use async_trait::async_trait;
use serde::Deserialize;
use serde_json::json;
use std::sync::Arc;
use uuid::Uuid;
use bamboo_agent_core::tools::{Tool, ToolError, ToolExecutionContext, ToolResult};
use bamboo_domain::session::runtime_state::ChildWaitPolicy;
use bamboo_domain::subagent::SubagentProfileRegistry;
use bamboo_domain::ReasoningEffort;
use bamboo_engine::session_app::child_session::{
self, ChildSessionError, ChildSessionPort, CreateChildInput, ModelCatalogPort,
SubagentResolutionPort,
};
// ---------------------------------------------------------------------------
// Args enum
// ---------------------------------------------------------------------------
#[derive(Debug, Deserialize)]
#[serde(tag = "action", rename_all = "snake_case")]
enum SubAgentArgs {
Create {
#[serde(default)]
title: Option<String>,
#[serde(default)]
description: String,
#[serde(default)]
responsibility: Option<String>,
prompt: String,
/// Subagent profile/role. Optional: defaults to `general-purpose` when
/// omitted or empty, so a missing value never hard-fails a create.
#[serde(default)]
subagent_type: Option<String>,
/// Working directory for the child. Optional: defaults to the parent
/// session's workspace when omitted.
#[serde(default)]
workspace: Option<String>,
#[serde(default)]
auto_run: Option<bool>,
/// When `true`, the parent suspends immediately and waits for THIS child
/// to finish (the legacy one-shot behavior). Defaults to `false`:
/// `create` runs the child in the background and returns right away so
/// the parent can spawn more children. Call `action=wait` once, after
/// spawning everything, to suspend until they finish.
#[serde(default)]
wait: Option<bool>,
/// Optional reasoning effort for the child session. When omitted,
/// the child stays at `None` so the provider's default applies
/// (it does NOT inherit the parent's reasoning_effort). The LLM
/// should pass an explicit value (e.g. `"low"` for cheap fan-outs,
/// `"high"`/`"max"` for hard reasoning) when it has a preference.
#[serde(default)]
reasoning_effort: Option<ReasoningEffort>,
/// Optional explicit model for the child, `"provider:model"`
/// (e.g. `"anthropic:claude-sonnet-4-6"`) or a bare model id (resolved
/// against the parent's provider, falling back to the default
/// provider). Takes precedence over per-`subagent_type` model routing.
/// Call `list_models` to see what is available.
#[serde(default)]
model: Option<String>,
/// Lifecycle: `"oneshot"` (default) creates a fresh throwaway child for
/// this task. `"resident"` reuses a long-lived agent identified by
/// `name` (scoped to this conversation): the FIRST resident create spins
/// one up; later creates with the same `name` route the new task to that
/// same agent instead of spawning another — so repeated similar work
/// (e.g. an "essayist" handling many essays) stays one agent/one entry,
/// not N. Use resident for recurring task types; one-shot for
/// independent throwaway work.
#[serde(default)]
lifecycle: Option<String>,
/// Resident reuse key (required when `lifecycle="resident"`; defaults to
/// `subagent_type`). The stable name of the resident agent to create or
/// reuse, e.g. `"essayist"`.
#[serde(default)]
name: Option<String>,
/// For a resident agent, how successive tasks treat prior context:
/// `"reset"` (default — each task is independent, prior context cleared)
/// or `"accumulate"` (the agent remembers earlier tasks). Set on first
/// create; honored on reuse.
#[serde(default)]
context: Option<String>,
},
/// Suspend the parent run until its background child sessions finish.
///
/// Spawn children with `action=create` (which no longer suspends), then call
/// this once. By default it waits on every currently-active child; pass
/// explicit `child_session_ids` to wait on a subset. If no children are
/// active it is a no-op (the parent keeps running).
Wait {
#[serde(default)]
child_session_ids: Option<Vec<String>>,
/// Wait policy: `all` (default) resumes when every tracked child is
/// terminal; `any` resumes on the first; `first_error` resumes early on
/// any error/timeout/cancel.
#[serde(default)]
wait_for: Option<ChildWaitPolicy>,
},
List,
Get {
child_session_id: String,
},
Update {
child_session_id: String,
#[serde(default)]
title: Option<String>,
#[serde(default)]
responsibility: Option<String>,
#[serde(default)]
prompt: Option<String>,
#[serde(default)]
subagent_type: Option<String>,
#[serde(default)]
reset_after_update: Option<bool>,
#[serde(default)]
auto_run: Option<bool>,
/// Optional reasoning effort to apply to the existing child session.
/// `Some(level)` overrides the current value; `None` (the default)
/// leaves it unchanged.
#[serde(default)]
reasoning_effort: Option<ReasoningEffort>,
},
Run {
child_session_id: String,
#[serde(default)]
reset_to_last_user: Option<bool>,
},
SendMessage {
child_session_id: String,
message: String,
#[serde(default)]
auto_run: Option<bool>,
#[serde(default)]
interrupt_running: Option<bool>,
},
Cancel {
child_session_id: String,
},
Delete {
child_session_id: String,
},
/// Enumerate the available subagent profiles (built-ins plus any
/// user/project overrides). Read-only; does not touch any session.
/// Useful both for the LLM (to discover roles before calling
/// `create`) and for the frontend (to populate a role dropdown).
ListProfiles,
/// Enumerate the models the parent can pin a child to via
/// `create.model`. Read-only; best-effort per configured provider.
ListModels,
}
// ---------------------------------------------------------------------------
// Normalization helpers (ported from legacy SpawnSessionTool)
// ---------------------------------------------------------------------------
fn normalize_required_text(value: Option<String>, field_name: &str) -> Result<String, ToolError> {
let Some(value) = value else {
return Err(ToolError::InvalidArguments(format!(
"{field_name} must be non-empty"
)));
};
let trimmed = value.trim();
if trimmed.is_empty() {
return Err(ToolError::InvalidArguments(format!(
"{field_name} must be non-empty"
)));
}
Ok(trimmed.to_string())
}
fn normalize_title(title: Option<String>, legacy_description: String) -> Result<String, ToolError> {
let title = title.and_then(|value| {
let trimmed = value.trim();
if trimmed.is_empty() {
None
} else {
Some(trimmed.to_string())
}
});
let legacy_description = {
let trimmed = legacy_description.trim();
if trimmed.is_empty() {
None
} else {
Some(trimmed.to_string())
}
};
normalize_required_text(title.or(legacy_description), "title")
}
fn tool_result(value: serde_json::Value) -> Result<ToolResult, ToolError> {
Ok(ToolResult {
success: true,
result: value.to_string(),
display_preference: Some("Collapsible".to_string()),
images: Vec::new(),
})
}
fn waiting_for_children_tool_result(mut value: serde_json::Value) -> Result<ToolResult, ToolError> {
if let Some(object) = value.as_object_mut() {
object.insert("runtime_control".to_string(), json!("waiting_for_children"));
// Don't clobber a caller-provided policy (e.g. action=wait with
// wait_for=any); only default it when absent.
object
.entry("wait_for".to_string())
.or_insert_with(|| json!("all"));
object.insert(
"note".to_string(),
json!("Child session queued. The parent run is suspended and will resume automatically when the child finishes or times out."),
);
}
Ok(ToolResult {
success: true,
result: value.to_string(),
display_preference: Some("runtime_control:waiting_for_children".to_string()),
images: Vec::new(),
})
}
/// Map a `ChildSessionError` to a `ToolError`.
fn tool_error_from_child_session(error: ChildSessionError) -> ToolError {
match error {
ChildSessionError::NotFound(id) => ToolError::Execution(format!("session not found: {id}")),
ChildSessionError::NotRootSession(id) => {
ToolError::Execution(format!("session is not a root session: {id}"))
}
ChildSessionError::InvalidArguments(msg) => ToolError::InvalidArguments(msg),
ChildSessionError::Execution(msg) => ToolError::Execution(msg),
other => ToolError::Execution(other.to_string()),
}
}
// ---------------------------------------------------------------------------
// Tool struct
// ---------------------------------------------------------------------------
pub struct SubAgentTool {
/// Child-session CRUD/lifecycle operations (load/save/run/cancel/…).
sessions: Arc<dyn ChildSessionPort>,
/// Subagent-type resolution (model, runtime metadata, prompt, active ids).
resolver: Arc<dyn SubagentResolutionPort>,
/// Registry consulted by `action=list_profiles`. Held as `Arc` so the
/// tool stays cheap to clone and share across executors.
profiles: Arc<SubagentProfileRegistry>,
/// Optional model catalog consulted by `action=list_models` and used to
/// resolve a bare `create.model` id to a provider. `None` keeps the tool
/// constructible without a live provider registry (tests, embedded use).
catalog: Option<Arc<dyn ModelCatalogPort>>,
}
impl SubAgentTool {
pub fn new(
sessions: Arc<dyn ChildSessionPort>,
resolver: Arc<dyn SubagentResolutionPort>,
profiles: Arc<SubagentProfileRegistry>,
) -> Self {
Self {
sessions,
resolver,
profiles,
catalog: None,
}
}
/// Attach a model catalog, enabling `action=list_models` and bare-model
/// resolution for `create.model`.
pub fn with_model_catalog(mut self, catalog: Arc<dyn ModelCatalogPort>) -> Self {
self.catalog = Some(catalog);
self
}
}
/// Parse an explicit `create.model` spec into a `ProviderModelRef`.
///
/// `"provider:model"` is explicit; a bare model id falls back to the parent
/// session's provider, then the catalog's default provider.
fn parse_model_spec(
spec: &str,
parent: &bamboo_agent_core::Session,
default_provider: Option<String>,
) -> Result<bamboo_domain::ProviderModelRef, ToolError> {
let spec = spec.trim();
if spec.is_empty() {
return Err(ToolError::InvalidArguments(
"model must be non-empty when provided".to_string(),
));
}
if let Some((provider, model)) = spec.split_once(':') {
let (provider, model) = (provider.trim(), model.trim());
if provider.is_empty() || model.is_empty() {
return Err(ToolError::InvalidArguments(format!(
"model '{spec}' must be 'provider:model' with both parts non-empty"
)));
}
return Ok(bamboo_domain::ProviderModelRef::new(provider, model));
}
// Bare model id: inherit the parent's provider, else the default provider.
let provider = parent
.model_ref
.as_ref()
.map(|r| r.provider.clone())
.filter(|p| !p.trim().is_empty())
.or(default_provider)
.ok_or_else(|| {
ToolError::InvalidArguments(format!(
"model '{spec}' has no provider prefix and no default provider is known; \
use 'provider:model' (see action=list_models)"
))
})?;
Ok(bamboo_domain::ProviderModelRef::new(provider, spec))
}
#[async_trait]
impl Tool for SubAgentTool {
fn name(&self) -> &str {
"SubAgent"
}
fn description(&self) -> &str {
"Create, inspect, and manage child sessions for explicitly requested delegated, parallel, or sub-agent work. A child session runs independently under the current root session with its own conversation context, can use a specialized subagent profile, streams progress back to the parent via sub_agent_* events, and can be reopened from the Sub-agents panel. \
PARALLEL FAN-OUT (important): action=create now runs the child in the BACKGROUND and returns immediately WITHOUT suspending the parent. To launch several agents in parallel, call create once per child (ideally several creates in a single turn), then call action=wait ONCE to suspend until they finish. Do NOT pass wait=true on each create for parallel work — that would serialize them (suspend after the first). action=wait defaults to waiting on every active child; if you forget to call it, the runtime auto-waits at the end of the turn so results are never lost. \
Use list/get to inspect existing children; use update/run/send_message/cancel/delete to manage existing children; use list_profiles to enumerate subagent roles. Use only when the user explicitly asks for delegation/parallelism or when a side task would otherwise flood the main context. Do not use for simple one-step tasks. Child sessions cannot spawn nested child sessions. IMPORTANT: When a child fails or needs redirection, prefer send_message over creating a duplicate child. Use list before create to avoid spawning redundant children."
}
fn parameters_schema(&self) -> serde_json::Value {
json!({
"type": "object",
"properties": {
"action": {
"type": "string",
"enum": ["create", "wait", "list", "get", "update", "run", "send_message", "cancel", "delete", "list_profiles", "list_models"],
"description": "Sub-agent lifecycle operation. To run work in parallel: call create once per child (this no longer suspends the parent — children run in the background), then call wait ONCE to suspend until they all finish. Use list/get to inspect; update/run/send_message/cancel/delete to manage existing children; list_profiles to enumerate available subagent roles before choosing subagent_type; list_models to enumerate the models you can pin a child to via create.model. \
A create call requires: title, responsibility, prompt, and subagent_type (workspace and subagent_type are optional and default to the parent's workspace / general-purpose). EXAMPLE create: {\"action\":\"create\",\"subagent_type\":\"researcher\",\"title\":\"Analyze auth module\",\"responsibility\":\"Map the auth flow and list its public API\",\"prompt\":\"Read crates/auth/src/lib.rs, summarize the login flow, and list every pub fn.\",\"workspace\":\"/abs/path/to/repo\"}. Then EXAMPLE wait: {\"action\":\"wait\"}."
},
"child_session_id": {
"type": "string",
"description": "Existing child session id. Required for get/update/run/send_message/cancel/delete."
},
"child_session_ids": {
"type": "array",
"items": { "type": "string" },
"description": "For wait: optional explicit subset of child sessions to wait on. Omit to wait on every currently-active child."
},
"wait_for": {
"type": "string",
"enum": ["all", "any", "first_error"],
"description": "For wait: resume policy. all (default) resumes when every tracked child is done; any resumes on the first; first_error resumes early on any error/timeout/cancel."
},
"wait": {
"type": "boolean",
"description": "For create: if true, suspend immediately and wait for just THIS child (legacy one-shot behavior). Defaults to false — create returns immediately and the child runs in the background; suspend later with action=wait."
},
"title": {
"type": "string",
"description": "Short title for a new or updated child session. Required for create. Displayed in the Sub-agents panel."
},
"description": {
"type": "string",
"description": "Legacy alias of title; prefer title."
},
"responsibility": {
"type": "string",
"description": "Single explicit responsibility for the child session. Required for create. Keep this narrow and non-overlapping with other child sessions."
},
"prompt": {
"type": "string",
"description": "Detailed task instructions, context, constraints, and expected output for the child session. Required for create; optional for update."
},
"subagent_type": {
"type": "string",
"description": "For create: the specialized child agent profile/role, e.g. general-purpose, researcher, coder, plan. Use plan/researcher for read-only exploration and coder/general-purpose for implementation when allowed. Optional — omitting it defaults to general-purpose — but you should pick the most fitting role. Call list_profiles to see the available roles."
},
"workspace": {
"type": "string",
"description": "For create: absolute path to the child session's working directory for file operations. Optional — defaults to the parent session's workspace when omitted."
},
"auto_run": {
"type": "boolean",
"description": "For create/send_message/update: whether to enqueue the child session immediately. Defaults to true for create/send_message and false for update."
},
"reset_after_update": {
"type": "boolean",
"description": "For update: whether to truncate messages after refreshed assignment. Defaults to true."
},
"reset_to_last_user": {
"type": "boolean",
"description": "For run: whether to truncate messages after the last user message before rerun. Defaults to true."
},
"message": {
"type": "string",
"description": "Follow-up instruction to append as a new user message for send_message. Required for send_message."
},
"interrupt_running": {
"type": "boolean",
"description": "For send_message/cancel: if true, cancel a currently running child session before appending or returning. Defaults to false for send_message. When false on a running child, the message is queued and will be picked up at the next turn boundary without canceling progress."
},
"reasoning_effort": {
"type": "string",
"enum": ["low", "medium", "high", "xhigh", "max"],
"description": "For create/update: reasoning effort level applied to the child session's own LLM calls. Use \"low\" for trivial fan-outs (e.g. simple lookups), \"medium\"/\"high\" for normal coding/analysis, \"xhigh\"/\"max\" for deep reasoning tasks. Omit to leave at provider default; the child does NOT inherit the parent's reasoning_effort."
},
"model": {
"type": "string",
"description": "For create: explicit model for the child as 'provider:model' (e.g. 'anthropic:claude-sonnet-4-6'), or a bare model id to use the parent's provider. Takes precedence over per-role model routing. Pick a cheaper/faster model for simple fan-outs and a stronger model for hard reasoning. Call list_models first to see what is available; omit to use the configured default for the chosen subagent_type."
},
"lifecycle": {
"type": "string",
"enum": ["oneshot", "resident"],
"description": "For create: 'oneshot' (default) spins up a fresh throwaway child for this task. 'resident' reuses ONE long-lived agent (identified by 'name', scoped to this conversation) across many tasks — the first resident create spins it up, later creates with the same name route the new task to that same agent instead of spawning another. Use resident for recurring task types (e.g. an 'essayist' that writes many essays — one agent, one panel entry, not N); use oneshot for independent throwaway work."
},
"name": {
"type": "string",
"description": "For create with lifecycle=resident: the resident agent's stable reuse key, e.g. 'essayist'. Required to reuse a resident; defaults to subagent_type when omitted. Reusing the same name routes the new task to the existing resident agent."
},
"context": {
"type": "string",
"enum": ["reset", "accumulate"],
"description": "For create with lifecycle=resident: how the resident treats prior tasks. 'reset' (default) makes each task independent (clears prior context). 'accumulate' makes the agent remember earlier tasks (useful for a researcher building up knowledge). Set on first create; honored on reuse."
}
},
"required": ["action"],
"additionalProperties": false
})
}
async fn execute(&self, args: serde_json::Value) -> Result<ToolResult, ToolError> {
self.execute_with_context(args, ToolExecutionContext::none("tool_call"))
.await
}
async fn execute_with_context(
&self,
args: serde_json::Value,
ctx: ToolExecutionContext<'_>,
) -> Result<ToolResult, ToolError> {
let parent_session_id = ctx.session_id.ok_or_else(|| {
ToolError::Execution("SubAgent requires a session_id in tool context".to_string())
})?;
// Backward compatibility: legacy SubAgent calls did not include an
// "action" field and always meant "create". If action is missing,
// default to "create" before deserializing the tagged enum.
let mut args = args;
if args.get("action").is_none() {
args["action"] = json!("create");
}
let parsed: SubAgentArgs = serde_json::from_value(args).map_err(|error| {
ToolError::InvalidArguments(format!("Invalid SubAgent args: {error}"))
})?;
// `list_profiles` is read-only and operates purely on the
// in-memory profile registry, so we short-circuit before doing
// any session lookup. This also lets the LLM call `list_profiles`
// safely from any context (root or otherwise).
if let SubAgentArgs::ListProfiles = parsed {
return tool_result(self.list_profiles_payload());
}
// `list_models` is likewise read-only and session-independent.
if let SubAgentArgs::ListModels = parsed {
let Some(catalog) = self.catalog.as_ref() else {
return Err(ToolError::Execution(
"model catalog is not configured on this server".to_string(),
));
};
let providers = catalog.list_models().await;
return tool_result(json!({
"default_provider": catalog.default_provider(),
"providers": providers,
"usage": "Pass create.model as 'provider:model' (or a bare model id to use the parent's provider).",
}));
}
let parent = self
.sessions
.as_ref()
.load_root_session(parent_session_id)
.await
.map_err(tool_error_from_child_session)?;
match parsed {
SubAgentArgs::Create {
title,
description,
responsibility,
prompt,
subagent_type,
workspace,
auto_run,
wait,
reasoning_effort,
model,
lifecycle,
name,
context,
} => {
let title = normalize_title(title, description)?;
let responsibility = normalize_required_text(responsibility, "responsibility")?;
let prompt = normalize_required_text(Some(prompt), "prompt")?;
// subagent_type is optional: an omitted/blank value falls back to
// the catch-all `general-purpose` profile (the same fallback
// resolve_subagent_prompt already applies), so the model never
// hard-fails a create just for leaving the role unspecified.
let subagent_type = subagent_type
.map(|value| value.trim().to_string())
.filter(|value| !value.is_empty())
.unwrap_or_else(|| "general-purpose".to_string());
// workspace is optional: default to the parent's workspace.
let workspace = workspace
.map(|value| value.trim().to_string())
.filter(|value| !value.is_empty())
.or_else(|| parent.workspace.clone())
.ok_or_else(|| {
ToolError::InvalidArguments(
"workspace must be non-empty (parent has no workspace to inherit)"
.to_string(),
)
})?;
if parent.model.trim().is_empty() {
return Err(ToolError::Execution(
"parent session model is empty".to_string(),
));
}
let should_auto_run = auto_run.unwrap_or(true);
// Resident routing: `lifecycle="resident"` reuses the existing
// resident agent of the same `name` in this root tree (one stable
// agent/entry for recurring work) instead of minting a new child.
// `reset` (default) replaces the resident's task and reruns it;
// `accumulate` appends the task to its history. The FIRST resident
// create (none found yet) falls through to a normal create tagged
// as resident.
let is_resident = lifecycle.as_deref().map(str::trim) == Some("resident");
let resident_name = is_resident.then(|| {
name.as_deref()
.map(str::trim)
.filter(|n| !n.is_empty())
.map(str::to_string)
.unwrap_or_else(|| subagent_type.clone())
});
let resident_context = context
.as_deref()
.map(str::trim)
.filter(|c| matches!(*c, "reset" | "accumulate"))
.unwrap_or("reset")
.to_string();
let existing_resident = match resident_name.as_deref() {
Some(rname) => {
// Children are created with `root_session_id == parent.id`
// (a flat tree; children cannot spawn children), so the
// parent's id is the tree root key for the lookup.
self.sessions.find_resident_child(&parent.id, rname).await
}
None => None,
};
let (child_session_id, child_model, reused) =
if let Some(existing_id) = existing_resident {
// A resident processes tasks serially. If it is still running
// a previous task, stop it first: otherwise `reset` would
// truncate the session while the runner writes back (a
// corrupting race + a possible duplicate spawn job), and
// `accumulate` would queue a message into a run that may end
// before picking it up (the task would never execute). After
// cancel the resident is idle, so both paths apply cleanly.
if self.sessions.is_child_running(&existing_id).await {
self.sessions
.cancel_child_run_and_wait(&existing_id)
.await
.map_err(tool_error_from_child_session)?;
}
// Reuse: reset => update (truncate + new task) then rerun;
// accumulate => send the task as a new message (auto-runs).
if resident_context == "accumulate" {
child_session::send_message_to_child_action(
self.sessions.as_ref(),
&parent,
existing_id.clone(),
format!("# Task: {title}\n\n{responsibility}\n\n{prompt}"),
Some(should_auto_run),
Some(false),
)
.await
.map_err(tool_error_from_child_session)?;
} else {
child_session::update_child_action(
self.sessions.as_ref(),
&parent.id,
existing_id.clone(),
Some(title.clone()),
Some(responsibility.clone()),
Some(prompt.clone()),
Some(subagent_type.clone()),
Some(true),
reasoning_effort,
)
.await
.map_err(tool_error_from_child_session)?;
if should_auto_run {
let child = self
.sessions
.load_child_for_parent(&parent.id, &existing_id)
.await
.map_err(tool_error_from_child_session)?;
self.sessions
.enqueue_child_run(&parent, &child)
.await
.map_err(tool_error_from_child_session)?;
}
}
let model = self
.sessions
.load_child_for_parent(&parent.id, &existing_id)
.await
.map(|c| c.model)
.unwrap_or_default();
(existing_id, model, true)
} else {
let child_id = Uuid::new_v4().to_string();
// Model precedence: explicit `model` arg > per-subagent_type
// routing (resolver) > engine defaults (None).
let model_ref_override =
match model.as_deref().map(str::trim).filter(|m| !m.is_empty()) {
Some(spec) => Some(parse_model_spec(
spec,
&parent,
self.catalog.as_ref().map(|c| c.default_provider()),
)?),
None => self.resolver.resolve_subagent_model(&subagent_type).await,
};
let model_override = model_ref_override
.as_ref()
.map(|model_ref| model_ref.model.clone());
let runtime_metadata =
self.resolver.resolve_runtime_metadata(&subagent_type).await;
let system_prompt_override =
Some(self.resolver.resolve_subagent_prompt(&subagent_type));
let result = child_session::create_child_action(
self.sessions.as_ref(),
CreateChildInput {
parent_session: parent.clone(),
child_id: child_id.clone(),
title: title.clone(),
responsibility: responsibility.clone(),
assignment_prompt: prompt.clone(),
subagent_type: subagent_type.clone(),
workspace: workspace.clone(),
model_override,
model_ref_override,
runtime_metadata,
system_prompt_override,
auto_run: should_auto_run,
reasoning_effort,
lifecycle: resident_name.as_ref().map(|_| "resident".to_string()),
resident_name: resident_name.clone(),
resident_context: resident_name
.as_ref()
.map(|_| resident_context.clone()),
},
)
.await
.map_err(tool_error_from_child_session)?;
(result.child_session_id, result.model, false)
};
// Ensure index entry is visible immediately (best-effort).
self.sessions.ensure_child_indexed(&child_session_id).await;
ctx.emit_tool_token(if reused {
format!("Reused resident agent: {child_session_id}")
} else {
format!("Spawned child session: {child_session_id}")
})
.await;
// `wait=true` preserves the legacy one-shot behavior: register a
// wait for THIS child and suspend now. Default (`wait=false`) runs
// the child in the background and returns immediately, so the
// parent can keep spawning; it suspends later via `action=wait`.
let should_wait = should_auto_run && wait.unwrap_or(false);
if should_wait {
self.sessions
.register_parent_wait_for_child(&parent.id, &child_session_id, None)
.await
.map_err(tool_error_from_child_session)?;
}
let status = if !should_auto_run {
"created"
} else if should_wait {
"queued"
} else {
"running_in_background"
};
let note = if should_wait {
"Child session queued (typically 30-120 seconds); the parent is suspended until it finishes. Use send_message (not create) to correct a child in place."
} else if should_auto_run {
"Child session is running in the background (typically 30-120 seconds). Spawn any other children you need, then call action=wait once to suspend until they finish. Use send_message (not create) to correct a child in place."
} else {
"Child session created (not started). Use action=run to start it. Use send_message (not create) to correct a child in place."
};
let payload = json!({
"title": title.clone(),
"description": title,
"responsibility": responsibility,
"prompt": prompt,
"subagent_type": subagent_type,
"child_session_id": child_session_id,
"parent_session_id": parent_session_id,
"model": child_model,
"reasoning_effort": reasoning_effort.map(|effort| effort.as_str()),
"status": status,
"lifecycle": resident_name.as_ref().map(|_| "resident"),
"resident_name": resident_name.clone(),
"reused": reused,
"note": note,
});
if should_wait {
waiting_for_children_tool_result(payload)
} else {
tool_result(payload)
}
}
SubAgentArgs::Wait {
child_session_ids,
wait_for,
} => {
let policy = wait_for.unwrap_or(ChildWaitPolicy::All);
// Default to every currently-active child; honor an explicit
// subset when provided.
let targets = match child_session_ids {
Some(ids) if !ids.is_empty() => ids,
_ => self.sessions.active_child_ids(&parent.id).await,
};
if targets.is_empty() {
// Nothing to wait on — never register an empty wait (that
// would suspend the parent with no child able to resume it).
return tool_result(json!({
"status": "no_active_children",
"parent_session_id": parent_session_id,
"note": "No active child sessions to wait for; the parent continues running.",
}));
}
let count = self
.sessions
.register_parent_wait_for_children(&parent.id, &targets, policy)
.await
.map_err(tool_error_from_child_session)?;
waiting_for_children_tool_result(json!({
"status": "waiting",
"parent_session_id": parent_session_id,
"child_session_ids": targets,
"wait_for": policy.as_str(),
"waiting_on": count,
}))
}
SubAgentArgs::List => {
let result =
child_session::list_children_action(self.sessions.as_ref(), &parent.id).await;
tool_result(result)
}
SubAgentArgs::Get { child_session_id } => {
let result = child_session::get_child_action(
self.sessions.as_ref(),
&parent.id,
child_session_id,
)
.await
.map_err(tool_error_from_child_session)?;
tool_result(result)
}
SubAgentArgs::Update {
child_session_id,
title,
responsibility,
prompt,
subagent_type,
reset_after_update,
auto_run,
reasoning_effort,
} => {
let result = child_session::update_child_action(
self.sessions.as_ref(),
&parent.id,
child_session_id.clone(),
title,
responsibility,
prompt,
subagent_type,
reset_after_update,
reasoning_effort,
)
.await
.map_err(tool_error_from_child_session)?;
let should_auto_run = auto_run.unwrap_or(false);
if should_auto_run {
let child = self
.sessions
.load_child_for_parent(&parent.id, &child_session_id)
.await
.map_err(tool_error_from_child_session)?;
self.sessions
.enqueue_child_run(&parent, &child)
.await
.map_err(tool_error_from_child_session)?;
// Re-running an existing child keeps its synchronous "wait for
// the answer" semantics: register the wait + suspend. (enqueue
// itself no longer registers — that is now explicit.)
self.sessions
.register_parent_wait_for_child(&parent.id, &child_session_id, None)
.await
.map_err(tool_error_from_child_session)?;
}
if should_auto_run {
waiting_for_children_tool_result(result)
} else {
tool_result(result)
}
}
SubAgentArgs::Run {
child_session_id,
reset_to_last_user,
} => {
let result = child_session::run_child_action(
self.sessions.as_ref(),
&parent,
child_session_id.clone(),
reset_to_last_user,
)
.await
.map_err(tool_error_from_child_session)?;
// `run` keeps the synchronous retry semantics: wait for this child.
self.sessions
.register_parent_wait_for_child(&parent.id, &child_session_id, None)
.await
.map_err(tool_error_from_child_session)?;
waiting_for_children_tool_result(result)
}
SubAgentArgs::SendMessage {
child_session_id,
message,
auto_run,
interrupt_running,
} => {
let should_auto_run = auto_run.unwrap_or(true);
let result = child_session::send_message_to_child_action(
self.sessions.as_ref(),
&parent,
child_session_id.clone(),
message,
auto_run,
interrupt_running,
)
.await
.map_err(tool_error_from_child_session)?;
let queued = should_auto_run
&& result
.get("status")
.and_then(|value| value.as_str())
.is_some_and(|status| status == "queued");
if queued {
// Sending + running keeps synchronous semantics: wait for the
// child's response. (enqueue no longer registers the wait.)
self.sessions
.register_parent_wait_for_child(&parent.id, &child_session_id, None)
.await
.map_err(tool_error_from_child_session)?;
waiting_for_children_tool_result(result)
} else {
tool_result(result)
}
}
SubAgentArgs::Cancel { child_session_id } => {
let result = child_session::cancel_child_action(
self.sessions.as_ref(),
&parent.id,
child_session_id,
)
.await
.map_err(tool_error_from_child_session)?;
tool_result(result)
}
SubAgentArgs::Delete { child_session_id } => {
let result = child_session::delete_child_action(
self.sessions.as_ref(),
&parent.id,
child_session_id,
)
.await
.map_err(tool_error_from_child_session)?;
tool_result(result)
}
// Already short-circuited above; kept here so the match stays
// exhaustive without a wildcard.
SubAgentArgs::ListProfiles => tool_result(self.list_profiles_payload()),
// Handled by the session-independent short-circuit above.
SubAgentArgs::ListModels => unreachable!("list_models short-circuits earlier"),
}
}
}
impl SubAgentTool {
/// Build the JSON payload returned by `action=list_profiles`.
///
/// Shape (kept stable as a public contract for the frontend and for
/// the LLM):
///
/// ```jsonc
/// {
/// "profiles": [
/// {
/// "id": "researcher",
/// "display_name": "Researcher",
/// "description": "...",
/// "tools": { "mode": "allowlist", "allow": ["Read", "Grep"] },
/// "model_hint": null,
/// "default_responsibility": null,
/// "ui": { "icon": "🔎", "color": "blue" }
/// // NOTE: `system_prompt` is intentionally omitted from the
/// // listing — it can be lengthy and is not needed for UI/LLM
/// // selection. Use `action=get` on a child to inspect the
/// // resolved prompt of an active session.
/// }
/// ],
/// "fallback_id": "general-purpose",
/// "count": 6
/// }
/// ```
fn list_profiles_payload(&self) -> serde_json::Value {
// Project each profile into a UI-friendly shape that excludes the
// (potentially large) `system_prompt`. This keeps the payload
// small for both the LLM context window and the frontend list.
let profiles: Vec<serde_json::Value> = self
.profiles
.iter()
.map(|p| {
json!({
"id": p.id,
"display_name": p.display_name,
"description": p.description,
"tools": p.tools,
"model_hint": p.model_hint,
"default_responsibility": p.default_responsibility,
"ui": p.ui,
})
})
.collect();
json!({
"profiles": profiles,
"fallback_id": self.profiles.fallback_id(),
"count": self.profiles.len(),
})
}
}
// ---------------------------------------------------------------------------
// Tests
//
// Pure unit tests for the framework-agnostic helpers live here. Integration
// tests that wire `SubAgentTool` to a real `ChildSessionAdapter` live in
// `bamboo-server` (`tools/sub_agent_tests.rs`), where the adapter + AppState
// types are available.
// ---------------------------------------------------------------------------
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn normalize_title_accepts_legacy_description() {
let title = normalize_title(None, "Search refs".to_string()).unwrap();
assert_eq!(title, "Search refs");
}
#[test]
fn normalize_title_prefers_title_over_description() {
let title =
normalize_title(Some("Real title".to_string()), "Legacy desc".to_string()).unwrap();
assert_eq!(title, "Real title");
}
#[test]
fn normalize_title_rejects_both_empty() {
let err = normalize_title(None, "".to_string()).unwrap_err();
assert!(matches!(err, ToolError::InvalidArguments(msg) if msg.contains("title")));
}
// ---- parse_model_spec ----
fn parent_session(
model_ref: Option<bamboo_domain::ProviderModelRef>,
) -> bamboo_agent_core::Session {
let mut session = bamboo_agent_core::Session::new("p1", "gpt-test");
session.model_ref = model_ref;
session
}
#[test]
fn model_spec_provider_colon_model_is_explicit() {
let parent = parent_session(None);
let r = parse_model_spec("anthropic:claude-sonnet-4-6", &parent, None).unwrap();
assert_eq!(r.provider, "anthropic");
assert_eq!(r.model, "claude-sonnet-4-6");
}
#[test]
fn model_spec_bare_inherits_parent_provider() {
let parent = parent_session(Some(bamboo_domain::ProviderModelRef::new(
"openai", "gpt-test",
)));
let r = parse_model_spec("o4-mini", &parent, Some("anthropic".to_string())).unwrap();
assert_eq!(r.provider, "openai"); // parent wins over default
assert_eq!(r.model, "o4-mini");
}
#[test]
fn model_spec_bare_falls_back_to_default_provider() {
let parent = parent_session(None);
let r =
parse_model_spec("claude-haiku-4-5", &parent, Some("anthropic".to_string())).unwrap();
assert_eq!(r.provider, "anthropic");
}
#[test]
fn model_spec_bare_without_any_provider_errors() {
let parent = parent_session(None);
let err = parse_model_spec("mystery-model", &parent, None).unwrap_err();
assert!(matches!(err, ToolError::InvalidArguments(msg) if msg.contains("provider")));
}
#[test]
fn model_spec_rejects_malformed() {
let parent = parent_session(None);
assert!(parse_model_spec(" ", &parent, None).is_err());
assert!(parse_model_spec("anthropic:", &parent, None).is_err());
assert!(parse_model_spec(":model", &parent, None).is_err());
}
}