broker: add stopped-preserving upgrade without actor execution

headless: register signal handlers synchronously, split tests, drop deployment examples and wording
Add --headless mode (no stdin read, runs until SIGTERM/exit) + systemd/NixOS examples
2026-10-22 06:49:58 +00:00 · 2026-10-18 15:31:41 +00:00 · 2026-10-18 15:24:20 +00:00 · 2026-10-17 17:22:50 +00:00 · 2026-10-17 16:36:08 +00:00 · 2026-10-17 16:33:31 +00:00
24 changed files with 1507 additions and 197 deletions
@@ -1,4 +1,4 @@
-# Runs on the Nix runner on nexus (label `nixos`, host executor).
+# Runs on a self-hosted Nix runner (label `nixos`, host executor).
 name: ci
 on: [push, pull_request]

@@ -678,6 +678,15 @@ dependencies = [
 "windows-sys 0.59.0",
 ]

+[[package]]
+name = "fsevent-sys"
+version = "4.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "76ee7a02da4d231650c7cea31349b889be2f45ddb3ef3032d2ec8185f6313fd2"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "futures"
 version = "0.3.34"
@@ -1066,6 +1075,26 @@ dependencies = [
 "serde_core",
 ]

+[[package]]
+name = "inotify"
+version = "0.11.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4cc00ea907cab49550b7da656f80ebb97be1b997d931fbcd28d39734e17ce592"
+dependencies = [
+ "bitflags",
+ "inotify-sys",
+ "libc",
+]
+
+[[package]]
+name = "inotify-sys"
+version = "0.1.8"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c033f80b2c113cdf91ab7a33faa9cbc014726dcad99880c8609af2a370edf37d"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "instant"
 version = "0.1.13"
@@ -1165,6 +1194,26 @@ dependencies = [
 "wasm-bindgen",
 ]

+[[package]]
+name = "kqueue"
+version = "1.2.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8d763e5b24120b4ddf50de6c92308156765aabfbbccebf401da7cff2d70a41ea"
+dependencies = [
+ "kqueue-sys",
+ "libc",
+]
+
+[[package]]
+name = "kqueue-sys"
+version = "1.1.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "07293a4e297ac234359b510362495713f75ea345d5307140414f20c69ffeb087"
+dependencies = [
+ "bitflags",
+ "libc",
+]
+
 [[package]]
 name = "leb128"
 version = "0.2.7"
@@ -1259,6 +1308,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1788edb87fdc09c7e26304471e2f5be8cdefb1b6930d6e3985fc02ff53bf86ee"
 dependencies = [
 "libc",
+ "log",
 "wasi",
 "windows-sys 0.61.2",
 ]
@@ -1269,6 +1319,33 @@ version = "1.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "743fb55ba31b18fb1ecef6bdc9aa2743314978ac084044301a7eee33fb99a20d"

+[[package]]
+name = "notify"
+version = "8.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4d3d07927151ff8575b7087f245456e549fea62edf0ec4e565a5ee50c8402bc3"
+dependencies = [
+ "bitflags",
+ "fsevent-sys",
+ "inotify",
+ "kqueue",
+ "libc",
+ "log",
+ "mio",
+ "notify-types",
+ "walkdir",
+ "windows-sys 0.60.2",
+]
+
+[[package]]
+name = "notify-types"
+version = "2.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "42b8cfee0e339a0337359f3c88165702ac6e600dc01c0cc9579a92d62b08477a"
+dependencies = [
+ "bitflags",
+]
+
 [[package]]
 name = "object"
 version = "0.40.0"
@@ -1580,6 +1657,15 @@ version = "1.0.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "cf54715a573b99ac80df0bc206da022bcd442c974952c7b9720069370852e21f"

+[[package]]
+name = "same-file"
+version = "1.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "93fc1dc3aaa9bfed95e02e6eadabb4baf7e3078b0bd1b4d7b6b0b68378900502"
+dependencies = [
+ "winapi-util",
+]
+
 [[package]]
 name = "sausage"
 version = "0.1.0"
@@ -1590,6 +1676,7 @@ dependencies = [
 "http",
 "http-body-util",
 "include_dir",
+ "notify",
 "serde",
 "serde_json",
 "tokio",
@@ -2075,6 +2162,16 @@ version = "1.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "317211a0dc0ceedd78fb2ca9a44aed3d7b9b26f81870d485c07122b4350673b7"

+[[package]]
+name = "walkdir"
+version = "2.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "29790946404f91d9c5d06f9874efddea1dc06c5efe94541a7d6863108e3a5e4b"
+dependencies = [
+ "same-file",
+ "winapi-util",
+]
+
 [[package]]
 name = "want"
 version = "0.3.2"
@@ -63,12 +63,14 @@ Each instance sees three folders:

 | Path | Access | Contents |
 | --- | --- | --- |
-| `/data` | read-write | the actor's state |
+| `/data` | read-write by its actor; readable by every actor | the actor's state |
 | `/public` | read-write | the files that the actor publishes |
 | `/actor` | read-only | the `actor/` folder of the world |

+An actor can write only its own `/data`, but can read every actor's `/data` through `/actor/<id>/data/`. `/data` is not private: any unprotected secret file there is readable by every actor.
+
 `/actor/<id>/` holds everything about an actor: `actor.wasm`, `actor.json`
-(installed version, running, subscriptions, tapped), `source/`, `public/`, `data/`,
+(installed version, running, subscriptions, tapped, watches), `source/`, `public/`, `data/`,
 `stderr.log`, and `versions/<n>/`. `ls /actor` lists all actors.

 Files are plain POSIX files: a write is visible at once. To publish a file
@@ -110,26 +112,29 @@ Any request can be answered with `{"type": "fs_error", "errno": "no-entry"}`.
 - No reply within 10 seconds, or an actor that does not exist, makes the
  operation fail (`io`, or `no-entry`). An actor that opens its own `fs/`
  waits for itself and fails after the timeout.
- `fs` is reserved: no actor can have that ID.
+- `fs` is reserved: no actor can have that ID. It also sends `fs_changed`
+  to watchers (see Watches).

 ## The broker

 Requests to `broker`. The answer is a message whose `reply_to` is the ID of
 the request: the request's `type` in past tense (`spawned`, `upgraded`,
 `rolled_back`, `stopped`, `started`, `removed`, `subscribed`, `unsubscribed`,
-`tapping`, `untapped`, `scanned`, `exiting`), or `{"type": "error", "reason": "..."}`.
+`tapping`, `untapped`, `watching`, `unwatched`, `exiting`), or `{"type": "error", "reason": "..."}`.

 | `type` | Fields | Effect |
 | --- | --- | --- |
 | `spawn` | `actor`, `wasm`, `source`?, `built_by`? | test, then start a new actor |
-| `upgrade` | `actor`, `wasm`, `source`?, `built_by`? | test, then swap the version |
+| `upgrade` | `actor`, `wasm`, `source`?, `built_by`?, `preserve_stopped`? (boolean, default `false`) | test, then swap the version; with `preserve_stopped: true`, a stopped actor is compiled and installed without execution and stays stopped |
 | `rollback` | `actor`, `version` | go back to an older version |
 | `stop`, `start` | `actor` | stop or start an actor |
 | `remove` | `actor` | stop the actor (normal shutdown), then delete its whole folder `/actor/<id>/` (all versions, `data/`, `public/`, logs); answer `{"type": "removed", "actor"}`. Messages to it are then undeliverable, and the ID can be spawned again. An unknown actor or a native one (`broker`, `console`, `cargo`, `fs`) is an `error` |
 | `subscribe`, `unsubscribe` | `actor` | receive (or not) what `actor` publishes |
 | `tap`, `untap` | | receive (or not) a copy of every message the broker routes |
-| `scan` | | boot new or changed `actor.wasm` files |
-| `exit` | | answer `exiting`, stop every actor, and end the `sausage` process (a wrapper like `nix run .#loop` starts it again, updated) |
+| `watch`, `unwatch` | `path` | receive (or not) `fs_changed` for every change of `path` or anything under it; answer `{"type": "watching" \| "unwatched", "path"}` with the path as an `/actor/...` path |
+| `exit` | | answer `exiting`, stop every actor, and end the `sausage` process with exit code 75 (a wrapper like `nix run .#loop` starts it again, updated) |
+
+**Headless.** `sausage run --headless [dir]` (or `sausage --headless`, or `SAUSAGE_HEADLESS=1`) turns the console input off: the host does not read stdin, so end of input (`/dev/null`) does not stop the world. `nix run .#loop -- --headless <world>` passes it through the loop. It runs until SIGTERM or SIGINT (stops every actor, exit code 0) or a broker `exit` request (exit code 75, as above). Without the flag, EOF on stdin or `/quit` ends the host.

 Paths are `/actor/...`, or the sender's `/public/...` or `/data/...`.

@@ -151,11 +156,37 @@ including its own), for debugging, logging, or a live view of the world:
 - Like subscriptions, a tap is saved in `actor.json` (`"tapped": true`) and
  survives restarts and upgrades. `untap` ends it.

+**Watches.** The host watches the whole world folder with the OS file
+watcher: every change inside it counts, whoever makes it (actors, a 9P mount,
+a shell or `git` on the host). Actor-served files (`/actor/<id>/fs/`) are not
+on disk and do not count. After `watch {path}`, the sender gets for each
+change of `path` or a path under it (`/actor/a` covers `/actor/a/x`, not
+`/actor/ab`):
+
+```json
+{"type": "fs_changed", "path": "/actor/notes/public/list.md", "op": "write"}
+```
+
+- `op` is `write`, `create`, `remove`, or `rename` (then `to` is the new
+  path; a watcher of either path gets it).
+- Changes of one path within about 100 ms become one event (a `write` after
+  a `create` stays a `create`).
+- The host's temporary files (names starting with `.sausage-tmp-`) are
+  skipped; renaming one onto a file is a `write` of that file.
+- Events come from `fs`. Like tap copies, they are not routed, a full
+  mailbox drops them, and they never block the writer. Never answer them.
+- If the OS drops events, every watcher gets one
+  `{"type": "fs_changed", "path": "/actor", "op": "rescan"}`: anything may
+  have changed.
+- A watch is saved in `actor.json` (`"watches": [...]`) and survives
+  restarts and upgrades. `unwatch` with the same path ends it.
+
 **Versions.** A new version (from `spawn`, `upgrade`, or a changed
-`actor.wasm` on disk) is tested first, while the old version keeps running.
+`actor.wasm` on disk) is tested first, while the old version keeps running
+(except the stopped-preserving upgrade described below).
 Only when its tests pass does the broker copy it to `versions/<n>/`
 (`actor.wasm`, `source/`, `meta.json`) and switch to it. So `versions/` holds
-exactly the versions that were installed, and instances always run from
+exactly the versions that were installed (including stopped installations), and instances always run from
 there. If the tests fail, nothing changes, and the sender gets an `error`
 with the reason. All versions share the actor's `data/` and `public/`: a
 new version must read the data of the old one.
@@ -184,16 +215,35 @@ instance. A version without test files passes.
 top-level `actor.wasm` and `source/` a copy of the new version, and starts
 it. A rollback does the same with the older version.

+**Stopped-preserving upgrade.** Send `upgrade` with `"preserve_stopped": true`
+to install a stopped actor's new program without starting it, running `main`
+(even as a candidate), or reading or sending any of its mail. The host compiles
+the WASM and copies the new version, source and metadata as usual, increments
+the version, and keeps `actor.json`'s `running: false`. Its mailbox, `data/`
+and `public/` stay untouched. Because candidate tests execute `main`, they are
+not run in this case: the `upgraded` reply contains
+`"tests": {"skipped": "preserve_stopped"}`. Test the build separately before
+installing it this way. An explicit `start` later runs the installed version
+and delivers the queued mail. No startup/crash probation runs for this stopped
+installation. Running actors still use the normal tested swap, even with this
+option. Omitting the option (or `false`) retains the existing behavior,
+including starting a stopped actor. Explicit stop/start requests serialize
+with upgrades of the same actor.
+
 **Crashes.** The broker restarts a crashed actor after 0.5 s, then doubles
 the delay up to 30 s. A new version that crashes 3 times within 30 seconds
 after an upgrade goes back to the previous version. The actor that asked for
 the upgrade gets `{"type": "rolled_back", "from", "to", "reason", "last_exit"}`.

-**The folder is the truth.** At start, and on `scan`, the broker boots every
-`actor/<id>/actor.wasm`. If the file differs from the installed version, it
-is a new version, tested with the tests in `actor/<id>/source/`. If its
-tests fail, the file stays where it is, the reason goes to `stderr.log`, and
-the next scan tries again.
+**The folder is the truth.** At start, the broker boots every
+`actor/<id>/actor.wasm`. Afterwards, the file watcher reports a new or
+changed `actor/<id>/actor.wasm` (or a new folder `actor/<id>/`) once it is
+quiet for about 100 ms: a new file becomes a new actor, a changed one an
+upgrade, tested first with the tests in `actor/<id>/source/`. A file equal
+to the installed version changes nothing. If its tests fail, the file stays
+where it is, the reason goes to `stderr.log`, and the next change of the
+file tries again. When the OS drops events, the broker checks every actor
+folder again, as at start.

 ## The `cargo` actor

@@ -201,9 +251,19 @@ the next scan tries again.
 {"type": "build", "source": "/actor/agent/public/drafts/notes"}
 ```

-The reply is `{"type": "built", "wasm": "/actor/cargo/public/builds/<n>/<name>.wasm", "built_by": "cargo", ...}`
-or `{"type": "build_failed", "errors": "..."}`. Pass `wasm`, `source`, and
-`built_by` on to `spawn` or `upgrade`.
+The build request may include `"skip_tests": true` to skip Rust unit tests.
+By default, `cargo` first runs `cargo test` for the host target (not WASI),
+so pure-logic unit tests in a `wasm32-wasip2` actor crate can run without a
+WASI test runner. If tests fail or cannot compile, the build fails before the
+WASI build. The reply carries `tests: {"passed": <count>, "failed": <count>,
+"output": "<cargo test stdout and stderr>"}`; counts sum Cargo's test-result
+summaries, and a test command/compilation failure without a summary counts as
+one failure. With `skip_tests: true`, counts are zero and output says tests
+were skipped.
+
+The reply is `{"type": "built", "wasm": "/actor/cargo/public/builds/<n>/<name>.wasm", "built_by": "cargo", "tests": {...}, ...}`
+or `{"type": "build_failed", "errors": "...", "tests": {...}}`. Pass
+`wasm`, `source`, and `built_by` on to `spawn` or `upgrade`.

 A Rust actor depends on the SDK with `sausage-actor = "0.1"`. `cargo` points
 that dependency at the SDK of the host, which it publishes at
@@ -10,28 +10,55 @@ change its own code.

 ## Start

-You need [Nix](https://nixos.org/download) with flakes.
+You need [Nix](https://nixos.org/download) with flakes, and an API token for
+an LLM (OpenAI Responses or Anthropic Messages API). No clone needed:

 ```sh
-nix run -- init     # creates ./world and builds the agent (a few minutes)
-nix run -- run      # starts the world and a console
+nix run git+https://gitea.surma.technology/surma/sausage
+```
+
+This creates a world in `./world` the first time (it builds the agent, which
+takes a few minutes), then starts it with a console. Later runs start the
+existing world. The two steps on their own:
+
+```sh
+nix run git+https://gitea.surma.technology/surma/sausage -- init   # creates ./world and builds the agent
+nix run git+https://gitea.surma.technology/surma/sausage -- run    # starts the world and a console
 ```

 `init` asks for the API shape (`openai-responses` or `anthropic-messages`),
 the base URL, the model, and the token. It reads `SAUSAGE_API`,
 `SAUSAGE_BASE_URL`, `SAUSAGE_MODEL`, and `SAUSAGE_API_TOKEN` first. The
 values go to `world/actor/agent/data/config.json`. Both commands take another
-folder as an argument, for example `nix run -- run my-world`.
+folder as an argument, for example `... -- run my-world`.

 To keep a world on the newest version, run it in a loop:

 ```sh
-nix run .#loop -- my-world            # or .#loop-internal (Gitea on the LAN)
+nix run git+https://gitea.surma.technology/surma/sausage#loop               # ./world, created if needed
+nix run git+https://gitea.surma.technology/surma/sausage#loop -- my-world   # an existing world
 ```

-The loop runs `nix run --refresh git+ssh://…/surma/sausage -- run my-world`.
+The loop runs `nix run --refresh git+https://gitea.surma.technology/surma/sausage -- run my-world`.
 When the host crashes or an actor sends `@broker {"type":"exit"}`, the loop
-fetches the newest commit and starts it again. `/quit` ends the loop.
+fetches the newest commit and starts it again. `/quit` ends the loop. Set
+`SAUSAGE_FLAKE` to run another flake URL instead, for example your fork.
+
+From a clone (`git clone https://gitea.surma.technology/surma/sausage`), the
+same commands are `nix run`, `nix run -- init`, `nix run -- run` and
+`nix run .#loop`.
+
+### Headless
+
+`--headless` (or `SAUSAGE_HEADLESS=1`) turns the console input off: the host
+never reads stdin, so end of input does not stop the world. It runs until
+SIGTERM/SIGINT (clean stop, exit code 0) or a broker `exit` request (exit
+code 75, so the loop starts the newest version). Create the world first
+(`init`), then pass the flag through the loop:
+
+```sh
+nix run git+https://gitea.surma.technology/surma/sausage#loop -- --headless my-world
+```

 In the console:

@@ -51,14 +78,15 @@ actor.wasm     the current version
 actor.json     installed version, running, subscriptions
 source/        the source of the current version
 public/        files that the actor publishes
-data/          the actor's state
+data/          actor state; owner-writable, world-readable via /actor/<id>/data/ (not private)
 stderr.log     what the actor and the broker write about it
 versions/<n>/  every installed version: actor.wasm, source/, meta.json
 ```

 To add or replace an actor by hand, put an `actor.wasm` (and its `source/`,
-with the tests in `source/tests/`) into its folder. Then restart the world or
-send `@broker {"type":"scan"}`. The broker tests the file before it uses it.
+with the tests in `source/tests/`) into its folder; the host watches the
+world folder and picks the file up once it stops changing. The broker tests
+the file before it uses it.
 If the tests fail, the old version keeps running, and `stderr.log` says why.

 ## More
@@ -14,11 +14,49 @@ pub struct Config {
    pub base_url: String,
    pub model: String,
    pub token: String,
-    /// Optional reasoning effort for OpenAI models ("low", "medium", ...).
+    /// Optional reasoning effort for OpenAI models only (wins over `thinking` there).
    #[serde(default)]
    pub reasoning_effort: Option<String>,
    #[serde(default)]
    pub max_tokens: Option<u32>,
+    /// Thinking effort for both APIs: "low", "medium", "high", "xhigh", "max" (`LEVELS`), or "off"/unset for none.
+    /// Anthropic: adaptive thinking + output_config.effort; OpenAI: reasoning.effort. A level the model rejects is
+    /// replaced by the nearest accepted one (logged), never by "off".
+    #[serde(default)]
+    pub thinking: Option<String>,
+}
+
+/// Thinking levels, lowest first.
+pub const LEVELS: [&str; 5] = ["low", "medium", "high", "xhigh", "max"];
+
+/// The level nearest to `want` (lower on a tie) among those a 400 error lists in quotes (else all levels), not
+/// yet `tried`. None: nothing left, so the request goes out without an effort field (model default).
+fn fallback_effort(want: &str, error: &str, tried: &[String]) -> Option<String> {
+    let listed: Vec<&str> = error.split('\'').skip(1).step_by(2).filter(|w| LEVELS.contains(w)).collect();
+    let pool = if listed.iter().any(|l| *l != want) { listed } else { LEVELS.to_vec() };
+    let at = LEVELS.iter().position(|l| *l == want).unwrap_or(1) as i32;
+    pool.into_iter()
+        .filter(|l| *l != want && !tried.iter().any(|t| t == l))
+        .min_by_key(|l| {
+            let i = LEVELS.iter().position(|x| x == l).unwrap_or(1) as i32;
+            ((i - at).abs(), i > at)
+        })
+        .map(str::to_string)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::fallback_effort;
+
+    #[test]
+    fn nearest_accepted_level() {
+        let anthropic = "HTTP 400: output_config.effort: Input should be 'low', 'medium', 'high' or 'max'";
+        assert_eq!(fallback_effort("xhigh", anthropic, &["xhigh".into()]).as_deref(), Some("high"));
+        let openai = "HTTP 400: Invalid value: 'ultra'. Supported values are: 'none', 'minimal', 'low', 'medium', 'high', 'xhigh', and 'max'.";
+        assert_eq!(fallback_effort("ultra", openai, &["ultra".into()]).as_deref(), Some("medium"));
+        // No list in the error: step to the neighbour, never to off.
+        assert_eq!(fallback_effort("max", "HTTP 400: effort max not supported", &["max".into()]).as_deref(), Some("xhigh"));
+    }
 }

 #[derive(Serialize, Deserialize, Clone, Debug)]
@@ -32,7 +70,7 @@ pub enum Entry {
    },
    ToolCall { id: String, name: String, arguments: String },
    ToolResult { id: String, output: String },
-    /// A provider item that only this API shape understands (OpenAI reasoning).
+    /// A provider item that only this API shape understands (OpenAI reasoning, Anthropic thinking).
    Opaque { api: String, item: Value },
 }

@@ -43,10 +81,28 @@ pub struct Tool {
 }

 pub async fn complete(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]) -> Result<Vec<Entry>, String> {
-    match cfg.api.as_str() {
-        "openai-responses" => openai(cfg, system, tools, history).await,
-        "anthropic-messages" => anthropic(cfg, system, tools, history).await,
-        other => Err(format!("unknown API shape {other:?}")),
+    let want = match cfg.api.as_str() {
+        "openai-responses" => cfg.reasoning_effort.clone().or_else(|| cfg.thinking.clone()),
+        _ => cfg.thinking.clone(),
+    }
+    .filter(|t| t != "off");
+    let mut effort = want.clone();
+    let mut tried = Vec::new();
+    loop {
+        let result = match cfg.api.as_str() {
+            "openai-responses" => openai(cfg, system, tools, history, effort.as_deref()).await,
+            "anthropic-messages" => anthropic(cfg, system, tools, history, want.is_some(), effort.as_deref()).await,
+            other => return Err(format!("unknown API shape {other:?}")),
+        };
+        match (result, effort.clone()) {
+            // The model rejected the effort level: retry with the nearest accepted one (thinking stays on).
+            (Err(e), Some(current)) if e.starts_with("HTTP 400") && e.contains("effort") => {
+                tried.push(current.clone());
+                effort = fallback_effort(want.as_deref().unwrap_or(""), &e, &tried);
+                eprintln!("agent: {} rejected thinking effort {current:?}; retrying with {effort:?}: {e}", cfg.model);
+            }
+            (result, _) => return result,
+        }
    }
 }

@@ -67,7 +123,7 @@ async fn post(url: &str, headers: &[(&str, String)], body: &Value) -> Result<Val
    serde_json::from_str(&text).map_err(|e| format!("invalid JSON in response: {e}"))
 }

-async fn openai(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]) -> Result<Vec<Entry>, String> {
+async fn openai(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry], effort: Option<&str>) -> Result<Vec<Entry>, String> {
    let mut input = Vec::new();
    for e in history {
        match e {
@@ -104,7 +160,7 @@ async fn openai(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]) -
    if let Some(max_tokens) = cfg.max_tokens {
        body["max_output_tokens"] = json!(max_tokens);
    }
-    if let Some(effort) = &cfg.reasoning_effort {
+    if let Some(effort) = effort {
        body["reasoning"] = json!({"effort": effort});
    }
    let url = format!("{}/v1/responses", cfg.base_url);
@@ -143,7 +199,14 @@ async fn openai(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]) -
    Ok(out)
 }

-async fn anthropic(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]) -> Result<Vec<Entry>, String> {
+async fn anthropic(
+    cfg: &Config,
+    system: &str,
+    tools: &[Tool],
+    history: &[Entry],
+    thinking: bool,
+    effort: Option<&str>,
+) -> Result<Vec<Entry>, String> {
    // Merge consecutive blocks of the same role into one message.
    let mut messages: Vec<(String, Vec<Value>)> = Vec::new();
    let mut push = |role: &str, block: Value| match messages.last_mut() {
@@ -163,6 +226,8 @@ async fn anthropic(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]
            Entry::ToolResult { id, output } => {
                push("user", json!({"type": "tool_result", "tool_use_id": id, "content": output}))
            }
+            // Thinking blocks go back unchanged (with their signature): required next to tool use.
+            Entry::Opaque { api, item } if api == "anthropic-messages" && thinking => push("assistant", item.clone()),
            Entry::Opaque { .. } => {}
        }
    }
@@ -171,13 +236,19 @@ async fn anthropic(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]
        .iter()
        .map(|t| json!({"name": t.name, "description": t.description, "input_schema": t.parameters}))
        .collect();
-    let body = json!({
+    let mut body = json!({
        "model": cfg.model,
        "system": system,
        "messages": messages,
        "tools": tools,
        "max_tokens": cfg.max_tokens.unwrap_or(32000),
    });
+    if thinking {
+        body["thinking"] = json!({"type": "adaptive"});
+        if let Some(effort) = effort {
+            body["output_config"] = json!({"effort": effort});
+        }
+    }
    let url = format!("{}/v1/messages", cfg.base_url);
    let resp = post(
        &url,
@@ -199,6 +270,9 @@ async fn anthropic(cfg: &Config, system: &str, tools: &[Tool], history: &[Entry]
                name: block["name"].as_str().unwrap_or_default().to_string(),
                arguments: block["input"].to_string(),
            }),
+            Some("thinking") | Some("redacted_thinking") => {
+                out.push(Entry::Opaque { api: "anthropic-messages".into(), item: block })
+            }
            _ => {}
        }
    }
@@ -3,3 +3,5 @@ People talk to you through other actors. Your text goes to your subscribers.
 Before writing an actor, read `/actor/broker/public/PROTOCOL.md` for messages, the broker, `cargo`, and tests.
 Read the SDK at `/actor/cargo/public/sdk/` before you write an actor.
 You can change your own source at `/actor/agent/source`.
+
+The global filesystem is readable by every actor, including every actor's `/data` at `/actor/<id>/data/`. An actor can write only its own `/data` and `/public`. `/data` is not private: any unprotected secret file there is readable by every actor.
@@ -15,6 +15,7 @@ clap = { version = "4", features = ["derive"] }
 http = "1"
 http-body-util = "0.1"
 include_dir = "0.7"
+notify = "8"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 tokio = { version = "1", features = ["full"] }
@@ -1,11 +1,12 @@
-//! The broker: message routing, mailboxes, subscriptions, taps, the version
-//! store, and the requests to `broker` (spawn, upgrade, rollback, stop, start,
-//! remove, subscribe, unsubscribe, tap, untap, scan, exit). `supervisor.rs`
-//! runs each WASM actor.
+//! The broker: message routing, mailboxes, subscriptions, taps, watches, the
+//! version store, and the requests to `broker` (spawn, upgrade, rollback,
+//! stop, start, remove, subscribe, unsubscribe, tap, untap, watch, unwatch,
+//! exit). `supervisor.rs` runs each WASM actor.
 //!
 //! The broker accepts only `.wasm` files. It does not know any source language.
-//! The world folder is the source of truth: `scan` boots every
-//! `actor/<id>/actor.wasm` and treats a changed file as an upgrade.
+//! The world folder is the source of truth: at start, the broker boots every
+//! `actor/<id>/actor.wasm`; afterwards the file watcher (`watch.rs`) reports
+//! changes: a new `actor.wasm` becomes an actor, a changed one an upgrade.

 use std::collections::{BTreeMap, HashMap};
 use std::path::Path;
@@ -167,14 +168,19 @@ impl Broker {
    /// they are never tapped themselves; a full mailbox drops the copy.
    fn tap(&self, tappers: &[String], message: Value) {
        let ts = std::time::SystemTime::now().duration_since(std::time::UNIX_EPOCH).map_or(0, |d| d.as_millis() as u64);
-        let body = json!({"type": "tapped", "message": message, "ts": ts});
+        self.inject(tappers, "broker", json!({"type": "tapped", "message": message, "ts": ts}));
+    }
+
+    /// Put a message of the host straight into the mailboxes of `to`: not
+    /// routed (so not tapped), never blocking, dropped if a mailbox is full.
+    fn inject(&self, to: &[String], from: &str, body: Value) {
        let mailboxes = self.mailboxes.lock().unwrap();
-        for tapper in tappers {
-            if let Some(tx) = mailboxes.get(tapper) {
+        for id in to {
+            if let Some(tx) = mailboxes.get(id) {
                let _ = tx.try_send(Envelope {
                    id: format!("m-{}", self.seq.fetch_add(1, Ordering::Relaxed)),
-                    from: "broker".into(),
-                    to: tapper.clone(),
+                    from: from.into(),
+                    to: id.clone(),
                    reply_to: None,
                    body: body.clone(),
                });
@@ -307,12 +313,16 @@ impl Broker {
        Ok(component)
    }

-    /// Test a new version where it lies. Only when the tests pass, copy it
-    /// to `versions/<n>/`: that folder holds only versions that run(ran).
-    /// Returns the component and the test report.
-    async fn try_version(&self, id: &str, meta: &VersionMeta, wasm: &Path, source: Option<&Path>) -> Result<(Component, Value)> {
+    /// Compile a new version where it lies and, normally, test it. Copy it
+    /// to `versions/<n>/` only on success. A stopped-preserving install skips
+    /// execution, including candidate tests. Return the component and report.
+    async fn try_version(&self, id: &str, meta: &VersionMeta, wasm: &Path, source: Option<&Path>, run_tests: bool) -> Result<(Component, Value)> {
        let component = self.load(id, wasm)?;
-        let tests = candidate::test(self, id, meta.version, &component, source).await;
+        let tests = if run_tests {
+            candidate::test(self, id, meta.version, &component, source).await
+        } else {
+            Ok(json!({"skipped": "preserve_stopped"}))
+        };
        if let Err(e) = &tests {
            self.log(id, &format!("version {} from {} failed its tests: {e:#}", meta.version, meta.wasm));
        }
@@ -326,11 +336,11 @@ impl Broker {
    }

    /// Make `version` the current `actor.wasm`, `source/`, and `actor.json`.
-    pub fn install(&self, id: &str, version: u64) -> Result<()> {
+    pub fn install(&self, id: &str, version: u64, running: bool) -> Result<()> {
        self.world.install_files(id, version)?;
        self.update_meta(id, |m| {
            m.version = Some(version);
-            m.running = true;
+            m.running = running;
        })
    }

@@ -353,26 +363,29 @@ impl Broker {
    }

    async fn create_version(self: &Arc<Self>, id: &str, meta: &VersionMeta, wasm: &Path, source: Option<&Path>) -> Result<Value> {
-        let (_, tests) = self.try_version(id, meta, wasm, source).await?;
-        self.install(id, meta.version)?;
+        let (_, tests) = self.try_version(id, meta, wasm, source, true).await?;
+        self.install(id, meta.version, true)?;
        self.start_supervisor(id, meta.version, true);
        Ok(json!({"type": "spawned", "actor": id, "version": meta.version, "tests": tests}))
    }

-    pub async fn upgrade_actor(self: &Arc<Self>, id: &str, wasm: &Path, source: Option<&Path>, meta: VersionMeta) -> Result<Value> {
+    pub async fn upgrade_actor(self: &Arc<Self>, id: &str, wasm: &Path, source: Option<&Path>, meta: VersionMeta, preserve_stopped: bool) -> Result<Value> {
        let lock = self.op_lock(id);
        let _guard = lock.lock().await;
        if !self.is_supervised(id) {
            bail!("no WASM actor with ID {id}");
        }
        let version = self.world.next_version(id);
-        self.upgrade_version(id, &VersionMeta { version, ..meta }, wasm, source).await
+        self.upgrade_version(id, &VersionMeta { version, ..meta }, wasm, source, preserve_stopped).await
    }

-    async fn upgrade_version(self: &Arc<Self>, id: &str, meta: &VersionMeta, wasm: &Path, source: Option<&Path>) -> Result<Value> {
-        let (component, tests) = self.try_version(id, meta, wasm, source).await?;
+    async fn upgrade_version(self: &Arc<Self>, id: &str, meta: &VersionMeta, wasm: &Path, source: Option<&Path>, preserve_stopped: bool) -> Result<Value> {
+        // Lifecycle requests hold the same op lock, so a stopped actor cannot
+        // start between this decision and the supervisor's installation.
+        let run_tests = !preserve_stopped || self.meta(id).is_some_and(|m| m.running);
+        let (component, tests) = self.try_version(id, meta, wasm, source, run_tests).await?;
        let (version, requester) = (meta.version, meta.requested_by.clone());
-        let mut result = self.control(id, |reply| Control::Upgrade { version, component, requester, reply }).await?;
+        let mut result = self.control(id, |reply| Control::Upgrade { version, component, requester, preserve_stopped, reply }).await?;
        result["tests"] = tests;
        Ok(result)
    }
@@ -406,31 +419,90 @@ impl Broker {
        Ok(json!({"type": "removed", "actor": id}))
    }

-    /// Boot the actors in `actor/`. Called at start and on `scan`.
+    /// Boot the actors in `actor/`: at start, and as a full pass when the
+    /// file watcher dropped events (`overflowed`).
    ///
    /// - An `actor.wasm` equal to the installed version starts as is.
    /// - A new or changed `actor.wasm` is tested where it lies. A new actor
    ///   is installed, a changed one is an upgrade. If its tests fail,
-    ///   nothing changes, and the file stays for the next scan.
+    ///   nothing changes, and the file stays until it changes again.
    pub async fn scan(self: &Arc<Self>) -> Vec<Value> {
        let mut report = Vec::new();
        for id in self.world.actor_ids().unwrap_or_default() {
-            if NATIVE.contains(&id.as_str()) || !self.world.wasm_path(&id).is_file() {
-                continue;
-            }
-            let lock = self.op_lock(&id);
-            let _guard = lock.lock().await;
-            if let Some(r) = self.scan_actor(&id).await {
+            if let Some(r) = self.boot_actor(&id).await {
                report.push(r);
            }
        }
        report
    }

+    /// The file watcher saw `actor/<id>/actor.wasm` (or the folder) appear or
+    /// change: start, install or upgrade `id` as at boot, and log the result.
+    pub async fn check_actor(self: Arc<Self>, id: String) {
+        if let Some(result) = self.boot_actor(&id).await {
+            self.log(&id, &format!("actor.wasm changed: {result}"));
+            eprintln!("sausage: {result}");
+        }
+    }
+
+    /// The OS file watcher dropped events: tell every watcher with one
+    /// `fs_changed {path: "/actor", op: "rescan"}`, then run a full pass.
+    pub async fn overflowed(self: Arc<Self>) -> Vec<Value> {
+        self.host_error("the file watcher dropped events; watchers get a rescan, actors a full pass");
+        let watchers: Vec<String> =
+            self.metas.lock().unwrap().iter().filter(|(_, m)| !m.watches.is_empty()).map(|(id, _)| id.clone()).collect();
+        self.inject(&watchers, vfs::ID, json!({"type": "fs_changed", "path": "/actor", "op": "rescan"}));
+        let report = self.scan().await;
+        for change in &report {
+            eprintln!("sausage: {change}");
+        }
+        report
+    }
+
+    /// Send `fs_changed` for one coalesced change (from `watch.rs`) to every
+    /// actor that watches `path` or `to`. Like taps: straight into the
+    /// mailbox, dropped if it is full.
+    pub fn notify_watchers(&self, path: &str, op: &str, to: Option<&str>) {
+        let watchers: Vec<String> = {
+            let metas = self.metas.lock().unwrap();
+            metas
+                .iter()
+                .filter(|(_, m)| m.watches.iter().any(|w| under(path, w) || to.is_some_and(|t| under(t, w))))
+                .map(|(id, _)| id.clone())
+                .collect()
+        };
+        if watchers.is_empty() {
+            return;
+        }
+        let mut body = json!({"type": "fs_changed", "path": path, "op": op});
+        if let Some(to) = to {
+            body["to"] = json!(to);
+        }
+        self.inject(&watchers, vfs::ID, body);
+    }
+
+    /// A problem of the host itself: the broker's `stderr.log` and the terminal.
+    pub fn host_error(&self, text: &str) {
+        self.log("broker", text);
+        eprintln!("sausage: {text}");
+    }
+
+    async fn boot_actor(self: &Arc<Self>, id: &str) -> Option<Value> {
+        if NATIVE.contains(&id) || !world::valid_id(id) {
+            return None;
+        }
+        let lock = self.op_lock(id);
+        let _guard = lock.lock().await;
+        if !self.world.wasm_path(id).is_file() {
+            return None;
+        }
+        self.scan_actor(id).await
+    }
+
    async fn scan_actor(self: &Arc<Self>, id: &str) -> Option<Value> {
-        let meta = self.world.read_meta(id);
+        // The broker's copy wins: actors change theirs (subscribe, ...) without the op lock.
+        let meta = self.metas.lock().unwrap().entry(id.to_string()).or_insert_with(|| self.world.read_meta(id)).clone();
        let installed = meta.version.filter(|v| self.world.version_dir(id, *v).join("actor.wasm").is_file());
-        self.metas.lock().unwrap().insert(id.to_string(), meta.clone());
        if let Some(v) = installed
            && !self.is_supervised(id)
        {
@@ -452,7 +524,7 @@ impl Broker {
        let source = source.is_dir().then_some(source.as_path());
        let result = match installed {
            None => self.create_version(id, &vmeta, &wasm, source).await,
-            Some(_) => self.upgrade_version(id, &vmeta, &wasm, source).await,
+            Some(_) => self.upgrade_version(id, &vmeta, &wasm, source, false).await,
        };
        Some(result.unwrap_or_else(|e| json!({"type": "error", "actor": id, "reason": format!("{e:#}")})))
    }
@@ -499,7 +571,11 @@ impl Broker {
            }
            "upgrade" => {
                let (wasm, source, meta) = version_meta()?;
-                self.upgrade_actor(&actor()?, &wasm, source.as_deref(), meta).await
+                let preserve_stopped = match b.get("preserve_stopped") {
+                    None => false,
+                    Some(value) => value.as_bool().ok_or_else(|| anyhow!("preserve_stopped must be a boolean"))?,
+                };
+                self.upgrade_actor(&actor()?, &wasm, source.as_deref(), meta, preserve_stopped).await
            }
            "rollback" => {
                let id = actor()?;
@@ -508,8 +584,15 @@ impl Broker {
                let _guard = lock.lock().await;
                self.control(&id, |reply| Control::Rollback { version, reply }).await
            }
-            "stop" => self.control(&actor()?, |reply| Control::Stop { reply }).await,
-            "start" => self.control(&actor()?, |reply| Control::Start { reply }).await,
+            "stop" | "start" => {
+                let id = actor()?;
+                let lock = self.op_lock(&id);
+                let _guard = lock.lock().await;
+                self.control(&id, |reply| match ty {
+                    "stop" => Control::Stop { reply },
+                    _ => Control::Start { reply },
+                }).await
+            },
            "remove" => self.remove_actor(&actor()?).await,
            "subscribe" | "unsubscribe" => {
                let publisher = actor()?;
@@ -529,14 +612,32 @@ impl Broker {
                let answer = if on { "tapping" } else { "untapped" };
                Ok(json!({"type": answer}))
            }
-            "scan" => Ok(json!({"type": "scanned", "changes": self.scan().await})),
+            "watch" | "unwatch" => {
+                let on = ty == "watch";
+                let path_s = b["path"].as_str().ok_or_else(|| anyhow!("missing \"path\""))?;
+                let global = self.world.global_path(&self.world.resolve(&env.from, path_s)?);
+                let path = global.ok_or_else(|| anyhow!("{path_s} is not under /actor"))?;
+                let path = match path.trim_end_matches('/') {
+                    "" => "/actor".to_string(),
+                    p => p.to_string(),
+                };
+                self.update_meta(&env.from, |m| {
+                    if on {
+                        m.watches.insert(path.clone());
+                    } else {
+                        m.watches.remove(&path);
+                    }
+                })?;
+                let answer = if on { "watching" } else { "unwatched" };
+                Ok(json!({"type": answer, "path": path}))
+            }
            "exit" => {
                self.log("broker", &format!("exit requested by {}", env.from));
                // notify_one keeps the permit if nobody waits yet.
                self.exit.notify_one();
                Ok(json!({"type": "exiting"}))
            }
-            other => bail!("unknown broker request {other:?}. Known: spawn, upgrade, rollback, stop, start, remove, subscribe, unsubscribe, tap, untap, scan, exit"),
+            other => bail!("unknown broker request {other:?}. Known: spawn, upgrade, rollback, stop, start, remove, subscribe, unsubscribe, tap, untap, watch, unwatch, exit"),
        }
    }

@@ -563,6 +664,11 @@ fn same_file(a: &Path, b: &Path) -> bool {
    matches!((std::fs::read(a), std::fs::read(b)), (Ok(x), Ok(y)) if x == y)
 }

+/// Whether the namespace path `path` is `prefix` or inside it.
+fn under(path: &str, prefix: &str) -> bool {
+    path.strip_prefix(prefix).is_some_and(|rest| rest.is_empty() || rest.starts_with('/'))
+}
+
 pub fn truncate(s: &str, max: usize) -> String {
    if s.len() <= max {
        return s.to_string();
@@ -1,8 +1,9 @@
 //! The native `cargo` actor: builds Rust source folders into `.wasm` files.
 //!
 //! Request:  `{"type": "build", "source": "/actor/<id>/public/<folder>"}`
-//! Replies:  `{"type": "built", "wasm": "/actor/cargo/public/builds/<n>/<name>.wasm", ...}`
-//!           `{"type": "build_failed", "errors": "..."}`
+//!           optional `"skip_tests": true` skips native `cargo test` (tests run by default).
+//! Replies:  `{"type": "built", "wasm": "/actor/cargo/public/builds/<n>/<name>.wasm", "tests": {"passed": 1, "failed": 0, "output": "..."}, ...}`
+//!           `{"type": "build_failed", "errors": "...", "tests": {"passed": 0, "failed": 1, "output": "..."}}`
 //!
 //! The broker knows nothing about Rust. This actor is one way to get a
 //! `.wasm` file. Other actors can add other languages.
@@ -32,11 +33,49 @@ pub const TARGET: &str = "wasm32-wasip2";
 /// The source of the `sausage-actor` SDK.
 static SDK: Dir<'_> = include_dir!("$CARGO_MANIFEST_DIR/../sdk");

+#[derive(Clone, Default, Debug)]
+pub struct TestResults {
+    pub passed: u64,
+    pub failed: u64,
+    pub output: String,
+}
+
+impl TestResults {
+    fn value(&self) -> Value {
+        json!({"passed": self.passed, "failed": self.failed, "output": self.output})
+    }
+}
+
 pub struct BuildOutput {
    /// Host path of the published `.wasm` file.
    pub wasm: PathBuf,
    pub seconds: f64,
    pub warnings: String,
+    pub tests: TestResults,
+}
+
+#[derive(Debug)]
+pub struct BuildFailure {
+    pub reason: String,
+    pub tests: TestResults,
+}
+
+impl std::fmt::Display for BuildFailure {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        if self.tests.failed > 0 && !self.tests.output.is_empty() {
+            write!(f, "{}:\n{}", self.reason, self.tests.output)
+        } else {
+            f.write_str(&self.reason)
+        }
+    }
+}
+
+impl std::error::Error for BuildFailure {}
+
+impl From<anyhow::Error> for BuildFailure {
+    fn from(error: anyhow::Error) -> Self {
+        Self { reason: format!("{error:#}"), tests: TestResults::default() }
+    }
 }

 /// All builds share one target folder, so they run one at a time.
@@ -51,51 +90,84 @@ impl Cargo {
        Ok(Self { world })
    }

-    /// Build the crate in `source` (a host path) and publish the `.wasm` file
-    /// under `/actor/cargo/public/builds/<n>/`.
-    pub async fn build(&self, source: &Path) -> Result<BuildOutput> {
+    /// Run native `cargo test` unless opted out, then build and publish the
+    /// crate's `wasm32-wasip2` binary under `/actor/cargo/public/builds/<n>/`.
+    pub async fn build(&self, source: &Path, skip_tests: bool) -> std::result::Result<BuildOutput, BuildFailure> {
        if !source.join("Cargo.toml").is_file() {
-            bail!("{} has no Cargo.toml", source.display());
+            return Err(BuildFailure {
+                reason: format!("{} has no Cargo.toml", source.display()),
+                tests: TestResults::default(),
+            });
        }
        let started = Instant::now();
        let data = self.world.data_dir(ID);
        let build_dir = data.join("work");
        world::replace_source(source, &build_dir)?;
        let manifest = build_dir.join("Cargo.toml");
-        std::fs::write(&manifest, prepare_manifest(&std::fs::read_to_string(&manifest)?, Some(&self.world.public_dir(ID).join("sdk")))?)?;
+        let original = std::fs::read_to_string(&manifest).context("read build manifest")?;
+        let prepared = prepare_manifest(&original, Some(&self.world.public_dir(ID).join("sdk")))?;
+        std::fs::write(&manifest, prepared).context("write prepared manifest")?;

-        let mut cmd = Command::new("cargo");
-        cmd.arg("build")
-            .arg("--release")
-            .arg("--target")
-            .arg(TARGET)
-            .arg("--message-format=json-render-diagnostics")
-            .arg("--target-dir")
-            .arg(data.join("target"))
-            .current_dir(&build_dir)
-            .kill_on_drop(true);
-        let out = cmd.output().await.context("cannot run cargo (is it on PATH, with the wasm32-wasip2 target?)")?;
-        let stderr = String::from_utf8_lossy(&out.stderr).to_string();
-        if !out.status.success() {
-            bail!("{}", strip_progress(&stderr));
-        }
-        let mut artifact = None;
-        for line in String::from_utf8_lossy(&out.stdout).lines() {
-            if let Ok(v) = serde_json::from_str::<Value>(line)
-                && v["reason"] == "compiler-artifact"
-                && let Some(exe) = v["executable"].as_str()
-            {
-                artifact = Some(PathBuf::from(exe));
+        let target_dir = data.join("target");
+        let tests = if skip_tests {
+            TestResults { output: "skipped (skip_tests=true)".into(), ..TestResults::default() }
+        } else {
+            match run_tests(&build_dir, &target_dir).await {
+                Ok(tests) => tests,
+                Err(error) => {
+                    let output = format!("cannot run cargo tests: {error:#}");
+                    return Err(BuildFailure {
+                        reason: output.clone(),
+                        tests: TestResults { failed: 1, output, ..TestResults::default() },
+                    });
+                }
            }
+        };
+        if tests.failed > 0 {
+            return Err(BuildFailure { reason: "unit tests failed".into(), tests });
        }
-        let artifact = artifact.ok_or_else(|| anyhow!("cargo built no binary. Does the crate have a src/main.rs?"))?;

-        let n = next_build_number(&self.world.public_dir(ID).join("builds"))?;
-        let dest_dir = self.world.public_dir(ID).join("builds").join(n.to_string());
-        let name = artifact.file_name().context("artifact has no file name")?;
-        let dest = dest_dir.join(name);
-        world::write_atomic(&dest, &std::fs::read(&artifact)?)?;
-        Ok(BuildOutput { wasm: dest, seconds: started.elapsed().as_secs_f64(), warnings: strip_progress(&stderr) })
+        let build_result: Result<BuildOutput> = async {
+            let mut cmd = Command::new("cargo");
+            cmd.arg("build")
+                .arg("--release")
+                .arg("--target")
+                .arg(TARGET)
+                .arg("--message-format=json-render-diagnostics")
+                .arg("--target-dir")
+                .arg(&target_dir)
+                .current_dir(&build_dir)
+                .kill_on_drop(true);
+            let out = cmd.output().await.context("cannot run cargo (is it on PATH, with the wasm32-wasip2 target?)")?;
+            let stderr = String::from_utf8_lossy(&out.stderr).to_string();
+            if !out.status.success() {
+                bail!("{}", strip_progress(&stderr));
+            }
+            let mut artifact = None;
+            for line in String::from_utf8_lossy(&out.stdout).lines() {
+                if let Ok(v) = serde_json::from_str::<Value>(line)
+                    && v["reason"] == "compiler-artifact"
+                    && let Some(exe) = v["executable"].as_str()
+                {
+                    artifact = Some(PathBuf::from(exe));
+                }
+            }
+            let artifact = artifact.ok_or_else(|| anyhow!("cargo built no binary. Does the crate have a src/main.rs?"))?;
+
+            let n = next_build_number(&self.world.public_dir(ID).join("builds"))?;
+            let dest_dir = self.world.public_dir(ID).join("builds").join(n.to_string());
+            let name = artifact.file_name().context("artifact has no file name")?;
+            let dest = dest_dir.join(name);
+            world::write_atomic(&dest, &std::fs::read(&artifact)?)?;
+            Ok(BuildOutput {
+                wasm: dest,
+                seconds: started.elapsed().as_secs_f64(),
+                warnings: strip_progress(&stderr),
+                tests: tests.clone(),
+            })
+        }
+        .await;
+        build_result.map_err(|error| BuildFailure { reason: format!("{error:#}"), tests })
    }

    /// The native `cargo` actor. Builds run one at a time, in order.
@@ -106,7 +178,7 @@ impl Cargo {
                continue;
            }
            let body = self.handle(&broker, &env).await.unwrap_or_else(|e| {
-                json!({"type": "build_failed", "source": env.body.get("source"), "errors": format!("{e:#}")})
+                json!({"type": "build_failed", "source": env.body.get("source"), "errors": format!("{e:#}"), "tests": TestResults::default().value()})
            });
            broker.send(ID, Some(&env.from), None, Some(&env.id), body);
        }
@@ -117,22 +189,76 @@ impl Cargo {
            Some("build") => {
                let source_s = env.body.get("source").and_then(Value::as_str).ok_or_else(|| anyhow!("missing \"source\""))?;
                let source = broker.world.resolve(&env.from, source_s)?;
-                let out = self.build(&source).await?;
-                let wasm = self.world.global_path(&out.wasm).unwrap_or_else(|| out.wasm.display().to_string());
-                Ok(json!({
-                    "type": "built",
-                    "wasm": wasm,
-                    "source": source_s,
-                    "built_by": ID,
-                    "seconds": (out.seconds * 10.0).round() / 10.0,
-                    "warnings": out.warnings,
-                }))
+                let skip_tests = env.body.get("skip_tests").and_then(Value::as_bool).unwrap_or(false);
+                match self.build(&source, skip_tests).await {
+                    Ok(out) => {
+                        let wasm = self.world.global_path(&out.wasm).unwrap_or_else(|| out.wasm.display().to_string());
+                        Ok(json!({
+                            "type": "built",
+                            "wasm": wasm,
+                            "source": source_s,
+                            "built_by": ID,
+                            "seconds": (out.seconds * 10.0).round() / 10.0,
+                            "warnings": out.warnings,
+                            "tests": out.tests.value(),
+                        }))
+                    }
+                    Err(failure) => Ok(json!({
+                        "type": "build_failed",
+                        "source": source_s,
+                        "errors": failure.reason,
+                        "tests": failure.tests.value(),
+                    })),
+                }
            }
            other => bail!("unknown request {other:?}. Known: build"),
        }
    }
 }

+/// Run tests for the host triple: WASI crates' pure-logic tests can use the
+/// normal Rust test harness without needing a WASI test runner.
+async fn run_tests(build_dir: &Path, target_dir: &Path) -> Result<TestResults> {
+    let mut cmd = Command::new("cargo");
+    cmd.arg("test")
+        .arg("--target-dir")
+        .arg(target_dir)
+        .current_dir(build_dir)
+        .kill_on_drop(true);
+    let out = cmd.output().await.context("cannot run cargo test")?;
+    let stdout = String::from_utf8_lossy(&out.stdout).to_string();
+    let stderr = String::from_utf8_lossy(&out.stderr).to_string();
+    let combined = match (stdout.is_empty(), stderr.is_empty()) {
+        (true, true) => String::new(),
+        (false, true) => stdout,
+        (true, false) => stderr,
+        (false, false) => format!("{stdout}\n{stderr}"),
+    };
+    let output = strip_progress(&combined);
+    let passed = test_count(&output, "passed");
+    let mut failed = test_count(&output, "failed");
+    // Compilation failures and runner errors have no `test result` summary,
+    // but still make the default test phase fail.
+    if !out.status.success() && failed == 0 {
+        failed = 1;
+    }
+    Ok(TestResults { passed, failed, output })
+}
+
+/// Count the `passed` or `failed` totals in cargo's per-target test summaries.
+fn test_count(output: &str, field: &str) -> u64 {
+    output
+        .lines()
+        .filter_map(|line| line.split_once("test result:").map(|(_, rest)| rest))
+        .flat_map(|summary| summary.split(';'))
+        .filter_map(|part| {
+            let words: Vec<_> = part.split_whitespace().collect();
+            let index = words.iter().position(|word| *word == field)?;
+            words.get(index.checked_sub(1)?)?.parse::<u64>().ok()
+        })
+        .sum()
+}
+
 /// Point a `sausage-actor` dependency at `sdk` (keeping its features), and
 /// add an empty `[workspace]`, so the crate never joins a parent workspace.
 fn prepare_manifest(text: &str, sdk: Option<&Path>) -> Result<String> {
@@ -3,7 +3,8 @@
 //! - `@<actor> <json>` sends a JSON body. `@<actor> <text>` sends
 //!   `{"type": "text", "text": ...}`.
 //! - A line without `@` goes to `agent` as text.
-//! - `/help`, `/quit`. EOF on stdin also quits.
+//! - `/help`, `/quit`. EOF on stdin also quits (not with `--headless`, which
+//!   never reads stdin).

 use std::sync::Arc;

@@ -1,7 +1,8 @@
 //! Sausage: grow a system of WASI actors by talking to it.
 //!
 //! The host runs the broker, the native actors (`broker`, `cargo`, and the
-//! console), and one supervisor per WASM actor. See `PROTOCOL.md`.
+//! console), one supervisor per WASM actor, and the file watcher of the
+//! world folder (`watch.rs`). See `PROTOCOL.md`.

 pub mod broker;
 pub mod candidate;
@@ -10,6 +11,7 @@ pub mod console;
 pub mod runtime;
 pub mod supervisor;
 pub mod vfs;
+pub mod watch;
 pub mod world;

 use std::path::Path;
@@ -30,12 +32,13 @@ pub struct Host {
    pub world: World,
    pub broker: Arc<Broker>,
    native: Vec<tokio::task::JoinHandle<()>>,
+    watcher: Option<watch::Watcher>,
    _lock: std::fs::File,
 }

 impl Host {
-    /// Start the broker, `cargo`, and every registered actor of the world at
-    /// `root`. Only one host can run a world at a time.
+    /// Start the broker, `cargo`, the file watcher, and every registered
+    /// actor of the world at `root`. Only one host can run a world at a time.
    pub async fn start(root: &Path) -> Result<Host> {
        let world = World::new(std::path::absolute(root)?);
        if !world.exists() {
@@ -53,10 +56,14 @@ impl Host {
            tokio::spawn(broker.clone().run_actor(broker.open_mailbox("broker"))),
            tokio::spawn(cargo.run_actor(broker.clone(), broker.open_mailbox(cargo_actor::ID))),
        ];
+        // Watch first, so that no change during the boot pass is lost.
+        let watcher = watch::start(broker.clone())
+            .inspect_err(|e| broker.host_error(&format!("no file watching: {e}")))
+            .ok();
        for change in broker.scan().await {
            eprintln!("sausage: {change}");
        }
-        Ok(Host { world, broker, native, _lock: lock })
+        Ok(Host { world, broker, native, watcher, _lock: lock })
    }

    /// Register a native actor with the ID `id` and return its mailbox.
@@ -64,9 +71,10 @@ impl Host {
        self.broker.open_mailbox(id)
    }

-    /// Stop the broker and `cargo` (with the requests they still run), then
-    /// every actor, gracefully.
-    pub async fn stop(self) {
+    /// Stop the file watcher, the broker and `cargo` (with the requests they
+    /// still run), then every actor, gracefully.
+    pub async fn stop(mut self) {
+        drop(self.watcher.take());
        for task in &self.native {
            task.abort();
        }
@@ -5,6 +5,7 @@ use anyhow::{Context, Result, bail};
 use clap::{Parser, Subcommand};
 use include_dir::{Dir, include_dir};
 use serde_json::json;
+use tokio::signal::unix::Signal;

 use sausage::world::{ActorMeta, World};
 use sausage::{Host, cargo_actor, console};
@@ -12,11 +13,18 @@ use sausage::{Host, cargo_actor, console};
 /// The source of the agent actor. `init` builds it like any other actor.
 static AGENT_SOURCE: Dir<'_> = include_dir!("$CARGO_MANIFEST_DIR/../agent");

+// Without a command, `sausage` runs the world in `./world` and creates it
+// first (like `init`) if it does not exist yet.
 #[derive(Parser)]
 #[command(version, about)]
 struct Cli {
+    /// Do not read stdin, so end of input does not stop the world. It runs
+    /// until SIGTERM (or SIGINT), or until an actor sends
+    /// `@broker {"type":"exit"}`. Also set by SAUSAGE_HEADLESS=1.
+    #[arg(long, global = true)]
+    headless: bool,
    #[command(subcommand)]
-    command: Cmd,
+    command: Option<Cmd>,
 }

 #[derive(Subcommand)]
@@ -83,19 +91,47 @@ async fn init(dir: PathBuf) -> Result<()> {
    std::fs::create_dir_all(&source)?;
    AGENT_SOURCE.extract(&source).context("extract agent source")?;
    println!("Building the agent actor (the first build takes a few minutes) …");
-    let out = cargo_actor::Cargo::new(world.clone())?.build(&source).await.context("build agent")?;
+    let out = cargo_actor::Cargo::new(world.clone())?.build(&source, false).await.context("build agent")?;
    std::fs::copy(&out.wasm, world.wasm_path("agent"))?;
    println!("Built in {:.1} s. Start the world with: sausage run {}", out.seconds, dir.display());
    Ok(())
 }

-async fn run(dir: PathBuf) -> Result<()> {
+/// Install the SIGTERM and SIGINT handlers now (not when first polled).
+fn termination_signals() -> Result<(Signal, Signal)> {
+    use tokio::signal::unix::{SignalKind, signal};
+    let term = signal(SignalKind::terminate()).context("install SIGTERM handler")?;
+    let int = signal(SignalKind::interrupt()).context("install SIGINT handler")?;
+    Ok((term, int))
+}
+
+/// Resolves on SIGTERM or SIGINT.
+async fn termination((mut term, mut int): (Signal, Signal)) {
+    tokio::select! {
+        _ = term.recv() => {}
+        _ = int.recv() => {}
+    }
+}
+
+async fn run(dir: PathBuf, headless: bool) -> Result<()> {
+    // Register the signal handlers before the host starts, so a signal that
+    // arrives early is never the default (kill) action.
+    let signals = if headless { Some(termination_signals()?) } else { None };
    let host = Host::start(&dir).await?;
    tokio::spawn(console::run_mailbox(host.mailbox(console::ID)));
    let broker = host.broker.clone();
-    let exit = tokio::select! {
-        _ = console::run_input(host.broker.clone()) => false,
-        _ = broker.exit_requested() => true,
+    let exit = match signals {
+        Some(signals) => {
+            println!("Sausage headless: running until SIGTERM or an exit request.");
+            tokio::select! {
+                _ = termination(signals) => false,
+                _ = broker.exit_requested() => true,
+            }
+        }
+        None => tokio::select! {
+            _ = console::run_input(host.broker.clone()) => false,
+            _ = broker.exit_requested() => true,
+        },
    };
    println!("Stopping actors …");
    host.stop().await;
@@ -111,8 +147,18 @@ async fn run(dir: PathBuf) -> Result<()> {

 #[tokio::main]
 async fn main() -> Result<()> {
-    match Cli::parse().command {
-        Cmd::Init { dir } => init(dir).await,
-        Cmd::Run { dir } => run(dir).await,
+    let cli = Cli::parse();
+    let headless = cli.headless || std::env::var("SAUSAGE_HEADLESS").is_ok_and(|v| !v.is_empty() && v != "0");
+    match cli.command {
+        Some(Cmd::Init { dir }) => init(dir).await,
+        Some(Cmd::Run { dir }) => run(dir, headless).await,
+        None => {
+            let dir = PathBuf::from("world");
+            if !dir.join("actor").is_dir() {
+                println!("No world in ./world yet, creating it (same as `sausage init`).");
+                init(dir.clone()).await?;
+            }
+            run(dir, headless).await
+        }
    }
 }
@@ -23,7 +23,7 @@ pub type Reply = oneshot::Sender<Result<Value>>;

 pub enum Control {
    /// Swap to `version` (already in `versions/`, compiled as `component`).
-    Upgrade { version: u64, component: Component, requester: String, reply: Reply },
+    Upgrade { version: u64, component: Component, requester: String, preserve_stopped: bool, reply: Reply },
    Rollback { version: u64, reply: Reply },
    Stop { reply: Reply },
    Start { reply: Reply },
@@ -221,24 +221,29 @@ impl Supervisor {
    /// Make `version` current. It is compiled first (unless `component` is
    /// given): if that fails, nothing changes. `data/` and `public/` stay as
    /// they are.
-    async fn install(&mut self, version: u64, component: Option<Component>) -> Result<()> {
+    async fn install(&mut self, version: u64, component: Option<Component>, running: bool) -> Result<()> {
        let component = match component {
            Some(c) => c,
            None => self.broker.load_version(&self.id, version)?,
        };
        self.stop().await;
-        self.broker.install(&self.id, version)?;
+        self.broker.install(&self.id, version, running)?;
        self.version = version;
        self.component = Some(component);
-        self.running = true;
+        self.running = running;
        self.restarts = 0;
        self.restart_at = None;
        Ok(())
    }

-    async fn upgrade(&mut self, version: u64, component: Component, requester: String) -> Result<Value> {
+    async fn upgrade(&mut self, version: u64, component: Component, requester: String, preserve_stopped: bool) -> Result<Value> {
        let previous = self.version;
-        self.install(version, Some(component)).await?;
+        let running = self.running || !preserve_stopped;
+        self.install(version, Some(component), running).await?;
+        if !running {
+            self.probation = None;
+            return Ok(json!({"type": "upgraded", "actor": self.id, "version": version, "previous": previous}));
+        }
        self.probation = Some(Probation { previous, since: Instant::now(), crashes: 0, requester });
        if let Err(e) = self.start().await {
            self.rollback(previous).await?;
@@ -249,7 +254,7 @@ impl Supervisor {

    async fn rollback(&mut self, version: u64) -> Result<Value> {
        let from = self.version;
-        self.install(version, None).await?;
+        self.install(version, None, true).await?;
        self.probation = None;
        self.start().await?;
        Ok(json!({"type": "rolled_back", "actor": self.id, "from": from, "to": version}))
@@ -258,8 +263,8 @@ impl Supervisor {
    /// Handle a control request. Returns `true` when the supervisor ends.
    async fn handle(&mut self, ctl: Control) -> bool {
        match ctl {
-            Control::Upgrade { version, component, requester, reply } => {
-                let _ = reply.send(self.upgrade(version, component, requester).await);
+            Control::Upgrade { version, component, requester, preserve_stopped, reply } => {
+                let _ = reply.send(self.upgrade(version, component, requester, preserve_stopped).await);
            }
            Control::Rollback { version, reply } => {
                let _ = reply.send(self.rollback(version).await);
@@ -328,14 +328,15 @@ impl VfsView<'_> {

    async fn native_directory_entries(&mut self, fd: Resource<types::Descriptor>) -> FsResult<Vec<DirectoryEntry>> {
        let stream = HostDescriptor::read_directory(&mut self.inner, fd).await?;
+        let stream_rep = stream.rep();
        let mut entries = Vec::new();
        loop {
-            match HostDirectoryEntryStream::read_directory_entry(&mut self.inner, stream.clone()).await? {
+            match HostDirectoryEntryStream::read_directory_entry(&mut self.inner, Resource::new_borrow(stream_rep)).await? {
                Some(entry) => entries.push(entry),
                None => break,
            }
        }
-        HostDirectoryEntryStream::drop(&mut self.inner, stream)?;
+        HostDirectoryEntryStream::drop(&mut self.inner, stream).map_err(|_| ErrorCode::Io)?;
        Ok(entries)
    }

@@ -587,10 +588,10 @@ impl HostDescriptor for VfsView<'_> {
            None => {
                let child_path = self.state.dirs.get(&fd.rep()).map(|base| join(base, &path));
                let opened = HostDescriptor::open_at(&mut self.inner, fd, path_flags, path, oflags, flags).await?;
-                if let Some(path) = child_path
-                    && HostDescriptor::get_type(&mut self.inner, opened.clone()).await? == DescriptorType::Directory
-                {
-                    self.state.dirs.insert(opened.rep(), path);
+                if let Some(path) = child_path {
+                    if HostDescriptor::get_type(&mut self.inner, Resource::new_borrow(opened.rep())).await? == DescriptorType::Directory {
+                        self.state.dirs.insert(opened.rep(), path);
+                    }
                }
                Ok(opened)
            }
@@ -0,0 +1,256 @@
+//! File watching: the host watches the whole world folder with the OS file
+//! watcher (`notify`, inotify on Linux), whoever writes: actors, a 9P mount,
+//! a shell or `git` on the host. Nothing outside the folder counts.
+//!
+//! Events are coalesced per path: once a path is quiet for `QUIET` (or has
+//! kept changing for `MAX_DELAY`), it becomes one change
+//! `{path, op: write | create | remove | rename, to?}` with a namespace path
+//! (`/actor/...`). Then the broker
+//!
+//! - sends `fs_changed` to the actors that watch the path (`watch` request),
+//! - starts or upgrades an actor whose `actor/<id>/actor.wasm` (or folder
+//!   `actor/<id>`) changed, tested first, as at boot (`Broker::check_actor`).
+//!
+//! The host's temporary files (`.sausage-tmp-*`) are skipped: a rename of one
+//! onto a file is a `write` of that file. If the OS drops events, the broker
+//! runs a full pass and tells every watcher (`Broker::overflowed`).
+
+use std::collections::HashMap;
+use std::path::Path;
+use std::sync::Arc;
+use std::time::Duration;
+
+use notify::event::{EventKind, ModifyKind, RenameMode};
+use notify::{RecursiveMode, Watcher as _};
+use tokio::sync::mpsc;
+use tokio::time::Instant;
+
+use crate::broker::Broker;
+use crate::world::TMP_PREFIX;
+
+/// A path that is quiet this long becomes one change.
+pub const QUIET: Duration = Duration::from_millis(100);
+/// A path that keeps changing becomes a change at least this often.
+pub const MAX_DELAY: Duration = Duration::from_secs(1);
+
+/// The running watcher. Dropping it stops watching.
+pub struct Watcher {
+    _os: notify::RecommendedWatcher,
+    task: tokio::task::JoinHandle<()>,
+}
+
+impl Drop for Watcher {
+    fn drop(&mut self) {
+        self.task.abort();
+    }
+}
+
+/// Watch the world folder of `broker` recursively. A failure to watch part
+/// of it (for example too few inotify watches) is logged, and the rest is
+/// watched.
+pub fn start(broker: Arc<Broker>) -> notify::Result<Watcher> {
+    let (tx, rx) = mpsc::unbounded_channel();
+    let mut os = notify::recommended_watcher(move |event| {
+        let _ = tx.send(event);
+    })?;
+    if let Err(e) = os.watch(&broker.world.root, RecursiveMode::Recursive) {
+        broker.host_error(&format!("cannot watch all of {}: {e}", broker.world.root.display()));
+    }
+    let task = tokio::spawn(run(broker, rx));
+    Ok(Watcher { _os: os, task })
+}
+
+/// A change waiting for its path to be quiet.
+struct Pending {
+    op: &'static str,
+    to: Option<String>,
+    first: Instant,
+    last: Instant,
+}
+
+impl Pending {
+    fn due(&self) -> Instant {
+        (self.last + QUIET).min(self.first + MAX_DELAY)
+    }
+}
+
+async fn run(broker: Arc<Broker>, mut events: mpsc::UnboundedReceiver<notify::Result<notify::Event>>) {
+    let root = broker.world.actors_root();
+    let mut pending: HashMap<String, Pending> = HashMap::new();
+    // Actor checks and full passes. Dropping the set (when the host stops) aborts them.
+    let mut tasks = tokio::task::JoinSet::new();
+    loop {
+        let due = pending.values().map(Pending::due).min();
+        tokio::select! {
+            event = events.recv() => match event {
+                Some(Ok(event)) => {
+                    if event.need_rescan() {
+                        let broker = broker.clone();
+                        tasks.spawn(async move {
+                            broker.overflowed().await;
+                        });
+                    }
+                    let now = Instant::now();
+                    for (path, op, to) in changes(&root, &event) {
+                        add(&mut pending, path, op, to, now);
+                    }
+                }
+                Some(Err(e)) => broker.host_error(&format!("file watcher: {e}")),
+                None => break,
+            },
+            _ = sleep_until(due) => {
+                let now = Instant::now();
+                let ready: Vec<String> = pending.iter().filter(|(_, p)| p.due() <= now).map(|(k, _)| k.clone()).collect();
+                for path in ready {
+                    let p = pending.remove(&path).unwrap();
+                    let op = if p.op == "moved-in" { "create" } else { p.op };
+                    broker.notify_watchers(&path, op, p.to.as_deref());
+                    if let Some(id) = new_wasm(&path, op, p.to.as_deref()) {
+                        tasks.spawn(broker.clone().check_actor(id));
+                    }
+                }
+            }
+            Some(_) = tasks.join_next() => {}
+        }
+    }
+}
+
+async fn sleep_until(t: Option<Instant>) {
+    match t {
+        Some(t) => tokio::time::sleep_until(t).await,
+        None => std::future::pending().await,
+    }
+}
+
+/// The namespace path (`/actor/...`) of a host path, unless it is outside
+/// `actor/` or a temporary file of the host.
+fn namespace(root: &Path, path: &Path) -> Option<String> {
+    let rel = path.strip_prefix(root).ok()?;
+    let mut out = String::from("/actor");
+    for part in rel.components() {
+        let name = part.as_os_str().to_str()?;
+        if name.starts_with(TMP_PREFIX) {
+            return None;
+        }
+        out.push('/');
+        out.push_str(name);
+    }
+    Some(out)
+}
+
+/// One change: namespace path, op, and the new path of a rename.
+type Change = (String, &'static str, Option<String>);
+
+/// The changes of one OS event. inotify reports a rename three times:
+/// `From` (the old name), `To` (the new name) and `Both`. `To` becomes the
+/// internal op `moved-in`, which `Both` replaces.
+fn changes(root: &Path, event: &notify::Event) -> Vec<Change> {
+    let path = |i: usize| event.paths.get(i).and_then(|p| namespace(root, p));
+    let one = |op: &'static str| -> Vec<Change> { path(0).map(|p| (p, op, None)).into_iter().collect() };
+    match event.kind {
+        EventKind::Create(_) => one("create"),
+        EventKind::Remove(_) => one("remove"),
+        EventKind::Modify(ModifyKind::Metadata(_)) => vec![],
+        EventKind::Modify(ModifyKind::Name(RenameMode::From)) => one("remove"),
+        EventKind::Modify(ModifyKind::Name(RenameMode::To)) => one("moved-in"),
+        EventKind::Modify(ModifyKind::Name(RenameMode::Both)) => match (path(0), path(1)) {
+            (Some(from), Some(to)) => vec![(from, "rename", Some(to))],
+            // Renamed to a temporary name: gone.
+            (Some(from), None) => vec![(from, "remove", None)],
+            // A temporary file renamed onto `to`: an atomic write.
+            (None, Some(to)) => vec![(to, "write", None)],
+            (None, None) => vec![],
+        },
+        EventKind::Modify(_) => one("write"),
+        _ => vec![],
+    }
+}
+
+/// Coalesce a change into `pending`. The last op wins, except that a write
+/// after a create stays a create. A rename replaces the `moved-in` of its
+/// new name.
+fn add(pending: &mut HashMap<String, Pending>, path: String, op: &'static str, to: Option<String>, now: Instant) {
+    if let Some(to) = &to
+        && pending.get(to).is_some_and(|p| p.op == "moved-in")
+    {
+        pending.remove(to);
+    }
+    match pending.get_mut(&path) {
+        Some(p) => {
+            if !(p.op == "create" && op == "write") {
+                p.op = op;
+                p.to = to;
+            }
+            p.last = now;
+        }
+        None => {
+            pending.insert(path, Pending { op, to, first: now, last: now });
+        }
+    }
+}
+
+/// The actor to check when `actor/<id>/actor.wasm` or the folder `actor/<id>`
+/// appeared or changed.
+fn new_wasm(path: &str, op: &str, to: Option<&str>) -> Option<String> {
+    let target = match op {
+        "remove" => return None,
+        "rename" => to?,
+        _ => path,
+    };
+    let rest = target.strip_prefix("/actor/")?;
+    let id = rest.strip_suffix("/actor.wasm").unwrap_or(rest);
+    (!id.is_empty() && !id.contains('/')).then(|| id.to_string())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn wasm_and_folder_changes_name_the_actor() {
+        assert_eq!(new_wasm("/actor/a/actor.wasm", "write", None).as_deref(), Some("a"));
+        assert_eq!(new_wasm("/actor/a", "create", None).as_deref(), Some("a"));
+        assert_eq!(new_wasm("/actor/x/y", "rename", Some("/actor/b/actor.wasm")).as_deref(), Some("b"));
+        assert_eq!(new_wasm("/actor/a/actor.wasm", "remove", None), None);
+        assert_eq!(new_wasm("/actor/a/versions/1/actor.wasm", "write", None), None);
+        assert_eq!(new_wasm("/actor/a/public/x", "write", None), None);
+        assert_eq!(new_wasm("/actor", "write", None), None);
+    }
+
+    #[test]
+    fn a_rename_is_one_change_and_temp_files_are_skipped() {
+        let root = Path::new("/w/actor");
+        let now = Instant::now();
+        let mut pending = HashMap::new();
+        let ev = |kind, paths: &[&str]| {
+            let mut e = notify::Event::new(kind);
+            for p in paths {
+                e = e.add_path(std::path::PathBuf::from(*p));
+            }
+            e
+        };
+        let name = |mode| EventKind::Modify(ModifyKind::Name(mode));
+        for e in [
+            ev(name(RenameMode::From), &["/w/actor/a/public/x"]),
+            ev(name(RenameMode::To), &["/w/actor/a/public/y"]),
+            ev(name(RenameMode::Both), &["/w/actor/a/public/x", "/w/actor/a/public/y"]),
+            ev(EventKind::Create(notify::event::CreateKind::File), &["/w/actor/a/.sausage-tmp-host-1-actor.json"]),
+            ev(name(RenameMode::Both), &["/w/actor/a/.sausage-tmp-host-1-actor.json", "/w/actor/a/actor.json"]),
+            ev(EventKind::Create(notify::event::CreateKind::File), &["/w/.lock"]),
+        ] {
+            for (path, op, to) in changes(root, &e) {
+                add(&mut pending, path, op, to, now);
+            }
+        }
+        let mut got: Vec<(String, &str, Option<String>)> =
+            pending.into_iter().map(|(k, p)| (k, p.op, p.to)).collect();
+        got.sort();
+        assert_eq!(
+            got,
+            vec![
+                ("/actor/a/actor.json".to_string(), "write", None),
+                ("/actor/a/public/x".to_string(), "rename", Some("/actor/a/public/y".to_string())),
+            ]
+        );
+    }
+}
@@ -1,11 +1,12 @@
 //! On-disk layout of a Sausage world. The folder is the source of truth:
-//! `sausage run` boots every actor that it finds in `actor/`.
+//! `sausage run` boots every actor that it finds in `actor/`, and then
+//! watches the folder (`watch.rs`).
 //!
 //! ```text
 //! <world>/
 //!   actor/<id>/
 //!     actor.wasm        the current version
-//!     actor.json        version, running, subscriptions, tapped (broker)
+//!     actor.json        version, running, subscriptions, tapped, watches (broker)
 //!     source/           source of the current version
 //!     public/           written by the actor, read-only for others
 //!     data/             state, writable only by the actor
@@ -46,6 +47,9 @@ pub struct ActorMeta {
    /// Whether this actor gets a copy of every routed message (`tap`).
    #[serde(default)]
    pub tapped: bool,
+    /// Namespace paths whose changes this actor gets as `fs_changed` (`watch`).
+    #[serde(default)]
+    pub watches: BTreeSet<String>,
 }

 fn yes() -> bool {
@@ -38,6 +38,10 @@ fn main() {
        panic!("version 2 crashes on startup");
    }

+    if actor == "stopped-upgrade" {
+        std::fs::write("/data/main-version", &version).unwrap();
+    }
+
    let mut open = HashMap::<String, std::fs::File>::new();
    // Files that this actor serves at /actor/<id>/fs/ (path -> text).
    let mut served = std::collections::BTreeMap::<String, String>::new();
@@ -0,0 +1,85 @@
+//! Tests of the `sausage` command line.
+
+use std::path::PathBuf;
+use std::process::{Command, Stdio};
+use std::time::{SystemTime, UNIX_EPOCH};
+
+fn empty_dir(label: &str) -> PathBuf {
+    let nonce = SystemTime::now().duration_since(UNIX_EPOCH).unwrap().as_nanos();
+    let dir = PathBuf::from(env!("CARGO_MANIFEST_DIR"))
+        .parent()
+        .unwrap()
+        .join("target/test-cli")
+        .join(format!("{label}-{}-{nonce}", std::process::id()));
+    std::fs::create_dir_all(&dir).unwrap();
+    dir
+}
+
+/// Bare `sausage` (as in `nix run <flake>`) without a world starts `init`
+/// for ./world instead of printing a usage error.
+#[test]
+fn bare_command_without_a_world_starts_init() {
+    let dir = empty_dir("bare");
+    let output = Command::new(env!("CARGO_BIN_EXE_sausage"))
+        .current_dir(&dir)
+        .env_remove("SAUSAGE_API")
+        .env_remove("SAUSAGE_BASE_URL")
+        .env_remove("SAUSAGE_MODEL")
+        .env_remove("SAUSAGE_API_TOKEN")
+        .stdin(Stdio::null())
+        .output()
+        .unwrap();
+    let stdout = String::from_utf8_lossy(&output.stdout);
+    let stderr = String::from_utf8_lossy(&output.stderr);
+    // With no answers on stdin, init stops at the first required value.
+    assert!(!output.status.success(), "stdout:\n{stdout}\nstderr:\n{stderr}");
+    assert!(stdout.contains("No world in ./world yet"), "stdout:\n{stdout}\nstderr:\n{stderr}");
+    assert!(stderr.contains("SAUSAGE_BASE_URL"), "stdout:\n{stdout}\nstderr:\n{stderr}");
+    assert!(!stderr.contains("Usage:"), "stderr:\n{stderr}");
+    // It asks before it creates anything.
+    assert!(!dir.join("world").exists());
+}
+
+/// Start a headless host on a fresh world with stdin at EOF; returns once it
+/// printed its ready line (printed after the signal handlers are installed).
+fn start_headless(label: &str) -> (std::process::Child, std::io::Lines<std::io::BufReader<std::process::ChildStdout>>) {
+    use std::io::{BufRead, BufReader};
+    let dir = empty_dir(label);
+    let world = dir.join("world");
+    sausage::World::create(&world).unwrap();
+    let mut child = Command::new(env!("CARGO_BIN_EXE_sausage"))
+        .args(["run", "--headless"])
+        .arg(&world)
+        .stdin(Stdio::null())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::null())
+        .spawn()
+        .unwrap();
+    let mut lines = BufReader::new(child.stdout.take().unwrap()).lines();
+    let line = lines.next().expect("host exited before it started").unwrap();
+    assert!(line.contains("headless"), "unexpected first line: {line}");
+    (child, lines)
+}
+
+/// With stdin at EOF (as soon as the host starts), a headless host stays up.
+#[test]
+fn headless_ignores_stdin_eof() {
+    let (mut child, _lines) = start_headless("headless-eof");
+    std::thread::sleep(std::time::Duration::from_millis(500));
+    let alive = child.try_wait().unwrap().is_none();
+    let _ = child.kill();
+    let _ = child.wait();
+    assert!(alive, "headless host exited on stdin EOF");
+}
+
+/// SIGTERM right after the ready line stops the actors and exits with 0.
+#[test]
+fn headless_stops_cleanly_on_sigterm_right_after_ready() {
+    let (mut child, lines) = start_headless("headless-term");
+    let pid = child.id().to_string();
+    assert!(Command::new("kill").args(["-TERM", &pid]).status().unwrap().success());
+    let rest: Vec<String> = lines.map(|l| l.unwrap()).collect();
+    let status = child.wait().unwrap();
+    assert!(status.success(), "exit status {status:?}, output {rest:?}");
+    assert!(rest.iter().any(|l| l.contains("Stopping actors")), "output {rest:?}");
+}
@@ -0,0 +1,6 @@
+[package]
+name = "cargo-test-fail"
+version = "0.1.0"
+edition = "2024"
+
+[workspace]
@@ -0,0 +1,9 @@
+fn main() {}
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn deliberate_unit_test_failure() {
+        assert_eq!(2 + 2, 5);
+    }
+}
@@ -1 +1,9 @@
 fn main() {}
+
+#[cfg(test)]
+mod tests {
+    #[test]
+    fn unit_tests_run_on_the_host_target() {
+        assert_eq!(2 + 2, 4);
+    }
+}
@@ -108,15 +108,15 @@ async fn routes_replies_publishes_and_reports_undeliverable() {
 async fn broker_does_not_answer_messages_with_reply_to() {
    let world = create_world("broker-reply-to");
    let (host, mut mailbox) = start(&world).await;
-    let scanned = request(&host, &mut mailbox, json!({"type": "scan"})).await;
-    assert_eq!(scanned.body["type"], "scanned");
+    let untapped = request(&host, &mut mailbox, json!({"type": "untap"})).await;
+    assert_eq!(untapped.body["type"], "untapped");

    host.broker.send(
        "console",
        Some("broker"),
        None,
        Some("upstream-message"),
-        json!({"type": "scan"}),
+        json!({"type": "untap"}),
    );
    assert!(
        tokio::time::timeout(Duration::from_secs(1), mailbox.recv())
@@ -436,6 +436,113 @@ async fn upgrades_test_candidates_keep_old_versions_and_queue_messages_in_order(
    host.stop().await;
 }

+#[tokio::test]
+async fn stopped_preserving_upgrade_installs_without_execution_and_keeps_all_mail() {
+    let world = create_world("stopped-upgrade");
+    let wasm = toolbox_wasm().clone();
+    let (host, mut mailbox) = start(&world).await;
+    let source = passing_source(&world, "initial");
+    spawn_type(&host, &mut mailbox, "stopped-upgrade", &wasm, &source).await;
+    ask_type(&host, &mut mailbox, "stopped-upgrade", json!({"type": "version"}), "version").await;
+    let stopped = request(&host, &mut mailbox, json!({"type": "stop", "actor": "stopped-upgrade"})).await;
+    assert_eq!(stopped.body["type"], "stopped");
+
+    // A distinct valid component: append an empty custom section named x.
+    let mut bytes = std::fs::read(&wasm).unwrap();
+    bytes.extend_from_slice(&[0, 2, 1, b'x']);
+    let new_wasm = world.public_dir("console").join("new.wasm");
+    std::fs::write(&new_wasm, &bytes).unwrap();
+    // This would fail if candidate main/tests were executed.
+    let unrun = source_with_test(&world, "unrun", json!({"steps": [{
+        "send": {"type": "echo"}, "expect": {"type": "must-not-run"}
+    }]}));
+    std::fs::write(unrun.join("new-source"), "version two").unwrap();
+    std::fs::write(world.public_dir("stopped-upgrade").join("keep"), "public state").unwrap();
+
+    let mut ids = Vec::new();
+    for n in 0..sausage::broker::MAILBOX_SIZE {
+        ids.push(host.broker.send("console", Some("stopped-upgrade"), None, None, json!({"type": "ordered", "n": n})));
+    }
+    let upgrade_id = host.broker.send("console", Some("broker"), None, None, json!({
+        "type": "upgrade", "actor": "stopped-upgrade", "wasm": new_wasm.display().to_string(),
+        "source": unrun.display().to_string(), "preserve_stopped": true
+    }));
+    // Do not silently discard unexpected actor mail while awaiting the reply.
+    let upgraded = receive_message(&mut mailbox, Duration::from_secs(10)).await;
+    assert_eq!(upgraded.reply_to.as_deref(), Some(upgrade_id.as_str()));
+    assert_eq!(upgraded.body["type"], "upgraded");
+    assert_eq!(upgraded.body["version"], 2);
+    assert_eq!(upgraded.body["previous"], 1);
+    assert_eq!(upgraded.body["tests"], json!({"skipped": "preserve_stopped"}));
+    let meta = world.read_meta("stopped-upgrade");
+    assert_eq!(meta.version, Some(2));
+    assert!(!meta.running);
+    assert_eq!(std::fs::read(world.wasm_path("stopped-upgrade")).unwrap(), bytes);
+    assert_eq!(std::fs::read(world.version_dir("stopped-upgrade", 2).join("actor.wasm")).unwrap(), bytes);
+    assert_ne!(std::fs::read(world.version_dir("stopped-upgrade", 1).join("actor.wasm")).unwrap(), bytes);
+    assert_eq!(std::fs::read_to_string(world.source_dir("stopped-upgrade").join("new-source")).unwrap(), "version two");
+    assert_eq!(std::fs::read_to_string(world.data_dir("stopped-upgrade").join("main-version")).unwrap(), "1");
+    assert_eq!(std::fs::read_to_string(world.public_dir("stopped-upgrade").join("keep")).unwrap(), "public state");
+    assert!(mailbox.try_recv().is_err());
+
+    // The full mailbox is still full: nothing was consumed or added by the upgrade.
+    let overflow_id = host.broker.send("console", Some("stopped-upgrade"), None, None, json!({"type": "echo"}));
+    let overflow = receive_message(&mut mailbox, Duration::from_secs(10)).await;
+    assert_eq!(overflow.reply_to.as_deref(), Some(overflow_id.as_str()));
+    assert_eq!(overflow.body["type"], "undeliverable");
+    assert!(overflow.body["reason"].as_str().unwrap().contains("full"));
+
+    // Only an explicit start executes v2. Every original envelope survives in order.
+    let start_id = host.broker.send("console", Some("broker"), None, None, json!({"type": "start", "actor": "stopped-upgrade"}));
+    let mut started = false;
+    let mut count = 0;
+    tokio::time::timeout(Duration::from_secs(10), async {
+        while !started || count < ids.len() {
+            let reply = mailbox.recv().await.unwrap();
+            if reply.reply_to.as_deref() == Some(start_id.as_str()) {
+                assert_eq!(reply.body["type"], "started");
+                started = true;
+            } else {
+                assert_eq!(reply.from, "stopped-upgrade");
+                assert_eq!(reply.reply_to.as_deref(), Some(ids[count].as_str()));
+                assert_eq!(reply.body, json!({"type": "ordered", "n": count, "version": "2"}));
+                count += 1;
+            }
+        }
+    }).await.expect("queued mail was not delivered after start");
+    assert_eq!(std::fs::read_to_string(world.data_dir("stopped-upgrade").join("main-version")).unwrap(), "2");
+    host.stop().await;
+}
+
+#[tokio::test]
+async fn stopped_preserving_option_keeps_running_upgrades_tested_and_default_upgrades_start() {
+    let world = create_world("preserve-running");
+    let wasm = toolbox_wasm().clone();
+    let (host, mut mailbox) = start(&world).await;
+    let passing = passing_source(&world, "passing");
+    let failing = source_with_test(&world, "failing", json!({"steps": [{
+        "send": {"type": "echo"}, "expect": {"type": "wrong"}
+    }]}));
+    spawn_type(&host, &mut mailbox, "running", &wasm, &passing).await;
+    for (option, source, expected) in [(json!("true"), &passing, "error"), (json!(true), &failing, "error"), (json!(true), &passing, "upgraded")] {
+        let reply = request(&host, &mut mailbox, json!({
+            "type": "upgrade", "actor": "running", "wasm": wasm.display().to_string(),
+            "source": source.display().to_string(), "preserve_stopped": option
+        })).await;
+        assert_eq!(reply.body["type"], expected);
+    }
+    assert!(world.read_meta("running").running);
+    let version = ask_type(&host, &mut mailbox, "running", json!({"type": "version"}), "version").await;
+    assert_eq!(version.body["version"], "2");
+    request(&host, &mut mailbox, json!({"type": "stop", "actor": "running"})).await;
+    let default = upgrade_actor(&host, &mut mailbox, "running", &wasm, &passing).await;
+    assert_eq!(default.body["type"], "upgraded");
+    assert!(world.read_meta("running").running);
+    let version = ask_type(&host, &mut mailbox, "running", json!({"type": "version"}), "version").await;
+    assert_eq!(version.body["version"], "3");
+    host.stop().await;
+}
+
 #[tokio::test]
 async fn rolls_back_a_crash_loop_to_the_previous_version_without_restoring_data() {
    let world = create_world("crash-loop");
@@ -611,14 +718,18 @@ async fn unknown_broker_request_lists_the_known_ones() {
    let unknown = request(&host, &mut mailbox, json!({"type": "frobnicate"})).await;
    assert_eq!(unknown.body["type"], "error");
    let reason = unknown.body["reason"].as_str().unwrap();
-    assert!(reason.contains("remove") && reason.contains("tap"), "{reason}");
+    assert!(reason.contains("remove") && reason.contains("tap") && reason.contains("watch"), "{reason}");
+    // `scan` is retired: the host watches the world folder instead.
+    let scan = request(&host, &mut mailbox, json!({"type": "scan"})).await;
+    assert_eq!(scan.body["type"], "error");
    host.stop().await;
 }

 #[tokio::test]
-async fn cargo_actor_builds_valid_crates_and_reports_compiler_errors() {
+async fn cargo_actor_runs_unit_tests_by_default_and_can_skip_them() {
    let world = create_world("cargo-actor");
    let (host, mut mailbox) = start(&world).await;
+
    let valid_source = cargo_source("cargo-valid");
    let valid_id = host.broker.send(
        "console",
@@ -627,12 +738,42 @@ async fn cargo_actor_builds_valid_crates_and_reports_compiler_errors() {
        None,
        json!({"type": "build", "source": valid_source.display().to_string()}),
    );
-    let built = receive_reply(&mut mailbox, &valid_id, Duration::from_secs(120)).await;
-    assert_eq!(built.body["type"], "built");
+    let built = receive_reply(&mut mailbox, &valid_id, Duration::from_secs(180)).await;
+    assert_eq!(built.body["type"], "built", "{built:?}");
+    assert_eq!(built.body["tests"]["passed"], 1);
+    assert_eq!(built.body["tests"]["failed"], 0);
+    assert!(built.body["tests"]["output"].as_str().unwrap().contains("unit_tests_run_on_the_host_target"));
    let wasm = built.body["wasm"].as_str().unwrap();
    assert!(wasm.ends_with(".wasm"));
    assert!(host.world.resolve("console", wasm).unwrap().is_file());

+    let failing_test_source = cargo_source("cargo-test-fail");
+    let failed_test_id = host.broker.send(
+        "console",
+        Some("cargo"),
+        None,
+        None,
+        json!({"type": "build", "source": failing_test_source.display().to_string()}),
+    );
+    let failed_test = receive_reply(&mut mailbox, &failed_test_id, Duration::from_secs(120)).await;
+    assert_eq!(failed_test.body["type"], "build_failed", "{failed_test:?}");
+    assert_eq!(failed_test.body["tests"]["passed"], 0);
+    assert_eq!(failed_test.body["tests"]["failed"], 1);
+    assert!(failed_test.body["tests"]["output"].as_str().unwrap().contains("deliberate_unit_test_failure"));
+
+    let skipped_id = host.broker.send(
+        "console",
+        Some("cargo"),
+        None,
+        None,
+        json!({"type": "build", "source": failing_test_source.display().to_string(), "skip_tests": true}),
+    );
+    let skipped = receive_reply(&mut mailbox, &skipped_id, Duration::from_secs(120)).await;
+    assert_eq!(skipped.body["type"], "built", "{skipped:?}");
+    assert_eq!(skipped.body["tests"]["passed"], 0);
+    assert_eq!(skipped.body["tests"]["failed"], 0);
+    assert_eq!(skipped.body["tests"]["output"], "skipped (skip_tests=true)");
+
    let invalid_source = cargo_source("cargo-fail");
    let invalid_id = host.broker.send(
        "console",
@@ -641,14 +782,10 @@ async fn cargo_actor_builds_valid_crates_and_reports_compiler_errors() {
        None,
        json!({"type": "build", "source": invalid_source.display().to_string()}),
    );
-    let failed = receive_reply(&mut mailbox, &invalid_id, Duration::from_secs(120)).await;
-    assert_eq!(failed.body["type"], "build_failed");
-    assert!(
-        failed.body["errors"]
-            .as_str()
-            .unwrap()
-            .contains("missing_function")
-    );
+    let invalid = receive_reply(&mut mailbox, &invalid_id, Duration::from_secs(120)).await;
+    assert_eq!(invalid.body["type"], "build_failed", "{invalid:?}");
+    assert_eq!(invalid.body["tests"]["failed"], 1);
+    assert!(invalid.body["tests"]["output"].as_str().unwrap().contains("missing_function"));

    host.stop().await;
 }
@@ -974,3 +1111,147 @@ async fn a_full_tapper_mailbox_drops_copies_without_blocking() {
    assert_eq!(ghost.try_recv().unwrap().body["type"], "tapped");
    host.stop().await;
 }
+
+/// Ask `actor` again and again (it may not exist yet) until `want` holds for
+/// the answer.
+async fn ask_until(
+    host: &Host,
+    mailbox: &mut mpsc::Receiver<Envelope>,
+    actor: &str,
+    body: Value,
+    want: impl Fn(&Value) -> bool,
+) -> Envelope {
+    tokio::time::timeout(Duration::from_secs(30), async {
+        loop {
+            let reply = ask(host, mailbox, actor, body.clone()).await;
+            if want(&reply.body) {
+                return reply;
+            }
+            tokio::time::sleep(Duration::from_millis(100)).await;
+        }
+    })
+    .await
+    .unwrap_or_else(|_| panic!("{actor} never gave the wanted answer"))
+}
+
+/// The `fs_changed` bodies among `messages` (all from `fs`).
+fn fs_events(messages: Vec<Envelope>) -> Vec<Value> {
+    messages
+        .into_iter()
+        .filter(|m| m.body["type"] == "fs_changed")
+        .map(|m| {
+            assert_eq!(m.from, "fs");
+            m.body
+        })
+        .collect()
+}
+
+#[tokio::test]
+async fn a_new_actor_wasm_on_disk_becomes_an_actor() {
+    let world = create_world("watch-spawn");
+    let (host, mut mailbox) = start(&world).await;
+    // Any writer counts: here the test itself, like a shell on the host.
+    let tests = world.source_dir("dropped").join("tests");
+    std::fs::create_dir_all(&tests).unwrap();
+    let test = json!({"steps": [{"send": {"type": "echo", "value": "candidate"}, "expect": {"type": "echo", "value": "candidate"}}]});
+    std::fs::write(tests.join("candidate.json"), test.to_string()).unwrap();
+    std::fs::copy(toolbox_wasm(), world.wasm_path("dropped")).unwrap();
+
+    let echo = json!({"type": "echo", "value": "hi"});
+    let reply = ask_until(&host, &mut mailbox, "dropped", echo, |b| b["type"] == "echo").await;
+    assert_eq!(reply.from, "dropped");
+    assert_eq!(world.read_meta("dropped").version, Some(1));
+    let meta = std::fs::read_to_string(world.version_dir("dropped", 1).join("meta.json")).unwrap();
+    assert_eq!(serde_json::from_str::<Value>(&meta).unwrap()["requested_by"], "scan");
+    host.stop().await;
+}
+
+#[tokio::test]
+async fn a_changed_actor_wasm_on_disk_is_an_upgrade() {
+    let world = create_world("watch-upgrade");
+    let wasm = toolbox_wasm().clone();
+    let (host, mut mailbox) = start(&world).await;
+    let source = passing_source(&world, "source");
+    spawn_type(&host, &mut mailbox, "changing", &wasm, &source).await;
+
+    // The same program plus a custom section: a different file.
+    let mut changed = std::fs::read(&wasm).unwrap();
+    changed.extend_from_slice(&[0, 4, 1, b'x', b'v', b'2']);
+    std::fs::write(world.wasm_path("changing"), changed).unwrap();
+
+    let version = json!({"type": "version"});
+    ask_until(&host, &mut mailbox, "changing", version, |b| b["version"] == "2").await;
+    assert_eq!(world.read_meta("changing").version, Some(2));
+    assert_eq!(world.versions("changing"), vec![1, 2]);
+    host.stop().await;
+}
+
+#[tokio::test]
+async fn watchers_get_one_event_per_change_from_any_writer() {
+    let (world, host, mut mailbox) = filesystem_world("watch-events").await;
+    let watch = json!({"type": "watch", "path": "/actor/writer/public/"});
+    let watching = request(&host, &mut mailbox, watch).await;
+    assert_eq!(watching.body, json!({"type": "watching", "path": "/actor/writer/public"}));
+    assert!(world.read_meta("console").watches.contains("/actor/writer/public"));
+    drain(&mut mailbox, Duration::from_millis(300)).await;
+
+    let quiet = Duration::from_millis(600);
+    let event = |path: &str, op: &str| json!({"type": "fs_changed", "path": format!("/actor/writer/public/{path}"), "op": op});
+
+    // An actor creates a file: one create, not create + write.
+    let write = json!({"type": "write", "path": "/public/a.txt", "text": "a"});
+    ask_type(&host, &mut mailbox, "writer", write, "ok").await;
+    assert_eq!(fs_events(drain(&mut mailbox, quiet).await), vec![event("a.txt", "create")]);
+
+    // The host (a shell, git, ...) changes it.
+    std::fs::write(world.public_dir("writer").join("a.txt"), "b").unwrap();
+    assert_eq!(fs_events(drain(&mut mailbox, quiet).await), vec![event("a.txt", "write")]);
+
+    // A rename is one event with `to`.
+    let rename = json!({"type": "rename", "path": "/public/a.txt", "to": "/public/b.txt"});
+    ask_type(&host, &mut mailbox, "writer", rename, "ok").await;
+    let mut renamed = event("a.txt", "rename");
+    renamed["to"] = json!("/actor/writer/public/b.txt");
+    assert_eq!(fs_events(drain(&mut mailbox, quiet).await), vec![renamed]);
+
+    std::fs::remove_file(world.public_dir("writer").join("b.txt")).unwrap();
+    assert_eq!(fs_events(drain(&mut mailbox, quiet).await), vec![event("b.txt", "remove")]);
+
+    // Nothing outside the watched path.
+    let outside = json!({"type": "write", "path": "/data/x", "text": "x"});
+    ask_type(&host, &mut mailbox, "writer", outside, "ok").await;
+    assert_eq!(fs_events(drain(&mut mailbox, quiet).await), Vec::<Value>::new());
+
+    let unwatch = json!({"type": "unwatch", "path": "/actor/writer/public"});
+    let unwatched = request(&host, &mut mailbox, unwatch).await;
+    assert_eq!(unwatched.body, json!({"type": "unwatched", "path": "/actor/writer/public"}));
+    assert!(world.read_meta("console").watches.is_empty());
+    let write = json!({"type": "write", "path": "/public/c.txt", "text": "c"});
+    ask_type(&host, &mut mailbox, "writer", write, "ok").await;
+    assert_eq!(fs_events(drain(&mut mailbox, quiet).await), Vec::<Value>::new());
+    host.stop().await;
+}
+
+#[tokio::test]
+async fn a_watcher_overflow_sends_one_rescan_and_runs_a_full_pass() {
+    let world = create_world("watch-overflow");
+    let wasm = toolbox_wasm().clone();
+    let (host, mut mailbox) = start(&world).await;
+    let mut bystander = host.mailbox("bystander");
+    let watch = json!({"type": "watch", "path": "/actor/elsewhere"});
+    assert_eq!(request(&host, &mut mailbox, watch).await.body["type"], "watching");
+
+    // A dropped actor that the full pass finds (before the watcher's quiet time).
+    let tests = world.source_dir("planted").join("tests");
+    std::fs::create_dir_all(&tests).unwrap();
+    let test = json!({"steps": [{"send": {"type": "echo", "value": "candidate"}, "expect": {"type": "echo", "value": "candidate"}}]});
+    std::fs::write(tests.join("candidate.json"), test.to_string()).unwrap();
+    std::fs::copy(&wasm, world.wasm_path("planted")).unwrap();
+    let report = host.broker.clone().overflowed().await;
+    assert!(report.iter().any(|r| r["type"] == "spawned" && r["actor"] == "planted"), "{report:?}");
+
+    let rescan = json!({"type": "fs_changed", "path": "/actor", "op": "rescan"});
+    assert_eq!(fs_events(drain(&mut mailbox, Duration::from_millis(600)).await), vec![rescan]);
+    assert!(bystander.try_recv().is_err(), "an actor without watches got the rescan");
+    host.stop().await;
+}
@@ -1,5 +1,20 @@
 {
  "nodes": {
+    "crane": {
+      "locked": {
+        "lastModified": 1791754810,
+        "narHash": "sha256-cN0Mgk+sQtCJcGWs9H4M0CnLRAuC4I8aCwiBTcBJs6E=",
+        "owner": "ipetkov",
+        "repo": "crane",
+        "rev": "bd63bf1d0a0f6dcacd7549c714a671ee63b0db6a",
+        "type": "github"
+      },
+      "original": {
+        "owner": "ipetkov",
+        "repo": "crane",
+        "type": "github"
+      }
+    },
    "flake-utils": {
      "inputs": {
        "systems": "systems"
@@ -36,6 +51,7 @@
    },
    "root": {
      "inputs": {
+        "crane": "crane",
        "flake-utils": "flake-utils",
        "nixpkgs": "nixpkgs",
        "rust-overlay": "rust-overlay"
@@ -6,9 +6,10 @@
    flake-utils.url = "github:numtide/flake-utils";
    rust-overlay.url = "github:oxalica/rust-overlay";
    rust-overlay.inputs.nixpkgs.follows = "nixpkgs";
+    crane.url = "github:ipetkov/crane";
  };

-  outputs = { nixpkgs, flake-utils, rust-overlay, ... }:
+  outputs = { nixpkgs, flake-utils, rust-overlay, crane, ... }:
    flake-utils.lib.eachSystem flake-utils.lib.allSystems (system:
      let
        pkgs = import nixpkgs {
@@ -20,34 +21,53 @@
          targets = [ "wasm32-wasip2" ];
          extensions = [ "rust-src" ];
        };
-        rustPlatform = pkgs.makeRustPlatform { cargo = rust; rustc = rust; };
-        sausage = rustPlatform.buildRustPackage {
+        # The runtime cargo actor needs Cargo, rustc (with wasm32-wasip2), and
+        # the native linker, even when the host itself is started with a bare
+        # PATH. Keep this same runtime path in the packaged host and exercise
+        # the installed binary below with PATH deliberately empty.
+        runtimePath = pkgs.lib.makeBinPath [ rust pkgs.stdenv.cc ];
+        craneLib = (crane.mkLib pkgs).overrideToolchain rust;
+        common = {
          pname = "sausage";
          version = "0.1.0";
          src = ./.;
-          cargoLock.lockFile = ./Cargo.lock;
-          cargoBuildFlags = [ "-p" "sausage" ];
+        };
+        # The test packages for `checks.tests`.
+        testPackages = "-p sausage -p sausage-actor -p agent";
+        cargoArtifacts = craneLib.buildDepsOnly (common // {
+          doCheck = false;
+          # The dependencies of the host binary and of the tests.
+          buildPhaseCargoCommand = ''
+            cargoWithProfile build --locked -p sausage
+            cargoWithProfile test --locked --no-run ${testPackages}
+          '';
+        });
+        sausage = craneLib.buildPackage (common // {
+          inherit cargoArtifacts;
+          doNotRemoveReferencesToRustToolchain = true;
+          cargoExtraArgs = "--locked -p sausage";
          doCheck = false;
          nativeBuildInputs = [ pkgs.makeWrapper ];
          # The `cargo` actor runs cargo at runtime. Build scripts need a C linker.
          postInstall = ''
-            wrapProgram $out/bin/sausage --prefix PATH : ${pkgs.lib.makeBinPath [ rust pkgs.stdenv.cc ]}
+            wrapProgram $out/bin/sausage --prefix PATH : ${runtimePath}
          '';
          meta.mainProgram = "sausage";
-        };
+        });
        # `nix run .#loop -- <world>` runs the newest `sausage run <world>`
-        # from the Gitea repo, again and again: after a crash or a broker
+        # (without a folder: bare `sausage`, i.e. ./world, created if needed)
+        # from the public repo, again and again: after a crash or a broker
        # `exit` request it refreshes the flake and starts the host again,
        # so the world updates itself. `/quit` (exit code 0) ends the loop.
-        # `--internal` (or `.#loop-internal`) uses Gitea's SSH port on the LAN.
-        publicUrl = "git+ssh://containeruser@gitea.surma.technology:2222/surma/sausage";
-        internalUrl = "git+ssh://containeruser@10.0.0.2:2222/surma/sausage";
-        loop = name: url: pkgs.writeShellScriptBin name ''
-          url="${url}"
-          if [ "''${1:-}" = "--internal" ]; then url="${internalUrl}"; shift; fi
-          if [ $# -eq 0 ]; then set -- world; fi
+        # SAUSAGE_FLAKE overrides where it fetches from (any flake URL,
+        # e.g. a fork or a local checkout).
+        repoUrl = "git+https://gitea.surma.technology/surma/sausage";
+        loop = pkgs.writeShellScriptBin "sausage-loop" ''
+          url="''${SAUSAGE_FLAKE:-${repoUrl}}"
+          # No folder: bare `sausage` (creates ./world first if needed).
+          if [ $# -gt 0 ]; then set -- run "$@"; fi
          while true; do
-            nix run --refresh "$url#default" -- run "$@"
+            nix run --refresh "$url#default" -- "$@"
            code=$?
            if [ "$code" -eq 0 ]; then exit 0; fi
            echo "sausage-loop: sausage exited with $code, starting the newest version in 2 s …" >&2
@@ -61,19 +81,85 @@
        # `nix flake check` runs every test of the host, the SDK, and the
        # agent. The tests build WASM actors with cargo, offline, from the
        # vendored crates of Cargo.lock.
-        checks.tests = sausage.overrideAttrs (old: {
-          pname = "sausage-tests";
-          doCheck = true;
-          cargoTestFlags = [ "-p" "sausage" "-p" "sausage-actor" "-p" "agent" ];
+        # The binary the host loop runs (`nix run`), built from the same
+        # cached dependencies as the tests.
+        checks.package = sausage;
+        checks.tests = craneLib.cargoTest (common // {
+          inherit cargoArtifacts;
+          cargoExtraArgs = "--locked ${testPackages}";
+          # Nothing builds on the test run's target folder: don't store it.
+          doInstallCargoArtifacts = false;
          preCheck = ''
            # Its own lock file pins crates that the vendored set may lack.
            rm -f cli/tests/actors/toolbox/Cargo.lock
          '';
        });
+        # Exercise the installed host exactly as the loop launches it: the
+        # caller has no Cargo on PATH, so only the packaged runtime toolchain
+        # can run the native tests and the wasm32-wasip2 build.
+        checks.runtime-toolchain = pkgs.runCommand "sausage-runtime-toolchain" { } ''
+          set -eu
+          root="$TMPDIR/runtime-toolchain"
+          world="$root/world"
+          emptyPath="$root/empty-bin"
+          mkdir -p "$emptyPath" "$root/home" "$root/cargo-home"
+          for actor in broker console cargo fs; do
+            mkdir -p "$world/actor/$actor/public" "$world/actor/$actor/data"
+          done
+          source="$world/actor/console/public/probe"
+          mkdir -p "$source/src" "$source/tests"
+          cat > "$source/Cargo.toml" <<'TOML'
+          [package]
+          name = "runtime-probe"
+          version = "0.1.0"
+          edition = "2024"
+          TOML
+          printf 'fn main() {}\n' > "$source/src/main.rs"
+          cat > "$source/tests/probe.rs" <<'RUST'
+          #[test]
+          fn native_test_harness_runs() {
+              assert_eq!(2 + 2, 4);
+          }
+          RUST
+          echo "runtime-check: run packaged host with an empty PATH"
+          coproc host {
+            ${pkgs.coreutils}/bin/env -i PATH="$emptyPath" HOME="$root/home" TMPDIR="$root" \
+              CARGO_HOME="$root/cargo-home" ${sausage}/bin/sausage run "$world"
+          }
+          hostInput=''${host[1]}
+          hostOutput=''${host[0]}
+          hostPid=$host_PID
+          trap 'kill "$hostPid" 2>/dev/null || true' EXIT
+          printf '%s\n' '@cargo {"type":"build","source":"/actor/console/public/probe"}' >&"$hostInput"
+          built=0
+          while IFS= read -r -t 300 line <&"$hostOutput"; do
+            echo "$line"
+            case "$line" in
+              *'"type":"build_failed"'*)
+                echo "packaged cargo actor build failed with an empty caller PATH" >&2
+                exit 1
+                ;;
+              *'"type":"built"'*)
+                case "$line" in
+                  *'"passed":1'*) built=1 ;;
+                  *) echo "native test did not pass" >&2; exit 1 ;;
+                esac
+                break
+                ;;
+            esac
+          done
+          if [ "$built" -ne 1 ]; then
+            echo "packaged host did not build the probe" >&2
+            exit 1
+          fi
+          printf '%s\n' '/quit' >&"$hostInput"
+          wait "$hostPid"
+          mkdir -p "$out"
+          touch "$out/passed"
+        '';

        apps.default = { type = "app"; program = "${sausage}/bin/sausage"; };
-        apps.loop = { type = "app"; program = "${loop "sausage-loop" publicUrl}/bin/sausage-loop"; };
-        apps.loop-internal = { type = "app"; program = "${loop "sausage-loop-internal" internalUrl}/bin/sausage-loop-internal"; };
+        apps.loop = { type = "app"; program = "${loop}/bin/sausage-loop"; };

        devShells.default = pkgs.mkShell {
          packages = [ rust pkgs.wasmtime pkgs.wasm-tools ];
Author	SHA1	Message	Date
Sausage Agent	0bd046ed9f	broker: add stopped-preserving upgrade without actor execution ci / check (push) Failing after 1m47s Details	2026-10-22 06:49:58 +00:00
Sausage Agent	a72e42a8c9	headless: register signal handlers synchronously, split tests, drop deployment examples and wording ci / check (pull_request) Successful in 1m36s Details ci / check (push) Successful in 11s Details	2026-10-18 15:31:41 +00:00
Sausage Agent	01a821ab72	Add --headless mode (no stdin read, runs until SIGTERM/exit) + systemd/NixOS examples ci / check (push) Successful in 2m44s Details	2026-10-18 15:24:20 +00:00
Sausage Agent	c204e1ef0c	docs: clarify actor data is world-readable ci / check (pull_request) Successful in 12s Details ci / check (push) Successful in 6s Details	2026-10-17 17:22:50 +00:00
Sausage Agent	6923ca3e60	preserve the packaged runtime Rust toolchain ci / check (push) Successful in 6s Details	2026-10-17 16:36:08 +00:00
Sausage Agent	c504532513	drive packaged host runtime check over pipes ci / check (push) Failing after 1m16s Details	2026-10-17 16:33:31 +00:00
Sausage Agent	eb35a79698	trace the packaged host runtime check ci / check (push) Failing after 1m47s Details	2026-10-17 16:29:02 +00:00
Sausage Agent	bfcd71e2f4	test packaged host cargo runtime with an empty PATH ci / check (push) Failing after 1m33s Details	2026-10-17 16:25:14 +00:00
agent-pool-12	172a776eec	cargo: annotate test summary count parsing ci / check (push) Successful in 5s Details	2026-10-17 16:00:54 +00:00
agent-pool-12	6a9d35ceb8	cargo: preserve build failures in result wrapper ci / check (push) Failing after 6m15s Details	2026-10-17 15:52:37 +00:00
agent-pool-12	22590c608d	cargo: run host unit tests before builds ci / check (push) Failing after 8m51s Details	2026-10-17 15:50:09 +00:00
Sausage Agent	3bc462ffd2	agent seed: thinking levels low\|medium\|high\|xhigh\|max\|off on both APIs (card #128 ) ci / check (push) Successful in 5s Details New config field `thinking` (OpenAI: reasoning_effort still wins): Anthropic adaptive thinking + output_config.effort (thinking blocks kept and sent back), OpenAI reasoning.effort. A 400 that rejects the effort is retried with the nearest accepted level (from the error's list, lower on a tie), logged, never dropped to off.	2026-10-17 15:28:44 +00:00
Sausage Agent	f008d0cc60	nix: lock crane; don't store the test run's target folder ci / check (push) Successful in 39s Details	2026-10-17 08:02:08 +00:00
Sausage Agent	81daddfd70	nix: build with crane, dependencies cached separately ci / check (push) Successful in 10m50s Details buildDepsOnly builds the dependencies of the host and the tests once (keyed only on Cargo.toml/Cargo.lock); packages.default and checks.tests reuse them, so a change to our crates no longer recompiles every crate. checks.package builds the binary the host loop runs.	2026-10-17 07:49:54 +00:00
agent-pool-9	e8df22fe06	host tests: reuse the existing drain helper ci / check (push) Successful in 8s Details	2026-10-17 07:23:30 +00:00
agent-pool-9	0995824c01	Watch the world folder: auto spawn/upgrade, broker watch/unwatch, retire scan ci / check (push) Failing after 7m36s Details The host watches the whole world folder recursively with the notify crate (any writer counts, .sausage-tmp-* skipped, ~100 ms coalescing per path). A new or changed actor/<id>/actor.wasm (or new folder actor/<id>) is booted as at start once quiet: new = spawn (tested first), changed = upgrade. The broker `scan` request is gone; the boot pass stays, plus a full pass when the OS watcher overflows. New broker requests watch/unwatch {path} (saved in actor.json `watches`) deliver fs_changed {path, op: write\|create\|remove\|rename, to?} from `fs`, not routed, dropped when the mailbox is full; overflow sends every watcher one fs_changed {path:"/actor", op:"rescan"}. Card #91 (multi-host step 1).	2026-10-17 07:12:04 +00:00
agent-pool-8	6a136c1b17	Borrow host descriptors while tracking VFS paths ci / check (push) Successful in 19s Details	2026-10-16 21:47:17 +00:00
agent-pool-8	3e90cd73d3	Keep discovery test file at served root	2026-10-16 21:47:17 +00:00
agent-pool-8	845c961f4f	Expose advertised actor-served files in directory listings	2026-10-16 21:47:17 +00:00
Sausage agent	d6d3c9a2f2	Bare `sausage` creates ./world if needed and runs it (default app) ci / check (push) Successful in 5s Details - cli: the subcommand is optional; without one, sausage runs ./world and runs init first when ./world is not a world yet (test: cli/tests/cli.rs). So `nix run git+https://gitea.surma.technology/surma/sausage` works. - flake: .#loop without a folder runs bare sausage. - README: mention the bare command and the bare loop.	2026-10-16 21:09:06 +00:00
Sausage Agent	0951ba18e1	README: run straight from the public URL (nix run git+https://...), clone optional ci / check (push) Successful in 8s Details	2026-10-16 20:55:39 +00:00
Sausage Agent	d0876626df	Public repo: .#loop fetches over public HTTPS, README for outsiders ci / check (push) Successful in 8s Details - flake.nix: .#loop runs git+https://gitea.surma.technology/surma/sausage (SAUSAGE_FLAKE overrides it); drop loop-internal / --internal (LAN SSH). - README: clone step, LLM token requirement, loop URL and SAUSAGE_FLAKE. - ci.yml: comment without internal host name.	2026-10-16 20:40:26 +00:00