Skip to content

Lightpanda: the headless browser designed for AI and automation

agent: keep the full tool list in prompt turns, fewer prompt-parse rules

Adrià Arrufatb5e939b54425
diff --git a/src/agent/Agent.zig b/src/agent/Agent.zig
index 1e1df0755..74b516b34 100644
--- a/src/agent/Agent.zig
+++ b/src/agent/Agent.zig
@@ -673,7 +673,7 @@ fn runTurn(self: *Agent, input: TurnInput) bool {
self.terminal.printInfo("(no response from model)", .{});
}
if (self.prompt_target) |target| {
- if (target.data) |d| self.printData(target.tool, d);
+ if (target.data) |d| self.printData(target.cmd.tool, d);
}
self.conversation.prune();
return true;
@@ -782,7 +782,7 @@ fn runRepl(self: *Agent) void {
continue :repl;
}
const prompt = promptCommandText(aa, p) catch continue :repl;
- self.prompt_target = .{ .tool = p.tool, .pinned = p.args, .arena = aa };
+ self.prompt_target = .{ .cmd = p, .arena = aa };
defer self.prompt_target = null;
_ = self.runTurn(.{ .prompt = prompt, .record_comment = line, .capture_for_save = true, .label = label });
},
@@ -1193,28 +1193,13 @@ const SavedCall = struct {
/// Active `/click the sign in…`-style turn: one target call, read-only tools only.
const PromptTarget = struct {
- tool: BrowserTool,
- pinned: ?std.json.Value,
+ cmd: Command.Prompt,
/// REPL line arena; outlives the turn.
arena: std.mem.Allocator,
done: bool = false,
data: ?[]const u8 = null,
};
-const prompt_read_tools = [_]BrowserTool{ .tree, .findElement, .nodeDetails, .interactiveElements, .detectForms, .markdown, .html };
-
-fn isPromptReadTool(tool: BrowserTool) bool {
- return std.mem.findScalar(BrowserTool, &prompt_read_tools, tool) != null;
-}
-
-fn promptTools(arena: std.mem.Allocator, target: BrowserTool) ![]const ProviderTool {
- const all = globalTools();
- const out = try arena.alloc(ProviderTool, prompt_read_tools.len + 1);
- for (prompt_read_tools, out[0..prompt_read_tools.len]) |t, *o| o.* = all[@backingInt(t)];
- out[prompt_read_tools.len] = all[@backingInt(target)];
- return out;
-}
-
fn urlArg(args: ?std.json.Value) ?std.json.Value {
const a = args orelse return null;
if (a != .object) return null;
@@ -1897,7 +1882,9 @@ fn processUserMessage(self: *Agent, input: TurnInput) !?[]const u8 {
ma,
.{ .context = @ptrCast(self), .callFn = handleToolCall },
.{
- .tools = if (self.prompt_target) |t| try promptTools(ma, t.tool) else globalTools(),
+ // The full list even in a prompt turn: a narrower one would change
+ // the cached prefix; `toolOutcome` refuses the rest.
+ .tools = globalTools(),
.before_request = .{ .context = @ptrCast(self), .callFn = beforeTurnRequest },
.after_response = .{ .context = @ptrCast(self), .callFn = afterResponse },
.max_turns = 100,
@@ -2146,14 +2133,14 @@ fn toolOutcome(self: *Agent, allocator: std.mem.Allocator, tool_name: []const u8
if (self.prompt_target) |t| {
if (t.done) return .{ .content = "Error: the request is already done. Reply without calling more tools.", .is_error = true };
const tool = std.meta.stringToEnum(BrowserTool, tool_name);
- is_target = tool == t.tool;
+ is_target = tool == t.cmd.tool;
if (is_target) {
- if (t.pinned != null) {
- args = try pinArgs(self.conversation.arena.allocator(), args, t.pinned);
+ if (t.cmd.args != null) {
+ args = try pinArgs(self.conversation.arena.allocator(), args, t.cmd.args);
saved.args = args;
}
- } else if (tool == null or !isPromptReadTool(tool.?)) {
- return .{ .content = allocator.print("Error: only {s} and the read-only tools are available for this request.", .{@tagName(t.tool)}) catch "Error: tool not available", .is_error = true };
+ } else if (tool == null or tool.?.group() != .reading) {
+ return .{ .content = allocator.print("Error: only {s} and the read-only tools are available for this request.", .{@tagName(t.cmd.tool)}) catch "Error: tool not available", .is_error = true };
} else if (tool.?.navigatesToUrl() and urlArg(args) != null) {
return .{ .content = "Error: read the current page; this request must not navigate.", .is_error = true };
}
@@ -2170,7 +2157,7 @@ fn toolOutcome(self: *Agent, allocator: std.mem.Allocator, tool_name: []const u8
if (is_target and !result.is_error) {
const t = &self.prompt_target.?;
t.done = true;
- if (t.tool.producesData()) t.data = try t.arena.dupe(u8, result.text);
+ if (t.cmd.tool.producesData()) t.data = try t.arena.dupe(u8, result.text);
}
const content = capToolOutput(allocator, tool_name, result.text);
return .{
diff --git a/src/script/Schema.zig b/src/script/Schema.zig
index c588ecaa9..65df02a4e 100644
--- a/src/script/Schema.zig
+++ b/src/script/Schema.zig
@@ -264,9 +264,8 @@ pub fn parsePrompt(self: Schema, arena: std.mem.Allocator, rest_raw: []const u8,
const text = std.mem.trimEnd(u8, rest[0..tail_start], &std.ascii.whitespace);
if (text.len == 0) return null;
- const whole_quoted = (text[0] == '\'' or text[0] == '"') and quotedSpanEnd(text, 0) == text.len;
return .{
- .prompt = if (whole_quoted) stripQuotes(text) else text,
+ .prompt = text,
.args = try self.parseArgs(arena, rest[tail_start..], diag, false),
};
}
@@ -485,13 +484,13 @@ fn buildOne(arena: std.mem.Allocator, tool: BrowserTool, td: BrowserTool.Definit
}
}
- info.hints = try buildHints(arena, info.required, info.fields, takesPrompt(tool));
+ info.hints = try buildHints(arena, info.required, info.fields, info.leadingPositionalField());
std.debug.assert(info.hints.len <= max_hint_slots);
return info;
}
-fn buildHints(arena: std.mem.Allocator, required: []const []const u8, fields: []const FieldEntry, keyword_only: bool) ![]const HintSlot {
+fn buildHints(arena: std.mem.Allocator, required: []const []const u8, fields: []const FieldEntry, positional: ?[]const u8) ![]const HintSlot {
if (fields.len == 0 and required.len == 0) return &.{};
var optional_count: usize = 0;
for (fields) |f| {
@@ -504,7 +503,7 @@ fn buildHints(arena: std.mem.Allocator, required: []const []const u8, fields: []
out[idx] = .{
.name = name,
.required = true,
- .fragment = if (keyword_only) try arena.print("{s}=…", .{name}) else try arena.print("<{s}>", .{name}),
+ .fragment = if (positional != null and std.mem.eql(u8, name, positional.?)) try arena.print("<{s}>", .{name}) else try arena.print("{s}=…", .{name}),
};
idx += 1;
}
@@ -958,7 +957,7 @@ test "parsePrompt: splits the prompt from the key=value tail" {
try testing.expectString("the email field", p.prompt);
try testing.expectString("a b", p.args.?.object.get("value").?.string);
- try testing.expectString("the email field", (try fill.parsePrompt(aa, "\"the email field\"", null)).?.prompt);
+ try testing.expectString("\"the email field\"", (try fill.parsePrompt(aa, "\"the email field\"", null)).?.prompt);
try testing.expectString("'a' then b", (try fill.parsePrompt(aa, "'a' then b", null)).?.prompt);
try testing.expect((try fill.parsePrompt(aa, "selector=#e value=x", null)) == null);
@@ -970,7 +969,7 @@ test "parsePrompt: splits the prompt from the key=value tail" {
try testing.expectString("valu", diag.bad_field);
const click = Schema.find(Schema.all(), "click").?;
- try testing.expectString("the item with id=5", (try click.parsePrompt(aa, "\"the item with id=5\"", null)).?.prompt);
+ try testing.expectString("\"the item with id=5\"", (try click.parsePrompt(aa, "\"the item with id=5\"", null)).?.prompt);
try testing.expectString("the user's menu", (try click.parsePrompt(aa, "the user's menu", null)).?.prompt);
const cause = (try fill.parsePrompt(aa, "'cause value=x", null)).?;
try testing.expectString("'cause", cause.prompt);
diff --git a/src/script/command.zig b/src/script/command.zig
index 90418ff4e..9a34ccc15 100644
--- a/src/script/command.zig
+++ b/src/script/command.zig
@@ -401,7 +401,7 @@ test "parse: leading text on a prompt tool is a prompt, key=value stays literal"
const quoted = (try Command.parse(aa, "/click 'Login'")).prompt;
try testing.expectEqual(BrowserTool.click, quoted.tool);
- try testing.expectString("Login", quoted.text);
+ try testing.expectString("'Login'", quoted.text);
try testing.expect(quoted.args == null);
try testing.expectString("the user's menu", (try Command.parse(aa, "/click the user's menu")).prompt.text);