fix(core): safely recover malformed tool input (#37698)

packages/ai/src/protocols/anthropic-messages.ts

@@ -703,7 +703,14 @@ const onContentBlockStart = (state: ParserState, event: AnthropicEvent): StepRes
           providerExecuted: block.type === "server_tool_use",
         }),
       },
-      [...events, LLMEvent.toolInputStart({ id: block.id ?? String(event.index), name: block.name ?? "" })],
+      [
+        ...events,
+        LLMEvent.toolInputStart({
+          id: block.id ?? String(event.index),
+          name: block.name ?? "",
+          providerExecuted: block.type === "server_tool_use" ? true : undefined,
+        }),
+      ],
     ]
   }
 

packages/ai/src/protocols/bedrock-converse.ts

@@ -561,7 +561,9 @@ const step = (state: ParserState, event: BedrockEvent) =>
       return [
         {
           ...state,
-          hasToolCalls: resultEvents.some(LLMEvent.is.toolCall) ? true : state.hasToolCalls,
+          hasToolCalls:
+            resultEvents.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
+            state.hasToolCalls,
           lifecycle,
           tools: result.tools,
           reasoningSignatures: Object.fromEntries(

packages/ai/src/protocols/openai-chat.ts

@@ -464,7 +464,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
     }
 
     // Finalize accumulated tool inputs eagerly when finish_reason arrives so
-    // JSON parse failures fail the stream at the boundary rather than at halt.
+    // valid calls and malformed local calls settle independently.
     const finished =
       finishReason !== undefined && state.finishReason === undefined && Object.keys(tools).length > 0
         ? yield* ToolStream.finishAll(ADAPTER, tools)

packages/ai/src/protocols/openai-responses.ts

@@ -835,7 +835,9 @@ const onOutputItemDone = Effect.fn("OpenAIResponses.onOutputItemDone")(function*
       {
         ...state,
         lifecycle,
-        hasFunctionCall: resultEvents.some(LLMEvent.is.toolCall) ? true : state.hasFunctionCall,
+        hasFunctionCall:
+          resultEvents.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
+          state.hasFunctionCall,
         tools: result.tools,
       },
       events,

packages/ai/src/protocols/utils/tool-stream.ts

@@ -1,5 +1,5 @@
 import { Effect } from "effect"
-import { LLMError, LLMEvent, type ProviderMetadata, type ToolCall } from "../../schema"
+import { LLMError, LLMEvent, type ProviderMetadata, type ToolCall, type ToolInputError } from "../../schema"
 import { eventError, parseToolInput, type ToolAccumulator } from "../shared"
 
 type StreamKey = string | number
@@ -53,6 +53,7 @@ const inputStart = (tool: PendingTool) =>
   LLMEvent.toolInputStart({
     id: tool.id,
     name: tool.name,
+    providerExecuted: tool.providerExecuted ? true : undefined,
     providerMetadata: tool.providerMetadata,
   })
 
@@ -63,19 +64,38 @@ const inputDelta = (tool: PendingTool, text: string) =>
     text,
   })
 
-const toolCall = (route: string, tool: PendingTool, inputOverride?: string) =>
-  parseToolInput(route, tool.name, inputOverride ?? tool.input).pipe(
-    Effect.map(
-      (input): ToolCall =>
-        LLMEvent.toolCall({
-          id: tool.id,
-          name: tool.name,
-          input,
-          providerExecuted: tool.providerExecuted ? true : undefined,
-          providerMetadata: tool.providerMetadata,
-        }),
+const toolCall = (route: string, tool: PendingTool, inputOverride?: string) => {
+  const raw = inputOverride ?? tool.input
+  return parseToolInput(route, tool.name, raw).pipe(
+    Effect.map((input): ToolCall | ToolInputError =>
+      LLMEvent.toolCall({
+        id: tool.id,
+        name: tool.name,
+        input,
+        providerExecuted: tool.providerExecuted ? true : undefined,
+        providerMetadata: tool.providerMetadata,
+      }),
+    ),
+    Effect.catch((error) =>
+      tool.providerExecuted
+        ? Effect.fail(error)
+        : Effect.succeed(
+            LLMEvent.toolInputError({
+              id: tool.id,
+              name: tool.name,
+              raw,
+              message: error.reason.message,
+              providerMetadata: tool.providerMetadata,
+            }),
+          ),
     ),
   )
+}
+
+const finishEvents = (tool: PendingTool, event: ToolCall | ToolInputError): ReadonlyArray<LLMEvent> =>
+  event.type === "tool-input-error"
+    ? [event]
+    : [LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }), event]
 
 /** Store the updated tool and produce the optional public delta event. */
 const appendTool = <K extends StreamKey>(
@@ -158,8 +178,9 @@ export const appendExisting = <K extends StreamKey>(
 
 /**
  * Finalize one pending tool call: parse the accumulated raw JSON, remove it
- * from state, and return the optional public `tool-call` event. Missing keys are
- * a no-op because some providers emit stop events for non-tool content blocks.
+ * from state, and return either a call or a non-executable local input error.
+ * Missing keys are a no-op because some providers emit stop events for
+ * non-tool content blocks.
  */
 export const finish = <K extends StreamKey>(route: string, tools: State<K>, key: K) =>
   Effect.gen(function* () {
@@ -167,10 +188,7 @@ export const finish = <K extends StreamKey>(route: string, tools: State<K>, key:
     if (!tool) return { tools }
     return {
       tools: withoutTool(tools, key),
-      events: [
-        LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }),
-        yield* toolCall(route, tool),
-      ],
+      events: finishEvents(tool, yield* toolCall(route, tool)),
     }
   })
 
@@ -185,17 +203,14 @@ export const finishWithInput = <K extends StreamKey>(route: string, tools: State
     if (!tool) return { tools }
     return {
       tools: withoutTool(tools, key),
-      events: [
-        LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }),
-        yield* toolCall(route, tool, input),
-      ],
+      events: finishEvents(tool, yield* toolCall(route, tool, input)),
     }
   })
 
 /**
  * Finalize every pending tool call at once. OpenAI Chat has this shape: it does
- * not emit per-tool stop events, so all accumulated calls finish when the choice
- * receives a terminal `finish_reason`.
+ * not emit per-tool stop events, so all accumulated calls finish independently
+ * when the choice receives a terminal `finish_reason`.
  */
 export const finishAll = <K extends StreamKey>(route: string, tools: State<K>) =>
   Effect.gen(function* () {
@@ -205,12 +220,7 @@ export const finishAll = <K extends StreamKey>(route: string, tools: State<K>) =
     return {
       tools: empty<K>(),
       events: yield* Effect.forEach(pending, (tool) =>
-        toolCall(route, tool).pipe(
-          Effect.map((call) => [
-            LLMEvent.toolInputEnd({ id: tool.id, name: tool.name, providerMetadata: tool.providerMetadata }),
-            call,
-          ]),
-        ),
+        toolCall(route, tool).pipe(Effect.map((event) => finishEvents(tool, event))),
       ).pipe(Effect.map((events) => events.flat())),
     }
   })

packages/ai/src/schema/events.ts

@@ -129,6 +129,7 @@ export const ToolInputStart = Schema.Struct({
   type: Schema.tag("tool-input-start"),
   id: ToolCallID,
   name: Schema.String,
+  providerExecuted: Schema.optional(Schema.Boolean),
   providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.ToolInputStart" })
 export type ToolInputStart = Schema.Schema.Type<typeof ToolInputStart>
@@ -149,6 +150,17 @@ export const ToolInputEnd = Schema.Struct({
 }).annotate({ identifier: "LLM.Event.ToolInputEnd" })
 export type ToolInputEnd = Schema.Schema.Type<typeof ToolInputEnd>
 
+/** A local tool call that could not be decoded. `raw` is diagnostic-only. */
+export const ToolInputError = Schema.Struct({
+  type: Schema.tag("tool-input-error"),
+  id: ToolCallID,
+  name: Schema.String,
+  raw: Schema.String,
+  message: Schema.String,
+  providerMetadata: Schema.optional(ProviderMetadata),
+}).annotate({ identifier: "LLM.Event.ToolInputError" })
+export type ToolInputError = Schema.Schema.Type<typeof ToolInputError>
+
 export const ToolCall = Schema.Struct({
   type: Schema.tag("tool-call"),
   id: ToolCallID,
@@ -216,6 +228,7 @@ const llmEventTagged = Schema.Union([
   ToolInputStart,
   ToolInputDelta,
   ToolInputEnd,
+  ToolInputError,
   ToolCall,
   ToolResult,
   ToolError,
@@ -253,6 +266,8 @@ export const LLMEvent = Object.assign(llmEventTagged, {
   toolInputDelta: (input: WithID<ToolInputDelta, ToolCallID>) =>
     ToolInputDelta.make({ ...input, id: toolCallID(input.id) }),
   toolInputEnd: (input: WithID<ToolInputEnd, ToolCallID>) => ToolInputEnd.make({ ...input, id: toolCallID(input.id) }),
+  toolInputError: (input: WithID<ToolInputError, ToolCallID>) =>
+    ToolInputError.make({ ...input, id: toolCallID(input.id) }),
   toolCall: (input: WithID<ToolCall, ToolCallID>) => ToolCall.make({ ...input, id: toolCallID(input.id) }),
   toolResult: (input: WithID<ToolResult, ToolCallID>) =>
     ToolResult.make({
@@ -283,6 +298,7 @@ export const LLMEvent = Object.assign(llmEventTagged, {
     toolInputStart: llmEventTagged.guards["tool-input-start"],
     toolInputDelta: llmEventTagged.guards["tool-input-delta"],
     toolInputEnd: llmEventTagged.guards["tool-input-end"],
+    toolInputError: llmEventTagged.guards["tool-input-error"],
     toolCall: llmEventTagged.guards["tool-call"],
     toolResult: llmEventTagged.guards["tool-result"],
     toolError: llmEventTagged.guards["tool-error"],
@@ -548,6 +564,10 @@ const reduceResponseState = (state: ResponseState, event: LLMEvent): ResponseSta
       return reduceToolInputDelta(next, event)
     case "tool-input-end":
       return reduceToolInputEnd(next, event)
+    case "tool-input-error": {
+      const { [event.id]: _finished, ...toolInputs } = next.toolInputs
+      return { ...next, toolInputs }
+    }
     case "tool-call":
       return reduceToolCall(next, event)
     case "tool-result":

packages/ai/test/provider/anthropic-messages.test.ts

@@ -484,6 +484,30 @@ describe("Anthropic Messages route", () => {
     }),
   )
 
+  it.effect("keeps malformed server tool input terminal", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        { type: "message_start", message: { usage: { input_tokens: 5 } } },
+        {
+          type: "content_block_start",
+          index: 0,
+          content_block: { type: "server_tool_use", id: "call_1", name: "web_search" },
+        },
+        {
+          type: "content_block_delta",
+          index: 0,
+          delta: { type: "input_json_delta", partial_json: '{"query":"partial' },
+        },
+        { type: "content_block_stop", index: 0 },
+      )
+
+      const error = yield* LLMClient.generate(request).pipe(Effect.provide(fixedResponse(body)), Effect.flip)
+
+      expect(error).toBeInstanceOf(LLMError)
+      expect(error.message).toContain("Invalid JSON input for anthropic-messages tool call web_search")
+    }),
+  )
+
   it.effect("fails with a typed provider error for stream error frames", () =>
     Effect.gen(function* () {
       const error = yield* LLMClient.generate(request).pipe(

packages/ai/test/provider/bedrock-converse.test.ts

@@ -303,6 +303,32 @@ describe("Bedrock Converse route", () => {
     }),
   )
 
+  it.effect("emits malformed tool input as an unexecuted tool error", () =>
+    Effect.gen(function* () {
+      const body = eventStreamBody(
+        ["messageStart", { role: "assistant" }],
+        [
+          "contentBlockStart",
+          {
+            contentBlockIndex: 0,
+            start: { toolUse: { toolUseId: "tool_1", name: "lookup" } },
+          },
+        ],
+        ["contentBlockDelta", { contentBlockIndex: 0, delta: { toolUse: { input: '{"query":"partial' } } }],
+        ["contentBlockStop", { contentBlockIndex: 0 }],
+        ["messageStop", { stopReason: "end_turn" }],
+      )
+      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))
+
+      expect(response.events.find((event) => event.type === "tool-input-error")).toMatchObject({
+        id: "tool_1",
+        name: "lookup",
+        raw: '{"query":"partial',
+      })
+      expect(response.finishReason).toBe("tool-calls")
+    }),
+  )
+
   it.effect("decodes reasoning deltas", () =>
     Effect.gen(function* () {
       const body = eventStreamBody(

packages/ai/test/provider/openai-responses.test.ts

@@ -1,7 +1,7 @@
 import { describe, expect } from "bun:test"
 import { ConfigProvider, Effect, Layer, Stream } from "effect"
 import { Headers, HttpClientRequest } from "effect/unstable/http"
-import { LLM, LLMError, Message, Model, ToolCallPart, Usage } from "../../src"
+import { LLM, LLMError, LLMEvent, Message, Model, ToolCallPart, Usage } from "../../src"
 import { Auth, LLMClient, RequestExecutor, WebSocketExecutor } from "../../src/route"
 import * as Azure from "../../src/providers/azure"
 import * as OpenAI from "../../src/providers/openai"
@@ -1259,6 +1259,71 @@ describe("OpenAI Responses route", () => {
     }),
   )
 
+  it.effect("emits malformed final function arguments as an unexecuted tool error", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        {
+          type: "response.output_item.added",
+          item: { type: "function_call", id: "item_1", call_id: "call_1", name: "lookup", arguments: "" },
+        },
+        { type: "response.function_call_arguments.delta", item_id: "item_1", delta: '{"query":"streamed"}' },
+        {
+          type: "response.output_item.done",
+          item: {
+            type: "function_call",
+            id: "item_1",
+            call_id: "call_1",
+            name: "lookup",
+            arguments: '{"query":"partial',
+          },
+        },
+        { type: "response.completed", response: { usage: { input_tokens: 5, output_tokens: 1 } } },
+      )
+      const response = yield* LLMClient.generate(
+        LLM.updateRequest(request, {
+          tools: [{ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } }],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events.find(LLMEvent.is.toolInputError)).toEqual({
+        type: "tool-input-error",
+        id: "call_1",
+        name: "lookup",
+        raw: '{"query":"partial',
+        message: "Invalid JSON input for openai-responses tool call lookup",
+        providerMetadata: { openai: { itemId: "item_1" } },
+      })
+      expect(response.finishReason).toBe("tool-calls")
+      expect(response.events.some(LLMEvent.is.toolCall)).toBeFalse()
+    }),
+  )
+
+  it.effect("settles malformed function arguments when output_item.added is absent", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        {
+          type: "response.output_item.done",
+          item: {
+            type: "function_call",
+            id: "item_1",
+            call_id: "call_1",
+            name: "lookup",
+            arguments: '{"query":"partial',
+          },
+        },
+        { type: "response.completed", response: { usage: { input_tokens: 5, output_tokens: 1 } } },
+      )
+      const response = yield* LLMClient.generate(request).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events.find(LLMEvent.is.toolInputError)).toMatchObject({
+        id: "call_1",
+        name: "lookup",
+        raw: '{"query":"partial',
+      })
+      expect(response.finishReason).toBe("tool-calls")
+    }),
+  )
+
   it.effect("decodes web_search_call as provider-executed tool-call + tool-result", () =>
     Effect.gen(function* () {
       const item = {

packages/ai/test/response.test.ts

@@ -95,4 +95,20 @@ describe("LLMResponse reducer", () => {
       { type: "tool-call", id: "call_1", name: "lookup", input: { query: "weather" } },
     ])
   })
+
+  test("clears malformed tool input without appending an executable call", () => {
+    const state = reduce([
+      LLMEvent.toolInputStart({ id: "call_1", name: "lookup" }),
+      LLMEvent.toolInputDelta({ id: "call_1", name: "lookup", text: '{"query":"partial' }),
+      LLMEvent.toolInputError({
+        id: "call_1",
+        name: "lookup",
+        raw: '{"query":"partial',
+        message: "Invalid JSON input",
+      }),
+    ])
+
+    expect(state.toolInputs).toEqual({})
+    expect(state.message.content).toEqual([])
+  })
 })

packages/ai/test/tool-stream.test.ts

@@ -64,6 +64,75 @@ describe("ToolStream", () => {
     }),
   )
 
+  it.effect("finalizes malformed local input as a non-executable tool error", () =>
+    Effect.gen(function* () {
+      const tools = ToolStream.start(ToolStream.empty<string>(), "item_1", {
+        id: "call_1",
+        name: "lookup",
+        input: '{"query":"partial',
+      })
+      const finished = yield* ToolStream.finish(ADAPTER, tools, "item_1")
+
+      expect(finished).toEqual({
+        tools: {},
+        events: [
+          {
+            type: "tool-input-error",
+            id: "call_1",
+            name: "lookup",
+            raw: '{"query":"partial',
+            message: "Invalid JSON input for test-route tool call lookup",
+          },
+        ],
+      })
+    }),
+  )
+
+  it.effect("preserves valid siblings when one parallel input is malformed", () =>
+    Effect.gen(function* () {
+      const valid = ToolStream.start(ToolStream.empty<number>(), 0, {
+        id: "call_valid",
+        name: "lookup",
+        input: '{"query":"weather"}',
+      })
+      const tools = ToolStream.start(valid, 1, {
+        id: "call_invalid",
+        name: "lookup",
+        input: '{"query":"partial',
+      })
+      const finished = yield* ToolStream.finishAll(ADAPTER, tools)
+
+      expect(finished).toEqual({
+        tools: {},
+        events: [
+          { type: "tool-input-end", id: "call_valid", name: "lookup" },
+          { type: "tool-call", id: "call_valid", name: "lookup", input: { query: "weather" } },
+          {
+            type: "tool-input-error",
+            id: "call_invalid",
+            name: "lookup",
+            raw: '{"query":"partial',
+            message: "Invalid JSON input for test-route tool call lookup",
+          },
+        ],
+      })
+    }),
+  )
+
+  it.effect("keeps malformed provider-executed input terminal", () =>
+    Effect.gen(function* () {
+      const tools = ToolStream.start(ToolStream.empty<string>(), "item_1", {
+        id: "call_1",
+        name: "web_search",
+        input: '{"query":"partial',
+        providerExecuted: true,
+      })
+      const result = yield* Effect.exit(ToolStream.finish(ADAPTER, tools, "item_1"))
+
+      expect(result._tag).toBe("Failure")
+    }),
+  )
+
   it.effect("preserves providerExecuted and clears all tools", () =>
     Effect.gen(function* () {
       const first: ToolStream.State<number> = ToolStream.start(ToolStream.empty<number>(), 0, {

packages/client/src/promise/generated/types.ts

@@ -1276,6 +1276,8 @@ export type SessionStepFailed = {
     error: SessionStructuredError
     cost?: MoneyUSD
     tokens?: TokenUsageInfo
+    snapshot?: string
+    files?: Array<string>
   }
 }
 

packages/core/src/aisdk.ts

@@ -30,6 +30,7 @@ import {
   type UsageInput,
 } from "@opencode-ai/ai"
 import { Auth, Endpoint, type AnyRoute } from "@opencode-ai/ai/route"
+import { ProviderShared } from "@opencode-ai/ai/protocols/shared"
 import { Cause, Context, Effect, Layer, Option, Schema, Scope, Stream } from "effect"
 import { ModelV2 } from "./model"
 import { ProviderV2 } from "./provider"
@@ -605,6 +606,7 @@ function streamPartEvents(
         LLMEvent.toolInputStart({
           id: event.id,
           name: event.toolName,
+          providerExecuted: event.providerExecuted,
           providerMetadata: providerMetadata(event.providerMetadata),
         }),
       ])
@@ -622,15 +624,30 @@ function streamPartEvents(
       ])
     case "tool-call":
       state.toolNames[event.toolCallId] = event.toolName
-      return Effect.succeed([
-        LLMEvent.toolCall({
-          id: event.toolCallId,
-          name: event.toolName,
-          input: parseToolInput(event.input),
-          providerExecuted: event.providerExecuted,
-          providerMetadata: providerMetadata(event.providerMetadata),
-        }),
-      ])
+      return ProviderShared.parseToolInput("aisdk", event.toolName, event.input).pipe(
+        Effect.map((input) => [
+          LLMEvent.toolCall({
+            id: event.toolCallId,
+            name: event.toolName,
+            input,
+            providerExecuted: event.providerExecuted,
+            providerMetadata: providerMetadata(event.providerMetadata),
+          }),
+        ]),
+        Effect.catch((error) =>
+          event.providerExecuted
+            ? Effect.fail(error)
+            : Effect.succeed([
+                LLMEvent.toolInputError({
+                  id: event.toolCallId,
+                  name: event.toolName,
+                  raw: event.input,
+                  message: error.reason.message,
+                  providerMetadata: providerMetadata(event.providerMetadata),
+                }),
+              ]),
+        ),
+      )
     case "tool-result":
       delete state.toolNames[event.toolCallId]
       return Effect.succeed([
@@ -685,14 +702,6 @@ function providerMetadata(value: unknown) {
   return Schema.is(ProviderMetadata)(value) ? value : undefined
 }
 
-function parseToolInput(value: string) {
-  try {
-    return JSON.parse(value) as unknown
-  } catch {
-    return value
-  }
-}
-
 function jsonObject(input: Record<string, unknown>) {
   return Object.fromEntries(Object.entries(input).map(([key, value]) => [key, jsonValue(value)]))
 }

packages/core/src/session/message-updater.ts

@@ -297,6 +297,12 @@ export function update(adapter: Adapter, event: SessionEvent.Event) {
             draft.cost = event.data.cost
             draft.tokens = castDraft(event.data.tokens)
           }
+          if (event.data.snapshot || event.data.files)
+            draft.snapshot = {
+              ...draft.snapshot,
+              end: event.data.snapshot,
+              files: event.data.files ? Array.from(event.data.files) : undefined,
+            }
         })
       },
       "session.text.started": (event) => {

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

@@ -80,6 +80,7 @@ const layer = Layer.effect(
       sessionID: SessionSchema.ID,
       promotion: SessionPending.Delivery | undefined,
       step: number,
+      recoverMalformedToolInput: boolean,
       recoverOverflow?: typeof compaction.compact,
       assistantMessageID?: SessionMessage.ID,
     ) {
@@ -195,25 +196,27 @@ const layer = Layer.effect(
         tokens: settlement.tokens,
       })
 
-      // Captures the end snapshot, diffs it against the step's start, and durably ends the
-      // assistant step.
+      const captureStepEnd = Effect.fnUntraced(function* () {
+        const snapshot = yield* snapshots.capture()
+        const files =
+          startSnapshot && snapshot
+            ? yield* snapshots
+                .files({ from: startSnapshot, to: snapshot })
+                .pipe(Effect.catch(() => Effect.succeed(undefined)))
+            : undefined
+        return { snapshot, files }
+      })
+
       const publishStepEnd = (settlement: NonNullable<ReturnType<typeof publisher.stepSettlement>>) =>
         Effect.gen(function* () {
-          const endSnapshot = yield* snapshots.capture()
-          const files =
-            startSnapshot && endSnapshot
-              ? yield* snapshots
-                  .files({ from: startSnapshot, to: endSnapshot })
-                  .pipe(Effect.catch(() => Effect.succeed(undefined)))
-              : undefined
+          const end = yield* captureStepEnd()
           yield* serialized(
             events.publish(SessionEvent.Step.Ended, {
               sessionID: session.id,
               assistantMessageID: yield* publisher.startAssistant(),
               finish: settlement.finish,
               ...stepUsage(settlement),
-              snapshot: endSnapshot,
-              files,
+              ...end,
             }),
           )
         })
@@ -253,7 +256,7 @@ const layer = Layer.effect(
                 step: currentStep,
               })
             }
-            yield* serialized(publisher.failAssistant(error))
+            yield* serialized(publisher.failAssistant(error, true))
           }
           // Provider error events only arrive from the stream, so the flag is final here.
           const providerFailed = publisher.hasProviderError()
@@ -274,7 +277,7 @@ const layer = Layer.effect(
           if (settled._tag === "Failure") yield* FiberSet.clear(toolFibers)
           if (userDeclined || streamInterrupted || toolsInterrupted) {
             yield* serialized(publisher.failUnsettledTools({ type: "aborted", message: "Tool execution interrupted" }))
-            yield* serialized(publisher.failAssistant({ type: "aborted", message: "Step interrupted" }))
+            yield* serialized(publisher.failAssistant({ type: "aborted", message: "Step interrupted" }, true))
           }
           // A settled tool fiber failure is one of two things. A defect from a tool
           // implementation becomes a failed tool call the model can read, and the step still
@@ -287,7 +290,7 @@ const layer = Layer.effect(
             const failure = infraError ?? Cause.squash(settledFailure)
             const error = toSessionError(failure)
             yield* serialized(publisher.failUnsettledTools(error))
-            if (infraError !== undefined) yield* serialized(publisher.failAssistant(error))
+            if (infraError !== undefined) yield* serialized(publisher.failAssistant(error, true))
           }
 
           // Fail unresolved calls before the terminal step event. Local calls have joined, so
@@ -315,27 +318,49 @@ const layer = Layer.effect(
               : false
           if (hostedResultMissing && !publisher.stepSettlement())
             yield* serialized(
-              publisher.failAssistant({
-                type: "tool.result-missing",
-                message: "Provider did not return a tool result",
-              }),
+              publisher.failAssistant(
+                {
+                  type: "tool.result-missing",
+                  message: "Provider did not return a tool result",
+                },
+                true,
+              ),
             )
 
           const stepFailure = publisher.stepFailure()
           const stepSettlement = publisher.stepSettlement()
           if (stepSettlement && !stepFailure) yield* publishStepEnd(stepSettlement)
-          if (stepFailure)
-            yield* serialized(publisher.publishStepFailure(stepSettlement ? stepUsage(stepSettlement) : undefined))
+          if (stepFailure) {
+            const end = yield* captureStepEnd()
+            yield* serialized(
+              publisher.publishStepFailure({
+                ...(stepSettlement ? stepUsage(stepSettlement) : {}),
+                ...end,
+              }),
+            )
+          }
+
+          const recoveredMalformedToolInput =
+            recoverMalformedToolInput &&
+            publisher.hasMalformedToolInput() &&
+            stream._tag === "Success" &&
+            stepSettlement !== undefined &&
+            !providerFailed &&
+            !streamInterrupted &&
+            !userDeclined &&
+            !toolsInterrupted &&
+            infraError === undefined
 
           if (stream._tag === "Failure") return yield* Effect.failCause(stream.cause)
           if (userDeclined) return yield* Effect.interrupt
           if ((toolsInterrupted || infraError !== undefined) && settledFailure)
             return yield* Effect.failCause(settledFailure)
           if (toolsInterrupted && settled._tag === "Failure") return yield* Effect.failCause(settled.cause)
-          if (stepFailure) return yield* new StepFailedError({ error: stepFailure })
+          if (stepFailure && !recoveredMalformedToolInput) return yield* new StepFailedError({ error: stepFailure })
           return {
             _tag: "Completed",
-            needsContinuation,
+            needsContinuation: needsContinuation || recoveredMalformedToolInput,
+            malformedToolInput: recoveredMalformedToolInput,
             step: currentStep,
           } as const
         }),
@@ -346,6 +371,7 @@ const layer = Layer.effect(
       sessionID: SessionSchema.ID,
       promotion: SessionPending.Delivery | undefined,
       step: number,
+      recoverMalformedToolInput: boolean,
     ) {
       // Compaction restarts rebuild the request from compacted history without re-promoting.
       // Overflow recovery is one-shot: a post-compaction attempt must not recover another
@@ -356,7 +382,14 @@ const layer = Layer.effect(
       let assistantMessageID: SessionMessage.ID | undefined
       while (true) {
         const attempt = yield* Effect.suspend(() =>
-          attemptStep(sessionID, currentPromotion, currentStep, recoverOverflow, assistantMessageID),
+          attemptStep(
+            sessionID,
+            currentPromotion,
+            currentStep,
+            recoverMalformedToolInput,
+            recoverOverflow,
+            assistantMessageID,
+          ),
         ).pipe(
           Effect.tapError((error) =>
             error instanceof SessionRunnerRetry.RetryableFailure
@@ -378,7 +411,12 @@ const layer = Layer.effect(
               .pipe(Effect.andThen(Effect.fail(error.cause)))
           }),
         )
-        if (attempt._tag === "Completed") return { needsContinuation: attempt.needsContinuation, step: attempt.step }
+        if (attempt._tag === "Completed")
+          return {
+            needsContinuation: attempt.needsContinuation,
+            malformedToolInput: attempt.malformedToolInput,
+            step: attempt.step,
+          }
         if (attempt._tag === "RestartAfterOverflowCompaction") recoverOverflow = undefined
         yield* Effect.yieldNow
         currentPromotion = undefined
@@ -436,12 +474,18 @@ const layer = Layer.effect(
       while (shouldRun) {
         let needsContinuation = true
         let step = 1
+        let canRecoverMalformedToolInput = true
         // Repeat steps while continuation is needed. A step needs continuation only
         // when it recorded local tool calls whose results the model has not yet seen;
         // a provider error suppresses it. Pending steers also continue the loop so
         // interjections are answered before the session goes idle.
         while (needsContinuation) {
-          const result = yield* runStep(input.sessionID, promotion, step)
+          const result = yield* runStep(
+            input.sessionID,
+            promotion,
+            step,
+            canRecoverMalformedToolInput,
+          )
           // Steer/queue promotion inside runStep has already made the pending input a visible
           // user message by this point, so the first-user-message check below is reliable.
           if (!titleAttempted.has(input.sessionID)) {
@@ -449,6 +493,7 @@ const layer = Layer.effect(
             forkTitle(title.generateForFirstPrompt(yield* getSession(input.sessionID)).pipe(Effect.ignore))
           }
           needsContinuation = result.needsContinuation
+          if (result.malformedToolInput) canRecoverMalformedToolInput = false
           step = result.step + 1
           if (needsContinuation) {
             yield* runPendingCompaction(input.sessionID)

packages/core/src/session/runner/publish-llm-event.ts

@@ -9,6 +9,7 @@ import { SessionError } from "@opencode-ai/schema/session-error"
 import { Money } from "@opencode-ai/schema/money"
 import { AgentV2 } from "../../agent"
 import { Snapshot } from "../../snapshot"
+import { RelativePath } from "../../schema"
 import { SessionUsage } from "../usage"
 
 type Input = {
@@ -60,6 +61,7 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
   let stepFailed = false
   let providerFailed = false
   let retryEvidence = false
+  let malformedToolInput = false
   let stepFailure: SessionError.Error | undefined
   let stepSettlement:
     | {
@@ -112,12 +114,12 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
         if (state !== undefined) current.state = { ...current.state, ...state }
         return Effect.succeed(current.ordinal)
       })
-    const end = Effect.fnUntraced(function* (id: string, state?: Record<string, unknown>) {
+    const end = Effect.fnUntraced(function* (id: string, state?: Record<string, unknown>, value?: string) {
       const current = chunks.get(id)
       if (!current) return yield* Effect.die(new Error(`${name} end before start: ${id}`))
       yield* ended(
         id,
-        current.values.join(""),
+        value ?? current.values.join(""),
         current.ordinal,
         state === undefined ? current.state : { ...current.state, ...state },
       )
@@ -175,7 +177,11 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
     yield* toolInput.flush()
   })
 
-  const startToolInput = Effect.fnUntraced(function* (event: { readonly id: string; readonly name: string }) {
+  const startToolInput = Effect.fnUntraced(function* (event: {
+    readonly id: string
+    readonly name: string
+    readonly providerExecuted?: boolean
+  }) {
     if (tools.has(event.id)) return yield* Effect.die(new Error(`Duplicate tool input start: ${event.id}`))
     const assistantMessageID = yield* startAssistant()
     tools.set(event.id, {
@@ -183,7 +189,7 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
       name: event.name,
       called: false,
       settled: false,
-      providerExecuted: false,
+      providerExecuted: event.providerExecuted === true,
     })
     yield* toolInput.start(event.id)
     yield* events.publish(SessionEvent.Tool.Input.Started, {
@@ -194,13 +200,44 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
     })
   })
 
-  const endToolInput = Effect.fnUntraced(function* (event: { readonly id: string; readonly name: string }) {
+  const endToolInput = Effect.fnUntraced(function* (
+    event: { readonly id: string; readonly name: string },
+    value?: string,
+  ) {
     const tool = tools.get(event.id)
     if (!tool) return yield* Effect.die(new Error(`Tool input end before start: ${event.id}`))
     if (tool.name !== event.name)
       return yield* Effect.die(new Error(`Tool input name changed for ${event.id}: ${tool.name} -> ${event.name}`))
     if (!toolInput.has(event.id)) return yield* Effect.die(new Error(`Duplicate tool input end: ${event.id}`))
-    yield* toolInput.end(event.id)
+    yield* toolInput.end(event.id, undefined, value)
+  })
+
+  const failMalformedToolInput = Effect.fnUntraced(function* (event: {
+    readonly id: string
+    readonly name: string
+    readonly raw: string
+    readonly message: string
+  }) {
+    if (!tools.has(event.id)) yield* startToolInput(event)
+    const tool = tools.get(event.id)
+    if (!tool || tool.called || tool.settled)
+      return yield* Effect.die(new Error(`Malformed tool input after call settlement: ${event.id}`))
+    if (tool.name !== event.name)
+      return yield* Effect.die(new Error(`Tool input name changed for ${event.id}: ${tool.name} -> ${event.name}`))
+    if (toolInput.has(event.id)) yield* endToolInput(event, event.raw)
+    tool.settled = true
+    malformedToolInput = true
+    yield* events.publish(SessionEvent.Tool.Failed, {
+      sessionID: input.sessionID,
+      assistantMessageID: tool.assistantMessageID,
+      callID: event.id,
+      error: {
+        type: "tool.input-json",
+        message: "Tool call arguments were malformed JSON and were not executed. Retry with valid JSON.",
+      },
+      executed: false,
+    })
+    if (stepFailure === undefined) stepFailure = { type: "provider.invalid-output", message: event.message }
   })
 
   const flush = Effect.fn("SessionRunner.flush")(function* () {
@@ -236,9 +273,11 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
     if (replace || stepFailure === undefined) stepFailure = error
   })
 
-  const publishStepFailure = Effect.fnUntraced(function* (usage?: {
-    readonly cost: Money.USD
-    readonly tokens: ReturnType<typeof SessionUsage.tokens>
+  const publishStepFailure = Effect.fnUntraced(function* (details?: {
+    readonly cost?: Money.USD
+    readonly tokens?: ReturnType<typeof SessionUsage.tokens>
+    readonly snapshot?: Snapshot.ID
+    readonly files?: readonly RelativePath[]
   }) {
     if (stepFailed || stepFailure === undefined) return
     const assistantMessageID = yield* startAssistant()
@@ -247,7 +286,7 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
       sessionID: input.sessionID,
       assistantMessageID,
       error: stepFailure,
-      ...usage,
+      ...details,
     })
   })
 
@@ -337,6 +376,10 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
       case "tool-input-end":
         yield* endToolInput(event)
         return
+      case "tool-input-error":
+        retryEvidence = true
+        yield* failMalformedToolInput(event)
+        return
       case "tool-call": {
         retryEvidence = true
         if (!tools.has(event.id)) yield* startToolInput(event)
@@ -439,6 +482,7 @@ export const createLLMEventPublisher = (events: Pick<EventV2.Interface, "publish
     publishStepFailure,
     failUnsettledTools,
     hasProviderError: () => providerFailed,
+    hasMalformedToolInput: () => malformedToolInput,
     hasRetryEvidence: () => retryEvidence,
     stepFailure: () => stepFailure,
     stepSettlement: () => stepSettlement,

packages/core/src/session/runner/to-llm-message.ts

@@ -136,14 +136,20 @@ const assistant = (message: SessionMessage.Assistant, model: ModelV2.Ref, provid
         : item.text.length > 0
           ? [{ type: "text", text: item.text }]
           : []
+    const reuseToolProviderMetadata =
+      sameModel &&
+      (message.error === undefined ||
+        (item.executed === true &&
+          (item.state.status === "completed" ||
+            (item.state.status === "error" && item.state.result !== undefined))))
     const call = toolCall(
       item,
-      reuseProviderMetadata ? providerMetadata(providerMetadataKey, item.providerState) : undefined,
+      reuseToolProviderMetadata ? providerMetadata(providerMetadataKey, item.providerState) : undefined,
     )
     if (item.executed !== true) return [call]
     const result = toolResult(
       item,
-      reuseProviderMetadata
+      reuseToolProviderMetadata
         ? providerMetadata(providerMetadataKey, item.providerResultState ?? item.providerState)
         : undefined,
     )

packages/core/test/aisdk.test.ts

@@ -1,11 +1,11 @@
-import type { LanguageModelV3CallOptions } from "@ai-sdk/provider"
+import type { LanguageModelV3, LanguageModelV3CallOptions, LanguageModelV3StreamPart } from "@ai-sdk/provider"
 import { AISDK } from "@opencode-ai/core/aisdk"
 import { ModelV2 } from "@opencode-ai/core/model"
 import { ProviderV2 } from "@opencode-ai/core/provider"
-import { LLM, Message } from "@opencode-ai/ai"
-import { LLMClient } from "@opencode-ai/ai/route"
+import { LLM, LLMError, LLMEvent, Message } from "@opencode-ai/ai"
+import { LLMClient, RequestExecutor } from "@opencode-ai/ai/route"
 import { expect } from "bun:test"
-import { Effect } from "effect"
+import { Effect, Layer } from "effect"
 import { testEffect } from "./lib/effect"
 
 const it = testEffect(AISDK.locationLayer)
@@ -19,6 +19,37 @@ const model = (packageName: string, settings: Record<string, unknown> = {}) =>
     limit: { context: 100, output: 20 },
   })
 
+const streamModel = (events: ReadonlyArray<LanguageModelV3StreamPart>): LanguageModelV3 => ({
+  specificationVersion: "v3",
+  provider: "test",
+  modelId: "test",
+  supportedUrls: {},
+  doGenerate: () => Promise.reject(new Error("Unexpected non-streaming request")),
+  doStream: () =>
+    Promise.resolve({
+      stream: new ReadableStream({
+        start(controller) {
+          events.forEach((event) => controller.enqueue(event))
+          controller.close()
+        },
+      }),
+    }),
+})
+
+const usage = {
+  inputTokens: { total: 1, noCache: 1, cacheRead: 0, cacheWrite: 0 },
+  outputTokens: { total: 1, text: 0, reasoning: 0 },
+} as const
+
+const client = LLMClient.layer.pipe(
+  Layer.provide(
+    Layer.succeed(
+      RequestExecutor.Service,
+      RequestExecutor.Service.of({ execute: () => Effect.die("Unexpected HTTP request") }),
+    ),
+  ),
+)
+
 it.effect("keys language models by package and flattened overlays", () =>
   Effect.gen(function* () {
     const aisdk = yield* AISDK.Service
@@ -238,3 +269,69 @@ it.effect("projects replay metadata onto AI SDK prompt parts", () =>
     ])
   }),
 )
+
+it.effect("emits malformed AI SDK tool input without executing it", () =>
+  Effect.gen(function* () {
+    const aisdk = yield* AISDK.Service
+    const raw = '{"query":"partial'
+    yield* aisdk.hook.sdk((event) => {
+      event.sdk = {
+        languageModel: () =>
+          streamModel([
+            { type: "tool-input-start", id: "call_1", toolName: "lookup" },
+            { type: "tool-input-delta", id: "call_1", delta: raw },
+            { type: "tool-input-end", id: "call_1" },
+            { type: "tool-call", toolCallId: "call_1", toolName: "lookup", input: raw },
+            { type: "finish", finishReason: { unified: "tool-calls", raw: "tool_calls" }, usage },
+          ]),
+      }
+    })
+
+    const resolved = yield* aisdk.model(model("test-ai-sdk"))
+    const response = yield* LLMClient.generate(LLM.request({ model: resolved, prompt: "Lookup" })).pipe(
+      Effect.provide(client),
+    )
+
+    expect(response.events.find(LLMEvent.is.toolInputError)).toMatchObject({
+      id: "call_1",
+      name: "lookup",
+      raw,
+      message: "Invalid JSON input for aisdk tool call lookup",
+    })
+    expect(response.events.some(LLMEvent.is.toolInputEnd)).toBeTrue()
+    expect(response.events.some(LLMEvent.is.toolCall)).toBeFalse()
+  }),
+)
+
+it.effect("keeps malformed provider-executed AI SDK input terminal", () =>
+  Effect.gen(function* () {
+    const aisdk = yield* AISDK.Service
+    const raw = '{"query":"partial'
+    yield* aisdk.hook.sdk((event) => {
+      event.sdk = {
+        languageModel: () =>
+          streamModel([
+            { type: "tool-input-start", id: "call_1", toolName: "web_search", providerExecuted: true },
+            { type: "tool-input-delta", id: "call_1", delta: raw },
+            { type: "tool-input-end", id: "call_1" },
+            {
+              type: "tool-call",
+              toolCallId: "call_1",
+              toolName: "web_search",
+              input: raw,
+              providerExecuted: true,
+            },
+          ]),
+      }
+    })
+
+    const resolved = yield* aisdk.model(model("hosted-test-ai-sdk"))
+    const error = yield* LLMClient.generate(LLM.request({ model: resolved, prompt: "Search" })).pipe(
+      Effect.provide(client),
+      Effect.flip,
+    )
+
+    expect(error).toBeInstanceOf(LLMError)
+    expect(error.message).toContain("Invalid JSON input for aisdk tool call web_search")
+  }),
+)

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

@@ -565,6 +565,22 @@ Recent work
               text: "Partial thought",
               state: { itemId: "rs_failed", reasoningEncryptedContent: null },
             }),
+            SessionMessage.AssistantTool.make({
+              type: "tool",
+              id: "hosted-completed",
+              name: "web_search",
+              executed: true,
+              providerState: { itemId: "call_completed" },
+              providerResultState: { itemId: "result_completed" },
+              state: SessionMessage.ToolStateCompleted.make({
+                status: "completed",
+                input: { query: "Effect" },
+                content: [],
+                structured: {},
+                result: { type: "json", value: { found: true } },
+              }),
+              time: { created, completed: created },
+            }),
             SessionMessage.AssistantTool.make({
               type: "tool",
               id: "hosted-failed",
@@ -592,6 +608,22 @@ Recent work
 
     expect(messages[0]?.content).toEqual([
       { type: "text", text: "Partial thought" },
+      {
+        type: "tool-call",
+        id: "hosted-completed",
+        name: "web_search",
+        input: { query: "Effect" },
+        providerExecuted: true,
+        providerMetadata: { provider: { itemId: "call_completed" } },
+      },
+      {
+        type: "tool-result",
+        id: "hosted-completed",
+        name: "web_search",
+        result: { type: "json", value: { found: true } },
+        providerExecuted: true,
+        providerMetadata: { provider: { itemId: "result_completed" } },
+      },
       {
         type: "tool-call",
         id: "hosted-failed",

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

@@ -9,6 +9,8 @@ import { SessionMessage } from "@opencode-ai/core/session/message"
 import { SessionV2 } from "@opencode-ai/core/session"
 import { ModelV2 } from "@opencode-ai/core/model"
 import { ProviderV2 } from "@opencode-ai/core/provider"
+import { RelativePath } from "@opencode-ai/core/schema"
+import { Snapshot } from "@opencode-ai/core/snapshot"
 import { createLLMEventPublisher } from "@opencode-ai/core/session/runner/publish-llm-event"
 
 const sessionID = SessionV2.ID.make("ses_tool_event_test")
@@ -229,6 +231,8 @@ test("content-filter finish retains failure evidence until step closeout", async
     publisher.publishStepFailure({
       cost: Money.USD.make(1.25),
       tokens: settlement.tokens,
+      snapshot: Snapshot.ID.make("tree-end"),
+      files: [RelativePath.make("src/changed.ts")],
     }),
   )
   expect(published.map((event) => event.type)).toEqual(["session.step.started.1", "session.step.failed.1"])
@@ -236,6 +240,8 @@ test("content-filter finish retains failure evidence until step closeout", async
     error: { type: "provider.content-filter", message: "Provider blocked the response" },
     cost: 1.25,
     tokens: { input: 8, output: 2, reasoning: 1 },
+    snapshot: "tree-end",
+    files: ["src/changed.ts"],
   })
 })
 

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

@@ -4164,6 +4164,299 @@ describe("SessionRunnerLLM", () => {
     }),
   )
 
+  it.effect("continues once after malformed local tool input without exposing raw arguments", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      yield* admit(session, "Recover malformed tool input")
+      const marker = "raw-malformed-marker"
+      const raw = `{"text":"${marker}`
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolInputStart({ id: "call-malformed", name: "echo" }),
+          LLMEvent.toolInputDelta({ id: "call-malformed", name: "echo", text: raw }),
+          LLMEvent.toolInputEnd({ id: "call-malformed", name: "echo" }),
+          LLMEvent.toolInputError({
+            id: "call-malformed",
+            name: "echo",
+            raw,
+            message: "Invalid JSON input for test tool call echo",
+          }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        reply.stop(),
+      ]
+
+      yield* session.resume(sessionID)
+
+      expect(requests).toHaveLength(2)
+      expect(executions).toEqual([])
+      expect(JSON.stringify(requests[1])).not.toContain(marker)
+      expect(requests[1]?.messages).toEqual(
+        expect.arrayContaining([
+          expect.objectContaining({
+            role: "assistant",
+            content: expect.arrayContaining([
+              expect.objectContaining({ type: "tool-call", id: "call-malformed", name: "echo", input: {} }),
+            ]),
+          }),
+          expect.objectContaining({
+            role: "tool",
+            content: expect.arrayContaining([
+              expect.objectContaining({
+                type: "tool-result",
+                id: "call-malformed",
+                result: expect.objectContaining({
+                  type: "error",
+                  value: expect.objectContaining({
+                    error: expect.objectContaining({
+                      message: "Tool call arguments were malformed JSON and were not executed. Retry with valid JSON.",
+                    }),
+                  }),
+                }),
+              }),
+            ]),
+          }),
+        ]),
+      )
+      const context = yield* session.context(sessionID)
+      const failed = context.find(
+        (message): message is SessionMessage.Assistant =>
+          message.type === "assistant" && message.content.some((item) => item.type === "tool"),
+      )
+      expect(failed).toMatchObject({
+        error: { type: "provider.invalid-output", message: "Invalid JSON input for test tool call echo" },
+        content: [
+          {
+            type: "tool",
+            id: "call-malformed",
+            executed: false,
+            state: {
+              status: "error",
+              input: {},
+              error: {
+                type: "tool.input-json",
+                message: "Tool call arguments were malformed JSON and were not executed. Retry with valid JSON.",
+              },
+            },
+          },
+        ],
+      })
+      if (!failed) throw new Error("Malformed tool assistant missing")
+      expect((yield* recordedStepSettlementEvents(sessionID, failed.id)).map((event) => event.type)).toEqual([
+        "session.step.started.1",
+        "session.tool.failed.1",
+        "session.step.failed.1",
+      ])
+      const database = (yield* Database.Service).db
+      const durable = yield* database
+        .select({ type: EventTable.type, data: EventTable.data })
+        .from(EventTable)
+        .where(eq(EventTable.aggregate_id, sessionID))
+        .all()
+        .pipe(Effect.orDie)
+      expect(durable.find((event) => event.type === "session.tool.input.ended.1")?.data).toMatchObject({
+        callID: "call-malformed",
+        text: raw,
+      })
+    }),
+  )
+
+  it.effect("settles a valid sibling before recovering malformed tool input", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      yield* admit(session, "Run parallel tools")
+      toolExecutionGate = yield* Deferred.make<void>()
+      toolExecutionsStarted = yield* Deferred.make<void>()
+      toolExecutionsReady = 1
+      responses = [
+        [
+          LLMEvent.stepStart({ index: 0 }),
+          LLMEvent.toolCall({ id: "call-valid", name: "echo", input: { text: "valid" } }),
+          LLMEvent.toolInputError({
+            id: "call-malformed",
+            name: "echo",
+            raw: '{"text":"partial',
+            message: "Invalid JSON input for test tool call echo",
+          }),
+          LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+          LLMEvent.finish({ reason: "tool-calls" }),
+        ],
+        reply.stop(),
+      ]
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(toolExecutionsStarted)
+      expect(requests).toHaveLength(1)
+      yield* Deferred.succeed(toolExecutionGate, undefined)
+      yield* Fiber.join(run)
+      toolExecutionGate = undefined
+      toolExecutionsStarted = undefined
+
+      expect(requests).toHaveLength(2)
+      expect(executions).toEqual(["valid"])
+      const request = requests[1]
+      if (!request) throw new Error("Malformed recovery request missing")
+      expect(request.messages.flatMap((message) => (message.role === "tool" ? message.content : []))).toEqual(
+        expect.arrayContaining([
+          expect.objectContaining({ id: "call-valid", type: "tool-result" }),
+          expect.objectContaining({ id: "call-malformed", type: "tool-result" }),
+        ]),
+      )
+    }),
+  )
+
+  it.effect("does not recover malformed input after sibling execution is interrupted", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      yield* admit(session, "Interrupt malformed recovery")
+      toolExecutionGate = yield* Deferred.make<void>()
+      toolExecutionsStarted = yield* Deferred.make<void>()
+      toolExecutionsReady = 1
+      response = [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolCall({ id: "call-valid", name: "echo", input: { text: "blocked" } }),
+        LLMEvent.toolInputError({
+          id: "call-malformed",
+          name: "echo",
+          raw: '{"text":"partial',
+          message: "Invalid JSON input for test tool call echo",
+        }),
+        LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+        LLMEvent.finish({ reason: "tool-calls" }),
+      ]
+
+      const run = yield* session.resume(sessionID).pipe(Effect.forkChild)
+      yield* Deferred.await(toolExecutionsStarted)
+      while (
+        !(yield* session.context(sessionID)).some(
+          (message) =>
+            message.type === "assistant" &&
+            message.content.some((item) => item.type === "tool" && item.id === "call-malformed"),
+        )
+      )
+        yield* Effect.yieldNow
+      yield* session.interrupt(sessionID)
+      toolExecutionGate = undefined
+      toolExecutionsStarted = undefined
+
+      expect(yield* Fiber.await(run)).toMatchObject({ _tag: "Failure" })
+      expect(requests).toHaveLength(1)
+      expect(yield* session.context(sessionID)).toMatchObject([
+        { type: "user", text: "Interrupt malformed recovery" },
+        {
+          type: "assistant",
+          error: { type: "aborted", message: "Step interrupted" },
+          content: [
+            { type: "tool", id: "call-valid", state: { status: "error", error: { type: "aborted" } } },
+            { type: "tool", id: "call-malformed", state: { status: "error" } },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("records malformed provider-executed input as executed", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      yield* admit(session, "Fail malformed hosted input")
+      const failure = new LLMError({
+        module: "test",
+        method: "stream",
+        reason: new InvalidProviderOutputReason({ message: "Invalid hosted tool input" }),
+      })
+      responseStream = Stream.fromIterable([
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolInputStart({ id: "call-hosted", name: "web_search", providerExecuted: true }),
+        LLMEvent.toolInputDelta({ id: "call-hosted", name: "web_search", text: '{"query":"partial' }),
+      ]).pipe(Stream.concat(Stream.fail(failure)))
+
+      expect(yield* session.resume(sessionID).pipe(Effect.flip)).toBe(failure)
+      expect(requireAssistant(yield* session.context(sessionID))).toMatchObject({
+        error: { type: "provider.invalid-output", message: "Invalid hosted tool input" },
+        content: [
+          {
+            type: "tool",
+            id: "call-hosted",
+            executed: true,
+            state: { status: "error", error: { type: "provider.invalid-output" } },
+          },
+        ],
+      })
+    }),
+  )
+
+  it.effect("replaces malformed input diagnosis with a later provider failure", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      yield* admit(session, "Fail after malformed input")
+      const failure = new LLMError({
+        module: "test",
+        method: "stream",
+        reason: new InvalidProviderOutputReason({ message: "Provider failed after malformed input" }),
+      })
+      responseStream = Stream.fromIterable([
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolInputError({
+          id: "call-malformed",
+          name: "echo",
+          raw: '{"text":"partial',
+          message: "Invalid JSON input for test tool call echo",
+        }),
+      ]).pipe(Stream.concat(Stream.fail(failure)))
+
+      expect(yield* session.resume(sessionID).pipe(Effect.flip)).toBe(failure)
+      expect(requireAssistant(yield* session.context(sessionID))).toMatchObject({
+        error: { type: "provider.invalid-output", message: "Provider failed after malformed input" },
+        content: [
+          {
+            type: "tool",
+            id: "call-malformed",
+            executed: false,
+            state: { status: "error", error: { type: "tool.input-json" } },
+          },
+        ],
+      })
+      expect(requests).toHaveLength(1)
+    }),
+  )
+
+  it.effect("does not reset the malformed recovery budget after a valid tool step", () =>
+    Effect.gen(function* () {
+      const session = yield* setup
+      yield* admit(session, "Keep producing malformed tools")
+      const malformed = (id: string) => [
+        LLMEvent.stepStart({ index: 0 }),
+        LLMEvent.toolInputError({
+          id,
+          name: "echo",
+          raw: '{"text":"partial',
+          message: "Invalid JSON input for test tool call echo",
+        }),
+        LLMEvent.stepFinish({ index: 0, reason: "tool-calls" }),
+        LLMEvent.finish({ reason: "tool-calls" }),
+      ]
+      responses = [
+        malformed("call-first"),
+        reply.tool("call-valid-between", "echo", { text: "valid" }),
+        malformed("call-second"),
+        reply.stop(),
+      ]
+
+      expect(yield* session.resume(sessionID).pipe(Effect.flip)).toMatchObject({
+        error: {
+          type: "provider.invalid-output",
+          message: "Invalid JSON input for test tool call echo",
+        },
+      })
+
+      expect(requests).toHaveLength(3)
+      expect(executions).toEqual(["valid"])
+      expect((yield* recordedEventTypes(sessionID)).filter((type) => type === "session.step.failed.1")).toHaveLength(2)
+    }),
+  )
+
   it.effect("does not continue automatically after a provider error follows a local tool call", () =>
     Effect.gen(function* () {
       const session = yield* setup

packages/schema/src/session-event.ts

@@ -276,6 +276,8 @@ export namespace Step {
       error: SessionError.Error,
       cost: Money.USD.pipe(optional),
       tokens: TokenUsage.Info.pipe(optional),
+      snapshot: Snapshot.ID.pipe(optional),
+      files: Schema.Array(RelativePath).pipe(optional),
     },
   })
   export type Failed = typeof Failed.Type