onAfterToolCall failure records a second, contradictory result for a successful server tool #1558
Send to my agentTanStack 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:
executeToolCalls()selects a server tool and delegates toexecuteServerTool().tool.execute()returns{ sent: true }.executeServerTool()pushes that success result intoresults.onAfterToolCall({ ok: true, result: { sent: true } })throwspost-hook failed.- The surrounding
catchtreats the hook error as a tool execution error. It pushes{ error: 'post-hook failed' }for the sametoolCallIdand callsonAfterToolCall({ ok: false, error }). - 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.
github-actions[bot](GitHub profile, opens in a new tab) assignedAlemTuzlak
github-actions[bot](GitHub profile, opens in a new tab) added labelshas-pr
github-actions[bot](GitHub profile, opens in a new tab) added labelswaiting-on: maintainer
dakjdakd(GitHub profile, opens in a new tab) linked this to #1559
Sign in to join this conversation.