agent: add /usage slash command to REPL

Adds a new `/usage` meta command to print cumulative token usage and
cache hit rate for the current session. Also updates zenai.

build.zig.zon

@@ -35,8 +35,8 @@
             .hash = "sqlite3-3.51.0-DMxLWssOAABZ8cAvU_LfBIbp0kZjm824PU8sSLXpEDdr",
         },
         .zenai = .{
-            .url = "git+https://github.com/lightpanda-io/zenai.git#f2031a854aa0533a8bf3dfa7d0ede6cd77dc59df",
-            .hash = "zenai-0.0.0-iOY_VOhjBAB5sEZuzUdokns-GYzvrbPcanneYdpJRW1V",
+            .url = "git+https://github.com/lightpanda-io/zenai.git#c6a4c11de53cc5c39c825a57513b3071bda154f6",
+            .hash = "zenai-0.0.0-iOY_VDVqBAA3mH3_90Lg7Vz3JK8oD0gIhcete_M9c1UC",
         },
         .isocline = .{
             .url = "git+https://github.com/arrufat/isocline?ref=lightpanda#75f26db298f679e8aec82d8fc50c7115158a63e8",

docs/agent.md

@@ -304,9 +304,10 @@ See [agent-script.md](agent-script.md) for the full script format reference.
   the provider's fetched model list, and bare `/provider`/`/model` print the
   current selection — `/save [file.js]` writes the session to a script and
   `/load <path>` runs one from disk (Tab completes file paths), `/quit` exits
-  the REPL, `/verbosity <low|medium|high>` tunes the log level, and
+  the REPL, `/verbosity <low|medium|high>` tunes the log level,
   `/effort <none|minimal|low|medium|high|xhigh>` sets the per-turn reasoning
-  budget (saved to `.lp-agent.zon`). These are REPL-only and never recorded.
+  budget (saved to `.lp-agent.zon`), and `/usage` prints cumulative token usage
+  and the cache hit rate for the session. These are REPL-only and never recorded.
   ```
   > /goto https://example.com
   > /findElement role=button name=Submit

src/agent/Agent.zig

@@ -633,6 +633,7 @@ fn handleMeta(self: *Agent, arena: std.mem.Allocator, meta: *const SlashCommand.
         .help => self.printSlashHelp(arena, rest),
         .verbosity => self.handleVerbosity(rest),
         .effort => self.handleEffort(rest),
+        .usage => self.handleUsage(),
         .save => self.handleSave(arena, rest),
         .load => self.handleLoad(rest),
         .model => self.handleModel(arena, rest),
@@ -668,6 +669,27 @@ fn handleEffort(self: *Agent, rest: []const u8) void {
     self.reportSaved("effort", @tagName(level));
 }
 
+/// Print cumulative token usage for the session, broken down so the cache's
+/// effect is visible — the REPL otherwise never surfaces the `$usage` line that
+/// `--task` prints. Reads `total_usage`, accumulated across every turn by
+/// `processUserMessage`; the 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()});
+    }
+}
+
 fn handleLoad(self: *Agent, rest: []const u8) void {
     const path = std.mem.trim(u8, rest, &std.ascii.whitespace);
     if (path.len == 0) {
@@ -1147,6 +1169,10 @@ fn printSlashHelp(self: *Agent, arena: std.mem.Allocator, target: []const u8) vo
                 "/effort <none|minimal|low|medium|high|xhigh> — 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",
+                .{},
+            ),
             .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.",
                 .{},

src/agent/SlashCommand.zig

@@ -17,7 +17,7 @@
 // along with this program.  If not, see <https://www.gnu.org/licenses/>.
 
 //! REPL-only meta slash commands (`/help`, `/quit`, `/verbosity`, `/effort`,
-//! `/model`, `/provider`). Meta commands aren't tool slash commands — they're handled
+//! `/usage`, `/model`, `/provider`). Meta commands aren't tool slash commands — they're handled
 //! by `Agent.handleMeta` and never reach the recorder. Tool slash-command
 //! schema primitives live in `lp.Schema`; consumers should import that
 //! directly.
@@ -47,7 +47,7 @@ pub const MetaCommand = struct {
 
     /// Dispatched by `Agent.handleMeta` via an exhaustive switch so adding
     /// a new meta command is a compile error until it's wired up there too.
-    const Tag = enum { help, quit, verbosity, effort, save, load, model, provider };
+    const Tag = enum { help, quit, verbosity, effort, usage, save, load, model, provider };
 };
 
 pub const meta_commands = [_]MetaCommand{
@@ -55,6 +55,7 @@ pub const meta_commands = [_]MetaCommand{
     .{ .tag = .quit, .name = "quit", .hint = "", .values = &.{}, .description = "Exit the REPL" },
     .{ .tag = .verbosity, .name = "verbosity", .hint = "<low|medium|high>", .values = &.{ "low", "medium", "high" }, .description = "Set agent verbosity" },
     .{ .tag = .effort, .name = "effort", .hint = "<none|minimal|low|medium|high|xhigh>", .values = &.{ "none", "minimal", "low", "medium", "high", "xhigh" }, .description = "Set per-turn reasoning effort" },
+    .{ .tag = .usage, .name = "usage", .hint = "", .values = &.{}, .description = "Show token usage and cache stats for this session" },
     .{ .tag = .save, .name = "save", .hint = "[filename.js] [prompt]", .values = &.{}, .description = "Save this session to a file" },
     .{ .tag = .load, .name = "load", .hint = "<path>", .values = &.{}, .description = "Load and run a script from disk" },
     .{ .tag = .model, .name = "model", .hint = "[name]", .values = &.{}, .description = "Change the model" },