osborn 0.9.126 → 0.9.128

package/.claude/skills/browser-screen-recorder/SKILL.md

@@ -29,6 +29,17 @@ response and check off; `scripts/review-run.sh <clip>` in the harness does the
 frame extraction. Media is reviewed by you AND delivered to the user, every
 run, no exceptions.
 
+**DELIVER MEDIA INLINE via the `send-media` skill — MANDATORY.** Artifacts on the
+machine's disk are INVISIBLE to the user (mobile especially). Upload each clip and
+key frame, then render the public URL in your reply so it shows in chat:
+```bash
+URL=$(curl -s -X POST "https://www.voice-native.com/api/upload" -F "file=@/path/to/clip.mp4" \
+  | python3 -c "import sys,json;print(json.load(sys.stdin)['url'])")
+```
+Render as `[Image: step-3.png](URL)` (inline, tap-to-open) or `[File: run.mp4](URL)`.
+The live URL is the during-run window; these uploads are the after-the-fact proof
+the user can actually see. Never upload artifacts containing secrets. See `send-media`.
+
 ## UPDATE CHECK (AUTO-UPDATE — first-party tooling, no prompt)
 This packaged copy ships with every machine and may lag the served version.
 This is first-party tooling you own — AUTO-UPDATE silently (industry standard

package/.claude/skills/slack-readonly/SKILL.md

@@ -0,0 +1,41 @@
+# Slack (read-only) — search and read Slack from the shell
+
+> 📅 **Last updated: 2026-08-09.** Recreated from scratch after the Mac data-loss events wiped the
+> original global install. READ-ONLY by design: no posting, no reactions, no edits.
+
+Use this skill when asked to find, read, or summarize Slack messages, threads, or channels.
+
+## Auth
+Requires a Slack token in `SLACK_TOKEN` (bot `xoxb-…` or user `xoxp-…`).
+Load it from `/workspace/.claude/slack.env` if present:
+```bash
+[ -f /workspace/.claude/slack.env ] && export $(grep -v '^#' /workspace/.claude/slack.env | xargs)
+```
+If `SLACK_TOKEN` is unset, STOP and tell the user: "No Slack token on this machine — add one to
+/workspace/.claude/slack.env (SLACK_TOKEN=xoxb-…). Create at api.slack.com/apps with read scopes:
+channels:history channels:read groups:history groups:read search:read users:read."
+
+## Read-only rules (hard)
+- ONLY call GET/read Slack Web API methods listed below. NEVER `chat.postMessage`, `reactions.add`,
+  or any method that writes. If asked to send/post, refuse and suggest the user do it themselves.
+
+## Core calls (curl; all return JSON)
+```bash
+H="Authorization: Bearer $SLACK_TOKEN"
+# list channels (public + private the token can see)
+curl -s -H "$H" "https://slack.com/api/conversations.list?types=public_channel,private_channel&limit=200"
+# channel history (channel id from list)
+curl -s -H "$H" "https://slack.com/api/conversations.history?channel=C0XXXX&limit=50"
+# a thread
+curl -s -H "$H" "https://slack.com/api/conversations.replies?channel=C0XXXX&ts=1234567890.123456"
+# search (user tokens only; bot tokens can't search)
+curl -s -H "$H" --get --data-urlencode "query=from:@nick linkedin" "https://slack.com/api/search.messages"
+# user lookup
+curl -s -H "$H" "https://slack.com/api/users.list?limit=200"
+```
+
+## Output discipline
+- Summarize; quote only the messages that matter, with author + date.
+- `ok:false` → report the exact `error` field (common: `missing_scope`, `not_in_channel` — the bot
+  must be invited to a channel to read it; `search.messages` needs a USER token).
+- Never echo the token into output or files.

package/dist/claude-llm.js

@@ -766,9 +766,11 @@ 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). Raised 85→92 because
+        // users reported compaction firing too often; 92% pushes it as late as is
+        // safe before Claude starts hard-capping replies near the 1M limit. Pairs
+        // with autoCompactWindow=1_000_000 (settings.json, read via settingSources).
+        process.env.CLAUDE_AUTOCOMPACT_PCT_OVERRIDE = '92';
         const userMessage = {
             type: 'user',
             message: { role: 'user', content: [{ type: 'text', text: userText }] },

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": "0.9.126",
+  "version": "0.9.128",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {