fix(core): keep tool result errors when importing threads

[kevin9327]

Summary

When a conversation is imported (file import or drag-and-drop), failed tool calls come in as successful:

• Anthropic Messages JSON: a tool_result block with is_error: true loses the flag.
• DeerFlow run-event JSONL: DeerFlow persists tool results as ToolMessage.model_dump(), so a failed call carries LangChain's status: "error". That field is ignored too.

The imported tool call then shows as completed rather than errored (getToolCallStatus() reads output.isError). Re-running the thread also replays that result to the model with isError: false (convertToPiContext()), so the model is told a failed call succeeded. For a workbench used to debug agents, that hides exactly the failure being investigated.

normalizeToThread() already resolves both shapes. The flag was simply not carried to _attachToolResult(). It now sets output.isError: true for failed results. Successful results keep the exact { content } output they had before, so already-imported or saved threads do not change shape.

Validation

bun test packages/core/tests/parsers/json-thread-parser.test.ts packages/core/tests/parsers/deerflow-jsonl-thread-parser.test.ts

Before the fix (new tests added, source unchanged):

json-thread-parser.test.ts: 4 pass, 1 fail

packages\core\tests\parsers\json-thread-parser.test.ts:
 96 |       expect(assistant?.role).toBe("assistant");
 97 |       if (assistant?.role === "assistant") {
 98 |         expect(assistant.toolCalls?.[0]?.output).toEqual({
 99 |           content: [{ type: "text", text: "hello" }],
100 |         });
101 |         expect(assistant.toolCalls?.[1]?.output).toEqual({
                                                       ^
error: expect(received).toEqual(expected)

@@ -7,3 +7,3 @@
    ],
-   "isError": true,
  }

- Expected  - 1
+ Received  + 0

      at <anonymous> (packages\core\tests\parsers\json-thread-parser.test.ts:101:50)
(fail) JsonThreadParser > keeps the error flag of imported Anthropic tool results [3.33ms]

 4 pass
 1 fail

deerflow-jsonl-thread-parser.test.ts: 5 pass, 1 fail

packages\core\tests\parsers\deerflow-jsonl-thread-parser.test.ts:
142 |       expect(assistant?.role).toBe("assistant");
143 |       if (assistant?.role === "assistant") {
144 |         expect(assistant.toolCalls?.[0]?.output).toEqual({
145 |           content: [{ type: "text", text: "alpha" }],
146 |         });
147 |         expect(assistant.toolCalls?.[1]?.output).toEqual({
                                                       ^
error: expect(received).toEqual(expected)

@@ -7,3 +7,3 @@
    ],
-   "isError": true,
  }

- Expected  - 1
+ Received  + 0

      at <anonymous> (packages\core\tests\parsers\deerflow-jsonl-thread-parser.test.ts:147:50)
(fail) DeerFlowJsonlThreadParser > keeps the error status of tool results [1.09ms]

 5 pass
 1 fail

After the fix: 11 pass, 0 fail across both files.

• In each new test, the successful result's assertion (toEqual({ content: [...] }), with no isError key) passes both before and after. The existing DeerFlow import test, which pins a successful tool output exactly, is also unchanged. This shows successful results are not rewritten.
• bun test packages/core/tests/parsers packages/core/tests/thread packages/core/tests/client: 138 pass, 0 fail.
• ESLint (--max-warnings 0) on the three touched files: clean on main and on this branch.
• tsc --noEmit -p tsconfig.json: clean.
• Prettier: normalize-thread.ts already has 11 unformatted lines on main; this change adds none, and both test files stay clean.

🤖 Generated with Claude Code