castle-web-cli 0.4.178 → 0.4.180

package/dist/agent.js

@@ -432,9 +432,13 @@ function handleModelCaps(req, res) {
 }
 // Build the headless CLI invocation for a spawning backend/role (smith never
 // comes through here -- it has no CLI process; see runAgentSmith). Cursor's
-// router runs in read-only ask mode; claude runs permission-mode auto for
-// both roles (NOT plan mode -- that makes it emit plan tool calls) at medium
-// effort. claudeModel "openrouter" keeps the SAME claude binary/flags and
+// router runs in read-only ask mode; claude runs permission-mode
+// bypassPermissions for both roles (NOT plan mode -- that makes it emit plan
+// tool calls) at medium effort. Not `auto`: auto mode runs a sonnet
+// classifier call beside every turn, on Anthropic's own key, whatever --model
+// says -- a second model's cost and a second billing line per turn. bypass
+// asks no classifier. A cli fix is expected upstream; until then bypass is
+// the mode. claudeModel "openrouter" keeps the SAME claude binary/flags and
 // swaps only the model slug and env (see envForOpenrouterSpawn) -- plus one
 // extra permission grant, see OPENROUTER_ALLOWED_TOOLS below.
 // claude CLI 2.1.198 SILENTLY resolves `--permission-mode auto` to "default"
@@ -507,11 +511,15 @@ function claudeSettingsArg(auth) {
 // new process, so that text (the kit guide: ~23k tokens on a kit deck) is sent
 // uncached on every call of every turn. The router already carries the deck's
 // quick reference and file tree in its cached system part, and can read a
-// guide on demand; tasks keep the CLI's own loading. Measured on a 30k-token
-// router prompt: a second turn goes from 40k tokens written to 4.6k.
+// guide on demand; tasks keep the CLI's own loading. Auto memory is off for the
+// same reason: the first process in a deck creates the memory store, the next
+// one's system prompt gains a memory block, and the cached prefix only matches
+// from the third turn on. Measured on a 30k-token router prompt: a second turn
+// goes from 50k tokens written to 6k.
 const ROUTER_CLAUDE_ENV = {
     CLAUDE_CODE_DISABLE_CLAUDE_MDS: '1',
     CLAUDE_CODE_DISABLE_GIT_INSTRUCTIONS: '1',
+    CLAUDE_CODE_DISABLE_AUTO_MEMORY: '1',
 };
 function buildAgentInvocation(backend, role, prompt, claudeModel, 
 // Already resolved for this role by the caller (router turns pass
@@ -549,7 +557,7 @@ systemPrompt) {
                 'stream-json',
                 '--include-partial-messages',
                 '--permission-mode',
-                'auto',
+                'bypassPermissions',
                 ...(viaOpenrouter ? [openrouterAllowedToolsArg(role)] : []),
                 '--model',
                 viaOpenrouter ? openrouterModel : claudeModel,
@@ -2269,6 +2277,8 @@ function persistTaskFile(tasksDir, task) {
 // written whether or not the deck has the plan doc switched on: it is a disk
 // receipt, never injected into a prompt, so it costs an off deck nothing.
 const TASK_INDEX_FILE = 'index.md';
+const TASK_FEED_FILE = 'feed.json';
+const TASK_FEED_CAP = 80;
 const TASK_INDEX_HEADER = '# tasks\n# <task-id> | <item> | <title> | <created> | <status> | <touching>\n';
 // A pipe-delimited row can't carry a pipe. Titles come from the router's fence
 // and can hold anything.
@@ -3224,8 +3234,27 @@ function asClientMessage(message, signPath) {
         attachments: message.attachments.map((name) => clientUrl(`${AGENT_ATTACHMENT_PREFIX}${name}`, signPath)),
     };
 }
-function createTaskFeeds(broadcast) {
+function loadTaskFeeds(tasksDir) {
     const map = new Map();
+    for (const entry of fs.existsSync(tasksDir) ? fs.readdirSync(tasksDir) : []) {
+        const loaded = readJsonFile(path.join(tasksDir, entry, TASK_FEED_FILE));
+        if (!Array.isArray(loaded))
+            continue;
+        const lines = loaded.filter((line) => typeof line === 'string');
+        if (lines.length === 0)
+            continue;
+        map.set(entry, lines.slice(-TASK_FEED_CAP));
+    }
+    return map;
+}
+function createTaskFeeds(broadcast, tasksDir) {
+    const map = loadTaskFeeds(tasksDir);
+    function persist(task) {
+        const feed = map.get(task.id);
+        if (!feed || feed.length === 0)
+            return;
+        atomicWriteFileSync(path.join(tasksDir, task.id, TASK_FEED_FILE), JSON.stringify(feed, null, 2) + '\n');
+    }
     function push(task, entry) {
         let feed = map.get(task.id);
         if (!feed) {
@@ -3235,11 +3264,11 @@ function createTaskFeeds(broadcast) {
         if (feed[feed.length - 1] === entry)
             return;
         feed.push(entry);
-        if (feed.length > 80)
-            feed.splice(0, feed.length - 80);
+        if (feed.length > TASK_FEED_CAP)
+            feed.splice(0, feed.length - TASK_FEED_CAP);
         broadcast({ type: 'task-feed', id: task.id, entry });
     }
-    return { map, push };
+    return { map, push, persist };
 }
 function createMessageThinking(broadcast) {
     const map = new Map();
@@ -4399,7 +4428,7 @@ export function createAgentServer(opts) {
         // Castle's), so the bar follows the change instead of the poll.
         usageFeed.refresh();
     };
-    const taskFeeds = createTaskFeeds(broadcast);
+    const taskFeeds = createTaskFeeds(broadcast, tasksDir);
     const messageThinking = createMessageThinking(broadcast);
     const taskStore = createTaskStore({
         deckDir,
@@ -4421,7 +4450,7 @@ export function createAgentServer(opts) {
         onUpdate: (task) => broadcast({ type: 'task-update', task: asClientTask(task, opts.signPath) }),
         onStarted: () => undefined,
         onRetry: (task, attempt) => addLog(`agent died, retrying (${attempt}/${MAX_TASK_ATTEMPTS}): ${task.title}`),
-        onFinished: (task) => taskFeeds.map.delete(task.id),
+        onFinished: (task) => taskFeeds.persist(task),
         onFeed: (task, entry) => taskFeeds.push(task, entry),
     });
     // The mid-run send queue (queue-by-default, "send now"/Stop interrupt, epoch