🐛 Bug Report: MCP instrumentation crashes on error results with non-text content

Aperta
#4,569 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

@notshriyansh ci sta già lavorando.

Dal 6/10/2026.

  • #4570 di @notshriyansh — aperta

Valutazione

Difficoltà
3/5
Tempo stimato
1-2 giorni
Idoneità per principianti
73/100
Tipo di issue
Bug
Chiarezza
Specificata chiaramente
Stato di attività
Attiva
Stack tecnologico
python
Ambito
observability

Direzione di ricerca

Read _execute_and_handle_result() in the MCP instrumentation and inspect how it handles CallToolResult errors; also check InstrumentedStreamWriter.send() for the similar streamed-content assumption. Reproduce with a CallToolResult containing ImageContent, then run the relevant instrumentation tests. Done means non-text error content no longer raises an instrumentation exception and the span records error.type = tool_error.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

Which component is this bug for?

All Packages

📜 Description

opentelemetry-instrumentation-mcp assumes that the first content block in an MCP CallToolResult with isError=True is always a TextContent.

In _execute_and_handle_result(), the instrumentation currently does:

result.content[0].text

when handling an MCP tool error.

However, MCP CallToolResult.content is a list of ContentBlock values and can contain other valid content types such as ImageContent, AudioContent, ResourceLink, or EmbeddedResource.

As a result, a valid MCP error result containing a non-text content block causes the instrumentation itself to raise AttributeError.

For example, this is a valid CallToolResult:

CallToolResult(
    content=[
        ImageContent(
            type="image",
            data="aGVsbG8=",
            mimeType="image/png",
        )
    ],
    isError=True,
)

Passing this result through _execute_and_handle_result() currently raises:

AttributeError: 'ImageContent' object has no attribute 'text'

The span is consequently recorded with error.type = AttributeError, rather than the MCP tool error being represented as a tool_error.

I also found a similar content[0]["text"] assumption in InstrumentedStreamWriter.send() when handling streamed error results, although I have not yet verified whether the same failure is reachable through the streaming path.

👟 Reproduction steps

This can be reproduced without an MCP server, network access, or credentials using a protocol-valid CallToolResult in the instrumentation's result-handling path:

from mcp.types import CallToolResult, ImageContent

result = CallToolResult(
    content=[
        ImageContent(
            type="image",
            data="aGVsbG8=",
            mimeType="image/png",
        )
    ],
    isError=True,
)

# Pass `result` through the MCP instrumentation's
# _execute_and_handle_result() path.

The current implementation attempts to access:

result.content[0].text

and raises:

AttributeError: 'ImageContent' object has no attribute 'text'

The span produced by the instrumentation has:

error.type = AttributeError

instead of treating the MCP result as a tool-level error.

👍 Expected behavior

The instrumentation should handle valid MCP error results regardless of the concrete ContentBlock type, without raising an exception itself.

The tool result should still produce an ERROR span with:

error.type = tool_error

and the instrumentation should not assume that the first content block is TextContent.

👎 Actual Behavior with Screenshots

A valid error result containing a non-text content block causes the instrumentation to raise:

AttributeError: 'ImageContent' object has no attribute 'text'

and the resulting span records:

error.type = AttributeError

This means the instrumentation can mask the actual MCP tool error.

🤖 Python Version

3.12.6

📃 Provide any additional context for the Bug.

Environment:

opentelemetry-instrumentation-mcp: 0.62.4
mcp: 1.26.0
fastmcp: 2.14.3
Python: 3.12.6

mcp==1.26.0 defines CallToolResult.content as a list of MCP content block types rather than a list containing only TextContent.

I have reproduced the failure directly against the current instrumentation code.

There is also a similar content[0]["text"] assumption in the streaming response writer path, which may need to be handled separately.

👀 Have you spent some time to check if this bug has been raised before?
  • I checked and didn't find similar issue
Are you willing to submit PR?

Yes I am willing to submit a PR!

Lingua principale
Python
Stelle
7.5k
Fork
1.2k
Merge medio
2g 6h
PR unite (30g)
2

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di traceloop/openllmetry

Tutte le issue di traceloop/openllmetry

Issue simili

Altre issue su Python

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.