castle-web-sdk 0.4.40 → 0.4.42

package/README.md

@@ -731,9 +731,11 @@ be coalesced, so it's safe to call on frequent events.
 ## Agent
 
 `Agent` asks a language model for one chat completion. The host sends it
-through Castle's oracle on the player's own Castle account, so the deck
-holds no key. Agent.chat is limited to Castle admins for now; every other
-player gets `'refused'`.
+through Castle's deck-agent service, so the deck holds no key. Castle pays
+for each player's calls out of a daily play allowance, which Castle
+characters draw from too; once a player has used it up, calls resolve
+`'allowance_exhausted'` until it resets at midnight UTC. Agent.chat is
+limited to Castle admins for now; every other player gets `'refused'`.
 
 ### `Agent.chat(params): Promise<AgentChatResult>`
 
@@ -745,9 +747,11 @@ player gets `'refused'`.
 - `tools` — optional function tools, in the OpenAI function-tool schema.
 - `toolChoice` — optional: `'auto'` (the default), `'none'`, `'required'`,
   or `{ type: 'function', function: { name } }` to force one tool.
-- `model` — optional OpenRouter model id. Defaults to
-  `'meta/muse-spark-1.3'`.
-- `maxTokens` — optional cap on the reply's length.
+- `model` — optional OpenRouter model id: `'meta/muse-spark-1.3-contributor'`
+  (the default), `'meta/muse-spark-1.3'` or `'openai/gpt-4.1-mini'`. Any
+  other model resolves `'error'`.
+- `maxTokens` — optional cap on the reply's length. Defaults to 4096; the
+  service caps it at 16384.
 - `reasoning` — optional `{ effort: 'minimal' | 'low' | 'medium' | 'high' }`,
   how hard the model thinks before it answers. Lower effort answers sooner.
   Left out, the model uses its own default. Muse can't turn reasoning off;
@@ -819,9 +823,12 @@ for (let round = 0; round < 8; round++) {
   `'tool_calls'` or `'length'`. `usage` carries `prompt_tokens`,
   `completion_tokens` and `total_tokens`.
 - `'refused'` — the player is not a Castle admin. The model was not called.
-- `'unavailable'` — this host has no route to the oracle.
+- `'allowance_exhausted'` — the player's play allowance for the day is used
+  up. The model was not called. It resets at midnight UTC.
+- `'unavailable'` — this host has no route to the deck-agent service, or
+  the service is down.
 - `'cancelled'` — the call's `signal` was aborted.
-- `'error'` — the oracle or the model failed; `reason` says how.
+- `'error'` — the service or the model failed; `reason` says how.
 
 Every status but `'ok'` carries a `reason`.
 

package/dist/agent.js

@@ -1,12 +1,15 @@
-// Agent — one chat completion from a language model, run by the host through
-// Castle's oracle with the player's own Castle account. Messages and tools
-// follow the OpenAI chat completions format. The reply streams: `onDelta` gets
-// text and reasoning as they arrive and each tool call once its arguments are
-// complete, and the promise resolves with the whole message. A deck runs any
-// tool calls itself and calls Agent.chat again with the results.
+// Agent — one chat completion from a language model, run through Castle's
+// deck-agent service. Castle funds each player's calls from a daily play
+// allowance that characters share. Messages and tools follow the OpenAI chat
+// completions format. The reply streams: `onDelta` gets text and reasoning as
+// they arrive and each tool call once its arguments are complete, and the
+// promise resolves with the whole message. A deck runs any tool calls itself
+// and calls Agent.chat again with the results.
 //
 // Agent.chat is limited to Castle admins for now: every other player gets
-// `refused`. A host with no route to the oracle answers `unavailable`.
+// `refused`. A player whose allowance for the day is spent gets
+// `allowance_exhausted`. A host with no route to the service answers
+// `unavailable`.
 import { CastleError } from "./errors";
 import { hostRequest } from "./transport";
 async function chat(params) {

package/dist/commands.d.ts

@@ -255,7 +255,7 @@ export type AgentChatResult = {
     finishReason: string | null;
     usage?: AgentUsage;
 } | {
-    status: "unavailable" | "refused" | "error" | "cancelled";
+    status: "unavailable" | "refused" | "allowance_exhausted" | "error" | "cancelled";
     reason: string;
 };
 export interface DeckSummary {

package/dist/commands.js

@@ -15,7 +15,7 @@ export const COMMAND_TIMEOUTS_MS = {
     "decks.saved": 30_000,
 };
 // Once a host streams a command, each stream envelope restarts the wait at this
-// length instead. The oracle sends a heartbeat every 15 seconds, and the host
+// length instead. The deck-agent service sends a heartbeat every 15 seconds, and the host
 // passes each one on.
 export const STREAM_IDLE_TIMEOUT_MS = 60_000;
 export function isResponseEnvelope(value) {

package/dist/runtime.js

@@ -8,6 +8,8 @@ const HOST_CAPTURE_REQUEST = "castle-capture-cover";
 const HOST_CAPTURE_RESULT = "castle-capture-cover-result";
 let ws = null;
 let logBuffer = [];
+// The shell's play log keeps the same number of lines.
+const MAX_BUFFERED_MESSAGES = 500;
 let nextRequestId = 1;
 const pendingRequests = new Map();
 // Local-dev command channel: the `castle-web serve` dev server is the host, so
@@ -306,6 +308,11 @@ function sendMsg(msg) {
     }
     else {
         logBuffer.push(msg);
+        // A deck logging every frame fills this while no socket is open, so the
+        // oldest messages are dropped.
+        if (logBuffer.length > MAX_BUFFERED_MESSAGES) {
+            logBuffer.splice(0, logBuffer.length - MAX_BUFFERED_MESSAGES);
+        }
     }
 }
 function sendLocalRequest(msg) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "castle-web-sdk",
-  "version": "0.4.40",
+  "version": "0.4.42",
   "repository": {
     "type": "git",
     "url": "git+https://github.com/castle-xyz/castle-experimental-web.git"