onAfterToolCall failure records a second, contradictory result for a successful server tool #1558

TanStack AI version: @tanstack/ai 0.63.0 at main commit 62bec34bb.

Framework/Library version: Node.js 24.11.1; framework-independent server tool path.

Describe the bug and the steps to reproduce it

When a server tool returns successfully and onAfterToolCall throws, one tool-call ID gets both a success result and an error result. TanStack AI also invokes the hook twice: first with ok: true, then with ok: false. If the hook's second invocation returns, the run continues as though the tool itself had failed.

This can mislead the model and downstream consumers. It also repeats a hook with side effects, such as an audit write or trace export.

The exact path on main is:

1. executeToolCalls() selects a server tool and delegates to executeServerTool().
2. tool.execute() returns { sent: true }. executeServerTool() pushes that success result into results.
3. onAfterToolCall({ ok: true, result: { sent: true } }) throws post-hook failed.
4. The surrounding catch treats the hook error as a tool execution error. It pushes { error: 'post-hook failed' } for the same toolCallId and calls onAfterToolCall({ ok: false, error }).
5. The second hook call returns, so executeToolCalls() returns both results and does not propagate the hook error to the run.

This differs from a real tool failure. If tool.execute() throws before a result exists, one error result and one onAfterToolCall({ ok: false }) call are expected.

Minimal reproduction

The following is the complete script used for the red/green check. Save it as repro.mts and run pnpm exec tsx repro.mts <absolute-checkout-root> after pnpm install. Run it once against clean main at 62bec34bb and once against the proposed fix.

import { pathToFileURL } from 'node:url'
import path from 'node:path'

const root = process.argv[2]
if (!root) throw new Error('Pass the checkout root')
const url = pathToFileURL(path.join(root, 'packages/ai/src/activities/chat/tools/tool-calls.ts')).href
const { executeToolCalls } = await import(url)
let executed = 0
let afterCalls = 0
const calls = [{ id: 'call-1', type: 'function', function: { name: 'sendEmail', arguments: '{}' } }]
const tools = [{ name: 'sendEmail', execute: () => { executed++; return { sent: true } } }]
const generator = executeToolCalls(calls, tools, new Map(), new Map(), undefined, {
  onAfterToolCall: async () => {
    afterCalls++
    if (afterCalls === 1) throw new Error('post-hook failed')
  },
})
let result
let error
try {
  for (;;) {
    const step = await generator.next()
    if (step.done) { result = step.value; break }
  }
} catch (caught) {
  error = caught instanceof Error ? caught.message : String(caught)
}
const observed = { executed, afterCalls, error: error ?? null, results: result?.results ?? null }
console.log(JSON.stringify(observed))
if (executed !== 1 || afterCalls !== 1 || error !== 'post-hook failed' || result !== undefined) {
  throw new Error('after-hook failure was reported as a tool result or hook ran twice')
}

Actual on clean main:

{"executed":1,"afterCalls":2,"error":null,"results":[{"toolCallId":"call-1","toolName":"sendEmail","result":{"sent":true},"input":{},"output":{"sent":true},"duration":0},{"toolCallId":"call-1","toolName":"sendEmail","result":{"error":"post-hook failed"},"input":{},"state":"output-error","duration":0}]}

The script exits with code 1 on main. The duration: 0 values are observations from this run; durations vary with the clock.

Expected: executed and afterCalls both equal 1. The hook error propagates, so error is post-hook failed. The chat run calls onError; its SSE stream ends with RUN_ERROR and no TOOL_CALL_RESULT for this call.

With the proposed fix, the same script returned {"executed":1,"afterCalls":1,"error":"post-hook failed","results":null} and exited with code 0. The chat-level regression test also checks that onError runs once.

Your Minimal, Reproducible Example - (Sandbox Highly Recommended)

Regression test on the fork branch and the standalone repro.mts above.

Screenshots or Videos (Optional)

Not applicable. The JSON output above shows the duplicate result.

Do you intend to try to help solve this bug with your own PR?

Yes, I am also opening a PR that solves the problem alongside this issue.

Related reports and PRs

I searched open and closed issues and PRs for onAfterToolCall, duplicate tool results, and after-tool errors. I found no report of this failure. #1547 concerns serial execution of independent tool calls. #1513 changes subagent return values near this function, but it does not address hook failures.

Terms & Code of Conduct

• I agree to follow this project's Code of Conduct.
• I understand that a bug without a reliable reproduction may be closed.