Pivot: focus Klonode on contextualization, drop chat UI

[smorchj]

The pivot

Klonode is not a Claude wrapper. It is a contextualizer that rides alongside Claude Code. Claude Code handles the conversation. Klonode watches, learns, and improves the routing. Two tools, one job each.

With recent Claude Code UI/framework improvements there is no reason to rebuild the chat experience. All our energy should go into making the contextualizer as good and solid as it can be.

Shipped

PR #78 — pivot foundation + live node graph

Phase 1: Foundational delete (-4377 lines). Removed the chat/CO/agent surface: ChatPanel, stores/chat, stores/agents, stores/settings, api/chat (+/stream, /classify), api/co, api/agents, core/agents/*.

Phase 2: Live node graph watcher (+875 lines). Server tail watcher on ~/.claude/projects/<path-hash>/*.jsonl, SSE endpoint, client EventSource wrapper, scope settings store, top bar watcher pill + Project/Machine scope toggle. Verified end-to-end: every Read into packages/core/... triggers a pulse ring on the packages node in real time. Also includes five bug fixes uncovered during verification (slash mismatch, stacking setIntervals, break→continue, lifetime tuning, project-scope ancestor walk).

Roadmap — open issues

The learning model is the heart of the contextualizer pivot. It works like human memory: repetition (boring, accumulative signals) builds confidence, emotion (rare, high-weight, time-decayed signals) builds urgency. Repetition is background hum; emotion is what makes the suggestions panel light up.

Foundation

• Persistent observation log — foundation for the contextualizer learning model #79 — Persistent observation log. Everything below needs a history; currently events flow through the 8s pulse lifetime and are lost.

Learning model

• Learning model: repetition signals (confidence per node) #80 — Repetition signals → confidence per folder node. Reads/edits over sessions, routing coverage, cost-per-use.
• Learning model: emotion signals (urgency per node, time-decayed) #81 — Emotion signals → urgency per folder node, time-decayed. User corrections, max_turns hits, long/expensive turns, hedges, explicit praise.

Analyzer trigger

• Session-end detection + post-session analyzer trigger #82 — Session-end detection + post-session analyzer. When a session ends, recompute learning scores and emit suggestions for score deltas above threshold.

User-facing output

• Suggestions panel — the user-facing output of the contextualizer #83 — Suggestions panel. Urgency-sorted list of proposed CONTEXT.md edits. One-click approve, dismiss, snooze. Never auto-applies.
• Visualize confidence and urgency on the graph view #84 — Visualize confidence (node size/opacity) and urgency (border color/thickness) on the graph. Makes the health of the routing legible at a glance.

Nice-to-have

• Past-session replay mode — scrub any session through the live node graph #85 — Past-session replay/scrub mode. Pick a recorded session from a dropdown, scrub through it on a timeline with play/pause/speed.

Dependency order

#78 (shipped) ──┐
                ├── #79 ── #80 ─┐
                │               ├── #82 ── #83
                └── #79 ── #81 ─┤         ── #84
                                │
                #79 ────────────┴── #85

What we deleted (reminder)

• ChatPanel as primary UI
• Compare mode (with/without Klonode)
• CO as a chat session
• /api/chat/classify endpoint
• Plan approval flow
• Multi-agent / spawn_subagent plans (never built, dropped from roadmap)
• sendMessage / approvePlan / sendComparison in stores/chat.ts
• Related API routes that only existed to serve the chat panel

Plus six redundant issues closed that were scoped around the old model: #73, #71, #64, #63, #62, #44.

What stays and gets better

• packages/core analyzer, content-extractor, generator, routing-graph, optimizer
• packages/ui GraphView, TreeView, ContextEditor, GitHubView
• CLI: klonode init, klonode optimize, klonode status