auto-model-router 0.37.0 → 0.38.0

package/.omp-plugin/marketplace.json

@@ -7,14 +7,14 @@
   },
   "metadata": {
     "description": "auto-model-router: a local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter",
-    "version": "0.37.0",
+    "version": "0.38.0",
     "pluginRoot": "."
   },
   "plugins": [
     {
       "name": "auto-model-router",
       "description": "Local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter. Runs in-process, routes per turn by price and task complexity, with budget caps, mid-stream escalation, and cache-aware hysteresis.",
-      "version": "0.37.0",
+      "version": "0.38.0",
       "author": {
         "name": "drewappling",
         "email": "drewappling@gmail.com"

package/README.md

@@ -1071,11 +1071,19 @@ Each task (`coding`, `vision`, `documentation`, `data`, `chat`) is a
 
 The model that produced the rejected output never serves the retry, at this
 tier or the next. Signals that indict the *provider* rather than the tier —
-`empty_completion`, `refusal`, and an error finish — first try a different
-model in the **same** tier (bounded, like a 5xx failover) and only then step
-up; structural signals (`malformed_tool_args`, `repeat_tool_call`, a truncated
-tool call) escalate a tier directly. A client that hangs up after the finish
-event has already arrived is treated as a completed turn, not an error.
+`empty_completion`, `refusal`, `content_filter`, and an error finish — first
+try a different model in the **same** tier (bounded, like a 5xx failover) and
+only then step up; structural signals (`malformed_tool_args`,
+`repeat_tool_call`, a truncated tool call) escalate a tier directly. A client
+that hangs up after the finish event has already arrived is treated as a
+completed turn, not an error.
+
+A `content_filter` finish (the provider's filter stopping the completion,
+including Anthropic's `refusal` stop reason, which maps to it) rides the
+`refusal` trigger: the same failure class — the provider declining the work —
+so a config that already opts into refusal escalation gets it without a
+change. A filter finish that produced NO text is an `empty_completion`
+instead: nothing was filtered, so a plain retry is the fix.
 
 ### `hysteresis` — cache-aware model stickiness
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "auto-model-router",
-  "version": "0.37.0",
+  "version": "0.38.0",
   "private": false,
   "description": "Local cost/complexity-aware model router for Oh My Pi, backed by OpenRouter",
   "type": "module",

package/src/router/escalate.ts

@@ -165,6 +165,17 @@ export function createProbe(
 		if (matchesRefusal() && triggers.has("refusal")) {
 			return escalate("refusal", "completion opens with a refusal");
 		}
+		if (finishReason === "content_filter") {
+			// The PROVIDER stopped the completion, not the model. Whatever text
+			// arrived ends mid-answer, and a different model — one whose filter
+			// does not read this content the same way — may serve the turn whole.
+			// It rides the `refusal` trigger: the same failure class, the provider
+			// declining the work rather than the model. A filter finish with NO
+			// text was never generated, so nothing was filtered; that is the
+			// empty_completion case above, and a plain retry is the fix.
+			if (triggers.has("refusal")) return escalate("content_filter", "the provider's content filter stopped the completion");
+			return commit("content_filter finish tolerated; signal disabled");
+		}
 		if (finishReason === "error" && triggers.has("upstream_error")) {
 			return escalate("upstream_error", "upstream reported an error finish");
 		}

package/src/router/types.ts

@@ -329,7 +329,8 @@ export type EscalationSignal =
 	| "repeat_tool_call"
 	| "length_stop"
 	| "missing_expected_tool_call"
-	| "upstream_error";
+	| "upstream_error"
+	| "content_filter";
 
 export type ProbeVerdict =
 	| { action: "commit"; reason: string }

package/src/server/turn.ts

@@ -54,7 +54,7 @@ const MAX_SAME_TIER_FAILOVERS = 2;
  * signals (malformed arguments, a repeated call) still escalate directly — a
  * stronger model is the remedy there.
  */
-const PROVIDER_SIGNALS: ReadonlySet<string> = new Set(["empty_completion", "refusal", "upstream_error"]);
+const PROVIDER_SIGNALS: ReadonlySet<string> = new Set(["empty_completion", "refusal", "upstream_error", "content_filter"]);
 
 /** A client hang-up: the request signal fired, or the transport reported the abort. */
 function isClientAbort(err: unknown, signal: AbortSignal): boolean {

package/test/escalate.test.ts

@@ -226,4 +226,29 @@ describe("createProbe", () => {
 		expect(verdict.action).toBe("escalate");
 		if (verdict.action === "escalate") expect(verdict.signal).toBe("empty_completion");
 	});
+
+	test("a content_filter finish with text escalates on the refusal trigger", async () => {
+		// The PROVIDER stopped the completion mid-answer; another model may serve
+		// the turn whole, so this is a provider signal (same-tier failover).
+		const p = createProbe(plan({ maxTokens: 1_000 }), req(), ALL_TRIGGERS);
+		expect(p.observe(text("The first part of the answer is"))).toBeNull();
+		const verdict = p.observe(chunk([{ type: "finish", reason: "content_filter" }]));
+		expect(verdict?.action).toBe("escalate");
+		if (verdict?.action === "escalate") expect(verdict.signal).toBe("content_filter");
+	});
+
+	test("a content_filter finish with NO text is an empty completion: nothing was filtered", async () => {
+		const p = createProbe(plan({ maxTokens: 1_000 }), req(), ALL_TRIGGERS);
+		const verdict = p.observe(chunk([{ type: "finish", reason: "content_filter" }]));
+		expect(verdict?.action).toBe("escalate");
+		if (verdict?.action === "escalate") expect(verdict.signal).toBe("empty_completion");
+	});
+
+	test("a content_filter finish commits when the refusal trigger is off", async () => {
+		const triggers = new Set([...ALL_TRIGGERS].filter((t) => t !== "refusal"));
+		const p = createProbe(plan({ maxTokens: 1_000 }), req(), triggers);
+		expect(p.observe(text("The first part of the answer is"))).toBeNull();
+		const verdict = p.observe(chunk([{ type: "finish", reason: "content_filter" }]));
+		expect(verdict?.action).toBe("commit");
+	});
 });