feat(v3): daemon + WS protocol + client SDK (Phases 0-3)

CHANGELOG.md

@@ -6,6 +6,43 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/), and this
 
 ## [Unreleased]
 
+## [0.4.0] - 2026-04-18
+
+### Added
+
+- **`arc chat` CLI** — interactive REPL with streaming responses, tool use, and permission modes (`read-only` / `supervised` / `autonomous`). REPL slash commands: `/exit`, `/save`, `/new`, `/mode`, `/clear`, `/sessions`, `/resume`, `/help`. One-shot mode via `--once "<prompt>"`.
+- **`ChatSession` per-profile persistence** — sessions stored at `~/.arc/profiles/<name>/chat-sessions/<id>.json` with atomic writes; `--session <id>` / `/resume <id>` to pick up where you left off.
+- **Roundtable orchestrator** — `RoundtableOrchestrator` class drives the existing roundtable hook over multiple profiles with adaptive delivery pacing (EMA latency) and a designated synthesizer returning a consensus score. Ported from Agent-Forge.
+- **Staged workflow state machine** — `StagedWorkflowManager` implements PLAN → EXEC → VERIFY with per-phase completion patterns and timeouts.
+- **Agent stall watchdog** — `AgentWatchdog` nudges agents at 3 min, marks them stalled at 5 min, and runs a decision protocol (ported from Agent-Forge).
+- **Agent client abstraction** — `packages/core/src/agent-client/` one-shot CLI invocation for Claude / Codex / Gemini with MCP config injection per `mcpMode` variant and per-tool stream parsers.
+- **Tool registry + agent loop** — `packages/core/src/agent/` with three permission modes and ~16 ARC tools (11 read, 4 write, 1 dangerous) wired to existing handlers.
+- **Knowledge endowment** — `packages/core/src/knowledge/` with static ARC catalog (architecture + 52-entry command reference + 16-term glossary), 33-entry feature index, and `buildSystemPrompt()` runtime composer under 4K tokens (~1.3K typical).
+- **Launch modes** — `launchMode: "native" | "worker"` field on Profile (default `native`). `--native` / `--worker` CLI overrides. TUI `m` key in ProfilesView toggles. Doctor check for deprecated `CLAUDE_CODE_NO_FLICKER` env var.
+- **Bare launch** — `arc run <tool>` and `arc launch --bare <tool>` skip the ARC overlay entirely (no env injection, no hook pipeline). Tool-name inference falls through to bare when no matching profile exists.
+- **Clearable active profile** — `arc profile switch none` and `arc profile clear-active` set `activeProfile` to `null`. Rendered as `(none)` in CLI and TUI.
+- **Agent instructions** — `instructions` / `instructionsFile` fields on Profile, injected as `ARC_AGENT_INSTRUCTIONS` env var at launch. `arc instructions` CLI: `show` / `set` / `edit` / `clear`.
+- **OpenAI-compatible providers** — `openai-compat` auth type + `ProviderConfig` (`baseUrl`, `model`, `apiKeyEnvVar`, `displayName`) on Profile; 7 presets (OpenRouter, Ollama, LM Studio, Together, Groq, MiniMax, DeepSeek). `arc provider` CLI: `set` / `show` / `clear` / `presets`.
+- **Backup / export / import** — `arc backup create/restore/list` for a gzipped `~/.arc/` archive (credentials excluded by default); `arc profile export` / `arc profile import-file` for single-profile transport with inlined instructions.
+- **Profile cloning** — `cloneProfile()` core function + `arc profile clone <src> <dst> [--no-copy-dir]` CLI + `Shift+C` inline clone in ProfilesView.
+- **Launch history** — `~/.arc/history.json` records each launch (profile, tool, timestamp, outcome, exitCode); DashView shows recent launches + activity log entries.
+- **Toast notifications** — `ToastProvider` + `useToast()` with auto-dismiss (2.5 s); mounted globally in the Dashboard.
+- **Interactive sidebar queue** — Enter on a profile row in the Sidebar quick-launches without switching views.
+
+### Documentation
+
+- `user-docs/guide/chat.md` — Chat Guide (quickstart, permission modes, REPL commands, session persistence, known limitations).
+- `user-docs/guide/roundtable.md` — Running Roundtables (concepts, programmatic API, adaptive pacing, worker-mode requirement).
+- `user-docs/guide/multi-agent-pipelines.md` — PLAN → EXEC → VERIFY state machine with completion patterns and timeouts.
+- `user-docs/architecture/index.md` — extended "Agent Client + Chat + Orchestration" section.
+
+### Coming in 0.4.x / 0.5.x
+
+- `arc roundtable` CLI with streaming transcript and per-agent color coding (Phase 6).
+- MCP tools: `arc_chat`, `arc_roundtable`, and the 6-tool `team_*` contract (Phase 6).
+- Dashboard chat view with per-session WebSocket streaming and tool-call visualization (Phase 7).
+- Dashboard roundtable + pipelines views (Phase 8).
+
 ## [0.2.0] - 2026-04-03
 
 All 25 phases of the [v2.0 spec](./docs/spec/SPEC.md) are now implemented. ARC has evolved from a profile manager into a unified agent runtime control plane, absorbing the [Axiom-Supervisor](https://github.com/Codename-11/axiom-supervisor) project.
@@ -186,6 +223,7 @@ All 25 phases of the [v2.0 spec](./docs/spec/SPEC.md) are now implemented. ARC h
 - **Light mode contrast** — WCAG AA compliant dimmed/border colors, explicit `colors.text` on import hint
 - **React hooks violation** — `useScreenSize()` moved above conditional returns in DashView
 
-[Unreleased]: https://github.com/Codename-11/ARC/compare/v0.2.0...HEAD
+[Unreleased]: https://github.com/Codename-11/ARC/compare/v0.4.0...HEAD
+[0.4.0]: https://github.com/Codename-11/ARC/compare/v0.2.0...v0.4.0
 [0.2.0]: https://github.com/Codename-11/ARC/compare/v0.1.0...v0.2.0
 [0.1.0]: https://github.com/Codename-11/ARC/releases/tag/v0.1.0

CLAUDE.md

@@ -29,8 +29,17 @@ ARC (Agent Runtime Control) is a CLI + TUI for managing multiple agent profiles
 - **Landing site:** `site/` — React 19 + Vite + Tailwind v4, Nothing-design marketing page
 - **Deployment:** Root `Dockerfile` + `nginx.conf` — multi-stage build merging `site/` at `/` and `user-docs/` at `/docs/` into single nginx container
 - **Web Dashboard:** 13 view components (Overview, Sessions, Traces, Risk, Tasks, Skills, Memory, Agents, Factory + Profiles, Diagnostics, Sync, Plugins)
-- **Orchestration layer:** Hook pipeline (8 hooks in priority order), roundtable multi-agent discussions, task delegation protocol, interagent routing, source classification
-- **Adapters:** Claude Code (SDK + plugin + hooks), Codex CLI, Gemini CLI, OpenClaw (native plugin), Hermes Agent (MCP bridge), Generic (fallback for any tool)
+- **Orchestration layer:** Hook pipeline (8 hooks in priority order), roundtable multi-agent discussions, task delegation protocol, interagent routing, source classification, `arc chat` REPL, `RoundtableOrchestrator`, `StagedWorkflowManager` (PLAN/EXEC/VERIFY), `AgentWatchdog` (stall detection)
+- **Adapters:** Claude Code (SDK + plugin + hooks), Codex CLI, Gemini CLI, OpenClaw (native plugin), Hermes Agent (MCP bridge), OpenAI Compatible (custom providers), Generic (fallback for any tool)
+- **Agent instructions:** `instructions` / `instructionsFile` fields on Profile; resolved at launch, injected as `ARC_AGENT_INSTRUCTIONS` env var; `arc instructions` CLI for show/set/edit/clear
+- **Custom providers:** `openai-compat` auth type + `ProviderConfig` (baseUrl, model, apiKeyEnvVar) on Profile; 7 presets (OpenRouter, Ollama, LM Studio, Together, Groq, MiniMax, DeepSeek); `arc provider` CLI for set/show/clear/presets
+- **Launch modes:** `launchMode: "native" | "worker"` on Profile (default `native`). Native uses full TTY handoff so the tool paints its own TUI; worker uses `spawnManagedProcess` for ARC-supervised orchestration. CLI flags `--native` / `--worker` override. TUI: `m` in ProfilesView toggles. Roundtable forces worker regardless.
+- **Bare launch:** `arc run <tool>` and `arc launch --bare <tool>` skip ARC overlay entirely (no env injection, no hook pipeline). Tool-name inference falls through to bare when no matching profile exists. `activeProfile` may be `null` — cleared via `arc profile switch none` or `arc profile clear-active`, rendered as `(none)`.
+- **Agent client:** `packages/core/src/agent-client/` — CLI-spawn clients for Claude/Codex/Gemini with MCP config injection per `mcpMode` variant and per-tool stream parsers. Substrate for `arc chat` + `RoundtableOrchestrator`. See `docs/plans/ai-and-roundtable.md`.
+- **Agent loop + tool registry:** `packages/core/src/agent/` — tool registry with read-only/supervised/autonomous permission modes, `runAgent` generator for tool-use dispatch, ~16 ARC tools wired to existing handlers (list_profiles, clone_profile, switch_active_profile, query_logs, etc.).
+- **Knowledge:** `packages/core/src/knowledge/` — static ARC catalog (architecture + 52-entry command reference + 16-term glossary) + 33-entry feature index + `buildSystemPrompt()` runtime composer under 4K tokens.
+- **Chat:** `packages/core/src/chat/` — `ChatSession` primitive + per-profile store at `~/.arc/profiles/<name>/chat-sessions/` (atomic writes, resume support). Consumed by `arc chat` CLI (`packages/cli/src/commands/chat.ts`).
+- **Orchestration:** `packages/core/src/orchestration/` — `RoundtableOrchestrator` (driver over the roundtable hook with adaptive pacing + synthesizer), `StagedWorkflowManager` (PLAN → EXEC → VERIFY with completion patterns + per-phase timeouts), `AgentWatchdog` (3-min nudge / 5-min stall), `AgentDeliveryPolicy` + EMA latency tracking (ported from Agent-Forge).
 
 ## Key Conventions
 

FEATURES.md

@@ -2,6 +2,10 @@
 
 Tracking file for planned features, enhancements, and ideas. Checked items are shipped. See `docs/expansion-ideas.md` for broader product direction and `docs/spec/SPEC.md` for the full v2.0 spec.
 
+## v3 — Daemon + many clients (planned)
+
+See [docs/plans/arc-v3-daemon.md](./docs/plans/arc-v3-daemon.md) for the full 14-phase plan. Targeting `1.0.0` (breaking). Daemon on :7272 + binary-mux WS protocol + client SDK; TUI/CLI/dashboard/Electron/mobile all become peer clients; E2E-encrypted relay for remote access; SQLite canonical store; provider `extends`; `arc loop`; chat rooms; enhanced roundtable + handoff; Docker server mode.
+
 ## Priority 1 — Core UX Gaps
 
 - [x] **Profile creation in TUI** — stepped overlay form (name -> tool -> auth type -> done) so users don't have to exit the TUI to create profiles
@@ -15,6 +19,8 @@ Tracking file for planned features, enhancements, and ideas. Checked items are s
 - [x] **Workspace-aware profile auto-selection** — `arc.json` in repo root specifies preferred profile/tool; workspace overrides applied on launch (Phase 9)
 - [x] **Workspace shell syntax highlighting** — tokenized input with color-coded `/commands` (green), `@profiles` (blue), `#tags` (dimmed); invalid tokens show in red
 - [x] **Workspace shell auto-complete** — suggestion overlay for `/` commands and `@profile` mentions; Tab/Enter accepts, arrows navigate, Escape dismisses
+- [x] **Launch modes (native / worker)** — `launchMode` field on Profile, `arc launch --native` / `--worker` CLI flags, `m` key toggle in ProfilesView, doctor check for deprecated `CLAUDE_CODE_NO_FLICKER`
+- [x] **Bare launch / clearable active profile** — `arc run <tool>`, `arc launch --bare <tool>`, tool-name inference when no matching profile exists, `arc profile switch none` / `arc profile clear-active`, `activeProfile: null` renders as `(none)`
 - [ ] **Quick profile switch overlay** — global `Ctrl+S` or palette action that shows a focused profile picker from any view
 - [x] **Doctor repair actions** — inline install hints, re-auth instructions, and PATH/shell fix hints on actionable diagnostics
 - [ ] **Profile search/filter** — `/` search in Profiles view and queue for scaling to 10+ profiles
@@ -27,19 +33,31 @@ Tracking file for planned features, enhancements, and ideas. Checked items are s
 - [x] **Tool-adapter architecture** — `RuntimeAdapter` interface with lifecycle methods; Claude, Codex, Gemini, OpenClaw, and Generic adapters (Phase 2, 5-6)
 - [x] **Profile inheritance** — `inherits` field + `resolveProfile()` engine for base + override resolution (Phase 9)
 - [x] **Project-local config** (`arc.json`) — preferred tool, profile, workspace overrides per repo (Phase 9)
+- [x] **Agent instructions** — `instructions` / `instructionsFile` fields on Profile, resolved at launch, injected as `ARC_AGENT_INSTRUCTIONS` env var; `arc instructions` CLI (show/set/edit/clear)
+- [x] **OpenAI-compatible providers** — `openai-compat` auth type + `ProviderConfig` on Profile (baseUrl, model, apiKeyEnvVar); 7 presets (OpenRouter, Ollama, LM Studio, Together, Groq, MiniMax, DeepSeek); `arc provider` CLI (set/show/clear/presets)
 - [ ] **Team/shared config** — repo-checked config with local secret overlays
-- [ ] **Backup/export/import** — move profiles and settings between machines
+- [x] **Backup/export/import** — `arc backup create/restore/list` (gzipped archive of `~/.arc/`, credentials excluded by default) + `arc profile export` / `arc profile import-file` (single-profile JSON transport with inlined instructions)
 - [x] **Managed updates** — self-update system with npm registry check and TUI update banner
+- [x] **Agent client foundation** — internal CLI-spawn agent client at `packages/core/src/agent-client/` (Claude/Codex/Gemini), MCP config injection per `mcpMode`, stream parsers. Plan Phase 1
+- [x] **Tool registry + agent loop** — `packages/core/src/agent/` with ~16 ARC tools spanning read / write / dangerous tiers; three permission modes (read-only / supervised / autonomous); `runAgent` generator. Plan Phase 2
+- [x] **Knowledge endowment** — `packages/core/src/knowledge/` system prompt composition (ARC architecture + 52-entry command catalog + 33-entry feature index + 16-term glossary + runtime state). Plan Phase 3
+- [x] **`arc chat` CLI** — terminal REPL using active profile's agent client with streaming output, permission-gated tool calls, per-profile session persistence at `~/.arc/profiles/<name>/chat-sessions/`, REPL slash commands. Plan Phase 4 (0.4.0)
+- [x] **Roundtable orchestrator** — `RoundtableOrchestrator` driving the existing roundtable hook with adaptive pacing (EMA latency) and synthesizer-driven consensus score. Plan Phase 5 (0.4.0)
+- [x] **Staged workflow state machine** — `StagedWorkflowManager` PLAN → EXEC → VERIFY with completion patterns and per-phase timeouts (ported from Agent-Forge)
+- [x] **Agent stall watchdog** — nudge at 3 min, mark stalled at 5 min, decision protocol (ported from Agent-Forge)
+- [x] **`arc roundtable` CLI + team MCP tools** — `arc roundtable <topic> --agents a,b,c` with streaming transcript; `arc_chat` / `arc_roundtable` / 6 `team_*` MCP tools. Plan Phase 6 (0.4.0)
+- [x] **Dashboard chat view** — per-session WS streaming, tool-call visualization, permission-mode toggle, confirmation modal. Plan Phase 7 (0.4.0)
+- [x] **Dashboard roundtable + pipelines view** — configure + run multi-agent flows from the browser with live transcript; per-run history persisted to `~/.arc/roundtables/<id>.json` and `~/.arc/pipelines/<id>.json`. Plan Phase 8 (0.4.0)
 
 ## Priority 4 — Observability & Polish
 
-- [ ] **Launch history on Dash** — recent launches list (`{ profile, tool, timestamp }`) in `~/.arc/history.json`, displayed on Dash after first session
+- [x] **Launch history on Dash** — `~/.arc/history.json` records each launch (profile, tool, timestamp, outcome, exitCode); DashView RightColumn shows recent launches + recent activity log entries (polled)
 - [x] **Shared layer visibility** — SettingsView shows per-profile sync details; ProfileList shows shared indicator column
-- [ ] **Toast notifications** — brief auto-dismiss messages for confirmations/errors that work across all views
-- [ ] **Interactive sidebar queue** — Enter on sidebar profile list to quick-launch without switching views
+- [x] **Toast notifications** — `ToastProvider` + `useToast()` hook with auto-dismiss (2.5s); `ToastContainer` mounted in Dashboard
+- [x] **Interactive sidebar queue** — combined nav+profile selection in Sidebar; `↑/↓` cycles through nav items then profiles; Enter on a profile row quick-launches without switching views
 - [x] **MCP server management** — MCP host manager with connect/disconnect/list/getTools + callTool with risk classification (Phase 8)
 - [x] **Policy layer** — three-tier permission model (coordinator/interactive/worker) with deny > ask > allow precedence (Phase 20)
-- [ ] **Profile cloning/duplication** — create a new profile from an existing one as template
+- [x] **Profile cloning/duplication** — `cloneProfile()` core fn + `arc profile clone <src> <dst> [--no-copy-dir]` CLI + `Shift+C` inline clone in ProfilesView
 - [x] **Usage/audit log** — structured JSONL log with `arc logs` CLI, level/component/profile filtering (Phase 3)
 
 ## v2.0 Spec Features (All 25 Phases Complete)
@@ -53,7 +71,8 @@ Tracking file for planned features, enhancements, and ideas. Checked items are s
 - [x] OpenClaw adapter (plugin manifest, RuntimeAdapter, 3 lifecycle hooks)
 - [x] Hermes Agent adapter (MCP bridge, lifecycle, process management)
 - [x] Generic adapter factory (fallback for any unknown tool, health monitoring)
-- [x] 48 adapter registry + generic adapter tests
+- [x] OpenAI Compatible adapter (custom provider endpoints, 7 presets)
+- [x] 50+ adapter registry + generic adapter tests
 
 ### Logging & Lifecycle (Phases 3-4)
 - [x] Structured JSONL log at `~/.arc/logs/structured.jsonl`
@@ -210,8 +229,3 @@ These items from the original v0.1 backlog are still open:
 - [ ] Profile search/filter in Profiles view
 - [ ] Environment preview before launch
 - [ ] Team/shared config (repo-checked config with local secret overlays)
-- [ ] Backup/export/import (move profiles between machines)
-- [ ] Launch history on Dash
-- [ ] Toast notifications
-- [ ] Interactive sidebar queue
-- [ ] Profile cloning/duplication