castle-web-cli 0.4.183 → 0.4.184

package/dist/agent-failures.d.ts

@@ -11,6 +11,7 @@ export interface AgentFailure {
     castleCreditsExhausted?: boolean;
     castleCreditsInsufficient?: boolean;
     castleSpendLimit?: boolean;
+    castleFreeModelLimit?: boolean;
 }
 export declare function failureForStatus(status: number, body: string, model?: string): AgentFailure | undefined;
 export declare function classifyProviderError(text: string | undefined, model?: string): AgentFailure | undefined;

package/dist/agent-failures.js

@@ -48,10 +48,32 @@ export function failureForStatus(status, body, model) {
         return config("no-credits");
     if (status === 403) {
         // Castle's proxy deliberately uses provider-shaped 403s for its own spend
-        // and model gates. A request-size reservation can refuse while the account
-        // still has a small positive balance, so pre-flight cannot catch this case.
-        // Keep the matches tied to the proxy's Castle-specific copy: a provider's
-        // ordinary permission/content 403 must continue to read as a refusal.
+        // and model gates. Keep the matches tied to the proxy's Castle-specific
+        // copy: a provider's ordinary permission/content 403 must continue to read
+        // as a refusal.
+        //
+        // The free model's own daily cap is matched FIRST. It is about neither the
+        // balance nor the account-wide cap, so whatever else the proxy's sentence
+        // happens to mention, any credit copy would send the user off to buy
+        // something that cannot help them.
+        if (/free model daily limit reached/i.test(body)) {
+            // The proxy names the reset as an ISO instant in its prose ("It resets at
+            // 2026-09-12T07:00:00.000Z."); it is the only machine-readable field the
+            // provider-shaped body has room for, and it is what lets the copy promise
+            // a time instead of just "today".
+            const resetAt = /resets at (\d{4}-\d{2}-\d{2}T[\d:.]+Z)/i.exec(body);
+            const resetAtMs = resetAt ? Date.parse(resetAt[1]) : Number.NaN;
+            return {
+                kind: "limit",
+                detail,
+                model,
+                verbose: `HTTP ${status}: ${body}`,
+                castleFreeModelLimit: true,
+                ...(Number.isFinite(resetAtMs) ? { resetAtMs } : {}),
+            };
+        }
+        // A request-size reservation can refuse while the account still has a
+        // small positive balance, so pre-flight cannot catch this case.
         if (/Castle AI request[^.]*balance can(?:not|'t) cover|add credits at castle\.xyz\/credits/i.test(body)) {
             return {
                 kind: "limit",
@@ -205,6 +227,11 @@ export function failureCopy(opts) {
         case "config":
             return `${configCopy(opts.failure)}${tasksNote}`;
         case "limit":
+            // Before the credit cases for the same reason it classifies first: this
+            // user is not spending credits, so no amount of buying them helps.
+            if (opts.failure.castleFreeModelLimit) {
+                return `Today's allowance for the free model is used up${resetsClause(opts.failure.resetAtMs)}. Pick another model in settings to keep going.${tasksNote}`;
+            }
             if (opts.failure.castleCreditsInsufficient) {
                 return `Not enough Operator credits for this request — get more at castle.xyz/credits${tasksNote}`;
             }

package/dist/agent-prompts.d.ts

@@ -83,7 +83,6 @@ export interface RouterPromptOpts {
     deckLabel: string;
     quickReference?: string;
     deckTree?: string;
-    deckContents?: string;
     messages: PromptMessage[];
     tasks: PromptTask[];
     plan?: RouterPlanOpts;
@@ -115,9 +114,7 @@ export declare function buildTaskPrompt(opts: {
     notesPath: string;
     handoffPath: string;
     depsSummary?: string;
-    backend?: "cursor" | "claude" | "smith";
     deckTree?: string;
-    deckContents?: string;
     quickReference?: string;
     siblings?: PromptSibling[];
     plan?: TaskPlanOpts;

package/dist/agent-prompts.js

@@ -3,7 +3,7 @@
 // actual deck edits. Kept separate from agent.ts so the orchestration code
 // stays readable.
 // How many trailing conversation messages get replayed into each router call.
-// Router calls are stateless (a fresh cursor-agent print run each time) so the
+// Router calls are stateless (a fresh headless run each time) so the
 // transcript is the only memory.
 const TRANSCRIPT_LIMIT = 40;
 // Byte ceiling on the replayed transcript, applied AFTER TRANSCRIPT_LIMIT.
@@ -16,9 +16,8 @@ const TRANSCRIPT_LIMIT = 40;
 // needs a byte budget.
 //
 // Sized to leave room for everything else the prompt carries: the rules
-// (~10KB), the deck's quick reference (~13KB), the file tree, the smith-only
-// deck contents (ROUTER_DECK_CONTENTS_BUDGET, 40KB), the task board, and this
-// turn's instruction.
+// (~10KB), the deck's quick reference (~13KB), the file tree, the task board,
+// and this turn's instruction.
 const TRANSCRIPT_BYTE_BUDGET = 32 * 1024;
 // The playtest steering's router half, measured as one added bullet on the t1
 // router ladder (memo 26-09-03-14-01): 15/18 written when asked, 6/6 withheld on
@@ -834,9 +833,6 @@ export function buildRouterPromptParts(opts) {
     const deckFiles = opts.deckTree?.trim()
         ? `\n\n== deck files ==\n${opts.deckTree.trim()}`
         : "";
-    const deckSource = opts.deckContents?.trim()
-        ? `\n\n== deck source (current contents -- no need to read_file these to see what's already there) ==\n${opts.deckContents.trim()}`
-        : "";
     // Sits with the board, not with the deck identity above: the plan changes on
     // roughly the turns the board does, so co-locating them means it invalidates
     // nothing the board wasn't already invalidating.
@@ -844,16 +840,16 @@ export function buildRouterPromptParts(opts) {
         ? `\n\n== plan ==\n${renderPlanDigest(opts.plan, opts.tasks)}`
         : "";
     // Two parts, split for prompt caching. `system` holds what is the same from
-    // one turn to the next within a serve -- the rules, the deck identity, the
-    // file tree and (smith) the deck source -- and each backend sends it as
-    // system text, which is the prefix the provider caches turn after turn. A
-    // file edit invalidates it; a chat turn does not. `user` holds what changes
-    // every turn: the transcript, the plan, the board and this instruction.
+    // one turn to the next within a serve -- the rules, the deck identity and
+    // the file tree -- and is sent as system text, which is the prefix the
+    // provider caches turn after turn. A file edit invalidates it; a chat turn
+    // does not. `user` holds what changes every turn: the transcript, the plan,
+    // the board and this instruction.
     return {
         system: `${routerRules(opts.plan !== undefined, opts.playtest === true)}
 
 == deck ==
-${opts.deckLabel}${quickReference}${deckFiles}${deckSource}`,
+${opts.deckLabel}${quickReference}${deckFiles}`,
         user: `== conversation so far ==
 ${renderTranscript(opts.messages)}${plan}
 
@@ -899,10 +895,7 @@ export function userTurnInstruction(opts) {
             .join("\n\n")}`);
     }
     if (opts.attachments && opts.attachments.length > 0) {
-        // Wording covers both delivery modes: smith gets the images inline as
-        // content blocks in this same message (its read tool rejects binaries),
-        // while the CLI backends open the saved files themselves.
-        parts.push(`The user attached image file(s), saved in the deck at: ${opts.attachments.join(", ")}. If the images are not already visible in this message, open them with your read tool. Take them into account; pass the paths along to task agents that need them.`);
+        parts.push(`The user attached image file(s), saved in the deck at: ${opts.attachments.join(", ")}. Open them with your read tool. Take them into account; pass the paths along to task agents that need them.`);
     }
     if (opts.refs && opts.refs.length > 0) {
         parts.push(`The user pointed at: ${opts.refs.map(refPhrase).join(", ")}. Read them with your read tool if you need them to reply. When you spawn a task one of these concerns, name it the same way in that task's prompt -- the path and the lines, or the path and the param -- so the agent reads that place; do not mention it to tasks it does not concern.`);
@@ -928,9 +921,6 @@ export function buildTaskPrompt(opts) {
     const layout = opts.deckTree?.trim()
         ? `\n\nDeck files (snapshot at task start; parallel sibling tasks may create files named in your prompt that are absent here):\n\n${opts.deckTree.trim()}\n`
         : "";
-    const deckSource = opts.deckContents?.trim()
-        ? `\n\nDeck source (complete current contents of the deck's text files, snapshot at task start -- you do NOT need to read these files before editing them, edit directly; a file marked "not inlined" is either binary/generated or didn't fit the budget, so read it yourself first if you need it):\n\n${opts.deckContents.trim()}\n`
-        : "";
     const planSliceText = opts.plan
         ? renderTaskPlanSlice(opts.plan.fileText ?? "", opts.plan.item, opts.plan.finishedWork)
         : "";
@@ -945,20 +935,16 @@ export function buildTaskPrompt(opts) {
             ? `\n- Before you finish, write these lines to ${opts.handoffPath} -- they are the ONLY record of this task once the task itself is forgotten:\n  - \`what: <what now exists, and where>\` -- an existence fact, not a report of activity: "a shop panel in ui/shop.js the bag button opens", never "updated the shop logic". This is what the plan shows as already built, so a later agent extends it instead of rebuilding it.\n  - \`files: <the files you created or changed, comma-separated>\`\n  - \`knobs: <the values worth tuning, comma-separated>\` -- names only, and only if there are real ones.\n  - \`durable: <fact>\` -- the single thing worth remembering that is NOT covered by the three above (a preference the user expressed, a constraint you ran into, an approach that did not work). At most 200 characters. Leave it out only if nothing genuinely qualifies.`
             : "",
     ].join("");
-    // Claude-only: collapsing the wrap-up (progress 90 + notes) into
-    // one shell call reliably saves 1-2 serial ~7s turns there. Cursor's
-    // composer sometimes reacts to the same rule with MORE calls, so it stays
-    // on the plain instructions.
+    // Collapsing the wrap-up (progress 90 + notes) into one shell call reliably
+    // saves 1-2 serial ~7s turns.
     // (We deliberately do NOT tell agents to overwrite files via `cat > … <<EOF`
     // -- blind whole-file rewrites made parallel agents clobber each other's
     // edits. Read-then-edit is slower but safe; that is the right tradeoff.)
     const wrapUpList = opts.plan
         ? "the 90-progress write, the `castle-web save-version` for your paths, writing the notes file, and the handoff lines"
         : "the 90-progress write, the `castle-web save-version` for your paths, and writing the notes file";
-    const wrapUp = opts.backend === "claude" || opts.backend === "smith"
-        ? `\n- Wrap up in ONE tool call, not several: once your last file edit is done, combine ${wrapUpList} into a single shell command (\`;\`-separated so the notes land even if an earlier part hiccups). Then stop -- no extra turns after it.`
-        : "";
-    return `You are a background build agent for the Castle deck "${opts.deckLabel}" (current directory). A separate conversation agent dispatched you with one task. Follow the deck's CLAUDE.md / AGENTS.md conventions. Do not reload the served deck -- the person applies your changes when they are ready. \`npm run restart\` exists, but it is HIGHLY flow breaking: it reloads every panel they have open, including the one they are working in. Use it only when a change clearly requires restarting their whole experience.${quickReference}${layout}${deckSource}
+    const wrapUp = `\n- Wrap up in ONE tool call, not several: once your last file edit is done, combine ${wrapUpList} into a single shell command (\`;\`-separated so the notes land even if an earlier part hiccups). Then stop -- no extra turns after it.`;
+    return `You are a background build agent for the Castle deck "${opts.deckLabel}" (current directory). A separate conversation agent dispatched you with one task. Follow the deck's CLAUDE.md / AGENTS.md conventions. Do not reload the served deck -- the person applies your changes when they are ready. \`npm run restart\` exists, but it is HIGHLY flow breaking: it reloads every panel they have open, including the one they are working in. Use it only when a change clearly requires restarting their whole experience.${quickReference}${layout}
 
 Your task (id ${opts.taskId}): ${opts.title}
 

package/dist/agent.d.ts

@@ -1,24 +1,13 @@
 import * as http from 'http';
 import { Duplex } from 'stream';
-import type { ORReasoningEffort, ORRoutingMode } from './native/openrouter.js';
 import { type PlaytestToolResult } from './native/playtest.js';
 export declare const AGENT_WS_PATH = "/__castle/agent";
-export declare const AGENT_MODEL_CAPS_PREFIX = "/__castle/agent/model-caps";
-export type AgentBackend = 'cursor' | 'claude' | 'smith';
 export type ClaudeModel = 'sonnet' | 'opus' | 'fable' | 'openrouter';
 interface AgentSettings {
-    router: AgentBackend;
-    tasks: AgentBackend;
     routerClaudeModel: ClaudeModel;
     tasksClaudeModel: ClaudeModel;
     routerOpenrouterModel: string;
     tasksOpenrouterModel: string;
-    routerReasoningEffort: ORReasoningEffort;
-    tasksReasoningEffort: ORReasoningEffort;
-    routerRouting: ORRoutingMode;
-    tasksRouting: ORRoutingMode;
-    routerProviderTier: string;
-    tasksProviderTier: string;
 }
 type OpenrouterSlugKey = 'routerOpenrouterModel' | 'tasksOpenrouterModel';
 type SettingsWarnings = Partial<Record<OpenrouterSlugKey, {
@@ -58,7 +47,6 @@ interface TaskRecord {
     durablePending?: boolean;
     blockedBy?: string[];
 }
-export declare function roleUsesOpenrouter(backend: AgentBackend, claudeModel: ClaudeModel): boolean;
 export interface DepsState {
     kind: 'ready' | 'waiting' | 'blocked';
     blockedBy?: string[];
@@ -73,7 +61,7 @@ export interface AgentServer {
     /**
      * Run one playtest on the serve's warm Chromium. Backs the `playtest_request`
      * control-WS message (serve.ts), which is how a claude task agent reaches
-     * this through its MCP server -- smith calls runPlaytest directly instead.
+     * this through its MCP server.
      */
     runPlaytestFor(req: {
         taskId?: string;