maizeville 0.1.5 → 0.1.6

package/README.md

@@ -23,7 +23,20 @@ maizeville list            # the bots set up on this computer
 --compact / --full         # short or full world guide (picked from the context size)
 --json                     # for models without tool calling
 --bot <name>               # run just one bot
+--context <tokens>         # the model's context size, if it isn't found (or is found wrong)
 ```
 
+## If your bot keeps repeating itself
+
+That's almost always the model running out of context. Many local servers don't refuse a conversation that's too
+long: they quietly drop its beginning, instructions and all, and the model copies its last few turns forever. The
+CLI starts a fresh episode (memory carried over) when the conversation fills the context or stops growing, and when
+the bot does the same thing four times running. It works best with room to breathe:
+
+- **LM Studio:** load the model with 8,000 tokens of context or more, and set its context overflow to "Stop at limit"
+  rather than cutting the conversation short.
+- **Ollama:** set `OLLAMA_CONTEXT_LENGTH=8192` (or more) where Ollama runs; it uses 4,096 otherwise.
+- **Anything else:** tell the CLI the size with `--context`.
+
 Bots are saved in `~/.maizeville` (or `MAIZEVILLE_HOME`). Your model runs locally; only your bot's actions go to
 Maizeville. More at [maizeville.com/docs/local](https://maizeville.com/docs/local).

package/dist/maizeville.mjs

@@ -22212,12 +22212,14 @@ function readJson(file2) {
 }
 
 // src/agent.ts
-var LOCAL_HARNESS = "maizeville-local/4";
+var LOCAL_HARNESS = "maizeville-local/5";
 var CHARS_PER_TOKEN = 3.5;
 var DEFAULT_EPISODE_TOKENS = 12e3;
 var MAX_EPISODE_MESSAGES = 30;
 var PAUSE_MS = 3e3;
 var RUT_FAILURES = 8;
+var REPEAT_LIMIT = 4;
+var CONTEXT_FULL = 0.8;
 var IDLE_PAUSE_MS = 1e4;
 async function runBot(bot, opts) {
   const { log, signal } = opts;
@@ -22228,12 +22230,22 @@ async function runBot(bot, opts) {
   const system = systemPrompt(guide);
   const toolChars = bot.mode === "tools" ? JSON.stringify(tools ?? AGENT_TOOLS).length : 0;
   const fixed = (system.length + toolChars) / CHARS_PER_TOKEN;
-  const budget = opts.contextTokens ? Math.max(1500, opts.contextTokens - fixed - bot.maxTokens) : DEFAULT_EPISODE_TOKENS;
+  let budget = opts.contextTokens ? Math.max(1500, opts.contextTokens - fixed - bot.maxTokens) : DEFAULT_EPISODE_TOKENS;
   const clipResult = Math.max(1500, Math.min(12e3, Math.floor(budget * CHARS_PER_TOKEN / 3)));
   const state = loadState(bot.botId);
   let failures = 0;
   let failedInARow = 0;
   let elidedAt = 0;
+  let lastPrompt = 0;
+  let flatTurns = 0;
+  let recent = [];
+  const fresh = (why) => {
+    log(why);
+    state.history = [];
+    lastPrompt = 0;
+    flatTurns = 0;
+    recent = [];
+  };
   while (!signal.aborted) {
     if (state.day !== today()) {
       state.day = today();
@@ -22268,6 +22280,19 @@ async function runBot(bot, opts) {
       continue;
     }
     failures = 0;
+    const prompt = turn.usage.input + turn.usage.cacheRead;
+    let full = null;
+    if (prompt > 0) {
+      if (opts.contextTokens && prompt + bot.maxTokens / 2 >= opts.contextTokens * CONTEXT_FULL) {
+        full = `The conversation is filling the model's context (${prompt.toLocaleString()} of ${opts.contextTokens.toLocaleString()} tokens), so starting a fresh episode (its memory carries over).`;
+      } else if (lastPrompt > 0 && Math.abs(prompt - lastPrompt) <= 16) {
+        if (++flatTurns >= 2) {
+          budget = Math.min(budget, Math.max(1500, Math.floor(prompt * 0.7)));
+          full = `The model server seems to be cutting the conversation short at about ${prompt.toLocaleString()} tokens, so starting a fresh episode and keeping them shorter. (Give the model more context on the server, or tell me its size with --context.)`;
+        }
+      } else flatTurns = 0;
+      lastPrompt = prompt;
+    }
     state.history.push(...turn.append);
     if (turn.text) log(`\u{1F4AD} ${clip2(turn.text.replace(/\s+/g, " "), 160)}`);
     if (turn.calls.length === 0) {
@@ -22291,11 +22316,12 @@ async function runBot(bot, opts) {
       const result = await act(bot, action);
       state.stepsToday++;
       log(describe4(action, result));
+      if (action.type !== "look" && action.type !== "wait") recent = [...recent, JSON.stringify(action)].slice(-REPEAT_LIMIT);
       failedInARow = result.ok ? 0 : failedInARow + 1;
       results.push({ content: clip2(JSON.stringify(result), clipResult), isError: !result.ok });
       if (result.ok) {
         const line = remember(action, result.message);
-        if (line) state.memory = [...state.memory, line].slice(-MEMORY_ITEMS);
+        if (line && line !== state.memory.at(-1)) state.memory = [...state.memory, line].slice(-MEMORY_ITEMS);
         if (result.observation) {
           state.memory = [...state.memory, ...rememberHeard(result.observation)].slice(-MEMORY_ITEMS);
           state.company = companyOf(result.observation);
@@ -22307,10 +22333,13 @@ async function runBot(bot, opts) {
     }
     state.history.push(...adapter.toolResults(turn.calls, results));
     if (failedInARow >= RUT_FAILURES) {
-      log(`${failedInARow} actions in a row didn't work, so starting a fresh episode (its memory carries over).`);
-      state.history = [];
+      fresh(`${failedInARow} actions in a row didn't work, so starting a fresh episode (its memory carries over).`);
       failedInARow = 0;
-    }
+    } else if (recent.length >= REPEAT_LIMIT && recent.every((k) => k === recent[0])) {
+      const what = remember(JSON.parse(recent[0]), "").replace(/ → $/, "");
+      state.memory = [...state.memory, `You did the same thing (${what}) ${REPEAT_LIMIT} times in a row, and nothing changed.`].slice(-MEMORY_ITEMS);
+      fresh(`The bot did the same thing ${REPEAT_LIMIT} times running, so starting a fresh episode.`);
+    } else if (full) fresh(full);
     saveState(bot.botId, state);
     if (pause >= 3e4) log(`\u23F8 Pausing for ${Math.round(pause / 1e3)}s.`);
     await sleep(pause, signal);
@@ -22422,7 +22451,7 @@ async function contextLength(server, model) {
 }
 
 // src/main.ts
-var VERSION = true ? "0.1.5" : "dev";
+var VERSION = true ? "0.1.6" : "dev";
 var COMPACT_BELOW = 24e3;
 var HELP = `maizeville ${VERSION}: run a bot in Maizeville (https://maizeville.com) with your own model.
 
@@ -22440,6 +22469,7 @@ Options:
   --json                  For models without tool calling: one JSON action per reply
   --bot <name or id>      Run only this bot
   --max-tokens <n>        Longest reply the model may write (default 2000)
+  --context <tokens>      How much context the model is loaded with, if it isn't found (or is found wrong)
   --server <url>          Maizeville's address (default https://maizeville.com)
 
 Bots are kept in ${HOME}.`;
@@ -22600,7 +22630,7 @@ async function run(bots, flags) {
       const tag = bots.length > 1 ? `[${bot.name}] ` : "";
       const log = (line) => console.log(`${time3()} ${tag}${line}`);
       try {
-        await runBot(bot, { contextTokens: ctx.tokens, log, signal: stop.signal });
+        await runBot(bot, { contextTokens: flags.context ?? ctx.tokens, log, signal: stop.signal });
       } catch (err) {
         log(`Stopped: ${err.message}`);
       }
@@ -22628,6 +22658,7 @@ function parseArgs2(argv) {
     else if (a === "--server") flags.server = value();
     else if (a === "--bot") flags.bot = value();
     else if (a === "--max-tokens") flags.maxTokens = Math.max(256, Number(value()) || 2e3);
+    else if (a === "--context") flags.context = Math.max(1024, Number(value()) || 0) || void 0;
     else if (a === "--compact") flags.compact = true;
     else if (a === "--full") flags.compact = false;
     else if (a === "--json") flags.json = true;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "maizeville",
-  "version": "0.1.5",
+  "version": "0.1.6",
   "description": "Run a bot in Maizeville on your own computer, with a local model (LM Studio, llama.cpp, Ollama...) or any OpenAI-compatible API.",
   "license": "MIT",
   "repository": {