mirror of
https://github.com/lightpanda-io/browser.git
synced 2026-06-11 01:25:53 -04:00
b141da30ca
Introduces a multi-step synthesis process for `/save` that derives a logical JSON output schema and uses a dry-run runtime to verify candidate scripts. The LLM can now run and self-correct its scripts using a new `run_script` tool before finalizing the save.
1933 lines
83 KiB
Zig
1933 lines
83 KiB
Zig
// Copyright (C) 2023-2026 Lightpanda (Selecy SAS)
|
||
//
|
||
// Francis Bouvier <francis@lightpanda.io>
|
||
// Pierre Tachoire <pierre@lightpanda.io>
|
||
//
|
||
// This program is free software: you can redistribute it and/or modify
|
||
// it under the terms of the GNU Affero General Public License as
|
||
// published by the Free Software Foundation, either version 3 of the
|
||
// License, or (at your option) any later version.
|
||
//
|
||
// This program is distributed in the hope that it will be useful,
|
||
// but WITHOUT ANY WARRANTY; without even the implied warranty of
|
||
// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
||
// GNU Affero General Public License for more details.
|
||
//
|
||
// You should have received a copy of the GNU Affero General Public License
|
||
// along with this program. If not, see <https://www.gnu.org/licenses/>.
|
||
|
||
const std = @import("std");
|
||
const zenai = @import("zenai");
|
||
const lp = @import("lightpanda");
|
||
const browser_tools = lp.tools;
|
||
const BrowserTool = browser_tools.Tool;
|
||
const ProviderTool = zenai.provider.Tool;
|
||
|
||
const log = lp.log;
|
||
const Config = lp.Config;
|
||
const Command = lp.Command;
|
||
const Schema = lp.Schema;
|
||
const Recorder = lp.Recorder;
|
||
const ScriptRuntime = lp.Runtime;
|
||
const Credentials = zenai.provider.Credentials;
|
||
|
||
const App = @import("../App.zig");
|
||
const CDPNode = @import("../cdp/Node.zig");
|
||
const Conversation = @import("Conversation.zig");
|
||
const Terminal = @import("Terminal.zig");
|
||
const SlashCommand = @import("SlashCommand.zig");
|
||
const settings = @import("settings.zig");
|
||
const welcome = @import("welcome.zig");
|
||
const string = @import("../string.zig");
|
||
|
||
const Agent = @This();
|
||
|
||
/// Raised by init/listModels after they've printed a user-facing message to
|
||
/// stderr; callers should exit non-zero without logging more.
|
||
pub const UserError = error{
|
||
MissingApiKey,
|
||
MissingProvider,
|
||
ConflictingFlags,
|
||
ModelNotAvailable,
|
||
};
|
||
|
||
pub fn isUserError(err: anyerror) bool {
|
||
inline for (@typeInfo(UserError).error_set.?) |e| {
|
||
if (err == @field(anyerror, e.name)) return true;
|
||
}
|
||
return false;
|
||
}
|
||
|
||
const default_system_prompt = browser_tools.driver_guidance ++
|
||
\\
|
||
\\Agent-specific behavior:
|
||
\\- Call a tool for every browser action. NEVER claim you performed an
|
||
\\ action, visited a page, or saw content without the corresponding tool
|
||
\\ call. If a task needs a capability Lightpanda lacks (images, PDFs,
|
||
\\ audio), say so rather than improvising.
|
||
\\- Verify before answering: when a task asks for a specific value, ranked
|
||
\\ list, or comparison, and your first source is ambiguous, incomplete,
|
||
\\ or the answer is non-obvious, cross-check on ONE more authoritative
|
||
\\ source before committing. For multi-candidate questions (yes/no,
|
||
\\ A/B/C, pick-N), commit to a choice — don't abstain when you have data
|
||
\\ to reason from.
|
||
\\- If the user asks for account-scoped data (karma, profile, inbox, …)
|
||
\\ and the page shows you're not signed in, log in proactively (per
|
||
\\ the Credentials section above) before reporting unavailable.
|
||
;
|
||
|
||
const synthesis_prompt =
|
||
\\You have used your tool budget or cannot finish the exploration.
|
||
\\Give your best final answer NOW based ONLY on what you actually observed
|
||
\\via tool calls in this conversation. Do NOT fall back to prior knowledge —
|
||
\\if your snapshots show only cookie banners, 403/access-denied pages,
|
||
\\blocked search results, or empty bodies, say that explicitly
|
||
\\(e.g. "the page was blocked by a cookie wall and I could not extract X").
|
||
\\Do not invent details that are not visible in the tool outputs above.
|
||
\\Do not call any more tools.
|
||
\\Respond with ONLY the answer — one word, one number, one short phrase,
|
||
\\or a brief honest explanation of why the page could not be read.
|
||
\\No prefix, no markdown.
|
||
;
|
||
|
||
allocator: std.mem.Allocator,
|
||
ai_client: ?zenai.provider.Client,
|
||
model_credentials: ?Credentials,
|
||
/// True when the no-LLM state is a persisted preference (remembered null
|
||
/// provider or runtime `/provider null`), so `reportSaved` writes
|
||
/// `provider = null`. A transient `--no-llm` run leaves it false so saving
|
||
/// other settings doesn't clobber the remembered provider.
|
||
no_llm_persisted: bool,
|
||
model_base_url: ?[:0]const u8,
|
||
/// Cached chat-model ids for the current provider, backed by
|
||
/// `model_completion_arena`; invalidated on `/provider` switch.
|
||
model_completions: ?ModelCompletions,
|
||
model_completion_arena: std.heap.ArenaAllocator,
|
||
notification: *lp.Notification,
|
||
browser: lp.Browser,
|
||
session: *lp.Session,
|
||
node_registry: CDPNode.Registry,
|
||
terminal: Terminal,
|
||
save_buffer: Recorder,
|
||
save_path: ?[]u8,
|
||
/// Backs `last_extract_json`; reset alongside `save_buffer`.
|
||
last_extract_arena: std.heap.ArenaAllocator,
|
||
/// The JSON the most recent successful `extract` returned this session — the
|
||
/// real data `/save` grounds and verifies its synthesized script against.
|
||
last_extract_json: ?[]const u8 = null,
|
||
/// Set for the duration of an LLM `/save` so the `run_script` tool can reach
|
||
/// the dry-run runtime it executes candidates on.
|
||
active_verify: ?*Verify = null,
|
||
script_runtime_mutex: std.Thread.Mutex = .{},
|
||
active_script_runtime: ?*ScriptRuntime = null,
|
||
conversation: Conversation,
|
||
model: []u8,
|
||
/// Per-turn reasoning budget for LLM turns. Mutable at runtime via `/effort`.
|
||
effort: Config.Effort,
|
||
script_file: ?[]const u8,
|
||
one_shot_task: ?[]const u8,
|
||
one_shot_attachments: ?[]const []const u8,
|
||
cancel_requested: std.atomic.Value(bool) = .init(false),
|
||
/// Shuts down the in-flight LLM socket on Ctrl-C so an agent turn aborts
|
||
/// mid-request instead of blocking until the model's full response arrives.
|
||
http_interrupt: zenai.http.Interrupt = .{},
|
||
synthetic_tool_call_id: u32 = 0,
|
||
/// Aggregate Anthropic/OpenAI/Gemini token usage across every model call.
|
||
/// Printed as a structured `$usage ...` line on stderr at the end of `--task`
|
||
/// (one-shot) mode so wrappers can capture per-task cost.
|
||
total_usage: zenai.provider.Usage = .{},
|
||
/// Set when the last turn ended in a model refusal (safety stop).
|
||
last_turn_refused: bool = false,
|
||
available_providers: []const []const u8,
|
||
|
||
pub fn init(allocator: std.mem.Allocator, app: *App, opts: Config.Agent) !*Agent {
|
||
var providers_buf: [@typeInfo(Config.AiProvider).@"enum".fields.len]Credentials = undefined;
|
||
const found_providers = settings.availableProviders(&providers_buf);
|
||
const available_providers = try allocator.alloc([]const u8, found_providers.len);
|
||
var provider_count: usize = 0;
|
||
errdefer {
|
||
for (available_providers[0..provider_count]) |p| allocator.free(p);
|
||
allocator.free(available_providers);
|
||
}
|
||
for (found_providers, 0..) |f, i| {
|
||
available_providers[i] = try allocator.dupe(u8, @tagName(f.provider));
|
||
provider_count = i + 1;
|
||
}
|
||
|
||
if (opts.task != null and opts.script_file != null) {
|
||
log.fatal(.app, "conflicting flags", .{
|
||
.hint = "--task runs a one-shot turn; drop the positional script or drop --task",
|
||
});
|
||
return error.ConflictingFlags;
|
||
}
|
||
if (opts.no_llm and opts.provider != null) {
|
||
log.warn(.app, "ignoring --provider", .{ .reason = "--no-llm takes precedence" });
|
||
}
|
||
if (opts.task == null and opts.attach.items.len > 0) {
|
||
log.warn(.app, "ignoring --attach", .{ .reason = "no --task; attachments are only consumed in one-shot mode" });
|
||
}
|
||
|
||
const is_one_shot = opts.task != null;
|
||
const will_repl = !is_one_shot and opts.script_file == null;
|
||
|
||
// Load remembered selection up front so a saved null provider can flip the
|
||
// REPL into basic mode before resolution. Pure script runs need nothing.
|
||
const remembered: ?settings.Remembered = if (will_repl or is_one_shot) settings.loadRemembered(allocator) else null;
|
||
defer if (remembered) |r| std.zon.parse.free(allocator, r);
|
||
|
||
// A remembered null provider means the user disabled the LLM via
|
||
// `/provider null`; honor it for the REPL only (one-shot --task and script
|
||
// runs always need a model). An explicit --provider overrides it.
|
||
const remembered_no_llm = will_repl and opts.provider == null and
|
||
remembered != null and remembered.?.provider == null;
|
||
|
||
// Basic-mode REPL (no LLM) must be opted into via --no-llm or a remembered
|
||
// null provider. Without it the REPL accepts natural language, so an absent
|
||
// API key would only surface at the first non-slash-command line — too late.
|
||
// Pure JavaScript script runs stay allowed: no REPL, no LLM.
|
||
const requires_llm = is_one_shot or (will_repl and !opts.no_llm and !remembered_no_llm);
|
||
|
||
// Skip resolve when no client is wanted — else resolveCredentials prints
|
||
// "No API key detected" for a run that does not need one.
|
||
const resolve = !opts.no_llm and requires_llm;
|
||
|
||
// Print the banner before provider resolution so it precedes any
|
||
// interactive "Select a provider" prompt. On error paths (missing key / no
|
||
// key detected) resolveCredentials prints its own message; banner skipped.
|
||
if (will_repl and (!resolve or settings.wouldResolve(allocator, opts, remembered))) {
|
||
welcome.print(resolve);
|
||
}
|
||
|
||
const resolved: ?settings.ResolvedProvider = if (resolve) try settings.resolveCredentials(allocator, opts, remembered, will_repl) else null;
|
||
const llm: ?Credentials = if (resolved) |r| r.credentials else null;
|
||
|
||
if (llm == null and requires_llm) {
|
||
if (opts.no_llm) {
|
||
std.debug.print("--no-llm forbids LLM use; drop it to run this mode.\n", .{});
|
||
}
|
||
return error.MissingProvider;
|
||
}
|
||
|
||
var model = try allocator.dupe(u8, settings.resolveModelName(opts, resolved, remembered));
|
||
errdefer allocator.free(model);
|
||
|
||
// The REPL skips this network round trip for snappy startup; an invalid
|
||
// model surfaces on the first turn instead.
|
||
if (llm) |l| if (!will_repl) {
|
||
const remembered_matches = remembered != null and remembered.?.provider == l.provider;
|
||
const explicit = opts.model != null or remembered_matches;
|
||
switch (try settings.reconcileModel(allocator, l, model, opts.base_url, explicit)) {
|
||
.use => |m| {
|
||
allocator.free(model);
|
||
model = m;
|
||
},
|
||
.abort => return error.ModelNotAvailable,
|
||
}
|
||
};
|
||
|
||
const effort = settings.resolveEffort(opts, remembered, will_repl);
|
||
const verbosity = settings.resolveVerbosity(opts, remembered);
|
||
|
||
if (resolved) |r| {
|
||
if (r.source == .picked) {
|
||
settings.saveRemembered(.{ .provider = r.credentials.provider, .model = model, .effort = effort, .verbosity = verbosity }) catch {};
|
||
}
|
||
// provider/model now live in the status bar; just space before the help
|
||
std.debug.print("\n", .{});
|
||
}
|
||
|
||
const notification: *lp.Notification = try .init(allocator);
|
||
errdefer notification.deinit();
|
||
|
||
const self = try allocator.create(Agent);
|
||
errdefer allocator.destroy(self);
|
||
|
||
const history_paths: ?Terminal.HistoryPaths = if (will_repl)
|
||
.{ .normal = ".lp-history", .js = ".lp-history.js" }
|
||
else
|
||
null;
|
||
|
||
self.* = .{
|
||
.allocator = allocator,
|
||
.ai_client = null,
|
||
.model_credentials = llm,
|
||
.no_llm_persisted = remembered_no_llm,
|
||
.model_base_url = opts.base_url,
|
||
.model_completions = null,
|
||
.model_completion_arena = .init(allocator),
|
||
.notification = notification,
|
||
.browser = undefined,
|
||
.session = undefined,
|
||
.node_registry = .init(allocator),
|
||
.terminal = .init(allocator, history_paths, verbosity, will_repl),
|
||
.save_buffer = .init(allocator),
|
||
.save_path = null,
|
||
.last_extract_arena = .init(allocator),
|
||
.conversation = .init(allocator, opts.system_prompt orelse default_system_prompt),
|
||
.model = model,
|
||
.effort = effort,
|
||
.script_file = opts.script_file,
|
||
.one_shot_task = opts.task,
|
||
.one_shot_attachments = if (opts.attach.items.len == 0) null else opts.attach.items,
|
||
.available_providers = available_providers,
|
||
};
|
||
errdefer self.node_registry.deinit();
|
||
errdefer self.terminal.deinit();
|
||
errdefer self.conversation.deinit();
|
||
self.terminal.installLogSink();
|
||
errdefer self.terminal.uninstallLogSink();
|
||
|
||
try self.browser.init(app, .{}, null);
|
||
errdefer self.browser.deinit();
|
||
|
||
try self.startSession();
|
||
|
||
self.ai_client = if (llm) |l| try zenai.provider.Client.init(allocator, l, .{ .base_url = opts.base_url, .retry_policy = .long_running }) else null;
|
||
errdefer if (self.ai_client) |c| c.deinit(allocator);
|
||
if (self.ai_client) |c| c.setInterrupt(&self.http_interrupt);
|
||
|
||
if (will_repl) {
|
||
self.terminal.attachCompleter();
|
||
self.terminal.completion_source = .{
|
||
.context = @ptrCast(self),
|
||
.providers = completionProviders,
|
||
.models = completionModels,
|
||
};
|
||
// The model-list cache fills lazily on the first `/model` completion,
|
||
// so startup never blocks on the network.
|
||
}
|
||
|
||
return self;
|
||
}
|
||
|
||
pub fn deinit(self: *Agent) void {
|
||
self.terminal.uninstallLogSink();
|
||
self.save_buffer.deinit();
|
||
self.last_extract_arena.deinit();
|
||
if (self.save_path) |p| self.allocator.free(p);
|
||
self.terminal.deinit();
|
||
self.conversation.deinit();
|
||
self.model_completion_arena.deinit();
|
||
self.node_registry.deinit();
|
||
self.browser.deinit();
|
||
self.notification.deinit();
|
||
if (self.ai_client) |ai_client| ai_client.deinit(self.allocator);
|
||
self.allocator.free(self.model);
|
||
for (self.available_providers) |p| self.allocator.free(p);
|
||
self.allocator.free(self.available_providers);
|
||
self.allocator.destroy(self);
|
||
}
|
||
|
||
/// Create a fresh browser session and wire its cancel hook back to this agent
|
||
/// so Ctrl-C aborts in-flight page work. Startup and `/reset`.
|
||
fn startSession(self: *Agent) !void {
|
||
self.session = try self.browser.newSession(self.notification);
|
||
self.session.cancel_hook = .{ .context = @ptrCast(self), .check = checkCancel };
|
||
}
|
||
|
||
// Compile-time constant; projected once per process to avoid rebuilding per call.
|
||
var global_tools_storage: [browser_tools.tool_defs.len]ProviderTool = undefined;
|
||
var global_tools_once = std.once(initGlobalTools);
|
||
|
||
fn initGlobalTools() void {
|
||
for (Schema.all(), 0..) |s, i| {
|
||
global_tools_storage[i] = .{ .name = s.tool_name, .description = s.description, .parameters = s.parameters };
|
||
}
|
||
}
|
||
|
||
fn globalTools() []const ProviderTool {
|
||
global_tools_once.call();
|
||
return global_tools_storage[0..browser_tools.tool_defs.len];
|
||
}
|
||
|
||
/// Called from the sighandler thread. Flips `cancel_requested` for the LLM
|
||
/// streaming/HTTP probe and any code polling `Session.isCancelled`, then asks
|
||
/// V8 to bail out of whatever JS is running. Both hooks are thread-safe
|
||
/// (`Env.terminate` takes a mutex); no terminal touches from this context.
|
||
pub fn requestCancel(self: *Agent) void {
|
||
self.cancel_requested.store(true, .release);
|
||
self.http_interrupt.fire();
|
||
{
|
||
self.script_runtime_mutex.lock();
|
||
defer self.script_runtime_mutex.unlock();
|
||
if (self.active_script_runtime) |runtime| {
|
||
runtime.terminate();
|
||
}
|
||
}
|
||
self.browser.env.terminate();
|
||
}
|
||
|
||
/// Lives in main's stack so it can be registered with the sighandler before the
|
||
/// agent thread exists. The agent attaches once constructed and detaches before
|
||
/// deinit, so the sighandler-thread listener can fire safely whether or not an
|
||
/// agent is currently up.
|
||
pub const SigBridge = struct {
|
||
agent: std.atomic.Value(?*Agent) = .init(null),
|
||
|
||
pub fn attach(self: *SigBridge, agent: *Agent) void {
|
||
self.agent.store(agent, .release);
|
||
}
|
||
|
||
pub fn detach(self: *SigBridge) void {
|
||
self.agent.store(null, .release);
|
||
}
|
||
|
||
pub fn onSignal(self: *SigBridge) void {
|
||
const a = self.agent.load(.acquire) orelse return;
|
||
a.requestCancel();
|
||
}
|
||
};
|
||
|
||
fn checkCancel(ctx: *anyopaque) bool {
|
||
const self: *Agent = @ptrCast(@alignCast(ctx));
|
||
return self.cancel_requested.load(.acquire);
|
||
}
|
||
|
||
/// Roll the agent back to `baseline` messages, clear the V8 termination flag,
|
||
/// drop the cancel signal, and surface `error.UserCancelled`. Caller handles
|
||
/// any spinner cleanup not already done on its path.
|
||
fn drainCancellation(self: *Agent, baseline: usize) error{UserCancelled} {
|
||
self.resetAfterCancel(baseline);
|
||
return error.UserCancelled;
|
||
}
|
||
|
||
/// The side effects of `drainCancellation` without surfacing the error, for
|
||
/// void callers (e.g. `/save` synthesis) that just need to clean up.
|
||
fn resetAfterCancel(self: *Agent, baseline: usize) void {
|
||
self.conversation.rollback(baseline);
|
||
self.browser.env.cancelTerminate();
|
||
self.cancel_requested.store(false, .release);
|
||
}
|
||
|
||
/// One agent turn: the prompt sent to the model, plus optional context — a
|
||
/// recorder comment to write before the turn, file attachments to bundle into
|
||
/// the first user message, and a display label used in error output.
|
||
pub const TurnInput = struct {
|
||
prompt: []const u8,
|
||
record_comment: ?[]const u8 = null,
|
||
capture_for_save: bool = false,
|
||
attachments: ?[]const []const u8 = null,
|
||
label: []const u8 = "Request",
|
||
};
|
||
|
||
/// Returns true on success.
|
||
pub fn run(self: *Agent) bool {
|
||
if (self.one_shot_task) |task| {
|
||
const ok = self.runTurn(.{
|
||
.prompt = task,
|
||
.attachments = self.one_shot_attachments,
|
||
});
|
||
self.printUsageSummary();
|
||
return ok;
|
||
}
|
||
if (self.script_file) |path| {
|
||
return self.runScript(path);
|
||
}
|
||
self.runRepl();
|
||
return true;
|
||
}
|
||
|
||
/// Print single-line cumulative token usage to stderr, so wrappers driving
|
||
/// `lightpanda agent --task ...` can capture per-task cost by `grep`-ing the
|
||
/// `$usage` prefix. Stable key=value format:
|
||
/// $usage prompt=N completion=N total=N cached=N cache_creation=N
|
||
/// Fields emit 0 when the provider didn't report them.
|
||
fn printUsageSummary(self: *Agent) void {
|
||
const u = self.total_usage;
|
||
std.debug.print(
|
||
"$usage prompt={d} completion={d} total={d} cached={d} cache_creation={d}\n",
|
||
.{
|
||
u.prompt_tokens orelse 0,
|
||
u.completion_tokens orelse 0,
|
||
u.total_tokens orelse 0,
|
||
u.cached_tokens orelse 0,
|
||
u.cache_creation_tokens orelse 0,
|
||
},
|
||
);
|
||
}
|
||
|
||
fn runTurn(self: *Agent, input: TurnInput) bool {
|
||
const text = self.processUserMessage(input) catch |err| switch (err) {
|
||
error.UnsupportedAttachment, error.AttachmentReadFailed => return false,
|
||
error.UserCancelled => {
|
||
self.terminal.printInfo("Interrupted.", .{});
|
||
self.conversation.prune();
|
||
return false;
|
||
},
|
||
else => {
|
||
self.terminal.printError("{s} failed: {s}", .{ input.label, @errorName(err) });
|
||
return false;
|
||
},
|
||
};
|
||
if (text) |t|
|
||
self.terminal.printAssistant(t)
|
||
else if (self.last_turn_refused)
|
||
self.terminal.printInfo("(model declined to respond — safety refusal)", .{})
|
||
else
|
||
self.terminal.printInfo("(no response from model)", .{});
|
||
self.conversation.prune();
|
||
return true;
|
||
}
|
||
|
||
fn runRepl(self: *Agent) void {
|
||
log.debug(.app, "tools loaded", .{ .count = globalTools().len });
|
||
|
||
if (self.ai_client != null) {
|
||
const a = Terminal.ansi;
|
||
std.debug.print(" model: {s}{s} {s}effort: {s}{s}{s}\n", .{ a.dim, self.model, a.reset, a.dim, @tagName(self.effort), a.reset });
|
||
}
|
||
|
||
repl: while (true) {
|
||
std.debug.print("\n", .{});
|
||
const line = Terminal.readLine("") orelse break;
|
||
defer Terminal.freeLine(line);
|
||
|
||
// Slash commands and idle Ctrl-C set the cancel flag without clearing
|
||
// V8's terminate state; drain both before the next turn.
|
||
if (self.cancel_requested.swap(false, .acq_rel)) {
|
||
self.browser.env.cancelTerminate();
|
||
}
|
||
|
||
const trimmed = std.mem.trim(u8, line, &std.ascii.whitespace);
|
||
if (trimmed.len == 0) {
|
||
self.terminal.clearPromptFrame();
|
||
continue;
|
||
}
|
||
std.debug.print("\n", .{});
|
||
|
||
var arena: std.heap.ArenaAllocator = .init(self.allocator);
|
||
defer arena.deinit();
|
||
const aa = arena.allocator();
|
||
|
||
if (self.terminal.jsMode()) {
|
||
// `line` keeps the `$LP_*` placeholder so the secret never reaches
|
||
// the recorder; only the evaluated copy is expanded.
|
||
const script = browser_tools.substituteEnvVars(aa, line) catch line;
|
||
const result = browser_tools.evalScript(aa, self.session, &self.node_registry, script) catch |err| {
|
||
self.terminal.printError("{s}", .{switch (err) {
|
||
error.OutOfMemory => "out of memory",
|
||
error.FrameNotLoaded => "no page loaded — run /goto <url> first (Esc exits JS mode)",
|
||
else => std.fmt.allocPrint(aa, "evaluate failed: {s}", .{@errorName(err)}) catch "evaluate failed",
|
||
}});
|
||
continue :repl;
|
||
};
|
||
// Surface console output: slash commands (and thus /consoleLogs)
|
||
// are unreachable in JS mode, so a console must echo logs itself.
|
||
const logs = std.mem.trimRight(u8, self.session.drainConsoleMessages(), "\n");
|
||
if (logs.len > 0) self.printData(logs);
|
||
if (result.is_error) {
|
||
self.terminal.printError("{s}", .{result.text});
|
||
} else {
|
||
self.printData(result.text);
|
||
self.recordSaveRaw(line);
|
||
}
|
||
continue :repl;
|
||
}
|
||
|
||
const slash_split: ?Schema.Split = Schema.parseSlashCommand(trimmed);
|
||
if (slash_split) |split| {
|
||
if (SlashCommand.findMeta(split.name)) |meta| {
|
||
if (self.handleMeta(aa, meta, split.rest)) break :repl;
|
||
continue :repl;
|
||
}
|
||
}
|
||
|
||
var diag: Schema.Diag = .{};
|
||
const cmd = Command.parseDiag(aa, line, &diag) catch |err| switch (err) {
|
||
error.NotASlashCommand => {
|
||
if (self.ai_client == null) {
|
||
self.terminal.printError("Basic REPL (LLM disabled) accepts only commands. Try /help, or " ++ llm_setup_hint ++ " to enable natural-language prompts.", .{});
|
||
continue :repl;
|
||
}
|
||
_ = self.runTurn(.{ .prompt = line, .record_comment = line, .capture_for_save = true });
|
||
continue :repl;
|
||
},
|
||
else => |e| {
|
||
const name = if (slash_split) |sp| sp.name else line;
|
||
self.terminal.printSlashParseError(e, name, &diag);
|
||
continue :repl;
|
||
},
|
||
};
|
||
|
||
if (cmd == .llm) {
|
||
var name_buf: [32]u8 = undefined;
|
||
const name = std.fmt.bufPrint(&name_buf, "/{s}", .{@tagName(cmd.llm)}) catch "/?";
|
||
if (!self.requireLlm(name)) continue :repl;
|
||
}
|
||
|
||
switch (cmd) {
|
||
.comment => continue :repl,
|
||
.llm => |lc| {
|
||
var label_buf: [32]u8 = undefined;
|
||
const label = std.fmt.bufPrint(&label_buf, "/{s}", .{@tagName(lc)}) catch "/?";
|
||
_ = self.runTurn(.{ .prompt = lc.prompt(), .record_comment = line, .capture_for_save = true, .label = label });
|
||
},
|
||
.tool_call => |tc| {
|
||
self.terminal.beginTool(tc.name(), slash_split.?.rest);
|
||
const result = self.runCommand(aa, cmd);
|
||
self.terminal.endTool();
|
||
self.printCommandResult(cmd, result);
|
||
if (!result.is_error) {
|
||
self.recordSaveCommand(cmd);
|
||
}
|
||
self.recordSlashToolCall(trimmed, tc.name(), tc.args, result) catch |err| {
|
||
self.terminal.printWarning("LLM conversation out of sync (/{s}: {s}); next prompt may not see this action", .{ tc.name(), @errorName(err) });
|
||
};
|
||
},
|
||
}
|
||
}
|
||
}
|
||
|
||
/// Handle a REPL-only meta slash command — not a tool slash command, never
|
||
/// reaches the browser tool dispatcher. Returns `true` if the user asked to quit.
|
||
fn handleMeta(self: *Agent, arena: std.mem.Allocator, meta: *const SlashCommand.MetaCommand, rest: []const u8) bool {
|
||
switch (meta.tag) {
|
||
.quit => return true,
|
||
.help => self.printSlashHelp(arena, rest),
|
||
.verbosity => self.setEnumOption("verbosity", &self.terminal.verbosity, rest),
|
||
.effort => self.setEnumOption("effort", &self.effort, rest),
|
||
.usage => self.handleUsage(),
|
||
.clear => self.handleClear(),
|
||
.reset => self.handleReset(),
|
||
.save => self.handleSave(arena, rest),
|
||
.load => self.handleLoad(rest),
|
||
.model => self.handleModel(arena, rest),
|
||
.provider => self.handleProvider(arena, rest),
|
||
}
|
||
return false;
|
||
}
|
||
|
||
/// Shared body of `/verbosity` and `/effort`: bare prints the current level; an
|
||
/// argument is parsed against the enum, stored in `target`, and persisted.
|
||
/// `name` drives the slash name, usage hint, and report label.
|
||
fn setEnumOption(self: *Agent, comptime name: []const u8, target: anytype, rest: []const u8) void {
|
||
const T = @typeInfo(@TypeOf(target)).pointer.child;
|
||
if (rest.len == 0) {
|
||
self.terminal.printInfo(name ++ ": {s}", .{@tagName(target.*)});
|
||
return;
|
||
}
|
||
const level = std.meta.stringToEnum(T, rest) orelse {
|
||
self.terminal.printError("usage: /" ++ name ++ " " ++ Config.tagHint(T) ++ " (got {s})", .{rest});
|
||
return;
|
||
};
|
||
target.* = level;
|
||
self.reportSaved(name, @tagName(level));
|
||
}
|
||
|
||
/// Print cumulative session token usage, broken down so the cache's effect is
|
||
/// visible — the REPL otherwise never surfaces the `$usage` line `--task`
|
||
/// prints. Reads `total_usage` (accumulated per turn by `processUserMessage`);
|
||
/// fresh/cache split semantics live on `Usage`.
|
||
fn handleUsage(self: *Agent) void {
|
||
const u = self.total_usage;
|
||
const input = u.inputTokens();
|
||
const output = u.completion_tokens orelse 0;
|
||
if (input == 0 and output == 0) {
|
||
self.terminal.printInfo("usage: no model turns yet this session", .{});
|
||
return;
|
||
}
|
||
self.terminal.printInfo(
|
||
"usage: input={d} (fresh={d} · cache read={d} · cache write={d}), output={d}",
|
||
.{ input, u.prompt_tokens orelse 0, u.cached_tokens orelse 0, u.cache_creation_tokens orelse 0, output },
|
||
);
|
||
if (input > 0) {
|
||
self.terminal.printInfo("cache: {d}% of input served from cache", .{u.cacheHitPercent()});
|
||
}
|
||
}
|
||
|
||
/// Drop everything tied to the conversation: history (system prompt re-seeds
|
||
/// lazily next turn), cumulative usage, the recorded action buffer, and DOM
|
||
/// node IDs. Shared by `/clear` and `/reset`.
|
||
fn clearConversation(self: *Agent) void {
|
||
self.conversation.rollback(0);
|
||
self.resetSaveBuffers();
|
||
self.total_usage = .{};
|
||
self.node_registry.reset();
|
||
}
|
||
|
||
/// Drop everything `/save` accumulates: the recorded action buffer and the
|
||
/// captured extract data that grounds synthesis.
|
||
fn resetSaveBuffers(self: *Agent) void {
|
||
self.save_buffer.reset();
|
||
_ = self.last_extract_arena.reset(.retain_capacity);
|
||
self.last_extract_json = null;
|
||
}
|
||
|
||
/// Forget the conversation while leaving the browser session live — loaded page
|
||
/// stays put, cookies/logins preserved.
|
||
fn handleClear(self: *Agent) void {
|
||
self.clearConversation();
|
||
self.terminal.printInfo("Cleared conversation, usage, and node IDs. Page and cookies kept.", .{});
|
||
}
|
||
|
||
/// Full clean slate: everything `/clear` drops, plus a fresh browser session,
|
||
/// so the loaded page, cookies, storage, and history are gone too.
|
||
fn handleReset(self: *Agent) void {
|
||
self.startSession() catch |err| {
|
||
self.terminal.printError("reset failed: {s}", .{@errorName(err)});
|
||
return;
|
||
};
|
||
self.clearConversation();
|
||
self.terminal.printInfo("Reset conversation and browser session. Page, cookies, and storage cleared.", .{});
|
||
}
|
||
|
||
fn handleLoad(self: *Agent, rest: []const u8) void {
|
||
const path = std.mem.trim(u8, rest, &std.ascii.whitespace);
|
||
if (path.len == 0) {
|
||
self.terminal.printError("usage: /load <path>", .{});
|
||
return;
|
||
}
|
||
_ = self.runScript(path);
|
||
}
|
||
|
||
const api_keys_hint = settings.api_keys_hint;
|
||
const llm_setup_hint = "set an API key (" ++ api_keys_hint ++ ") and run /provider <name>";
|
||
|
||
/// `/provider <keyword>` disables the LLM and persists it; shared by command
|
||
/// parser, autocomplete, and save report so they can't drift apart.
|
||
const provider_off_keyword = "null";
|
||
|
||
fn requireLlm(self: *Agent, name: []const u8) bool {
|
||
if (self.model_credentials == null) {
|
||
self.terminal.printError("{s} requires an LLM — " ++ llm_setup_hint ++ ".", .{name});
|
||
return false;
|
||
}
|
||
return true;
|
||
}
|
||
|
||
fn handleModel(self: *Agent, _: std.mem.Allocator, rest: []const u8) void {
|
||
if (!self.requireLlm("/model")) return;
|
||
|
||
const trimmed = std.mem.trim(u8, rest, &std.ascii.whitespace);
|
||
if (trimmed.len == 0) {
|
||
self.terminal.printInfo("Current model: {s} (Tab to list)", .{self.model});
|
||
return;
|
||
}
|
||
const ids = completionModels(self, self.allocator);
|
||
// Empty list = fetch failed or unlisted local models; can't confirm, allow.
|
||
if (ids.len != 0 and !string.isOneOf(trimmed, ids)) {
|
||
self.terminal.printError("unknown model: {s} (Tab to list)", .{trimmed});
|
||
return;
|
||
}
|
||
self.setModel(trimmed) catch |err| {
|
||
self.terminal.printError("failed to set model: {s}", .{@errorName(err)});
|
||
};
|
||
}
|
||
|
||
/// Persist provider/model/effort/verbosity to `.lp-agent.zon` and report it as
|
||
/// "<label>: <value>", appending "(saved to …)" on write success. With no model
|
||
/// credentials it persists `provider = null` only when that's an intentional
|
||
/// preference (`no_llm_persisted`); a transient --no-llm run reports without saving.
|
||
fn reportSaved(self: *Agent, label: []const u8, value: []const u8) void {
|
||
const provider: ?Config.AiProvider = if (self.model_credentials) |c| c.provider else null;
|
||
// A transient --no-llm run has no provider and no intent to persist one;
|
||
// report without saving so we don't clobber the remembered selection.
|
||
if (provider == null and !self.no_llm_persisted) {
|
||
self.terminal.printInfo("{s}: {s}", .{ label, value });
|
||
return;
|
||
}
|
||
if (settings.saveRemembered(.{ .provider = provider, .model = self.model, .effort = self.effort, .verbosity = self.terminal.verbosity })) {
|
||
self.terminal.printInfo("{s}: {s} (saved to {s})", .{ label, value, settings.remembered_path });
|
||
} else |_| {
|
||
self.terminal.printInfo("{s}: {s}", .{ label, value });
|
||
}
|
||
}
|
||
|
||
fn setModel(self: *Agent, model: []const u8) !void {
|
||
const new_model = try self.allocator.dupe(u8, model);
|
||
self.allocator.free(self.model);
|
||
self.model = new_model;
|
||
self.reportSaved("model", self.model);
|
||
}
|
||
|
||
fn handleProvider(self: *Agent, _: std.mem.Allocator, rest: []const u8) void {
|
||
const trimmed = std.mem.trim(u8, rest, &std.ascii.whitespace);
|
||
|
||
if (trimmed.len == 0) {
|
||
if (self.model_credentials) |c| {
|
||
self.terminal.printInfo("Current provider: {s} (Tab to list, /provider null to disable)", .{@tagName(c.provider)});
|
||
} else {
|
||
self.terminal.printInfo("Current provider: none — LLM disabled (/provider <name> to enable)", .{});
|
||
}
|
||
return;
|
||
}
|
||
|
||
if (std.mem.eql(u8, trimmed, provider_off_keyword)) {
|
||
self.disableProvider();
|
||
return;
|
||
}
|
||
|
||
const provider = std.meta.stringToEnum(Config.AiProvider, trimmed) orelse {
|
||
self.terminal.printError("unknown provider: {s} (or 'null' to disable the LLM)", .{trimmed});
|
||
return;
|
||
};
|
||
if (self.model_credentials) |current| if (provider == current.provider) {
|
||
self.terminal.printInfo("provider: {s}", .{@tagName(provider)});
|
||
return;
|
||
};
|
||
const key = zenai.provider.envApiKey(provider) orelse {
|
||
self.terminal.printError("no API key for {s}; set {s}", .{ @tagName(provider), zenai.provider.envVarName(provider) });
|
||
return;
|
||
};
|
||
self.setProvider(.{ .provider = provider, .key = key }) catch |err| {
|
||
self.terminal.printError("failed to set provider: {s}", .{@errorName(err)});
|
||
};
|
||
}
|
||
|
||
/// Tear down the LLM client and persist a null provider so the next REPL launch
|
||
/// starts in basic mode without re-prompting. Inverse of `setProvider`.
|
||
fn disableProvider(self: *Agent) void {
|
||
if (self.ai_client) |client| client.deinit(self.allocator);
|
||
self.ai_client = null;
|
||
self.model_credentials = null;
|
||
self.model_completions = null;
|
||
self.no_llm_persisted = true;
|
||
self.reportSaved("provider", provider_off_keyword);
|
||
}
|
||
|
||
fn setProvider(self: *Agent, credentials: Credentials) !void {
|
||
const new_client = try zenai.provider.Client.init(self.allocator, credentials, .{ .base_url = self.model_base_url, .retry_policy = .long_running });
|
||
errdefer new_client.deinit(self.allocator);
|
||
|
||
const new_model = try self.allocator.dupe(u8, zenai.provider.defaultModel(credentials.provider));
|
||
if (self.ai_client) |client| client.deinit(self.allocator);
|
||
new_client.setInterrupt(&self.http_interrupt);
|
||
self.ai_client = new_client;
|
||
self.model_credentials = credentials;
|
||
self.no_llm_persisted = false;
|
||
self.model_completions = null;
|
||
self.allocator.free(self.model);
|
||
self.model = new_model;
|
||
self.terminal.printInfo("provider: {s}", .{@tagName(credentials.provider)});
|
||
self.reportSaved("model", self.model);
|
||
_ = completionModels(self, self.allocator);
|
||
}
|
||
|
||
const SaveMode = enum { append, replace };
|
||
|
||
const PathAndMode = struct { path: []const u8, mode: SaveMode };
|
||
|
||
fn resolveSavePathAndMode(self: *Agent, arena: std.mem.Allocator, filename: ?[]const u8) ?PathAndMode {
|
||
if (self.save_path) |saved| {
|
||
if (filename) |name| {
|
||
if (!std.mem.eql(u8, saved, name)) {
|
||
self.terminal.printError("already saving to {s}; use /save without a filename to append to it", .{saved});
|
||
return null;
|
||
}
|
||
}
|
||
return .{ .path = saved, .mode = .append };
|
||
} else if (filename) |name| {
|
||
const exists = fileExists(name) catch |err| {
|
||
self.terminal.printError("failed to inspect {s}: {s}", .{ name, @errorName(err) });
|
||
return null;
|
||
};
|
||
const mode = if (exists)
|
||
self.promptSaveMode(name) orelse return null
|
||
else
|
||
.replace;
|
||
return .{ .path = name, .mode = mode };
|
||
} else {
|
||
const path = randomSaveFilename(arena) catch |err| {
|
||
self.terminal.printError("failed to choose save filename: {s}", .{@errorName(err)});
|
||
return null;
|
||
};
|
||
return .{ .path = path, .mode = .replace };
|
||
}
|
||
}
|
||
|
||
fn handleSave(self: *Agent, arena: std.mem.Allocator, rest: []const u8) void {
|
||
const parsed = parseSaveCommand(rest) catch |err| {
|
||
const msg: []const u8 = switch (err) {
|
||
error.UnterminatedQuote => "unterminated filename quote",
|
||
error.EmptyFilename => "filename cannot be empty",
|
||
error.InvalidFilename => "filename must be a local file name, not a path",
|
||
};
|
||
self.terminal.printError("{s}", .{msg});
|
||
return;
|
||
};
|
||
|
||
if (self.ai_client != null) {
|
||
self.synthesizeSave(arena, parsed.filename, parsed.prompt);
|
||
return;
|
||
}
|
||
|
||
if (parsed.prompt != null) {
|
||
self.terminal.printWarning("prompt ignored without an LLM; saving the recorded commands as-is", .{});
|
||
}
|
||
const resolved = self.resolveSavePathAndMode(arena, parsed.filename) orelse return;
|
||
const path = resolved.path;
|
||
const mode = resolved.mode;
|
||
|
||
// `path` aliases either an arena-owned string (first save) or
|
||
// `self.save_path` (subsequent saves to the same destination); only the
|
||
// former needs persisting into agent-owned memory.
|
||
var new_save_path: ?[]u8 = if (self.save_path == null)
|
||
self.allocator.dupe(u8, path) catch |err| {
|
||
self.terminal.printError("failed to remember save destination {s}: {s}", .{ path, @errorName(err) });
|
||
return;
|
||
}
|
||
else
|
||
null;
|
||
defer if (new_save_path) |p| self.allocator.free(p);
|
||
|
||
self.writeSaveFile(path, mode) catch |err| {
|
||
self.terminal.printError("failed to save {s}: {s}", .{ path, @errorName(err) });
|
||
return;
|
||
};
|
||
|
||
if (new_save_path) |p| {
|
||
self.save_path = p;
|
||
new_save_path = null;
|
||
}
|
||
const saved_lines = self.save_buffer.lines;
|
||
self.resetSaveBuffers();
|
||
self.terminal.printInfo("Saved {d} line(s) to {s}", .{ saved_lines, self.save_path.? });
|
||
}
|
||
|
||
const SaveCommand = struct { filename: ?[]const u8, prompt: ?[]const u8 };
|
||
|
||
/// Split `/save` arguments into an optional filename and an optional trailing
|
||
/// natural-language prompt. A quoted leading token is always a filename; an
|
||
/// unquoted one is a filename only if it ends in `.js` (else the whole argument
|
||
/// is the prompt, and a name is chosen automatically).
|
||
fn parseSaveCommand(rest: []const u8) !SaveCommand {
|
||
const trimmed = std.mem.trim(u8, rest, &std.ascii.whitespace);
|
||
if (trimmed.len == 0) return .{ .filename = null, .prompt = null };
|
||
|
||
if (trimmed[0] == '\'' or trimmed[0] == '"') {
|
||
const quote = trimmed[0];
|
||
const end = std.mem.indexOfScalarPos(u8, trimmed, 1, quote) orelse return error.UnterminatedQuote;
|
||
const name = trimmed[1..end];
|
||
try validateSaveFilename(name);
|
||
const rest_prompt = std.mem.trim(u8, trimmed[end + 1 ..], &std.ascii.whitespace);
|
||
return .{ .filename = name, .prompt = if (rest_prompt.len == 0) null else rest_prompt };
|
||
}
|
||
|
||
const tok_end = std.mem.indexOfAny(u8, trimmed, &std.ascii.whitespace) orelse trimmed.len;
|
||
const first = trimmed[0..tok_end];
|
||
if (std.mem.endsWith(u8, first, ".js")) {
|
||
try validateSaveFilename(first);
|
||
const rest_prompt = std.mem.trim(u8, trimmed[tok_end..], &std.ascii.whitespace);
|
||
return .{ .filename = first, .prompt = if (rest_prompt.len == 0) null else rest_prompt };
|
||
}
|
||
return .{ .filename = null, .prompt = trimmed };
|
||
}
|
||
|
||
fn validateSaveFilename(name: []const u8) !void {
|
||
if (name.len == 0) return error.EmptyFilename;
|
||
if (std.fs.path.isAbsolute(name)) return error.InvalidFilename;
|
||
if (std.mem.indexOfScalar(u8, name, '/') != null) return error.InvalidFilename;
|
||
if (std.mem.indexOfScalar(u8, name, '\\') != null) return error.InvalidFilename;
|
||
if (std.mem.eql(u8, name, ".") or std.mem.eql(u8, name, "..")) return error.InvalidFilename;
|
||
}
|
||
|
||
fn randomSaveFilename(arena: std.mem.Allocator) ![]const u8 {
|
||
for (0..100) |_| {
|
||
const n = std.crypto.random.int(u64);
|
||
const path = try std.fmt.allocPrint(arena, "session-{x}.js", .{n});
|
||
if (!(try fileExists(path))) return path;
|
||
}
|
||
return error.NameCollision;
|
||
}
|
||
|
||
fn fileExists(path: []const u8) !bool {
|
||
std.fs.cwd().access(path, .{}) catch |err| switch (err) {
|
||
error.FileNotFound => return false,
|
||
else => return err,
|
||
};
|
||
return true;
|
||
}
|
||
|
||
fn promptSaveMode(self: *Agent, path: []const u8) ?SaveMode {
|
||
var header_buf: [256]u8 = undefined;
|
||
const header = std.fmt.bufPrint(&header_buf, "{s} already exists. Pick save mode:", .{path}) catch
|
||
"File already exists. Pick save mode:";
|
||
const idx = Terminal.promptNumberedChoice(header, std.meta.fieldNames(SaveMode), 0) catch {
|
||
self.terminal.printInfo("Save cancelled.", .{});
|
||
return null;
|
||
};
|
||
return @enumFromInt(idx);
|
||
}
|
||
|
||
fn writeSaveFile(self: *Agent, path: []const u8, mode: SaveMode) !void {
|
||
return writeContentFile(path, self.save_buffer.bytes(), mode);
|
||
}
|
||
|
||
fn writeContentFile(path: []const u8, content: []const u8, mode: SaveMode) !void {
|
||
const file = try std.fs.cwd().createFile(path, .{ .truncate = mode == .replace });
|
||
defer file.close();
|
||
if (mode == .append) {
|
||
try file.seekFromEnd(0);
|
||
const pos = try file.getPos();
|
||
if (pos > 0 and content.len > 0) try file.writeAll("\n");
|
||
}
|
||
try file.writeAll(content);
|
||
if (content.len > 0 and content[content.len - 1] != '\n') try file.writeAll("\n");
|
||
}
|
||
|
||
fn failSave(self: *Agent, reason: []const u8) void {
|
||
self.terminal.printError("save failed: {s}", .{reason});
|
||
}
|
||
|
||
/// Roll the in-flight save turn back out of the conversation, then report the
|
||
/// failure — so a doomed `/save` synthesis never leaks its messages into history.
|
||
fn abortSave(self: *Agent, baseline: usize, reason: []const u8) void {
|
||
self.conversation.rollback(baseline);
|
||
self.failSave(reason);
|
||
}
|
||
|
||
/// In-flight `/save` verification harness: the dry-run runtime the `run_script`
|
||
/// tool executes candidates on, plus the last source it ran (a fallback script
|
||
/// if the model finishes the loop without re-emitting it as text).
|
||
const Verify = struct {
|
||
runtime: *ScriptRuntime,
|
||
last_source: ?[]const u8 = null,
|
||
};
|
||
|
||
/// Agent-only addendum (kept out of the shared `save_synthesis_prompt`) telling
|
||
/// the model to derive every value at runtime and check the result with run_script.
|
||
const save_verify_addendum =
|
||
\\Read data with the recorded extract(...), not evaluate() — extract can read a
|
||
\\card's whole text via an empty selector (""). Reshape its result in plain JS so the
|
||
\\completion value matches the schema exactly (same keys, parsed numbers); don't
|
||
\\return the raw extract or hard-code values.
|
||
\\Before finalizing, test with run_script: it runs your FULL script for real from a
|
||
\\blank page, so it must goto(...) first (missing goto → "no page loaded", a wrong
|
||
\\selector → null). Confirm every field is populated, then reply with ONLY the final
|
||
\\JavaScript source.
|
||
;
|
||
|
||
/// Cap on the captured extract sample shown in the synthesis prompt (the full
|
||
/// data still feeds the dry run); keeps a large result from dominating context.
|
||
const save_sample_cap = 8 * 1024;
|
||
|
||
/// LLM-synthesized `/save`. Pin the output shape first — derive the session's
|
||
/// intent, then a typed output schema from it — so the script's result shape is
|
||
/// stable across runs, then synthesize the script honoring that schema. Each
|
||
/// step degrades gracefully: a null schema falls back to plain synthesis.
|
||
fn synthesizeSave(self: *Agent, arena: std.mem.Allocator, filename: ?[]const u8, prompt: ?[]const u8) void {
|
||
self.conversation.ensureSystemPrompt() catch return self.failSave("out of memory");
|
||
const baseline = self.conversation.messages.items.len;
|
||
|
||
const anchor = prompt orelse self.one_shot_task;
|
||
const schema = self.deriveOutputSchema(arena, baseline, anchor);
|
||
if (self.cancel_requested.load(.acquire)) {
|
||
self.resetAfterCancel(baseline);
|
||
return;
|
||
}
|
||
|
||
self.synthesizeScript(arena, filename, prompt, schema);
|
||
}
|
||
|
||
/// Steps 1–2 of `/save`: intent (over the session) → typed output schema. Both
|
||
/// turns leave the conversation as they found it; returns null if either turn
|
||
/// produced nothing usable (the caller then synthesizes without a schema).
|
||
fn deriveOutputSchema(self: *Agent, arena: std.mem.Allocator, baseline: usize, anchor: ?[]const u8) ?[]const u8 {
|
||
const intent = self.deriveIntent(arena, baseline, anchor) orelse return null;
|
||
if (self.cancel_requested.load(.acquire)) return null;
|
||
return self.deriveSchema(arena, intent);
|
||
}
|
||
|
||
/// One-sentence intent from the session turns. Runs over the live conversation
|
||
/// (so the model sees the session) but rolls back to `baseline`, keeping the
|
||
/// turn out of history. An explicit anchor is folded in as authoritative.
|
||
fn deriveIntent(self: *Agent, arena: std.mem.Allocator, baseline: usize, anchor: ?[]const u8) ?[]const u8 {
|
||
const ma = self.conversation.arena.allocator();
|
||
var out: std.Io.Writer.Allocating = .init(ma);
|
||
out.writer.writeAll(browser_tools.save_intent_prompt) catch return null;
|
||
if (anchor) |a| {
|
||
out.writer.print("\nThe user described the goal as: {s}\nTreat that as authoritative and reconcile it with the session.", .{a}) catch return null;
|
||
}
|
||
self.conversation.messages.append(self.allocator, .{ .role = .user, .content = out.written() }) catch return null;
|
||
defer self.conversation.rollback(baseline);
|
||
return self.runTextTurn(&self.conversation.messages, arena, self.allocator, ma, 512, "understanding the task");
|
||
}
|
||
|
||
/// Typed output schema from the intent. Runs over a throwaway message list —
|
||
/// not the conversation — so the schema is derived from the logical intent
|
||
/// alone, blind to the page structure and how the data was fetched.
|
||
fn deriveSchema(self: *Agent, arena: std.mem.Allocator, intent: []const u8) ?[]const u8 {
|
||
var msgs: std.ArrayList(zenai.provider.Message) = .empty;
|
||
const msg = std.fmt.allocPrint(arena, "{s} {s}", .{ browser_tools.save_schema_prompt, intent }) catch return null;
|
||
msgs.append(arena, .{ .role = .user, .content = msg }) catch return null;
|
||
const raw = self.runTextTurn(&msgs, arena, arena, arena, 1024, "designing the output schema") orelse return null;
|
||
return string.stripCodeFence(raw);
|
||
}
|
||
|
||
/// Run a single no-tools text turn over `messages` and return the model's text
|
||
/// duped into `dest` (so it survives any rollback of `messages`), or null on
|
||
/// cancel, error, or empty output. Shared by the intent and schema steps.
|
||
fn runTextTurn(
|
||
self: *Agent,
|
||
messages: *std.ArrayList(zenai.provider.Message),
|
||
dest: std.mem.Allocator,
|
||
list_alloc: std.mem.Allocator,
|
||
data_alloc: std.mem.Allocator,
|
||
max_tokens: i32,
|
||
status: []const u8,
|
||
) ?[]const u8 {
|
||
self.terminal.spinner.start();
|
||
self.terminal.spinner.setStatus(status);
|
||
var result = self.ai_client.?.runTools(
|
||
self.model,
|
||
messages,
|
||
list_alloc,
|
||
data_alloc,
|
||
.{ .context = @ptrCast(self), .callFn = handleToolCall },
|
||
.{
|
||
.tools = &.{},
|
||
.max_turns = 1,
|
||
.max_tokens = max_tokens,
|
||
.tool_choice = .none,
|
||
.effort = .low,
|
||
.cancel = .{ .context = @ptrCast(self), .checkFn = checkCancel },
|
||
},
|
||
) catch |err| {
|
||
self.terminal.spinner.cancel();
|
||
if (!self.cancel_requested.load(.acquire)) log.err(.app, "AI save schema turn error", .{ .err = err });
|
||
return null;
|
||
};
|
||
self.terminal.spinner.stop();
|
||
defer result.deinit();
|
||
self.total_usage.add(result.usage);
|
||
if (result.cancelled) return null;
|
||
const text = std.mem.trim(u8, result.text orelse return null, &std.ascii.whitespace);
|
||
if (text.len == 0) return null;
|
||
return dest.dupe(u8, text) catch null;
|
||
}
|
||
|
||
/// Step 3 of `/save`: hand the model the builtin catalog, the full conversation,
|
||
/// the deterministic record of what ran, and the required output schema, then
|
||
/// write the idiomatic script it returns.
|
||
fn synthesizeScript(self: *Agent, arena: std.mem.Allocator, filename: ?[]const u8, prompt: ?[]const u8, schema: ?[]const u8) void {
|
||
const provider_client = self.ai_client.?;
|
||
|
||
const resolved = self.resolveSavePathAndMode(arena, filename) orelse return;
|
||
const path = resolved.path;
|
||
|
||
self.conversation.ensureSystemPrompt() catch return self.failSave("out of memory");
|
||
|
||
const ma = self.conversation.arena.allocator();
|
||
const baseline = self.conversation.messages.items.len;
|
||
|
||
// When the session captured extract data, let the model test candidates on
|
||
// it via `run_script`; otherwise fall back to a single no-tools synthesis.
|
||
var verify: Verify = .{ .runtime = undefined };
|
||
var run_tools: [1]ProviderTool = undefined;
|
||
const verifying = blk: {
|
||
// Gate on a captured extract: it means the session loaded the page and
|
||
// left it in a state worth verifying against (and gives a prompt sample).
|
||
if (self.last_extract_json == null) break :blk false;
|
||
run_tools[0] = browser_tools.runScriptToolDef(ma) catch break :blk false;
|
||
const runtime = ScriptRuntime.init(self.allocator, self.browser.app, self.session, &self.node_registry) catch break :blk false;
|
||
verify.runtime = runtime;
|
||
self.active_verify = &verify;
|
||
self.script_runtime_mutex.lock();
|
||
self.active_script_runtime = runtime;
|
||
self.script_runtime_mutex.unlock();
|
||
break :blk true;
|
||
};
|
||
defer if (verifying) {
|
||
self.script_runtime_mutex.lock();
|
||
self.active_script_runtime = null;
|
||
self.script_runtime_mutex.unlock();
|
||
self.active_verify = null;
|
||
verify.runtime.cancelTerminate();
|
||
verify.runtime.deinit();
|
||
};
|
||
|
||
const sample: ?[]const u8 = if (verifying) blk: {
|
||
const d = self.last_extract_json.?;
|
||
break :blk d[0..@min(d.len, save_sample_cap)];
|
||
} else null;
|
||
const user_msg = self.buildSaveSynthesisMessage(ma, prompt, schema, sample) catch return self.failSave("out of memory");
|
||
self.conversation.messages.append(self.allocator, .{ .role = .user, .content = user_msg }) catch return self.failSave("out of memory");
|
||
|
||
self.terminal.spinner.start();
|
||
self.terminal.spinner.setStatus(if (verifying) "writing and testing the script" else "writing the script");
|
||
var result = provider_client.runTools(
|
||
self.model,
|
||
&self.conversation.messages,
|
||
self.allocator,
|
||
ma,
|
||
.{ .context = @ptrCast(self), .callFn = handleToolCall },
|
||
.{
|
||
.tools = if (verifying) run_tools[0..1] else &.{},
|
||
.max_turns = if (verifying) 6 else 1,
|
||
.max_tokens = 8192,
|
||
.tool_choice = if (verifying) .auto else .none,
|
||
.effort = .medium,
|
||
.cancel = .{ .context = @ptrCast(self), .checkFn = checkCancel },
|
||
},
|
||
) catch |err| {
|
||
self.terminal.spinner.cancel();
|
||
if (self.cancel_requested.load(.acquire)) {
|
||
self.resetAfterCancel(baseline);
|
||
return;
|
||
}
|
||
log.err(.app, "AI save synthesis error", .{ .err = err });
|
||
return self.abortSave(baseline, @errorName(err));
|
||
};
|
||
self.terminal.spinner.stop();
|
||
defer result.deinit();
|
||
self.total_usage.add(result.usage);
|
||
|
||
if (result.cancelled) {
|
||
self.resetAfterCancel(baseline);
|
||
return;
|
||
}
|
||
|
||
// Prefer the last candidate that ran cleanly — it's verified, pure JS, with
|
||
// none of the commentary the model sometimes wraps its final message in. Fall
|
||
// back to the final text only when nothing ran (no extract data, or it never
|
||
// called run_script).
|
||
const raw: []const u8 = blk: {
|
||
if (verifying) {
|
||
if (verify.last_source) |s| break :blk s;
|
||
}
|
||
if (result.text) |t| {
|
||
if (std.mem.trim(u8, t, &std.ascii.whitespace).len > 0) break :blk t;
|
||
}
|
||
return self.abortSave(baseline, "the model returned no script");
|
||
};
|
||
|
||
// `raw` lives in the conversation arena, freed by the rollback below; copy
|
||
// into the command arena first (scrubbing may return its input as-is).
|
||
const owned = arena.dupe(u8, string.stripCodeFence(raw)) catch return self.abortSave(baseline, "out of memory");
|
||
const script = browser_tools.reverseSubstituteEnvVars(arena, owned) catch return self.abortSave(baseline, "out of memory");
|
||
|
||
// The save turn is a meta-action; keep it out of the ongoing conversation.
|
||
self.conversation.rollback(baseline);
|
||
|
||
writeContentFile(path, script, resolved.mode) catch |err| {
|
||
self.terminal.printError("failed to save {s}: {s}", .{ path, @errorName(err) });
|
||
return;
|
||
};
|
||
|
||
self.rememberSavePath(path);
|
||
self.resetSaveBuffers();
|
||
self.terminal.printInfo("Saved synthesized script to {s}", .{path});
|
||
}
|
||
|
||
/// `run_script` tool handler: execute `source` on the dry-run runtime and hand
|
||
/// the model back the completion value (or the error), so it can judge and fix
|
||
/// its own script against real data.
|
||
fn runScriptTool(self: *Agent, allocator: std.mem.Allocator, arguments: ?std.json.Value) zenai.provider.Client.ToolHandler.Result {
|
||
const verify = self.active_verify.?;
|
||
const args = browser_tools.parseArgsOrDefault(struct { source: []const u8 = "" }, allocator, arguments) catch
|
||
return .{ .content = "invalid run_script arguments", .is_error = true };
|
||
const source = args.source;
|
||
if (source.len == 0) return .{ .content = "run_script requires a non-empty \"source\" string", .is_error = true };
|
||
|
||
// Start each candidate from a blank page, exactly like a standalone replay —
|
||
// so a script that forgets to goto(...) fails here instead of silently relying
|
||
// on the page the session left loaded.
|
||
if (self.session.hasPage()) self.session.removePage();
|
||
|
||
const outcome = verify.runtime.runSourceCapture(source, "candidate.js") catch
|
||
return .{ .content = "out of memory running candidate", .is_error = true };
|
||
if (outcome.err) |e| {
|
||
self.terminal.agentVerifyRun(oneLinePreview(allocator, e, 120), false);
|
||
return .{ .content = std.fmt.allocPrint(allocator, "Script threw: {s}", .{e}) catch "Script threw an error", .is_error = true };
|
||
}
|
||
|
||
// Keep the last source that ran cleanly — it's the verified, prose-free
|
||
// artifact `synthesizeScript` saves, instead of the model's final message
|
||
// (which may wrap the script in commentary).
|
||
verify.last_source = self.conversation.arena.allocator().dupe(u8, source) catch source;
|
||
|
||
const body = if (outcome.output.len == 0) "(completion value is empty/undefined)" else outcome.output;
|
||
self.terminal.agentVerifyRun(oneLinePreview(allocator, body, 120), true);
|
||
const content = std.fmt.allocPrint(allocator, "Completion value:\n{s}", .{body}) catch body;
|
||
return .{ .content = string.truncateWithMarker(allocator, content, tool_output_max_bytes), .is_error = false };
|
||
}
|
||
|
||
/// Collapse `text` to a single trimmed line capped at `max` cells (with an
|
||
/// ellipsis when cut) — a compact preview for the verify-run trace bullet.
|
||
fn oneLinePreview(arena: std.mem.Allocator, text: []const u8, max: usize) []const u8 {
|
||
const trimmed = std.mem.trim(u8, text, &std.ascii.whitespace);
|
||
const first = trimmed[0 .. std.mem.indexOfScalar(u8, trimmed, '\n') orelse trimmed.len];
|
||
if (first.len <= max) return first;
|
||
const cut = string.truncateUtf8(first, max);
|
||
return std.fmt.allocPrint(arena, "{s}…", .{cut}) catch cut;
|
||
}
|
||
|
||
/// Persist `path` as the destination reused by a subsequent bare `/save`.
|
||
fn rememberSavePath(self: *Agent, path: []const u8) void {
|
||
if (self.save_path) |old| {
|
||
if (std.mem.eql(u8, old, path)) return;
|
||
}
|
||
const dup = self.allocator.dupe(u8, path) catch return;
|
||
if (self.save_path) |old| self.allocator.free(old);
|
||
self.save_path = dup;
|
||
}
|
||
|
||
fn buildSaveSynthesisMessage(self: *Agent, arena: std.mem.Allocator, prompt: ?[]const u8, schema: ?[]const u8, sample: ?[]const u8) ![]const u8 {
|
||
var out: std.Io.Writer.Allocating = .init(arena);
|
||
const w = &out.writer;
|
||
try w.writeAll(browser_tools.save_synthesis_prompt);
|
||
try w.writeAll("\n\nBuiltin functions (call them as JS functions). extract is the main way to read data — use it for every value you need; the rest navigate or act on the page:\n");
|
||
try renderBuiltinCatalog(w);
|
||
const recorded = self.save_buffer.bytes();
|
||
if (recorded.len > 0) {
|
||
try w.writeAll("\nCommands and JS that actually ran this session:\n");
|
||
try w.writeAll(recorded);
|
||
}
|
||
if (schema) |s| {
|
||
try w.writeAll("\nThe completion value must match this output schema (types are examples):\n");
|
||
try w.writeAll(s);
|
||
}
|
||
if (sample) |data| {
|
||
try w.writeAll("\nWhat a recorded extract returned this session, for reference:\n");
|
||
try w.writeAll(data);
|
||
try w.writeAll("\n\n");
|
||
try w.writeAll(save_verify_addendum);
|
||
}
|
||
if (prompt) |p| {
|
||
try w.writeAll("\nThe user's instruction for this script:\n");
|
||
try w.writeAll(p);
|
||
}
|
||
return out.written();
|
||
}
|
||
|
||
/// Document the recorded browser tools — the subset callable from a saved
|
||
/// script — with full descriptions, so the model gets each function's argument
|
||
/// dialect (e.g. `extract`'s schema format) without the tool schemas a no-tools
|
||
/// synthesis turn omits.
|
||
fn renderBuiltinCatalog(w: *std.Io.Writer) !void {
|
||
// The primary builtins first; `evaluate` is held back and framed as a last
|
||
// resort below, so it isn't presented as a peer way to read data.
|
||
for (Schema.all()) |s| {
|
||
if (!s.tool.isRecorded() or s.tool == .evaluate) continue;
|
||
try renderBuiltinEntry(w, s);
|
||
}
|
||
for (Schema.all()) |s| {
|
||
if (s.tool != .evaluate) continue;
|
||
try w.writeAll("\nEscape hatch for advanced page interaction or page-side logic no builtin above can express — not for reading data extract can read:\n");
|
||
try renderBuiltinEntry(w, s);
|
||
}
|
||
}
|
||
|
||
fn renderBuiltinEntry(w: *std.Io.Writer, s: Schema) !void {
|
||
try w.print("\n{s}(", .{s.tool_name});
|
||
for (s.required, 0..) |req, i| {
|
||
if (i != 0) try w.writeAll(", ");
|
||
try w.writeAll(req);
|
||
}
|
||
try w.print("):\n{s}\n", .{s.description});
|
||
}
|
||
|
||
fn logSaveBufferError(self: *Agent, err: anyerror) void {
|
||
self.terminal.printError("save buffer disabled: {s}", .{@errorName(err)});
|
||
}
|
||
|
||
fn recordSaveCommand(self: *Agent, cmd: Command) void {
|
||
self.save_buffer.record(cmd) catch |err| self.logSaveBufferError(err);
|
||
}
|
||
|
||
fn recordSaveComment(self: *Agent, comment: []const u8) void {
|
||
self.save_buffer.recordComment(comment) catch |err| self.logSaveBufferError(err);
|
||
}
|
||
|
||
fn recordSaveRaw(self: *Agent, line: []const u8) void {
|
||
self.save_buffer.recordRaw(line) catch |err| self.logSaveBufferError(err);
|
||
}
|
||
|
||
fn printSlashHelp(self: *Agent, arena: std.mem.Allocator, target: []const u8) void {
|
||
if (target.len == 0) {
|
||
const all = Schema.all();
|
||
const browser = arena.alloc(SlashCommand.Help, all.len) catch return;
|
||
for (all, browser) |*s, *e| e.* = .{ .name = s.tool_name, .description = s.summary };
|
||
self.terminal.printHelpSection("Browser commands:", browser);
|
||
|
||
if (self.ai_client != null) {
|
||
const llm = arena.alloc(SlashCommand.Help, SlashCommand.llm_commands.len) catch return;
|
||
@memcpy(llm, &SlashCommand.llm_commands);
|
||
self.terminal.printHelpSection("\nLLM commands:", llm);
|
||
}
|
||
|
||
const meta = arena.alloc(SlashCommand.Help, SlashCommand.meta_commands.len) catch return;
|
||
for (SlashCommand.meta_commands, meta) |m, *e| e.* = .{ .name = m.name, .description = m.description };
|
||
self.terminal.printHelpSection("\nMeta commands:", meta);
|
||
return;
|
||
}
|
||
if (SlashCommand.findMeta(target)) |meta| {
|
||
switch (meta.tag) {
|
||
.help => self.terminal.printInfo("/help [name] — show help for a command, or list all when [name] is omitted", .{}),
|
||
.quit => self.terminal.printInfo("/quit — exit the REPL", .{}),
|
||
.verbosity => self.terminal.printInfo(
|
||
"/verbosity " ++ Config.tagHint(Config.AgentVerbosity) ++ " — set REPL agent verbosity (currently: {s}). Bare /verbosity prints the level.",
|
||
.{@tagName(self.terminal.verbosity)},
|
||
),
|
||
.effort => self.terminal.printInfo(
|
||
"/effort " ++ Config.tagHint(Config.Effort) ++ " — set per-turn reasoning effort (currently: {s}); saved to {s}. Bare /effort prints the level.",
|
||
.{ @tagName(self.effort), settings.remembered_path },
|
||
),
|
||
.usage => self.terminal.printInfo(
|
||
"/usage — show cumulative token usage and cache hit rate for this session",
|
||
.{},
|
||
),
|
||
.clear => self.terminal.printInfo(
|
||
"/clear — forget the conversation (history, usage, recorded actions, node IDs); keeps the loaded page and cookies",
|
||
.{},
|
||
),
|
||
.reset => self.terminal.printInfo(
|
||
"/reset — full reset: everything /clear does plus a new browser session, dropping the page, cookies, storage, and history",
|
||
.{},
|
||
),
|
||
.save => self.terminal.printInfo(
|
||
"/save [filename.js] [prompt] — save the session to [filename.js] (a random session-*.js if omitted). With an LLM, synthesizes an idiomatic script from the session and the optional prompt; with --no-llm, dumps the recorded actions verbatim.",
|
||
.{},
|
||
),
|
||
.load => self.terminal.printInfo(
|
||
"/load <path> — read a script from disk and run it against the current session; Tab completes file paths",
|
||
.{},
|
||
),
|
||
.model => self.terminal.printInfo(
|
||
"/model [name] — change the model; Tab completes the provider's models, bare /model shows the current one",
|
||
.{},
|
||
),
|
||
.provider => self.terminal.printInfo(
|
||
"/provider [name|null] — change the provider, or 'null' to disable the LLM (persisted, so the next launch starts in basic mode); Tab completes detected providers, bare /provider shows the current one",
|
||
.{},
|
||
),
|
||
}
|
||
return;
|
||
}
|
||
if (self.ai_client != null) {
|
||
for (SlashCommand.llm_commands) |row| {
|
||
if (std.ascii.eqlIgnoreCase(row.name, target)) {
|
||
self.terminal.printInfo("/{s} — {s}", .{ row.name, row.description });
|
||
return;
|
||
}
|
||
}
|
||
}
|
||
const tool_schema = Schema.findByName(target) orelse {
|
||
if (Terminal.closestCommand(target)) |near| {
|
||
self.terminal.printError("unknown command: {s}. Did you mean " ++ Terminal.highlightCmd("/help {s}") ++ "?", .{ target, near });
|
||
} else {
|
||
self.terminal.printError("unknown command: {s}", .{target});
|
||
}
|
||
return;
|
||
};
|
||
self.terminal.printInfo("/{s} — {s}", .{ tool_schema.tool_name, tool_schema.description });
|
||
}
|
||
|
||
/// Caller contract: `cmd` must be `.tool_call` — `.comment` and `.llm` are
|
||
/// filtered upstream, having no tool mapping.
|
||
fn runCommand(self: *Agent, arena: std.mem.Allocator, cmd: Command) browser_tools.ToolResult {
|
||
const tc = switch (cmd) {
|
||
.tool_call => |t| t,
|
||
else => return .{ .text = "internal: command has no tool mapping", .is_error = true },
|
||
};
|
||
return browser_tools.call(arena, self.session, &self.node_registry, tc.name(), tc.args) catch |err| .{
|
||
.text = switch (err) {
|
||
error.OutOfMemory => "out of memory",
|
||
error.FrameNotLoaded => "no page loaded — run /goto <url> first",
|
||
else => std.fmt.allocPrint(arena, "{s} failed: {s}", .{ tc.name(), @errorName(err) }) catch "tool failed",
|
||
},
|
||
.is_error = true,
|
||
};
|
||
}
|
||
|
||
/// Data output (/extract, /evaluate, /markdown, /tree, …) → plain stdout on
|
||
/// success so a caller can pipe it. Everything else routes through
|
||
/// `printToolOutcome`, which lays down the green ● / red ● dot shared with the
|
||
/// LLM tool-call path. Callers only invoke this for `.tool_call` commands (the
|
||
/// comment/login/acceptCookies branches take other paths).
|
||
fn printCommandResult(self: *Agent, cmd: Command, result: browser_tools.ToolResult) void {
|
||
const tc = switch (cmd) {
|
||
.tool_call => |t| t,
|
||
else => return,
|
||
};
|
||
if (cmd.producesData() and !result.is_error) {
|
||
self.printData(result.text);
|
||
return;
|
||
}
|
||
self.terminal.printToolOutcome(tc.name(), result.text, result.is_error);
|
||
}
|
||
|
||
/// Re-indent JSON for the terminal; MCP keeps renderJson's compact form.
|
||
fn printData(self: *Agent, text: []const u8) void {
|
||
var arena = std.heap.ArenaAllocator.init(self.allocator);
|
||
defer arena.deinit();
|
||
self.terminal.printAssistant(Terminal.reindentJson(arena.allocator(), text) orelse text);
|
||
}
|
||
|
||
/// Tracks whether a `/load`-run script emitted any `console.*` output, deciding
|
||
/// how `runScript` ends: one that printed nothing freezes the spinner into a
|
||
/// `/goto`-style bullet; one that printed leaves its output as the result.
|
||
const ScriptOutput = struct {
|
||
terminal: *Terminal,
|
||
emitted: bool = false,
|
||
|
||
/// `Runtime.ConsoleObserver` callback: on the first line, clear the live
|
||
/// spinner so output starts clean instead of colliding with the indicator.
|
||
fn observe(context: *anyopaque) void {
|
||
const self: *ScriptOutput = @ptrCast(@alignCast(context));
|
||
if (self.emitted) return;
|
||
self.emitted = true;
|
||
self.terminal.endTool();
|
||
}
|
||
};
|
||
|
||
fn runScript(self: *Agent, path: []const u8) bool {
|
||
var script_arena: std.heap.ArenaAllocator = .init(self.allocator);
|
||
defer script_arena.deinit();
|
||
|
||
const content = std.fs.cwd().readFileAlloc(script_arena.allocator(), path, 10 * 1024 * 1024) catch |err| {
|
||
self.terminal.printError("Failed to read script '{s}': {s}", .{ path, @errorName(err) });
|
||
return false;
|
||
};
|
||
|
||
const runtime = ScriptRuntime.init(self.allocator, self.browser.app, self.session, &self.node_registry) catch |err| {
|
||
self.terminal.printError("Failed to initialize script runtime: {s}", .{@errorName(err)});
|
||
return false;
|
||
};
|
||
defer runtime.deinit();
|
||
self.script_runtime_mutex.lock();
|
||
self.active_script_runtime = runtime;
|
||
self.script_runtime_mutex.unlock();
|
||
defer {
|
||
self.script_runtime_mutex.lock();
|
||
self.active_script_runtime = null;
|
||
self.script_runtime_mutex.unlock();
|
||
runtime.cancelTerminate();
|
||
self.browser.env.cancelTerminate();
|
||
self.cancel_requested.store(false, .release);
|
||
}
|
||
|
||
var output: ScriptOutput = .{ .terminal = &self.terminal };
|
||
runtime.console_observer = .{ .context = @ptrCast(&output), .notify = ScriptOutput.observe };
|
||
self.terminal.beginTool("script", path);
|
||
const result = runtime.runSource(content, path);
|
||
self.terminal.endTool();
|
||
|
||
if (result catch |err| {
|
||
self.terminal.printError("Script failed: {s}", .{@errorName(err)});
|
||
return false;
|
||
}) |message| {
|
||
self.terminal.printError("{s}", .{message});
|
||
return false;
|
||
}
|
||
|
||
// A script that printed nothing leaves no trace, so freeze the spinner into
|
||
// a green bullet (like /goto); one that printed already showed its result.
|
||
if (!output.emitted) self.terminal.printScriptDone("script", path);
|
||
return true;
|
||
}
|
||
|
||
/// Mirror a user-typed slash command into `self.conversation.messages` as if the
|
||
/// LLM had called the tool itself, so the next natural-language turn sees the
|
||
/// same conversation shape either way.
|
||
fn recordSlashToolCall(
|
||
self: *Agent,
|
||
user_input: []const u8,
|
||
tool_name: []const u8,
|
||
args: ?std.json.Value,
|
||
result: browser_tools.ToolResult,
|
||
) !void {
|
||
if (self.ai_client == null) return;
|
||
try self.conversation.ensureSystemPrompt();
|
||
|
||
const ma = self.conversation.arena.allocator();
|
||
self.synthetic_tool_call_id += 1;
|
||
|
||
const user_content = try ma.dupe(u8, user_input);
|
||
|
||
const tool_calls = try ma.alloc(zenai.provider.ToolCall, 1);
|
||
tool_calls[0] = .{
|
||
.id = try std.fmt.allocPrint(ma, "lp-slash-{d}", .{self.synthetic_tool_call_id}),
|
||
.name = try ma.dupe(u8, tool_name),
|
||
.arguments = if (args) |v| try zenai.json.dupeValue(ma, v) else null,
|
||
};
|
||
|
||
// truncateWithMarker returns its input unchanged under the cap; dupe so
|
||
// content doesn't alias the caller's per-iteration arena.
|
||
const capped = string.truncateWithMarker(ma, result.text, tool_output_max_bytes);
|
||
const content = if (capped.ptr == result.text.ptr) try ma.dupe(u8, capped) else capped;
|
||
|
||
const tool_results = try ma.alloc(zenai.provider.ToolResult, 1);
|
||
tool_results[0] = .{
|
||
.id = try ma.dupe(u8, tool_calls[0].id),
|
||
.name = try ma.dupe(u8, tool_calls[0].name),
|
||
.content = content,
|
||
.is_error = result.is_error,
|
||
};
|
||
|
||
const baseline = self.conversation.messages.items.len;
|
||
errdefer self.conversation.messages.shrinkRetainingCapacity(baseline);
|
||
// User turn before the assistant tool_call satisfies Gemini's rule
|
||
// that a function call must follow a user or function-response turn.
|
||
try self.conversation.messages.append(self.allocator, .{
|
||
.role = .user,
|
||
.content = user_content,
|
||
});
|
||
try self.conversation.messages.append(self.allocator, .{
|
||
.role = .assistant,
|
||
.tool_calls = tool_calls,
|
||
});
|
||
try self.conversation.messages.append(self.allocator, .{
|
||
.role = .tool,
|
||
.tool_results = tool_results,
|
||
});
|
||
}
|
||
|
||
/// Returned text lives in `conversation.arena`, valid only until the next prune.
|
||
/// Caller must call `conversation.prune()` after consuming it — pruning earlier
|
||
/// frees the arena the slice points into. `null` means the model emitted nothing
|
||
/// even after the synthesis turn.
|
||
fn processUserMessage(self: *Agent, input: TurnInput) !?[]const u8 {
|
||
const ma = self.conversation.arena.allocator();
|
||
|
||
try self.conversation.ensureSystemPrompt();
|
||
|
||
// Attachments only ride on the first user turn (just after the system prompt).
|
||
const turn_attachments: ?[]const []const u8 =
|
||
if (self.conversation.messages.items.len == 1) input.attachments else null;
|
||
|
||
// Roll-back baseline: on API failure the failed user turn would otherwise
|
||
// stay in history and replay on the next attempt.
|
||
const msg_baseline = self.conversation.messages.items.len;
|
||
|
||
if (turn_attachments) |paths| {
|
||
const parts = try self.buildUserMessageParts(ma, input.prompt, paths);
|
||
try self.conversation.messages.append(self.allocator, .{
|
||
.role = .user,
|
||
.parts = parts,
|
||
});
|
||
} else {
|
||
try self.conversation.messages.append(self.allocator, .{
|
||
.role = .user,
|
||
.content = try ma.dupe(u8, input.prompt),
|
||
});
|
||
}
|
||
|
||
const provider_client = self.ai_client orelse return error.NoAiClient;
|
||
|
||
self.terminal.spinner.start();
|
||
var result = provider_client.runTools(
|
||
self.model,
|
||
&self.conversation.messages,
|
||
self.allocator,
|
||
ma,
|
||
.{ .context = @ptrCast(self), .callFn = handleToolCall },
|
||
.{
|
||
.tools = globalTools(),
|
||
.max_turns = 100,
|
||
.max_tool_calls = 200,
|
||
.max_tokens = 4096,
|
||
.tool_choice = .auto,
|
||
// Per-turn reasoning budget; resolved from --effort / .lp-agent.zon
|
||
// / mode default, adjustable at runtime via /effort. Ignored by
|
||
// non-thinking models.
|
||
.effort = self.effort,
|
||
.cancel = .{ .context = @ptrCast(self), .checkFn = checkCancel },
|
||
},
|
||
) catch |err| {
|
||
self.terminal.spinner.cancel();
|
||
// Ctrl-C can land while runTools unwinds an HTTP error — surface
|
||
// UserCancelled, not ApiError, so the user sees the outcome they asked for.
|
||
if (self.cancel_requested.load(.acquire)) return self.drainCancellation(msg_baseline);
|
||
log.err(.app, "AI API error", .{ .err = err });
|
||
self.conversation.rollback(msg_baseline);
|
||
return error.ApiError;
|
||
};
|
||
self.terminal.spinner.stop();
|
||
defer result.deinit();
|
||
self.total_usage.add(result.usage);
|
||
|
||
if (result.cancelled) return self.drainCancellation(msg_baseline);
|
||
|
||
if (input.capture_for_save) {
|
||
// When the LLM tries multiple `extract` schemas in one turn, only the
|
||
// last successful one is the answer; earlier probes are noise.
|
||
var last_extract_idx: ?usize = null;
|
||
for (result.tool_calls_made, 0..) |tc, i| {
|
||
const t = std.meta.stringToEnum(BrowserTool, tc.name) orelse continue;
|
||
if (!tc.is_error and t == .extract) last_extract_idx = i;
|
||
}
|
||
|
||
// Keep the latest extract's real result so `/save` can ground and
|
||
// verify its synthesized post-processing against actual data.
|
||
if (last_extract_idx) |idx| {
|
||
_ = self.last_extract_arena.reset(.retain_capacity);
|
||
self.last_extract_json = self.last_extract_arena.allocator().dupe(u8, result.tool_calls_made[idx].result) catch null;
|
||
}
|
||
|
||
var recorded_any = false;
|
||
for (result.tool_calls_made, 0..) |tc, i| {
|
||
if (tc.is_error) continue;
|
||
const tool = std.meta.stringToEnum(BrowserTool, tc.name) orelse continue;
|
||
if (last_extract_idx) |idx| {
|
||
if (tool == .extract and idx != i) continue;
|
||
}
|
||
const args = browser_tools.normalizeArgKeys(self.conversation.arena.allocator(), tool, tc.arguments) catch tc.arguments;
|
||
const cmd = Command.fromToolCall(tool, args);
|
||
if (!cmd.isRecorded()) continue;
|
||
if (!recorded_any) {
|
||
if (input.record_comment) |c| self.recordSaveComment(c);
|
||
recorded_any = true;
|
||
}
|
||
self.recordSaveCommand(cmd);
|
||
}
|
||
}
|
||
|
||
// Dupe into the conversation arena — RunToolsResult arenas deinit below.
|
||
self.last_turn_refused = result.finish_reason == .safety;
|
||
const final_text: ?[]const u8 = blk: {
|
||
if (result.text) |text| {
|
||
if (std.mem.trim(u8, text, " \t\r\n").len > 0) break :blk try ma.dupe(u8, text);
|
||
}
|
||
|
||
// A refusal is deterministic; re-prompting just refuses again.
|
||
if (self.last_turn_refused) break :blk null;
|
||
|
||
// Without a synthesis turn forbidding tools+pretraining, models
|
||
// confabulate when the page was blocked or empty.
|
||
log.info(.app, "synthesizing final answer", .{});
|
||
const synth_baseline = self.conversation.messages.items.len;
|
||
try self.conversation.messages.append(self.allocator, .{
|
||
.role = .user,
|
||
.content = try ma.dupe(u8, synthesis_prompt),
|
||
});
|
||
|
||
var synth = provider_client.runTools(
|
||
self.model,
|
||
&self.conversation.messages,
|
||
self.allocator,
|
||
ma,
|
||
.{ .context = @ptrCast(self), .callFn = handleToolCall },
|
||
.{
|
||
.tools = &.{},
|
||
.max_turns = 1,
|
||
.max_tokens = 4096,
|
||
.tool_choice = .none,
|
||
// .low (≈512 tokens) so reasoning models still pick an answer
|
||
// but can't burn the whole turn on thinking and emit nothing.
|
||
.effort = .low,
|
||
.cancel = .{ .context = @ptrCast(self), .checkFn = checkCancel },
|
||
},
|
||
) catch |err| {
|
||
if (self.cancel_requested.load(.acquire)) return self.drainCancellation(msg_baseline);
|
||
log.err(.app, "AI synthesis error", .{ .err = err });
|
||
self.conversation.rollback(synth_baseline);
|
||
break :blk null;
|
||
};
|
||
defer synth.deinit();
|
||
self.total_usage.add(synth.usage);
|
||
|
||
if (synth.cancelled) return self.drainCancellation(msg_baseline);
|
||
|
||
break :blk if (synth.text) |text| try ma.dupe(u8, text) else null;
|
||
};
|
||
|
||
// NB: pruning is deferred to the caller. `final_text` is in the conversation
|
||
// arena, and `conversation.prune()` may rebuild that arena — running it here
|
||
// would hand the caller a dangling slice.
|
||
return final_text;
|
||
}
|
||
|
||
/// Build a `parts`-based user message when `--attach` was given. Text-ish files
|
||
/// are inlined into the text prefix (surrounded by clear markers); binary files
|
||
/// (image/audio/pdf) are base64-encoded as provider inline-data parts. Unknown
|
||
/// extensions error out so the caller fails loudly instead of silently dropping
|
||
/// the attachment.
|
||
fn buildUserMessageParts(
|
||
self: *Agent,
|
||
ma: std.mem.Allocator,
|
||
user_input: []const u8,
|
||
paths: []const []const u8,
|
||
) ![]const zenai.provider.ContentPart {
|
||
var text_prefix: std.ArrayList(u8) = .empty;
|
||
var inline_parts: std.ArrayList(zenai.provider.ContentPart) = .empty;
|
||
|
||
for (paths) |path| {
|
||
const mime = zenai.provider.inferInlineMimeType(path) orelse {
|
||
log.err(.app, "unsupported attachment", .{ .path = path });
|
||
self.terminal.printError("unsupported attachment type: {s}", .{path});
|
||
return error.UnsupportedAttachment;
|
||
};
|
||
|
||
if (std.mem.startsWith(u8, mime, "text/")) {
|
||
const bytes = std.fs.cwd().readFileAlloc(ma, path, 512 * 1024) catch |err| {
|
||
log.err(.app, "read attachment failed", .{ .path = path, .err = err });
|
||
self.terminal.printError("could not read attachment: {s}", .{path});
|
||
return error.AttachmentReadFailed;
|
||
};
|
||
try text_prefix.writer(ma).print(
|
||
"[Attached file: {s}]\n{s}\n[End of attachment]\n\n",
|
||
.{ path, bytes },
|
||
);
|
||
} else {
|
||
const raw = std.fs.cwd().readFileAlloc(ma, path, 20 * 1024 * 1024) catch |err| {
|
||
log.err(.app, "read attachment failed", .{ .path = path, .err = err });
|
||
self.terminal.printError("could not read attachment: {s}", .{path});
|
||
return error.AttachmentReadFailed;
|
||
};
|
||
const b64_len = std.base64.standard.Encoder.calcSize(raw.len);
|
||
const b64 = try ma.alloc(u8, b64_len);
|
||
_ = std.base64.standard.Encoder.encode(b64, raw);
|
||
try inline_parts.append(ma, .{ .image = .{
|
||
.data = b64,
|
||
.mime_type = try ma.dupe(u8, mime),
|
||
} });
|
||
}
|
||
}
|
||
|
||
var parts: std.ArrayList(zenai.provider.ContentPart) = .empty;
|
||
try text_prefix.appendSlice(ma, user_input);
|
||
try parts.append(ma, .{ .text = try text_prefix.toOwnedSlice(ma) });
|
||
for (inline_parts.items) |p| try parts.append(ma, p);
|
||
return parts.toOwnedSlice(ma);
|
||
}
|
||
|
||
// Cap per-call tool output so heavy pages don't balloon the message arena (and
|
||
// the next request body) without bound.
|
||
const tool_output_max_bytes: usize = 1 * 1024 * 1024;
|
||
|
||
fn handleToolCall(ctx: *anyopaque, allocator: std.mem.Allocator, tool_name: []const u8, arguments: ?std.json.Value) zenai.provider.Client.ToolHandler.Result {
|
||
const self: *Agent = @ptrCast(@alignCast(ctx));
|
||
// `run_script`'s only arg is the whole candidate script — too long and noisy
|
||
// to render, so suppress it and let the label/phase carry the context.
|
||
const is_run_script = self.active_verify != null and std.mem.eql(u8, tool_name, browser_tools.run_script_tool_name);
|
||
// The spinner doesn't render args, and `agentToolDone` skips the body line
|
||
// at low verbosity — don't pay for the stringify when nobody reads it.
|
||
const needs_args = !is_run_script and (self.terminal.spinner.isEnabled() or self.terminal.verbosity != .low);
|
||
// Stringify the pre-substitution args so $LP_* placeholders the model
|
||
// emitted stay redacted in the UI.
|
||
const args_str: []const u8 = if (needs_args) (if (arguments) |v|
|
||
std.json.Stringify.valueAlloc(allocator, v, .{}) catch ""
|
||
else
|
||
"") else "";
|
||
self.terminal.spinner.setTool(tool_name, args_str);
|
||
defer self.terminal.spinner.setThinking();
|
||
|
||
const outcome: zenai.provider.Client.ToolHandler.Result = if (is_run_script)
|
||
self.runScriptTool(allocator, arguments)
|
||
else if (browser_tools.call(allocator, self.session, &self.node_registry, tool_name, arguments)) |result|
|
||
.{ .content = string.truncateWithMarker(allocator, result.text, tool_output_max_bytes), .is_error = result.is_error }
|
||
else |err|
|
||
.{ .content = std.fmt.allocPrint(allocator, "Error: {s}", .{@errorName(err)}) catch "Error: tool execution failed", .is_error = true };
|
||
|
||
// run_script emits its own always-visible trace inside `runScriptTool`.
|
||
if (!is_run_script) self.terminal.agentToolDone(tool_name, args_str, !outcome.is_error);
|
||
if (self.terminal.verbosity == .high) self.terminal.printToolOutcome(tool_name, outcome.content, outcome.is_error);
|
||
return outcome;
|
||
}
|
||
|
||
/// One-shot for `--list-models`: resolve provider+key, fetch chat-capable model
|
||
/// IDs, print to stdout one per line.
|
||
pub fn listModels(allocator: std.mem.Allocator, opts: Config.Agent) !void {
|
||
if (opts.no_llm) {
|
||
log.fatal(.app, "list-models needs LLM", .{
|
||
.hint = "--no-llm and --list-models conflict; drop --no-llm",
|
||
});
|
||
return error.ConflictingFlags;
|
||
}
|
||
if (opts.task != null or opts.script_file != null) {
|
||
log.fatal(.app, "list-models is exclusive", .{
|
||
.hint = "--list-models only takes --provider/--model/--base-url",
|
||
});
|
||
return error.ConflictingFlags;
|
||
}
|
||
const resolved = (try settings.resolveCredentials(allocator, opts, null, false)) orelse return error.MissingProvider;
|
||
const llm = resolved.credentials;
|
||
|
||
var arena: std.heap.ArenaAllocator = .init(allocator);
|
||
defer arena.deinit();
|
||
const ids = try zenai.provider.listChatModelIds(allocator, arena.allocator(), llm.provider, llm.key, opts.base_url);
|
||
|
||
var stdout_file = std.fs.File.stdout().writer(&.{});
|
||
const w = &stdout_file.interface;
|
||
for (ids) |id| try w.print("{s}\n", .{id});
|
||
try w.flush();
|
||
}
|
||
|
||
const ModelCompletions = struct {
|
||
provider: Config.AiProvider,
|
||
/// Empty when the fetch failed — cached so the per-keystroke hinter doesn't
|
||
/// re-hit the network each press.
|
||
ids: []const []const u8,
|
||
};
|
||
|
||
/// `CompletionSource.providers`. Reuses pre-detected available providers to
|
||
/// avoid reading environment variables on each autocomplete keypress.
|
||
fn completionProviders(context: *anyopaque, arena: std.mem.Allocator) []const []const u8 {
|
||
const self: *Agent = @ptrCast(@alignCast(context));
|
||
const names = arena.alloc([]const u8, self.available_providers.len + 1) catch return &.{};
|
||
for (self.available_providers, 0..) |p, i| {
|
||
names[i] = arena.dupe(u8, p) catch return &.{};
|
||
}
|
||
names[self.available_providers.len] = provider_off_keyword;
|
||
return names;
|
||
}
|
||
|
||
/// `CompletionSource.models`. Blocks on a one-time fetch per provider, caching
|
||
/// success or empty so the per-keystroke hinter pays the round-trip once.
|
||
fn completionModels(context: *anyopaque, _: std.mem.Allocator) []const []const u8 {
|
||
const self: *Agent = @ptrCast(@alignCast(context));
|
||
const llm = self.model_credentials orelse return &.{};
|
||
if (self.model_completions) |c| if (c.provider == llm.provider) return c.ids;
|
||
|
||
_ = self.model_completion_arena.reset(.retain_capacity);
|
||
const ids = zenai.provider.listChatModelIds(
|
||
self.allocator,
|
||
self.model_completion_arena.allocator(),
|
||
llm.provider,
|
||
llm.key,
|
||
self.model_base_url,
|
||
) catch &.{};
|
||
self.model_completions = .{ .provider = llm.provider, .ids = ids };
|
||
return ids;
|
||
}
|
||
|
||
test "parseSaveCommand: filename only" {
|
||
const r = try parseSaveCommand("out.js");
|
||
try std.testing.expectEqualStrings("out.js", r.filename.?);
|
||
try std.testing.expect(r.prompt == null);
|
||
}
|
||
|
||
test "parseSaveCommand: filename and prompt" {
|
||
const r = try parseSaveCommand("out.js summarize the login flow");
|
||
try std.testing.expectEqualStrings("out.js", r.filename.?);
|
||
try std.testing.expectEqualStrings("summarize the login flow", r.prompt.?);
|
||
}
|
||
|
||
test "parseSaveCommand: quoted filename keeps trailing prompt" {
|
||
const r = try parseSaveCommand("\"my flow.js\" do X");
|
||
try std.testing.expectEqualStrings("my flow.js", r.filename.?);
|
||
try std.testing.expectEqualStrings("do X", r.prompt.?);
|
||
}
|
||
|
||
test "parseSaveCommand: prompt only when first token is not a .js name" {
|
||
const r = try parseSaveCommand("make a login script");
|
||
try std.testing.expect(r.filename == null);
|
||
try std.testing.expectEqualStrings("make a login script", r.prompt.?);
|
||
}
|
||
|
||
test "parseSaveCommand: empty is all null" {
|
||
const r = try parseSaveCommand(" ");
|
||
try std.testing.expect(r.filename == null);
|
||
try std.testing.expect(r.prompt == null);
|
||
}
|
||
|
||
test "parseSaveCommand: rejects path-like filenames" {
|
||
try std.testing.expectError(error.InvalidFilename, parseSaveCommand("../evil.js"));
|
||
try std.testing.expectError(error.InvalidFilename, parseSaveCommand("/tmp/x.js"));
|
||
try std.testing.expectError(error.UnterminatedQuote, parseSaveCommand("\"unclosed.js"));
|
||
}
|
||
|
||
test "renderBuiltinCatalog: lists recorded tools, omits read-only ones" {
|
||
var out: std.Io.Writer.Allocating = .init(std.testing.allocator);
|
||
defer out.deinit();
|
||
try renderBuiltinCatalog(&out.writer);
|
||
const text = out.written();
|
||
try std.testing.expect(std.mem.indexOf(u8, text, "goto(") != null);
|
||
try std.testing.expect(std.mem.indexOf(u8, text, "extract(") != null);
|
||
try std.testing.expect(std.mem.indexOf(u8, text, "click(") != null);
|
||
// tree/markdown are read-only and not callable from a saved script.
|
||
try std.testing.expect(std.mem.indexOf(u8, text, "tree(") == null);
|
||
try std.testing.expect(std.mem.indexOf(u8, text, "markdown(") == null);
|
||
}
|