runa/README.md

# Runa — Rune Actors

Runa (short for **Rune Actors**) is an experimental project to build a
persistent, live programming world around isolated actors written in Rune.
The long-term vision combines an inspectable graphical environment with a
programmatic API and a VS Code extension, all connected to the same running
image.

## Status

Runa is at an early prototype stage. The executable runs a Phase 0 feasibility
experiment using Rune 0.14.2: script execution, native calls, async suspension,
instruction budgets, stepping, and compiled-code sharing. It also runs a minimal
single-worker actor demonstration with bounded mailboxes and explicit message
effects. Source and compiled implementation identities are BLAKE3 content
addresses with an in-memory compile cache. Automatic crash recovery,
canonicalization, and remote network access are not implemented yet. The
initial native graphical world client can browse and inspect actors in a
running local image.
A minimal VS Code extension can connect to the local image, list actors, and
inspect state, queued/reserved messages, and execution telemetry. Inspection
JSON is editable for inactive paused/failed actors; saving restores
protected fields with a warning. The image owns mutable, directory-based Rune source files with
independent commit history. The CLI exposes source writes, commit/history/rollback,
reload/become, graceful shutdown, kill, restart, pause/resume, and cancellation through a trusted image
API. A versioned transport-neutral protocol contract is now defined, but no
authentication is implemented. The `runa` server provides an unauthenticated
WebSocket listener restricted to localhost, with bounded request sizes.
The `runa-cli` REPL and VS Code connect with
the `runa-debug` WebSocket subprotocol, so clients share one server-owned image.
Server change notifications automatically refresh VS Code's actor tree and inspectors.

The prototype uses worker-local invocation stacks and cooperative instruction
slices, and actor snapshots report basic active-execution progress. Native
callbacks are not preemptible. The debugger foundation exposes bounded active Rune
stack frames and exact loaded source through read-only protocol inspection and
VS Code's **Inspect Execution Stack...** actor action. Actor-local source
breakpoints and bounded step-into/over/out are implemented through the protocol and VS
Code actor actions. **Debug Actor** attaches a native VS Code debug session with
gutter breakpoints, the call-stack view, and bounded toolbar stepping. Continue
resumes the actor; scheduler ticks remain explicit. The Variables view exposes
a read-only committed actor snapshot with expandable state and mailbox.
Watch and Debug Console accept snapshot paths such as `actor/committed_state`;
they never execute actor code. Frame locals and Rune expression evaluation
remain unimplemented.
Failed actors can also be attached for exception details and retained-snapshot
inspection without restarting them; post-mortem stacks are not retained.
Explicit disk image save/restore is implemented at handler boundaries, including
sources/history, actor identities, committed state, mailboxes and retained logs.

## Design goals

- Keep actor identity, mailbox, private state, behavior, and incarnation
  separate.
- Use message passing rather than shared mutable state for actor interaction.
- Keep behavior implementations immutable and support safe, versioned updates.
- Expose the same runtime capabilities to in-world tools and external clients.
- Be explicit about execution, delivery, and persistence guarantees.

See [the architecture spec](docs/architecture.md) for the full vision,
architecture, roadmap, and open questions.

## Getting started

You need a Rust toolchain that supports the workspace's Rust 2024 edition.

Start the live image from the repository root:

```sh
mise run start
```

In another terminal, connect the REPL:

```sh
mise run repl
```

To open the initial graphical world client against the same running image:

```sh
mise run world
```

The world client connects to `127.0.0.1:3737` by default. Set `RUNA_CONNECT` or
pass `--connect 127.0.0.1:4000` to select another loopback address. It provides
actor browsing and inspection, message sending, lifecycle controls, explicit
scheduler ticks, bounded debugger stepping, source breakpoints,
execution-stack inspection, and retained logs. A resizable source file tree on
the left opens behavior files and selects the source used by Spawn. A resizable
actor tree on the right expands actors into metadata/inspector, state, mailbox,
log, and `behavior.rn` entries, opening the corresponding view in the center.
Its source view supports
editing image sources with optimistic conflict checks, compilation diagnostics,
explicit history commits, and revision preview/rollback. Rollback requires
confirmation and creates a new current source version without deleting history.
The actor inspector can also load and edit an actor-local temporary behavior,
save it with source-identity checks, reload the behavior from the image source,
or explicitly confirm updating the shared source from the temporary version.
The events view polls a bounded, non-destructive event history using its own
cursor, so reading events does not consume them for other clients. The server
retains up to 4096 events in memory; this history is process-local and is not
included in image snapshots. Rune frame locals and expression evaluation are
not available. The client does not create or embed a separate image.

The REPL defaults to `127.0.0.1:3737`. To select another loopback endpoint,
use `cargo run -p runa-cli -- --connect 127.0.0.1:4000 repl` and start the server with
the corresponding `--listen` address. Quitting the REPL disconnects only that
client; actors and source files survive until the server exits. Ctrl+C and
Ctrl+D also exit the REPL and complete the WebSocket close handshake.
Reconnect and use `ps` to recover their references. REPL bindings and
command history are session-local and are not restored on reconnect.

### Save and restore an image

```sh
cargo run -p runa-cli -- save /path/to/image.runa.json
cargo run -p runa-cli -- save --overwrite /path/to/image.runa.json
cargo run -p runa-cli -- restore /path/to/image.runa.json
```

To open the snapshot when starting the server, use:

```sh
cargo run -p runa -- --restore /path/to/image.runa.json
# Alternatively:
RUNA_RESTORE=/path/to/image.runa.json mise run start
```

`--restore` overrides `RUNA_RESTORE`. Startup restoration validates the snapshot
before accepting clients and does not add or overwrite bundled sources.
Missing, malformed or incompatible snapshots fail startup rather than silently
starting a fresh image. Without either option, startup loads the bundled counter
source as before. Restored actors follow the same status rules as live restore;
the server does not execute messages until scheduler ticks are requested.

The REPL also accepts `save <file>`, `save --overwrite <file>`, and
`restore <file>`. VS Code provides
**Runa: Save Image...** and **Runa: Restore Image...**. Paths refer to the
server's filesystem, not the client's. Saving refuses an existing destination
unless overwrite is explicitly enabled; VS Code asks before replacing one.
Replacement atomically swaps in a fully written and synced snapshot.
The server captures all actors at a handler boundary and leaves runnable actors
paused after a successful save. Saving is rejected without changing the image
if any handler is unfinished, including paused or awaiting execution.

Restore replaces the whole image after validating and recompiling the snapshot.
Actors that were runnable when saved resume scheduling eligibility; previously
paused and failed actors stay that way. Execution still requires explicit ticks.
Actor references survive, but incarnations advance and old debug sessions end.
Restore also refuses to replace an image with unfinished handlers.
Only saved data survives server exit; this is not an automatic checkpoint,
durable per-message transaction log, or execution-stack serialization.
See [runtime persistence guarantees](docs/runtime.md#image-save-and-restore).

### Commands and connections

Both executables use Clap for `--help`, `--version`, and argument validation.
`runa` listens on `127.0.0.1:3737` without arguments. Set `RUNA_LISTEN` to change
the server address and `RUNA_CONNECT` to change the REPL address; explicit
`--listen` and `--connect` flags override their respective environment variables.
For example:

```sh
RUNA_LISTEN=127.0.0.1:4000 mise run start
RUNA_CONNECT=127.0.0.1:4000 mise run repl
```

`runa-cli` requires a subcommand. Use `runa-cli repl` for an interactive session,
or run an image operation directly:

```sh
cargo run -p runa-cli -- ps
actor=$(cargo run -p runa-cli -- spawn actors/counter)
cargo run -p runa-cli -- send "$actor" '{"op":"inc"}'
cargo run -p runa-cli -- tick 100
```

Each image operation in the REPL has a matching subcommand; use
`runa-cli <command> --help` for arguments. Quote JSON as a single shell argument.
One-shot commands use actor references rather than session-local bindings,
print results without a REPL banner, and exit unsuccessfully on errors.
`let name = spawn ...` is session-local REPL syntax. `quit` (alias `exit`)
exits either interface without stopping the image. `cancel` is no longer a
REPL command; the underlying runtime cancellation API remains available.

The server is WebSocket-only; stdio transport has been removed. By default it listens on
`ws://127.0.0.1:3737/` and requires the `runa-debug` subprotocol, with up to 16
simultaneous clients sharing one live image. The listener rejects non-loopback
addresses and has no authentication; any local process that can connect has
full access to the development API. Use it only on a trusted local machine.

The `runa-client` crate connects over loopback
WebSockets and exchange typed protocol requests with request/response correlation.
Network connection establishment is limited to 5 seconds and response waits/writes to 30
seconds. Transport failures disconnect the client and are never automatically
retried: an operation may have committed even if its response was lost.
The VS Code extension connects over the network to this same live image; actors remain in
the image regardless of how clients connect. Remote network access requires
authentication and authorization, which are not implemented.

The server loads the bundled `source/actors/counter.rn` working file at startup.
Connecting a client does not modify source or create a commit:

```text
runa> let counter = spawn actors/counter
counter = @<actor-reference>
runa> send counter {"op": "inc"}
message queued
runa> tick 100
runa> pause counter
runa> inspect counter
runa> events
```

Actor references are generated NanoIDs rather than sequential numbers. The
`let` form binds each spawned actor to a local name; references can also be
entered directly with an `@` prefix. The counter treats missing/null state as
zero on its first message, increments and logs its value, then queues another
`inc` message to itself. It runs only when ticks advance; pause/resume controls
its loop. An explicit integer initial state is preserved. The REPL provides Tab
completion and Up/Down command history. `write actors/counter <local-file>`
updates the image's working source. `commit actors/counter` records a snapshot
with an automatically generated readable revision ID; `history actors/counter`
lists commits and `rollback actors/counter <revision>` restores a working file.
Committing is version control only: spawn always compiles the current image
file, including uncommitted changes. Invalid working code fails spawn without
creating an actor; commits can preserve invalid work in progress.
`reload <actor>` loads current source at its existing path, and
`become <actor> <path>` switches scripts, preserving state and queued messages
after the active handler completes. `stop` runs optional `on_shutdown(state)`
after the active handler and then removes the actor; `kill` removes it
immediately without shutdown. Removal discards identity, state, mailbox, and
retained logs. Failed actors remain inspectable until killed. `restart` creates
a new incarnation with the same state and loaded script, without reloading.
Restart is independent of stop/kill; a removed actor cannot be restarted.
Pause/resume remain cooperative controls between scheduler slices. Use `help` in
the session for the complete command list.
State and messages support null, booleans, signed 64-bit integers, strings,
lists, records, and distinct actor-reference values. `spawn <path> <json>`
sets initial state; `send <actor> <json>` sends any supported value. JSON
objects of the form `{"$actor": "name-or-@reference"}` embed an actor reference,
including inside a record or list. See the [runtime guide](docs/runtime.md)
for Rune conversions and value limits. Protocol version 5 uses
`handle(ctx, state, sender, message) -> new_state`; `ctx.send(recipient, message)`
records sends with the calling actor as sender, and
`ctx.send_from(sender, recipient, message)` supports forwarding.
Returning state is optional: an omitted return, `return;`, or `()` retains the
previously committed state, discarding invocation-local state mutations.
Valid sends still commit and the message is acknowledged. The same rule applies
to `on_shutdown(state)`. To replace state, return a non-unit portable value.
Unit can no longer be used to reset state to null; use host state editing instead.
Sender metadata is separate from message data; host-originated messages have
unit/null sender. System messages are trusted runtime commands, not data an
actor can construct.

See [the feasibility report](docs/rune-feasibility.md) for verified Rune
capabilities and limitations, and [the runtime guide](docs/runtime.md) for
the runtime contract.

## VS Code client

To install into your regular VS Code instance, run `npm --prefix vscode ci`
followed by `mise run install-extension`, then reload the window. This builds the
Rust debug adapter and packages/installs the extension using the local `code`
CLI. Set `VSCODE_CLI` to choose another executable, such as `code-insiders`.
See [extension setup](vscode/README.md#run-from-source) for debugger configuration.

With Node.js 22+ and npm installed, run `npm --prefix vscode ci`. In VS Code,
open this repository and launch **Run Runa Extension** from Run and Debug
(F5). Start `mise run start`, then use **Runa: Connect to Local Image**
in the Extension Development Host. Open the **Runa** Activity Bar sidebar.
Its **Actors** view lists
actors created through the REPL or another client as expandable directories.
Open `meta.json` for read-only identity/status/behavior-path metadata,
`config.json` for live log-level configuration, `state.json` for natural JSON
state, `mailbox.json` for reserved/queued messages, or `actor.log` for retained logs.
State and mailbox JSON use two-space indentation and preserve exact integers.
The `.log` file uses syslog-style lines and VS Code's built-in log highlighting.
State and mailbox files are writable only for inactive paused/failed actors.
Server notifications automatically refresh the tree and open files;
**Runa: Refresh Image** is also available on demand.

The **Image Sources** view in the Runa sidebar automatically shows the image's virtual
filesystem beneath **source**, including `actors/counter.rn`, without adding a
workspace folder. Open a file and save edits to update the image without
committing. Right-click a source file to **Spawn** (null state) or **Spawn with
State** (which prompts for JSON), or to commit, browse history, and roll back. Historical revisions
remain read-only.
The source toolbar and context menus support creating, renaming/moving, and
deleting files. Renaming an in-use behavior asks to update every actor using its
path; declining cancels the rename. Deleting a source used by a running actor
is rejected with a stop-first error.

**Edit Current Behavior** on an actor opens its loaded script (or existing
temporary script). Saving compiles an actor-specific override and installs it
at the next handler boundary without changing shared source, state, or mailbox.
Every actor directory includes `behavior.rn`, even without a temporary override.
Set breakpoints there and use **Debug Actor** to debug that actor with VS Code's
Run and Debug controls. Build the standalone Rust DAP server with
`cargo build -p runa-dap` and set `runa.debugAdapterPath` to its absolute path,
or install it on PATH with `cargo install --path crates/runa-dap`.
**Reload Actor Source**
clears the override, **Go to Behavior** opens the shared source, and **Update Behavior**
asks before copying the temporary text into shared working source without committing.

Use **Runa: Spawn Actor** to choose a source path and supply JSON initial state.
Actor context menus offer structured messages and system lifecycle commands.
**Runa: Advance Scheduler** explicitly executes ticks; sending alone only queues
a message. The extension does not start or shut down the image. Disconnecting
leaves actors in the server. Detailed runtime event subscriptions and source-level
debugging are not implemented yet. Rune compilation errors and warnings
appear inline and in Problems for open image sources and actor-specific behavior
documents, including unsaved edits. Validation does not save, commit, or install
the editor text. See the [extension guide](vscode/README.md)
for connection restrictions, snapshot semantics, and tests.

## Repairing actors and reading logs

State and messages can be edited only on paused or failed actors with
no active invocation. An active handler must finish or be cancelled through the
runtime cancellation API; kill removes the actor rather than making it editable. State
edits preserve messages; message edits preserve state. Neither restarts an actor.
Reserved messages are retried first on restart; clearing the reserved array
removes the failed input.

```text
set-state @<actor-ref> 1
set-messages @<actor-ref> {"reserved":[],"queued":[{"op":"add","value":2}]}
log-level @<actor-ref> debug
logs @<actor-ref>
logs -f
```

These are REPL commands and `runa-cli` subcommands (quote JSON when using a shell).
REPL log filters also accept session-local actor names. Omitting the filter selects all
actors. `logs` prints retained history; `logs -f` additionally subscribes to
future entries until Ctrl+C, which returns to the REPL or closes a one-shot CLI.
In VS Code, actor context menus offer **Edit Actor State...**,
**Edit Reserved and Queued Messages...**, and **Set Actor Log Level...**.
The actor directory provides editable state/mailbox files and a `config.json`
that remains writable while running. The **Inspect Actor** command also retains
the combined JSON editor; protected fields there are restored with a warning.
The Actors toolbar advances the scheduler,
and **Copy Actor Reference** copies a reference from the actor context menu.

Rune handlers can call `ctx.log("info", "message")`, or `ctx.trace`, `ctx.debug`,
`ctx.info`, `ctx.warn`, `ctx.error`, and `ctx.fatal` with a string message.
Severity order is trace < debug < info < warn < error < fatal; the default
threshold is info. Only entries at or above the actor's threshold are retained
or emitted. Logs are immediate diagnostics, not transactional effects: they
survive handler failure and may repeat on retries. Each actor keeps its latest
1,000 accepted messages in memory, each limited to 512 UTF-8 bytes. Restart
preserves logs and the threshold. Server exit loses them. Rebuild the server,
CLI, and extension together for protocol v5.

The read-only `ctx.this` field is the current actor's typed reference:
`ctx.send(ctx.this, message)` queues a normal transactional message to itself.

## Development

Install [mise](https://mise.jdx.dev/) and run `mise install` to install Tombi.
Rust, rustfmt, Clippy, Node.js 22+, and npm must also be available on your PATH.
Install the extension's locked dependencies with `npm --prefix vscode ci`.

```sh
mise run format
mise run lint
mise run test
mise run pre-commit
```

The pre-commit task runs formatting, linting, and tests in that order, stopping
on the first failure. Run it successfully before committing and review any
formatter changes. This is a task, not an automatically installed Git hook.

Tombi formats TOML files offline; rustfmt formats all workspace Rust code;
Prettier formats the extension. Clippy checks all workspace targets with
warnings treated as errors. Extension checks include strict TypeScript and
Prettier. The test task runs the Rust workspace tests, extension unit tests,
and a TypeScript-to-Rust WebSocket interoperability test.
`mise run test-extension-host` additionally downloads VS Code 1.96.4 into an
ignored test cache and tests the actual inspector commands in an isolated
Extension Development Host. It requires a graphical display; on headless Linux,
use `xvfb-run -a mise run test-extension-host`.

## Project structure

```text
.
├── Cargo.toml
├── crates/
│   ├── runa/             # Image server, protocol dispatcher, WebSocket transport
│   ├── runa-cli/         # Interactive REPL client
│   ├── runa-dap/         # Standalone stdio Debug Adapter Protocol server
│   ├── runa-core/        # Shared stable identities and portable actor values
│   ├── runa-actor/       # Actor value contracts
│   ├── runa-code/        # Source identities and compile cache
│   ├── runa-rune/        # Phase 0 Rune experiments
│   ├── runa-runtime/     # Single-worker actor prototype
│   ├── runa-image-api/   # Trusted in-process image operations
│   ├── runa-protocol/    # Versioned transport-neutral wire DTOs
│   └── runa-client/      # Loopback WebSocket client
├── docs/
│   ├── architecture.md
│   ├── protocol.md
│   └── decisions/
├── runa-std/
│   ├── core/
│   ├── actors/
│   ├── ui/
│   └── tools/
├── scripts/
├── vscode/
├── examples/
└── tests/
```

`runa-std/` is reserved for the default Rune scripts loaded into an image.
`scripts/` is for general-purpose development automation. Root-level `tests/`
is reserved for system and end-to-end tests; it is not automatically discovered
by Cargo. Add an explicit runner when those tests are introduced, keeping Rust
crate integration tests within each crate.

Only the Phase 0 and minimal Phase 1 crates are present. Remaining crates will
be added as their functionality is implemented.