castle-web-cli 0.4.32 → 0.4.34

package/dist/agent-prompts.d.ts

@@ -28,5 +28,6 @@ export declare function buildTaskPrompt(opts: {
     progressPath: string;
     notesPath: string;
     depsSummary?: string;
+    backend?: 'cursor' | 'claude';
 }): string;
 export declare const CLAUDE_TASK_SYSTEM_REMINDER = "Castle background task agent: work autonomously to completion -- never pause to ask questions or wait for confirmation; finish the task end-to-end, then stop. Prefer the quickest viable change that fully does the job.";

package/dist/agent-prompts.js

@@ -21,8 +21,10 @@ Hard rules:
 short imperative title on the first line
 after: comma-separated titles or ids this task must wait for (optional line)
 supersedes: comma-separated titles or ids this task replaces (optional line)
-Then a detailed, self-contained prompt for the implementing agent:
-what to build or fix, which files matter, and what "done" looks like.
+Then a SHORT self-contained prompt -- one tight paragraph (aim under 100
+words): what to build or fix and what "done" looks like. Task agents read
+the deck's own docs for framework/API detail, so never restate recipes,
+file layouts, or implementation steps.
 \`\`\`
 
 - Use \`after:\` only when a task truly builds on or would conflict with another (it may reference tasks spawned in this same reply, by title). Independent tasks must NOT wait on each other.
@@ -101,6 +103,16 @@ export function buildTaskPrompt(opts) {
     const deps = opts.depsSummary
         ? `\n\nThis task waited on earlier tasks:\n${opts.depsSummary}\n`
         : '';
+    // Claude-only: collapsing the wrap-up (progress 90 + restart + notes) into
+    // one shell call reliably saves 1-2 serial ~7s turns there. Cursor's
+    // composer sometimes reacts to the same rule with MORE calls, so it stays
+    // on the plain instructions. Read-before-overwrite likewise targets
+    // claude's Write-tool precondition (a blind Write to an existing file is
+    // rejected, forcing a wasted failed-call + Read + retry cycle).
+    const wrapUp = opts.backend === 'claude'
+        ? `\n- NEVER point your Write tool at a file that already exists (scenes/main.scene especially): it rejects files it has not Read, and the failed call + forced Read wastes two turns. Replace existing files with ONE shell command instead (\`cat > scenes/main.scene <<'EOF' ... EOF\`); reserve the Write tool for brand-new files.
+- Wrap up in ONE tool call, not several: once your last file edit is done, combine the 90-progress write, the final \`npm run restart\`, and writing the notes file into a single shell command (\`;\`-separated so the notes land even if the restart hiccups). Then stop -- no extra turns after it.`
+        : '';
     return `You are a background build agent for the Castle deck "${opts.deckLabel}" (current directory). A separate conversation agent dispatched you with one task. Read the deck's CLAUDE.md / AGENTS.md first and follow its workflow (including reloading the served deck after changes, e.g. \`npm run restart\`).
 
 Your task (id ${opts.taskId}): ${opts.title}
@@ -113,7 +125,7 @@ Operating rules:
 - The moment implementation is complete and you switch to verifying, write 90 to the progress file -- verification time must not read as stalled progress.
 - Do this one task completely, then stop. Do not expand scope.
 - Update your progress VERY frequently: write a bare integer 0-100 to ${opts.progressPath} (e.g. \`echo 30 > ${opts.progressPath}\`) every time you advance -- at least every 10 points, or every 20 for properly small tasks. Start near 10, write 90 just before wrapping up. Never let it sit stale while you work.
-- Before finishing, write ${opts.notesPath}: a SHORT test guide for the user -- 2-4 brief sentences. Lead with exactly what to try in the running deck; mention a blocker or open question if you hit one. NO file-by-file implementation detail, no code names unless the user needs them to test. The user reads this verbatim when checking your work off.
+- Before finishing, write ${opts.notesPath}: a SHORT test guide for the user -- 2-4 brief sentences. Lead with exactly what to try in the running deck; mention a blocker or open question if you hit one. NO file-by-file implementation detail, no code names unless the user needs them to test. The user reads this verbatim when checking your work off.${wrapUp}
 - If you are truly blocked, write the blocker to the notes file and stop rather than guessing wildly.
 - Never touch files under .castle/ other than those two paths.`;
 }

package/dist/agent.js

@@ -428,10 +428,15 @@ async function runTaskAgentIn(ctx, task) {
         progressPath: path.join(relDir, 'progress'),
         notesPath: path.join(relDir, 'notes.md'),
         depsSummary: ctx.depsSummary,
+        backend: ctx.backend,
     });
-    // Claude task runs get /goal (claude code's built-in commit-to-completion
-    // command) plus the system-prompt autonomy reminder.
-    const invocation = buildAgentInvocation(ctx.backend, 'task', ctx.backend === 'claude' ? `/goal ${taskPrompt}` : taskPrompt, ctx.claudeModel);
+    // No /goal wrapper: it makes a fresh evaluator re-check the WHOLE task
+    // prompt (including user-only "done when you reach wave 5"-style play
+    // criteria) after every turn, which traps the agent in an endless
+    // can't-verify-live-play loop. A single message + the autonomy reminder
+    // appended to the system prompt (CLAUDE_TASK_SYSTEM_REMINDER) is enough to
+    // get the work done; the agent stops when the code is in, the user verifies.
+    const invocation = buildAgentInvocation(ctx.backend, 'task', taskPrompt, ctx.claudeModel);
     let result = { ok: false, finalText: '', error: 'not run' };
     let lineBuf = '';
     const flushFeedLines = (delta) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "castle-web-cli",
-  "version": "0.4.32",
+  "version": "0.4.34",
   "type": "module",
   "bin": {
     "castle-web": "./dist/index.js"