Skip Creative chest grind and report Gemini token usage

gemini_bot.js

@@ -8,7 +8,7 @@ const mineflayer = require("mineflayer");
 const { pathfinder, Movements, goals: { GoalNear } } = require("mineflayer-pathfinder");
 const mcData = require("minecraft-data");
 const Vec3 = require("vec3");
-const { DIRECTIONS, parseDecision, extractCliResponse, makePrompt } = require("./gemini_bot_logic");
+const { DIRECTIONS, parseDecision, extractCliResponse, tokenUsage, makePrompt } = require("./gemini_bot_logic");
 
 const USERNAME = "Gemini_Bot";
 const ADMIN = "IffyBinkie6134";
@@ -19,7 +19,7 @@ const WORKSPACE = process.env.GEMINI_BOT_WORKSPACE;
 const PORT = Number(process.env.GEMINI_BOT_STATUS_PORT || "3001");
 const state = { ready: false, busy: false, reconnect_count: 0, completed_actions: 0,
   failed_actions: 0, cli_calls: 0, last_action: null, last_error: null, last_elapsed_s: null,
-  prompt_tokens: null, completion_tokens: null, model: MODEL, bot_username: USERNAME };
+  prompt_tokens: null, completion_tokens: null, models_used: [], model: MODEL, bot_username: USERNAME };
 let bot = null;
 let reconnectTimer = null;
 let busy = false;
@@ -72,9 +72,10 @@ function callCli(prompt) {
         }
         try {
           const result = extractCliResponse(stdout);
-          const usage = result.stats.models && result.stats.models[MODEL] && result.stats.models[MODEL].tokens;
-          state.prompt_tokens = usage && (usage.prompt || usage.input) || null;
-          state.completion_tokens = usage && (usage.candidates || usage.output) || null;
+          const usage = tokenUsage(result.stats);
+          state.prompt_tokens = usage.prompt_tokens;
+          state.completion_tokens = usage.completion_tokens;
+          state.models_used = usage.models_used;
           resolve(result.decision);
         } catch (_) { reject(new Error("Gemini CLI response malformed")); }
       });

gemini_bot_logic.js

@@ -46,6 +46,18 @@ function extractCliResponse(output) {
   return { decision: parseDecision(envelope.response), stats: envelope.stats || {} };
 }
 
+function tokenUsage(stats) {
+  const models = stats && stats.models && typeof stats.models === "object" ? stats.models : {};
+  const tokenSets = Object.values(models).map(value => value && value.tokens || {});
+  if (!tokenSets.length && stats && stats.tokens) tokenSets.push(stats.tokens);
+  const sum = names => tokenSets.reduce((total, tokens) => {
+    const key = names.find(name => Number.isFinite(tokens[name]));
+    return total + (key ? tokens[key] : 0);
+  }, 0);
+  return { models_used: Object.keys(models), prompt_tokens: sum(["prompt", "input", "inputTokens"]),
+    completion_tokens: sum(["candidates", "output", "outputTokens"]) };
+}
+
 function makePrompt(observation, request) {
   const instruction = request ? `Admin request: ${JSON.stringify(request.slice(0, 240))}` :
     "Autonomous turn: choose explore, mine, or place to make one small real change. Use wait only if blocked; save chat for admin requests.";
@@ -63,4 +75,4 @@ function makePrompt(observation, request) {
   ].join("\n");
 }
 
-module.exports = { DIRECTIONS, PLACE_PALETTE, MINE_PALETTE, parseDecision, extractCliResponse, makePrompt };
+module.exports = { DIRECTIONS, PLACE_PALETTE, MINE_PALETTE, parseDecision, extractCliResponse, tokenUsage, makePrompt };

tests/test_gemini_bot.js

@@ -1,7 +1,7 @@
 "use strict";
 const assert = require("node:assert/strict");
 const { test } = require("node:test");
-const { parseDecision, extractCliResponse, makePrompt } = require("../gemini_bot_logic");
+const { parseDecision, extractCliResponse, tokenUsage, makePrompt } = require("../gemini_bot_logic");
 
 test("accepts one bounded action", () => {
   assert.deepEqual(parseDecision('{"action":"explore","direction":"north","blocks":16}'),
@@ -33,6 +33,13 @@ test("reads CLI envelope, and fails closed on API errors", () => {
   assert.throws(() => extractCliResponse("invalid"));
 });
 
+test("aggregates per-model CLI token metrics", () => {
+  assert.deepEqual(tokenUsage({ models: { "gemini-3.1-flash-lite": { tokens: { prompt: 120, candidates: 25 } } } }),
+    { models_used: ["gemini-3.1-flash-lite"], prompt_tokens: 120, completion_tokens: 25 });
+  assert.deepEqual(tokenUsage({ tokens: { input: 10, output: 2 } }),
+    { models_used: [], prompt_tokens: 10, completion_tokens: 2 });
+});
+
 test("prompt contains one observed world and the short admin request", () => {
   const prompt = makePrompt({ position: { x: 0, y: 70, z: 0 }, nearby_blocks: ["stone"] }, "segue o Naldo");
   assert.match(prompt, /Naldo_Bot/);

tests/test_voyager_patch.py

@@ -125,7 +125,7 @@ def test_resume():
 def check_curriculum_behavior(source):
     tree = ast.parse(source)
     cls = next(n for n in tree.body if isinstance(n, ast.ClassDef) and n.name == "CurriculumAgent")
-    names = {"get_task_context", "propose_next_ai_task", "parse_ai_message", "render_human_message", "render_observation"}
+    names = {"get_task_context", "propose_next_ai_task", "propose_next_task", "parse_ai_message", "render_human_message", "render_observation"}
     methods = [n for n in cls.body if isinstance(n, ast.FunctionDef) and n.name in names]
     ns = {"AIMessage": SimpleNamespace, "HumanMessage": SimpleNamespace,
           "choose_curriculum_task": choose_curriculum_task,
@@ -180,6 +180,13 @@ def check_curriculum_behavior(source):
     task, _ = agent.propose_next_ai_task(messages=[SimpleNamespace(content="Nearby blocks: dirt, stone")])
     assert task == "Mine 1 dirt" and agent.llm.call_count == 3
     assert task_spec(task) not in {task_spec(old) for old in agent.completed_tasks + agent.failed_tasks}
+    agent.mode = "auto"
+    agent.render_system_message = Mock(return_value=SimpleNamespace(content="Creative"))
+    agent.render_human_message = Mock(return_value=SimpleNamespace(content="inventory full"))
+    agent.propose_next_ai_task = Mock(return_value=("Place 1 stone", "Creative action"))
+    full = {"status": {"inventoryUsed": 35}, "inventory": {}}
+    assert agent.propose_next_task(events=[["observe", full]], chest_observation="Chests: None\n\n") == ("Place 1 stone", "Creative action")
+    agent.propose_next_ai_task.assert_called_once()
     print("PASS valid, composite, duplicate, Survival, unlabeled and exhausted Curriculum outputs")
 
 
@@ -304,6 +311,7 @@ def check_clean_upstream(repo):
         curriculum = (stack.SRC / "voyager/agents/curriculum.py").read_text()
         assert "Creative deployment: skip Survival QA" in curriculum
         assert "Curriculum fallback after bounded retries" in curriculum
+        assert "Craft 1 chest" not in curriculum
         check_curriculum_behavior(curriculum)
         check_action_behavior((stack.SRC / "voyager/agents/action.py").read_text())
         exploration = (stack.SRC / "voyager/control_primitives/exploreUntil.js").read_text()

voyager_stack.py

@@ -701,8 +701,18 @@ async function exploreUntil(bot, direction, maxTime = 60, callback) {
             r'(?ms)^        if self.progress >= self.warm_up\["context"\]:\n.*?(?=^        for key in self.curriculum_observations:)',
             '', curriculum_text, count=1,
         )
-        if qa_count != 1:
-            raise RuntimeError("Voyager Creative observation QA patch did not apply")
+        if qa_count != 1:
+            raise RuntimeError("Voyager Creative observation QA patch did not apply")
+        # Upstream's near-full inventory shortcut forces an endless Survival
+        # "Craft 1 chest" task at 33 slots. Creative has no such prerequisite:
+        # the bounded curriculum must choose a world action from observation.
+        curriculum_text, chest_count = re.subn(
+            r'(?ms)^        # hard code task when inventory is almost full\n.*?(?=^        messages = \[)',
+            '        # Creative deployment skips the Survival chest shortcut.\n\n',
+            curriculum_text, count=1,
+        )
+        if chest_count != 1:
+            raise RuntimeError("Voyager Survival chest shortcut marker changed")
         curriculum_text = curriculum_text.replace('from __future__ import annotations', 'from __future__ import annotations\nfrom voyager_verification import validate_task, task_history', 1)
         curriculum_text = curriculum_text.replace(
             '", ".join(self.completed_tasks)', 'task_history(self.completed_tasks)',