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 f24876a
Browse filesBrowse the repository at this point in the historyBrowse files
feat(composer): skip predictions the agent will answer itself, edit the prompt in Labs
Two thirds of predictions on live data followed a turn that ended while the
agent waited for its own tests, builds or subagents; the agent carried on by
itself, so they read a long thread from cache for nothing or suggested
"ok, waiting". A turn whose thread still has work the agent will be woken by
now gets no prediction.
The prompt now treats a question anywhere in the last message, or a closing
plan of what the assistant will do once approved, as waiting on the user,
and rules out acknowledgments that ask for nothing.
The prompt can be edited under Settings → Labs once reply suggestions are on;
an empty setting keeps the built-in prompt, which moves to contracts so the
editor can show it.
description="After a Claude turn finishes, predict your next message and show it as ghost text in an empty composer: Tab or Right arrow fills the draft, Esc hides it. The prediction reuses the finished turn's prompt cache, so it pays only the new text, and it never runs once that cache is gone."
191
+
description="After a Claude turn finishes, predict your next message and show it as ghost text in an empty composer: Tab or Right arrow fills the draft, Esc hides it. The prediction reads the thread from the finished turn's prompt cache, so it never runs once that cache is gone."
177
192
control={
178
193
<ScopedSwitch
179
194
settingKeys={["composerReplySuggestions"]}
@@ -188,6 +203,55 @@ function ComposerReplySuggestionsRow() {
"[Composer prediction request from the app, not a message from the user. Do not answer the user, continue the task, or call tools.]",
1208
+
"",
1209
+
'Predict what the USER would type next, not what you think they should do. The test: would they think "I was just about to type that"?',
1210
+
"",
1211
+
"Start from the user's original request and their recent messages.",
1212
+
'- They asked for several steps and one is done: the next step ("run the tests").',
1213
+
'- Code was just written: "try it out".',
1214
+
"- The assistant offered options: the one this user would pick, judging by the conversation.",
1215
+
'- The assistant asked whether to continue: "yes" or "go ahead".',
1216
+
'- The task is complete with an obvious follow-up: "commit this", "push it".',
1217
+
"- The question or proposal waiting on the user can sit anywhere in the assistant's last message, not only at its end. A message that closes with what the assistant will do once approved is asking for approval.",
1218
+
"- After an error or a misunderstanding: stay silent, they need to assess it themselves.",
1219
+
"- The assistant is waiting for its own background work (tests, builds, other agents) and will continue without the user: stay silent.",
1220
+
'Be specific: "run the tests" beats "continue".',
1221
+
"",
1222
+
'Never suggest: praise or thanks ("looks good"), an acknowledgment that asks for nothing ("ok", "waiting", "got it"), a question, anything in the assistant\'s voice ("Let me...", "I\'ll...", "Here\'s..."), a new idea the user did not ask about, or more than one sentence. Stay silent when the next step is not obvious from what the user said, or when the suggestion could be unsafe or touches credentials, private data, or security incidents.',
1223
+
"",
1224
+
"Format: 2-12 words in the user's own language and style. Output only the message text: no role label, no quotes, no explanation. To stay silent, output exactly NONE.",
1225
+
].join("\n");
1226
+
1201
1227
exportconstServerSettings=Schema.Struct({
1202
1228
// Null means unmigrated; empty is authoritative and must not resurrect cleared preferences.
0 commit comments