fix(mcp): sessionId loses precision in JavaScript JSON consumers
Mantenedores costumam responder em até 2 dias
Ninguém assumiu esta issue ainda.
Avaliação
- Dificuldade
- 4/5
- Tempo estimado
- 3-5 dias
- Facilidade para iniciantes
- 48/100
- Tipo de issue
- Bug
- Clareza
- Razoavelmente clara
- Status de atividade
- Ativa
- Stack de tecnologia
- javascript, json
- Domínio
- api
Direção de pesquisa
The report identifies structuredContent.openItem.sessionId in the GUI MCP server response, but the server implementation is shipped inside the proprietary application and no source file or test is named. Start by locating the code that builds the bn_binary_view_set_active response and checking how it serializes sessionId; reproduce with an identifier above JavaScript’s safe-integer limit. Done means JavaScript can recover the exact identifier while opaque handle selection still works.
Escrita pelo modelo de indexação a partir do texto da issue.
Descrição
Summary
The GUI MCP server emits structuredContent.openItem.sessionId as a bare JSON integer above JavaScript's safe-integer range. Parsing the raw HTTP reply with JSON.parse changes its exact integer value. This is an interoperability defect in emitted metadata, not a claim that the JSON syntax is invalid or that opaque openItem handles stop working.
Steps to reproduce
- Save the C source below as
fixture.c, compile it, and openfixturein Binary Ninja. Enable and start the built-in GUI MCP server, then connect an MCP client and runtools/list.
#include <stdio.h>
volatile int fixture_global = 7;
__attribute__((noinline, used)) int fixture_leaf(int value) { puts("FIXTURE_LEAF"); return value + fixture_global; }
__attribute__((noinline, used)) int fixture_branch(int value) { return value > 3 ? fixture_leaf(value) : fixture_leaf(-value); }
int main(int argc, char **argv) { return fixture_branch(argc); }
clang -arch arm64 -O0 -g -fno-inline fixture.c -o fixture
-
Call
bn_binary_view_list, select the returned Mach-O view withbn_binary_view_set_active, and wait untilbn_analysis_statusreportsIdleState. Use the returned handle unchanged; handle values are session-scoped. -
Capture a raw HTTP reply from
bn_binary_view_set_activebefore a JavaScript JSON parser touches it. The following Python standard-library script initializes a separate MCP connection, retainsMCP-Session-Id, lists tools and views, selects the returned Mach-O handle, and saves the response bytes asreply.raw. Run it in a disposable directory while the fixture is open. It does not execute the target or save a database.
import json
import urllib.request
from pathlib import Path
url = "http://127.0.0.1:24642/mcp"
headers = {"Content-Type": "application/json", "Accept": "application/json, text/event-stream"}
def send(method, params=None, request_id=1):
message = {"jsonrpc": "2.0", "method": method}
if request_id is not None:
message["id"] = request_id
if params is not None:
message["params"] = params
request = urllib.request.Request(url, json.dumps(message).encode(), headers)
with urllib.request.urlopen(request) as response:
return response.headers, response.read()
def decode(raw):
text = raw.decode()
if text.startswith(("event:", "data:")):
text = "\n".join(line[5:].strip() for line in text.splitlines() if line.startswith("data:"))
return json.loads(text)
response_headers, _ = send("initialize", {"protocolVersion": "2025-11-25", "capabilities": {}, "clientInfo": {"name": "repro", "version": "1"}})
headers["MCP-Session-Id"] = response_headers["MCP-Session-Id"]
send("notifications/initialized", request_id=None)
send("tools/list", {})
_, listing = send("tools/call", {"name": "bn_binary_view_list", "arguments": {}})
text = decode(listing)["result"]["content"][0]["text"]
rows = [[cell.strip() for cell in line.strip().strip("|").split("|")] for line in text.splitlines() if line.startswith("|")]
handle = next(row[0] for row in rows if row[2] == "Mach-O")
_, raw = send("tools/call", {"name": "bn_binary_view_set_active", "arguments": {"binaryView": handle}})
Path("reply.raw").write_bytes(raw)
- Run this JavaScript against those saved bytes. It isolates the numeric token in
structuredContentand compares its exact integer value with the already-parsed JavaScript number, without publishing the generated identifier.
const fs = require('fs');
const raw = fs.readFileSync('reply.raw', 'utf8');
const json = raw.startsWith('event:') || raw.startsWith('data:')
? raw.split('\n').filter(line => line.startsWith('data:')).map(line => line.slice(5).trim()).join('\n')
: raw;
const token = json.match(/"structuredContent"[\s\S]*?"sessionId"\s*:\s*(\d+)/)[1];
const id = JSON.parse(json).result.structuredContent.openItem.sessionId;
console.log('raw sessionId above MAX_SAFE_INTEGER:', BigInt(token) > BigInt(Number.MAX_SAFE_INTEGER));
console.log('Number.isSafeInteger(parsed):', Number.isSafeInteger(id));
console.log('BigInt(parsed) === BigInt(raw):', BigInt(id) === BigInt(token));
Expected behavior
An identifier exposed to MCP JSON clients should retain its exact value through supported JSON consumption. The guide explicitly formats output addresses as strings to avoid losing 64-bit precision, but does not document a precision-loss exception for sessionId. RFC 8259 section 6 describes the interoperable exact-integer range for binary64 consumers; it does not prohibit larger JSON number tokens.
Actual behavior
Raw HTTP bytes confirmed that the property was a bare decimal JSON number, not a quoted decimal string. The generated value exceeded Number.MAX_SAFE_INTEGER. The JavaScript check printed:
raw sessionId above MAX_SAFE_INTEGER: true
Number.isSafeInteger(parsed): false
BigInt(parsed) === BigInt(raw): false
The text field of the same response contained the original digits inside its JSON-encoded string, and bn_open_item_list displayed those digits in its Markdown table. The loss occurs when parsing the numeric structuredContent field with JavaScript's normal JSON parser. This was observed for one disposable fixture's generated identifier; frequency across other identifiers is unmeasured. Generated identifier values and local paths are intentionally omitted from this public report.
Evidence
- Expected source: JSON interoperability guidance above and the MCP guide's output-address precision convention, linked under Runtime or environment;
sessionIdhas no separate documented precision contract. - Failure source: Raw HTTP response bytes from the view-selection tool and the executed JavaScript token-versus-parsed-integer comparison shown above.
- Evidence provenance: observed
- Local verification: reproduced
- Reproduction completeness: complete
Restoration check
Capture the same response before JSON parsing and verify that a normal JavaScript JSON consumer can recover the exact emitted identifier value. Include a generated identifier above the binary64 safe-integer range as a control; passing only with a smaller identifier would not cover the observed failure. Preserve normal opaque-handle selection behavior.
Runtime or environment
- Binary Ninja Version: Free 6.0.10905-Stable; bundle version
6.0.10905_free; built-in GUI HTTP MCP serverbinaryninja_ui_mcpversion0.1.0. - Edition: Free. OS: macOS 27.2; CPU architecture: arm64. Target: thin arm64 Mach-O built with the command above. Analysis status was
IdleStatebefore the observations. - The official stable release metadata matched the installed build. A fresh profile enabled
ui.mcp.enabled; no user plugins or existing databases were involved. The target was analyzed, never executed, and no database was saved. - The stable MCP guide and the current default-branch MCP guide were compared. Their relevant calling conventions agree. The implementation ships inside the proprietary application, so public default-branch source inspection does not verify its runtime behavior.
- The issue template requests development-build testing. That was not performed: Free cannot switch to the development channel. Personal/Commercial editions, development builds, Windows, and Linux are untested.
Triage assessment
- Impact level: P2
- Assessment status: supported
- Impact basis: Normal JavaScript JSON consumption silently changes the exact identifier in structured metadata for the observed value. No failed handle operation, collision, database corruption, or analysis error was observed.
- Workaround status: available
- Workaround basis: The exact digits remained in the reply's text string and in the open-item list's Markdown table; a client can preserve them as text. Python's arbitrary-precision JSON parser also retained the raw integer. These avoid the observed numeric precision loss without changing the binary or server.
- Linguagem predominante
- C++
- Estrelas
- 1.3k
- Forks
- 304
- Merge médio
- 2d 20h
- PRs com merge (30d)
- 10
Preparar o ambiente
Este projeto não oferece contêiner de desenvolvimento, Dockerfile nem guia de contribuição, então a configuração fica por sua conta: comece pelo README e veja nosso guia da primeira contribuição para os passos gerais.
Primeiros passos
- Leia a issue inteira e depois o guia de contribuição do projeto.
- Comente na issue dizendo que vai assumir — evita que duas pessoas façam o mesmo trabalho.
- Faça um fork do repositório e trabalhe em uma branch.
- Abra um pull request que referencie o número da issue.
Mais de Vector35/binaryninja-api
-
Dificuldade 1/5 Menos de uma hora Facilidade para iniciantes 84/100
Vector35/binaryninja-api#8715 ·
Mantenedores costumam responder em até 2 dias
-
Dificuldade 2/5 1-3 horas Facilidade para iniciantes 78/100
Vector35/binaryninja-api#8714 ·
Mantenedores costumam responder em até 2 dias
-
Dificuldade 1/5 Menos de uma hora Facilidade para iniciantes 85/100
Vector35/binaryninja-api#8713 ·
Mantenedores costumam responder em até 2 dias
-
Dificuldade 1/5 1-3 horas Facilidade para iniciantes 78/100
Vector35/binaryninja-api#8689 ·
Mantenedores costumam responder em até 2 dias
-
Component: Python API
Dificuldade 2/5 1-3 horas Facilidade para iniciantes 70/100
Vector35/binaryninja-api#8649 ·
Mantenedores costumam responder em até 2 dias
Todas as issues de Vector35/binaryninja-api
Issues semelhantes
-
Bug Core VIF
Dificuldade 2/5 1-3 horas Facilidade para iniciantes 68/100
Mantenedores costumam responder em até 1 dia
-
Dificuldade 2/5 1-3 horas Facilidade para iniciantes 80/100
ESPresense/ESPresense#2680 · 2 comentários ·
Mantenedores costumam responder em até 1 dia
-
Dificuldade 2/5 Menos de uma hora Facilidade para iniciantes 75/100
stew675/llama-cpp-rdna-boosts#348 ·
Mantenedores costumam responder em até 1 dia
-
Documentation: Add comprehensive newcomer onboarding guide, project structure, and local usage to READMETalvez já em andamento @Ash-Blanc assumiu hoje. Abertaarea/other status/bot-triaged status/need-triage
Dificuldade 2/5 Meio dia Facilidade para iniciantes 85/100
-
bug
Dificuldade 1/5 Menos de uma hora Facilidade para iniciantes 85/100
Mantenedores costumam responder em até 1 dia