swarm: per-agent inference settings (model, effort, subagent_models) in KV

swarm-controller keeps each agent's inference settings in a new
`agent-inference` KV bucket, one key per agent, and is its only writer
(`GET`/`PUT`/`DELETE /api/agents/{name}/inference`; `swarmctl agent
inference get|set|clear`). Every field is optional. Each agent reads its
own key every 10 s through a direct-get grant on that key alone, and a
set `model` or `effort` replaces the Bus value, so the next turn runs on
it and the badge follows.

The swarm value is applied without persisting it: `Bus::set_model` and
`set_effort` are split into a persisting setter and a non-persisting
`apply_*`, and the poller uses the latter. Otherwise the swarm value
would overwrite `hyperhive-model`/`hyperhive-effort`, and clearing it, or
a restart, would no longer find the agent's own value. An unset field
changes nothing; a field cleared after being applied goes back to the
local value (`harness_state::local_model`/`local_effort`, the same
resolution `Bus::new` uses), read again at the clear.

`subagent_models` is validated, stored and shown by `get`; no agent
applies it yet.

Refs #5034

docs/README.md

@@ -84,6 +84,8 @@ declarations.
   sealed?** → [`swarm/bao.md`](swarm/bao.md).
 - **How does an agent get a prompt at set times, and why doesn't it fire
   twice after downtime?** → [`swarm/schedules.md`](swarm/schedules.md).
+- **How do I set an agent's model or effort from the swarm, and what happens
+  to its own pick?** → [`swarm/inference.md`](swarm/inference.md).
 
 ## Scheduler, CI, observability
 

docs/swarm/README.md

@@ -23,6 +23,8 @@ grafana, the controller, authelia, …) →
   graph, agent creation. → [below](#swarm-controller)
 - **scheduled prompts** — named prompts for one or more agents at set times,
   kept by the swarm and fired in each agent's container. → [`schedules.md`](schedules.md)
+- **inference settings** — the model and effort the swarm sets for an agent,
+  over its own. → [`inference.md`](inference.md)
 - **swarm UI** — the operator's day-to-day surface, on the swarm apex,
   `admins` only. → [`ui.md`](ui.md)
 - **shared services** — one forge, homeserver, SSO, queue and metrics/logs

docs/swarm/inference.md

@@ -0,0 +1,68 @@
+# Per-agent inference settings
+
+The swarm can set the model and reasoning effort each agent's turns run on,
+plus the models it may start subagents on. A field you set wins over the
+agent's own value; a field you leave unset keeps the agent on its own value,
+so an agent you set nothing for runs as it does today.
+
+## Setting them
+
+```sh
+swarmctl agent inference set iris --model opus --effort high
+swarmctl agent inference get iris
+swarmctl agent inference clear iris
+```
+
+- `set` changes only the flags you pass; the other fields keep what's
+  stored. A hive must declare the agent.
+- `--effort` is one of `low`, `medium`, `high`, `xhigh`, `max`.
+- `--subagent-models` takes a comma-separated list. It's stored and shown,
+  but agents don't apply it yet.
+- `get` shows each field, naming the unset ones.
+- `clear` removes every field; the agent goes back to its own values.
+
+The full flag reference is [`swarmctl-cli.md`](../tools/swarmctl-cli.md).
+
+## When an agent picks a change up
+
+Each agent reads its settings every 10 seconds and switches on the next
+read: its next turn runs on the new model or effort, and its web UI badge
+shows it. A turn already running keeps what it started with.
+
+- **Its own value is the one it runs without the swarm**, from a `/model`
+  or `/effort` pick in its web UI or from its agent.nix. Setting a swarm
+  value doesn't overwrite it.
+- **Clearing a field** (or the whole key) puts the agent back on its own
+  value on the next read.
+- **A `/model` or `/effort` pick while the swarm sets that field** changes
+  the agent's own value, but the swarm value comes back on the next read.
+- **After a restart**, the agent runs its own values until its first read
+  of its settings succeeds.
+
+<details><summary>Implementation</summary>
+
+The settings live in the `agent-inference` KV bucket, one key per agent,
+which swarm-controller alone writes:
+
+```json
+{ "model": "opus", "effort": "high", "subagent_models": ["haiku", "sonnet"] }
+```
+
+Every field is optional, and `{}` reads the same as no key.
+
+| method   | path                            | body                                                                                                                                 |
+| -------- | ------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ |
+| `GET`    | `/api/agents/<agent>/inference` | —; answers the stored value, `{}` when none                                                                                          |
+| `PUT`    | `/api/agents/<agent>/inference` | the whole value, replacing the stored one; the controller refuses unknown fields, an unknown effort, or a model name with whitespace |
+| `DELETE` | `/api/agents/<agent>/inference` | —; answers 204                                                                                                                       |
+
+An agent's own queue credential grants it
+`$JS.API.DIRECT.GET.KV_agent-inference.$KV.agent-inference.<agent>`, the
+direct read of its own key, and nothing else in this bucket — the same shape
+as its schedule key ([`schedules.md`](schedules.md)).
+
+The harness (`hive-agent`'s `swarm_inference`) applies a set field to the
+in-memory model or effort without writing `hyperhive-model` or
+`hyperhive-effort`, which keep the agent's own value.
+
+</details>

docs/swarm/schedules.md

@@ -132,8 +132,8 @@ Each prompt they answer with has `id`, `name`, `agents` (sorted), `body`,
 `next_fire_unix`: the next time on the prompt's grid, as of the moment the
 controller answers, `null` when paused or a one-shot whose time has passed.
 
-An agent's own queue credential grants exactly one subject in the swarm's
+An agent's own queue credential grants exactly one subject in this bucket:
-JetStream: `$JS.API.DIRECT.GET.KV_agent-schedules.$KV.agent-schedules.<agent>`,
+`$JS.API.DIRECT.GET.KV_agent-schedules.$KV.agent-schedules.<agent>`,
 the direct read of its own key. It can't read another agent's key, open or
 watch the bucket, or write to it. A KV watch would need a consumer, and a
 consumer create subject doesn't name a key, so granting one would let the

docs/tools/swarmctl-cli.md

@@ -9,6 +9,10 @@ This document contains the help content for the `swarmctl` command-line program.
 * [`swarmctl agent create`↴](#swarmctl-agent-create)
 * [`swarmctl agent mint-identity`↴](#swarmctl-agent-mint-identity)
 * [`swarmctl agent mint-forge-token`↴](#swarmctl-agent-mint-forge-token)
+* [`swarmctl agent inference`↴](#swarmctl-agent-inference)
+* [`swarmctl agent inference get`↴](#swarmctl-agent-inference-get)
+* [`swarmctl agent inference set`↴](#swarmctl-agent-inference-set)
+* [`swarmctl agent inference clear`↴](#swarmctl-agent-inference-clear)
 * [`swarmctl schedule`↴](#swarmctl-schedule)
 * [`swarmctl schedule add`↴](#swarmctl-schedule-add)
 * [`swarmctl schedule list`↴](#swarmctl-schedule-list)
@@ -59,6 +63,7 @@ Manage agents across the swarm
 * `create` — Queue creation of a new agent on a hive in this swarm
 * `mint-identity` — Queue a re-mint of an existing agent's identity at the swarm's secret store
 * `mint-forge-token` — Check one agent's forge token, and mint it if it's missing or stale
+* `inference` — Show or change the model, effort and subagent models the swarm sets for an agent
 
 
 
@@ -137,6 +142,82 @@ Queues and returns, the same way `agent create` does — watch the swarm UI's jo
 
 
 
+## `swarmctl agent inference`
+
+Show or change the model, effort and subagent models the swarm sets for an agent.
+
+A field the swarm sets wins over the agent's own value (from its `/model` or `/effort` pick or its agent.nix) from its next turn on; the agent reads its settings every ten seconds. A field the swarm doesn't set, or clears, leaves the agent on its own value.
+
+**Usage:** `swarmctl agent inference <COMMAND>`
+
+###### **Subcommands:**
+
+* `get` — Show what the swarm sets for an agent, naming the fields it doesn't set
+* `set` — Set fields for an agent. Only the flags given change; the others keep what's stored
+* `clear` — Clear every field for an agent, so it goes back to its own values
+
+
+
+## `swarmctl agent inference get`
+
+Show what the swarm sets for an agent, naming the fields it doesn't set
+
+**Usage:** `swarmctl agent inference get [OPTIONS] <AGENT>`
+
+###### **Arguments:**
+
+* `<AGENT>` — Agent name
+
+###### **Options:**
+
+* `--controller-socket <PATH>` — swarm-controller's unix socket.
+
+   Supplied by the nix module that installs this binary, from the same `socketPath` option the daemon binds; falls back to `SWARM_CONTROLLER_SOCKET`.
+
+
+
+## `swarmctl agent inference set`
+
+Set fields for an agent. Only the flags given change; the others keep what's stored
+
+**Usage:** `swarmctl agent inference set [OPTIONS] <AGENT>`
+
+###### **Arguments:**
+
+* `<AGENT>` — Agent name. A hive must declare it
+
+###### **Options:**
+
+* `--model <MODEL>` — The model its turns run on, as claude's `--model` takes it
+* `--effort <LEVEL>` — The reasoning effort its turns run at
+
+  Possible values: `low`, `medium`, `high`, `xhigh`, `max`
+
+* `--subagent-models <MODELS>` — The models it may start subagents on, comma-separated. Stored and shown; the agent doesn't apply it yet
+* `--controller-socket <PATH>` — swarm-controller's unix socket.
+
+   Supplied by the nix module that installs this binary, from the same `socketPath` option the daemon binds; falls back to `SWARM_CONTROLLER_SOCKET`.
+
+
+
+## `swarmctl agent inference clear`
+
+Clear every field for an agent, so it goes back to its own values
+
+**Usage:** `swarmctl agent inference clear [OPTIONS] <AGENT>`
+
+###### **Arguments:**
+
+* `<AGENT>` — Agent name
+
+###### **Options:**
+
+* `--controller-socket <PATH>` — swarm-controller's unix socket.
+
+   Supplied by the nix module that installs this binary, from the same `socketPath` option the daemon binds; falls back to `SWARM_CONTROLLER_SOCKET`.
+
+
+
 ## `swarmctl schedule`
 
 Manage scheduled prompts.

docs/turn-loop/claude-invocation.md

@@ -54,7 +54,9 @@ back to `"haiku"` if unset). The operator can flip it at runtime
 with `/model <name>` in the web terminal — the next turn picks it
 up. The harness persists the choice to `/harness/hyperhive-model` so it
 survives restart; override path: `HYPERHIVE_MODEL_FILE` env var
-for tests.
+for tests. A model or effort the swarm sets for the agent replaces the
+`Bus` value, and the harness doesn't persist it; clearing it goes back to
+the persisted or nix value → [`swarm/inference.md`](../swarm/inference.md).
 
 `harness_state::context_window_tokens(model)` looks up
 context-window size per-model. Resolution order (first

hive-agent/Cargo.toml

@@ -37,10 +37,14 @@ rusqlite.workspace = true
 schemars.workspace = true
 serde.workspace = true
 serde_json.workspace = true
-# `agent-schedules` alone: the schedule ticker reads its own key with one direct
+# `agent-schedules` and `agent-inference`: the schedule ticker and the
-# get. `kv`/`notices` name buckets and streams this harness opens neither end
+# inference poller each read their own key with one direct get. `kv`/`notices`
-# of, and the terminal publisher is a plain core-subject publish.
+# name buckets and streams this harness opens neither end of, and the terminal
-swarm-queue-client = { workspace = true, features = ["agent-schedules"] }
+# publisher is a plain core-subject publish.
+swarm-queue-client = { workspace = true, features = [
+  "agent-schedules",
+  "agent-inference",
+] }
 # This agent's own queue secret, read from the store under its own certificate.
 swarm-secret-client.workspace = true
 tokio.workspace = true

hive-agent/src/events.rs

@@ -19,9 +19,8 @@ use serde::{Deserialize, Serialize};
 use tokio::sync::broadcast;
 
 use crate::harness_state::{
-    DEFAULT_EFFORT, DEFAULT_MODEL, configured_effort, configured_model, context_window_tokens,
+    configured_effort, configured_model, context_window_tokens, local_effort, local_model,
-    load_effort, load_model, persist_effort, persist_model, read_harness_state,
+    persist_effort, persist_model, read_harness_state, write_harness_state,
-    write_harness_state,
 };
 
 const CHANNEL_CAPACITY: usize = 256;
@@ -394,21 +393,8 @@ impl Bus {
             }
         };
         let (tx, _) = broadcast::channel(CHANNEL_CAPACITY);
-        // Priority: HIVE_DEFAULT_MODEL (from services.hyperhive.agent.model in agent.nix) >
+        let initial_model = local_model(configured_model());
-        // persisted runtime override > compiled-in DEFAULT_MODEL.
+        let initial_effort = local_effort(configured_effort());
-        // The nix config always wins on rebuild; the persisted file is kept
-        // for within-session tracking only (see persist_model / set_model).
-        let initial_model = configured_model().map_or_else(
-            || load_model().unwrap_or_else(|| DEFAULT_MODEL.to_owned()),
-            str::to_owned,
-        );
-        // Effort precedence is the inverse of model: a persisted operator
-        // pick wins over the nix `HIVE_DEFAULT_EFFORT` baseline so the
-        // runtime choice survives harness restart (it only resets on
-        // `--purge`). nix is the default for a never-touched agent.
-        let initial_effort = load_effort()
-            .or_else(|| configured_effort().map(str::to_owned))
-            .unwrap_or_else(|| DEFAULT_EFFORT.to_owned());
         // Restore rate_limited (and needs_login) from the consolidated
         // harness state file so the dashboard shows the correct status
         // on cold load if the harness crashed while parked.
@@ -590,10 +576,21 @@ impl Bus {
     /// Panics if the internal lock is poisoned.
     pub fn set_model(&self, name: impl Into<String>) {
         let value: String = name.into();
-        self.model.lock().unwrap().clone_from(&value);
         if let Err(e) = persist_model(&value) {
             tracing::warn!(error = ?e, "model: persist failed");
         }
+        self.apply_model(value);
+    }
+
+    /// Switch the model for future turns without persisting it, so
+    /// `hyperhive-model` keeps the agent's local value. For a value the
+    /// swarm sets (`crate::swarm_inference`).
+    ///
+    /// # Panics
+    ///
+    /// Panics if the internal lock is poisoned.
+    pub fn apply_model(&self, value: String) {
+        self.model.lock().unwrap().clone_from(&value);
         // Mirror the resolved model into hyperhive-harness.json so
         // hive-c0re can surface it on the dashboard badge without reading
         // the hyperhive-model override file directly.
@@ -624,10 +621,21 @@ impl Bus {
     /// Panics if the internal lock is poisoned.
     pub fn set_effort(&self, level: impl Into<String>) {
         let value: String = level.into();
-        self.effort.lock().unwrap().clone_from(&value);
         if let Err(e) = persist_effort(&value) {
             tracing::warn!(error = ?e, "effort: persist failed");
         }
+        self.apply_effort(value);
+    }
+
+    /// Switch the effort level without persisting it, so `hyperhive-effort`
+    /// keeps the agent's local value. For a value the swarm sets
+    /// (`crate::swarm_inference`).
+    ///
+    /// # Panics
+    ///
+    /// Panics if the internal lock is poisoned.
+    pub fn apply_effort(&self, value: String) {
+        self.effort.lock().unwrap().clone_from(&value);
         self.emit(LiveEvent::EffortChanged { effort: value });
     }
 

hive-agent/src/harness_state.rs

@@ -53,6 +53,27 @@ pub(crate) fn load_effort() -> Option<String> {
     }
 }
 
+/// The model this agent runs when the swarm sets none. Priority:
+/// `configured` (`HIVE_DEFAULT_MODEL`, from `services.hyperhive.agent.model`
+/// in agent.nix) > the persisted runtime override > [`DEFAULT_MODEL`].
+pub(crate) fn local_model(configured: Option<&str>) -> String {
+    configured.map_or_else(
+        || load_model().unwrap_or_else(|| DEFAULT_MODEL.to_owned()),
+        str::to_owned,
+    )
+}
+
+/// The effort this agent runs at when the swarm sets none. The inverse of
+/// [`local_model`]: a persisted operator pick wins over `configured`
+/// (`HIVE_DEFAULT_EFFORT`), so the runtime choice survives a harness restart
+/// (it only resets on `--purge`); nix is the default for a never-touched
+/// agent.
+pub(crate) fn local_effort(configured: Option<&str>) -> String {
+    load_effort()
+        .or_else(|| configured.map(str::to_owned))
+        .unwrap_or_else(|| DEFAULT_EFFORT.to_owned())
+}
+
 pub(crate) fn persist_effort(level: &str) -> std::io::Result<()> {
     let path = effort_file_path();
     if let Some(parent) = path.parent() {

hive-agent/src/main.rs

@@ -32,6 +32,7 @@ mod state_entry_watch;
 mod stats;
 mod swarm_agent_icon;
 mod swarm_agent_state;
+mod swarm_inference;
 mod swarm_input;
 mod swarm_queue;
 mod swarm_term;
@@ -532,6 +533,8 @@ async fn serve_main<S: Surface>(socket: &Path, poll_ms: u64) -> Result<()> {
     swarm_agent_state::spawn(&bus);
     // Once per start: a new icon only arrives with a new harness process.
     swarm_agent_icon::spawn();
+    // The model and effort the swarm sets for this agent, over its local ones.
+    tokio::spawn(swarm_inference::run(bus.clone()));
     // Set by the web UI's `/api/cancel` on a successful SIGINT, read-and-
     // cleared by `handle_turn` before building the next wake prompt — see
     // `hive_sh4re::inbox::INTERRUPTED_HINT`. Shared between the web server task and

hive-agent/src/swarm_inference.rs

@@ -0,0 +1,316 @@
+//! Applies the model and effort the swarm sets for this agent.
+//!
+//! swarm-controller keeps the agent's settings under its key in the
+//! `agent-inference` KV bucket. This task reads that key every
+//! [`POLL_INTERVAL`] with the one subject the agent's own queue credential is
+//! granted there (`swarm_queue_client::agent_inference::read_own`).
+//!
+//! **A set field wins**: it replaces the Bus value, so the next turn runs on
+//! it and the web UI's badge follows. It is not persisted: `hyperhive-model`
+//! and `hyperhive-effort` keep the agent's local value.
+//!
+//! **An unset field changes nothing**, so an agent the swarm sets nothing for
+//! runs on its local value ([`crate::harness_state::local_model`] and
+//! [`crate::harness_state::local_effort`]). A field that was set and then
+//! cleared goes back to that local value, read again at that moment.
+//!
+//! A failed read changes nothing. Until the first successful read after a
+//! start, the agent runs on its local values.
+//!
+//! `subagent_models` is not read here.
+
+use std::time::Duration;
+
+use swarm_queue_client::agent_inference::{AgentInference, Effort, Read, read_own};
+
+use crate::events::Bus;
+use crate::harness_state::{configured_effort, configured_model, local_effort, local_model};
+
+/// How often the key is read. Also the longest a change can take to reach
+/// the agent's next turn.
+const POLL_INTERVAL: Duration = Duration::from_secs(10);
+
+/// Where the settings are applied, and where the local values come from.
+trait Target {
+    fn model(&self) -> String;
+    fn effort(&self) -> String;
+    fn apply_model(&self, value: String);
+    fn apply_effort(&self, value: String);
+    fn local_model(&self) -> String;
+    fn local_effort(&self) -> String;
+}
+
+/// The harness's Bus, with the nix values read once at start.
+struct Harness {
+    bus: Bus,
+    nix_model: Option<&'static str>,
+    nix_effort: Option<&'static str>,
+}
+
+impl Target for Harness {
+    fn model(&self) -> String {
+        self.bus.model()
+    }
+    fn effort(&self) -> String {
+        self.bus.effort()
+    }
+    fn apply_model(&self, value: String) {
+        self.bus.apply_model(value);
+    }
+    fn apply_effort(&self, value: String) {
+        self.bus.apply_effort(value);
+    }
+    fn local_model(&self) -> String {
+        local_model(self.nix_model)
+    }
+    fn local_effort(&self) -> String {
+        local_effort(self.nix_effort)
+    }
+}
+
+/// Which fields currently carry the swarm's value.
+#[derive(Debug, Default)]
+struct Overridden {
+    model: bool,
+    effort: bool,
+}
+
+/// What one field should become: the swarm's value while it sets one, the
+/// local value on the read after it stops, and otherwise `None` (left alone).
+/// Also `None` when the value is already what it should be.
+fn next(
+    swarm: Option<&str>,
+    overridden: &mut bool,
+    current: &str,
+    local: impl FnOnce() -> String,
+) -> Option<String> {
+    let wanted = match swarm {
+        Some(value) => {
+            *overridden = true;
+            value.to_owned()
+        }
+        None if std::mem::take(overridden) => local(),
+        None => return None,
+    };
+    (wanted != current).then_some(wanted)
+}
+
+impl Overridden {
+    /// Apply one read of the key to `target`.
+    fn apply(&mut self, read: &AgentInference, target: &impl Target) {
+        if let Some(model) = next(
+            read.model.as_deref(),
+            &mut self.model,
+            &target.model(),
+            || target.local_model(),
+        ) {
+            tracing::info!(%model, swarm = self.model, "inference: switching model");
+            target.apply_model(model);
+        }
+        if let Some(effort) = next(
+            read.effort.map(Effort::as_str),
+            &mut self.effort,
+            &target.effort(),
+            || target.local_effort(),
+        ) {
+            tracing::info!(%effort, swarm = self.effort, "inference: switching effort");
+            target.apply_effort(effort);
+        }
+    }
+}
+
+/// Run the poller for the harness's lifetime, if this agent has a label to
+/// name its key with.
+pub async fn run(bus: Bus) {
+    let agent = crate::identity::label();
+    if agent.is_empty() {
+        tracing::warn!("this agent has no label; not reading its swarm inference settings");
+        return;
+    }
+    let target = Harness {
+        bus,
+        nix_model: configured_model(),
+        nix_effort: configured_effort(),
+    };
+    let client = crate::swarm_queue::client().await;
+    tracing::info!(%agent, "reading swarm inference settings from the swarm queue");
+    let mut overridden = Overridden::default();
+    loop {
+        if let Some(read) = read(&client, &agent).await {
+            overridden.apply(&read, &target);
+        }
+        tokio::time::sleep(POLL_INTERVAL).await;
+    }
+}
+
+/// The key's current value, or `None` when it could not be read.
+async fn read(client: &async_nats::Client, agent: &str) -> Option<AgentInference> {
+    // An unconnected client does not fail a JetStream request, it hangs on it.
+    swarm_queue_client::ensure_connected(client).ok()?;
+    match read_own(client, agent).await {
+        Ok(Read::Current(current)) => Some(current),
+        Ok(Read::NoBucket) => {
+            tracing::debug!("inference: the agent-inference bucket does not answer yet");
+            None
+        }
+        Err(e) => {
+            tracing::warn!(error = %e, "inference: reading this agent's settings failed");
+            None
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use std::cell::RefCell;
+
+    use super::*;
+
+    /// A Bus stand-in recording every apply.
+    struct Fake {
+        model: RefCell<String>,
+        effort: RefCell<String>,
+        local_model: RefCell<String>,
+        local_effort: String,
+        applies: RefCell<usize>,
+    }
+
+    impl Fake {
+        fn new() -> Self {
+            Self {
+                model: RefCell::new("sonnet".to_owned()),
+                effort: RefCell::new("medium".to_owned()),
+                local_model: RefCell::new("sonnet".to_owned()),
+                local_effort: "medium".to_owned(),
+                applies: RefCell::new(0),
+            }
+        }
+        fn now(&self) -> (String, String) {
+            (self.model.borrow().clone(), self.effort.borrow().clone())
+        }
+    }
+
+    impl Target for Fake {
+        fn model(&self) -> String {
+            self.model.borrow().clone()
+        }
+        fn effort(&self) -> String {
+            self.effort.borrow().clone()
+        }
+        fn apply_model(&self, value: String) {
+            *self.applies.borrow_mut() += 1;
+            *self.model.borrow_mut() = value;
+        }
+        fn apply_effort(&self, value: String) {
+            *self.applies.borrow_mut() += 1;
+            *self.effort.borrow_mut() = value;
+        }
+        fn local_model(&self) -> String {
+            self.local_model.borrow().clone()
+        }
+        fn local_effort(&self) -> String {
+            self.local_effort.clone()
+        }
+    }
+
+    fn set(model: Option<&str>, effort: Option<Effort>) -> AgentInference {
+        AgentInference {
+            model: model.map(str::to_owned),
+            effort,
+            subagent_models: None,
+        }
+    }
+
+    fn pair(model: &str, effort: &str) -> (String, String) {
+        (model.to_owned(), effort.to_owned())
+    }
+
+    #[test]
+    fn a_set_field_replaces_the_bus_value() {
+        let fake = Fake::new();
+        let mut o = Overridden::default();
+        o.apply(&set(Some("opus"), Some(Effort::High)), &fake);
+        assert_eq!(fake.now(), pair("opus", "high"));
+        // Re-reading the same value applies nothing again.
+        o.apply(&set(Some("opus"), Some(Effort::High)), &fake);
+        assert_eq!(*fake.applies.borrow(), 2);
+    }
+
+    /// An absent key, and a key with every field unset, leave the agent on
+    /// whatever it runs now — even when that differs from its local value.
+    #[test]
+    fn an_absent_key_leaves_the_bus_alone() {
+        let fake = Fake::new();
+        *fake.model.borrow_mut() = "haiku".to_owned();
+        let mut o = Overridden::default();
+        for _ in 0..3 {
+            o.apply(&AgentInference::default(), &fake);
+        }
+        assert_eq!(fake.now(), pair("haiku", "medium"));
+        assert_eq!(*fake.applies.borrow(), 0);
+    }
+
+    /// Only the field the swarm sets moves; the other keeps its value.
+    #[test]
+    fn an_unset_field_beside_a_set_one_is_left_alone() {
+        let fake = Fake::new();
+        *fake.effort.borrow_mut() = "max".to_owned();
+        let mut o = Overridden::default();
+        o.apply(&set(Some("opus"), None), &fake);
+        assert_eq!(fake.now(), pair("opus", "max"));
+    }
+
+    #[test]
+    fn a_cleared_field_goes_back_to_the_local_value() {
+        let fake = Fake::new();
+        let mut o = Overridden::default();
+        o.apply(&set(Some("opus"), Some(Effort::Xhigh)), &fake);
+        assert_eq!(fake.now(), pair("opus", "xhigh"));
+        // The effort is cleared, the model still set.
+        o.apply(&set(Some("opus"), None), &fake);
+        assert_eq!(fake.now(), pair("opus", "medium"));
+        // The local value is read at the clear, not at start.
+        "haiku".clone_into(&mut fake.local_model.borrow_mut());
+        o.apply(&AgentInference::default(), &fake);
+        assert_eq!(fake.now(), pair("haiku", "medium"));
+        // Once reverted, a later unset read changes nothing.
+        *fake.model.borrow_mut() = "sonnet".to_owned();
+        o.apply(&AgentInference::default(), &fake);
+        assert_eq!(fake.now(), pair("sonnet", "medium"));
+    }
+
+    /// On the real Bus: the swarm's value reaches the Bus but not the
+    /// persisted files, so clearing it, or a restart, finds the local value.
+    #[test]
+    fn the_swarm_value_is_not_persisted_over_the_local_one() {
+        let dir = tempfile::tempdir().unwrap();
+        let prev = std::env::var_os("HYPERHIVE_HARNESS_DIR");
+        // SAFETY: no other test in this crate reads or sets this variable.
+        unsafe { std::env::set_var("HYPERHIVE_HARNESS_DIR", dir.path()) };
+        let bus = Bus::new();
+        bus.set_model("sonnet");
+        bus.set_effort("low");
+        let target = Harness {
+            bus: bus.clone(),
+            nix_model: None,
+            nix_effort: None,
+        };
+        let mut o = Overridden::default();
+        o.apply(&set(Some("opus"), Some(Effort::Max)), &target);
+        let swarm = (bus.model(), bus.effort());
+        let files = (
+            crate::harness_state::load_model(),
+            crate::harness_state::load_effort(),
+        );
+        o.apply(&AgentInference::default(), &target);
+        let cleared = (bus.model(), bus.effort());
+        match prev {
+            // SAFETY: as above.
+            Some(v) => unsafe { std::env::set_var("HYPERHIVE_HARNESS_DIR", v) },
+            None => unsafe { std::env::remove_var("HYPERHIVE_HARNESS_DIR") },
+        }
+        assert_eq!(swarm, pair("opus", "max"));
+        assert_eq!(files, (Some("sonnet".to_owned()), Some("low".to_owned())));
+        assert_eq!(cleared, pair("sonnet", "low"));
+    }
+}

nix/host-modules/swarm-nats.nix

@@ -739,6 +739,9 @@ in
                 # `CONSUMER.CREATE` (a watch's create subject carries no key,
                 # so it would cover every agent's) and no `$KV` write.
                 "--agent-token-publish-subject ${lib.escapeShellArg "\$\$JS.API.DIRECT.GET.KV_agent-schedules.\$\$KV.agent-schedules.{agent}"}"
+                # Its own key of the `agent-inference` KV bucket
+                # (`swarm_queue_client::agent_inference`), on the same terms.
+                "--agent-token-publish-subject ${lib.escapeShellArg "\$\$JS.API.DIRECT.GET.KV_agent-inference.\$\$KV.agent-inference.{agent}"}"
                 "--store-cert-role ${lib.escapeShellArg authCertRole}"
               ];
               # The store identity arrives by `LoadCredential` and is named

nix/module-eval/nats-authelia.nix

@@ -295,7 +295,7 @@ let
       # The whole per-agent grant, compared as a list rather than by infix: a
       # wider subject added beside these (`$$JS.API.>`, another stream's name)
       # would pass every presence check above.
-      name = "a verified agent's grant is exactly its own subjects, and its one JetStream API subject reads its own schedule key";
+      name = "a verified agent's grant is exactly its own subjects, and its JetStream API subjects read its own schedule and inference keys";
       ok =
         let
           args = lib.splitString " " (responderOf natsOldPath).serviceConfig.ExecStart;
@@ -314,6 +314,7 @@ let
           "'$$KV.agent-icons.{agent}'"
           "'$$SWARM.term.{agent}.sub.>'"
           "'$$JS.API.DIRECT.GET.KV_agent-schedules.$$KV.agent-schedules.{agent}'"
+          "'$$JS.API.DIRECT.GET.KV_agent-inference.$$KV.agent-inference.{agent}'"
         ];
     }
     {

swarm-controller/README.md

@@ -46,6 +46,10 @@ The swarm's control plane. The swarm UI and `swarmctl` are its clients.
   `agent-schedules`, both of which this daemon alone writes
   (`GET`/`POST /api/schedules`, `PATCH`/`DELETE /api/schedules/{id}`); each
   agent fires them itself → [`docs/swarm/schedules.md`](../docs/swarm/schedules.md).
+- **Inference settings** — each agent's model, effort and subagent models,
+  one key per agent in the `agent-inference` KV bucket, which this daemon
+  alone writes (`GET`/`PUT`/`DELETE /api/agents/{name}/inference`); each
+  agent applies them itself → [`docs/swarm/inference.md`](../docs/swarm/inference.md).
 - **Relays** — each agent's terminal and turn-state header as SSE, agent
   icons, the cross-repo issue report, and the UI's quick links.
 

swarm-controller/src/inference.rs

@@ -0,0 +1,295 @@
+//! `/api/agents/{name}/inference` — the model, effort and subagent models the
+//! swarm sets for one agent.
+//!
+//! Each agent's settings are one value under its own key of `agent-inference`
+//! (`swarm_queue_client::agent_inference`), which this daemon alone writes.
+//! The agent reads its key and applies each set field over its local value;
+//! an unset field leaves the local value in force.
+
+use std::collections::BTreeSet;
+
+use axum::Json;
+use axum::extract::{Path, State};
+use axum::http::StatusCode;
+use problem_details::ProblemDetails;
+use serde::{Deserialize, Serialize};
+use swarm_queue_client::agent_inference::{self, AgentInference, Effort};
+use utoipa::ToSchema;
+
+use crate::schedules::{live, queue, read, refused, write};
+use crate::{AppState, error_problem, wanted_error_status};
+
+/// How many times a losing writer re-reads and re-applies, as in
+/// `wanted::MAX_ATTEMPTS`.
+const MAX_ATTEMPTS: usize = 5;
+
+/// One agent's inference settings. An absent field is not set at swarm level,
+/// and the agent keeps its own value for it.
+#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize, ToSchema)]
+#[serde(deny_unknown_fields)]
+pub struct InferenceSettings {
+    /// The model the agent's turns run on, as claude's `--model` takes it.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    model: Option<String>,
+    /// One of `low`, `medium`, `high`, `xhigh`, `max`.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    #[schema(value_type = Option<String>)]
+    effort: Option<Effort>,
+    /// The models the agent may start its subagents on. At least one.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    subagent_models: Option<Vec<String>>,
+}
+
+impl From<AgentInference> for InferenceSettings {
+    fn from(v: AgentInference) -> Self {
+        Self {
+            model: v.model,
+            effort: v.effort,
+            subagent_models: v.subagent_models,
+        }
+    }
+}
+
+/// A model name as the agent passes it to claude: one non-empty word.
+fn check_model(name: &str, what: &str) -> Result<(), String> {
+    if name.is_empty() {
+        return Err(format!("the {what} is empty"));
+    }
+    if name.chars().any(|c| c.is_whitespace() || c.is_control()) {
+        return Err(format!("the {what} {name:?} contains whitespace"));
+    }
+    Ok(())
+}
+
+impl InferenceSettings {
+    /// The value to store, or why it is refused.
+    fn into_stored(self) -> Result<AgentInference, String> {
+        if let Some(model) = &self.model {
+            check_model(model, "model")?;
+        }
+        if let Some(models) = &self.subagent_models {
+            if models.is_empty() {
+                return Err(
+                    "subagent_models is empty; leave it out to keep the agent's own list"
+                        .to_owned(),
+                );
+            }
+            let mut seen = BTreeSet::new();
+            for model in models {
+                check_model(model, "subagent model")?;
+                if !seen.insert(model) {
+                    return Err(format!("subagent model {model:?} is listed twice"));
+                }
+            }
+        }
+        Ok(AgentInference {
+            model: self.model,
+            effort: self.effort,
+            subagent_models: self.subagent_models,
+        })
+    }
+}
+
+fn agent_name(name: &str) -> Result<String, (StatusCode, String)> {
+    hive_types::Ident::parse(name)
+        .map(hive_types::Ident::into_string)
+        .map_err(|reason| (StatusCode::BAD_REQUEST, reason.to_owned()))
+}
+
+fn queue_problem(e: &anyhow::Error) -> ProblemDetails {
+    tracing::warn!(error = %format!("{e:#}"), "inference: queue failure");
+    error_problem(wanted_error_status(e), &format!("{e:#}"))
+}
+
+async fn open(
+    client: &async_nats::Client,
+) -> Result<async_nats::jetstream::kv::Store, ProblemDetails> {
+    agent_inference::open_or_create(client)
+        .await
+        .map_err(|e| queue_problem(&anyhow::Error::new(e)))
+}
+
+#[utoipa::path(
+    get,
+    path = "/api/agents/{name}/inference",
+    params(("name" = String, Path, description = "agent name")),
+    responses(
+        (status = 200, description = "the agent's settings; `{}` when none is set",
+            body = InferenceSettings),
+        (status = 400, description = "not an identifier (problem+json)", body = String),
+        (status = 503, description = "no swarm queue is configured on this host, or it is \
+            not connected (problem+json)", body = String),
+        (status = 500, description = "the settings could not be read (problem+json)",
+            body = String),
+    ),
+    tag = "agents"
+)]
+pub(crate) async fn get_agent_inference(
+    State(state): State<AppState>,
+    Path(name): Path<String>,
+) -> Result<Json<InferenceSettings>, ProblemDetails> {
+    let agent = agent_name(&name).map_err(refused)?;
+    let client = queue(&state).map_err(refused)?;
+    let store = open(&client).await?;
+    let (current, _) = read::<AgentInference>(&store, &agent)
+        .await
+        .map_err(|e| queue_problem(&e))?;
+    Ok(Json(current.into()))
+}
+
+#[utoipa::path(
+    put,
+    path = "/api/agents/{name}/inference",
+    params(("name" = String, Path, description = "agent name")),
+    request_body = InferenceSettings,
+    responses(
+        (status = 200, description = "the settings as stored; they replace the previous \
+            ones whole", body = InferenceSettings),
+        (status = 400, description = "not an identifier, not declared on any hive, or a \
+            model name is empty, has whitespace, or is listed twice (problem+json)",
+            body = String),
+        (status = 422, description = "an unknown field, or an effort that is not one of \
+            the five levels"),
+        (status = 503, description = "no swarm queue is configured on this host, or it is \
+            not connected (problem+json)", body = String),
+        (status = 500, description = "the settings could not be written (problem+json)",
+            body = String),
+    ),
+    tag = "agents"
+)]
+pub(crate) async fn put_agent_inference(
+    State(state): State<AppState>,
+    Path(name): Path<String>,
+    Json(req): Json<InferenceSettings>,
+) -> Result<Json<InferenceSettings>, ProblemDetails> {
+    let agent = agent_name(&name).map_err(refused)?;
+    let value = req
+        .into_stored()
+        .map_err(|why| error_problem(StatusCode::BAD_REQUEST, &why))?;
+    let client = queue(&state).map_err(refused)?;
+    if !live(&state).await?.contains(&agent) {
+        return Err(error_problem(
+            StatusCode::BAD_REQUEST,
+            &format!("no hive's wanted state declares an agent {agent:?} that is not destroyed"),
+        ));
+    }
+    let store = open(&client).await?;
+    for _ in 0..MAX_ATTEMPTS {
+        let (current, revision) = read::<AgentInference>(&store, &agent)
+            .await
+            .map_err(|e| queue_problem(&e))?;
+        if current == value
+            || write(&store, &agent, &value, revision)
+                .await
+                .map_err(|e| queue_problem(&e))?
+        {
+            tracing::info!(%agent, ?value, "set inference settings");
+            return Ok(Json(value.into()));
+        }
+    }
+    Err(queue_problem(&anyhow::anyhow!(
+        "gave up writing the inference settings of {agent} after {MAX_ATTEMPTS} conflicting \
+         writes"
+    )))
+}
+
+#[utoipa::path(
+    delete,
+    path = "/api/agents/{name}/inference",
+    params(("name" = String, Path, description = "agent name")),
+    responses(
+        (status = 204, description = "cleared; the agent goes back to its own values"),
+        (status = 400, description = "not an identifier (problem+json)", body = String),
+        (status = 503, description = "no swarm queue is configured on this host, or it is \
+            not connected (problem+json)", body = String),
+        (status = 500, description = "the settings could not be cleared (problem+json)",
+            body = String),
+    ),
+    tag = "agents"
+)]
+pub(crate) async fn delete_agent_inference(
+    State(state): State<AppState>,
+    Path(name): Path<String>,
+) -> Result<StatusCode, ProblemDetails> {
+    let agent = agent_name(&name).map_err(refused)?;
+    let client = queue(&state).map_err(refused)?;
+    let store = open(&client).await?;
+    store
+        .delete(&agent)
+        .await
+        .map_err(|e| queue_problem(&anyhow::Error::new(e)))?;
+    tracing::info!(%agent, "cleared inference settings");
+    Ok(StatusCode::NO_CONTENT)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn settings(json: &str) -> serde_json::Result<InferenceSettings> {
+        serde_json::from_str(json)
+    }
+
+    #[test]
+    fn a_full_put_is_stored_as_sent() {
+        let stored = settings(r#"{"model":"opus","effort":"high","subagent_models":["haiku"]}"#)
+            .unwrap()
+            .into_stored()
+            .unwrap();
+        assert_eq!(
+            stored,
+            AgentInference {
+                model: Some("opus".to_owned()),
+                effort: Some(Effort::High),
+                subagent_models: Some(vec!["haiku".to_owned()]),
+            }
+        );
+        assert_eq!(
+            serde_json::to_value(InferenceSettings::from(stored)).unwrap(),
+            serde_json::json!({"model": "opus", "effort": "high", "subagent_models": ["haiku"]})
+        );
+    }
+
+    /// Each field is optional, and one left out stays out of the stored value.
+    #[test]
+    fn a_partial_put_leaves_the_other_fields_unset() {
+        let stored = settings(r#"{"effort":"max"}"#)
+            .unwrap()
+            .into_stored()
+            .unwrap();
+        assert_eq!(
+            stored,
+            AgentInference {
+                effort: Some(Effort::Max),
+                ..AgentInference::default()
+            }
+        );
+    }
+
+    #[test]
+    fn an_unknown_field_or_effort_does_not_decode() {
+        for json in [
+            r#"{"modle":"opus"}"#,
+            r#"{"available_models":["opus"]}"#,
+            r#"{"effort":"extreme"}"#,
+            r#"{"effort":"HIGH"}"#,
+        ] {
+            assert!(settings(json).is_err(), "{json}");
+        }
+        // Control: the same shape with a known field and level decodes.
+        assert!(settings(r#"{"model":"opus","effort":"xhigh"}"#).is_ok());
+    }
+
+    #[test]
+    fn bad_model_names_are_refused() {
+        for json in [
+            r#"{"model":""}"#,
+            r#"{"model":"opus 4"}"#,
+            r#"{"subagent_models":[]}"#,
+            r#"{"subagent_models":["haiku",""]}"#,
+            r#"{"subagent_models":["haiku","haiku"]}"#,
+        ] {
+            assert!(settings(json).unwrap().into_stored().is_err(), "{json}");
+        }
+    }
+}

swarm-controller/src/main.rs

@@ -52,6 +52,7 @@ mod config_pr;
 mod forge;
 mod forge_account;
 mod github_account;
+mod inference;
 mod issue_report;
 mod linked_accounts;
 mod matrix_account;
@@ -3211,6 +3212,11 @@ fn build_app(state: AppState) -> axum::Router {
         .routes(routes!(get_jobq_rollup))
         .routes(routes!(get_agent_config_pr))
         .routes(routes!(get_agent_icon))
+        .routes(routes!(
+            inference::get_agent_inference,
+            inference::put_agent_inference,
+            inference::delete_agent_inference
+        ))
         .routes(routes!(get_config_prs))
         .routes(routes!(create_agent))
         .routes(routes!(mint_agent_identity))
@@ -5479,6 +5485,59 @@ mod tests {
         );
     }
 
+    /// Every inference verb refuses a name that is not an identifier with a
+    /// 400 and answers 503 without a queue; a PUT whose body is refused is a
+    /// 400 before the queue is asked.
+    #[tokio::test]
+    async fn the_inference_routes_check_the_name_and_body_before_the_queue() {
+        use super::inference::{delete_agent_inference, get_agent_inference, put_agent_inference};
+        use axum::extract::{Path, State};
+        use axum::http::StatusCode;
+
+        fn status<T>(r: Result<T, problem_details::ProblemDetails>) -> Option<StatusCode> {
+            r.map(|_| ()).expect_err("must not succeed").status
+        }
+        let (state, _sched) = state_with_roster();
+        let body = |json: &str| axum::Json(serde_json::from_str(json).expect("decodes"));
+        for name in ["iris", "Not An Ident"] {
+            let want = Some(if name == "iris" {
+                StatusCode::SERVICE_UNAVAILABLE
+            } else {
+                StatusCode::BAD_REQUEST
+            });
+            let path = || Path(name.to_owned());
+            assert_eq!(
+                status(get_agent_inference(State(state.clone()), path()).await),
+                want,
+                "GET {name}"
+            );
+            assert_eq!(
+                status(
+                    put_agent_inference(State(state.clone()), path(), body(r#"{"model":"opus"}"#))
+                        .await
+                ),
+                want,
+                "PUT {name}"
+            );
+            assert_eq!(
+                status(delete_agent_inference(State(state.clone()), path()).await),
+                want,
+                "DELETE {name}"
+            );
+        }
+        assert_eq!(
+            status(
+                put_agent_inference(
+                    State(state),
+                    Path("iris".to_owned()),
+                    body(r#"{"model":"opus 4"}"#)
+                )
+                .await
+            ),
+            Some(StatusCode::BAD_REQUEST)
+        );
+    }
+
     #[tokio::test]
     async fn listing_subagents_needs_a_queue_and_an_identifier() {
         let (state, _sched) = state_with_roster();

swarm-controller/src/schedules.rs

@@ -404,7 +404,7 @@ async fn open_or_create_definitions(client: &async_nats::Client) -> anyhow::Resu
 
 /// The value of `key` decoded, with the revision to write it back against.
 /// An absent or deleted key is the default value.
-async fn read<T: Default + serde::de::DeserializeOwned>(
+pub(crate) async fn read<T: Default + serde::de::DeserializeOwned>(
     store: &Store,
     key: &str,
 ) -> anyhow::Result<(T, Option<u64>)> {
@@ -425,7 +425,7 @@ async fn read<T: Default + serde::de::DeserializeOwned>(
 
 /// Write `value` to `key` against `revision`. `Ok(false)` when another write
 /// landed first.
-async fn write<T: Serialize>(
+pub(crate) async fn write<T: Serialize>(
     store: &Store,
     key: &str,
     value: &T,
@@ -549,7 +549,7 @@ pub(crate) fn spawn_reconcile(
 }
 
 /// The connected queue client, or the problem.
-fn queue(state: &AppState) -> Result<async_nats::Client, (StatusCode, String)> {
+pub(crate) fn queue(state: &AppState) -> Result<async_nats::Client, (StatusCode, String)> {
     let status = state.status.as_ref().ok_or_else(|| {
         (
             StatusCode::SERVICE_UNAVAILABLE,
@@ -566,12 +566,12 @@ fn queue(state: &AppState) -> Result<async_nats::Client, (StatusCode, String)> {
     Ok(client)
 }
 
-fn refused((status, detail): (StatusCode, String)) -> ProblemDetails {
+pub(crate) fn refused((status, detail): (StatusCode, String)) -> ProblemDetails {
     error_problem(status, &detail)
 }
 
 /// Every agent some hive declares and has not destroyed.
-async fn live(state: &AppState) -> Result<BTreeSet<String>, ProblemDetails> {
+pub(crate) async fn live(state: &AppState) -> Result<BTreeSet<String>, ProblemDetails> {
     let Some(wanted) = state.wanted.as_ref() else {
         return Err(error_problem(
             StatusCode::SERVICE_UNAVAILABLE,

swarm-nats-auth/src/main.rs

@@ -507,12 +507,14 @@ mod tests {
             "$SWARM.term.{agent}.sub.>",
             "--agent-token-publish-subject",
             "$JS.API.DIRECT.GET.KV_agent-schedules.$KV.agent-schedules.{agent}",
+            "--agent-token-publish-subject",
+            "$JS.API.DIRECT.GET.KV_agent-inference.$KV.agent-inference.{agent}",
             "--store-cert-role",
             "swarm-nats-auth",
         ])
         .expect("the unit's own argument vector must parse");
         assert_eq!(args.agent_client_suffix, "-agent");
-        assert_eq!(args.agent_token_publish_subjects.len(), 5);
+        assert_eq!(args.agent_token_publish_subjects.len(), 6);
     }
 
     /// A request with no `Nats-Server-Xkey` is dropped before anything reads

swarm-nats-auth/src/policy.rs

@@ -1440,6 +1440,39 @@ mod tests {
         assert!(covers("$KV.*.>", &key_subject("argus")));
     }
 
+    /// The inference template as `swarm-nats.nix` spells it reads one key, the
+    /// agent's own, and nothing that writes, watches or opens the bucket.
+    #[test]
+    fn an_agent_may_read_only_its_own_inference_key() {
+        use swarm_queue_client::agent_inference::{STREAM, direct_get_subject, key_subject};
+
+        let p = policy()
+            .with_agent_token_subjects(vec![
+                "$JS.API.DIRECT.GET.KV_agent-inference.$KV.agent-inference.{agent}".to_owned(),
+            ])
+            .expect("a per-agent template is valid");
+        let atlas = p.agent_token_permissions("atlas").expect("configured");
+        assert_eq!(atlas.publish, vec![direct_get_subject("atlas")]);
+        let argus = p.agent_token_permissions("argus").expect("configured");
+        assert_eq!(argus.publish, vec![direct_get_subject("argus")]);
+
+        for refused in [
+            direct_get_subject("argus"),
+            format!("$JS.API.DIRECT.GET.{STREAM}"),
+            format!("$JS.API.STREAM.INFO.{STREAM}"),
+            format!("$JS.API.CONSUMER.CREATE.{STREAM}"),
+            key_subject("atlas"),
+            swarm_queue_client::agent_schedules::direct_get_subject("atlas"),
+        ] {
+            assert!(
+                !atlas.publish.iter().any(|g| covers(g, &refused)),
+                "atlas's grant covers {refused}"
+            );
+        }
+        // Control: the matcher matches the one subject granted.
+        assert!(covers(&atlas.publish[0], &direct_get_subject("atlas")));
+    }
+
     #[test]
     fn an_agent_token_subject_without_the_placeholder_is_refused() {
         let err = policy()

swarm-queue-client/Cargo.toml

@@ -37,6 +37,9 @@ subagent-term = ["async-nats/jetstream"]
 # schedule key without opening the bucket. Explicit for the same reason as
 # `notices`.
 agent-schedules = ["async-nats/jetstream"]
+# `agent_inference::read_own`, for the agent harness, which reads its own
+# inference key the same way. Explicit for the same reason as `notices`.
+agent-inference = ["async-nats/jetstream"]
 
 [dependencies]
 # Bare (no `kv`/`jetstream`) unless a consumer opts into the `kv` feature

swarm-queue-client/src/agent_inference.rs

@@ -0,0 +1,235 @@
+//! The per-agent inference-settings KV bucket: one key per agent, holding the
+//! model, effort and subagent model list the swarm sets for it.
+//!
+//! **swarm-controller is the only writer**, and creates the bucket on its
+//! first write. Every field is optional: an absent field leaves the agent on
+//! its own local value, so an agent with no key runs exactly as it would with
+//! no swarm.
+//!
+//! **An agent reads with one `DIRECT.GET` on its own key, not a watch**, for
+//! the reason [`crate::agent_schedules`] gives: the direct get names the key
+//! in its subject ([`crate::agent_inference::direct_get_subject`]), which is
+//! what lets the queue grant scope the read to one agent.
+//!
+//! Keyed by the agent alone, like [`crate::agent_schedules`]: an agent can
+//! move between hives.
+
+/// The KV bucket inference settings are published into, keyed by agent name.
+pub const BUCKET: &str = "agent-inference";
+
+/// The stream backing [`BUCKET`]. `create_key_value` prefixes `KV_`.
+pub const STREAM: &str = "KV_agent-inference";
+
+/// The subject `agent`'s entry is stored under.
+#[must_use]
+pub fn key_subject(agent: &str) -> String {
+    format!("$KV.{BUCKET}.{agent}")
+}
+
+/// The request subject that reads `agent`'s current entry, and so the one
+/// subject the queue grants that agent in this bucket.
+#[must_use]
+pub fn direct_get_subject(agent: &str) -> String {
+    format!("$JS.API.DIRECT.GET.{STREAM}.{}", key_subject(agent))
+}
+
+/// One agent's inference settings: the value under `BUCKET`/`<agent>`.
+/// `None` is "not set at swarm level", never "use a default".
+#[derive(Debug, Clone, Default, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+pub struct AgentInference {
+    /// The model the agent's turns run on.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub model: Option<String>,
+    /// The reasoning effort the agent's turns run at.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub effort: Option<Effort>,
+    /// The models the agent may start its subagents on.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub subagent_models: Option<Vec<String>>,
+}
+
+/// A claude `--effort` level, ascending.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(rename_all = "lowercase")]
+pub enum Effort {
+    Low,
+    Medium,
+    High,
+    Xhigh,
+    Max,
+}
+
+impl Effort {
+    /// The level as claude's `--effort` takes it, and as it is stored.
+    #[must_use]
+    pub const fn as_str(self) -> &'static str {
+        match self {
+            Self::Low => "low",
+            Self::Medium => "medium",
+            Self::High => "high",
+            Self::Xhigh => "xhigh",
+            Self::Max => "max",
+        }
+    }
+}
+
+/// Open the inference bucket for writing, creating it if nothing has yet.
+///
+/// **Controller-side only.** `history: 1`: an agent wants its current value.
+#[cfg(feature = "kv")]
+pub async fn open_or_create(
+    client: &async_nats::Client,
+) -> Result<async_nats::jetstream::kv::Store, crate::Error> {
+    let js = async_nats::jetstream::new(client.clone());
+    crate::kv::open_or_create(
+        &js,
+        BUCKET,
+        "agent-inference",
+        "Inference settings the swarm controller sets for each agent".to_owned(),
+    )
+    .await
+}
+
+/// What reading one agent's key found.
+#[cfg(feature = "agent-inference")]
+#[derive(Debug)]
+pub enum Read {
+    /// The bucket answered; an absent or deleted key has every field unset.
+    Current(AgentInference),
+    /// Nothing answers for the bucket's stream: no setting has been written
+    /// yet, or `JetStream` is not serving it right now. Not an unset value,
+    /// since the second case says nothing about the agent's settings.
+    NoBucket,
+}
+
+/// Read `agent`'s current settings with the one subject its grant covers.
+///
+/// # Errors
+///
+/// A failed request or an undecodable value. A request the queue refused for
+/// permissions reaches the client as a timeout, not as a refusal.
+#[cfg(feature = "agent-inference")]
+pub async fn read_own(
+    client: &async_nats::Client,
+    agent: &str,
+) -> Result<Read, Box<dyn std::error::Error + Send + Sync>> {
+    use async_nats::jetstream::stream::DirectGetErrorKind;
+
+    let js = async_nats::jetstream::new(client.clone());
+    let stream = js.get_stream_no_info(STREAM).await?;
+    let message = match stream.direct_get_last_for_subject(key_subject(agent)).await {
+        Ok(message) => message,
+        Err(e) if matches!(e.kind(), DirectGetErrorKind::NotFound) => {
+            return Ok(Read::Current(AgentInference::default()));
+        }
+        Err(e) if is_no_responders(&e) => return Ok(Read::NoBucket),
+        Err(e) => return Err(e.into()),
+    };
+    // A KV delete or purge is stored as an empty marker carrying this header.
+    let removed = message
+        .headers
+        .get("KV-Operation")
+        .is_some_and(|op| matches!(op.as_str(), "DEL" | "PURGE"));
+    if removed {
+        return Ok(Read::Current(AgentInference::default()));
+    }
+    Ok(Read::Current(serde_json::from_slice(&message.payload)?))
+}
+
+#[cfg(feature = "agent-inference")]
+fn is_no_responders(e: &async_nats::jetstream::stream::DirectGetError) -> bool {
+    matches!(
+        e.kind(),
+        async_nats::jetstream::stream::DirectGetErrorKind::ErrorResponse(status, _)
+            if status == async_nats::StatusCode::NO_RESPONDERS
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The names the grant in `swarm-nats.nix` spells with `{agent}`.
+    #[test]
+    fn names_match_the_queue_grant() {
+        assert_eq!(STREAM, format!("KV_{BUCKET}"));
+        assert_eq!(key_subject("iris"), "$KV.agent-inference.iris");
+        assert_eq!(
+            direct_get_subject("iris"),
+            "$JS.API.DIRECT.GET.KV_agent-inference.$KV.agent-inference.iris"
+        );
+    }
+
+    #[test]
+    fn a_full_value_round_trips() {
+        let value = AgentInference {
+            model: Some("opus".to_owned()),
+            effort: Some(Effort::Xhigh),
+            subagent_models: Some(vec!["haiku".to_owned(), "sonnet".to_owned()]),
+        };
+        let encoded = serde_json::to_string(&value).unwrap();
+        assert_eq!(
+            encoded,
+            r#"{"model":"opus","effort":"xhigh","subagent_models":["haiku","sonnet"]}"#
+        );
+        assert_eq!(
+            serde_json::from_str::<AgentInference>(&encoded).unwrap(),
+            value
+        );
+    }
+
+    /// An unset field is left out on the wire, and `{}` is every field unset.
+    #[test]
+    fn unset_fields_are_absent() {
+        let value = AgentInference {
+            effort: Some(Effort::Low),
+            ..AgentInference::default()
+        };
+        assert_eq!(
+            serde_json::to_string(&value).unwrap(),
+            r#"{"effort":"low"}"#
+        );
+        assert_eq!(
+            serde_json::from_str::<AgentInference>("{}").unwrap(),
+            AgentInference::default()
+        );
+    }
+
+    #[test]
+    fn every_effort_is_stored_as_its_cli_spelling() {
+        for effort in [
+            Effort::Low,
+            Effort::Medium,
+            Effort::High,
+            Effort::Xhigh,
+            Effort::Max,
+        ] {
+            assert_eq!(
+                serde_json::to_string(&effort).unwrap(),
+                format!("\"{}\"", effort.as_str())
+            );
+        }
+    }
+
+    #[test]
+    fn an_unknown_effort_is_refused() {
+        for doc in [
+            r#"{"effort":"extreme"}"#,
+            r#"{"effort":"High"}"#,
+            r#"{"effort":""}"#,
+        ] {
+            assert!(
+                serde_json::from_str::<AgentInference>(doc).is_err(),
+                "{doc}"
+            );
+        }
+    }
+
+    /// Fields a newer controller adds must not cost an older agent its value.
+    #[test]
+    fn decoding_tolerates_unknown_fields() {
+        let decoded: AgentInference =
+            serde_json::from_str(r#"{"model":"opus","temperature":1}"#).unwrap();
+        assert_eq!(decoded.model.as_deref(), Some("opus"));
+    }
+}

swarm-queue-client/src/lib.rs

@@ -217,6 +217,11 @@ pub mod agent_icon;
 /// writing needs `kv` and reading needs `agent-schedules`.
 pub mod agent_schedules;
 
+/// The bucket swarm-controller sets each agent's model, effort and subagent
+/// models in — one key per agent, read by that agent alone. The names are
+/// unconditional; writing needs `kv` and reading needs `agent-inference`.
+pub mod agent_inference;
+
 #[cfg(feature = "kv")]
 mod kv;
 

swarmctl/src/inference.rs

@@ -0,0 +1,182 @@
+//! `swarmctl agent inference` — the model, effort and subagent models the
+//! swarm sets for one agent, through swarm-controller's
+//! `/api/agents/{name}/inference` routes.
+//!
+//! The agent picks a change up on its next read of its key, within ten
+//! seconds, and its next turn runs on it. A field the swarm does not set
+//! leaves the agent on its own value.
+//!
+//! The shape below mirrors the controller's `InferenceSettings`, for the
+//! reason [`crate::agent`]'s module doc gives.
+
+use std::path::Path;
+
+use anyhow::{Context as _, Result, bail};
+use serde::{Deserialize, Serialize};
+
+use crate::agent::{parse_ident, send};
+
+/// One agent's settings; an absent field is not set at swarm level.
+#[derive(Debug, Default, PartialEq, Eq, Serialize, Deserialize)]
+struct Settings {
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    model: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    effort: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    subagent_models: Option<Vec<String>>,
+}
+
+/// What `agent inference set` was asked to change.
+pub(crate) struct SetFlags {
+    pub model: Option<String>,
+    pub effort: Option<String>,
+    pub subagent_models: Option<Vec<String>>,
+}
+
+impl SetFlags {
+    /// `current` with each given flag written over it.
+    fn merge(self, current: Settings) -> Result<Settings> {
+        if self.model.is_none() && self.effort.is_none() && self.subagent_models.is_none() {
+            bail!("nothing to change: pass --model, --effort or --subagent-models");
+        }
+        Ok(Settings {
+            model: self.model.or(current.model),
+            effort: self.effort.or(current.effort),
+            subagent_models: self.subagent_models.or(current.subagent_models),
+        })
+    }
+}
+
+fn runtime() -> Result<tokio::runtime::Runtime> {
+    tokio::runtime::Builder::new_current_thread()
+        .enable_io()
+        .build()
+        .context("starting a tokio runtime for the controller request")
+}
+
+/// The agent's settings path. The name is checked first, so it cannot
+/// reshape the URI.
+fn member(agent: &str) -> Result<String> {
+    let agent = parse_ident(agent, "agent name")?;
+    Ok(format!("/api/agents/{agent}/inference"))
+}
+
+/// Run `swarmctl agent inference get`.
+pub(crate) fn get(socket: &Path, agent: &str) -> Result<()> {
+    let uri = member(agent)?;
+    let current: Settings = runtime()?.block_on(send(
+        socket,
+        hyper::Method::GET,
+        &uri,
+        None::<&()>,
+        "inference-get",
+    ))?;
+    print!("{}", render(&current));
+    Ok(())
+}
+
+/// Run `swarmctl agent inference set`: read the agent's settings, write the
+/// flags over them, and store the result.
+pub(crate) fn set(socket: &Path, agent: &str, flags: SetFlags) -> Result<()> {
+    let uri = member(agent)?;
+    let rt = runtime()?;
+    let current: Settings = rt.block_on(send(
+        socket,
+        hyper::Method::GET,
+        &uri,
+        None::<&()>,
+        "inference-get",
+    ))?;
+    let wanted = flags.merge(current)?;
+    let stored: Settings = rt.block_on(send(
+        socket,
+        hyper::Method::PUT,
+        &uri,
+        Some(&wanted),
+        "inference-set",
+    ))?;
+    print!("{}", render(&stored));
+    Ok(())
+}
+
+/// Run `swarmctl agent inference clear`.
+pub(crate) fn clear(socket: &Path, agent: &str) -> Result<()> {
+    let uri = member(agent)?;
+    runtime()?.block_on(send::<(), ()>(
+        socket,
+        hyper::Method::DELETE,
+        &uri,
+        None,
+        "inference-clear",
+    ))?;
+    println!("cleared; {agent} goes back to its own model and effort");
+    Ok(())
+}
+
+/// One line per field, naming the unset ones.
+fn render(s: &Settings) -> String {
+    const UNSET: &str = "(not set; the agent's own)";
+    format!(
+        "model: {}\neffort: {}\nsubagent_models: {}\n",
+        s.model.as_deref().unwrap_or(UNSET),
+        s.effort.as_deref().unwrap_or(UNSET),
+        s.subagent_models
+            .as_ref()
+            .map_or_else(|| UNSET.to_owned(), |m| m.join(", "))
+    )
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn flags() -> SetFlags {
+        SetFlags {
+            model: None,
+            effort: None,
+            subagent_models: None,
+        }
+    }
+
+    /// A flag replaces its field and every other field keeps what is stored.
+    #[test]
+    fn set_merges_into_the_stored_value() {
+        let current: Settings =
+            serde_json::from_str(r#"{"model":"sonnet","subagent_models":["haiku"]}"#).unwrap();
+        let merged = SetFlags {
+            effort: Some("high".to_owned()),
+            ..flags()
+        }
+        .merge(current)
+        .unwrap();
+        assert_eq!(
+            serde_json::to_string(&merged).unwrap(),
+            r#"{"model":"sonnet","effort":"high","subagent_models":["haiku"]}"#
+        );
+        let merged = SetFlags {
+            model: Some("opus".to_owned()),
+            ..flags()
+        }
+        .merge(merged)
+        .unwrap();
+        assert_eq!(merged.model.as_deref(), Some("opus"));
+        assert_eq!(merged.effort.as_deref(), Some("high"));
+    }
+
+    #[test]
+    fn set_with_no_flag_is_refused() {
+        assert!(flags().merge(Settings::default()).is_err());
+    }
+
+    #[test]
+    fn an_empty_value_renders_every_field_as_unset() {
+        let row = render(&serde_json::from_str("{}").unwrap());
+        assert_eq!(row.lines().count(), 3);
+        assert!(
+            row.lines()
+                .all(|l| l.ends_with("(not set; the agent's own)")),
+            "{row}"
+        );
+    }
+}

swarmctl/src/main.rs

@@ -29,6 +29,7 @@
 
 mod agent;
 mod forge;
+mod inference;
 mod schedule;
 mod users;
 
@@ -210,6 +211,55 @@ enum AgentVerb {
     /// Queues and returns, the same way `agent create` does — watch the
     /// swarm UI's job view for the outcome.
     MintForgeToken(AgentMintForgeTokenArgs),
+    /// Show or change the model, effort and subagent models the swarm sets
+    /// for an agent.
+    ///
+    /// A field the swarm sets wins over the agent's own value (from its
+    /// `/model` or `/effort` pick or its agent.nix) from its next turn on;
+    /// the agent reads its settings every ten seconds. A field the swarm
+    /// doesn't set, or clears, leaves the agent on its own value.
+    Inference {
+        #[command(subcommand)]
+        command: InferenceVerb,
+    },
+}
+
+#[derive(Subcommand)]
+enum InferenceVerb {
+    /// Show what the swarm sets for an agent, naming the fields it doesn't
+    /// set.
+    Get(InferenceAgentArgs),
+    /// Set fields for an agent. Only the flags given change; the others keep
+    /// what's stored.
+    Set(InferenceSetArgs),
+    /// Clear every field for an agent, so it goes back to its own values.
+    Clear(InferenceAgentArgs),
+}
+
+#[derive(Args)]
+struct InferenceAgentArgs {
+    /// Agent name.
+    agent: String,
+    #[command(flatten)]
+    socket: ControllerSocketArg,
+}
+
+#[derive(Args)]
+struct InferenceSetArgs {
+    /// Agent name. A hive must declare it.
+    agent: String,
+    /// The model its turns run on, as claude's `--model` takes it.
+    #[arg(long, value_name = "MODEL")]
+    model: Option<String>,
+    /// The reasoning effort its turns run at.
+    #[arg(long, value_name = "LEVEL", value_parser = ["low", "medium", "high", "xhigh", "max"])]
+    effort: Option<String>,
+    /// The models it may start subagents on, comma-separated. Stored and
+    /// shown; the agent doesn't apply it yet.
+    #[arg(long, value_name = "MODELS", value_delimiter = ',')]
+    subagent_models: Option<Vec<String>>,
+    #[command(flatten)]
+    socket: ControllerSocketArg,
 }
 
 #[derive(Args)]
@@ -496,6 +546,9 @@ fn main() -> Result<()> {
             let socket = path_from(args.controller_socket, "SWARM_CONTROLLER_SOCKET")?;
             agent::mint_forge_token(&socket, &args.name)
         }
+        Verb::Agent {
+            command: AgentVerb::Inference { command },
+        } => inference(command),
         // Same socket-resolution reasoning as `Create` above.
         Verb::Forge {
             command: ForgeVerb::MakeAdmin(args),
@@ -544,6 +597,22 @@ fn main() -> Result<()> {
     }
 }
 
+fn inference(command: InferenceVerb) -> Result<()> {
+    match command {
+        InferenceVerb::Get(a) => inference::get(&a.socket.resolve()?, &a.agent),
+        InferenceVerb::Set(a) => inference::set(
+            &a.socket.resolve()?,
+            &a.agent,
+            inference::SetFlags {
+                model: a.model,
+                effort: a.effort,
+                subagent_models: a.subagent_models,
+            },
+        ),
+        InferenceVerb::Clear(a) => inference::clear(&a.socket.resolve()?, &a.agent),
+    }
+}
+
 fn schedule(command: ScheduleVerb) -> Result<()> {
     match command {
         ScheduleVerb::Add(a) => schedule::add(