osborn 0.9.27 → 0.9.29

package/.claude/skills/user-context/SKILL.md

@@ -0,0 +1,108 @@
+---
+name: user-context
+description: Learn and maintain a living context document about this specific user — their vocabulary, communication style, recurring topics, and decision patterns. Adapted from grill-with-docs for voice-native conversations rather than code. Use when the user wants to update their context, or run passively after meaningful sessions to capture what was learned.
+---
+
+<what-to-do>
+
+Interview the user to build and sharpen a shared vocabulary between them and the agent.
+Ask one question at a time and wait for a response before continuing.
+
+The goal is NOT a coding spec. The goal is understanding HOW THIS PERSON communicates,
+what words they use naturally, what they mean when they say things, and what matters to them.
+
+Update `~/.claude/skills/user-context/CONTEXT.md` inline as terms and preferences are resolved.
+Do not batch — capture as they happen.
+
+</what-to-do>
+
+<trigger-phrases>
+- "grill me"
+- "learn my language"
+- "update my context"
+- "learn how I talk"
+- "standardise my language"
+- "update user context"
+</trigger-phrases>
+
+<supporting-info>
+
+## What to capture
+
+### Vocabulary
+When the user uses a term consistently, record it. When they use multiple words
+for the same thing, help them pick one. Example: if they say "the cloud machine",
+"the fly machine", and "the server" interchangeably — resolve it.
+
+### Communication style
+How formal or casual are they? How much technical depth do they expect?
+Do they prefer short answers or thorough ones? Do they interrupt and redirect often?
+
+### Recurring topics
+What subjects come up repeatedly? What projects, people, or concepts
+are central to their world?
+
+### Decision patterns
+How do they make decisions? Do they want options presented or a direct recommendation?
+Do they want to understand the why, or just the what?
+
+## During the session
+
+### Challenge fuzzy language
+When the user uses vague or inconsistent terms, name it: "You've said both 
+'voice session' and 'room' — are those the same thing to you, or different?"
+
+### Reflect back
+Periodically summarise what you've learned: "So when you say 'the agent', 
+you mean the cloud process on the fly machine, not me (Claude) — is that right?"
+
+### Update CONTEXT.md inline
+When a term or preference is clarified, write it to 
+`~/.claude/skills/user-context/CONTEXT.md` immediately using the format below.
+
+## CONTEXT.md format
+
+```md
+# User Context — {user name or handle}
+
+{One or two sentences about who this person is and what they're building.}
+
+## Language
+
+**Voice session**: An active connection between the user's browser and the fly machine via LiveKit. 
+_Avoid_: "room" (unless referring specifically to the LiveKit room name)
+
+**The agent**: The osborn Node.js process running on the fly machine (not Claude).
+_Avoid_: "Claude", "AI", "bot" — those refer to the LLM, not the agent process.
+
+**The dashboard**: voice-native.com — the web app frontend.
+_Avoid_: "the front end", "the UI" (too generic)
+
+## Communication style
+
+- Prefers direct recommendations over options when the answer is clear
+- Talks fast and interrupts — short answers are better unless depth is asked for
+- Uses voice transcription — expects the agent to handle garbled/incomplete input gracefully
+- Casual tone, first-name basis, profanity is fine
+
+## Recurring topics
+
+- Osborn: voice AI coding assistant built on LiveKit + Claude Agent SDK
+- Fly machines: the cloud sandbox infrastructure replacing Sprites
+- Sessions: Claude Code JSONL files stored in ~/.claude/projects/
+
+## Decision patterns
+
+- Values speed over perfection for first implementations
+- Prefers understanding the root cause before applying a fix
+- Wants to be told when something is irreversible before proceeding
+```
+
+## Rules
+
+- Keep definitions tight — one or two sentences.
+- Only record terms specific to this user's world. Generic concepts don't belong.
+- Update the file live during the session, not at the end.
+- Re-read the existing CONTEXT.md at the start of every session that touches it.
+
+</supporting-info>

package/dist/claude-llm.js

@@ -1095,6 +1095,35 @@ class ClaudeLLMStream extends llm.LLMStream {
                                         catch (blErr) {
                                             console.error('⚠️ PostCompact: BEHAVIORAL_LEARNINGS write failed:', blErr instanceof Error ? blErr.message : blErr);
                                         }
+                                        // ── Section 5: USER_CONTEXT — merge into user-context/CONTEXT.md ──
+                                        // Passive UX: agent learns the user's language automatically after every
+                                        // compaction. No explicit "grill me" command needed. The compact summary
+                                        // captures vocabulary/style observations confirmed during the session.
+                                        try {
+                                            const userCtx = extractSection('=== USER_CONTEXT ===');
+                                            const hasContent = userCtx.length >= 30 && !userCtx.includes('(none this session)');
+                                            if (hasContent) {
+                                                progress('Updating user context', `${userCtx.length} chars`);
+                                                const ctxFolder = join(skillDir, '.claude', 'skills', 'user-context');
+                                                const ctxPath = join(ctxFolder, 'CONTEXT.md');
+                                                mkdirSync(ctxFolder, { recursive: true });
+                                                // Append new observations to existing CONTEXT.md under a dated section.
+                                                // The grill-me skill can consolidate/clean up on explicit request.
+                                                const existing = existsSyncFs(ctxPath) ? readSyncFs(ctxPath, 'utf-8') : '# User Context\n\n';
+                                                const entry = `\n## Observations — ${today} (session ${sessionId.substring(0, 8)})\n\n${userCtx}\n`;
+                                                writeSyncFs(ctxPath, existing + entry, 'utf-8');
+                                                console.log(`🧠 PostCompact: updated user context at ${ctxPath}`);
+                                                skillsWritten++;
+                                                skillNames.push('user-context');
+                                                progress('Wrote skill', 'user-context');
+                                            }
+                                            else {
+                                                console.log('🧠 PostCompact: no USER_CONTEXT observations this session — skipping');
+                                            }
+                                        }
+                                        catch (ucErr) {
+                                            console.error('⚠️ PostCompact: USER_CONTEXT write failed:', ucErr instanceof Error ? ucErr.message : ucErr);
+                                        }
                                         this.#opts.onCompactionEvent?.({ type: 'compaction_complete', skillsWritten, skillNames });
                                         console.log(`🧠 PostCompact: complete — ${skillsWritten} skill file(s) written: [${skillNames.join(', ')}]`);
                                     }

package/dist/prompts/compact-learnings-instruction.md

@@ -56,4 +56,26 @@ If the EXISTING LEARNED SKILLS section is shown below, merge — update outdated
 
 ---
 
-IMPORTANT: All four sections MUST appear at the end of the compact summary even if some are empty. Empty sections should have a single line: (none this session)
+---
+
+## Section 5: USER_CONTEXT
+Format: === USER_CONTEXT ===
+Capture what you learned about THIS USER's language and communication style this session.
+Only include observations that were CONFIRMED — things the user actually said, corrected, or repeated.
+Skip if nothing new was observed.
+
+VOCABULARY:
+- List any terms the user used consistently with a specific meaning
+- List any terms they used interchangeably that should be standardised
+- Format: **term**: what they mean by it. _Avoid_: alternatives they used
+
+STYLE:
+- One or two sentences describing how they communicate (tone, pace, depth preference)
+- Only include if a clear pattern emerged this session
+
+RECURRING:
+- Topics, projects, or concepts that came up repeatedly
+
+---
+
+IMPORTANT: All five sections MUST appear at the end of the compact summary even if some are empty. Empty sections should have a single line: (none this session)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.27",
+  "version": "0.9.29",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {