fix(core): recover partial provider failures

packages/core/src/session/runner/llm.ts

@@ -61,10 +61,13 @@ const isDecline = (
 ): error is Permission.DeclinedError | QuestionTool.CancelledError =>
   error._tag === "Permission.DeclinedError" || error._tag === "QuestionTool.CancelledError"
 
-const isInterruptedStream = (failure: AIError) => {
-  if (failure.reason._tag === "InvalidProviderOutput")
-    return failure.reason.classification === "incomplete-stream"
+const shouldContinueAfterFailure = (failure: AIError, record: StepRecord) => {
+  if (failure.reason._tag === "InvalidProviderOutput") return failure.reason.classification === "incomplete-stream"
   if (failure.reason._tag === "Transport") return failure.reason.operation === "read"
+  // A fresh generation can continue partial output, but not cross a tool boundary
+  // whose provider-side effects may be ambiguous.
+  if (failure.reason._tag === "ProviderInternal")
+    return SessionRunnerRetry.isRetryable(failure) && record.calls.length === 0
   return false
 }
 
@@ -120,7 +123,7 @@ const classifyToolExits = (
 const TOOLS_INTERRUPTED = { type: "aborted", message: "Tool execution interrupted" } as const
 const STEP_INTERRUPTED = { type: "aborted", message: "Step interrupted" } as const
 const RESULT_MISSING = { type: "tool.result-missing", message: "Provider did not return a tool result" } as const
-const CONTINUE_AFTER_INCOMPLETE_STREAM =
+const CONTINUE_AFTER_PARTIAL_FAILURE =
   "The previous response was interrupted. Continue from where you left off without repeating completed content."
 
 const layer = Layer.effect(
@@ -302,7 +305,7 @@ const layer = Layer.effect(
           ).pipe(Pull.catchDone(() => Effect.fail(outcome.cause)))
           yield* bus.publish(SessionEvent.Synthetic, {
             sessionID,
-            text: CONTINUE_AFTER_INCOMPLETE_STREAM,
+            text: CONTINUE_AFTER_PARTIAL_FAILURE,
           })
           assistantMessageID = SessionMessage.ID.create()
         }
@@ -597,7 +600,7 @@ const layer = Layer.effect(
           if (
             llmFailure &&
             llmError &&
-            isInterruptedStream(llmFailure) &&
+            shouldContinueAfterFailure(llmFailure, record) &&
             record.outputStarted &&
             tools.declines.length === 0 &&
             !tools.interrupted

packages/core/test/session-runner.test.ts

@@ -10,6 +10,7 @@ import {
   TransportReason,
   InvalidProviderOutputReason,
   InvalidRequestReason,
+  ProviderInternalReason,
   RateLimitReason,
 } from "@opencode-ai/ai"
 import * as OpenAIChat from "@opencode-ai/ai/protocols/openai-chat"
@@ -572,6 +573,13 @@ const streamDisconnected = () =>
     }),
   })
 
+const providerInternal = () =>
+  new AIError({
+    module: "test",
+    method: "stream",
+    reason: new ProviderInternalReason({ message: "The provider failed while generating a response" }),
+  })
+
 const continuationRejected = (recovery: "retry-full" | "rotate-and-retry-full") =>
   new AIError({
     module: "test",
@@ -596,7 +604,7 @@ const incompleteStream = () =>
     }),
   })
 
-const INCOMPLETE_STREAM_CONTINUATION =
+const PARTIAL_FAILURE_CONTINUATION =
   "The previous response was interrupted. Continue from where you left off without repeating completed content."
 
 const invalidRequest = () =>
@@ -4591,7 +4599,7 @@ describe("SessionRunnerLLM", () => {
         content: [
           {
             type: "text",
-            text: INCOMPLETE_STREAM_CONTINUATION,
+            text: PARTIAL_FAILURE_CONTINUATION,
           },
         ],
       })
@@ -4606,7 +4614,7 @@ describe("SessionRunnerLLM", () => {
         },
         {
           type: "synthetic",
-          text: INCOMPLETE_STREAM_CONTINUATION,
+          text: PARTIAL_FAILURE_CONTINUATION,
         },
         { type: "assistant", finish: "stop", content: [{ type: "text", text: " continuation" }] },
       ])
@@ -4641,12 +4649,12 @@ describe("SessionRunnerLLM", () => {
       expect(requests).toHaveLength(2)
       expect(requests[1]?.messages.at(-1)).toMatchObject({
         role: "user",
-        content: [{ type: "text", text: INCOMPLETE_STREAM_CONTINUATION }],
+        content: [{ type: "text", text: PARTIAL_FAILURE_CONTINUATION }],
       })
       expect(yield* session.context(sessionID)).toMatchObject([
         { type: "user" },
         { type: "assistant", finish: "error", content: [{ type: "text", text: "Partial" }] },
-        { type: "synthetic", text: INCOMPLETE_STREAM_CONTINUATION },
+        { type: "synthetic", text: PARTIAL_FAILURE_CONTINUATION },
         { type: "assistant", finish: "stop", content: [{ type: "text", text: " continuation" }] },
       ])
     }),
@@ -4680,7 +4688,7 @@ describe("SessionRunnerLLM", () => {
         content: [
           {
             type: "text",
-            text: INCOMPLETE_STREAM_CONTINUATION,
+            text: PARTIAL_FAILURE_CONTINUATION,
           },
         ],
       })
@@ -4720,7 +4728,7 @@ describe("SessionRunnerLLM", () => {
       expect(yield* recordedEventTypes(sessionID)).toContain("session.retry.scheduled.1")
       expect(requests[1]?.messages.slice(-2)).toMatchObject([
         { role: "user", content: [{ type: "text", text: "Recover disconnected reasoning" }] },
-        { role: "user", content: [{ type: "text", text: INCOMPLETE_STREAM_CONTINUATION }] },
+        { role: "user", content: [{ type: "text", text: PARTIAL_FAILURE_CONTINUATION }] },
       ])
       expect(yield* session.context(sessionID)).toMatchObject([
         { type: "user" },
@@ -4735,12 +4743,77 @@ describe("SessionRunnerLLM", () => {
             },
           ],
         },
-        { type: "synthetic", text: INCOMPLETE_STREAM_CONTINUATION },
+        { type: "synthetic", text: PARTIAL_FAILURE_CONTINUATION },
         { type: "assistant", finish: "stop", content: [{ type: "text", text: "Recovered" }] },
       ])
     }),
   )
 
+  it.effect("continues a provider failure after reasoning without tool activity", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      yield* admit(session, "Recover failed reasoning")
+      const failure = providerInternal()
+      yield* TestLLM.push(
+        TestLLM.failAfter(
+          failure,
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.reasoningStart({ id: "failed-reasoning" }),
+          LLMEvent.reasoningEnd({
+            id: "failed-reasoning",
+            providerMetadata: {
+              openai: { itemId: "rs_failed", reasoningEncryptedContent: "failed-encrypted-state" },
+            },
+          }),
+        ),
+      )
+      yield* TestLLM.push(TestLLM.text("Recovered", "provider-failure-recovery"))
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* TestLLM.wait(1)
+      yield* TestClock.adjust("2400 millis")
+      yield* Fiber.join(run)
+
+      expect(requests).toHaveLength(2)
+      expect(requests[1]?.messages).toHaveLength(2)
+      expect(requests[1]?.messages.slice(-2)).toMatchObject([
+        { role: "user", content: [{ type: "text", text: "Recover failed reasoning" }] },
+        { role: "user", content: [{ type: "text", text: PARTIAL_FAILURE_CONTINUATION }] },
+      ])
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user" },
+        {
+          type: "assistant",
+          finish: "error",
+          error: { type: "provider.internal" },
+          content: [{ type: "reasoning", text: "" }],
+        },
+        { type: "synthetic", text: PARTIAL_FAILURE_CONTINUATION },
+        { type: "assistant", finish: "stop", content: [{ type: "text", text: "Recovered" }] },
+      ])
+      expect(yield* recordedEventTypes(sessionID)).toContain("session.retry.scheduled.1")
+    }),
+  )
+
+  it.effect("does not continue a provider failure after tool activity", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      const failure = providerInternal()
+      yield* TestLLM.push(
+        TestLLM.failAfter(
+          failure,
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-before-provider-failure", name: "echo", input: { text: "settled" } }),
+        ),
+      )
+
+      expect(yield* runPrompt(session, "Do not repeat tool activity").pipe(Effect.flip)).toBe(failure)
+      expect(requests).toHaveLength(1)
+      expect(executions).toEqual(["settled"])
+      expect(yield* recordedEventTypes(sessionID)).not.toContain("session.retry.scheduled.1")
+    }),
+  )
+
   it.effect("continues an incomplete stream after settling a local tool", () =>
     Effect.gen(function* () {
       const session = yield* setup
@@ -4772,7 +4845,7 @@ describe("SessionRunnerLLM", () => {
           content: [
             {
               type: "text",
-              text: INCOMPLETE_STREAM_CONTINUATION,
+              text: PARTIAL_FAILURE_CONTINUATION,
             },
           ],
         },
@@ -4812,7 +4885,7 @@ describe("SessionRunnerLLM", () => {
             },
           ],
         },
-        { type: "synthetic", text: INCOMPLETE_STREAM_CONTINUATION },
+        { type: "synthetic", text: PARTIAL_FAILURE_CONTINUATION },
         { type: "assistant", finish: "stop", content: [{ type: "text", text: "Recovered" }] },
       ])
     }),