Architecture

A walk through how eVi's pieces fit together. Read this once when you join the project; refer back to the per-module docs when you need detail.

One core, three frontends

          CLI (evi/apps/cli/main.py)
                    │
        Web (evi/apps/web/server.py)─── browser (evi/apps/web/static/)
                    │
       Desktop (desktop/) ──────────── Tauri webview → Web UI
                    │
                    ▼
        ┌─────────────────────┐
        │       evi.Agent     │   evi/llm/agent.py
        │  ┌───────────────┐  │
        │  │ system prompt │  │   composed from base + memory + skills + project
        │  ├───────────────┤  │
        │  │  chat history │  │
        │  ├───────────────┤  │
        │  │ tool dispatch │  │   permission → hooks → tool.call → hooks
        │  └───────────────┘  │
        └────┬───────┬────────┘
             │       │
             ▼       ▼
     ┌──────────┐ ┌──────────────────────────────┐
     │ Backend  │ │  Tool registry               │
     │ (OpenAI- │ │  evi.tools.base.REGISTRY     │
     │ compat   │ │  decorator builds JSON-schema │
     │ chat)    │ │  from type hints              │
     └──────────┘ └──────────────────────────────┘

The three frontends create their own Agent instance; nothing is shared in-process across frontends. The web UI keeps one Agent per session_id in a dict; the CLI has one per evi chat invocation; Tauri runs the same web server underneath, so it gets web-mode sessions.

The agent loop

Agent.chat(user_msg) is a generator that yields Events (TextDelta, ToolCall, ToolResult, Done, Error). Each turn:

1. Optionally prepend a [ongoing goal: …] reminder if agent.goal is set.
2. Optionally append a plan-only suffix and pass tools=None if plan_mode_once is set.
3. Open an OpenAI streaming completion against the configured backend.
4. As deltas arrive: text → TextDelta, tool-call deltas accumulate.
5. If the model finishes with tool calls, dispatch each through _invoke_tool: permission check → before-hooks (may veto) → tool call → after-hooks → result back to the LLM. Loop.
6. Else yield Done.

Per-turn cap: max_turns=6 for the main agent, max_turns=8 for the dream agent. Prevents tool-call ping-pong loops.

System prompt composition

Agent._compose_system_prompt() stitches together:

1. The base prompt ("You are eVi, …").
2. ## Memory index — one-line summaries of every entry in ~/.evi/memory/. Pulled live so dreaming + manual edits show up.
3. ## Available skills — name + description of every ~/.evi/skills/<name>/SKILL.md. The body is loaded on demand via the invoke_skill tool.
4. ## Project context (<path>) — full body of the nearest EVI.md walking up from cwd.

Refreshed on Agent.__init__ and Agent.reset(). Not refreshed mid-turn — if you edit memory while a session is running, the new state appears on the next turn.

Tool framework

@tool(description="Read a UTF-8 text file.", category="fs")
def read_file(path: str) -> str:
    return Path(path).read_text(encoding="utf-8")

The decorator inspects the function's type hints and docstring, builds an OpenAI-shaped JSON schema, and registers a Tool dataclass in REGISTRY. The agent passes [t.openai_schema() for t in tools] to the LLM each turn.

Tool categories govern:

• Whether the tool is enabled (tools.<category> in config.toml)
• Whether tool calls auto-approve (auto.auto_approve list)

New tool? Decorate a function with @tool(category="…"), import the module from evi/apps/cli/main.py and evi/apps/web/server.py for the side effect, add a config toggle if you want one.

Permission flow

Agent.permission_callback: (name, args_json, category) → bool.

If the tool's category is in auto.auto_approve (or auto_all is set for the session), the callback is skipped. Otherwise it's invoked synchronously inside _invoke_tool. Returning False surfaces "PERMISSION DENIED" as the tool result; the model adapts on the next turn.

CLI: callback is _cli_permission_prompt which uses console.input.

Web: callback runs on a worker thread (the agent loop is sync). It generates a decision_id, pushes a PermissionRequest SSE event with it, and blocks on a threading.Event. The browser POSTs to /api/decide; the endpoint flips the event; the worker unblocks.

Scheduler / scripted runs: no callback set → default-allow.

Backends

evi.backends.Backend is the ABC. Four implementations:

get_backend(settings) dispatches by llm.backend string. The chat client is OpenAI-SDK across all four; model-management methods on non-Ollama backends raise NotImplementedError and the CLI tells the user to use hf:<repo> direct downloads instead.

Storage layout (~/.evi/)

config.toml         primary config; profiles overlay this
profiles/*.toml     per-machine overlays selected by EVI_PROFILE / --profile
memory/             markdown notes, .attic/ holds soft-deleted
skills/<name>/      SKILL.md + skill-local assets
commands/<name>.md  user-defined slash command templates with {args}
scheduled/<id>.json one file per saved scheduled prompt
hooks.toml          before/after_tool_call hook entries
mcp.json            MCP server launch configs
images/             ComfyUI output cache
models/             huggingface_hub downloads
screenshots/        computer-use screenshots
transcripts/<day>/<session>.jsonl   session logs feeding the dream engine
logs/dreams/        per-dream audit logs
logs/scheduled/     per-task run logs

Subagents

evi.llm.subagent.run_subagent(...) builds a scoped Agent with a focused system prompt and a restricted tool category set, runs it to completion, returns concatenated text plus a brief tool trace. Used by:

• evi/tools/subagent.py — delegate_explore (fs only) and delegate_plan (no tools) tools the main agent can call.
• evi/dream.py — the dream agent itself is a subagent with category ("memory", "fs") and the dream system prompt.

Lifecycles

MCP — CLI lazy-starts on first _build_agent, atexit cleans up. Web starts it inside the FastAPI lifespan context.

Scheduler — CLI: evi scheduler runs as a foreground daemon. Web: also starts in the lifespan so evi web covers chat + scheduled jobs in one process.

Tauri desktop — Two modes:

• Local (default): spawns py -3.13 -m uvicorn evi.apps.web.server:app as a child, polls /api/health, opens webview at the local port.
• Remote (EVI_REMOTE_URL set): skips the spawn, just navigates.

Testing

pytest against the in-process Agent with stubbed OpenAI clients + httpx.MockTransport for HTTP boundaries. Web tests use Starlette's TestClient; permission-flow integration is unit-tested instead of end-to-end because TestClient buffers SSE responses.