Files
zesdex/src/app/subagent/engine.rs
T

210 lines
8.9 KiB
Rust
Raw Normal View History

//! Subagent execution loop: drive an LLM conversation, gate tool calls
//! against the context's allowlist, run tools, and stream progress events
//! to the parent via an mpsc channel.
use tokio::sync::mpsc;
use crate::dto::chat::message::ChatMessage;
use crate::dto::provider::request::ToolDef;
use crate::tool::{all_tools, tool_defs, tool_is_risky};
use super::context::SubagentContext;
use super::event::SubagentEvent;
/// Upper bound on agent loop steps; effectively unbounded (`usize::MAX`).
#[allow(dead_code)]
pub const MAX_AGENT_STEPS: usize = usize::MAX;
/// Maps a subagent's allowed tool names to concrete Tool trait objects and
/// OpenAI-style tool definitions.
///
/// Flow: load `all_tools()` → if `allowed_tools` is empty, use all; else
/// filter by membership → derive `ToolDef`s for the LLM.
///
/// Why: an empty allowlist means "no restriction" (matches
/// `build_subagent_context`'s default for non-reviewer roles).
///
/// Return: `(tool impls, schema defs)` for the subagent to use.
fn build_subagent_tools(allowed_tools: &[String]) -> (Vec<Box<dyn crate::tool::Tool>>, Vec<ToolDef>) {
let all = all_tools();
let filtered: Vec<Box<dyn crate::tool::Tool>> = if allowed_tools.is_empty() {
all
} else {
all.into_iter()
.filter(|t| allowed_tools.contains(&t.name().to_string()))
.collect()
};
let defs = tool_defs(&filtered);
(filtered, defs)
}
/// Resolve the API key, model, and base URL from persisted app config.
///
/// Flow: try the settings key for the active provider → fall back to the
/// provider's `api_key_env` env-var → fall back to the provider's
/// `default_api_key` → fall back to an empty string.
///
/// Why: matches the main agent's credential resolution exactly, so
/// subagents automatically inherit the same provider settings.
///
/// Return: `(api_key, model, optional_base_url)`.
fn resolve_provider_config() -> (String, String, Option<String>) {
let settings = crate::model::settings::Settings::load();
let app_config = crate::model::app_config::AppConfig::load();
let mut api_key = settings.api_keys.get(&settings.provider).cloned().unwrap_or_else(|| {
tracing::warn!("[subagent] no API key for provider '{}' in settings, trying env/default", settings.provider);
String::new()
});
let model = settings.model.clone();
let base_url = app_config.providers.get(&settings.provider)
.map(|p| p.api_base.clone());
if api_key.is_empty() {
if let Some(provider_cfg) = app_config.providers.get(&settings.provider) {
api_key = provider_cfg.api_key_env.as_ref()
.and_then(|env| std::env::var(env).ok())
.or_else(|| provider_cfg.default_api_key.clone())
.unwrap_or_else(|| {
tracing::warn!("[subagent] all API key resolution paths exhausted for '{}'", settings.provider);
String::new()
});
}
}
(api_key, model, base_url)
}
/// Synchronous subagent entry point: run up to `ctx.max_steps` iterations
/// of the LLM tool loop.
///
/// Flow: inject system prompt → for each step: resolve provider config,
/// build an LLM client, call `chat_with_tools_non_streaming`, process tool
/// calls or collect text output → send `SubagentEvent`s on `tx` → break on
/// first text-only (non-empty) response.
///
/// Why: runs synchronously on a dedicated thread so the main async event
/// loop is not blocked. Tool gating prevents restricted or risky tools from
/// executing unless explicitly allowed.
///
/// Return: the concatenated text output, or an `anyhow::Error` if the LLM
/// call fails at any step.
pub fn run_subagent(ctx: SubagentContext, tx: mpsc::Sender<SubagentEvent>) -> anyhow::Result<String> {
let mut output = String::new();
let mut messages: Vec<ChatMessage> = Vec::new();
messages.push(ChatMessage::system(ctx.system_prompt.clone()));
let tool_ctx = crate::tool::ToolCtx::builder()
.session_dir(ctx.session_dir.clone())
.workspaces(ctx.workspaces.clone())
.origin(crate::app::state::types::Origin::SubAgent)
.build();
// Build tool list once before the loop
let (tools, tdefs) = build_subagent_tools(&ctx.allowed_tools);
let tdefs_opt: Option<Vec<ToolDef>> = if tdefs.is_empty() { None } else { Some(tdefs) };
// Cache provider config once before the loop instead of re-resolving
// from disk on every step (Settings::load + AppConfig::load each parse
// JSON files, and the config cannot change between steps).
let (api_key, model, base_url) = resolve_provider_config();
let client = crate::service::provider::LlmClient::new(api_key, model, base_url);
for step in 0..ctx.max_steps {
// Use the structured tool-calling API so the LLM can request tools with
// proper arguments, exactly like the main agent does.
let (response, _usage) = match client.chat_with_tools_non_streaming(&messages, tdefs_opt.clone()) {
Ok(result) => result,
Err(e) => {
let _ = tx.blocking_send(SubagentEvent::StepFailed {
_step: step,
_error: e.to_string(),
});
anyhow::bail!("subagent call failed at step {}: {}", step, e);
}
};
let has_tool_calls = response.tool_calls.is_some()
&& response.tool_calls.as_ref().is_some_and(|tc| !tc.is_empty());
let content = response.content.clone().unwrap_or_default();
if has_tool_calls {
let tool_calls = response.tool_calls.clone().unwrap_or_default();
// Push the assistant message with tool_calls into the conversation
messages.push(response);
for tool_call in &tool_calls {
let tool_name = &tool_call.function.name;
let args = crate::dto::chat::tool::sanitize_tool_arguments(&tool_call.function.arguments);
let explicitly_allowed = ctx.allowed_tools.contains(tool_name);
let generally_allowed = ctx.allowed_tools.is_empty() || explicitly_allowed;
let _ = tx.blocking_send(SubagentEvent::ToolCall {
_tool: tool_name.clone(),
_args: args.clone(),
});
if !generally_allowed {
let msg = format!("tool '{}' not allowed for this subagent", tool_name);
messages.push(ChatMessage::tool_result(tool_call.id.clone(), msg.clone()));
let _ = tx.blocking_send(SubagentEvent::ToolResult {
_tool: tool_name.clone(),
_output: msg,
});
continue;
}
if tool_is_risky(tool_name) && !explicitly_allowed {
let msg = format!("risky tool '{}' requires explicit permission; not allowed for this subagent", tool_name);
messages.push(ChatMessage::tool_result(tool_call.id.clone(), msg.clone()));
let _ = tx.blocking_send(SubagentEvent::ToolResult {
_tool: tool_name.clone(),
_output: msg,
});
continue;
}
let result = match tools.iter().find(|t| t.name() == tool_name.as_str()) {
Some(tool) => tool.run(&tool_ctx, &args),
None => Err(anyhow::anyhow!("tool '{}' not found", tool_name)),
};
match result {
Ok(output_text) => {
messages.push(ChatMessage::tool_result(tool_call.id.clone(), output_text.clone()));
let _ = tx.blocking_send(SubagentEvent::ToolResult {
_tool: tool_name.clone(),
_output: output_text,
});
}
Err(e) => {
let msg = format!("tool '{}' failed: {}", tool_name, e);
messages.push(ChatMessage::tool_result(tool_call.id.clone(), msg.clone()));
let _ = tx.blocking_send(SubagentEvent::ToolResult {
_tool: tool_name.clone(),
_output: msg,
});
}
}
}
} else {
// Text-only response — accumulate and finish
if !content.is_empty() {
output.push_str(&content);
output.push('\n');
}
let _ = tx.blocking_send(SubagentEvent::StepCompleted {
_step: step,
_output: content.clone(),
});
// Break only when we got real content; empty means something went wrong
if !content.is_empty() {
break;
}
}
}
let _ = tx.blocking_send(SubagentEvent::Completed { _output: output.clone() });
Ok(output)
}