* claim: #8220 * Agent chat: Custom instructions in Preferences, added to the system note (#8220) Preferences > AI gets a multi-line Custom instructions field, saved with the other AI preferences (agent_instructions, kept by page resets like other user-written text). The chat adds the text after its own system note when a session starts, by the routes #9811 set up: ACP session/new (the Claude adapter's _meta key, the standard systemPrompt field, the Antigravity key) and the system message or field of every API protocol. Connect and New chat set it, so an edit applies from the next chat and never changes the one running. Blank text adds nothing; at most 4000 characters are kept, in the field, over ui.agent_chat and from a hand- edited settings file. ui.agent_chat takes and reports `instructions`. * Custom instructions: typing in the field stays in the field The text area did not report itself to Preferences search as a key sink, so with the field focused, search took the typed characters and moved focus away. The field now notes the sink while it has focus, like the other preference fields, and the #8220 test types into it. --------- Co-authored-by: MateuszSychXpl <151755914+MateuszSychXpl@users.noreply.github.com>
This commit is contained in:
@@ -0,0 +1,26 @@
|
||||
{
|
||||
"en": {
|
||||
"ai.prefs.instructions": "Custom instructions",
|
||||
"ai.prefs.instructions_note": "Sent to the agent at the start of every chat, after SolidCraft's own instructions: how you like to work, rules to keep to. A change applies from the next chat (Connect or New chat)."
|
||||
},
|
||||
"ar": {
|
||||
"ai.prefs.instructions": "تعليمات مخصصة",
|
||||
"ai.prefs.instructions_note": "تُرسَل إلى الوكيل في بداية كل محادثة، بعد تعليمات SolidCraft الخاصة: طريقتك المفضلة في العمل والقواعد التي يجب الالتزام بها. يسري أي تغيير من المحادثة التالية (اتصال أو محادثة جديدة)."
|
||||
},
|
||||
"he": {
|
||||
"ai.prefs.instructions": "הוראות מותאמות אישית",
|
||||
"ai.prefs.instructions_note": "נשלחות לסוכן בתחילת כל צ'אט, אחרי ההוראות של SolidCraft עצמה: דרך העבודה המועדפת עליך וכללים שיש להקפיד עליהם. שינוי חל מהצ'אט הבא (התחברות או צ'אט חדש)."
|
||||
},
|
||||
"ja": {
|
||||
"ai.prefs.instructions": "カスタム指示",
|
||||
"ai.prefs.instructions_note": "各チャットの開始時に、SolidCraft 自身の指示の後に続けてエージェントへ送られます。作業の進め方や守ってほしいルールを書いてください。変更は次のチャット(接続または新しいチャット)から適用されます。"
|
||||
},
|
||||
"vi": {
|
||||
"ai.prefs.instructions": "Hướng dẫn tùy chỉnh",
|
||||
"ai.prefs.instructions_note": "Được gửi cho agent khi bắt đầu mỗi cuộc trò chuyện, sau hướng dẫn riêng của SolidCraft: cách bạn muốn làm việc, các quy tắc cần tuân theo. Thay đổi có hiệu lực từ cuộc trò chuyện tiếp theo (Kết nối hoặc Cuộc trò chuyện mới)."
|
||||
},
|
||||
"zh": {
|
||||
"ai.prefs.instructions": "自定义指令",
|
||||
"ai.prefs.instructions_note": "在每次对话开始时发送给代理,位于 SolidCraft 自身的指令之后:你喜欢的工作方式、需要遵守的规则。更改从下一次对话(连接或新对话)起生效。"
|
||||
}
|
||||
}
|
||||
@@ -235,6 +235,21 @@ each stage compare `measure` with them. Prefer one `batch` over many small calls
|
||||
the chat, and each message allows a limited number of changes. Don't use shell or file tools: work through the SolidCraft tools. \
|
||||
Keep answers short.";
|
||||
|
||||
/// The most of the user's custom instructions (#8220) a session gets, in characters: the
|
||||
/// Preferences field takes no more, and a longer text from an edited settings file is cut here.
|
||||
pub const INSTRUCTIONS_LIMIT: usize = 4000;
|
||||
|
||||
/// [`SYSTEM_NOTE`] with the user's custom instructions (#8220) after it. Blank instructions leave
|
||||
/// the note as it is.
|
||||
pub fn system_note(instructions: &str) -> String {
|
||||
let text: String = instructions.trim().chars().take(INSTRUCTIONS_LIMIT).collect();
|
||||
let text = text.trim_end();
|
||||
if text.is_empty() {
|
||||
return SYSTEM_NOTE.to_string();
|
||||
}
|
||||
format!("{SYSTEM_NOTE}\n\nThe user's custom instructions, set in SolidCraft's preferences:\n{text}")
|
||||
}
|
||||
|
||||
/// The Claude Code adapter's own tools that reach the computer rather than the design: turned off
|
||||
/// for the session (`_meta.claudeCode.options.disallowedTools`), so it works through SolidCraft's.
|
||||
pub const CLAUDE_DISALLOWED_TOOLS: &[&str] =
|
||||
@@ -242,11 +257,12 @@ pub const CLAUDE_DISALLOWED_TOOLS: &[&str] =
|
||||
|
||||
/// The `session/new` parameters. `init` is the agent's `initialize` reply: agents read the system
|
||||
/// note from different places (#9811), and the standard field may only go to an agent offering it.
|
||||
pub fn session_new_params(cwd: &std::path::Path, bridge: Option<&McpBridge>, init: &Value) -> Value {
|
||||
/// `note` is the system note ([`system_note`]).
|
||||
pub fn session_new_params(cwd: &std::path::Path, bridge: Option<&McpBridge>, init: &Value, note: &str) -> Value {
|
||||
let servers: Vec<Value> = bridge.map(McpBridge::to_json).into_iter().collect();
|
||||
let mut meta = json!({
|
||||
// The Claude adapter's own key, for its versions without the standard field.
|
||||
"systemPrompt": {"append": SYSTEM_NOTE},
|
||||
"systemPrompt": {"append": note},
|
||||
// Hosts may opt a session out of "bypass permissions" (the Claude adapter reads it):
|
||||
// every tool call the agent asks about reaches the panel.
|
||||
// Its own shell and file tools are off: the agent works through SolidCraft's tools.
|
||||
@@ -255,7 +271,7 @@ pub fn session_new_params(cwd: &std::path::Path, bridge: Option<&McpBridge>, ini
|
||||
if takes_antigravity_note(init)
|
||||
&& let Some(meta) = meta.as_object_mut()
|
||||
{
|
||||
meta.insert("antigravity".into(), json!({"systemPrompt": {"append": SYSTEM_NOTE}}));
|
||||
meta.insert("antigravity".into(), json!({"systemPrompt": {"append": note}}));
|
||||
}
|
||||
let mut params = json!({
|
||||
"cwd": cwd.to_string_lossy(),
|
||||
@@ -265,7 +281,7 @@ pub fn session_new_params(cwd: &std::path::Path, bridge: Option<&McpBridge>, ini
|
||||
if takes_standard_note(init)
|
||||
&& let Some(params) = params.as_object_mut()
|
||||
{
|
||||
params.insert("systemPrompt".into(), json!({"mode": "append", "content": [{"type": "text", "text": SYSTEM_NOTE}]}));
|
||||
params.insert("systemPrompt".into(), json!({"mode": "append", "content": [{"type": "text", "text": note}]}));
|
||||
}
|
||||
params
|
||||
}
|
||||
@@ -770,29 +786,31 @@ mod tests {
|
||||
#[test]
|
||||
fn the_session_gets_the_bridged_mcp_server_without_a_token_or_any_path() {
|
||||
let b = McpBridge { cli: "/opt/solidcraft/solidcraft-cli".into(), port: 41234 };
|
||||
let p = session_new_params(std::path::Path::new("/home/u/designs"), Some(&b), &json!({}));
|
||||
let p = session_new_params(std::path::Path::new("/home/u/designs"), Some(&b), &json!({}), SYSTEM_NOTE);
|
||||
let s = &p["mcpServers"][0];
|
||||
assert_eq!(s["name"], BRIDGE_NAME);
|
||||
assert_eq!(s["args"], json!(["mcp", "--connect", "127.0.0.1:41234"]));
|
||||
let text = p.to_string();
|
||||
assert!(!text.contains("allow-any-path") && !text.contains("TOKEN") && !text.contains("token"), "{text}");
|
||||
assert!(p["_meta"]["claudeCode"]["options"]["disallowedTools"].as_array().is_some_and(|t| t.iter().any(|x| x == "Bash")));
|
||||
assert_eq!(session_new_params(std::path::Path::new("/tmp"), None, &json!({}))["mcpServers"], json!([]));
|
||||
assert_eq!(session_new_params(std::path::Path::new("/tmp"), None, &json!({}), SYSTEM_NOTE)["mcpServers"], json!([]));
|
||||
}
|
||||
|
||||
/// #9811: the note goes by the routes an agent's `initialize` reply offers, and the standard
|
||||
/// field never goes to an agent that doesn't offer it, as the protocol asks.
|
||||
#[test]
|
||||
fn the_system_note_takes_every_route_the_agent_offers() {
|
||||
let params = |init: Value| session_new_params(std::path::Path::new("/tmp"), None, &init);
|
||||
let standard = json!({"mode": "append", "content": [{"type": "text", "text": SYSTEM_NOTE}]});
|
||||
// With the user's custom instructions (#8220), which go wherever the note goes.
|
||||
let note = system_note("Sketch on XZ with explicit plane vectors.");
|
||||
let params = |init: Value| session_new_params(std::path::Path::new("/tmp"), None, &init, ¬e);
|
||||
let standard = json!({"mode": "append", "content": [{"type": "text", "text": note}]});
|
||||
|
||||
let offers_standard = params(json!({"agentCapabilities": {"sessionCapabilities": {"systemPrompt": {}}}}));
|
||||
assert_eq!(offers_standard["systemPrompt"], standard);
|
||||
assert!(offers_standard["_meta"].get("antigravity").is_none());
|
||||
|
||||
let offers_antigravity = params(json!({"agentCapabilities": {"_meta": {"antigravity": {"systemPrompt": {"append": true}}}}}));
|
||||
assert_eq!(offers_antigravity["_meta"]["antigravity"]["systemPrompt"]["append"], SYSTEM_NOTE);
|
||||
assert_eq!(offers_antigravity["_meta"]["antigravity"]["systemPrompt"]["append"], note);
|
||||
assert!(offers_antigravity.get("systemPrompt").is_none());
|
||||
|
||||
for none in [
|
||||
@@ -808,11 +826,24 @@ mod tests {
|
||||
// Every session keeps the key older Claude adapters read, and the tools they turn off.
|
||||
for init in [json!({}), json!({"agentCapabilities": {"sessionCapabilities": {"systemPrompt": {}}}})] {
|
||||
let p = params(init);
|
||||
assert_eq!(p["_meta"]["systemPrompt"]["append"], SYSTEM_NOTE);
|
||||
assert_eq!(p["_meta"]["systemPrompt"]["append"], note);
|
||||
assert!(p["_meta"]["claudeCode"]["options"]["disallowedTools"].is_array());
|
||||
}
|
||||
}
|
||||
|
||||
/// #8220: the user's custom instructions follow the note, trimmed and cut at the limit; blank
|
||||
/// ones leave the note as it is.
|
||||
#[test]
|
||||
fn custom_instructions_follow_the_note_trimmed_and_capped() {
|
||||
assert_eq!(system_note(""), SYSTEM_NOTE);
|
||||
assert_eq!(system_note(" \n\t "), SYSTEM_NOTE);
|
||||
let note = system_note("\n Use M3 screws.\nPrint flat side down. \n");
|
||||
assert!(note.starts_with(SYSTEM_NOTE), "{note}");
|
||||
assert!(note.ends_with("preferences:\nUse M3 screws.\nPrint flat side down."), "{note}");
|
||||
let note = system_note(&"ż".repeat(INSTRUCTIONS_LIMIT + 10));
|
||||
assert_eq!(note.chars().filter(|c| *c == 'ż').count(), INSTRUCTIONS_LIMIT);
|
||||
}
|
||||
|
||||
/// #5546: an agent started through `wsl.exe` gets the session's paths as WSL shows them.
|
||||
#[test]
|
||||
fn an_agent_inside_wsl_gets_wsl_paths() {
|
||||
@@ -833,7 +864,7 @@ mod tests {
|
||||
assert_eq!(wsl(r"\\server\share\x"), r"\\server\share\x");
|
||||
|
||||
let b = McpBridge { cli: r"C:\Program Files\SolidCraft\solidcraft-cli.exe".into(), port: 41234 };
|
||||
let p = session_new_params(&wsl_path(std::path::Path::new(r"C:\Users\u\work")), Some(&b.in_wsl()), &json!({}));
|
||||
let p = session_new_params(&wsl_path(std::path::Path::new(r"C:\Users\u\work")), Some(&b.in_wsl()), &json!({}), SYSTEM_NOTE);
|
||||
assert_eq!(p["cwd"], "/mnt/c/Users/u/work");
|
||||
assert_eq!(p["mcpServers"][0]["command"], "/mnt/c/Program Files/SolidCraft/solidcraft-cli.exe");
|
||||
assert_eq!(p["mcpServers"][0]["args"], json!(["mcp", "--connect", "127.0.0.1:41234"]));
|
||||
|
||||
@@ -35,6 +35,8 @@ pub struct AcpBackend {
|
||||
bridge: Option<acp::McpBridge>,
|
||||
/// The agent's `initialize` reply: which routes the system note can take into a session.
|
||||
init: Value,
|
||||
/// The system note the next session gets ([`acp::system_note`]).
|
||||
note: String,
|
||||
next_id: u64,
|
||||
pending: HashMap<u64, Pending>,
|
||||
session: Option<String>,
|
||||
@@ -64,6 +66,7 @@ impl AcpBackend {
|
||||
cwd,
|
||||
bridge,
|
||||
init: Value::Null,
|
||||
note: acp::SYSTEM_NOTE.to_string(),
|
||||
next_id: 0,
|
||||
pending: HashMap::new(),
|
||||
session: None,
|
||||
@@ -91,7 +94,7 @@ impl AcpBackend {
|
||||
fn new_session(&mut self) {
|
||||
self.config_pending = None;
|
||||
self.pending.retain(|_, request| *request != Pending::SetConfig);
|
||||
let params = acp::session_new_params(&self.cwd, self.bridge.as_ref(), &self.init);
|
||||
let params = acp::session_new_params(&self.cwd, self.bridge.as_ref(), &self.init, &self.note);
|
||||
self.request(Pending::SessionNew, "session/new", params);
|
||||
}
|
||||
|
||||
@@ -288,6 +291,10 @@ impl ChatBackend for AcpBackend {
|
||||
self.new_session();
|
||||
}
|
||||
|
||||
fn set_instructions(&mut self, text: &str) {
|
||||
self.note = acp::system_note(text);
|
||||
}
|
||||
|
||||
fn answer_permission(&mut self, key: u64, option: Option<&str>) {
|
||||
if let Some(rpc) = self.permissions.remove(&key) {
|
||||
self.proc.send(&acp::permission_reply(rpc, option));
|
||||
@@ -362,7 +369,7 @@ pub fn test_agent(argv: &[String], cwd: &std::path::Path) -> Result<String, Test
|
||||
let init = wait_for(&p, 1);
|
||||
let session_cwd = if acp::runs_in_wsl(argv) { acp::wsl_path(cwd) } else { cwd.to_path_buf() };
|
||||
let result = init.and_then(|v| {
|
||||
p.send(&acp::request(2, "session/new", acp::session_new_params(&session_cwd, None, &v)));
|
||||
p.send(&acp::request(2, "session/new", acp::session_new_params(&session_cwd, None, &v, acp::SYSTEM_NOTE)));
|
||||
wait_for(&p, 2).map(|_| {
|
||||
let who = acp::agent_info(&v);
|
||||
crate::i18n::format("ui.ai_test_acp_ok", &[("agent", if who.is_empty() { "?" } else { &who })])
|
||||
|
||||
@@ -658,6 +658,7 @@ enum Cmd {
|
||||
SetModel(String),
|
||||
SetRequestTimeout(u64),
|
||||
SetReasoningEffort(String),
|
||||
SetInstructions(String),
|
||||
}
|
||||
|
||||
pub struct ApiBackend {
|
||||
@@ -696,6 +697,7 @@ impl ApiBackend {
|
||||
tools: Vec::new(),
|
||||
tool_names: Vec::new(),
|
||||
history: Vec::new(),
|
||||
note: super::acp::SYSTEM_NOTE.to_string(),
|
||||
session_id,
|
||||
gemini_sigs: Default::default(),
|
||||
events: ev_tx,
|
||||
@@ -737,6 +739,11 @@ impl ChatBackend for ApiBackend {
|
||||
let _ = c.send(Cmd::SetReasoningEffort(effort.to_string()));
|
||||
}
|
||||
}
|
||||
fn set_instructions(&mut self, text: &str) {
|
||||
if let Some(c) = &self.cmds {
|
||||
let _ = c.send(Cmd::SetInstructions(text.to_string()));
|
||||
}
|
||||
}
|
||||
fn new_session(&mut self) {
|
||||
// A running turn is cancelled; the worker resets the history once it has ended.
|
||||
self.cancel.store(true, Ordering::Relaxed);
|
||||
@@ -774,6 +781,8 @@ struct Worker {
|
||||
tool_names: Vec<String>,
|
||||
/// The conversation in the endpoint's shape, without the system note (sent separately).
|
||||
history: Vec<Value>,
|
||||
/// The system note every request carries ([`super::acp::system_note`]).
|
||||
note: String,
|
||||
/// OpenCode Go's stable session ID for this chat (one per chat, not per request).
|
||||
session_id: String,
|
||||
/// Thought signatures of Gemini function-call parts, by kept call id, for the
|
||||
@@ -828,6 +837,7 @@ impl Worker {
|
||||
// the endpoint's, not the user's, and must not silently override the choice.
|
||||
self.no_reasoning.set(false);
|
||||
}
|
||||
Cmd::SetInstructions(text) => self.note = super::acp::system_note(&text),
|
||||
Cmd::NewChat => {
|
||||
self.history.clear();
|
||||
// A new conversation gets a fresh Go session ID for routing and caching.
|
||||
@@ -1135,7 +1145,7 @@ impl Worker {
|
||||
match protocol {
|
||||
GoProtocol::Chat => self.body(),
|
||||
GoProtocol::Responses => {
|
||||
let mut body = json!({"model": self.cfg.model, "input": go_responses_input(super::acp::SYSTEM_NOTE, &self.history)});
|
||||
let mut body = json!({"model": self.cfg.model, "input": go_responses_input(&self.note, &self.history)});
|
||||
if let Some(tools) = self.go_tools(protocol) {
|
||||
body["tools"] = tools;
|
||||
}
|
||||
@@ -1145,7 +1155,7 @@ impl Worker {
|
||||
let mut body = json!({
|
||||
"model": self.cfg.model,
|
||||
"max_tokens": ANTHROPIC_MAX_TOKENS,
|
||||
"system": super::acp::SYSTEM_NOTE,
|
||||
"system": self.note,
|
||||
"messages": go_anthropic_messages(&self.history),
|
||||
});
|
||||
if let Some(tools) = self.go_tools(protocol) {
|
||||
@@ -1155,7 +1165,7 @@ impl Worker {
|
||||
}
|
||||
GoProtocol::Gemini => {
|
||||
let mut body = json!({"contents": go_gemini_contents(&self.history, &self.gemini_sigs.borrow())});
|
||||
body["systemInstruction"] = json!({"parts": [{"text": super::acp::SYSTEM_NOTE}]});
|
||||
body["systemInstruction"] = json!({"parts": [{"text": self.note}]});
|
||||
if let Some(tools) = self.go_tools(protocol) {
|
||||
body["tools"] = tools;
|
||||
}
|
||||
@@ -1180,7 +1190,7 @@ impl Worker {
|
||||
match self.cfg.protocol {
|
||||
Protocol::OpenAi => {
|
||||
let mut messages = Vec::with_capacity(self.history.len() + 1);
|
||||
messages.push(json!({"role": "system", "content": super::acp::SYSTEM_NOTE}));
|
||||
messages.push(json!({"role": "system", "content": self.note}));
|
||||
messages.extend(self.history.iter().cloned());
|
||||
let mut body = json!({"model": self.cfg.model, "messages": messages});
|
||||
if let Some(effort) = self.reasoning_effort() {
|
||||
@@ -1200,7 +1210,7 @@ impl Worker {
|
||||
let mut body = json!({
|
||||
"model": self.cfg.model,
|
||||
"max_tokens": ANTHROPIC_MAX_TOKENS,
|
||||
"system": super::acp::SYSTEM_NOTE,
|
||||
"system": self.note,
|
||||
"messages": &self.history,
|
||||
});
|
||||
if !self.tools.is_empty() {
|
||||
@@ -1998,7 +2008,8 @@ mod tests {
|
||||
}
|
||||
|
||||
/// New chat resets the model's history to the system note: the next request carries nothing
|
||||
/// of the conversation before it.
|
||||
/// of the conversation before it. The note carries the custom instructions set for each
|
||||
/// conversation (#8220): the first one's, then none after they were cleared.
|
||||
#[test]
|
||||
fn a_new_chat_resets_the_history_to_the_system_note() {
|
||||
let seen = Arc::new(std::sync::Mutex::new(Vec::<Value>::new()));
|
||||
@@ -2033,12 +2044,16 @@ mod tests {
|
||||
std::thread::sleep(Duration::from_millis(10));
|
||||
}
|
||||
};
|
||||
b.set_instructions("Use M3 screws.");
|
||||
run(&mut b, &|b| b.prompt(&[PromptPart::Text("one".into())]), 2);
|
||||
b.set_instructions("");
|
||||
run(&mut b, &|b| b.new_session(), 1);
|
||||
run(&mut b, &|b| b.prompt(&[PromptPart::Text("two".into())]), 1);
|
||||
let sent = seen.lock().unwrap();
|
||||
assert_eq!(sent[0]["messages"][0]["content"], super::super::acp::system_note("Use M3 screws."));
|
||||
let roles: Vec<&str> = sent[1]["messages"].as_array().unwrap().iter().map(|m| m["role"].as_str().unwrap_or("")).collect();
|
||||
assert_eq!(roles, ["system", "user"]);
|
||||
assert_eq!(sent[1]["messages"][0]["content"], super::super::acp::SYSTEM_NOTE);
|
||||
assert_eq!(sent[1]["messages"][1]["content"], "two");
|
||||
b.shutdown(true);
|
||||
}
|
||||
@@ -2767,6 +2782,7 @@ mod tests {
|
||||
tools: vec![json!({"type": "function", "function": {"name": "measure"}})],
|
||||
tool_names: vec!["measure".into()],
|
||||
history: vec![json!({"role": "user", "content": "hi"})],
|
||||
note: super::super::acp::SYSTEM_NOTE.to_string(),
|
||||
session_id: "test-session".to_string(),
|
||||
gemini_sigs: Default::default(),
|
||||
events,
|
||||
|
||||
@@ -144,6 +144,9 @@ pub trait ChatBackend {
|
||||
/// Start a fresh conversation on the running backend (no restart); [`Event::Ready`] follows.
|
||||
/// A message being answered is cancelled first.
|
||||
fn new_session(&mut self);
|
||||
/// The user's custom instructions (#8220), added to the system note of every session that
|
||||
/// starts after this call. The panel sets them before the first session and before New chat.
|
||||
fn set_instructions(&mut self, _text: &str) {}
|
||||
/// Answer an [`Event::Permission`] with one of its options (`None`: cancelled).
|
||||
fn answer_permission(&mut self, key: u64, option: Option<&str>);
|
||||
/// Ask the backend to sign in by itself with one of its methods (it opens the browser).
|
||||
|
||||
@@ -1041,7 +1041,8 @@ pub fn connect(app: &mut SolveApp, ctx: &egui::Context) {
|
||||
},
|
||||
};
|
||||
match started {
|
||||
Ok(b) => {
|
||||
Ok(mut b) => {
|
||||
b.set_instructions(&app.preferences.agent_instructions);
|
||||
app.preferences.agent_last = entry.id.clone();
|
||||
chat.backend = Some(b);
|
||||
chat.argv = entry.argv.clone();
|
||||
@@ -1489,6 +1490,8 @@ pub fn new_chat(app: &mut SolveApp) {
|
||||
chat.restore_mode = Some(chat.mode.clone()).filter(|m| !m.is_empty());
|
||||
chat.state = Some(State::Starting);
|
||||
if let Some(b) = chat.backend.as_mut() {
|
||||
// Instructions edited since the last session apply from this one on.
|
||||
b.set_instructions(&app.preferences.agent_instructions);
|
||||
b.new_session();
|
||||
}
|
||||
}
|
||||
@@ -2361,7 +2364,8 @@ fn set_config_option_checked(app: &mut SolveApp, id: &str, value: &str, listed_o
|
||||
}
|
||||
|
||||
/// `ui.agent_chat` on the control channel (tests and screenshots): `enabled`, `command`,
|
||||
/// `connect_at_start` and `open_at_start` (the launch preferences), `open`,
|
||||
/// `connect_at_start` and `open_at_start` (the launch preferences), `instructions` (the custom
|
||||
/// instructions, #8220), `open`,
|
||||
/// `side`, `minimized`, `width`, `refresh`, `connect`, `stop`, `attach` (a file's path, for the
|
||||
/// next message), `remove_file` (index), `send`, `attach_view`, `attach_selection`, `pick`,
|
||||
/// `permission` (an option id, or `"cancel"`), `approve` (`true`/`false`, the oldest held
|
||||
@@ -2379,6 +2383,10 @@ pub fn control(app: &mut SolveApp, ctx: &egui::Context, p: &Value) -> Result<Val
|
||||
if let Some(b) = p.get("open_at_start").and_then(Value::as_bool) {
|
||||
app.preferences.agent_chat_open_at_start = b;
|
||||
}
|
||||
if let Some(text) = p.get("instructions") {
|
||||
let Some(text) = text.as_str() else { return Err("instructions must be text".into()) };
|
||||
app.preferences.agent_instructions = text.chars().take(acp::INSTRUCTIONS_LIMIT).collect();
|
||||
}
|
||||
if let Some(m) = p.get("permissions") {
|
||||
let Some(mode) = m.as_str().and_then(agents::Mode::parse) else { return Err("permissions must be ask, edits or all".into()) };
|
||||
app.preferences.ai_permissions = mode;
|
||||
@@ -2630,6 +2638,7 @@ pub fn status(app: &SolveApp) -> Value {
|
||||
if let Some(state) = state.as_object_mut() {
|
||||
state.insert("modelsNote".into(), json!(c.models_note));
|
||||
state.insert("customModels".into(), json!(custom_models_accepted(app)));
|
||||
state.insert("instructions".into(), json!(app.preferences.agent_instructions));
|
||||
}
|
||||
state
|
||||
}
|
||||
|
||||
@@ -471,6 +471,7 @@ pub fn page(app: &mut SolveApp, ui: &mut egui::Ui) {
|
||||
}
|
||||
agents_section(app, ui);
|
||||
}
|
||||
instructions_section(app, ui);
|
||||
panel_section(app, ui);
|
||||
connecting_section(app, ui);
|
||||
});
|
||||
@@ -496,6 +497,21 @@ fn connecting_section(app: &mut SolveApp, ui: &mut egui::Ui) {
|
||||
});
|
||||
}
|
||||
|
||||
/// Custom instructions (#8220): text every new chat's system note carries.
|
||||
fn instructions_section(app: &mut SolveApp, ui: &mut egui::Ui) {
|
||||
ui_kit::settings_section(ui, "instructions", tr("ai.prefs.instructions"), |ui| {
|
||||
let id = egui::Id::new("prefs_agent_instructions");
|
||||
let text = &mut app.preferences.agent_instructions;
|
||||
let response = ui_kit::text_area(ui, tr("ai.prefs.instructions"), id, text, 6, acp::INSTRUCTIONS_LIMIT);
|
||||
if response.has_focus() {
|
||||
// Typed text belongs to the field, not to Preferences search.
|
||||
crate::prefs_search::note_key_sink(ui.ctx());
|
||||
}
|
||||
crate::scenario::publish_target("prefs:agent_instructions", response.rect);
|
||||
ui_kit::text(ui, tr("ai.prefs.instructions_note"), ui_kit::Text::Hint("instructions"));
|
||||
});
|
||||
}
|
||||
|
||||
fn permissions_section(app: &mut SolveApp, ui: &mut egui::Ui) {
|
||||
ui_kit::settings_section(ui, "permissions", tr("ui.ai_permissions"), |ui| {
|
||||
let mut selected = Mode::ALL.iter().position(|mode| *mode == app.preferences.ai_permissions).unwrap_or(0);
|
||||
@@ -1355,7 +1371,7 @@ mod card_tests {
|
||||
assert!(scope_visible, "the full localized permission scope remains visible without hovering, including with AI off");
|
||||
assert!(privacy_visible, "the localized privacy note remains visible without hovering, including with AI off");
|
||||
let mut previous: Option<f32> = None;
|
||||
for key in ["assistant_availability", "permissions", "agents", "panel", "assistant_connecting", "agent_pointer"] {
|
||||
for key in ["assistant_availability", "permissions", "agents", "instructions", "panel", "assistant_connecting", "agent_pointer"] {
|
||||
let section = scenario::rect_of(&format!("prefs:section:{key}")).unwrap();
|
||||
assert!((section.left() - body.left()).abs() <= 1.0 && (section.right() - body.right()).abs() <= 1.0);
|
||||
assert!((section.height() - (crate::theme::Tokens::get().control_height + crate::theme::Tokens::get().space_2)).abs() <= 1.0);
|
||||
|
||||
@@ -1614,7 +1614,7 @@ done
|
||||
}
|
||||
|
||||
/// #9811: an agent that offers the standard `systemPrompt` field gets the note in it, on the first
|
||||
/// session and on the one New chat opens.
|
||||
/// session and on the one New chat opens, with the custom instructions set before it (#8220).
|
||||
#[cfg(unix)]
|
||||
#[test]
|
||||
fn an_acp_agent_offering_the_standard_field_gets_the_system_note_in_every_session() {
|
||||
@@ -1647,6 +1647,7 @@ done
|
||||
}
|
||||
};
|
||||
until(&mut app, &|a| a.agent_chat.state() == State::Ready);
|
||||
app.preferences.agent_instructions = "Print flat side down.".into();
|
||||
new_chat(&mut app);
|
||||
until(&mut app, &|a| a.agent_chat.state() == State::Ready);
|
||||
let sessions: Vec<Value> = std::fs::read_to_string(&log)
|
||||
@@ -1656,13 +1657,88 @@ done
|
||||
.filter(|v| v["method"] == "session/new")
|
||||
.collect();
|
||||
assert_eq!(sessions.len(), 2);
|
||||
for session in &sessions {
|
||||
assert_eq!(session.pointer("/params/systemPrompt/content/0/text").and_then(Value::as_str), Some(acp::SYSTEM_NOTE));
|
||||
let notes = [acp::SYSTEM_NOTE.to_string(), acp::system_note("Print flat side down.")];
|
||||
for (session, note) in sessions.iter().zip(¬es) {
|
||||
assert_eq!(session.pointer("/params/systemPrompt/content/0/text").and_then(Value::as_str), Some(note.as_str()));
|
||||
}
|
||||
disconnect(&mut app);
|
||||
let _ = std::fs::remove_dir_all(&dir);
|
||||
}
|
||||
|
||||
/// #8220: Preferences → AI shows the Custom instructions field; `ui.agent_chat` sets them (cut at
|
||||
/// the limit) and reports them; Connect sends them in the endpoint's system note, and an edit
|
||||
/// reaches the conversation New chat opens.
|
||||
#[cfg(not(target_arch = "wasm32"))]
|
||||
#[test]
|
||||
fn custom_instructions_reach_the_system_note_of_each_new_session() {
|
||||
let seen = Arc::new(std::sync::Mutex::new(Vec::<Value>::new()));
|
||||
let log = seen.clone();
|
||||
let http: backend::HttpFn = Arc::new(move |r: &backend::HttpRequest| {
|
||||
if r.url.ends_with("/chat/completions") {
|
||||
log.lock().unwrap().push(serde_json::from_slice(r.body.as_deref().unwrap_or(b"null")).unwrap());
|
||||
return Ok((200, json!({"choices": [{"message": {"role": "assistant", "content": "ok"}}]}).to_string().into_bytes()));
|
||||
}
|
||||
Ok((200, b"{\"data\":[]}".to_vec()))
|
||||
});
|
||||
let mut h = crate::scenario::Harness::new();
|
||||
h.app.home.open = false;
|
||||
h.app.services.agent_http = Some(http);
|
||||
crate::prefs::open_at(&mut h.app, crate::prefs::PAGE_ASSISTANT);
|
||||
h.frames(4);
|
||||
assert!(crate::scenario::rect_of("prefs:agent_instructions").is_some(), "the field is on the AI page");
|
||||
h.app.preferences.ai_features = true;
|
||||
h.ctx.memory_mut(|m| m.request_focus(egui::Id::new("prefs_agent_instructions")));
|
||||
h.frames(2);
|
||||
h.step(&json!({"text": "M3"})).unwrap();
|
||||
h.frames(2);
|
||||
assert_eq!(h.app.preferences.agent_instructions, "M3", "typed text lands in the field");
|
||||
assert!(h.app.prefs_window.filter.is_empty(), "Preferences search does not take the typing");
|
||||
let long = "x".repeat(acp::INSTRUCTIONS_LIMIT + 5);
|
||||
assert_eq!(
|
||||
h.call("ui.agent_chat", json!({"instructions": long}))["result"]["instructions"].as_str().map(str::len),
|
||||
Some(acp::INSTRUCTIONS_LIMIT)
|
||||
);
|
||||
assert!(!h.call("ui.agent_chat", json!({"instructions": 3}))["error"].is_null());
|
||||
assert_eq!(h.call("ui.agent_chat", json!({"instructions": "Use M3 screws."}))["result"]["instructions"], "Use M3 screws.");
|
||||
|
||||
h.app.preferences.agents.push(agents::AgentConfig {
|
||||
id: "api-1".into(),
|
||||
kind: agents::Kind::Api,
|
||||
base_url: "http://localhost:11434/v1".into(),
|
||||
model: "m".into(),
|
||||
..Default::default()
|
||||
});
|
||||
h.app.agent_chat.agent_id = Some("api-1".into());
|
||||
let ctx = crate::fonts::test_context();
|
||||
let until = |app: &mut SolveApp, done: &dyn Fn(&SolveApp) -> bool| {
|
||||
let t = std::time::Instant::now();
|
||||
while !done(app) {
|
||||
assert!(t.elapsed() < std::time::Duration::from_secs(20), "timed out");
|
||||
poll(app, &ctx);
|
||||
std::thread::sleep(std::time::Duration::from_millis(10));
|
||||
}
|
||||
};
|
||||
let sent = |n: usize| {
|
||||
let seen = seen.clone();
|
||||
move |a: &SolveApp| a.agent_chat.state() == State::Ready && seen.lock().unwrap().len() == n
|
||||
};
|
||||
connect(&mut h.app, &ctx);
|
||||
until(&mut h.app, &|a| a.agent_chat.state() == State::Ready);
|
||||
send(&mut h.app, &ctx, "one".into());
|
||||
until(&mut h.app, &sent(1));
|
||||
// Editing the field leaves the running conversation's note alone.
|
||||
h.app.preferences.agent_instructions = "Print flat side down.".into();
|
||||
send(&mut h.app, &ctx, "two".into());
|
||||
until(&mut h.app, &sent(2));
|
||||
new_chat(&mut h.app);
|
||||
until(&mut h.app, &|a| a.agent_chat.state() == State::Ready);
|
||||
send(&mut h.app, &ctx, "three".into());
|
||||
until(&mut h.app, &sent(3));
|
||||
let notes: Vec<Value> = seen.lock().unwrap().iter().map(|body| body["messages"][0]["content"].clone()).collect();
|
||||
assert_eq!(notes, [acp::system_note("Use M3 screws."), acp::system_note("Use M3 screws."), acp::system_note("Print flat side down.")]);
|
||||
disconnect(&mut h.app);
|
||||
}
|
||||
|
||||
fn models_reply(current: &str) -> Event {
|
||||
let m = |id: &str| acp::ModelInfo { id: id.into(), name: id.to_uppercase(), description: String::new() };
|
||||
Event::Models { current: current.into(), available: vec![m("fast"), m("smart")] }
|
||||
|
||||
@@ -1557,6 +1557,9 @@ preference_settings! {
|
||||
agent_chat_window: Option<[f32; 4]> = None, PAGE_ASSISTANT;
|
||||
/// Changes the agent may make for one message; 0 is no limit.
|
||||
agent_chat_turn_limit: u32 = crate::agent_chat::gate::DEFAULT_TURN_LIMIT, PAGE_ASSISTANT;
|
||||
/// Custom instructions (#8220): text the agent chat adds to its system note when a session
|
||||
/// starts (`agent_chat::acp::system_note`). No page reset discards this user-authored text.
|
||||
agent_instructions: String = String::new();
|
||||
|
||||
// ---- Sketch / selection modifier bindings (classic defaults; alternate mapping is opt-in) ----
|
||||
/// Modifier that suppresses sketch auto-snap while drawing or dragging a sketch point.
|
||||
|
||||
@@ -428,6 +428,19 @@ The agent reaches the open design only through SolidCraft's own tools (the
|
||||
an attached selection (and are shown the same way), and the box is empty again afterwards.
|
||||
**Cancel** keeps what was picked before.
|
||||
|
||||
## Custom instructions
|
||||
|
||||
Preferences → AI → **Custom instructions** holds your own text for the agent: how you like to
|
||||
work, rules to keep to ("don't start a new design from the chat", "sketch on XZ with explicit
|
||||
plane vectors"). Every chat starts with it, after SolidCraft's own note about where the agent
|
||||
is, by the same route that note takes: the system prompt of an endpoint's requests, or the
|
||||
session an ACP agent opens. An agent that reads neither doesn't get it. A change applies from
|
||||
the next chat (**Connect** or **New chat**), not to the one running. Blank text adds nothing;
|
||||
at most 4000 characters are kept.
|
||||
|
||||
They are for this chat only: an MCP client you run yourself keeps its own instruction files,
|
||||
and one text applies to every design.
|
||||
|
||||
## Permissions
|
||||
|
||||
Preferences → AI → **Permissions** decides what the agent may do without asking, for every agent;
|
||||
|
||||
@@ -653,7 +653,7 @@ no box selection.
|
||||
| `engine.execute` → `browser.expand_all` | `path?`: occurrence ids; `component?`: id; omit both for the whole tree | Open every Browser row under that scope, including folders that start closed (Origin, Named Views, Features). Display only: not saved in the design, not an undo step. |
|
||||
| `engine.execute` → `browser.collapse_all` | same | Close those rows. The design row stays open when the whole tree is collapsed. |
|
||||
| `engine.execute` → `view.history_layout` | `width?`, `dock?`: `bottom`, `left`, `under_browser`, `collapsed?` | Resize/dock the history without opening a dialog or changing the design. Contents respond to measured width; browser-stacked history shares its width and splits the height at the remembered share (`browser_share`, half until the seam between the two cards is dragged). One folded header gives the expanded card the rest; two folded headers stay at the top. |
|
||||
| `ui.agent_chat` | `features?`, `toolbar_button?`, `connect_at_start?` and `open_at_start?` (connect the agent started last, and open the panel, when SolidCraft starts; both off by default, never with `--control`), `permissions?` (`ask`, `edits`, `all`), `add_agent?` ({`name`, `kind`: `acp`, `api` or `anthropic` (`api` is OpenAI-compatible, `anthropic` Anthropic-compatible), `command` or `base_url` + `model` + `request_timeout_s?` (seconds one model request may take, 300 by default, rounded and clamped to 30 to 3600)}; no keys), `add_dialog?` (`installed`, `custom`, `api`, `anthropic`, or `false` to close), `edit_dialog?` (the id of an agent you added: the same dialog filled with what it holds, as its **Edit** button opens it; the key is never shown), `default_agent?`, `agent?` (the conversation's agent, by id), `mode?` (the agent's own mode), `provider?` (which provider's group the model picker opens first, when the models come from several), `model?` (one of its models, or see below), `open?`, `side?` (`left`, `right`), `minimized?`, `popped?` (in its own window), `width?`, `picker?`, `refresh?`, `connect?`, `stop?`, `attach?`: a file's path (a picture, a text file or a CAD file, for the next message; CAD files are imported at once), `remove_file?` (index of a file on the next message), `send?`: text (files alone make a message too), `attach_view?`, `attach_selection?`, `pick?` (true: open the context picker), `permission?`: an option id or `"cancel"` (the newest question of the agent's own), `approve?`: true, false or `"always"` (the oldest request held for Allow/Deny), `cancel?` (Stop: the queue stays), `send_now?` (stop the running step, then send the queued messages), `remove_pending?` (index), `new_chat?` (clear the chat, new session), `undo?` (the last message's changes) | the chat's state: `features`, `agents` (found and added), `currentAgent`, `state` (off, starting, ready, busy, auth, exited, failed: the agent never got ready, and Connect reads Retry), `connectAtStart`, `openAtStart`, `lastAgent` (the agent started last), `popped`, `window` (its place and size), `agent`, `modes`, `mode`, `models`, `model`, `modelWarning` (the model capability warning text, or `null` when unknown or tool support is reported; checked asynchronously and never blocks sending), `turnWarning` (the latest completed turn's no-tool notice, otherwise `null`; cleared on the next turn or new chat), `providers` (`name` and `count` of each, empty with one provider), `provider` (the one listed), `authMethods`, `toolsPort`, `pid`, `waiting`, `turn`, `pending` (queued messages, with their `files`), `attachments` (the files on the next message, by name), `loading` (pictures still being read for it, by name), `stopped`, `detection` and the conversation (`items`). With `features` off only the settings apply. Desktop only; see [Agent chat](agent-chat.md) |
|
||||
| `ui.agent_chat` | `features?`, `toolbar_button?`, `connect_at_start?` and `open_at_start?` (connect the agent started last, and open the panel, when SolidCraft starts; both off by default, never with `--control`), `instructions?` (the custom instructions every new chat's system note carries, Preferences → AI; at most 4000 characters are kept), `permissions?` (`ask`, `edits`, `all`), `add_agent?` ({`name`, `kind`: `acp`, `api` or `anthropic` (`api` is OpenAI-compatible, `anthropic` Anthropic-compatible), `command` or `base_url` + `model` + `request_timeout_s?` (seconds one model request may take, 300 by default, rounded and clamped to 30 to 3600)}; no keys), `add_dialog?` (`installed`, `custom`, `api`, `anthropic`, or `false` to close), `edit_dialog?` (the id of an agent you added: the same dialog filled with what it holds, as its **Edit** button opens it; the key is never shown), `default_agent?`, `agent?` (the conversation's agent, by id), `mode?` (the agent's own mode), `provider?` (which provider's group the model picker opens first, when the models come from several), `model?` (one of its models, or see below), `open?`, `side?` (`left`, `right`), `minimized?`, `popped?` (in its own window), `width?`, `picker?`, `refresh?`, `connect?`, `stop?`, `attach?`: a file's path (a picture, a text file or a CAD file, for the next message; CAD files are imported at once), `remove_file?` (index of a file on the next message), `send?`: text (files alone make a message too), `attach_view?`, `attach_selection?`, `pick?` (true: open the context picker), `permission?`: an option id or `"cancel"` (the newest question of the agent's own), `approve?`: true, false or `"always"` (the oldest request held for Allow/Deny), `cancel?` (Stop: the queue stays), `send_now?` (stop the running step, then send the queued messages), `remove_pending?` (index), `new_chat?` (clear the chat, new session), `undo?` (the last message's changes) | the chat's state: `features`, `agents` (found and added), `currentAgent`, `state` (off, starting, ready, busy, auth, exited, failed: the agent never got ready, and Connect reads Retry), `connectAtStart`, `openAtStart`, `instructions`, `lastAgent` (the agent started last), `popped`, `window` (its place and size), `agent`, `modes`, `mode`, `models`, `model`, `modelWarning` (the model capability warning text, or `null` when unknown or tool support is reported; checked asynchronously and never blocks sending), `turnWarning` (the latest completed turn's no-tool notice, otherwise `null`; cleared on the next turn or new chat), `providers` (`name` and `count` of each, empty with one provider), `provider` (the one listed), `authMethods`, `toolsPort`, `pid`, `waiting`, `turn`, `pending` (queued messages, with their `files`), `attachments` (the files on the next message, by name), `loading` (pictures still being read for it, by name), `stopped`, `detection` and the conversation (`items`). With `features` off only the settings apply. Desktop only; see [Agent chat](agent-chat.md) |
|
||||
| `ui.resize` | `width`, `height` | |
|
||||
| `app.quit` | `force?` (quit without asking about unsaved designs) | `quitting` (the window closes), `asking` (a "save changes?" prompt is open) |
|
||||
|
||||
|
||||
Reference in New Issue
Block a user