diff --git a/.planning/PROJECT.md b/.planning/PROJECT.md
new file mode 100644
index 0000000..80c7488
--- /dev/null
+++ b/.planning/PROJECT.md
@@ -0,0 +1,73 @@
+# Flynn — Operator DX Milestone
+
+## What This Is
+
+A focused quality-of-life milestone for Flynn's operator (you). Flynn is a mature multi-channel AI assistant daemon with 10 model providers, 5 channel adapters, 40+ tools, a web dashboard, automation, sandboxing, and 1077 passing tests. This milestone targets the three biggest friction points in developing and operating Flynn: the monolithic daemon wiring file, lack of multi-environment config management, and limited runtime observability.
+
+## Core Value
+
+Make Flynn easier to reason about, configure, and monitor — so that adding features and diagnosing issues takes minutes, not hours.
+
+## Requirements
+
+### Validated
+
+- ✓ Multi-channel AI assistant daemon with tool loop — existing
+- ✓ 10 model providers (Anthropic, OpenAI, Gemini, Bedrock, GitHub, OpenRouter, Zhipu, xAI, Ollama, llama.cpp) — existing
+- ✓ 5 channel adapters (Telegram, Discord, Slack, WhatsApp, WebChat) — existing
+- ✓ WebSocket gateway with JSON-RPC protocol — existing
+- ✓ Web dashboard SPA (dashboard, chat, sessions, settings, usage) — existing
+- ✓ YAML config with Zod validation and env var expansion — existing
+- ✓ SQLite session persistence with TTL pruning — existing
+- ✓ Memory system with hybrid keyword + vector search — existing
+- ✓ Docker sandboxing per session — existing
+- ✓ Multi-agent routing with per-agent config — existing
+- ✓ Automation: cron, webhooks, heartbeat, Gmail watcher — existing
+- ✓ MCP tool server integration — existing
+- ✓ Skills system (bundled/managed/workspace) — existing
+- ✓ Media pipeline (image analysis, audio transcription, outbound attachments) — existing
+- ✓ Context compaction with memory extraction — existing
+- ✓ Tool policy profiles with allow/deny lists — existing
+- ✓ 1077 tests passing — existing
+
+### Active
+
+- [ ] Decompose daemon/index.ts into focused service modules
+- [ ] Multi-environment config system (base + overlays)
+- [ ] Live operational dashboard with real-time metrics
+
+### Out of Scope
+
+- New channel adapters (Signal, Matrix, Teams, Google Chat) — not the focus of this milestone
+- Companion apps (macOS, iOS, Android) — massive scope, different project
+- Structured logging framework — would complement the dashboard but adds complexity; evaluate after dashboard reveals what metrics matter
+- Agent intelligence features (sub-agent spawning, planning) — separate milestone
+- ESLint / type safety cleanup — worthwhile but not blocking current development
+
+## Context
+
+Flynn has been through rapid feature development (P0-P8, Tiers 1-4 all completed in ~7 days). The codebase grew fast and the wiring layer absorbed complexity. Key context:
+
+- **daemon/index.ts** is 1087 lines — it handles model client creation, channel setup, agent factory, memory initialization, vector indexer, session pruning, lifecycle management, and graceful shutdown. Every new feature touches this file.
+- **Config** is a single YAML file validated by a 400+ line Zod schema. Managing dev vs Docker vs production requires manual YAML duplication. No layering, no environment-specific overrides.
+- **Observability** is currently console.log/error/warn. The web dashboard shows basic stats but no real-time metrics: no message trace, no queue depth, no model call latency, no error stream. Debugging requires reading source and tailing stdout.
+- The existing web dashboard (vanilla JS SPA) is functional and can be extended rather than rewritten.
+
+## Constraints
+
+- **Tech stack**: TypeScript, Node.js >= 22, pnpm. No new frameworks (keep vanilla JS for dashboard).
+- **Backwards compatibility**: Existing config files must continue to work. Decomposition must not change public API or behavior.
+- **Test coverage**: Maintain 1077+ passing tests. New modules need tests.
+- **Single-operator**: This is a personal tool. Don't over-engineer for multi-tenant or team scenarios.
+
+## Key Decisions
+
+| Decision | Rationale | Outcome |
+|----------|-----------|---------|
+| Decompose god file, not rewrite | Preserves working code, reduces risk, can be done incrementally | -- Pending |
+| Config overlays over separate files | Environment-specific overrides are less error-prone than maintaining N complete configs | -- Pending |
+| Extend existing web dashboard | Already works, no framework dependencies, familiar codebase | -- Pending |
+| Skip structured logging for now | Dashboard will reveal what metrics actually matter; avoid premature abstraction | -- Pending |
+
+---
+*Last updated: 2026-02-09 after initialization*