Dodaios is an open-source agent OS for your desktop: it runs your coding agents, browser and apps by voice or chat, on your own machine. Formerly Personal Jarvis.
Python agent-orchestrationai-agentai-coding-assistantclaude-codecoding-agentcomputer-usedesktop-applocal-firstmcppersonal-jarvisvoice-assistantvoice-coding
docs(agent-chat): two surfaces, three runners, one permission ladder
rubenluetke10-beepc727be005f16
diff --git a/docs/architecture-overview.md b/docs/architecture-overview.mdindex 09db325fbe..fb807012a9 100644--- a/docs/architecture-overview.md+++ b/docs/architecture-overview.md@@ -120,6 +120,57 @@ Ack-Brain (`jarvis/brain/ack_brain/`) emits a sub-second butler-style preamble b**Persona / custom system prompt:** the live persona comes from `jarvis/brain/persona_loader.py`. `base_persona_prompt()` returns an editable override (`data/custom_system_prompt.md`, written atomically via the Settings UI / `settings_routes.py`) when present, else the packaged `JARVIS_PERSONA.md`. `load_effective_persona_prompt()` is what the brains call: the base **plus the active mode's character block**. Edits apply on the **next turn** (no restart); `invalidate_cache()` clears the in-process cache. Never hardcode the persona string elsewhere.+### The typed chat: two surfaces, three runners (2026-08-25)++`jarvis/agent_chat/` serves two chats that look the same and mean different+things — every session carries a `surface` (`store.SURFACES`, mirrored in+Pydantic and TypeScript, parity-tested):++| Surface | What a typed turn is | Runner per seat | Permission ladder |+|---|---|---|---|+| `"jarvis"` — the front page's Chat section | **Jarvis with a keyboard**: the same assistant the microphone talks to (soul, persona, profile, identity card, memory, wiki, skills, ALL tools through the risk tiers) on the model the composer picked for THIS chat | API-key / local rows → `runner_brain` (`BrainManager.generate` + a `TurnOverride`); subscription CLIs → `runner_cli` with the Jarvis identity | ONE ladder in Claude Code's words — `ask` / `accept-edits` / `bypass` + `plan` — folded onto each CLI's own spelling |+| `"agent"` — the Agentic IDE's chat mode | a plain coding-agent session in a folder | `runner_api` (tool loop) or a plain `runner_cli` | the vendor's own ladder |++**The per-turn override contract (binding).** A chat's pick of provider /+model / effort NEVER moves the live brain: `BrainManager.generate(turn_override=…)`+carries it in a ContextVar (`jarvis/brain/turn_override.py`), read in exactly+three places — the chain (exactly the pick, no cross-provider stand-in; the+intelligent-router lead stays for a tool-incapable pick), the tool surface+(`tools_extra` merged after every gate, `tool_filter` last) and the dispatcher+(`reasoning_effort`, `tool_context`, `max_turns`) — and it writes nothing on the+manager: not `_active_name`, not the provider config, not the dead-lists. The+pick runs on its own cached instance (`_get_brain(scope=)`), so its credential+(the Agents-tab key, a task-local `override_provider_secrets`) is never the+voice's. `switch(persist=False)` / `apply_provider_model` per chat turn is the+mistake that got the first version reverted (ce3fb9673).++**Hands and approvals.** The folder chip is Jarvis' working folder for the+chat: `folder_tools.py` wraps Read/Write/Edit/Ls/Glob/Grep/RunCommand as real+`Tool`s (reads `safe`, writes and the shell `ask`) so they run through+`ToolExecutor` like every other Jarvis tool — one safety path (AP-3), the+blacklist matching `"RunCommand <cmd>"`. A gate the executor raises is answered+by the chat's clickable card, not the voice's two-turn "yes": the turn's tool+context declares `approval_surface: interactive` + an `approval_ref`+(`"agent-chat:<session_id>"`) + `approval_timeout_s`, the executor echoes the+ref on `ActionApprovalRequired`, and `approval_bridge.ChatApprovalBridge` (one+persistent subscriber, grants keyed by ref, never awaiting the person inside a+handler) turns the stance into a decision: `ask` cards everything consequential,+`accept-edits` waves the folder's Write/Edit through, `bypass` answers every+gate itself; `plan` offers reading hands only. The blacklist is never reached+from here in any stance.++**A CLI seat as Jarvis.** A subscription CLI keeps its own loop (that is what+the subscription pays for; only the vendor's official binary spends it) but+runs with Jarvis' head: `BrainManager.render_surface_prompt` renders the+voice turn's prompt layers read-only, `jarvis_harness.build_identity` adds the+chat's transcript on a fresh conversation and the surface addendum, and each+CLI takes it the way it can (Claude Code: `--append-system-prompt-file`; Codex+and agy: in front of the stdin prompt; Grok Build: a compact cut on argv). The+CLI's MCP config names the session in an `X-Jarvis-Chat-Session` header;+`mcp_server_routes` reads it into `jarvis_tools_server.CHAT_SESSION_REF`, and a+Jarvis tool the CLI reaches over MCP asks on the chat's card instead of failing+fast as unattended.+### Risk-Tier systemFour levels: `safe` / `monitor` / `ask` / `block`. Priority is **blacklist > whitelist > tool default** (`jarvis/safety/risk_tier.py`). Whitelist downgrades a tier to `safe` with `approved_by="whitelist"` — this is the anti-confirmation-fatigue contract. **Direct calls to `Tool.execute()` are a bug**; only `ToolExecutor.execute()` is authorized.diff --git a/docs/os-parity.md b/docs/os-parity.mdindex 5ec3fb75ca..fe8fc5b774 100644--- a/docs/os-parity.md+++ b/docs/os-parity.md@@ -210,6 +210,7 @@ implementations, not stubs.| Voice / audio (capture, playback, VAD, wake, STT, TTS, realtime) | Clean; headless disables voice honestly; WASAPI logic is inert-by-data off Windows || Core (launcher, config, keyring, restart, autostart, tray, elevation, paths) | Clean; per-OS autostart (Registry / LaunchAgent / XDG `.desktop`), keyring falls back to a 0600 file on headless hosts || Data / agents (wiki, contacts, telephony, sessions, missions, skills, self-mod, channels, MCP) | Clean; mission workers run on POSIX with a real process-group reaper |+| Typed chat on the Jarvis surface (brain runner, folder tools, approval card, CLI seats as Jarvis) | Clean; pure asyncio + SQLite, no OS API. Every CLI spawn keeps `NO_WINDOW_CREATIONFLAGS` and UTF-8 stdio. The identity for a Claude Code seat travels as a FILE under the app data dir (`jarvis_harness.write_identity_file`, removed after the turn) because Windows caps a command line at 32 767 characters; Codex and agy take it on stdin (no limit), Grok Build a compact cut on argv (`COMPACT_MAX_CHARS`). The MCP session header and the approval bridge are transport-level and OS-neutral |## Open parity gaps