id) ?? null;
+  const row = db()
+    .prepare("SELECT * FROM config_audit_log WHERE id = @id")
+    .get({ id }) as ConfigAuditRow | undefined;
+  return row ? rowToEntry(row) : null;
 }
 
 /**
@@ -260,19 +321,23 @@ export function getAuditSummary(): {
   const byAction: Record<string, number> = {};
   const bySource: Record<string, number> = {};
 
-  for (const entry of auditLog) {
-    byTarget[entry.target] = (byTarget[entry.target] || 0) + 1;
-    byAction[entry.action] = (byAction[entry.action] || 0) + 1;
-    bySource[entry.source] = (bySource[entry.source] || 0) + 1;
+  const rows = db()
+    .prepare("SELECT * FROM config_audit_log ORDER BY datetime(timestamp) DESC, id DESC")
+    .all() as ConfigAuditRow[];
+
+  for (const row of rows) {
+    byTarget[row.target] = (byTarget[row.target] || 0) + 1;
+    byAction[row.action] = (byAction[row.action] || 0) + 1;
+    bySource[row.source] = (bySource[row.source] || 0) + 1;
   }
 
   return {
-    totalEntries: auditLog.length,
+    totalEntries: rows.length,
     byTarget,
     byAction,
     bySource,
-    oldestEntry: auditLog.length > 0 ? auditLog[0].timestamp : null,
-    newestEntry: auditLog.length > 0 ? auditLog[auditLog.length - 1].timestamp : null,
+    oldestEntry: rows.length > 0 ? rows[rows.length - 1].timestamp : null,
+    newestEntry: rows.length > 0 ? rows[0].timestamp : null,
   };
 }
 
@@ -280,6 +345,6 @@ export function getAuditSummary(): {
  * Reset the audit log. Useful for testing.
  */
 export function resetAuditLog(): void {
-  auditLog = [];
+  db().prepare("DELETE FROM config_audit_log").run();
   idCounter = 0;
 }
diff --git a/src/i18n/messages/en.json b/src/i18n/messages/en.json
index ec091011e9..ff778753c6 100644
--- a/src/i18n/messages/en.json
+++ b/src/i18n/messages/en.json
@@ -5692,8 +5692,8 @@
     "aggregatorsGateways": "Aggregators Gateways",
     "enterpriseCloud": "Enterprise & Cloud",
     "apiFormatLabel": "Api Format Label",
-    "apiKeyOptionalHint": "Api Key Optional Hint",
-    "apiKeyOptionalLabel": "Api Key Optional Label",
+    "apiKeyOptionalHint": "Leave empty if your local setup or provider does not require authentication.",
+    "apiKeyOptionalLabel": "API Key (optional)",
     "apiRegionChina": "Api Region China",
     "apiRegionHint": "Api Region Hint",
     "apiRegionInternational": "Api Region International",
diff --git a/src/i18n/messages/vi.json b/src/i18n/messages/vi.json
index ca70aafdef..e04b206661 100644
--- a/src/i18n/messages/vi.json
+++ b/src/i18n/messages/vi.json
@@ -6271,18 +6271,18 @@
     "webSessionGuideStep4": "Dán vào đây và kiểm tra kết nối. Nếu nó ngừng hoạt động, hãy đăng nhập lại và thay thế bằng một giá trị mới.",
     "harImportButtonLabel": "Nhập tệp .har",
     "harImportButtonBusy": "Đang nhập…",
-    "harImportButtonHint": "Xuất từ thẻ Network trong DevTools sau khi gửi ít nhất một tin nhắn trò chuyện.",
-    "harImportStatusValid": "Đã nhập — còn hiệu lực trong khoảng {minutes} phút.",
-    "harImportStatusExpiringSoon": "Đã nhập — chỉ còn hiệu lực trong khoảng {minutes} phút nữa.",
-    "harImportStatusExpired": "Đã nhập, nhưng token này đã hết hạn ({minutes} phút trước) — hãy xuất một tệp HAR mới.",
-    "harImportStatusUnknownExpiry": "Đã nhập. Không đọc được thời gian hết hạn.",
-    "harImportErrorNotJson": "Tệp đó không phải JSON hợp lệ — nó có thực sự là tệp xuất .har không?",
-    "harImportErrorNoEntries": "Tệp HAR này không có bản ghi mạng nào được lưu.",
-    "harImportErrorNoChathubUrl": "Không tìm thấy kết nối trò chuyện Copilot trong tệp HAR này. Hãy gửi ít nhất một tin nhắn trò chuyện trong m365.cloud.microsoft trước khi xuất.",
-    "harImportErrorUnparsableUrl": "Đã tìm thấy kết nối trò chuyện, nhưng không đọc được URL của nó.",
-    "harImportErrorMissingFields": "Đã tìm thấy kết nối trò chuyện, nhưng trong đó thiếu token.",
-    "harImportErrorReadFailed": "Không thể đọc tệp đó.",
-    "harImportErrorUnknown": "Không thể trích xuất thông tin xác thực từ tệp HAR đó.",
+    "harImportButtonHint": "Xuất từ thẻ Network của DevTools sau khi gửi ít nhất một tin nhắn chat.",
+    "harImportStatusValid": "Đã nhập — hợp lệ trong ~{minutes} phút.",
+    "harImportStatusExpiringSoon": "Đã nhập — chỉ còn hợp lệ ~{minutes} phút nữa.",
+    "harImportStatusExpired": "Đã nhập, nhưng token này đã hết hạn ({minutes} phút trước) — hãy xuất một HAR mới.",
+    "harImportStatusUnknownExpiry": "Đã nhập. Không đọc được thời hạn.",
+    "harImportErrorNotJson": "Tệp đó không phải JSON hợp lệ — có đúng là bản xuất .har không?",
+    "harImportErrorNoEntries": "HAR này không có mục network nào được ghi lại.",
+    "harImportErrorNoChathubUrl": "Không tìm thấy kết nối Copilot chat trong HAR này. Hãy gửi ít nhất một tin nhắn chat trong m365.cloud.microsoft trước khi xuất.",
+    "harImportErrorUnparsableUrl": "Tìm thấy kết nối chat, nhưng không đọc được URL của nó.",
+    "harImportErrorMissingFields": "Tìm thấy kết nối chat, nhưng token bị thiếu trong đó.",
+    "harImportErrorReadFailed": "Không đọc được tệp đó.",
+    "harImportErrorUnknown": "Không trích xuất được thông tin xác thực từ tệp HAR đó.",
     "webSessionSecurityHint": "Hãy coi đây như mật khẩu: nó có thể truy cập vào tài khoản web đã đăng nhập của bạn cho đến khi hết hạn hoặc bị thu hồi.",
     "webNoAuthGuideTitle": "Không yêu cầu thông tin xác thực",
     "webNoAuthGuideBody": "{provider} không cần khóa API hoặc cookie. Lưu kết nối để sử dụng endpoint web miễn phí của nó.",
@@ -12221,7 +12221,7 @@
       },
       "omni-webhooks": {
         "name": "Webhook",
-        "description": "Đăng ký, liệt kê, kiểm tra và xóa các endpoint webhook. Cấu hình đăng ký sự kiện (request.completed, request.failed, quota.exceeded, v.v.) và quản lý việc thử lại khi gửi."
+        "description": "Đăng ký, liệt kê, kiểm thử và xoá các endpoint webhook. Cấu hình đăng ký sự kiện (request.completed, request.failed, quota.exceeded, v.v.) và quản lý thử lại giao hàng."
       },
       "omni-mcp": {
         "name": "Máy chủ MCP",
diff --git a/src/lib/config/runtimeSettings.ts b/src/lib/config/runtimeSettings.ts
index dc5e9d720d..bf615749ba 100644
--- a/src/lib/config/runtimeSettings.ts
+++ b/src/lib/config/runtimeSettings.ts
@@ -1,5 +1,9 @@
 import { clearHealthCheckLogCache } from "@/lib/tokenHealthCheck";
 import { setCustomBannedSignals } from "@omniroute/open-sse/services/accountFallback.ts";
+import {
+  setOperatorProviderErrorRules,
+  type OperatorProviderErrorRule,
+} from "@omniroute/open-sse/config/providerErrorRules.ts";
 import { isAutomatedTestProcess } from "@/shared/utils/testProcess";
 
 type JsonRecord = Record<string, unknown>;
@@ -46,6 +50,7 @@ interface RuntimeSettingsSnapshot {
   systemTransforms: unknown;
   authzBypass: AuthzBypassSnapshot;
   customBannedSignals: string[];
+  providerErrorRules: Record<string, OperatorProviderErrorRule[]> | null;
 }
 
 // Default bypass policy: kill-switch on, `/api/mcp/` bypassable. Mirrors the
@@ -72,6 +77,7 @@ const DEFAULT_RUNTIME_SETTINGS_SNAPSHOT: RuntimeSettingsSnapshot = {
   systemTransforms: null,
   authzBypass: DEFAULT_AUTHZ_BYPASS_SNAPSHOT,
   customBannedSignals: [],
+  providerErrorRules: null,
 };
 
 let lastAppliedSnapshot: RuntimeSettingsSnapshot | null = null;
@@ -138,6 +144,34 @@ function normalizeStringArray(value: unknown): string[] {
   );
 }
 
+/**
+ * Defensive shape-check of operator-declared error rules pulled from settings.
+ * The settings schema already validates this on write; this guard prevents a
+ * malformed stored value (or an unexpected shape) from crashing the
+ * error-classification hot path. Returns null when the value is missing or not
+ * a record of non-empty rule arrays.
+ */
+function normalizeOperatorProviderErrorRules(
+  value: unknown
+): Record<string, OperatorProviderErrorRule[]> | null {
+  if (value === null || typeof value !== "object") return null;
+  const record = value as Record<string, unknown>;
+  const result: Record<string, OperatorProviderErrorRule[]> = {};
+  for (const [provider, list] of Object.entries(record)) {
+    if (!Array.isArray(list) || list.length === 0) continue;
+    const rules = list.filter(
+      (entry): entry is OperatorProviderErrorRule =>
+        !!entry &&
+        typeof entry === "object" &&
+        typeof (entry as OperatorProviderErrorRule).status === "number" &&
+        typeof (entry as OperatorProviderErrorRule).match === "string" &&
+        typeof (entry as OperatorProviderErrorRule).scope === "string"
+    );
+    if (rules.length > 0) result[provider.toLowerCase()] = rules;
+  }
+  return Object.keys(result).length > 0 ? result : null;
+}
+
 function normalizeStringRecord(value: unknown): Record<string, string> {
   const record = toRecord(parseStoredJson(value, "modelAliases"));
   const entries = Object.entries(record)
@@ -244,6 +278,7 @@ export function buildRuntimeSettingsSnapshot(
     systemTransforms: parseStoredJson(settings.systemTransforms, "systemTransforms"),
     authzBypass: normalizeAuthzBypass(settings),
     customBannedSignals: normalizeStringArray(settings.customBannedSignals),
+    providerErrorRules: normalizeOperatorProviderErrorRules(settings.providerErrorRules),
   };
 }
 
@@ -540,6 +575,13 @@ export async function applyRuntimeSettings(
     markChanged("bannedSignals");
   }
 
+  if (
+    force ||
+    hasChanged(currentSnapshot.providerErrorRules, previousSnapshot.providerErrorRules)
+  ) {
+    setOperatorProviderErrorRules(currentSnapshot.providerErrorRules ?? undefined);
+  }
+
   lastAppliedSnapshot = currentSnapshot;
   return changes;
 }
diff --git a/src/lib/db/cleanup.ts b/src/lib/db/cleanup.ts
index 617bf4c226..60e288bf4b 100644
--- a/src/lib/db/cleanup.ts
+++ b/src/lib/db/cleanup.ts
@@ -193,6 +193,31 @@ export async function cleanupMcpAudit(): Promise<CleanupResult> {
   return result;
 }
 
+/**
+ * Clean up old config_audit_log based on retention settings.
+ */
+export async function cleanupConfigAudit(retentionDays = getRetentionSettings().configAudit): Promise<CleanupResult> {
+  const db = getDbInstance();
+  const result: CleanupResult = { deleted: 0, errors: 0 };
+
+  try {
+    const stmt = db.prepare(
+      "DELETE FROM config_audit_log WHERE datetime(timestamp) < datetime('now', '-' || ? || ' days')"
+    );
+    const runResult = stmt.run(String(retentionDays));
+    result.deleted = runResult.changes;
+
+    console.log(
+      `[Cleanup] Deleted ${result.deleted} config_audit_log older than ${retentionDays} days`
+    );
+  } catch (err: unknown) {
+    console.error("[Cleanup] Error cleaning config_audit_log:", err);
+    result.errors++;
+  }
+
+  return result;
+}
+
 /**
  * Clean up old a2a_task_events based on retention settings.
  */
@@ -420,6 +445,7 @@ export async function runAutoCleanup(): Promise<{
     usageHistory: await cleanupUsageHistory(),
     compressionAnalytics: await cleanupCompressionAnalytics(),
     mcpAudit: await cleanupMcpAudit(),
+    configAudit: await cleanupConfigAudit(),
     a2aEvents: await cleanupA2aEvents(),
     memoryEntries: await cleanupMemoryEntries(),
     domainCostHistory: await cleanupDomainCostHistory(),
diff --git a/src/lib/db/databaseSettings.ts b/src/lib/db/databaseSettings.ts
index e18f2a66bd..0a12729242 100644
--- a/src/lib/db/databaseSettings.ts
+++ b/src/lib/db/databaseSettings.ts
@@ -46,6 +46,7 @@ const LEGACY_FLAT_KEYS: {
     quotaSnapshots: ["quotaSnapshots"],
     compressionAnalytics: ["compressionAnalytics"],
     mcpAudit: ["mcpAudit"],
+    configAudit: ["configAudit"],
     a2aEvents: ["a2aEvents"],
     callLogs: ["callLogs"],
     usageHistory: ["usageHistory"],
diff --git a/src/lib/db/migrations/046_database_settings.sql b/src/lib/db/migrations/046_database_settings.sql
index 57fd15c903..6fb9864390 100644
--- a/src/lib/db/migrations/046_database_settings.sql
+++ b/src/lib/db/migrations/046_database_settings.sql
@@ -25,6 +25,7 @@ INSERT OR IGNORE INTO key_value (namespace, key, value) VALUES ('databaseSetting
 INSERT OR IGNORE INTO key_value (namespace, key, value) VALUES ('databaseSettings', 'quotaSnapshots', '90');
 INSERT OR IGNORE INTO key_value (namespace, key, value) VALUES ('databaseSettings', 'compressionAnalytics', '30');
 INSERT OR IGNORE INTO key_value (namespace, key, value) VALUES ('databaseSettings', 'mcpAudit', '30');
+INSERT OR IGNORE INTO key_value (namespace, key, value) VALUES ('databaseSettings', 'configAudit', '30');
 INSERT OR IGNORE INTO key_value (namespace, key, value) VALUES ('databaseSettings', 'a2aEvents', '30');
 INSERT OR IGNORE INTO key_value (namespace, key, value) VALUES ('databaseSettings', 'callLogs', '90');
 INSERT OR IGNORE INTO key_value (namespace, key, value) VALUES ('databaseSettings', 'usageHistory', '365');
diff --git a/src/lib/db/migrations/161_config_audit_log.sql b/src/lib/db/migrations/161_config_audit_log.sql
new file mode 100644
index 0000000000..0aad91ace2
--- /dev/null
+++ b/src/lib/db/migrations/161_config_audit_log.sql
@@ -0,0 +1,15 @@
+CREATE TABLE IF NOT EXISTS config_audit_log (
+  id TEXT PRIMARY KEY,
+  timestamp TEXT NOT NULL,
+  action TEXT NOT NULL,
+  target TEXT NOT NULL,
+  target_id TEXT NOT NULL,
+  target_name TEXT NOT NULL,
+  before_json TEXT,
+  after_json TEXT,
+  diff_json TEXT NOT NULL,
+  source TEXT NOT NULL,
+  note TEXT
+);
+CREATE INDEX IF NOT EXISTS idx_config_audit_log_target_created ON config_audit_log(target, timestamp);
+CREATE INDEX IF NOT EXISTS idx_config_audit_log_created ON config_audit_log(timestamp);
diff --git a/src/lib/db/providers.ts b/src/lib/db/providers.ts
index 3b02933e11..fbab6817fd 100644
--- a/src/lib/db/providers.ts
+++ b/src/lib/db/providers.ts
@@ -627,15 +627,26 @@ export async function createProviderConnection(data: JsonRecord) {
   // to no-overrides) keeps the field present on the returned object so the
   // UI can tell "field was read, no overrides" apart from "field absent."
   if ("quotaWindowThresholds" in connection) {
-    connection.quotaWindowThresholds = sanitizeQuotaWindowThresholds(
-      connection.quotaWindowThresholds
-    );
+    const result = sanitizeQuotaWindowThresholds(connection.quotaWindowThresholds);
+    if (result.rejected.length > 0) {
+      throw new Error(
+        `Refusing to persist quotaWindowThresholds with rejected keys: ${result.rejected.join(", ")}`
+      );
+    }
+    connection.quotaWindowThresholds = result.sanitized;
   }
 
   // Same sanitization for rateLimitOverrides — keep in-memory representation
-  // in sync with what gets persisted.
+  // in sync with what gets persisted. Reject (don't silently drop) invalid
+  // keys/values so a direct DB writer can't lose operator intent.
   if ("rateLimitOverrides" in connection) {
-    connection.rateLimitOverrides = sanitizeRateLimitOverrides(connection.rateLimitOverrides);
+    const result = sanitizeRateLimitOverrides(connection.rateLimitOverrides);
+    if (result.rejected.length > 0) {
+      throw new Error(
+        `Refusing to persist rateLimitOverrides with rejected keys: ${result.rejected.join(", ")}`
+      );
+    }
+    connection.rateLimitOverrides = result.sanitized;
   }
 
   _insertConnectionRow(db, encryptConnectionFields({ ...connection }));
@@ -849,13 +860,24 @@ export async function updateProviderConnection(id: string, data: JsonRecord) {
   // Mirror the sanitization the create path applies — keep the returned
   // object in lockstep with what we persist.
   if ("quotaWindowThresholds" in merged) {
-    const sanitized = sanitizeQuotaWindowThresholds(merged.quotaWindowThresholds);
+    const result = sanitizeQuotaWindowThresholds(merged.quotaWindowThresholds);
+    if (result.rejected.length > 0) {
+      throw new Error(
+        `Refusing to persist quotaWindowThresholds with rejected keys: ${result.rejected.join(", ")}`
+      );
+    }
     // For updates we always carry the key forward (even as null) so the read
-    // path surfaces the cleared state to callers that just patched it.
-    merged.quotaWindowThresholds = sanitized;
+    // path surfaces the cleared state to callers that merged it.
+    merged.quotaWindowThresholds = result.sanitized;
   }
   if ("rateLimitOverrides" in merged) {
-    merged.rateLimitOverrides = sanitizeRateLimitOverrides(merged.rateLimitOverrides);
+    const result = sanitizeRateLimitOverrides(merged.rateLimitOverrides);
+    if (result.rejected.length > 0) {
+      throw new Error(
+        `Refusing to persist rateLimitOverrides with rejected keys: ${result.rejected.join(", ")}`
+      );
+    }
+    merged.rateLimitOverrides = result.sanitized;
   }
   const existingRecord = toRecord(existing);
 
diff --git a/src/lib/db/providers/columns.ts b/src/lib/db/providers/columns.ts
index b32f653f7c..f06f948cd2 100644
--- a/src/lib/db/providers/columns.ts
+++ b/src/lib/db/providers/columns.ts
@@ -64,20 +64,37 @@ export function normalizeBooleanColumn(value: unknown, fallback: boolean): boole
   return fallback;
 }
 
+// Result of sanitizing a per-connection overrides/threshold map. `sanitized`
+// is the cleaned value (or null when it collapses to nothing); `rejected`
+// lists every key that was refused so callers can fail loudly
+// instead of silently dropping the operator's input.
+export type SanitizeResult = {
+  sanitized: Record<string, number> | null;
+  rejected: string[];
+};
+
 // Sanitize the per-connection rate limit overrides map: keep only known
-// fields with valid numeric values. Called once at each write-path boundary.
-export function sanitizeRateLimitOverrides(value: unknown): Record<string, number> | null {
-  if (value === null || value === undefined) return null;
-  if (typeof value !== "object" || Array.isArray(value)) return null;
+// fields with valid non-negative integer values. Called once at each
+// write-path boundary. Unknown keys and invalid values go into `rejected`
+// rather than being dropped in silence.
+export function sanitizeRateLimitOverrides(value: unknown): SanitizeResult {
+  if (value === null || value === undefined) return { sanitized: null, rejected: [] };
+  if (typeof value !== "object" || Array.isArray(value)) return { sanitized: null, rejected: [] };
   const allowedKeys = new Set(["rpm", "tpm", "tpd", "minTime", "maxConcurrent"]);
+  const rejected: string[] = [];
   const map: Record<string, number> = {};
   for (const [key, v] of Object.entries(value as Record<string, unknown>)) {
-    if (!allowedKeys.has(key)) continue;
+    if (!allowedKeys.has(key)) {
+      rejected.push(key);
+      continue;
+    }
     if (typeof v === "number" && Number.isInteger(v) && v >= 0) {
       map[key] = v;
+    } else {
+      rejected.push(key);
     }
   }
-  return Object.keys(map).length === 0 ? null : map;
+  return { sanitized: Object.keys(map).length === 0 ? null : map, rejected };
 }
 
 // Serialize an already-sanitized map for SQLite TEXT storage.
@@ -91,20 +108,29 @@ export function toRecord(value: unknown): JsonRecord {
   return value && typeof value === "object" ? (value as JsonRecord) : {};
 }
 
-// Sanitize the per-window threshold map: keep only 0-100 integer values.
-// Called once at each write-path boundary (createProviderConnection +
-// updateProviderConnection) so both the in-memory return and the persisted
-// row share the same shape. Serialization below trusts this output.
-export function sanitizeQuotaWindowThresholds(value: unknown): Record<string, number> | null {
-  if (value === null || value === undefined) return null;
-  if (typeof value !== "object" || Array.isArray(value)) return null;
+// Sanitize the per-window threshold map: keep only 0-100 integer values with
+// keys no longer than 64 chars. Called once at each write-path boundary
+// (createProviderConnection + updateProviderConnection) so both the in-memory
+// return and the persisted row share the same shape. Serialization below
+// trusts this output. Invalid keys/values go into `rejected` rather than being
+// dropped in silence.
+export function sanitizeQuotaWindowThresholds(value: unknown): SanitizeResult {
+  if (value === null || value === undefined) return { sanitized: null, rejected: [] };
+  if (typeof value !== "object" || Array.isArray(value)) return { sanitized: null, rejected: [] };
+  const rejected: string[] = [];
   const map: Record<string, number> = {};
   for (const [key, v] of Object.entries(value as Record<string, unknown>)) {
+    if (key.length > 64) {
+      rejected.push(key);
+      continue;
+    }
     if (typeof v === "number" && Number.isInteger(v) && v >= 0 && v <= 100) {
       map[key] = v;
+    } else {
+      rejected.push(key);
     }
   }
-  return Object.keys(map).length === 0 ? null : map;
+  return { sanitized: Object.keys(map).length === 0 ? null : map, rejected };
 }
 
 export function toStringOrNull(value: unknown): string | null {
diff --git a/src/lib/monitoring/comboHealthAutopilot.ts b/src/lib/monitoring/comboHealthAutopilot.ts
index 71a1049483..24eba32a63 100644
--- a/src/lib/monitoring/comboHealthAutopilot.ts
+++ b/src/lib/monitoring/comboHealthAutopilot.ts
@@ -16,6 +16,7 @@ import type {
   ComboForecastMetrics,
   ComboForecastResponse,
   ComboForecastRiskLevel,
+  ProviderAutopilotReport,
   ComboHealthMetrics,
   ComboHealthResponse,
   ComboRecord,
@@ -34,6 +35,7 @@ export interface ComboHealthAutopilotOptions {
   combos?: ComboRecord[];
   healthResponse?: ComboHealthResponse;
   forecastResponse?: ComboForecastResponse;
+  providerHealthResponse?: ProviderAutopilotReport;
 }
 
 type ProviderIssueView = {
@@ -103,7 +105,12 @@ function actionSet(
       case "open_combo_editor":
         return action(type, "Open combo editor", target, "/dashboard/combos");
       case "run_combo_test":
-        return action(type, "Run combo test", target, "/dashboard/combos");
+        return action(
+          type,
+          "Run combo test",
+          target,
+          `/dashboard/combos?test=${encodeURIComponent(target.comboId)}`
+        );
       case "open_provider_health_autopilot":
         return action(type, "Open provider autopilot", target, "/dashboard/health");
       case "review_quota_limits":
@@ -447,7 +454,8 @@ export async function buildComboHealthAutopilotReport(
         now: options.now,
         combos: combosSnapshot,
       }),
-    buildProviderHealthAutopilotReport({ includeHealthy: false, includeActions: false }),
+    options.providerHealthResponse ??
+      buildProviderHealthAutopilotReport({ includeHealthy: false, includeActions: false }),
   ]);
 
   const forecastsByComboId = new Map(forecast.combos.map((entry) => [entry.comboId, entry]));
@@ -470,7 +478,7 @@ export async function buildComboHealthAutopilotReport(
   const degradedCount = allCombos.filter((combo) => combo.state === "degraded").length;
   const healthyCount = allCombos.filter((combo) => combo.state === "healthy").length;
   const issueCount = allCombos.reduce((sum, combo) => sum + combo.issues.length, 0);
-  const actionableCount = allCombos.reduce(
+  const suggestionCount = allCombos.reduce(
     (sum, combo) =>
       sum + combo.issues.reduce((issueSum, issue) => issueSum + issue.actions.length, 0),
     0
@@ -487,7 +495,8 @@ export async function buildComboHealthAutopilotReport(
       degradedCount,
       downCount,
       issueCount,
-      actionableCount,
+      suggestionCount,
+      actionableCount: suggestionCount,
     },
     combos,
   };
diff --git a/src/shared/constants/providers.ts b/src/shared/constants/providers.ts
index 23f9c53b39..6b4fd86a50 100644
--- a/src/shared/constants/providers.ts
+++ b/src/shared/constants/providers.ts
@@ -237,7 +237,6 @@ export function isSelfHostedChatProvider(providerId: unknown): boolean {
 const EXPLICIT_OPTIONAL_APIKEY_PROVIDER_IDS = new Set([
   "searxng-search",
   "firecrawl",
-  "pollinations",
   "copilot-web",
   "hackclub",
   "g4f-groq",
diff --git a/src/shared/services/opencodeConfig.ts b/src/shared/services/opencodeConfig.ts
index a4bbd65591..a91ca5069c 100644
--- a/src/shared/services/opencodeConfig.ts
+++ b/src/shared/services/opencodeConfig.ts
@@ -84,11 +84,29 @@ export const buildOpenCodeProviderConfig = ({
   };
 };
 
+export const buildOpenCodeV2ProviderConfig = (
+  input: OpenCodeConfigInput
+): Record<string, any> => {
+  const v1Config = buildOpenCodeProviderConfig(input);
+  return {
+    name: v1Config.name,
+    package: "@opencode-ai/ai/providers/openai-compatible",
+    settings: {
+      baseURL: v1Config.options.baseURL,
+      apiKey: v1Config.options.apiKey,
+    },
+    models: v1Config.models,
+  };
+};
+
 export const buildOpenCodeConfigDocument = (input: OpenCodeConfigInput) => ({
   $schema: "https://opencode.ai/config.json",
   provider: {
     omniroute: buildOpenCodeProviderConfig(input),
   },
+  providers: {
+    omniroute: buildOpenCodeV2ProviderConfig(input),
+  },
 });
 
 export const mergeOpenCodeConfig = (
@@ -100,18 +118,18 @@ export const mergeOpenCodeConfig = (
       ? existingConfig
       : {};
 
-  // Same guard as the root above, one level down. Spreading a non-object here
-  // does not throw, it splays the value into index keys: an existing
-  // `"provider": ["a", "b"]` merged to `{"0": "a", "1": "b", omniroute: ... }`
-  // and a string was exploded one character per key. mergeOpenCodeConfigText
-  // refuses the same input outright, so the two disagreed on what to do with a
-  // malformed config.
   const existingProvider = (safeConfig as Record<string, unknown>).provider;
   const safeProvider =
     existingProvider && typeof existingProvider === "object" && !Array.isArray(existingProvider)
       ? (existingProvider as Record<string, unknown>)
       : {};
 
+  const existingProviders = (safeConfig as Record<string, unknown>).providers;
+  const safeProviders =
+    existingProviders && typeof existingProviders === "object" && !Array.isArray(existingProviders)
+      ? (existingProviders as Record<string, unknown>)
+      : {};
+
   return {
     ...safeConfig,
     $schema: safeConfig.$schema || "https://opencode.ai/config.json",
@@ -119,6 +137,10 @@ export const mergeOpenCodeConfig = (
       ...safeProvider,
       omniroute: buildOpenCodeProviderConfig(input),
     },
+    providers: {
+      ...safeProviders,
+      omniroute: buildOpenCodeV2ProviderConfig(input),
+    },
   };
 };
 
@@ -127,6 +149,7 @@ export const mergeOpenCodeConfigText = (
   input: OpenCodeConfigInput
 ) => {
   const providerConfig = buildOpenCodeProviderConfig(input);
+  const v2ProviderConfig = buildOpenCodeV2ProviderConfig(input);
   const content = typeof existingText === "string" ? existingText : "";
   const trimmedContent = content.trim();
 
@@ -161,6 +184,11 @@ export const mergeOpenCodeConfigText = (
   const providerEdits = modify(nextText, ["provider", "omniroute"], providerConfig, {
     formattingOptions: { insertSpaces: true, tabSize: 2 },
   });
+  nextText = applyEdits(nextText, providerEdits);
 
-  return applyEdits(nextText, providerEdits);
+  const v2ProviderEdits = modify(nextText, ["providers", "omniroute"], v2ProviderConfig, {
+    formattingOptions: { insertSpaces: true, tabSize: 2 },
+  });
+
+  return applyEdits(nextText, v2ProviderEdits);
 };
diff --git a/src/shared/types/utilization.ts b/src/shared/types/utilization.ts
index 36f0d7dabc..0a1e9dc900 100644
--- a/src/shared/types/utilization.ts
+++ b/src/shared/types/utilization.ts
@@ -260,7 +260,9 @@ export interface ComboAutopilotReport {
     degradedCount: number;
     downCount: number;
     issueCount: number;
-    actionableCount: number;
+    suggestionCount: number;
+    /** @deprecated Use suggestionCount instead. Kept as an alias for backward compatibility; remove after 2 releases. */
+    actionableCount?: number;
   };
   combos: ComboAutopilotCombo[];
 }
diff --git a/src/shared/validation/helpers.ts b/src/shared/validation/helpers.ts
index 4e486c7287..c811d6d4bf 100644
--- a/src/shared/validation/helpers.ts
+++ b/src/shared/validation/helpers.ts
@@ -4,6 +4,10 @@ import { z } from "zod";
 type ValidationErrorDetail = {
   field: string;
   message: string;
+  // Present for `unrecognized_keys` issues: the unknown key names that were
+  // refused (e.g. a typo'd override key). Surfaced by callers so clients can
+  // tell exactly which keys were rejected.
+  keys?: string[];
 };
 
 type ValidationErrorPayload = {
@@ -45,6 +49,9 @@ export function validateBody<TSchema extends z.ZodTypeAny>(
       details: issues.map((e) => ({
         field: e.path.join("."),
         message: e.message,
+        ...(("keys" in e && (e as { keys?: string[] }).keys)
+          ? { keys: (e as { keys: string[] }).keys }
+          : {}),
       })),
     },
   };
diff --git a/src/shared/validation/schemas/provider.ts b/src/shared/validation/schemas/provider.ts
index b3272f70fb..4cfe7afff4 100644
--- a/src/shared/validation/schemas/provider.ts
+++ b/src/shared/validation/schemas/provider.ts
@@ -420,6 +420,25 @@ export const providerNodeValidateSchema = z.object({
   modelId: z.string().trim().max(200).optional().or(z.literal("")),
 });
 
+// rate-limit override numeric fields must reject operator intent loss.
+// `z.coerce.number()` silently turns "" into 0 and "60abc" into NaN, which
+// would drop or distort the value instead of rejecting it. Preprocess first so
+// an empty/non-numeric string fails validation (surfaced as a 400), while still
+// coercing legit numeric strings like "60".
+function rateLimitOverrideNumber(max: number) {
+  return z.preprocess(
+    (raw) => {
+      if (typeof raw === "string") {
+        if (raw.trim() === "") return NaN;
+        const parsed = Number(raw);
+        return Number.isNaN(parsed) ? raw : parsed;
+      }
+      return raw;
+    },
+    z.coerce.number().int().min(0).max(max)
+  );
+}
+
 export const updateProviderConnectionSchema = z
   .object({
     name: z.string().max(200).optional(),
@@ -468,17 +487,24 @@ export const updateProviderConnectionSchema = z
     projectId: z.union([z.string(), z.null()]).optional(),
     // Per-connection rate limit overrides — overrides the global RequestQueueSettings
     // for this connection. Set to null to clear all overrides.
+    // Per-connection rate limit overrides — overrides the global
+    // RequestQueueSettings for this connection. Set to null to clear all
+    // overrides. `.strict()` rejects unknown keys (e.g. a typo'd `tmp`) with a
+    // 400 instead of silently stripping them: the operator's intent is
+    // never dropped without an error. `.nullable()` (rather than a
+    // `z.union([z.null(), …])`) keeps the `unrecognized_keys` issue at the top
+    // level so the rejected key name survives into the 400 response.
     rateLimitOverrides: z
-      .union([
-        z.null(),
-        z.object({
-          rpm: z.coerce.number().int().min(0).max(1_000_000).optional(),
-          tpm: z.coerce.number().int().min(0).max(100_000_000).optional(),
-          tpd: z.coerce.number().int().min(0).max(10_000_000_000).optional(),
-          minTime: z.coerce.number().int().min(0).max(60_000).optional(),
-          maxConcurrent: z.coerce.number().int().min(0).max(10_000).optional(),
-        }),
-      ])
+      .object({
+        rpm: rateLimitOverrideNumber(1_000_000).optional(),
+        tpm: rateLimitOverrideNumber(100_000_000).optional(),
+        tpd: rateLimitOverrideNumber(10_000_000_000).optional(),
+        minTime: rateLimitOverrideNumber(60_000).optional(),
+        maxConcurrent: rateLimitOverrideNumber(10_000).optional(),
+      })
+      .partial()
+      .strict()
+      .nullable()
       .optional(),
     proxyEnabled: z.boolean().optional(),
     perKeyProxyEnabled: z.boolean().optional(),
diff --git a/src/shared/validation/settingsSchemas.ts b/src/shared/validation/settingsSchemas.ts
index 388e3f73c7..beefb5fc8f 100644
--- a/src/shared/validation/settingsSchemas.ts
+++ b/src/shared/validation/settingsSchemas.ts
@@ -259,6 +259,48 @@ export const updateSettingsSchema = z.object({
       })
     )
     .optional(),
+  /**
+   * Operator-declared per-provider error rules. Consulted BEFORE the built-in
+   * `providerRuleRegistry` in open-sse/config/providerErrorRules.ts so an
+   * operator can add a scope/cooldown/reason override for a provider without
+   * editing the catalog. Matches are plain case-insensitive SUBSTRINGS of the
+   * error body (never RegExp) to keep the classification hot path ReDoS-safe.
+   * Bounded to 50 rules total so a misconfigured setting cannot blow up the
+   * matcher.
+   */
+  providerErrorRules: z
+    .record(
+      z.string().trim().min(1).max(100),
+      z.array(
+        z.object({
+          status: z.number().int().min(100).max(599),
+          match: z.string().min(1).max(200),
+          scope: z.enum(["model", "provider", "connection"]),
+          reason: z
+            .enum([
+              "auth_error",
+              "quota_exhausted",
+              "rate_limit_exceeded",
+              "model_capacity",
+              "server_error",
+              "unknown",
+            ])
+            .optional(),
+          cooldownMs: z.number().int().min(0).max(86_400_000).optional(),
+        })
+      )
+    )
+    .optional()
+    .superRefine((value, ctx) => {
+      if (!value) return;
+      const total = Object.values(value).reduce((n, rules) => n + rules.length, 0);
+      if (total > 50) {
+        ctx.addIssue({
+          code: z.ZodIssueCode.custom,
+          message: `providerErrorRules: at most 50 rules total, got ${total}`,
+        });
+      }
+    }),
   // #6168: global session-stickiness opt-out (per-combo config overrides this).
   disableSessionStickiness: z.boolean().optional(),
   /** Keep eligible combo targets close to the provider-side prompt cache. */
diff --git a/src/types/databaseSettings.ts b/src/types/databaseSettings.ts
index 6bcc210e5f..2b4820e198 100644
--- a/src/types/databaseSettings.ts
+++ b/src/types/databaseSettings.ts
@@ -44,6 +44,7 @@ export interface DatabaseSettings {
     quotaSnapshots: number;
     compressionAnalytics: number;
     mcpAudit: number;
+    configAudit: number;
     a2aEvents: number;
     callLogs: number;
     usageHistory: number;
@@ -114,6 +115,7 @@ export const DEFAULT_DATABASE_SETTINGS: Omit<DatabaseSettings, "location" | "sta
     quotaSnapshots: 7,
     compressionAnalytics: 30,
     mcpAudit: 30,
+    configAudit: 30,
     a2aEvents: 30,
     callLogs: 30,
     usageHistory: 30,
diff --git a/stryker.conf.json b/stryker.conf.json
index 240e8419e5..9658a82d51 100644
--- a/stryker.conf.json
+++ b/stryker.conf.json
@@ -272,6 +272,7 @@
       "tests/unit/model-lockout-max-cooldown.test.ts",
       "tests/unit/no-memory-header.test.ts",
       "tests/unit/noauth-autocombo-lockout-7623.test.ts",
+      "tests/unit/ollama-404-model-lockout-11071.test.ts",
       "tests/unit/non-streaming-sse-terminal-typescan-4459.test.ts",
       "tests/unit/nvidia-410-model-scope.test.ts",
       "tests/unit/nvidia-passthrough-models-6773.test.ts",
@@ -283,6 +284,7 @@
       "tests/unit/ollama-cloud-weekly-quota-cooldown-3709.test.ts",
       "tests/unit/openapi-security-tiers.test.ts",
       "tests/unit/opencode-autocombo-search-pair.test.ts",
+      "tests/unit/opencode-v2-config-11070.test.ts",
       "tests/unit/openrouter-free-model-credits-exhausted.test.ts",
       "tests/unit/openrouter-passthrough-models.test.ts",
       "tests/unit/openrouter-quota-6842.test.ts",
diff --git a/tests/unit/build/optional-transformers-dependency.test.ts b/tests/unit/build/optional-transformers-dependency.test.ts
index 7ad60fb09c..cf5997a5f5 100644
--- a/tests/unit/build/optional-transformers-dependency.test.ts
+++ b/tests/unit/build/optional-transformers-dependency.test.ts
@@ -9,40 +9,52 @@ function readJson<T = Record<string, unknown>>(relPath: string): T {
   return JSON.parse(readFileSync(join(repoRoot, relPath), "utf8")) as T;
 }
 
-test("@huggingface/transformers is a regular dependency so npm ci never skips it", () => {
-  // #9962 deliberately moved @huggingface/transformers out of optionalDependencies:
-  // as an optional dep, npm silently skipped the whole subtree on Node 24/26 (old
-  // pin dragged onnxruntime-node@1.21.0 whose NAN build no longer compiles), which
-  // broke `npm ci`/`next build` with "Can't resolve @huggingface/transformers"
-  // (lazy import in src/lib/memory/embedding/transformersLocal.ts). As a regular
-  // dep with onnxruntime-node@~1.24.3 (napi prebuilds, no node-gyp) it stays
-  // installable and the memory embedding path requires() cleanly.
+test("ONNX chain (@huggingface/transformers + onnxruntime-node) stays optional so Termux/Android installs succeed", () => {
+  // #11095: onnxruntime-node declares os ["win32","darwin","linux"], so while
+  // these lived in `dependencies` every npm install on Android/Termux aborted
+  // with a fatal EBADPLATFORM. As optionalDependencies npm skips only the
+  // unsupported-platform subtree (with a warning) and installs normally
+  // everywhere else. This deliberately reverses the MECHANISM of #9962 while
+  // keeping its goal: #9962's skip happened because the old onnxruntime-node@
+  // 1.21.0 pin built from source (NAN) and failed to compile on Node 24/26;
+  // the current 1.24.3 pin ships napi prebuilds, so on supported platforms the
+  // chain always installs and `npm ci`/`next build` keep resolving it. On
+  // platforms where it IS skipped, both consumers degrade gracefully via lazy/
+  // dynamic imports (asserted below).
   const pkg = readJson<{
     dependencies?: Record<string, string>;
     optionalDependencies?: Record<string, string>;
+    overrides?: Record<string, string>;
   }>("package.json");
 
   assert.equal(
     pkg.dependencies?.["@huggingface/transformers"],
+    undefined,
+    "transformers must NOT be a hard dependency (fatal EBADPLATFORM on Android)"
+  );
+  assert.equal(
+    pkg.optionalDependencies?.["@huggingface/transformers"],
     "^4.2.0",
-    "transformers must be a regular dependency (never optional) so npm ci cannot skip it"
+    "transformers must be an optionalDependency"
   );
-  assert.equal(pkg.optionalDependencies?.["@huggingface/transformers"], undefined);
-});
-
-test("transformers + onnxruntime-node are regular dependencies (not optional)", () => {
-  const pkg = readJson<{
-    dependencies?: Record<string, string>;
-    optionalDependencies?: Record<string, string>;
-  }>("package.json");
-
   assert.equal(
     pkg.dependencies?.["onnxruntime-node"],
-    "1.24.3",
-    "onnxruntime-node is a regular dep (napi prebuilds, installable on Node 24/26)"
+    undefined,
+    "onnxruntime-node must NOT be a hard dependency (fatal EBADPLATFORM on Android)"
   );
-  assert.equal(pkg.optionalDependencies?.["onnxruntime-node"], undefined);
+  assert.equal(
+    pkg.optionalDependencies?.["onnxruntime-node"],
+    "1.24.3",
+    "onnxruntime-node must be an optionalDependency pinned in lockstep with the overrides pin"
+  );
+  assert.equal(
+    pkg.overrides?.["onnxruntime-node"],
+    "1.24.3",
+    "the overrides pin must stay aligned with @huggingface/transformers' own pin (single-copy invariant)"
+  );
+});
 
+test("lockfile marks the whole ONNX chain optional", () => {
   const lock = readJson<{
     packages: Record<
       string,
@@ -51,26 +63,61 @@ test("transformers + onnxruntime-node are regular dependencies (not optional)",
         dependencies?: Record<string, string>;
         optionalDependencies?: Record<string, string>;
       }
-    >;
+      >;
   }>("package-lock.json");
 
   assert.equal(
-    lock.packages[""]?.dependencies?.["@huggingface/transformers"],
+    lock.packages[""]?.optionalDependencies?.["@huggingface/transformers"],
     "^4.2.0",
-    "root lock dependencies must hold transformers as a regular (non-optional) dep"
+    "root lock optionalDependencies must hold transformers"
   );
-  // Optional flag is only written `true` for genuinely optional packages;
-  // regular deps leave it absent/null. Assert each is NOT optional.
-  assert.ok(
-    !lock.packages["node_modules/@huggingface/transformers"]?.optional,
-    "transformers must not be marked optional in the lockfile"
+  assert.equal(
+    lock.packages[""]?.optionalDependencies?.["onnxruntime-node"],
+    "1.24.3",
+    "root lock optionalDependencies must hold onnxruntime-node"
   );
   assert.ok(
-    !lock.packages["node_modules/onnxruntime-node"]?.optional,
-    "onnxruntime-node must not be marked optional in the lockfile"
+    lock.packages["node_modules/@huggingface/transformers"]?.optional,
+    "transformers must be marked optional in the lockfile"
   );
   assert.ok(
-    !lock.packages["node_modules/onnxruntime-common"]?.optional,
-    "onnxruntime-common must not be marked optional in the lockfile"
+    lock.packages["node_modules/onnxruntime-node"]?.optional,
+    "onnxruntime-node must be marked optional in the lockfile"
+  );
+  assert.ok(
+    lock.packages["node_modules/onnxruntime-common"]?.optional,
+    "onnxruntime-common must be marked optional in the lockfile"
+  );
+});
+
+test("every @huggingface/transformers consumer loads it lazily so absent installs degrade gracefully", () => {
+  // If any module ever switches to a STATIC import of the optional chain,
+  // startup crashes on platforms where npm skipped it (Android/Termux).
+  // transformersLocal.ts must keep its lazy await import() (D8/D25);
+  // onnxWorker.ts must keep its runtime-variable dynamicImport indirection.
+
+  const embeddingSrc = readFileSync(
+    join(repoRoot, "src/lib/memory/embedding/transformersLocal.ts"),
+    "utf8"
+  );
+  assert.doesNotMatch(
+    embeddingSrc,
+    /^\s*import\s+(?:[^'"]*?\s+from\s+)?["']@huggingface\/transformers["']/m,
+    "transformersLocal.ts must not statically import @huggingface/transformers"
+  );
+  assert.match(
+    embeddingSrc,
+    /await import\(["']@huggingface\/transformers["']\)/,
+    "transformersLocal.ts must load @huggingface/transformers via await import()"
+  );
+
+  const workerSrc = readFileSync(
+    join(repoRoot, "open-sse/services/compression/engines/llmlingua/onnxWorker.ts"),
+    "utf8"
+  );
+  assert.doesNotMatch(
+    workerSrc,
+    /^\s*import\s+(?:[^'"]*?\s+from\s+)?["']@huggingface\/transformers["']/m,
+    "onnxWorker.ts must not statically import @huggingface/transformers"
   );
 });
diff --git a/tests/unit/cline-model-format-11099.test.ts b/tests/unit/cline-model-format-11099.test.ts
new file mode 100644
index 0000000000..3e5a87de43
--- /dev/null
+++ b/tests/unit/cline-model-format-11099.test.ts
@@ -0,0 +1,39 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { getModelsByProviderId } from "../../open-sse/config/providerModels.ts";
+import { parseClineRecommendedModels } from "../../open-sse/services/clinepassModels.ts";
+
+test("#11099: Cline provider catalog model IDs use valid modelType/model format", () => {
+  const models = getModelsByProviderId("cline");
+  assert.ok(models.length > 0, "cline provider must expose models");
+
+  for (const model of models) {
+    assert.match(
+      model.id,
+      /^[a-z0-9-]+-?[a-z0-9-]*\/[a-z0-9._:-]+$/i,
+      `Model ID '${model.id}' must follow provider/model format`
+    );
+    assert.notEqual(
+      model.id.split("/")[0],
+      "zai",
+      "Model ID must use 'z-ai' instead of invalid 'zai'"
+    );
+  }
+});
+
+test("#11099: parseClineRecommendedModels correctly extracts recommended/free models", () => {
+  const mockPayload = {
+    recommended: [
+      { id: "moonshotai/kimi-k3", name: "kimi-k3" },
+      { id: "x-ai/grok-4.5", name: "grok-4.5" },
+    ],
+    free: [{ id: "deepseek/deepseek-v4-flash", name: "deepseek-v4-flash" }],
+  };
+
+  const parsed = parseClineRecommendedModels(mockPayload);
+  assert.equal(parsed.length, 3);
+  assert.equal(parsed[0].id, "moonshotai/kimi-k3");
+  assert.equal(parsed[1].id, "x-ai/grok-4.5");
+  assert.equal(parsed[2].id, "deepseek/deepseek-v4-flash");
+});
diff --git a/tests/unit/columns-validation.test.ts b/tests/unit/columns-validation.test.ts
new file mode 100644
index 0000000000..41f70cddeb
--- /dev/null
+++ b/tests/unit/columns-validation.test.ts
@@ -0,0 +1,24 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import {
+  sanitizeRateLimitOverrides,
+  sanitizeQuotaWindowThresholds,
+} from "@/lib/db/providers/columns";
+
+test("sanitizeRateLimitOverrides surfaces rejected keys (blocking, not silent)", () => {
+  const r = sanitizeRateLimitOverrides({ rpm: 10, foo: 1, tpm: -1 });
+  assert.deepEqual(r.sanitized, { rpm: 10 });
+  assert.deepEqual(r.rejected.sort(), ["foo", "tpm"]);
+});
+
+test("sanitizeQuotaWindowThresholds surfaces key-too-long and out-of-range", () => {
+  const r = sanitizeQuotaWindowThresholds({ ["a".repeat(65)]: 50, win: 101 });
+  assert.ok(r.rejected.length >= 1);
+  assert.ok(r.rejected.includes("win"));
+});
+
+test("valid input yields no rejected keys", () => {
+  const r = sanitizeRateLimitOverrides({ rpm: 10, tpm: 20 });
+  assert.deepEqual(r.rejected, []);
+  assert.deepEqual(r.sanitized, { rpm: 10, tpm: 20 });
+});
diff --git a/tests/unit/combo-health-autopilot-counter.test.ts b/tests/unit/combo-health-autopilot-counter.test.ts
new file mode 100644
index 0000000000..1bd8330753
--- /dev/null
+++ b/tests/unit/combo-health-autopilot-counter.test.ts
@@ -0,0 +1,108 @@
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+
+import type {
+  ComboForecastResponse,
+  ComboHealthResponse,
+  ProviderAutopilotReport,
+} from "../../src/shared/types/utilization.ts";
+import { buildComboHealthAutopilotReport } from "../../src/lib/monitoring/comboHealthAutopilot.ts";
+
+function healthResponse(): ComboHealthResponse {
+  return {
+    timeRange: "24h",
+    combos: [
+      {
+        comboId: "c1",
+        comboName: "my-combo",
+        strategy: "fallback",
+        models: [],
+        targetHealth: [
+          {
+            executionKey: "e1",
+            stepId: "s1",
+            model: "m",
+            provider: "p",
+            connectionId: null,
+            label: null,
+            requests: 5,
+            successRate: 90,
+            avgLatencyMs: 100,
+            lastStatus: "error",
+            lastUsedAt: null,
+            quotaRemainingPct: 50,
+            quotaIsExhausted: false,
+            quotaTrend: "stable",
+            quotaScope: "provider",
+          },
+        ],
+        quotaHealth: { providers: [], worstRemainingPct: 100 },
+        usageSkew: { modelDistribution: [], giniCoefficient: 0 },
+        performance: { avgLatencyMs: 100, successRate: 1.0, totalRequests: 10 },
+      },
+    ],
+  };
+}
+
+function forecastResponse(): ComboForecastResponse {
+  return {
+    timeRange: "24h",
+    horizon: "30d",
+    asOf: new Date(0).toISOString(),
+    method: "linear_history",
+    combos: [],
+  };
+}
+
+function providerHealthResponse(): ProviderAutopilotReport {
+  return { providers: [] } as unknown as ProviderAutopilotReport;
+}
+
+function buildOptions() {
+  return {
+    range: "24h" as const,
+    horizon: "30d" as const,
+    healthResponse: healthResponse(),
+    forecastResponse: forecastResponse(),
+    providerHealthResponse: providerHealthResponse(),
+  };
+}
+
+describe("combo health autopilot counter", () => {
+  it("exposes suggestionCount and keeps actionableCount alias", async () => {
+    const report = await buildComboHealthAutopilotReport(buildOptions());
+    assert.equal(typeof report.summary.suggestionCount, "number");
+    assert.equal(report.summary.actionableCount, report.summary.suggestionCount);
+    const expected = report.combos.reduce(
+      (sum, combo) =>
+        sum + combo.issues.reduce((issueSum, issue) => issueSum + issue.actions.length, 0),
+      0
+    );
+    assert.equal(report.summary.suggestionCount, expected);
+  });
+
+  it("run_combo_test action links the dashboard with the combo id", async () => {
+    const report = await buildComboHealthAutopilotReport(buildOptions());
+    const actions = report.combos.flatMap((combo) => combo.issues.flatMap((i) => i.actions));
+    const runTest = actions.find((a) => a.type === "run_combo_test");
+    assert.ok(runTest, "run_combo_test action should exist");
+    assert.equal(typeof runTest.href, "string");
+    assert.ok(runTest.href?.includes("c1"), "href must carry the combo id");
+    assert.equal(
+      runTest.href?.includes("/api/combos/test?comboId="),
+      false,
+      "href must not target the GET-only API route (405)"
+    );
+  });
+
+  it("keeps every action in manual mode", async () => {
+    const report = await buildComboHealthAutopilotReport(buildOptions());
+    for (const combo of report.combos) {
+      for (const issue of combo.issues) {
+        for (const action of issue.actions) {
+          assert.equal(action.mode, "manual");
+        }
+      }
+    }
+  });
+});
diff --git a/tests/unit/config-audit-persistence.test.ts b/tests/unit/config-audit-persistence.test.ts
new file mode 100644
index 0000000000..4bad88bc5d
--- /dev/null
+++ b/tests/unit/config-audit-persistence.test.ts
@@ -0,0 +1,124 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-config-audit-"));
+process.env.DATA_DIR = TEST_DATA_DIR;
+process.env.DISABLE_SQLITE_AUTO_BACKUP = "true";
+
+const core = await import("../../src/lib/db/core.ts");
+const cleanup = await import("../../src/lib/db/cleanup.ts");
+const audit = await import("../../src/domain/configAudit.ts");
+
+type CountRow = { c: number };
+
+function resetStorage() {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+  fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
+}
+
+function countRows(): number {
+  const db = core.getDbInstance();
+  const row = db.prepare("SELECT COUNT(*) AS c FROM config_audit_log").get() as CountRow;
+  return row.c;
+}
+
+function insertOldRow(id: string, daysAgo: number) {
+  const db = core.getDbInstance();
+  const old = new Date(Date.now() - daysAgo * 24 * 60 * 60 * 1000).toISOString();
+  db.prepare(
+    `INSERT INTO config_audit_log
+       (id, timestamp, action, target, target_id, target_name, before_json, after_json, diff_json, source, note)
+     VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`
+  ).run(
+    id,
+    old,
+    "update",
+    "provider",
+    "p1",
+    "P1",
+    null,
+    null,
+    JSON.stringify({ added: [], removed: [], changed: [], isEmpty: true }),
+    "api",
+    null
+  );
+}
+
+test.beforeEach(() => {
+  resetStorage();
+});
+
+test.after(() => {
+  resetStorage();
+});
+
+test("recordChange persists to SQLite, not memory", () => {
+  const db = core.getDbInstance();
+  const tableRow = db
+    .prepare("SELECT count(*) as c FROM sqlite_master WHERE type='table' AND name='config_audit_log'")
+    .get() as CountRow;
+  assert.equal(tableRow.c, 1);
+
+  const e = audit.recordChange("update", "provider", "p1", "My Provider", { a: 1 }, { a: 2 }, "api", null);
+  assert.equal(countRows(), 1);
+
+  const { entries, total } = audit.getAuditLog({ target: "provider" });
+  assert.equal(total, 1);
+  assert.equal(entries[0].id, e.id);
+  assert.deepEqual(entries[0].diff.changed, [{ key: "a", from: 1, to: 2 }]);
+});
+
+test("pagination + filters read from SQLite", () => {
+  audit.recordChange("create", "combo", "c1", "C1", null, { models: ["m1"] }, "dashboard");
+  audit.recordChange("update", "combo", "c1", "C1", { models: ["m1"] }, { models: ["m1", "m2"] }, "api");
+
+  const { entries, total } = audit.getAuditLog({ target: "combo", limit: 1, offset: 0 });
+  assert.equal(total, 2);
+  assert.equal(entries.length, 1);
+});
+
+test("getRollbackState returns the before snapshot", () => {
+  const e = audit.recordChange("update", "policy", "pol1", "Pol", { x: 1 }, { x: 2 }, "api");
+  assert.deepEqual(audit.getRollbackState(e.id), { x: 1 });
+});
+
+test("computeDiff stays pure", () => {
+  const d = audit.computeDiff({ a: 1 }, { a: 2, b: 3 });
+  assert.deepEqual(d.added, ["b"]);
+  assert.deepEqual(d.changed, [{ key: "a", from: 1, to: 2 }]);
+});
+
+test("resetAuditLog clears persisted rows", () => {
+  audit.recordChange("update", "provider", "p1", "P1", { a: 1 }, { a: 2 }, "api");
+  assert.equal(countRows(), 1);
+  audit.resetAuditLog();
+  assert.equal(countRows(), 0);
+});
+
+test("cleanupConfigAudit prunes rows beyond retentionDays", async () => {
+  insertOldRow("audit-old", 40);
+  const r = await cleanup.cleanupConfigAudit(30);
+  assert.equal(r.deleted, 1);
+  assert.equal(countRows(), 0);
+});
+
+test("cleanupConfigAudit keeps recent rows within retention", async () => {
+  insertOldRow("audit-recent", 5);
+  const r = await cleanup.cleanupConfigAudit(30);
+  assert.equal(r.deleted, 0);
+  assert.equal(countRows(), 1);
+});
+
+test("runAutoCleanup includes a configAudit result", async () => {
+  insertOldRow("audit-old-2", 40);
+  const result = await cleanup.runAutoCleanup();
+  assert.ok(result.results.configAudit);
+  assert.equal(typeof result.results.configAudit.deleted, "number");
+  assert.equal(typeof result.results.configAudit.errors, "number");
+  assert.equal(result.results.configAudit.deleted, 1);
+  assert.equal(countRows(), 0);
+});
diff --git a/tests/unit/db-providers-split.test.ts b/tests/unit/db-providers-split.test.ts
index 220544ba2d..284a5193c8 100644
--- a/tests/unit/db-providers-split.test.ts
+++ b/tests/unit/db-providers-split.test.ts
@@ -52,26 +52,38 @@ describe("providers/columns — normalizeBooleanColumn", () => {
 });
 
 describe("providers/columns — sanitizeRateLimitOverrides", () => {
-  it("returns null for nullish / non-object / array input", () => {
-    assert.equal(sanitizeRateLimitOverrides(null), null);
-    assert.equal(sanitizeRateLimitOverrides(undefined), null);
-    assert.equal(sanitizeRateLimitOverrides("x"), null);
-    assert.equal(sanitizeRateLimitOverrides([1, 2]), null);
+  it("returns {sanitized:null,rejected:[]} for nullish / non-object / array input", () => {
+    assert.deepEqual(sanitizeRateLimitOverrides(null), { sanitized: null, rejected: [] });
+    assert.deepEqual(sanitizeRateLimitOverrides(undefined), { sanitized: null, rejected: [] });
+    assert.deepEqual(sanitizeRateLimitOverrides("x"), { sanitized: null, rejected: [] });
+    assert.deepEqual(sanitizeRateLimitOverrides([1, 2]), { sanitized: null, rejected: [] });
   });
-  it("keeps only allowed keys with non-negative integers", () => {
-    assert.deepEqual(sanitizeRateLimitOverrides({ rpm: 10, bogus: 5, tpm: -1 }), { rpm: 10 });
+  it("keeps only allowed keys with non-negative integers, reports the rest as rejected", () => {
+    assert.deepEqual(sanitizeRateLimitOverrides({ rpm: 10, bogus: 5, tpm: -1 }), {
+      sanitized: { rpm: 10 },
+      rejected: ["bogus", "tpm"],
+    });
   });
-  it("returns null when nothing valid remains", () => {
-    assert.equal(sanitizeRateLimitOverrides({ rpm: 1.5, nope: 3 }), null);
+  it("returns {sanitized:null} when nothing valid remains, with rejected keys", () => {
+    assert.deepEqual(sanitizeRateLimitOverrides({ rpm: 1.5, nope: 3 }), {
+      sanitized: null,
+      rejected: ["rpm", "nope"],
+    });
   });
 });
 
 describe("providers/columns — sanitizeQuotaWindowThresholds", () => {
-  it("keeps only 0-100 integers", () => {
-    assert.deepEqual(sanitizeQuotaWindowThresholds({ a: 50, b: 120, c: 0 }), { a: 50, c: 0 });
+  it("keeps only 0-100 integers, reports the rest as rejected", () => {
+    assert.deepEqual(sanitizeQuotaWindowThresholds({ a: 50, b: 120, c: 0 }), {
+      sanitized: { a: 50, c: 0 },
+      rejected: ["b"],
+    });
   });
-  it("returns null when empty", () => {
-    assert.equal(sanitizeQuotaWindowThresholds({ a: 200 }), null);
+  it("returns {sanitized:null} when nothing valid remains, with rejected keys", () => {
+    assert.deepEqual(sanitizeQuotaWindowThresholds({ a: 200 }), {
+      sanitized: null,
+      rejected: ["a"],
+    });
   });
 });
 
diff --git a/tests/unit/http-status-unprocessable-entity.test.ts b/tests/unit/http-status-unprocessable-entity.test.ts
new file mode 100644
index 0000000000..636e0ebe37
--- /dev/null
+++ b/tests/unit/http-status-unprocessable-entity.test.ts
@@ -0,0 +1,7 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { HTTP_STATUS } from "../../open-sse/config/constants.ts";
+
+test("HTTP_STATUS declares UNPROCESSABLE_ENTITY as 422", () => {
+  assert.equal(HTTP_STATUS.UNPROCESSABLE_ENTITY, 422);
+});
diff --git a/tests/unit/is-local-provider-11091.test.ts b/tests/unit/is-local-provider-11091.test.ts
new file mode 100644
index 0000000000..e5adc57b15
--- /dev/null
+++ b/tests/unit/is-local-provider-11091.test.ts
@@ -0,0 +1,38 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { isLocalProvider } from "../../open-sse/config/providerRegistry.ts";
+
+test("isLocalProvider detects RFC1918, CGNAT/Tailscale, and mDNS private hosts", () => {
+  // Local / loopback
+  assert.equal(isLocalProvider("http://localhost:11434/v1"), true);
+  assert.equal(isLocalProvider("http://127.0.0.1:11434/v1"), true);
+
+  // Docker 172.16/12
+  assert.equal(isLocalProvider("http://172.18.0.2:11434/v1"), true);
+
+  // RFC1918 LAN hosts (Issue #11091)
+  assert.equal(isLocalProvider("http://192.168.1.50:11434/v1"), true);
+  assert.equal(isLocalProvider("http://10.0.0.5:11434/v1"), true);
+
+  // Tailscale / CGNAT (100.64/10)
+  assert.equal(isLocalProvider("http://100.64.1.2:11434/v1"), true);
+
+  // Link-local (169.254/16)
+  assert.equal(isLocalProvider("http://169.254.1.1:11434/v1"), true);
+
+  // mDNS / private suffixes
+  assert.equal(isLocalProvider("http://studio.local:11434/v1"), true);
+  assert.equal(isLocalProvider("http://mybox.internal:11434/v1"), true);
+
+  // Public hosts (should be false)
+  assert.equal(isLocalProvider("https://api.openai.com/v1"), false);
+  assert.equal(isLocalProvider("https://api.anthropic.com/v1"), false);
+  assert.equal(isLocalProvider("http://8.8.8.8:8080/v1"), false);
+
+  // Fails open on missing or unparseable input (Issue #11091 review finding)
+  assert.equal(isLocalProvider(null), false);
+  assert.equal(isLocalProvider(undefined), false);
+  assert.equal(isLocalProvider(""), false);
+  assert.equal(isLocalProvider("not a url"), false);
+  assert.equal(isLocalProvider("file:///models"), false);
+});
diff --git a/tests/unit/learned-reasoning-effort-caps.test.ts b/tests/unit/learned-reasoning-effort-caps.test.ts
new file mode 100644
index 0000000000..5d69a342c3
--- /dev/null
+++ b/tests/unit/learned-reasoning-effort-caps.test.ts
@@ -0,0 +1,126 @@
+import { test, after, beforeEach } from "node:test";
+import assert from "node:assert/strict";
+import {
+  REASONING_EFFORT_ORDER,
+  parseReasoningEffortEnum,
+  recordLearnedReasoningEffort,
+  getLearnedReasoningEffort,
+  __test_resetLearnedReasoningEffortCaps,
+} from "../../open-sse/services/learnedReasoningEffortCaps.ts";
+
+beforeEach(() => {
+  __test_resetLearnedReasoningEffortCaps();
+});
+
+after(() => {
+  __test_resetLearnedReasoningEffortCaps();
+});
+
+// ── REASONING_EFFORT_ORDER ──────────────────────────────────────────────────
+
+test("REASONING_EFFORT_ORDER is none < minimal < low < medium < high < xhigh < max", () => {
+  assert.deepEqual(REASONING_EFFORT_ORDER, [
+    "none",
+    "minimal",
+    "low",
+    "medium",
+    "high",
+    "xhigh",
+    "max",
+  ]);
+});
+
+// ── parseReasoningEffortEnum ────────────────────────────────────────────────
+
+test("parseReasoningEffortEnum extracts the real OVH 422 enum (backtick-quoted)", () => {
+  const err =
+    "Failed to deserialize the JSON body into the target type: reasoning_effort: " +
+    "unknown variant `xhigh`, expected one of `none`, `high`, `medium`, `low`, `minimal`";
+  assert.deepEqual(parseReasoningEffortEnum(err), ["none", "high", "medium", "low", "minimal"]);
+});
+
+test("parseReasoningEffortEnum extracts a bare comma/and-joined enum with annotations", () => {
+  const err =
+    "Unexpected reasoning effort high. Supported types are xhigh (default), medium, and low.";
+  assert.deepEqual(parseReasoningEffortEnum(err), ["xhigh", "medium", "low"]);
+});
+
+test("parseReasoningEffortEnum drops unrecognized tokens", () => {
+  const err = "expected one of `none`, `turbo`, `high`";
+  assert.deepEqual(parseReasoningEffortEnum(err), ["none", "high"]);
+});
+
+test("parseReasoningEffortEnum returns null for unrelated error text", () => {
+  assert.equal(parseReasoningEffortEnum("connection refused"), null);
+  assert.equal(parseReasoningEffortEnum(""), null);
+  assert.equal(parseReasoningEffortEnum(null), null);
+  assert.equal(parseReasoningEffortEnum(undefined), null);
+});
+
+test("parseReasoningEffortEnum returns null when the list has no recognized token", () => {
+  assert.equal(parseReasoningEffortEnum("expected one of `foo`, `bar`"), null);
+});
+
+// ── recordLearnedReasoningEffort / getLearnedReasoningEffort ───────────────
+
+test("records the highest recognized value from the accepted list", () => {
+  const learned = recordLearnedReasoningEffort("ovh", "qwen3-coder-30b-a3b-instruct", [
+    "none",
+    "high",
+    "medium",
+    "low",
+    "minimal",
+  ]);
+  assert.equal(learned, "high");
+  assert.equal(getLearnedReasoningEffort("ovh", "qwen3-coder-30b-a3b-instruct"), "high");
+});
+
+test("returns null and stores nothing when acceptedValues has no recognized token", () => {
+  const learned = recordLearnedReasoningEffort("acme", "model-x", ["foo", "bar"]);
+  assert.equal(learned, null);
+  assert.equal(getLearnedReasoningEffort("acme", "model-x"), null);
+});
+
+test("monotonic decrease: a later, higher accepted-list never ratchets the cap back up", () => {
+  recordLearnedReasoningEffort("acme", "model-x", ["none", "low", "medium"]);
+  const learned = recordLearnedReasoningEffort("acme", "model-x", [
+    "none",
+    "low",
+    "medium",
+    "high",
+    "xhigh",
+  ]);
+  assert.equal(learned, "medium");
+  assert.equal(getLearnedReasoningEffort("acme", "model-x"), "medium");
+});
+
+test("a later, lower accepted-list does ratchet the cap down", () => {
+  recordLearnedReasoningEffort("acme", "model-x", ["none", "low", "medium", "high"]);
+  const learned = recordLearnedReasoningEffort("acme", "model-x", ["none", "low"]);
+  assert.equal(learned, "low");
+  assert.equal(getLearnedReasoningEffort("acme", "model-x"), "low");
+});
+
+test("getLearnedReasoningEffort returns null for unknown provider+model", () => {
+  assert.equal(getLearnedReasoningEffort("acme", "unknown-model"), null);
+});
+
+test("getLearnedReasoningEffort is keyed case-insensitively on provider+model", () => {
+  recordLearnedReasoningEffort("OVH", "Qwen3-Coder-30B", ["none", "high"]);
+  assert.equal(getLearnedReasoningEffort("ovh", "qwen3-coder-30b"), "high");
+  assert.equal(getLearnedReasoningEffort("OVH", "QWEN3-CODER-30B"), "high");
+});
+
+test("different providers for the same model id have independent caps", () => {
+  recordLearnedReasoningEffort("ovh", "shared-model", ["none", "high"]);
+  assert.equal(getLearnedReasoningEffort("openrouter", "shared-model"), null);
+});
+
+test("handles empty/null provider or model gracefully", () => {
+  assert.equal(getLearnedReasoningEffort("", "m"), null);
+  assert.equal(getLearnedReasoningEffort("p", ""), null);
+  assert.equal(getLearnedReasoningEffort(null, "m"), null);
+  assert.equal(getLearnedReasoningEffort("p", null), null);
+  assert.equal(recordLearnedReasoningEffort("", "m", ["high"]), null);
+  assert.equal(recordLearnedReasoningEffort("p", "", ["high"]), null);
+});
diff --git a/tests/unit/ollama-404-model-lockout-11071.test.ts b/tests/unit/ollama-404-model-lockout-11071.test.ts
new file mode 100644
index 0000000000..f58c29c07b
--- /dev/null
+++ b/tests/unit/ollama-404-model-lockout-11071.test.ts
@@ -0,0 +1,66 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import fs from "node:fs";
+import os from "node:os";
+import path from "node:path";
+
+const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-ollama-404-"));
+process.env.DATA_DIR = TEST_DATA_DIR;
+
+const core = await import("../../src/lib/db/core.ts");
+const providersDb = await import("../../src/lib/db/providers.ts");
+const auth = await import("../../src/sse/services/auth.ts");
+const { hasPerModelQuota, isModelLocked } = await import("../../open-sse/services/accountFallback.ts");
+
+async function resetStorage() {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+  fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
+}
+
+test.after(() => {
+  core.resetDbInstance();
+  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
+});
+
+test("hasPerModelQuota returns true for ollama-local and ollama providers", () => {
+  assert.equal(hasPerModelQuota("ollama-local"), true);
+  assert.equal(hasPerModelQuota("ollama"), true);
+});
+
+test("markAccountUnavailable locks only the missing model on a 404 from ollama-local", async () => {
+  await resetStorage();
+
+  const connection = await providersDb.createProviderConnection({
+    provider: "ollama-local",
+    authType: "none",
+    baseUrl: "http://127.0.0.1:11434/v1",
+    isActive: true,
+  });
+
+  const result = await auth.markAccountUnavailable(
+    connection.id,
+    404,
+    "model 'model-b' not found",
+    "ollama-local",
+    "model-b"
+  );
+
+  assert.equal(result.shouldFallback, true);
+
+  // The missing model must be locked
+  assert.equal(isModelLocked("ollama-local", connection.id, "model-b"), true);
+
+  // The connection in DB must remain active / not marked unavailable for sibling models
+  const connInDb = await providersDb.getProviderConnectionById(connection.id);
+  assert.notEqual(connInDb?.testStatus, "unavailable", "connection should not be marked unavailable connection-wide on a 404 model-not-found error");
+
+  // getProviderCredentials must still serve sibling models
+  const selectedForSibling = await auth.getProviderCredentials(
+    "ollama-local",
+    null,
+    null,
+    "model-a"
+  );
+  assert.ok(selectedForSibling && !("allExpired" in selectedForSibling), "sibling model-a must still be selected on the same connection");
+});
diff --git a/tests/unit/opencode-v2-config-11070.test.ts b/tests/unit/opencode-v2-config-11070.test.ts
new file mode 100644
index 0000000000..71bd8a9537
--- /dev/null
+++ b/tests/unit/opencode-v2-config-11070.test.ts
@@ -0,0 +1,40 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const opencodeConfig = await import("../../src/shared/services/opencodeConfig.ts");
+
+test("buildOpenCodeConfigDocument includes both V1 (provider) and V2 (providers) definitions", () => {
+  const doc = opencodeConfig.buildOpenCodeConfigDocument({
+    baseUrl: "http://localhost:20128/v1",
+    apiKey: "{env:OMNIROUTE_API_KEY}",
+    models: ["auto/best-coding"],
+  });
+
+  assert.ok(doc.provider?.omniroute, "V1 provider.omniroute must be present");
+  assert.equal(doc.provider.omniroute.npm, "@ai-sdk/openai-compatible");
+  assert.equal(doc.provider.omniroute.options.baseURL, "http://localhost:20128/v1");
+
+  assert.ok(doc.providers?.omniroute, "V2 providers.omniroute must be present");
+  assert.equal(doc.providers.omniroute.package, "@opencode-ai/ai/providers/openai-compatible");
+  assert.equal(doc.providers.omniroute.settings.baseURL, "http://localhost:20128/v1");
+  assert.equal(doc.providers.omniroute.settings.apiKey, "{env:OMNIROUTE_API_KEY}");
+  assert.ok(doc.providers.omniroute.models["auto/best-coding"].limit, "V2 model limit must be present");
+});
+
+test("mergeOpenCodeConfig preserves existing properties and updates both provider and providers", () => {
+  const existing = {
+    $schema: "https://opencode.ai/config.json",
+    customField: "keep-me",
+  };
+
+  const merged = opencodeConfig.mergeOpenCodeConfig(existing, {
+    baseUrl: "http://localhost:20128/v1",
+    apiKey: "sk_test_key",
+    models: ["auto/best-coding"],
+  });
+
+  assert.equal(merged.customField, "keep-me");
+  assert.ok(merged.provider?.omniroute);
+  assert.ok(merged.providers?.omniroute);
+  assert.equal(merged.providers.omniroute.settings.apiKey, "sk_test_key");
+});
diff --git a/tests/unit/opencode-zen-go-shared-models.test.ts b/tests/unit/opencode-zen-go-shared-models.test.ts
index 8c3079ad15..c1fb49e97d 100644
--- a/tests/unit/opencode-zen-go-shared-models.test.ts
+++ b/tests/unit/opencode-zen-go-shared-models.test.ts
@@ -24,3 +24,16 @@ test("every OPENCODE_ZEN_GO_SHARED_MODELS entry is present, unmodified, exactly
 test("OPENCODE_ZEN_GO_SHARED_MODELS is frozen (no accidental cross-registry mutation)", () => {
   assert.ok(Object.isFrozen(OPENCODE_ZEN_GO_SHARED_MODELS));
 });
+
+test("referenced non-shared model ids remain present", () => {
+  const goIds = new Set(opencode_goProvider.models.map((m) => m.id));
+  const zenIds = new Set(opencode_zenProvider.models.map((m) => m.id));
+  for (const id of ["minimax-m3", "glm-5.1"]) {
+    assert.ok(goIds.has(id) || zenIds.has(id), `expected ${id} in go or zen`);
+  }
+});
+
+test("models[0] is the intended dashboard default", () => {
+  assert.equal(opencode_goProvider.models[0].id, "glm-5.2");
+  assert.equal(opencode_zenProvider.models[0].id, "big-pickle");
+});
diff --git a/tests/unit/perplexity-discovery-filter.test.ts b/tests/unit/perplexity-discovery-filter.test.ts
new file mode 100644
index 0000000000..8533646746
--- /dev/null
+++ b/tests/unit/perplexity-discovery-filter.test.ts
@@ -0,0 +1,63 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import { PROVIDER_MODELS_CONFIG } from "../../src/app/api/providers/[id]/models/discovery/providerModelsConfig.ts";
+
+// Regression guard for #11060 — Perplexity's /v1/models endpoint lists the
+// Agent API catalog (vendor-prefixed ids like "anthropic/claude-fable-5"), but
+// chat requests always go to the classic /chat/completions endpoint, which only
+// accepts the Sonar family. Without a PROVIDER_MODELS_CONFIG entry, generic
+// model import pulled those agent-style ids into the connection's chat model
+// list and every routed request failed with 400 "Invalid model". The discovery
+// entry must exist and its parseResponse must keep only Sonar-family ids.
+
+test("perplexity has a discovery entry in PROVIDER_MODELS_CONFIG", () => {
+  const cfg = PROVIDER_MODELS_CONFIG.perplexity;
+  assert.ok(cfg, "expected a perplexity entry in PROVIDER_MODELS_CONFIG");
+  assert.equal(cfg.method, "GET");
+  assert.equal(cfg.url, "https://api.perplexity.ai/v1/models");
+  assert.equal(typeof cfg.parseResponse, "function");
+});
+
+test("perplexity parseResponse keeps only the Sonar family (#11060)", () => {
+  const cfg = PROVIDER_MODELS_CONFIG.perplexity;
+  const models = cfg.parseResponse({
+    object: "list",
+    data: [
+      { id: "anthropic/claude-fable-5", object: "model", owned_by: "anthropic" },
+      { id: "sonar-pro", object: "model", owned_by: "perplexity" },
+      { id: "sonar", object: "model", owned_by: "perplexity" },
+    ],
+  }) as Array<{ id: string }>;
+
+  assert.deepEqual(
+    models.map((model) => model.id),
+    ["sonar-pro", "sonar"]
+  );
+});
+
+test("perplexity parseResponse keeps every Sonar variant and drops non-Sonar ids", () => {
+  const cfg = PROVIDER_MODELS_CONFIG.perplexity;
+  const models = cfg.parseResponse({
+    data: [
+      { id: "sonar-deep-research" },
+      { id: "sonar-reasoning-pro" },
+      { id: "sonar-pro" },
+      { id: "sonar" },
+      { id: "openai/gpt-5" },
+      { id: "sonarish" },
+    ],
+  }) as Array<{ id: string }>;
+
+  assert.deepEqual(
+    models.map((model) => model.id),
+    ["sonar-deep-research", "sonar-reasoning-pro", "sonar-pro", "sonar"]
+  );
+});
+
+test("perplexity parseResponse tolerates empty and malformed payloads", () => {
+  const cfg = PROVIDER_MODELS_CONFIG.perplexity;
+  assert.deepEqual(cfg.parseResponse({ data: [] }), []);
+  assert.deepEqual(cfg.parseResponse(undefined), []);
+  assert.deepEqual(cfg.parseResponse({}), []);
+});
diff --git a/tests/unit/pollinations-api-key-required-11096.test.ts b/tests/unit/pollinations-api-key-required-11096.test.ts
new file mode 100644
index 0000000000..8b69fb8331
--- /dev/null
+++ b/tests/unit/pollinations-api-key-required-11096.test.ts
@@ -0,0 +1,11 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+import { providerAllowsOptionalApiKey } from "../../src/shared/constants/providers.js";
+
+test("pollinations provider requires an API key and does not allow optional API key", () => {
+  assert.equal(
+    providerAllowsOptionalApiKey("pollinations"),
+    false,
+    "pollinations must require an API key because anonymous completions are no longer supported"
+  );
+});
diff --git a/tests/unit/provider-connections-quota-threshold.test.ts b/tests/unit/provider-connections-quota-threshold.test.ts
index 5caaeea369..e7812557e9 100644
--- a/tests/unit/provider-connections-quota-threshold.test.ts
+++ b/tests/unit/provider-connections-quota-threshold.test.ts
@@ -106,18 +106,36 @@ test("updateProviderConnection with explicit null clears the column entirely", a
   assert.ok(reread.quotaWindowThresholds === null || reread.quotaWindowThresholds === undefined);
 });
 
-test("DB serializer drops out-of-range values silently", async () => {
-  // The DB module sanitizes the map on the way in; values outside 0-100 or
-  // non-integers are pruned. This is a defense in depth — the Zod schema
-  // already rejects them at the API boundary, but the DB shouldn't trust.
-  const created = await providersDb.createProviderConnection({
-    provider: "codex",
-    authType: "apikey",
-    name: "Codex Sanitize",
-    apiKey: "sk-san",
-    quotaWindowThresholds: { window5h: 95, bogus: 999, fractional: 1.5 },
-  });
-  assert.deepEqual(created.quotaWindowThresholds, { window5h: 95 });
+test("DB serializer refuses out-of-range / invalid values instead of dropping silently", async () => {
+  // the DB module must refuse the write (throw) rather than
+  // silently prune invalid keys/values on the way in, so operator intent is
+  // never lost without an error. The Zod schema already rejects at the API
+  // boundary; this is defense in depth for direct DB writers (seed/scripts).
+  await assert.rejects(
+    () =>
+      providersDb.createProviderConnection({
+        provider: "codex",
+        authType: "apikey",
+        name: "Codex Sanitize",
+        apiKey: "sk-san",
+        quotaWindowThresholds: { window5h: 95, bogus: 999, fractional: 1.5 },
+      }),
+    /rejected keys/
+  );
+});
+
+test("DB serializer refuses unknown rate-limit override keys instead of dropping silently", async () => {
+  await assert.rejects(
+    () =>
+      providersDb.createProviderConnection({
+        provider: "codex",
+        authType: "apikey",
+        name: "Codex Sanitize RLO",
+        apiKey: "sk-san-2",
+        rateLimitOverrides: { rpm: 10, bogus: 999, tpm: -1 },
+      }),
+    /rejected keys/
+  );
 });
 
 test("updateProviderConnectionSchema accepts a valid window map", () => {
diff --git a/tests/unit/provider-error-rules-operator.test.ts b/tests/unit/provider-error-rules-operator.test.ts
new file mode 100644
index 0000000000..6f8c9e52ef
--- /dev/null
+++ b/tests/unit/provider-error-rules-operator.test.ts
@@ -0,0 +1,135 @@
+import { describe, it, beforeEach } from "node:test";
+import assert from "node:assert/strict";
+import {
+  getProviderErrorRuleMatch,
+  setOperatorProviderErrorRules,
+  resolveRuleMatchBody,
+  honorsRuleLockScope,
+  type OperatorProviderErrorRule,
+} from "../../open-sse/config/providerErrorRules.ts";
+
+describe("operator error rules", () => {
+  beforeEach(() => {
+    // Isolate each test from the settings-backed cache.
+    setOperatorProviderErrorRules(undefined);
+  });
+
+  it("operator rule overrides the catalog registry for a provider", () => {
+    const op: Record<string, OperatorProviderErrorRule[]> = {
+      nvidia: [{ status: 404, match: "Not found for account", scope: "model", cooldownMs: 1000 }],
+    };
+    const m = getProviderErrorRuleMatch("nvidia", 404, null, "Not found for account id 123", op);
+    assert.ok(m, "operator rule should match");
+    assert.equal(m.scope, "model");
+    assert.equal(m.cooldownMs, 1000);
+  });
+
+  it("operator rule wins even when a catalog rule would also match", () => {
+    const op: Record<string, OperatorProviderErrorRule[]> = {
+      openrouter: [{ status: 402, match: "credits exhausted", scope: "model" }],
+    };
+    const m = getProviderErrorRuleMatch("openrouter", 402, null, "credits exhausted on key", op);
+    assert.ok(m);
+    // Catalog rule for openrouter/402 uses scope "connection"; the operator
+    // override must take precedence.
+    assert.equal(m.scope, "model");
+  });
+
+  it("operator can reclassify a 401 before the global permanent rule", () => {
+    const op: Record<string, OperatorProviderErrorRule[]> = {
+      acme: [
+        { status: 401, match: "transient quota", scope: "connection", reason: "quota_exhausted" },
+      ],
+    };
+    const m = getProviderErrorRuleMatch("acme", 401, null, "transient quota — retry shortly", op);
+    assert.ok(m);
+    assert.equal(m.scope, "connection");
+    assert.equal(m.reason, "quota_exhausted");
+  });
+
+  it("unknown provider with no operator rule returns null (no throw)", () => {
+    const m = getProviderErrorRuleMatch("unknown-provider", 402, null, "anything");
+    assert.equal(m, null);
+  });
+
+  it("substring match is case-insensitive", () => {
+    const op: Record<string, OperatorProviderErrorRule[]> = {
+      nvidia: [{ status: 404, match: "NOT FOUND", scope: "model" }],
+    };
+    const m = getProviderErrorRuleMatch("nvidia", 404, null, "Body says Not Found Here", op);
+    assert.ok(m);
+    assert.equal(m.scope, "model");
+  });
+
+  it("status must match before the substring is considered", () => {
+    const op: Record<string, OperatorProviderErrorRule[]> = {
+      nvidia: [{ status: 404, match: "not found", scope: "model" }],
+    };
+    // 500 with the same body text must NOT match a 404 rule.
+    const m = getProviderErrorRuleMatch("nvidia", 500, null, "not found for account", op);
+    assert.equal(m, null);
+  });
+
+  it("without an operator override the catalog registry is intact", () => {
+    const m = getProviderErrorRuleMatch("openrouter", 402, null, "credits exhausted on key");
+    assert.ok(m);
+    assert.equal(m.scope, "connection");
+    assert.equal(m.cooldownMs, 2 * 60 * 1000);
+  });
+
+  it("reads the settings-backed cache via setOperatorProviderErrorRules", () => {
+    setOperatorProviderErrorRules({
+      nvidia: [{ status: 404, match: "Not found", scope: "model" }],
+    });
+    const m = getProviderErrorRuleMatch("nvidia", 404, null, "Not found for account");
+    assert.ok(m);
+    assert.equal(m.scope, "model");
+    // Provider key lookup is case-insensitive.
+    const m2 = getProviderErrorRuleMatch("NVIDIA", 404, null, "Not found here");
+    assert.ok(m2);
+    assert.equal(m2.scope, "model");
+  });
+
+  // Regression coverage for #11104's original gap: an operator rule for any
+  // provider outside the built-in FULL_TEXT_RULE_PROVIDERS/
+  // HONORS_RULE_LOCK_SCOPE_PROVIDERS allowlists was silently text-blind (only
+  // {code,type} reached the matcher) and had its declared scope dropped by the
+  // persistence layer. Declaring an operator rule for a provider must be
+  // sufficient by itself — no separate allowlist entry required.
+  describe("operator rule bypasses the built-in allowlists", () => {
+    it("resolveRuleMatchBody hands the full error text once an operator rule exists for the provider", () => {
+      setOperatorProviderErrorRules({
+        acme: [{ status: 404, match: "model withdrawn", scope: "model" }],
+      });
+      const body = resolveRuleMatchBody("acme", { code: "not_found" }, "Model withdrawn upstream");
+      assert.equal(body, "Model withdrawn upstream");
+    });
+
+    it("resolveRuleMatchBody keeps returning the structured error for a provider with no operator rule", () => {
+      const body = resolveRuleMatchBody("acme", { code: "not_found" }, "Model withdrawn upstream");
+      assert.deepEqual(body, { code: "not_found" });
+    });
+
+    it("honorsRuleLockScope is true once an operator rule exists for the provider", () => {
+      assert.equal(honorsRuleLockScope("acme"), false);
+      setOperatorProviderErrorRules({
+        acme: [{ status: 404, match: "model withdrawn", scope: "model" }],
+      });
+      assert.equal(honorsRuleLockScope("acme"), true);
+    });
+
+    it("an operator rule for a non-allowlisted provider matches on raw body text end to end", () => {
+      setOperatorProviderErrorRules({
+        acme: [{ status: 404, match: "model withdrawn", scope: "model" }],
+      });
+      const body = resolveRuleMatchBody(
+        "acme",
+        { code: "not_found" },
+        "Error: model withdrawn upstream"
+      );
+      const m = getProviderErrorRuleMatch("acme", 404, null, body);
+      assert.ok(m, "operator rule should match once resolveRuleMatchBody hands it the raw text");
+      assert.equal(m.scope, "model");
+    });
+  });
+});
diff --git a/tests/unit/provider-route-schemas.test.ts b/tests/unit/provider-route-schemas.test.ts
index ca61c4f28d..30b5d4bd8c 100644
--- a/tests/unit/provider-route-schemas.test.ts
+++ b/tests/unit/provider-route-schemas.test.ts
@@ -5,17 +5,19 @@ const { createProviderSchema, providersBatchTestSchema } =
   await import("../../src/shared/validation/schemas.ts");
 const { providerAllowsOptionalApiKey } = await import("../../src/shared/constants/providers.ts");
 
-test("Pollinations is treated as a keyless-capable provider", () => {
-  assert.equal(providerAllowsOptionalApiKey("pollinations"), true);
+// #11117: Pollinations no longer serves anonymous requests (401 without a key),
+// so it left EXPLICIT_OPTIONAL_APIKEY_PROVIDER_IDS — key is now required.
+test("Pollinations requires an API key", () => {
+  assert.equal(providerAllowsOptionalApiKey("pollinations"), false);
 });
 
-test("createProviderSchema allows Pollinations without apiKey", () => {
+test("createProviderSchema rejects Pollinations without apiKey", () => {
   const result = createProviderSchema.safeParse({
     provider: "pollinations",
     name: "Pollinations",
   });
 
-  assert.equal(result.success, true);
+  assert.equal(result.success, false);
 });
 
 test("providersBatchTestSchema accepts cloud-agent batch mode", () => {
diff --git a/tests/unit/providers-patch-400.test.ts b/tests/unit/providers-patch-400.test.ts
new file mode 100644
index 0000000000..8ec1cbdf3b
--- /dev/null
+++ b/tests/unit/providers-patch-400.test.ts
@@ -0,0 +1,42 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import { updateProviderConnectionSchema } from "@/shared/validation/schemas/provider";
+
+test("PATCH rateLimitOverrides {rpm:\"60\"} coerces to a valid number", () => {
+  const r = updateProviderConnectionSchema.safeParse({ rateLimitOverrides: { rpm: "60" } });
+  assert.equal(r.success, true);
+});
+
+test("unknown key in rateLimitOverrides is rejected (no silent drop)", () => {
+  const r = updateProviderConnectionSchema.safeParse({ rateLimitOverrides: { rpm: 10, foo: 1 } });
+  assert.equal(r.success, false);
+  const flaggedFoo = r.error!.issues.some(
+    (i) => i.path.includes("foo") || (i as { keys?: string[] }).keys?.includes("foo") || i.message.includes("foo")
+  );
+  assert.ok(
+    flaggedFoo,
+    `expected an issue flagging "foo", got: ${JSON.stringify(r.error!.issues)}`
+  );
+});
+
+test("quotaWindowThresholds key longer than 64 chars is rejected", () => {
+  const r = updateProviderConnectionSchema.safeParse({
+    quotaWindowThresholds: { ["a".repeat(65)]: 50 },
+  });
+  assert.equal(r.success, false);
+});
+
+test("empty string rate limit value is rejected (coerce \"\"→0 trap)", () => {
+  const r = updateProviderConnectionSchema.safeParse({ rateLimitOverrides: { rpm: "" } });
+  assert.equal(r.success, false);
+});
+
+test("non-numeric rate limit value is rejected", () => {
+  const r = updateProviderConnectionSchema.safeParse({ rateLimitOverrides: { rpm: "60abc" } });
+  assert.equal(r.success, false);
+});
+
+test("quotaWindowThresholds value outside 0-100 is rejected", () => {
+  const r = updateProviderConnectionSchema.safeParse({ quotaWindowThresholds: { win: 101 } });
+  assert.equal(r.success, false);
+});
diff --git a/tests/unit/reasoning-cache.test.ts b/tests/unit/reasoning-cache.test.ts
index e13dfec7a9..aded469ca8 100644
--- a/tests/unit/reasoning-cache.test.ts
+++ b/tests/unit/reasoning-cache.test.ts
@@ -664,6 +664,7 @@ describe("Reasoning Replay Cache — Translator Replay", () => {
       {
         type: "reasoning",
         content: [{ type: "reasoning_text", text: "Cached Chat continuation reasoning" }],
+        summary: [],
       }
     );
   });
diff --git a/tests/unit/reasoning-effort-clamp-and-retry.test.ts b/tests/unit/reasoning-effort-clamp-and-retry.test.ts
new file mode 100644
index 0000000000..a97ac16df0
--- /dev/null
+++ b/tests/unit/reasoning-effort-clamp-and-retry.test.ts
@@ -0,0 +1,110 @@
+import { test, after, beforeEach } from "node:test";
+import assert from "node:assert/strict";
+import { BaseExecutor } from "../../open-sse/executors/base.ts";
+import {
+  getLearnedReasoningEffort,
+  recordLearnedReasoningEffort,
+  __test_resetLearnedReasoningEffortCaps,
+} from "../../open-sse/services/learnedReasoningEffortCaps.ts";
+
+const OVH_422_BODY = JSON.stringify({
+  error: {
+    message:
+      "Failed to deserialize the JSON body into the target type: reasoning_effort: " +
+      "unknown variant `xhigh`, expected one of `none`, `high`, `medium`, `low`, `minimal`",
+  },
+});
+
+// Passthrough executor: returns the body unchanged so we assert on exactly what
+// base.ts sends upstream.
+class SimpleExecutor extends BaseExecutor {
+  constructor() {
+    super("openai-compatible-chat-eaff6869", {
+      baseUrls: ["https://oai.endpoints.kepler.ai.cloud.ovh.net/v1/chat/completions"],
+    });
+  }
+  async transformRequest(_model: string, body: Record<string, unknown>) {
+    return { ...body };
+  }
+}
+
+beforeEach(() => {
+  __test_resetLearnedReasoningEffortCaps();
+});
+
+after(() => {
+  __test_resetLearnedReasoningEffortCaps();
+});
+
+test("422 'unknown variant xhigh, expected one of ...' clamps reasoning_effort and retries once", async () => {
+  const executor = new SimpleExecutor();
+  const originalFetch = globalThis.fetch;
+  const capturedBodies: Record<string, unknown>[] = [];
+
+  globalThis.fetch = async (_url: string | URL | Request, init: RequestInit = {}) => {
+    const body = JSON.parse(String(init.body));
+    capturedBodies.push(body);
+    if (capturedBodies.length === 1) {
+      return new Response(OVH_422_BODY, {
+        status: 422,
+        headers: { "Content-Type": "application/json" },
+      });
+    }
+    return new Response(JSON.stringify({ ok: true }), {
+      status: 200,
+      headers: { "Content-Type": "application/json" },
+    });
+  };
+
+  try {
+    const result = await executor.execute({
+      model: "qwen3-coder-30b-a3b-instruct",
+      body: { reasoning_effort: "xhigh" },
+      stream: false,
+      credentials: {},
+    });
+    assert.equal(capturedBodies.length, 2);
+    assert.equal(capturedBodies[0].reasoning_effort, "xhigh");
+    assert.equal(capturedBodies[1].reasoning_effort, "high");
+    assert.equal(
+      getLearnedReasoningEffort("openai-compatible-chat-eaff6869", "qwen3-coder-30b-a3b-instruct"),
+      "high"
+    );
+    assert.equal(result.response.status, 200);
+  } finally {
+    globalThis.fetch = originalFetch;
+  }
+});
+
+test("a second request for the same provider+model sends the learned value on the first try", async () => {
+  const executor = new SimpleExecutor();
+  const originalFetch = globalThis.fetch;
+  const capturedBodies: Record<string, unknown>[] = [];
+
+  globalThis.fetch = async (_url: string | URL | Request, init: RequestInit = {}) => {
+    const body = JSON.parse(String(init.body));
+    capturedBodies.push(body);
+    return new Response(JSON.stringify({ ok: true }), {
+      status: 200,
+      headers: { "Content-Type": "application/json" },
+    });
+  };
+
+  try {
+    recordLearnedReasoningEffort(
+      "openai-compatible-chat-eaff6869",
+      "qwen3-coder-30b-a3b-instruct",
+      ["none", "high", "medium", "low", "minimal"]
+    );
+    await executor.execute({
+      model: "qwen3-coder-30b-a3b-instruct",
+      body: { reasoning_effort: "xhigh" },
+      stream: false,
+      credentials: {},
+    });
+    assert.equal(capturedBodies.length, 1);
+    assert.equal(capturedBodies[0].reasoning_effort, "high");
+  } finally {
+    globalThis.fetch = originalFetch;
+  }
+});
diff --git a/tests/unit/reasoning-effort-learned-capability.test.ts b/tests/unit/reasoning-effort-learned-capability.test.ts
new file mode 100644
index 0000000000..210a451341
--- /dev/null
+++ b/tests/unit/reasoning-effort-learned-capability.test.ts
@@ -0,0 +1,91 @@
+import { test, after, beforeEach } from "node:test";
+import assert from "node:assert/strict";
+import { sanitizeReasoningEffortForProvider } from "../../open-sse/executors/base/reasoningEffort.ts";
+import {
+  recordLearnedReasoningEffort,
+  __test_resetLearnedReasoningEffortCaps,
+} from "../../open-sse/services/learnedReasoningEffortCaps.ts";
+
+beforeEach(() => {
+  __test_resetLearnedReasoningEffortCaps();
+});
+
+after(() => {
+  __test_resetLearnedReasoningEffortCaps();
+});
+
+test("unregistered/custom provider+model: no learned cap yet sends xhigh unchanged", () => {
+  const body = { reasoning_effort: "xhigh" };
+  const result = sanitizeReasoningEffortForProvider(
+    body,
+    "openai-compatible-chat-eaff6869",
+    "qwen3-coder-30b-a3b-instruct"
+  ) as { reasoning_effort: string };
+  assert.equal(result.reasoning_effort, "xhigh");
+});
+
+test("unregistered/custom provider+model: a learned cap clamps xhigh down to it", () => {
+  recordLearnedReasoningEffort("openai-compatible-chat-eaff6869", "qwen3-coder-30b-a3b-instruct", [
+    "none",
+    "high",
+    "medium",
+    "low",
+    "minimal",
+  ]);
+  const body = { reasoning_effort: "xhigh" };
+  const result = sanitizeReasoningEffortForProvider(
+    body,
+    "openai-compatible-chat-eaff6869",
+    "qwen3-coder-30b-a3b-instruct"
+  ) as { reasoning_effort: string };
+  assert.equal(result.reasoning_effort, "high");
+});
+
+test("learned cap only clamps when the requested effort is above it", () => {
+  recordLearnedReasoningEffort("acme", "model-x", ["none", "low", "medium"]);
+  const body = { reasoning_effort: "low" };
+  const result = sanitizeReasoningEffortForProvider(body, "acme", "model-x") as {
+    reasoning_effort: string;
+  };
+  assert.equal(result.reasoning_effort, "low");
+});
+
+test("registry says supportsXHighEffort:false (and no supportsMax path) with a learned cap below 'high': uses the learned cap, not the hardcoded 'high'", () => {
+  // claude-haiku-4-5 is registered with supportsXHighEffort:false
+  // (open-sse/config/providers/registry/claude/index.ts) and its family is
+  // excluded from supportsClaudeMaxEffort (CLAUDE_MAX_EFFORT_UNSUPPORTED_FAMILY_PATTERNS
+  // in providerModels.ts), so it reaches the hardcoded-"high" line today —
+  // a real registry-covered case. Teach a lower cap and confirm it wins.
+  recordLearnedReasoningEffort("claude", "claude-haiku-4-5-20251001", ["none", "low", "medium"]);
+  const body = { reasoning_effort: "xhigh" };
+  const result = sanitizeReasoningEffortForProvider(
+    body,
+    "claude",
+    "claude-haiku-4-5-20251001"
+  ) as {
+    reasoning_effort: string;
+  };
+  assert.equal(result.reasoning_effort, "medium");
+});
+
+test("registry says supportsXHighEffort:false with no learned cap: falls back to hardcoded 'high' (unchanged behavior)", () => {
+  const body = { reasoning_effort: "xhigh" };
+  const result = sanitizeReasoningEffortForProvider(
+    body,
+    "claude",
+    "claude-haiku-4-5-20251001"
+  ) as {
+    reasoning_effort: string;
+  };
+  assert.equal(result.reasoning_effort, "high");
+});
+
+test("deepseek's non-ordinal max<->xhigh translation is untouched by the learned-cap catch-all", () => {
+  recordLearnedReasoningEffort("deepseek", "deepseek-v4", ["none", "low"]);
+  const body = { reasoning_effort: "xhigh" };
+  const result = sanitizeReasoningEffortForProvider(body, "deepseek", "deepseek-v4") as {
+    reasoning_effort: string;
+  };
+  // deepseek's special case returns early — xhigh -> max, never reaches the catch-all.
+  assert.equal(result.reasoning_effort, "max");
+});
diff --git a/tests/unit/reasoning-input-policy-summary-11108.test.ts b/tests/unit/reasoning-input-policy-summary-11108.test.ts
new file mode 100644
index 0000000000..4317225dfb
--- /dev/null
+++ b/tests/unit/reasoning-input-policy-summary-11108.test.ts
@@ -0,0 +1,145 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+
+const { applyReasoningInputPolicy } =
+  await import("../../open-sse/services/reasoningInputPolicy.ts");
+
+test("#11108 applyReasoningInputPolicy defaults summary on a kept opaque reasoning item", () => {
+  const body: Record<string, unknown> = {
+    input: [
+      {
+        type: "reasoning",
+        id: "rs_example",
+        encrypted_content: "opaque-blob",
+      },
+    ],
+  };
+
+  applyReasoningInputPolicy(body, "responses", {
+    provider: "opencode",
+    preserveEncryptedReasoning: true,
+  });
+
+  const input = body.input as Record<string, unknown>[];
+  assert.equal(input.length, 1);
+  assert.deepEqual(input[0].summary, []);
+});
+
+test("#11108 applyReasoningInputPolicy preserves an existing summary on a kept reasoning item", () => {
+  const body: Record<string, unknown> = {
+    input: [
+      {
+        type: "reasoning",
+        id: "rs_example",
+        encrypted_content: "opaque-blob",
+        summary: [{ type: "summary_text", text: "Planning." }],
+      },
+    ],
+  };
+
+  applyReasoningInputPolicy(body, "responses", {
+    provider: "opencode",
+    preserveEncryptedReasoning: true,
+  });
+
+  const input = body.input as Record<string, unknown>[];
+  assert.deepEqual(input[0].summary, [{ type: "summary_text", text: "Planning." }]);
+});
+
+test("#11108 applyReasoningInputPolicy defaults summary on an opaque item surviving incompatible-drop", () => {
+  // Mixed item (plaintext + opaque) on an opaque-only transport is incompatible;
+  // dropIncompatibleResponsesReasoning() strips the plaintext content but keeps
+  // the opaque item alive — it must still get a default `summary`.
+  const body: Record<string, unknown> = {
+    input: [
+      {
+        type: "reasoning",
+        id: "rs_mixed",
+        content: [{ type: "reasoning_text", text: "inspect first" }],
+        encrypted_content: "opaque-blob",
+      },
+    ],
+  };
+
+  const result = applyReasoningInputPolicy(body, "responses", {
+    provider: "codex",
+    onIncompatibleReasoning: "drop",
+  });
+
+  assert.equal(result.incompatibleReasoning, false);
+  const input = body.input as Record<string, unknown>[];
+  assert.equal(input.length, 1);
+  assert.equal(input[0].content, undefined);
+  assert.equal(input[0].encrypted_content, "opaque-blob");
+  assert.deepEqual(input[0].summary, []);
+});
+
+test("#11108 applyReasoningInputPolicy strips a non-string id on a kept opaque reasoning item", () => {
+  // Same gap class as the summary fix above: opencode/zen also omits `id`
+  // entirely (surfaced by the client as `id: null`) on opaque-only reasoning
+  // items instead of a `rs_...` string. Replaying that shape verbatim trips
+  // strict Responses-API validators with "Expected 'id' to be a string."
+  const body: Record<string, unknown> = {
+    input: [
+      {
+        type: "reasoning",
+        id: null,
+        encrypted_content: "opaque-blob",
+      },
+    ],
+  };
+
+  applyReasoningInputPolicy(body, "responses", {
+    provider: "opencode",
+    preserveEncryptedReasoning: true,
+  });
+
+  const input = body.input as Record<string, unknown>[];
+  assert.equal(input.length, 1);
+  assert.equal("id" in input[0], false);
+});
+
+test("#11108 applyReasoningInputPolicy strips a non-string id on a non-reasoning item (function_call)", () => {
+  // Same gap class, generic branch: any non-"reasoning" input item (function_call,
+  // message, ...) only stripped `id` when it was already a valid string, so a
+  // malformed `id` (e.g. `null`, mirroring the opencode/zen omission pattern)
+  // on a function_call item survived replay untouched.
+  const body: Record<string, unknown> = {
+    input: [
+      {
+        type: "function_call",
+        id: null,
+        call_id: "call_abc",
+        name: "bash",
+        arguments: "{}",
+      },
+    ],
+  };
+
+  applyReasoningInputPolicy(body, "responses", { provider: "opencode" });
+
+  const input = body.input as Record<string, unknown>[];
+  assert.equal(input.length, 1);
+  assert.equal("id" in input[0], false);
+  assert.equal(input[0].call_id, "call_abc");
+});
+
+test("#11108 applyReasoningInputPolicy preserves a valid string id on a kept opaque reasoning item", () => {
+  const body: Record<string, unknown> = {
+    input: [
+      {
+        type: "reasoning",
+        id: "rs_example",
+        encrypted_content: "opaque-blob",
+      },
+    ],
+  };
+
+  applyReasoningInputPolicy(body, "responses", {
+    provider: "opencode",
+    preserveEncryptedReasoning: true,
+  });
+
+  const input = body.input as Record<string, unknown>[];
+  assert.equal(input[0].id, "rs_example");
+});
diff --git a/tests/unit/responses-input-sanitizer-name.test.ts b/tests/unit/responses-input-sanitizer-name.test.ts
index 7ac09604c7..4eec2d6c10 100644
--- a/tests/unit/responses-input-sanitizer-name.test.ts
+++ b/tests/unit/responses-input-sanitizer-name.test.ts
@@ -73,6 +73,23 @@ test("keeps valid server reasoning item ids", () => {
   assert.equal(result[0].id, "rs_123");
 });
 
+test("strips a non-string reasoning item id instead of passing it through (#11108)", () => {
+  // Same gap class fixed in reasoningInputPolicy.ts: some upstreams (e.g.
+  // opencode/zen) send `id: null` instead of omitting it. The previous
+  // `typeof record.id !== "string"` guard returned the record unchanged in
+  // that case, letting a malformed id reach a strict Responses-API upstream.
+  const items = [
+    {
+      id: null,
+      type: "reasoning",
+      summary: [{ type: "summary_text", text: "cached reasoning" }],
+    },
+  ];
+  const result = sanitizeResponsesInputItems(items) as Array<Record<string, unknown>>;
+  assert.equal("id" in result[0], false);
+  assert.equal(result[0].type, "reasoning");
+});
+
 test("normalizes user image_url content parts to input_image", () => {
   const items = [
     {
diff --git a/tests/unit/responses-parallel-tool-calls-index.test.ts b/tests/unit/responses-parallel-tool-calls-index.test.ts
new file mode 100644
index 0000000000..f825f8700b
--- /dev/null
+++ b/tests/unit/responses-parallel-tool-calls-index.test.ts
@@ -0,0 +1,294 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+const { openaiResponsesToOpenAIResponse } =
+  await import("../../open-sse/translator/response/openai-responses.ts");
+
+// Issue: 2+ `function_call` items opened (response.output_item.added) before any of
+// them closes (response.output_item.done) — a genuine parallel tool-call dispatch —
+// causes `state.toolCallIndex` (only incremented in the `.done` handler) to stay at 0
+// for every "added" header chunk. Clients that key their tool-call accumulator by
+// `delta.tool_calls[].index` (e.g. opencode's github-copilot chat-language-model
+// stream parser) then see the *first* `.done` argument chunk at index 1/2 with no
+// prior header and no `id`, and throw "Expected 'id' to be a string."
+test("Responses -> OpenAI: parallel function_call items get distinct index+id on the added header", () => {
+  const state = {};
+
+  const added0 = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      item: { type: "function_call", call_id: "call_0", name: "task" },
+    },
+    state
+  );
+  const added1 = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      item: { type: "function_call", call_id: "call_1", name: "task" },
+    },
+    state
+  );
+  const added2 = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      item: { type: "function_call", call_id: "call_2", name: "task" },
+    },
+    state
+  );
+
+  const headers = [added0, added1, added2].map((r) => r.choices[0].delta.tool_calls[0]);
+
+  assert.deepEqual(
+    headers.map((h) => h.index),
+    [0, 1, 2],
+    "each parallel tool call must get its own header index, not all 0"
+  );
+  assert.deepEqual(
+    headers.map((h) => h.id),
+    ["call_0", "call_1", "call_2"]
+  );
+
+  const done0 = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_0", name: "task", arguments: '{"i":0}' },
+    },
+    state
+  );
+  const done1 = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_1", name: "task", arguments: '{"i":1}' },
+    },
+    state
+  );
+  const done2 = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_2", name: "task", arguments: '{"i":2}' },
+    },
+    state
+  );
+
+  assert.deepEqual(
+    [done0, done1, done2].map((r) => r.choices[0].delta.tool_calls[0].index),
+    [0, 1, 2],
+    "argument chunks must reuse the SAME index assigned at .added time for each call_id"
+  );
+});
+
+test("Responses -> OpenAI: parallel calls closed out of order keep their own index", () => {
+  const state = {};
+
+  for (const callId of ["call_a", "call_b", "call_c"]) {
+    openaiResponsesToOpenAIResponse(
+      {
+        type: "response.output_item.added",
+        item: { type: "function_call", call_id: callId, name: "task" },
+      },
+      state
+    );
+  }
+
+  // Close in reverse order: c, then a, then b.
+  const doneC = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_c", name: "task", arguments: "{}" },
+    },
+    state
+  );
+  const doneA = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_a", name: "task", arguments: "{}" },
+    },
+    state
+  );
+  const doneB = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_b", name: "task", arguments: "{}" },
+    },
+    state
+  );
+
+  assert.equal(doneC.choices[0].delta.tool_calls[0].index, 2);
+  assert.equal(doneA.choices[0].delta.tool_calls[0].index, 0);
+  assert.equal(doneB.choices[0].delta.tool_calls[0].index, 1);
+});
+
+test("Responses -> OpenAI: argument deltas interleaved across 2 parallel calls do not get glued together", () => {
+  const state = {};
+
+  openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      item: { type: "function_call", call_id: "call_x", name: "Read", id: "fc_call_x" },
+    },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      item: { type: "function_call", call_id: "call_y", name: "Read", id: "fc_call_y" },
+    },
+    state
+  );
+
+  // Interleave argument deltas by item_id — x, y, x, y — before either closes.
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", item_id: "fc_call_x", delta: '{"filePath"' },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", item_id: "fc_call_y", delta: '{"filePath"' },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", item_id: "fc_call_x", delta: ':"/a.txt"}' },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", item_id: "fc_call_y", delta: ':"/b.txt"}' },
+    state
+  );
+
+  const doneX = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_x", name: "Read" },
+    },
+    state
+  );
+  const doneY = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_y", name: "Read" },
+    },
+    state
+  );
+
+  assert.equal(doneX.choices[0].delta.tool_calls[0].function.arguments, '{"filePath":"/a.txt"}');
+  assert.equal(doneY.choices[0].delta.tool_calls[0].function.arguments, '{"filePath":"/b.txt"}');
+});
+
+test("Responses -> OpenAI: a deferred (nameless) call that never resolves a name never consumes an index", () => {
+  const state = {};
+
+  openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      item: { type: "function_call", call_id: "call_deferred", name: "" },
+    },
+    state
+  );
+  const done = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_deferred", name: " " },
+    },
+    state
+  );
+
+  assert.equal(done, null);
+  assert.equal(state.toolCallIndex, 0);
+});
+
+test("Responses -> OpenAI: argument deltas interleaved across 2 parallel calls resolve by output_index when the upstream omits item_id", () => {
+  const state = {};
+
+  openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      output_index: 0,
+      item: { type: "function_call", call_id: "call_p", name: "Read" },
+    },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      output_index: 1,
+      item: { type: "function_call", call_id: "call_q", name: "Read" },
+    },
+    state
+  );
+
+  // No item_id on any of these deltas — only output_index, which the Responses API
+  // guarantees on every streamed event regardless of whether item_id is also sent.
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", output_index: 0, delta: '{"filePath"' },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", output_index: 1, delta: '{"filePath"' },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", output_index: 0, delta: ':"/p.txt"}' },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", output_index: 1, delta: ':"/q.txt"}' },
+    state
+  );
+
+  const doneP = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_p", name: "Read" },
+    },
+    state
+  );
+  const doneQ = openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.done",
+      item: { type: "function_call", call_id: "call_q", name: "Read" },
+    },
+    state
+  );
+
+  assert.equal(doneP.choices[0].delta.tool_calls[0].function.arguments, '{"filePath":"/p.txt"}');
+  assert.equal(doneQ.choices[0].delta.tool_calls[0].function.arguments, '{"filePath":"/q.txt"}');
+});
+
+test("Responses -> OpenAI: 2 parallel Agent calls still open at stream end each get their own flush chunk", () => {
+  const state = {};
+
+  openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      item: { type: "function_call", call_id: "call_agent0", name: "Agent" },
+    },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", output_index: 0, delta: '{"task":"a"}' },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    {
+      type: "response.output_item.added",
+      item: { type: "function_call", call_id: "call_agent1", name: "Agent" },
+    },
+    state
+  );
+  openaiResponsesToOpenAIResponse(
+    { type: "response.function_call_arguments.delta", output_index: 1, delta: '{"task":"b"}' },
+    state
+  );
+
+  // Stream ends (chunk === null) before either call's output_item.done arrives.
+  const flushed = openaiResponsesToOpenAIResponse(null, state);
+
+  assert.ok(Array.isArray(flushed));
+  const argChunks = flushed.filter((c) => c.choices[0].delta.tool_calls);
+  assert.deepEqual(
+    argChunks.map((c) => c.choices[0].delta.tool_calls[0].index).sort(),
+    [0, 1],
+    "each still-open parallel call must get its own flush chunk, at its own index"
+  );
+  const finalChunk = flushed[flushed.length - 1];
+  assert.equal(finalChunk.choices[0].finish_reason, "tool_calls");
+});
diff --git a/tests/unit/stream-payload-collector.test.ts b/tests/unit/stream-payload-collector.test.ts
index e8869bb8bf..63b96c5eaf 100644
--- a/tests/unit/stream-payload-collector.test.ts
+++ b/tests/unit/stream-payload-collector.test.ts
@@ -2,6 +2,7 @@ import test from "node:test";
 import assert from "node:assert/strict";
 
 const collector = await import("../../open-sse/utils/streamPayloadCollector.ts");
+import { splitConcatenatedToolCallArguments } from "../../open-sse/utils/streamPayloadCollector.ts";
 
 test("compactStructuredStreamPayload returns null for null input", () => {
   assert.equal(collector.compactStructuredStreamPayload(null), null);
@@ -413,3 +414,33 @@ test("#9315: getSummary() returns undefined when no format was configured (unaff
   c.push({ choices: [{ index: 0, delta: { content: "hi" } }] });
   assert.equal(c.getSummary(), undefined);
 });
+
+test("splitConcatenatedToolCallArguments — two back-to-back JSON objects", () => {
+  const a = JSON.stringify({ tool: "x", args: "1" });
+  const b = JSON.stringify({ tool: "y", args: "2" });
+  const out = splitConcatenatedToolCallArguments(a + b);
+  assert.deepEqual(out, [a, b]); // >=2 valid values -> split (array of parts)
+});
+
+test("splitConcatenatedToolCallArguments — nested object + escaped quotes stay single JSON", () => {
+  const a = JSON.stringify({ a: 'he said "hi"', b: { c: 1 } });
+  const single = a; // a is ONE valid JSON object -> no split
+  const out = splitConcatenatedToolCallArguments(single);
+  assert.equal(out, null); // single valid JSON -> untouched (null)
+});
+
+test("splitConcatenatedToolCallArguments — braces/quotes inside strings exercise escaped scanner", () => {
+  // Two valid JSON values whose string bodies contain braces and escaped quotes.
+  // Concatenated they reach the inString/escaped state machine (not the JSON.parse
+  // fast path), so this covers the case the owner asked about.
+  const a = JSON.stringify({ cmd: 'echo "}{" ; x' });
+  const b = JSON.stringify({ cmd: "{[not json]}" });
+  const out = splitConcatenatedToolCallArguments(a + b);
+  assert.deepEqual(out, [a, b]); // >=2 valid values -> split into parts
+});
+
+test("splitConcatenatedToolCallArguments — top-level array is single value", () => {
+  const arr = JSON.stringify([{ tool: "x" }, { tool: "y" }]);
+  const out = splitConcatenatedToolCallArguments(arr);
+  assert.equal(out, null); // one value boundary (array) -> not split
+});
diff --git a/tests/unit/stream-recovery-toolcall.test.ts b/tests/unit/stream-recovery-toolcall.test.ts
new file mode 100644
index 0000000000..7f3a765343
--- /dev/null
+++ b/tests/unit/stream-recovery-toolcall.test.ts
@@ -0,0 +1,178 @@
+import { describe, it } from "node:test";
+import assert from "node:assert/strict";
+import {
+  createRecoverableStream,
+  TruncatedStreamError,
+  scanOpenAiSseText,
+} from "../../open-sse/services/streamRecovery.ts";
+
+const enc = new TextEncoder();
+
+// Deliver the SSE chunk on the first read, then error on the second read so the
+// holdback window has committed (post-commit truncation) before the cut.
+function makeStream(sse: string): ReadableStream<Uint8Array> {
+  let n = 0;
+  return new ReadableStream<Uint8Array>({
+    pull(c) {
+      n += 1;
+      if (n === 1) {
+        c.enqueue(enc.encode(sse));
+        return;
+      }
+      c.error(new TruncatedStreamError());
+    },
+  });
+}
+
+// A clock that jumps past HOLDBACK_MS on the second read so the very first pushed
+// chunk commits the holdback window immediately (post-commit truncation path).
+function jumpingClock(): () => number {
+  let t = 0;
+  return () => (t += 1000);
+}
+
+describe("scanOpenAiSseText: terminal vs in-flight tool call", () => {
+  it("tool_calls without finish_reason → inFlight true, terminal false", () => {
+    const sse =
+      'data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_1","function":{"name":"lookup"}}]}}]}\n\n';
+    const r = scanOpenAiSseText(sse);
+    assert.equal(r.sawToolCall, true);
+    assert.equal(r.sawToolCallInFlight, true);
+    assert.equal(r.terminal, false);
+  });
+
+  it("complete tool_calls + finish_reason + [DONE] → terminal true, inFlight false", () => {
+    const sse =
+      'data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_1","function":{"name":"lookup","arguments":"{}"}}]}}]}\n' +
+      'data: {"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}\n' +
+      "data: [DONE]\n\n";
+    const r = scanOpenAiSseText(sse);
+    assert.equal(r.sawToolCall, true);
+    assert.equal(r.terminal, true);
+    assert.equal(r.sawToolCallInFlight, false);
+  });
+
+  it("plain text → no tool call", () => {
+    const sse = 'data: {"choices":[{"index":0,"delta":{"content":"hello"}}]}\n\n';
+    const r = scanOpenAiSseText(sse);
+    assert.equal(r.sawToolCall, false);
+    assert.equal(r.sawToolCallInFlight, false);
+    assert.equal(r.terminal, false);
+  });
+
+  it("complete tool_calls WITHOUT [DONE] → terminal false, inFlight false (the actual fix)", () => {
+    // This is the case the original plan promised to unblock: the tool call itself is
+    // done (finish_reason: "tool_calls"), but the overall stream/turn has not sent its
+    // own terminal marker yet — a truncation right here is recoverable.
+    const sse =
+      'data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_1","function":{"name":"lookup","arguments":"{}"}}]}}]}\n' +
+      'data: {"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}\n\n';
+    const r = scanOpenAiSseText(sse);
+    assert.equal(r.sawToolCall, true);
+    assert.equal(r.sawToolCallInFlight, false);
+    assert.equal(r.terminal, false);
+  });
+});
+
+describe("stream recovery does not duplicate an in-flight tool call", () => {
+  it("truncation with an in-flight tool call → no continuation", async () => {
+    let continued = false;
+    const sse =
+      'data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"c1","function":{"name":"f"}}]}}]}\n\n';
+    const wrapped = createRecoverableStream(makeStream(sse), async () => null, {
+      finalize: () => {},
+      now: jumpingClock(),
+      continueStream: async () => {
+        continued = true;
+        return null;
+      },
+    });
+    const reader = wrapped.getReader();
+    try {
+      for (;;) {
+        const r = await reader.read();
+        if (r.done) break;
+      }
+    } catch {
+      // the in-flight tool call makes the stream close without continuing
+    }
+    assert.equal(continued, false);
+  });
+
+  it("truncation right after a completed tool call → continuation attempted (the real 91% gain)", async () => {
+    // Text was emitted, THEN the tool call completed (finish_reason: "tool_calls"), THEN
+    // the connection drops before a [DONE]/other terminal marker. Before this fix, the
+    // blunt `emittedToolCall` guard blocked recovery here even though the call itself is
+    // done and only trailing prose was lost — this is the exact case the plan promised
+    // to unblock and the pre-fix table proved was a no-op.
+    let continued = false;
+    const sse =
+      'data: {"choices":[{"index":0,"delta":{"content":"Let me check that. "}}]}\n' +
+      'data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"c1","function":{"name":"f","arguments":"{}"}}]}}]}\n' +
+      'data: {"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}\n\n';
+    const wrapped = createRecoverableStream(makeStream(sse), async () => null, {
+      finalize: () => {},
+      now: jumpingClock(),
+      continueStream: async () => {
+        continued = true;
+        return null;
+      },
+    });
+    const reader = wrapped.getReader();
+    try {
+      for (;;) {
+        const r = await reader.read();
+        if (r.done) break;
+      }
+    } catch {
+      // no-op
+    }
+    assert.equal(continued, true);
+  });
+
+  it("truncation of plain text → continuation attempted", async () => {
+    let continued = false;
+    const sse = 'data: {"choices":[{"index":0,"delta":{"content":"hello "}}]}\n\n';
+    const wrapped = createRecoverableStream(makeStream(sse), async () => null, {
+      finalize: () => {},
+      now: jumpingClock(),
+      continueStream: async () => {
+        continued = true;
+        return null;
+      },
+    });
+    const reader = wrapped.getReader();
+    try {
+      for (;;) {
+        const r = await reader.read();
+        if (r.done) break;
+      }
+    } catch {
+      // no-op
+    }
+    assert.equal(continued, true);
+  });
+
+  it("naive removal of the tool-call guard would duplicate a partial tool call", () => {
+    // The blunt `sawToolCall` flag is true for BOTH a complete tool call and a
+    // partial (in-flight) one. The new `sawToolCallInFlight` flag is the only
+    // signal that tells them apart: a naive guard keyed on `sawToolCall` would
+    // block the complete call AND let the partial one through to the
+    // continuation, where trimContinuationOverlap (text-only) cannot de-duplicate
+    // the replayed tool_calls arguments.
+    const ssePartial =
+      'data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_1","function":{"name":"lookup","arguments":"{\\"q\\""}}]}}]}\n\n';
+    const sseFull =
+      'data: {"choices":[{"index":0,"delta":{"tool_calls":[{"id":"call_1","function":{"name":"lookup","arguments":"{\\"q\\":\\"x\\"}"}}]}}]}\n' +
+      'data: {"choices":[{"index":0,"delta":{},"finish_reason":"tool_calls"}]}\n\n';
+    const scanPartial = scanOpenAiSseText(ssePartial);
+    const scanFull = scanOpenAiSseText(sseFull);
+    // The blunt flag cannot distinguish them.
+    assert.equal(scanPartial.sawToolCall, true);
+    assert.equal(scanFull.sawToolCall, true);
+    // The in-flight flag can — and that is what keeps canContinue false only for
+    // the partial tool call, so the continuation never replays it.
+    assert.equal(scanPartial.sawToolCallInFlight, true);
+    assert.equal(scanFull.sawToolCallInFlight, false);
+  });
+});
diff --git a/tests/unit/strip-reasoning-blobs-agentic-context-1599.test.ts b/tests/unit/strip-reasoning-blobs-agentic-context-1599.test.ts
index cab5e9c37f..25eae7a839 100644
--- a/tests/unit/strip-reasoning-blobs-agentic-context-1599.test.ts
+++ b/tests/unit/strip-reasoning-blobs-agentic-context-1599.test.ts
@@ -274,7 +274,11 @@ test("explicit custom target opt-in remains an opaque transport override", () =>
   const result = applyReasoningInputPolicy(body, "responses", { preserveEncryptedReasoning: true });
 
   assert.equal(result.incompatibleReasoning, false);
-  assert.deepEqual(body.input, [{ type: "reasoning", encrypted_content: "encrypted-blob" }]);
+  // #11108: a kept opaque item defaults `summary` when the source omitted it —
+  // some upstreams reject `input[]` reasoning items missing the field entirely.
+  assert.deepEqual(body.input, [
+    { type: "reasoning", encrypted_content: "encrypted-blob", summary: [] },
+  ]);
 });
 
 test("preserved opaque reasoning remains redacted from log copies", () => {
diff --git a/tests/unit/translator-openai-responses-req.test.ts b/tests/unit/translator-openai-responses-req.test.ts
index e0d0a51773..da87b26116 100644
--- a/tests/unit/translator-openai-responses-req.test.ts
+++ b/tests/unit/translator-openai-responses-req.test.ts
@@ -475,6 +475,7 @@ test("Chat -> Responses defaults unannotated targets to plaintext reasoning", ()
     {
       type: "reasoning",
       content: [{ type: "reasoning_text", text: "Inspect the repository first" }],
+      summary: [],
     },
     {
       type: "function_call",
@@ -513,6 +514,7 @@ test("Chat -> DeepSeek Responses accepts the plaintext reasoning alias", () => {
   assert.deepEqual(result.input[0], {
     type: "reasoning",
     content: [{ type: "reasoning_text", text: "Alias plaintext reasoning" }],
+    summary: [],
   });
 });
 
diff --git a/tests/unit/webhook-discord-dispatcher.test.ts b/tests/unit/webhook-discord-dispatcher.test.ts
index 4a49c3ba2e..605f4aa86b 100644
--- a/tests/unit/webhook-discord-dispatcher.test.ts
+++ b/tests/unit/webhook-discord-dispatcher.test.ts
@@ -1,5 +1,6 @@
 import test from "node:test";
 import assert from "node:assert/strict";
+import { WEBHOOK_EVENT_VALUES } from "../../src/lib/webhooks/eventDescriptions.ts";
 
 const { buildDiscordPayload } = await import("../../src/lib/webhooks/integrations/discord.ts");
 
@@ -14,15 +15,7 @@ test("buildDiscordPayload — request.failed produces embed with model", () => {
 });
 
 test("buildDiscordPayload — all WEBHOOK_EVENTS return object with content or embeds", () => {
-  const events = [
-    "request.completed",
-    "request.failed",
-    "provider.error",
-    "provider.recovered",
-    "quota.exceeded",
-    "combo.switched",
-    "test.ping",
-  ] as const;
+  const events = WEBHOOK_EVENT_VALUES;
   for (const event of events) {
     const payload = buildDiscordPayload(event, {});
     assert.ok(
@@ -33,7 +26,7 @@ test("buildDiscordPayload — all WEBHOOK_EVENTS return object with content or e
 });
 
 test("buildDiscordPayload — embeds have title and color fields", () => {
-  const payload = buildDiscordPayload("provider.error", { provider: "openai" });
+  const payload = buildDiscordPayload("request.failed", { provider: "openai" });
   assert.ok(Array.isArray(payload.embeds) && payload.embeds.length > 0, "should have embeds");
   const embed = payload.embeds![0];
   assert.ok(typeof embed.title === "string" && embed.title.length > 0, "embed must have title");
diff --git a/tests/unit/webhook-slack-dispatcher.test.ts b/tests/unit/webhook-slack-dispatcher.test.ts
index 44b34626cf..ad91a774e3 100644
--- a/tests/unit/webhook-slack-dispatcher.test.ts
+++ b/tests/unit/webhook-slack-dispatcher.test.ts
@@ -1,5 +1,6 @@
 import test from "node:test";
 import assert from "node:assert/strict";
+import { WEBHOOK_EVENT_VALUES } from "../../src/lib/webhooks/eventDescriptions.ts";
 
 const { buildSlackPayload } = await import("../../src/lib/webhooks/integrations/slack.ts");
 
@@ -30,8 +31,8 @@ test("buildSlackPayload — test.ping produces a ping/test message", () => {
   );
 });
 
-test("buildSlackPayload — provider.error includes provider context", () => {
-  const payload = buildSlackPayload("provider.error", { provider: "openai", model: "gpt-4" });
+test("buildSlackPayload — request.failed includes provider context", () => {
+  const payload = buildSlackPayload("request.failed", { provider: "openai" });
   const combined = JSON.stringify(payload);
   assert.ok(
     combined.includes("Provider") ||
@@ -43,15 +44,7 @@ test("buildSlackPayload — provider.error includes provider context", () => {
 });
 
 test("buildSlackPayload — all WEBHOOK_EVENTS produce valid payloads with text field", () => {
-  const events = [
-    "request.completed",
-    "request.failed",
-    "provider.error",
-    "provider.recovered",
-    "quota.exceeded",
-    "combo.switched",
-    "test.ping",
-  ] as const;
+  const events = WEBHOOK_EVENT_VALUES;
   for (const event of events) {
     const payload = buildSlackPayload(event, {});
     assert.ok(
diff --git a/tests/unit/webhook-telegram-dispatcher.test.ts b/tests/unit/webhook-telegram-dispatcher.test.ts
index eeb5ee2c92..29bf2f586c 100644
--- a/tests/unit/webhook-telegram-dispatcher.test.ts
+++ b/tests/unit/webhook-telegram-dispatcher.test.ts
@@ -1,5 +1,6 @@
 import test from "node:test";
 import assert from "node:assert/strict";
+import { WEBHOOK_EVENT_VALUES } from "../../src/lib/webhooks/eventDescriptions.ts";
 
 const { buildTelegramPayload, buildTelegramUrl } =
   await import("../../src/lib/webhooks/integrations/telegram.ts");
@@ -77,15 +78,7 @@ test("buildTelegramPayload — chat_id matches provided value for groups", () =>
 });
 
 test("buildTelegramPayload — all WEBHOOK_EVENTS produce valid payloads with chat_id", () => {
-  const events = [
-    "request.completed",
-    "request.failed",
-    "provider.error",
-    "provider.recovered",
-    "quota.exceeded",
-    "combo.switched",
-    "test.ping",
-  ] as const;
+  const events = WEBHOOK_EVENT_VALUES;
   for (const event of events) {
     const payload = buildTelegramPayload(event, {}, "99999");
     assert.equal(payload.chat_id, "99999");
diff --git a/tests/unit/webhooks-ghost-events.test.ts b/tests/unit/webhooks-ghost-events.test.ts
index 427b0467a7..a5d26e6019 100644
--- a/tests/unit/webhooks-ghost-events.test.ts
+++ b/tests/unit/webhooks-ghost-events.test.ts
@@ -30,4 +30,16 @@ describe("webhook catalogue", () => {
     const { notifyWebhookEvent } = await import("../../src/lib/webhookDispatcher.ts");
     assert.equal(typeof notifyWebhookEvent, "function");
   });
+
+  it("every builder accepts every value in WEBHOOK_EVENT_VALUES without throwing", async () => {
+    const { buildDiscordPayload } = await import("../../src/lib/webhooks/integrations/discord.ts");
+    const { buildSlackPayload } = await import("../../src/lib/webhooks/integrations/slack.ts");
+    const { buildTelegramPayload } =
+      await import("../../src/lib/webhooks/integrations/telegram.ts");
+    for (const event of WEBHOOK_EVENT_VALUES) {
+      assert.doesNotThrow(() => buildDiscordPayload(event, {}));
+      assert.doesNotThrow(() => buildSlackPayload(event, {}));
+      assert.doesNotThrow(() => buildTelegramPayload(event, {}, "99999"));
+    }
+  });
 });