@stitchdb/cli 0.10.0 → 0.11.0

package/dist/cli.js

@@ -629,13 +629,28 @@ async function cmdHook(args) {
             const stitch = client(cfg);
             const projectTag = (threadName.split('/')[0] || threadName).toLowerCase();
             const baseUrl = cfg.baseUrl || 'https://db.stitchdb.com';
-            const [thread, memHits, workspaces, fileSummaries, aboutMems] = await Promise.all([
+            // Adaptive sizing — scale what we inject by how much the project
+            // actually has. A 5k-memory pool surfacing 8 random memories is
+            // noise; better to inject 0 and tell Claude to `recall` on demand.
+            const [thread, allProjectMems, workspaces, fileSummaries, aboutMems] = await Promise.all([
                 stitch.thread(threadName).recall({ last: 5 }).catch(() => ({ thread_id: '', recent: [], semantic: [] })),
-                stitch.recall(projectTag, { k: 8 }).catch(() => []),
+                stitch.list({ limit: 1 }).catch(() => []), // probe the pool size cheaply
                 stitch.workspaces.list().catch(() => []),
                 stitch.list({ limit: 12 }).then((all) => all.filter((m) => m.tags.some((t) => t.startsWith('file:')))).catch(() => []),
                 stitch.list({ tag: 'workspace:about', limit: 1 }).catch(() => []),
             ]);
+            // Estimate the workspace's total memory count from a single page
+            // header. We just need to know "small / medium / huge" — exact
+            // count is overkill. (D1 doesn't expose count in list — approximate
+            // by listing 200 once and capping.)
+            const sample = await stitch.list({ limit: 200 }).catch(() => []);
+            const poolSize = Array.isArray(sample) ? sample.length : 0;
+            // Tier the injection: tiny / small / medium / huge.
+            const memSliceK = poolSize <= 50 ? 8 : poolSize < 500 ? 8 : 0;
+            const turnSliceK = (thread.recent?.length ?? 0) < 5 ? 5 : (thread.recent?.length ?? 0) < 100 ? 5 : 2;
+            const memHits = memSliceK > 0
+                ? await stitch.recall(projectTag, { k: memSliceK }).catch(() => [])
+                : [];
             // Look up the current workspace using the client's resolved id; fall
             // back to first if resolveWorkspace failed (shouldn't normally).
             const currentWsId = await stitch.resolveWorkspace().catch(() => null);
@@ -655,8 +670,15 @@ async function cmdHook(args) {
             catch { /* ignore */ }
             const lines = [];
             lines.push('<stitch-context>');
-            lines.push(`Project: ${threadName} · Workspace: ${currentWs?.name || '(unknown)'} · Stitch MCP tools: recall, remember, recall_global, remember_global, thread_recall, thread_append, workspace_setup, file_summary, file_summary_save.`);
+            lines.push(`Project: ${threadName} · Workspace: ${currentWs?.name || '(unknown)'} · ${poolSize} memories indexed · Stitch MCP tools: recall, remember, recall_global, remember_global, thread_recall, thread_append, workspace_setup, file_summary, file_summary_save.`);
             lines.push('');
+            // Huge-pool mode: don't pre-load any memories. Tell Claude what's
+            // available and that it should pull what's relevant on demand.
+            if (poolSize >= 200 && memHits.length === 0) {
+                lines.push('### Memory pool');
+                lines.push(`This project has ${poolSize}+ memories. Don't try to load them all — call \`recall(query)\` whenever the user references prior decisions, code patterns, or preferences. Cheaper and more relevant than reading them upfront.`);
+                lines.push('');
+            }
             // Nudge the AI to set a meaningful workspace name once.
             if (currentWs?.name === 'default') {
                 lines.push('### ⚠ Workspace is still named "default"');
@@ -705,13 +727,17 @@ async function cmdHook(args) {
             }
             if (thread.recent && thread.recent.length > 0) {
                 lines.push('### Most recent turns (continue from here)');
-                for (const t of thread.recent.slice(-5)) {
+                for (const t of thread.recent.slice(-turnSliceK)) {
                     const txt = String(t.content || '').replace(/\n+/g, ' ').slice(0, 300);
                     lines.push(`- **${t.role}**: ${txt}`);
                 }
                 lines.push('');
             }
-            lines.push('Call `recall` for project memory, `recall_global` for cross-project user prefs, `thread_recall` for older turns, `file_summary` BEFORE reading any non-trivial file. Save user-level habits/preferences with `remember_global`; project facts with `remember`.');
+            lines.push('## How to use this memory layer');
+            lines.push('- The user mentions prior work, decisions, or "what we discussed" → call `recall(query)` (or `thread_recall(thread, semantic: ...)` for verbatim turns) BEFORE answering. Don\'t guess from the snippets above; they\'re a teaser, not the truth.');
+            lines.push('- About to open a non-trivial file → call `file_summary(path)` first. If hash matches the cached summary, you can skip the full read entirely.');
+            lines.push('- Learn a durable user-level habit ("I always use Postgres", "commit before reverting") → save with `remember_global` so it surfaces across every project. Project-specific facts → `remember`.');
+            lines.push('- A surfaced memory has `_(from thread, N turns)_` next to it → those turns are queryable via `thread_recall`. Reach for them when challenged on a fact.');
             lines.push('</stitch-context>');
             process.stdout.write(lines.join('\n'));
         }
@@ -735,19 +761,97 @@ async function cmdHook(args) {
     content = content.trim();
     if (!content)
         return;
+    // For UserPromptSubmit: log the turn AND fetch task-relevant memory in
+    // parallel, so the model sees only what's actually relevant to *this*
+    // prompt — not 8 random preloaded memories from session start. Task
+    // changes self-surface because the recall hits change with the prompt.
+    if (eventName === 'UserPromptSubmit') {
+        await Promise.all([
+            (async () => {
+                try {
+                    const stitch = client(cfg);
+                    await stitch.thread(threadName).append({ role: role, content });
+                }
+                catch { /* silent */ }
+            })(),
+            handleUserPromptInjection(cfg, content),
+        ]);
+        return;
+    }
+    // Stop: log the assistant turn, then maybe-distill.
     try {
         const stitch = client(cfg);
-        await stitch.thread(threadName).append({ role, content });
-    }
-    catch {
-        /* silent */
+        await stitch.thread(threadName).append({ role: role, content });
     }
-    // After Stop, opportunistically kick off a distillation pass in the
-    // background (fire-and-forget). Won't block the session; debouncing
-    // (cooldown + min-new-turns) is enforced inside maybeAutoDistill.
-    if (eventName === 'Stop') {
-        maybeAutoDistill(threadName).catch(() => { });
+    catch { /* silent */ }
+    maybeAutoDistill(threadName).catch(() => { });
+}
+/**
+ * Per-prompt smart context injection. Runs `recall(prompt)` against the
+ * project workspace + the user's `_global` workspace, score-gates the hits,
+ * and writes a Claude-Code-flavoured `additionalContext` JSON object to
+ * stdout. Time-bounded so a slow proxy doesn't delay the user's prompt.
+ *
+ * Why per-prompt: a static SessionStart block must guess what'll matter
+ * across the whole session. A per-prompt recall sees the actual question
+ * and pulls memory tailored to *that* — much higher signal per token.
+ * A task switch ("now let's do the dashboard") naturally surfaces the
+ * dashboard memories; an unrelated follow-up ("ok run the tests") returns
+ * low scores and we inject nothing — no noise.
+ */
+const PROMPT_INJECTION_TIMEOUT_MS = 2000;
+const PROMPT_INJECTION_MIN_SCORE = 0.4;
+async function handleUserPromptInjection(cfg, prompt) {
+    if (!prompt || prompt.length < 10)
+        return;
+    try {
+        const baseUrl = cfg.baseUrl || 'https://db.stitchdb.com';
+        const stitch = client(cfg);
+        const tOut = new Promise((_, rej) => setTimeout(() => rej(new Error('timeout')), PROMPT_INJECTION_TIMEOUT_MS));
+        // Project recall + global recall in parallel; tolerate either failing.
+        const projectHitsP = stitch.recall(prompt, { k: 3 }).catch(() => []);
+        const globalHitsP = (async () => {
+            try {
+                const ws = await stitch.workspaces.list();
+                const g = ws.find((w) => w.name === '_global');
+                if (!g)
+                    return [];
+                const gc = new Stitch({ apiKey: cfg.apiKey, baseUrl, workspace: g.id });
+                return gc.recall(prompt, { k: 2 }).catch(() => []);
+            }
+            catch {
+                return [];
+            }
+        })();
+        const [projectHits, globalHits] = await Promise.race([
+            Promise.all([projectHitsP, globalHitsP]),
+            tOut,
+        ]);
+        const project = projectHits.filter((h) => h.score >= PROMPT_INJECTION_MIN_SCORE).slice(0, 2);
+        const global = globalHits.filter((h) => h.score >= PROMPT_INJECTION_MIN_SCORE).slice(0, 1);
+        if (project.length === 0 && global.length === 0)
+            return;
+        const lines = ['<stitch-recall>'];
+        if (global.length > 0) {
+            lines.push('User-level rules relevant here:');
+            for (const h of global) {
+                const txt = String(h.content || '').replace(/\n+/g, ' ').slice(0, 250);
+                lines.push(`- [${h.kind}] ${txt}`);
+            }
+        }
+        if (project.length > 0) {
+            lines.push('Project memory relevant to this prompt:');
+            for (const h of project) {
+                const txt = String(h.content || '').replace(/\n+/g, ' ').slice(0, 300);
+                const src = h.source_thread_id ? ' _(thread receipt available — call thread_recall to dig)_' : '';
+                lines.push(`- [${h.kind}] (score ${Number(h.score).toFixed(2)}) ${txt}${src}`);
+            }
+        }
+        lines.push('</stitch-recall>');
+        const payload = { hookSpecificOutput: { hookEventName: 'UserPromptSubmit', additionalContext: lines.join('\n') } };
+        process.stdout.write(JSON.stringify(payload));
     }
+    catch { /* silent — never break a prompt */ }
 }
 /**
  * Derive a thread name for the project at `cwd`. Strategy:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stitchdb/cli",
-  "version": "0.10.0",
+  "version": "0.11.0",
   "description": "Stitch CLI — manage memory + run agents from your terminal",
   "type": "module",
   "bin": {