osborn 0.9.127 → 0.9.129

package/dist/claude-llm.js

@@ -483,7 +483,12 @@ export class ClaudeLLM extends llm.LLM {
         return 'claude.agent-sdk';
     }
     get model() {
-        return this.#opts.model || 'claude-opus-4-8'; // Opus 4.8 orchestrator with named sub-agents
+        // The [1m] suffix opts into Opus 4.8's 1M context window (Claude Code's
+        // context-1m-2025-08-07 beta). WITHOUT it the SDK runs opus at its 200k
+        // base, so auto-compaction fired at ~153k every ~10 min (confirmed in the
+        // live agent log: compact_boundary pre_tokens≈153k). With [1m] the window
+        // is 1M and compaction happens far later. Overridable via opts.model.
+        return this.#opts.model || 'claude-opus-4-8[1m]';
     }
     get sessionId() {
         return this.#sessionId;
@@ -766,9 +771,14 @@ export class ClaudeLLM extends llm.LLM {
     pushMessage(userText, sdkOptions, callbacks) {
         // Auto-compact threshold — fires PreCompact when context fills to this %.
         // Higher = uses more of the window before compacting (more context retained
-        // per turn, but less headroom for the next reply). 85% is the sweet spot
-        // before Claude starts hard-capping replies near the limit.
-        process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE = '85';
+        // per turn, but less headroom for the next reply). 92% pushes it as late as
+        // is safe before Claude starts hard-capping replies near the 1M limit.
+        process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE = '92';
+        // Compaction WINDOW in thousands of tokens (min(actual_context, this)). Now
+        // that the model runs with [1m] (1M context), set this to 1000 (=1,000,000)
+        // so the compaction budget matches the real window instead of being clamped
+        // to opus' 200k base — the actual cause of the ~153k early compaction.
+        process.env.CLAUDE_CODE_AUTO_COMPACT_WINDOW = '1000';
         const userMessage = {
             type: 'user',
             message: { role: 'user', content: [{ type: 'text', text: userText }] },
@@ -1002,7 +1012,7 @@ class ClaudeLLMStream extends llm.LLMStream {
                 permissionMode: this.#opts.permissionMode,
                 allowedTools,
                 // model: this.#opts.model || 'haiku', // haiku for speed with limited tools, sonnet for full research capabilities (including tool use trace in response)
-                model: this.#opts.model || 'claude-opus-4-8', // Opus 4.8 orchestrator with named sub-agents (Haiku tested but ignored delegation rules)
+                model: this.#opts.model || 'claude-opus-4-8[1m]', // Opus 4.8 + [1m] → 1M context (see get model() note); prevents ~153k early compaction
                 enableFileCheckpointing: true,
                 settingSources: ['project', 'user'],
                 extraArgs: { 'replay-user-messages': null },

package/dist/index.js

@@ -53,6 +53,58 @@ import { z } from 'zod';
 //   - Voice LLM with tool calling (ask_agent, respond_permission)
 //   - Routes tasks to Claude agents for execution
 // ============================================================
+// Build an enriched tool-use event for the frontend Logs drawer so it can
+// render Claude-style review cards (Read/Edited/Ran with file names, +/- line
+// counts, and an expandable diff) instead of a bare tool name. Best-effort:
+// any field that can't be derived is simply omitted and the card degrades to
+// the tool name. `input` is the raw tool arguments already carried on the
+// tool_use/tool_result events emitted by claude-llm.
+function buildToolLogEvent(name, input, status, agentRole) {
+    const ev = { type: 'tool_use', tool: name, status, agentRole };
+    const inp = input || {};
+    const basename = (p) => String(p).split('/').pop() || String(p);
+    const fp = inp.file_path || inp.path || inp.notebook_path;
+    if (fp) {
+        ev.filePath = fp;
+        ev.fileName = basename(fp);
+    }
+    if (typeof inp.command === 'string')
+        ev.command = inp.command;
+    if (typeof inp.pattern === 'string')
+        ev.pattern = inp.pattern;
+    if (typeof inp.url === 'string')
+        ev.url = inp.url;
+    if (typeof inp.description === 'string')
+        ev.description = inp.description;
+    try {
+        if (name === 'Edit' && typeof inp.old_string === 'string' && typeof inp.new_string === 'string') {
+            ev.linesRemoved = inp.old_string ? inp.old_string.split('\n').length : 0;
+            ev.linesAdded = inp.new_string ? inp.new_string.split('\n').length : 0;
+            ev.diff = createPatch(ev.fileName || 'edit', inp.old_string, inp.new_string, '', '', { context: 2 });
+        }
+        else if (name === 'MultiEdit' && Array.isArray(inp.edits)) {
+            let a = 0, r = 0;
+            const chunks = [];
+            for (const e of inp.edits) {
+                const o = String(e?.old_string ?? ''), n = String(e?.new_string ?? '');
+                r += o ? o.split('\n').length : 0;
+                a += n ? n.split('\n').length : 0;
+                chunks.push(createPatch(ev.fileName || 'edit', o, n, '', '', { context: 2 }));
+            }
+            ev.editCount = inp.edits.length;
+            ev.linesRemoved = r;
+            ev.linesAdded = a;
+            ev.diff = chunks.join('\n');
+        }
+        else if (name === 'Write' && typeof inp.content === 'string') {
+            ev.linesAdded = inp.content.split('\n').length;
+            ev.linesRemoved = 0;
+            ev.diff = inp.content.split('\n').slice(0, 60).map((l) => '+' + l).join('\n');
+        }
+    }
+    catch { /* diff is best-effort — omit on any failure */ }
+    return ev;
+}
 // Load skills list with name + description for frontend display
 function loadSkillsList(agentDir) {
     const skillsDir = join(agentDir, '.claude', 'skills');
@@ -2542,7 +2594,7 @@ async function main() {
             // Inline chat bubble — reuses the existing claude_output path that's already working.
             if (event.type === 'compaction_started') {
                 const triggerLabel = event.trigger ? ` (${event.trigger})` : '';
-                const text = `✨ _Learning from this session — saving your preferences and decisions…_${triggerLabel}`;
+                const text = `✨ _Teaching Osborn from this session — saving your preferences and decisions…_${triggerLabel}`;
                 sendToFrontend({
                     type: 'claude_output',
                     text,
@@ -2555,7 +2607,7 @@ async function main() {
                 const names = Array.isArray(event.skillNames) && event.skillNames.length > 0
                     ? ` — ${event.skillNames.join(', ')}`
                     : '';
-                const text = `✨ Done learning — ${n} skill${n === 1 ? '' : 's'} updated${names}. I'll carry this forward.`;
+                const text = `✨ Osborn learned — ${n} skill${n === 1 ? '' : 's'} updated${names}. I'll carry these forward.`;
                 sendToFrontend({
                     type: 'claude_output',
                     text,
@@ -2630,11 +2682,11 @@ async function main() {
         // Wire up events from the Claude SDK wrapper to frontend
         directLLM.events.on('tool_use', (data) => {
             console.log(`🔧 Claude: ${data.name}`);
-            sendToFrontend({ type: 'tool_use', tool: data.name, agentRole: 'direct' });
+            sendToFrontend(buildToolLogEvent(data.name, data.input, 'running', 'direct'));
         });
         directLLM.events.on('tool_result', (data) => {
             console.log(`✅ Done: ${data.name}`);
-            sendToFrontend({ type: 'tool_use', tool: data.name, status: 'completed', agentRole: 'direct' });
+            sendToFrontend(buildToolLogEvent(data.name, data.input, 'completed', 'direct'));
             // Detect research artifact writes (session workspace or legacy research dir)
             if ((data.name === 'Write' || data.name === 'Edit') && data.input?.file_path) {
                 const fp = data.input.file_path;
@@ -3069,11 +3121,11 @@ async function main() {
         // Wire up Claude events to frontend
         realtimeClaudeHandler.events.on('tool_use', (data) => {
             console.log(`🔧 Claude: ${data.name}`);
-            sendToFrontend({ type: 'tool_use', tool: data.name, agentRole: 'realtime' });
+            sendToFrontend(buildToolLogEvent(data.name, data.input, 'running', 'realtime'));
         });
         realtimeClaudeHandler.events.on('tool_result', (data) => {
             console.log(`✅ Done: ${data.name}`);
-            sendToFrontend({ type: 'tool_use', tool: data.name, status: 'completed', agentRole: 'realtime' });
+            sendToFrontend(buildToolLogEvent(data.name, data.input, 'completed', 'realtime'));
             // Detect research artifact writes (session workspace or legacy research dir)
             if ((data.name === 'Write' || data.name === 'Edit') && data.input?.file_path) {
                 const fp = data.input.file_path;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.127",
+  "version": "0.9.129",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {