MythEclipse/zesdex
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

refactor: streamline agent turn handling and background review process

Browse Source
This commit is contained in:
asepharyana
2026-07-20 17:25:23 +07:00
parent 4c186b62d4
commit dd7825b481
14 changed files with 409 additions and 327 deletions
Download Patch File Download Diff File Expand all files Collapse all files
apps/infrastructure/src/subagent/auto/engine.rs
+267 -242
View File
@@ -1,280 +1,305 @@
//! Auto-review engine — automatically checks git diff after file edits
//! using an LLM subagent.
//! Auto-review engine — after edits, spawns a subagent that reviews AND
//! automatically fixes issues using tool access (edit/write/grep).
//!
//! Flow: after each write/edit tool execution in the agent turn, the runner
//! calls `trigger_auto_review` which:
//! 1. Runs `git diff --cached` and `git diff` to get working-tree changes
//! 2. Sends the diff to a lightweight LLM call for quick review
//! 3. Emits findings as `TurnEvent::SystemNote` on the event queue
//! Flow: after an agent turn with edits completes:
//! 1. Emit `WorkflowAgentUpdate(Running)` → visible in workflow sidebar
//! 2. Run `git diff` to get the changed files
//! 3. Spawn the subagent engine with Write-tier tools + directive to
//! review & fix
//! 4. The subagent finds issues and applies fixes using edit/write tools
//! 5. Results stream as `TurnEvent::SystemNote` events
//! 6. Emit `WorkflowAgentUpdate(Completed)` when done
use std::collections::VecDeque;
use std::path::Path;
use std::path::PathBuf;
use std::process::Command;
use std::sync::{Arc, Mutex};
use anyhow::Result;
use tracing::{debug, info, instrument, warn};
use crate::llm::provider::LlmClient;
use crate::subagent::gating::should_review;
use crate::TurnEvent;
use crate::subagent::context::SubagentContext;
use crate::subagent::division::AccessTier;
use crate::subagent::engine::run_agent;
use crate::tools::ToolCtx;
use crate::{AgentStatus, TurnEvent};
/// Trigger an auto-review of recent git changes.
const REVIEW_AGENT_ID: &str = "auto-review";
/// Spawn a review subagent that reviews changes and auto-fixes issues.
///
/// Flow:
/// 1. Check gating conditions (edit count, consecutive empty reviews)
/// 2. Run `git diff --cached` to get staged changes
/// 3. Run `git diff` to get unstaged changes
/// 4. If there are changes, call LLM for a quick review
/// 5. Emit findings as TurnEvent::SystemNote
/// The subagent runs inline on the current background thread (no extra
/// thread spawn) with its own tokio runtime and Write-tier tool access
/// (edit, write, grep, read, glob). It receives the git diff as context
/// and is directed to:
/// 1. Read changed files
/// 2. Check for typos, missing imports, syntax errors, logic bugs
/// 3. Fix any issues found using edit/write tools
/// 4. Report what was fixed
///
/// Returns `(had_findings, total_findings)` tuple.
#[instrument(skip(turn_events, llm_client))]
pub fn trigger_auto_review(
workspace_root: &Path,
edit_count: u32,
consecutive_empty_reviews: &mut u32,
max_skip: u32,
turn_events: &Arc<Mutex<VecDeque<TurnEvent>>>,
llm_client: Option<&LlmClient>,
) -> Result<(bool, usize)> {
// Gating check
if !should_review(edit_count, *consecutive_empty_reviews, max_skip) {
debug!("auto-review skipped by gating");
return Ok((false, 0));
}
/// All findings stream as TurnEvent events consumed by the TUI event loop.
#[instrument(skip(turn_events))]
pub fn spawn_background_review(
workspace_roots: Vec<PathBuf>,
turn_events: Arc<Mutex<VecDeque<TurnEvent>>>,
api_key: String,
model: String,
api_base: Option<String>,
) {
let agent_id = REVIEW_AGENT_ID.to_string();
let agent_name = "Auto-Review".to_string();
info!("triggering auto-review");
std::thread::spawn(move || {
let root = match workspace_roots.first() {
Some(r) => r.clone(),
None => {
debug!("auto-review: no workspace root, skipping");
return;
}
};
// Run git diff to get changes
let diff = match get_git_diff(workspace_root) {
Ok(d) => d,
Err(e) => {
debug!(error = %e, "auto-review: git diff failed (not a git repo?)");
return Ok((false, 0));
}
};
info!("auto-review: starting");
if diff.is_empty() {
debug!("auto-review: no changes to review");
*consecutive_empty_reviews = consecutive_empty_reviews.saturating_add(1);
return Ok((false, 0));
}
// Mark Running in workflow panel
push_event(
&turn_events,
TurnEvent::WorkflowAgentUpdate {
agent_id: agent_id.clone(),
agent_name: agent_name.clone(),
status: AgentStatus::Running,
},
);
// If we have an LLM client, do a real review
let review_result = if let Some(client) = llm_client {
perform_llm_review(client, &diff)?
} else {
// Fallback: simple heuristic review without LLM
perform_heuristic_review(&diff)
};
// 1. Get the git diff to know what changed
let diff = match get_git_diff(&root) {
Ok(d) if !d.is_empty() => d,
Ok(_) => {
debug!("auto-review: no changes detected");
push_event(
&turn_events,
TurnEvent::WorkflowAgentUpdate {
agent_id,
agent_name,
status: AgentStatus::Completed,
},
);
return;
}
Err(e) => {
debug!(error = %e, "auto-review: git diff failed");
push_event(
&turn_events,
TurnEvent::WorkflowAgentUpdate {
agent_id,
agent_name,
status: AgentStatus::Failed(e),
},
);
return;
}
};
let had_findings = !review_result.is_empty();
let finding_count = review_result.len();
// Truncate very large diffs for the prompt
const MAX_DIFF_CHARS: usize = 5000;
let truncated_diff = if diff.len() > MAX_DIFF_CHARS {
push_event(
&turn_events,
TurnEvent::SystemNote {
kind: "review".into(),
message: format!(
"📐 Diff is large ({} chars), reviewing first {} chars...",
diff.len(),
MAX_DIFF_CHARS
),
},
);
format!(
"{}...\n[diff truncated at {} characters]",
&diff[..MAX_DIFF_CHARS],
MAX_DIFF_CHARS
)
} else {
diff.to_string()
};
if had_findings {
*consecutive_empty_reviews = 0;
info!(finding_count, "auto-review produced findings");
// 2. Build directive: review AND fix issues using tools
let directive = format!(
"You are an auto-review subagent. Complete the following:\n\n\
1. Review this git diff for:\n\
- Typos and spelling errors\n\
- Missing imports or undefined references\n\
- Syntax errors or type mismatches\n\
- Logic bugs or off-by-one errors\n\
- Missing error handling\n\
- Security issues\n\n\
2. FIX any issues you find using the available tools:\n\
- `read` to check file contents\n\
- `edit` to fix specific text blocks\n\
- `write` to replace files if needed\n\
- `grep` to find related patterns\n\n\
3. Be conservative: only fix CLEAR, CONFIRMED issues. \
Don't change logic, style, or formatting.\n\
4. Report what you fixed at the end.\n\n\
Git diff of changes:\n\n```diff\n{truncated_diff}\n```"
);
// Emit findings as SystemNote events
for finding in &review_result {
let note = TurnEvent::SystemNote {
kind: "info".to_string(),
message: format!("🔍 Auto-Review: {finding}"),
};
if let Ok(mut q) = turn_events.lock() {
q.push_back(note);
// 3. Emit progress note
push_event(
&turn_events,
TurnEvent::SystemNote {
kind: "review".into(),
message: "🔍 Auto-review: examining and fixing issues...".into(),
},
);
// 4. Build minimal ToolCtx
let tool_ctx = ToolCtx::builder()
.session_dir(root.join(".zesdex").join("sessions").join("auto-review"))
.workspaces(workspace_roots.clone())
.turn_events(turn_events.clone())
.build();
// 5. Determine base URL
let base_url = api_base.unwrap_or_else(|| {
std::env::var("OPENAI_API_BASE")
.unwrap_or_else(|_| "https://opencode.ai/zen/v1".to_string())
});
// 6. Build SubagentContext
let subagent_ctx = SubagentContext::new(
directive,
tool_ctx.clone(),
"write".to_string(),
base_url,
api_key,
model,
);
// 7. Run the subagent engine directly on this thread
// (creates its own tokio runtime, calls run_agent with Write tools)
let rt = match tokio::runtime::Runtime::new() {
Ok(r) => r,
Err(e) => {
warn!(error = %e, "auto-review: failed to create tokio runtime");
push_event(
&turn_events,
TurnEvent::WorkflowAgentUpdate {
agent_id,
agent_name,
status: AgentStatus::Failed(format!("runtime error: {e}")),
},
);
return;
}
};
let result = rt.block_on(run_agent(
subagent_ctx,
"Auto-review and fix issues in the changed files",
AccessTier::Write,
tool_ctx,
));
// 8. Report results
match result {
Ok(report) => {
let trimmed = report.trim();
if trimmed.is_empty() || trimmed.contains("no issues") {
push_event(
&turn_events,
TurnEvent::SystemNote {
kind: "review".into(),
message: "✅ Auto-review: no issues found.".into(),
},
);
info!("auto-review: no issues found");
} else {
push_event(
&turn_events,
TurnEvent::SystemNote {
kind: "review_finding".into(),
message: format!("📋 Auto-review complete:\n{}", trimmed),
},
);
info!("auto-review: completed with findings");
}
push_event(
&turn_events,
TurnEvent::WorkflowAgentUpdate {
agent_id,
agent_name,
status: AgentStatus::Completed,
},
);
}
Err(e) => {
warn!(error = %e, "auto-review subagent failed");
push_event(
&turn_events,
TurnEvent::SystemNote {
kind: "review".into(),
message: format!("⚠️ Auto-review encountered an error: {e}"),
},
);
push_event(
&turn_events,
TurnEvent::WorkflowAgentUpdate {
agent_id,
agent_name,
status: AgentStatus::Failed(e.to_string()),
},
);
}
}
} else {
*consecutive_empty_reviews = consecutive_empty_reviews.saturating_add(1);
info!("auto-review: no issues found");
}
Ok((had_findings, finding_count))
});
}
/// Run `git diff` to get workspace changes (both staged and unstaged).
fn get_git_diff(workspace_root: &Path) -> Result<String> {
// Check if this is a git repo
fn get_git_diff(workspace_root: &PathBuf) -> Result<String, String> {
let git_dir = workspace_root.join(".git");
if !git_dir.exists() {
return Ok(String::new());
return Err("not a git repository".to_string());
}
// Get unstaged diff
let unstaged = Command::new("git")
.arg("diff")
.current_dir(workspace_root)
.output()
.map_err(|e| anyhow::anyhow!("git diff failed: {e}"))?;
// Get staged diff
let staged = Command::new("git")
.arg("diff")
.arg("--cached")
.current_dir(workspace_root)
.output()
.map_err(|e| anyhow::anyhow!("git diff --cached failed: {e}"))?;
let mut combined = String::new();
let staged_out = String::from_utf8_lossy(&staged.stdout).trim().to_string();
if !staged_out.is_empty() {
combined.push_str("=== Staged Changes ===\n");
combined.push_str(&staged_out);
combined.push('\n');
}
let unstaged_out = String::from_utf8_lossy(&unstaged.stdout).trim().to_string();
if !unstaged_out.is_empty() {
combined.push_str("=== Unstaged Changes ===\n");
combined.push_str(&unstaged_out);
combined.push('\n');
}
// Run `git diff --stat` for summary
let stat = Command::new("git")
.arg("diff")
.arg("--stat")
// Unstaged diff
if let Ok(out) = Command::new("git")
.args(["diff"])
.current_dir(workspace_root)
.output()
.map_err(|e| anyhow::anyhow!("git diff --stat failed: {e}"))?;
let stat_out = String::from_utf8_lossy(&stat.stdout).trim().to_string();
if !stat_out.is_empty() {
combined.push_str("=== Summary ===\n");
combined.push_str(&stat_out);
combined.push('\n');
{
let stdout = String::from_utf8_lossy(&out.stdout).trim().to_string();
if !stdout.is_empty() {
combined.push_str("=== Unstaged Changes ===\n");
combined.push_str(&stdout);
combined.push('\n');
}
}
// Staged diff
if let Ok(out) = Command::new("git")
.args(["diff", "--cached"])
.current_dir(workspace_root)
.output()
{
let stdout = String::from_utf8_lossy(&out.stdout).trim().to_string();
if !stdout.is_empty() {
combined.push_str("=== Staged Changes ===\n");
combined.push_str(&stdout);
combined.push('\n');
}
}
if combined.is_empty() {
return Err("no changes".to_string());
}
Ok(combined)
}
/// Perform an LLM-based review of the git diff.
///
/// Sends the diff to the model with a focused prompt asking it to find
/// typos, missing imports, syntax errors, and other issues.
fn perform_llm_review(client: &LlmClient, diff: &str) -> Result<Vec<String>> {
// Truncate very large diffs to avoid token limits
const MAX_DIFF_CHARS: usize = 8000;
let truncated_diff = if diff.len() > MAX_DIFF_CHARS {
format!("{}...\n[diff truncated at {} characters]",
&diff[..MAX_DIFF_CHARS], MAX_DIFF_CHARS)
} else {
diff.to_string()
};
let messages = vec![
zesdex_domain::core::ChatMessage::system(
"You are a focused code reviewer. Review the following git diff for:\n\
1. Typos and spelling errors\n\
2. Missing imports or undefined references\n\
3. Syntax errors or type mismatches\n\
4. Logic bugs or off-by-one errors\n\
5. Missing error handling\n\
6. Security issues\n\n\
Be concise. List each issue on a new line with severity [HIGH], [MEDIUM], or [LOW].\n\
If no issues are found, reply with exactly: NO_ISSUES_FOUND"
.to_string(),
),
zesdex_domain::core::ChatMessage::user(format!(
"Review this git diff:\n\n```diff\n{truncated_diff}\n```"
)),
];
match client.chat_with_tools_non_streaming(&messages, None, Some(1024), Some(0.3), None) {
Ok((response, _usage)) => {
let text = response.content.unwrap_or_default().trim().to_string();
if text.contains("NO_ISSUES_FOUND") {
return Ok(Vec::new());
}
// Parse findings line by line
let findings: Vec<String> = text
.lines()
.map(|l| l.trim().to_string())
.filter(|l| {
!l.is_empty()
&& !l.starts_with("Here")
&& !l.starts_with("Let me")
&& !l.starts_with("I've")
&& !l.starts_with("The diff")
})
.collect();
Ok(findings)
}
Err(e) => {
warn!(error = %e, "auto-review LLM call failed");
Ok(Vec::new())
}
/// Push a TurnEvent onto the shared event queue.
fn push_event(queue: &Arc<Mutex<VecDeque<TurnEvent>>>, event: TurnEvent) {
if let Ok(mut q) = queue.lock() {
q.push_back(event);
}
}
/// Perform a simple heuristic-based review without an LLM call.
///
/// This is a fallback when no LLM client is available. It checks for:
/// - Missing semicolons
/// - Unclosed brackets
/// - `todo!()` or `unimplemented!()` macros left in code
/// - Debug print statements
/// - Extremely long functions
fn perform_heuristic_review(diff: &str) -> Vec<String> {
let mut findings = Vec::new();
// Check for added lines (lines starting with +)
let added_lines: Vec<&str> = diff
.lines()
.filter(|l| l.starts_with('+') && !l.starts_with("+++"))
.collect();
let added_content: String = added_lines
.iter()
.map(|l| &l[1..]) // Strip leading +
.collect::<Vec<&str>>()
.join("\n");
// Check for todo! and unimplemented!
if added_content.contains("todo!()") {
findings.push("[MEDIUM] `todo!()` found in new code — replace with implementation".to_string());
}
if added_content.contains("unimplemented!()") {
findings.push("[MEDIUM] `unimplemented!()` found in new code — replace with implementation".to_string());
}
// Check for debug print statements
if added_content.contains("println!") || added_content.contains("dbg!") {
findings.push("[LOW] Debug print statements (println!/dbg!) found — consider removing before finalizing".to_string());
}
if added_content.contains("eprintln!") {
findings.push("[LOW] Debug eprintln! found — consider removing before finalizing".to_string());
}
// Check for unreachable or panic statements
if added_content.contains("panic!(\"reached") || added_content.contains("panic!(\"not implemented") {
findings.push("[HIGH] Unreachable code / panic found — implement the missing logic".to_string());
}
// Check for very long lines (>120 chars)
for (i, line) in added_lines.iter().enumerate() {
let content = &line[1..]; // Strip leading +
if content.len() > 120 && !content.trim_start().starts_with("//") {
let preview: String = content.chars().take(80).collect();
findings.push(format!(
"[LOW] Very long line ({} chars, line {} in diff) — consider breaking up:\n `{}…`",
content.len(),
i + 1,
preview
));
}
}
// Count new functions to detect very long additions
let fn_count = added_content.matches("fn ").count();
if fn_count > 5 {
findings.push("[INFO] Large number of new functions ({fn_count}) — consider whether this should be split into separate modules".to_string());
}
findings
}