fix(plugin): constrain Anthropic prefill guard

Tighten the assistant-tail repair after review so only the Anthropic provider is treated as prefill-rejecting, and so assistant-tail model metadata is evaluated independently from the last user model.

Add regression coverage for an allowed user model followed by a rejecting Anthropic assistant tail, plus a non-Anthropic provider carrying a Claude-looking model id.

Plan: plans/fix-anthropic-assistant-prefill-tail.md

src/plugin/messages-transform.test.ts

@@ -286,6 +286,48 @@ describe("createMessagesTransformHandler", () => {
     })
   })
 
+  it("#given the assistant tail identifies a rejecting Anthropic model after an allowed user model #when messages transform runs #then it appends a synthetic user recovery turn", async () => {
+    //#given
+    const messages: TestMessage[] = [
+      {
+        info: {
+          id: "msg_user_allowed_then_rejecting_assistant",
+          role: "user",
+          sessionID: "ses_allowed_then_rejecting_assistant",
+          agent: "sisyphus",
+          model: { providerID: "openai", modelID: "gpt-5.4" },
+        },
+        parts: [{ type: "text", text: "continue" }],
+      },
+      {
+        info: {
+          id: "msg_assistant_rejecting_metadata",
+          role: "assistant",
+          sessionID: "ses_allowed_then_rejecting_assistant",
+          model: { providerID: "anthropic", modelID: "claude-opus-4-6" },
+        },
+        parts: [{ type: "text", text: "done" }],
+      },
+    ]
+
+    //#when
+    await runHandler(makeHooks({}), messages)
+
+    //#then
+    expect(messages).toHaveLength(3)
+    expect(messages.at(-1)?.info).toMatchObject({
+      role: "user",
+      sessionID: "ses_allowed_then_rejecting_assistant",
+      agent: "sisyphus",
+      model: { providerID: "openai", modelID: "gpt-5.4" },
+    })
+    expect(messages.at(-1)?.parts[0]).toMatchObject({
+      type: "text",
+      text: "[internal] Continue from the previous assistant state.",
+      synthetic: true,
+    })
+  })
+
   it("#given rejecting model metadata uses direct provider and model fields #when messages transform runs #then it appends a synthetic user recovery turn", async () => {
     //#given
     const messages: TestMessage[] = [
@@ -353,6 +395,13 @@ describe("createMessagesTransformHandler", () => {
         name: "missing model",
         userInfo: { role: "user" },
       },
+      {
+        name: "non-anthropic provider",
+        userInfo: {
+          role: "user",
+          model: { providerID: "opencode", modelID: "claude-opus-4-7" },
+        },
+      },
     ]
 
     for (const scenario of scenarios) {

src/plugin/messages-transform.ts

@@ -5,11 +5,7 @@ import { normalizeModelID } from "../shared/model-normalization"
 import type { CreatedHooks } from "../create-hooks"
 
 const ASSISTANT_PREFILL_RECOVERY_TEXT = "[internal] Continue from the previous assistant state."
-const ASSISTANT_PREFILL_UNSUPPORTED_PROVIDERS = new Set([
-  "anthropic",
-  "google-vertex-anthropic",
-  "opencode",
-])
+const ASSISTANT_PREFILL_UNSUPPORTED_PROVIDER = "anthropic"
 const ASSISTANT_PREFILL_UNSUPPORTED_MODEL_PREFIXES = [
   "claude-opus-4-7",
   "claude-opus-4-6",
@@ -98,7 +94,7 @@ function shouldRepairAssistantPrefillForModel(model: ModelIdentifier | undefined
   }
 
   const providerID = model.providerID.toLowerCase()
-  if (!ASSISTANT_PREFILL_UNSUPPORTED_PROVIDERS.has(providerID)) {
+  if (providerID !== ASSISTANT_PREFILL_UNSUPPORTED_PROVIDER) {
     return false
   }
 
@@ -162,7 +158,8 @@ function ensureUserTurnAfterAssistantTail(output: MessagesTransformOutput): void
   }
 
   const shouldRepairAssistantTail = hasInternalContinuationTrigger(output.messages) ||
-    shouldRepairAssistantPrefillForModel(findLastUserModel(output.messages) ?? readModelIdentifier(lastMessage.info))
+    shouldRepairAssistantPrefillForModel(findLastUserModel(output.messages)) ||
+    shouldRepairAssistantPrefillForModel(readModelIdentifier(lastMessage.info))
   if (!shouldRepairAssistantTail) {
     return
   }