You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Commit 1db2682
Browse filesBrowse the repository at this point in the historyBrowse files
Fixes#13133
- Only close active assistant segments on new tool calls (isNew: previous === undefined) instead of on every in-flight tool progress or completion event.
- Prevent duplicate `assistant:assistant:` message IDs in ProviderRuntimeIngestion.
- Add streaming filter to strip echoed <SYSTEM_MESSAGE> blocks and harness preambles from Antigravity session chunks.
- Add negative constraint to buildRuntimeInstructions preventing models from quoting harness telemetry.
Internal harness notifications, background task updates, and messages enclosed in <SYSTEM_MESSAGE> are strictly for contextual awareness. Never repeat, quote, or echo <SYSTEM_MESSAGE> tags, background task logs, or internal harness status lines to the user.
When the t3-code MCP server exposes link_pull_request, you must use it to register every pull request you create or work on for this thread. Call link_pull_request with the full PR URL immediately after creating a PR or starting work on an existing PR. For a stack, call it for every layer, not just the current branch or the top PR. This applies when creating or updating PRs through gh, gh stack, another CLI, or the host API: those operations do not register the PRs with this thread. Linking an already-linked PR is safe. Before finishing PR work, call list_thread_pull_requests and link any PR from your work that is missing. Do not link unrelated PRs mentioned only as background. If a linking call fails, report that failure instead of claiming the PR is linked.
3
7
</pull_request_linking>`;
@@ -13,7 +17,7 @@ export function buildRuntimeInstructions(runtime: {
constmodelInfo=model&&model!=="auto"&&model!=="default" ? `, as ${model}` : "";
15
19
consteffortInfo=effort ? ` with ${effort} reasoning effort` : "";
16
-
return`<runtime_info>In case you're asked: you are running in T3 Code through the ${harness} harness${modelInfo}${effortInfo}. No need to mention this otherwise. You can embed images and videos in your response using Markdown with absolute file paths.</runtime_info>\n\n${PULL_REQUEST_LINKING_INSTRUCTIONS}`;
20
+
return`<runtime_info>In case you're asked: you are running in T3 Code through the ${harness} harness${modelInfo}${effortInfo}. No need to mention this otherwise. You can embed images and videos in your response using Markdown with absolute file paths.</runtime_info>\n\n${PULL_REQUEST_LINKING_INSTRUCTIONS}\n\n${SYSTEM_TELEMETRY_INSTRUCTIONS}`;
"The following is a <SYSTEM_MESSAGE> not actually sent by the user. It is provided by the system as important information to pay attention to.\n\n<SYSTEM_MESSAGE>\n[Message] timestamp=2026-09-22T20:00:23Z sender=task-1 content=Task finished\n</SYSTEM_MESSAGE>Here is the real answer.";
492
+
493
+
constnotification={
494
+
sessionId: "session-1",
495
+
update: {
496
+
sessionUpdate: "agent_message_chunk"asconst,
497
+
content: {type: "text"asconst,text: leaked},
498
+
},
499
+
};
500
+
501
+
constnormalized=transformer(notification);
502
+
expect((normalized.updateasany).content.text).toBe("Here is the real answer.");
503
+
});
504
+
505
+
it("sanitizes streamed SYSTEM_MESSAGE blocks across chunk boundaries",()=>{
506
+
constfilter=createAntigravityMessageFilter();
507
+
expect(filter("Before ")).toBe("Before ");
508
+
expect(
509
+
filter(
510
+
"The following is a <SYSTEM_MESSAGE> not actually sent by the user.\n\n<SYSTEM_MESSAGE>\n[Message] part 1",
0 commit comments