fix(security): harden untrusted web tool transcripts

This commit is contained in:
Peter Steinberger
2026-02-13 00:46:11 +01:00
parent 4543c401b4
commit da55d70fb0
13 changed files with 484 additions and 18 deletions

View File

@@ -322,6 +322,25 @@ export function applyGoogleTurnOrderingFix(params: {
return { messages: sanitized, didPrepend };
}
function stripToolResultDetails(messages: AgentMessage[]): AgentMessage[] {
let touched = false;
const out: AgentMessage[] = [];
for (const msg of messages) {
if (!msg || typeof msg !== "object" || (msg as { role?: unknown }).role !== "toolResult") {
out.push(msg);
continue;
}
if (!("details" in msg)) {
out.push(msg);
continue;
}
const { details: _details, ...rest } = msg as unknown as Record<string, unknown>;
touched = true;
out.push(rest as unknown as AgentMessage);
}
return touched ? out : messages;
}
export async function sanitizeSessionHistory(params: {
messages: AgentMessage[];
modelApi?: string | null;
@@ -353,6 +372,7 @@ export async function sanitizeSessionHistory(params: {
const repairedTools = policy.repairToolUseResultPairing
? sanitizeToolUseResultPairing(sanitizedToolCalls)
: sanitizedToolCalls;
const sanitizedToolResults = stripToolResultDetails(repairedTools);
const isOpenAIResponsesApi =
params.modelApi === "openai-responses" || params.modelApi === "openai-codex-responses";
@@ -368,8 +388,8 @@ export async function sanitizeSessionHistory(params: {
: false;
const sanitizedOpenAI =
isOpenAIResponsesApi && modelChanged
? downgradeOpenAIReasoningBlocks(repairedTools)
: repairedTools;
? downgradeOpenAIReasoningBlocks(sanitizedToolResults)
: sanitizedToolResults;
if (hasSnapshot && (!priorSnapshot || modelChanged)) {
appendModelSnapshot(params.sessionManager, {