perf(agent): stabilkan async & parallel — satu runtime, bounded concurrency, isolasi error

Seperti Claude Code: satu runtime shared, concurrency dibatasi, error
subagent terisolasi (satu node gagal tidak menggagalkan cycle).

- feat(runtime): global tokio runtime via OnceLock — ganti 9+ titik
  Runtime::new() per tool call (spawn, parallel_delegate, workflow,
  explore, dir_cache, daemon handler). Hemat resource, hilangkan panic
  path Runtime::new().expect() di daemon compaction.
- fix(workflow): execute_cycle ganti try_join_all (fail-fast) →
  buffer_unordered(8) + isolasi error per node; node gagal di-log dan
  diganti [ERROR], hasil node lain tetap dipakai (Claude Code-style).
- fix(parallel_delegate): spawn subagent dibatasi per batch max_parallel
  (tidak unbounded threads).
- perf(subagent): run_agent adaptif max_tokens (800/1600/4096), temp 0.2,
  truncate tool output 12k, error-recovery note utk tool error berulang.
- test: runtime singleton + block_on (2 test).
This commit is contained in:
asepharyana
2026-08-27 23:25:44 +07:00
parent 3847c0e6fd
commit 6a98d52d54
11 changed files with 225 additions and 78 deletions
Generated
+11 -11
View File
@@ -4862,7 +4862,7 @@ dependencies = [
[[package]]
name = "zesdex-api"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"argon2",
@@ -4885,7 +4885,7 @@ dependencies = [
[[package]]
name = "zesdex-application"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"base64",
@@ -4902,7 +4902,7 @@ dependencies = [
[[package]]
name = "zesdex-bootstrap"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"chrono",
@@ -4919,7 +4919,7 @@ dependencies = [
[[package]]
name = "zesdex-daemon"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"base64",
@@ -4943,7 +4943,7 @@ dependencies = [
[[package]]
name = "zesdex-domain"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"base64",
@@ -4959,7 +4959,7 @@ dependencies = [
[[package]]
name = "zesdex-gateway"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"axum",
@@ -4986,7 +4986,7 @@ dependencies = [
[[package]]
name = "zesdex-grpc"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"axum",
@@ -5003,7 +5003,7 @@ dependencies = [
[[package]]
name = "zesdex-infrastructure"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"argon2",
@@ -5051,7 +5051,7 @@ dependencies = [
[[package]]
name = "zesdex-tui"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"base64",
@@ -5077,7 +5077,7 @@ dependencies = [
[[package]]
name = "zesdex-web"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"axum",
@@ -5097,7 +5097,7 @@ dependencies = [
[[package]]
name = "zesdex-ws"
version = "1.18.4"
version = "1.19.0"
dependencies = [
"anyhow",
"axum",
@@ -15,7 +15,7 @@
//! read of up to 3 relevant files).
//! 4. Join the result and return a concise bullet summary as a tool message.
use anyhow::{Context, Result};
use anyhow::Result;
use serde_json::{json, Value};
use tracing::{info, warn};
@@ -118,7 +118,7 @@ impl Tool for ExploreCodebase {
model,
);
let rt = tokio::runtime::Runtime::new().context("create explore tokio runtime")?;
let rt = crate::runtime::runtime();
let result = rt.block_on(run_agent(
subagent_ctx,
&directive,
+1
View File
@@ -34,6 +34,7 @@ pub mod llm;
pub mod mcp;
pub mod middleware;
pub mod persistence;
pub mod runtime;
pub mod subagent;
pub mod tools;
pub mod utils;
+62
View File
@@ -0,0 +1,62 @@
//! Process-wide shared Tokio runtime for sync → async bridging.
//!
//! Many `Tool::run` implementations are synchronous but need to drive async
//! work (LLM calls, subagent execution). Creating a fresh
//! [`tokio::runtime::Runtime`] on every call is expensive (spawns a thread
//! pool + runtime each time) and can fail randomly under thread pressure.
//!
//! # Flow
//!
//! [`runtime()`] returns a lazily-initialised process-wide runtime created
//! exactly once via [`std::sync::OnceLock`]. Callers use
//! `runtime().block_on(...)` exactly like they would with a local runtime —
//! the only difference is the runtime is shared, so the cost is paid once per
//! process instead of once per tool call.
//!
//! # Safety
//!
//! `block_on` panics if called from within a running Tokio runtime. The
//! tools that use this helper are synchronous (`Tool::run`), so this is safe
//! in practice. Async code should never call `runtime().block_on`.
use std::sync::OnceLock;
/// Maximum worker threads for the shared runtime. Kept modest — tools are
/// mostly I/O-bound and rarely need more concurrency than this.
const RUNTIME_WORKER_THREADS: usize = 8;
static SHARED_RUNTIME: OnceLock<tokio::runtime::Runtime> = OnceLock::new();
/// Return the process-wide shared Tokio runtime, initialising it on first use.
///
/// The runtime is configured with `worker_threads = 8` and
/// `enable_all()` (time + IO drivers) so streams, timers, and network calls
/// all work. If initialisation fails (extremely rare — resource exhaustion at
/// startup), the process aborts with a clear message rather than returning
/// an error on every subsequent call.
pub fn runtime() -> &'static tokio::runtime::Runtime {
SHARED_RUNTIME.get_or_init(|| {
tokio::runtime::Builder::new_multi_thread()
.worker_threads(RUNTIME_WORKER_THREADS)
.thread_name("zesdex-shared-rt")
.enable_all()
.build()
.expect("failed to create shared tokio runtime")
})
}
#[cfg(test)]
mod tests {
use super::runtime;
#[test]
fn runtime_is_singleton() {
assert!(std::ptr::eq(runtime(), runtime()));
}
#[test]
fn runtime_blocks_and_resolves() {
let val = runtime().block_on(async { 6 * 7 });
assert_eq!(val, 42);
}
}
+62 -3
View File
@@ -23,6 +23,37 @@ use zesdex_domain::subagent_directive;
/// Maximum number of tool-call iterations before the engine gives up.
const MAX_ITERATIONS: u32 = 25;
/// A single tool-result message is truncated before entering the subagent's
/// context so it cannot blow the window (matches the main turn service).
const TOOL_OUTPUT_MAX_CHARS: usize = 12_000;
/// Maximum consecutive identical tool errors before the engine injects a
/// recovery note steering the model to a different approach.
const MAX_CONSECUTIVE_TOOL_ERRORS: usize = 3;
/// Pick a `max_tokens` budget proportional to the directive's length.
fn adaptive_max_tokens(directive_len: usize) -> u32 {
if directive_len <= 80 {
800
} else if directive_len <= 400 {
1600
} else {
4096
}
}
fn truncate_tool_output(output: String) -> String {
if output.len() <= TOOL_OUTPUT_MAX_CHARS {
return output;
}
let mut result: String = output.chars().take(TOOL_OUTPUT_MAX_CHARS).collect();
result.push_str(&format!(
"\n...[truncated {} chars]",
output.len() - TOOL_OUTPUT_MAX_CHARS
));
result
}
/// Emit an `AgentProgress` event onto the turn-event queue, if one is
/// configured in the `ToolCtx`.
fn report_progress(tool_ctx: &ToolCtx, progress: AgentProgress) {
@@ -85,11 +116,17 @@ pub async fn run_agent(
Some(ctx.base_url.clone()),
);
let max_tokens = adaptive_max_tokens(directive.len());
// Track repeated tool errors so the agent can recover from a dead end.
let mut consecutive_errors = 0usize;
let mut last_tool = String::new();
// Limited iteration loop so we don't run forever
for iteration in 0..MAX_ITERATIONS {
use zesdex_application::ports::ProviderService;
let (response_msg, _usage) = client
.chat(&messages, Some(defs.clone()), Some(4096), None)
.chat(&messages, Some(defs.clone()), Some(max_tokens), Some(0.2))
.await?;
let content = response_msg.content.clone().unwrap_or_default();
@@ -113,7 +150,7 @@ pub async fn run_agent(
&tool_ctx,
AgentProgress::running(
"subagent",
format!("{}:{}", directive, tool_name),
format!("{}:{tool_name}", directive),
Some(tool_name.clone()),
),
);
@@ -127,7 +164,29 @@ pub async fn run_agent(
format!("Unknown tool: {tool_name}")
};
messages.push(ChatMessage::tool(tc.id.clone(), result));
// Error-recovery: if the same tool keeps failing, inject a
// system note steering the model to a different approach.
if result.starts_with("Error:") {
if last_tool.as_str() == tool_name.as_str() {
consecutive_errors += 1;
} else {
consecutive_errors = 1;
last_tool = tool_name.to_string();
}
if consecutive_errors >= MAX_CONSECUTIVE_TOOL_ERRORS {
messages.push(ChatMessage::system(
zesdex_domain::agent::prompt::error_recovery_note(tool_name, &result),
));
consecutive_errors = 0;
}
} else {
consecutive_errors = 0;
}
messages.push(ChatMessage::tool(
tc.id.clone(),
truncate_tool_output(result),
));
}
// Add assistant response if there was text content
+1 -2
View File
@@ -32,7 +32,6 @@ pub fn spawn_subagent(
) -> thread::JoinHandle<Result<String>> {
info!("Spawning subagent: {directive}");
thread::spawn(move || {
let rt = tokio::runtime::Runtime::new()?;
rt.block_on(run_agent(ctx, &directive, access, tool_ctx))
crate::runtime::runtime().block_on(run_agent(ctx, &directive, access, tool_ctx))
})
}
@@ -123,7 +123,7 @@ impl Tool for ParallelDelegate {
.collect()
} else {
// Auto-split using LLM
let rt = tokio::runtime::Runtime::new()?;
let rt = crate::runtime::runtime();
let directives = rt.block_on(auto_split_task(
&task,
max_parallel,
@@ -146,9 +146,13 @@ impl Tool for ParallelDelegate {
"parallel delegation: starting subagents"
);
// Spawn agents in parallel
let mut handles = Vec::new();
for (i, (directive, access)) in directives.iter().enumerate() {
// Spawn agents in parallel — bounded: never more than `max_parallel`
// subagent threads in flight at once (Claude Code-style isolation).
let mut results: Vec<(usize, String, String)> = Vec::new();
for batch in directives.chunks(max_parallel) {
let mut handles = Vec::with_capacity(batch.len());
for (i, (directive, access)) in batch.iter().enumerate() {
let global_idx = results.len() + i;
let subagent_ctx = SubagentContext::new(
directive.clone(),
ctx.clone(),
@@ -158,13 +162,12 @@ impl Tool for ParallelDelegate {
model.clone(),
);
debug!(agent_index = i, access = ?access, "spawning parallel agent");
debug!(agent_index = global_idx, access = ?access, "spawning parallel agent");
let handle = spawn_subagent(subagent_ctx, directive.clone(), *access, ctx.clone());
handles.push((i, handle));
handles.push((global_idx, handle));
}
// Join all results
let mut results: Vec<(usize, String, String)> = Vec::new();
// Join this batch before spawning the next.
for (i, handle) in handles {
match handle.join() {
Ok(Ok(output)) => {
@@ -185,10 +188,11 @@ impl Tool for ParallelDelegate {
}
}
}
}
// Consolidate results
if synthesize && results.len() > 1 {
let rt = tokio::runtime::Runtime::new()?;
let rt = crate::runtime::runtime();
let consolidated =
rt.block_on(consolidate_results(&results, &base_url, &api_key, &model))?;
Ok(format!(
@@ -63,7 +63,7 @@ impl crate::tools::Tool for DirCacheUpdate {
// Persist the resolved paths into the shared DirCache so the TUI
// and other tools can read the cached listing without re-scanning.
let dc = ctx.dir_cache.clone();
let rt = tokio::runtime::Runtime::new()?;
let rt = crate::runtime::runtime();
rt.block_on(async { dc.write().await.set(resolved).await });
info!(count, "directory cache updated");
+2 -2
View File
@@ -65,7 +65,7 @@ impl Tool for WorkflowRun {
zesdex_domain::agent::defaults::DEFAULT_MODEL.to_string(),
None,
);
let rt = tokio::runtime::Runtime::new()?;
let rt = crate::runtime::runtime();
let result: Vec<String> =
rt.block_on(async { execute_workflow(&script, ctx, &llm_client).await })?;
@@ -229,7 +229,7 @@ impl Tool for HiveMind {
.ok_or_else(|| anyhow::anyhow!("missing 'cycles' array"))?;
info!("Hive mind starting with {} cycles", cycles_val.len());
let rt = tokio::runtime::Runtime::new()?;
let rt = crate::runtime::runtime();
let mut all_node_outputs = Vec::new();
for (cycle_idx, cycle_val) in cycles_val.iter().enumerate() {
@@ -1,11 +1,15 @@
//! Hive-mind cycle execution — run one cycle of parallel nodes.
//!
//! Flow: load settings → resolve LLM credentials → run all directives in the
//! cycle concurrently via try_join_all → collect Vec<NodeOutput>.
//! cycle concurrently via a BOUNDED buffer (`buffer_unordered(MAX)`) → collect
//! `Vec<NodeOutput>`. Unlike `try_join_all`, a single failing node does NOT
//! fail the whole cycle — failed nodes are logged and replaced with an
//! `[ERROR]` output so the remaining results are preserved (like Claude
//! Code's isolated subagents).
use anyhow::Result;
use futures_util::future::try_join_all;
use tracing::info;
use futures_util::stream::StreamExt;
use tracing::{info, warn};
use zesdex_domain::cms::{AppConfigRepository, SettingsRepository};
use zesdex_domain::core::Store;
@@ -15,16 +19,21 @@ use crate::subagent::context::SubagentContext;
use crate::subagent::division::AccessTier;
use crate::subagent::engine::run_agent;
use crate::tools::ToolCtx;
use zesdex_domain::workflow::{CognitiveCycle, NodeOutput};
use zesdex_domain::workflow::{CognitiveCycle, NodeDirective, NodeOutput};
/// Maximum number of hive-mind nodes running concurrently per cycle.
/// Keeps thread/runtime pressure bounded (Claude Code-style).
const MAX_CONCURRENT_NODES: usize = 8;
/// Execute one cycle: run each node directive and collect outputs.
///
/// Flow:
/// 1. Load `Settings` and `AppConfig` from the store directory.
/// 2. Resolve provider, model, base_url, and api_key.
/// 3. Spawn all directives concurrently — each builds a `SubagentContext`
/// and calls `run_agent` (Full access).
/// 4. `try_join_all` waits for all to complete, then collect `NodeOutput`s.
/// 3. Spawn directives with bounded concurrency — each builds a
/// `SubagentContext` and calls `run_agent`.
/// 4. Collect `NodeOutput`s; failed nodes are logged and replaced with an
/// `[ERROR]` placeholder so the cycle still completes.
pub async fn execute_cycle(cycle: &CognitiveCycle, tool_ctx: &ToolCtx) -> Result<Vec<NodeOutput>> {
info!(
"Executing cycle {} with {} directives",
@@ -53,10 +62,9 @@ pub async fn execute_cycle(cycle: &CognitiveCycle, tool_ctx: &ToolCtx) -> Result
let cycle_index = cycle.index;
use zesdex_domain::workflow::NodeDirective;
// Run all directives in this cycle concurrently.
let handles: Vec<_> = cycle
// Run all directives with bounded concurrency. Each node is its own
// future; failures are collected, not propagated (isolated errors).
let tasks: Vec<_> = cycle
.directives
.iter()
.enumerate()
@@ -79,18 +87,33 @@ pub async fn execute_cycle(cycle: &CognitiveCycle, tool_ctx: &ToolCtx) -> Result
_ => AccessTier::Read,
};
let node_id = format!("Node-{}-{}", cycle_index, i);
async move {
let result = run_agent(ctx, &dir, access, tc).await?;
Ok::<NodeOutput, anyhow::Error>(NodeOutput {
id: format!("Node-{}-{}", cycle_index, i),
match run_agent(ctx, &dir, access, tc).await {
Ok(output) => Ok::<NodeOutput, anyhow::Error>(NodeOutput {
id: node_id.clone(),
directive: dir,
output: result,
output,
}),
Err(e) => {
warn!(node = %node_id, error = %e, "hive-mind node failed (isolated)");
Ok::<NodeOutput, anyhow::Error>(NodeOutput {
id: node_id,
directive: dir,
output: format!("[ERROR] {e}"),
})
}
}
}
})
.collect();
let results = try_join_all(handles).await?;
// Bounded concurrency: run at most MAX_CONCURRENT_NODES futures at once.
let mut stream = futures_util::stream::iter(tasks).buffer_unordered(MAX_CONCURRENT_NODES);
let mut results = Vec::with_capacity(cycle.directives.len());
while let Some(node) = stream.next().await {
results.push(node?);
}
Ok(results)
}
+1 -2
View File
@@ -471,8 +471,7 @@ fn handle_compact(state: &mut AppStateRest) {
let client = zesdex_infrastructure::llm::provider::LlmClient::new(api_key, model, api_base);
if let Some(ref mut rt) = state.session_runtime {
let tokio_rt =
tokio::runtime::Runtime::new().expect("create tokio runtime for AI compaction");
let tokio_rt = zesdex_infrastructure::runtime::runtime();
if let Ok(()) = tokio_rt.block_on(
zesdex_application::agent::turn_service::compact_messages_with_ai(
&mut rt.messages,