call_api_with_tools_sync() -> src/agent/oneshot.rs
This commit is contained in:
parent
da24e02159
commit
cbf7653cdf
9 changed files with 638 additions and 234 deletions
|
|
@ -1,228 +0,0 @@
|
|||
// agents/api.rs — Direct API backend for agent execution
|
||||
//
|
||||
// Uses poc-agent's OpenAI-compatible API client to call models directly
|
||||
// (vllm, llama.cpp, OpenRouter, etc.) instead of shelling out to claude CLI.
|
||||
// Implements the tool loop: send prompt → if tool_calls, execute them →
|
||||
// send results back → repeat until text response.
|
||||
//
|
||||
// Activated when config has api_base_url set.
|
||||
|
||||
use crate::agent::api::ApiClient;
|
||||
use crate::agent::api::types::*;
|
||||
use crate::agent::tools::{self as agent_tools};
|
||||
|
||||
use std::sync::OnceLock;
|
||||
|
||||
static API_CLIENT: OnceLock<ApiClient> = OnceLock::new();
|
||||
|
||||
fn get_client() -> Result<&'static ApiClient, String> {
|
||||
Ok(API_CLIENT.get_or_init(|| {
|
||||
let config = crate::config::get();
|
||||
let base_url = config.api_base_url.as_deref().unwrap_or("");
|
||||
let api_key = config.api_key.as_deref().unwrap_or("");
|
||||
let model = config.api_model.as_deref().unwrap_or("qwen-2.5-27b");
|
||||
ApiClient::new(base_url, api_key, model)
|
||||
}))
|
||||
}
|
||||
|
||||
/// Run agent prompts through the direct API with tool support.
|
||||
/// For multi-step agents, each prompt is injected as a new user message
|
||||
/// after the previous step's tool loop completes. The conversation
|
||||
/// context carries forward naturally between steps.
|
||||
/// Returns the final text response after all steps complete.
|
||||
pub async fn call_api_with_tools(
|
||||
agent: &str,
|
||||
prompts: &[String],
|
||||
phases: &[String],
|
||||
temperature: Option<f32>,
|
||||
priority: i32,
|
||||
tools: &[agent_tools::Tool],
|
||||
bail_fn: Option<&(dyn Fn(usize) -> Result<(), String> + Sync)>,
|
||||
log: &dyn Fn(&str),
|
||||
) -> Result<String, String> {
|
||||
let client = get_client()?;
|
||||
|
||||
// Tools are already filtered by the caller
|
||||
// Provenance tracks which agent:phase is making writes.
|
||||
// Updated between steps by the bail function via set_provenance().
|
||||
let first_phase = phases.first().map(|s| s.as_str()).unwrap_or("");
|
||||
let provenance = std::cell::RefCell::new(
|
||||
if first_phase.is_empty() { format!("agent:{}", agent) }
|
||||
else { format!("agent:{}:{}", agent, first_phase) }
|
||||
);
|
||||
|
||||
// Start with the first prompt as a user message
|
||||
let mut messages = vec![Message::user(&prompts[0])];
|
||||
let mut next_prompt_idx = 1; // index of next prompt to inject
|
||||
let reasoning = crate::config::get().api_reasoning.clone();
|
||||
|
||||
let max_turns = 50 * prompts.len();
|
||||
for turn in 0..max_turns {
|
||||
log(&format!("\n=== TURN {} ({} messages) ===\n", turn, messages.len()));
|
||||
|
||||
let mut last_err = None;
|
||||
let mut msg_opt = None;
|
||||
let mut usage_opt = None;
|
||||
for attempt in 0..5 {
|
||||
let sampling = crate::agent::api::SamplingParams {
|
||||
temperature: temperature.unwrap_or(0.6),
|
||||
top_p: 0.95,
|
||||
top_k: 20,
|
||||
};
|
||||
match client.chat_completion_stream_temp(
|
||||
&messages,
|
||||
tools,
|
||||
&reasoning,
|
||||
sampling,
|
||||
Some(priority),
|
||||
).await {
|
||||
Ok((msg, usage)) => {
|
||||
msg_opt = Some(msg);
|
||||
usage_opt = usage;
|
||||
break;
|
||||
}
|
||||
Err(e) => {
|
||||
let err_str = e.to_string();
|
||||
let is_transient = err_str.contains("IncompleteMessage")
|
||||
|| err_str.contains("connection closed")
|
||||
|| err_str.contains("connection reset")
|
||||
|| err_str.contains("timed out")
|
||||
|| err_str.contains("Connection refused");
|
||||
if is_transient && attempt < 4 {
|
||||
log(&format!("transient error (attempt {}): {}, retrying...",
|
||||
attempt + 1, err_str));
|
||||
tokio::time::sleep(std::time::Duration::from_secs(2 << attempt)).await;
|
||||
last_err = Some(e);
|
||||
continue;
|
||||
}
|
||||
let msg_bytes: usize = messages.iter()
|
||||
.map(|m| m.content_text().len())
|
||||
.sum();
|
||||
return Err(format!(
|
||||
"API error on turn {} (~{}KB payload, {} messages, {} attempts): {}",
|
||||
turn, msg_bytes / 1024, messages.len(), attempt + 1, e));
|
||||
}
|
||||
}
|
||||
}
|
||||
let msg = msg_opt.unwrap();
|
||||
if let Some(ref e) = last_err {
|
||||
log(&format!("succeeded after retry (previous error: {})", e));
|
||||
}
|
||||
|
||||
if let Some(u) = &usage_opt {
|
||||
log(&format!("tokens: {} prompt + {} completion",
|
||||
u.prompt_tokens, u.completion_tokens));
|
||||
}
|
||||
|
||||
let has_content = msg.content.is_some();
|
||||
let has_tools = msg.tool_calls.as_ref().is_some_and(|tc| !tc.is_empty());
|
||||
|
||||
if has_tools {
|
||||
// Push the assistant message with tool calls.
|
||||
// Sanitize arguments: vllm re-parses them as JSON when
|
||||
// preprocessing the conversation, so invalid JSON from the
|
||||
// model crashes the next request.
|
||||
let mut sanitized = msg.clone();
|
||||
if let Some(ref mut calls) = sanitized.tool_calls {
|
||||
for call in calls {
|
||||
if serde_json::from_str::<serde_json::Value>(&call.function.arguments).is_err() {
|
||||
log(&format!("sanitizing malformed args for {}: {}",
|
||||
call.function.name, &call.function.arguments));
|
||||
call.function.arguments = "{}".to_string();
|
||||
}
|
||||
}
|
||||
}
|
||||
messages.push(sanitized);
|
||||
|
||||
// Execute each tool call
|
||||
for call in msg.tool_calls.as_ref().unwrap() {
|
||||
log(&format!("\nTOOL CALL: {}({})",
|
||||
call.function.name,
|
||||
&call.function.arguments));
|
||||
|
||||
let args: serde_json::Value = match serde_json::from_str(&call.function.arguments) {
|
||||
Ok(v) => v,
|
||||
Err(_) => {
|
||||
log(&format!("malformed tool call args: {}", &call.function.arguments));
|
||||
messages.push(Message::tool_result(
|
||||
&call.id,
|
||||
"Error: your tool call had malformed JSON arguments. Please retry with valid JSON.",
|
||||
));
|
||||
continue;
|
||||
}
|
||||
};
|
||||
|
||||
let output = agent_tools::dispatch(&call.function.name, &args).await;
|
||||
|
||||
if std::env::var("POC_AGENT_VERBOSE").is_ok() {
|
||||
log(&format!("TOOL RESULT ({} chars):\n{}", output.len(), output));
|
||||
} else {
|
||||
let preview: String = output.lines().next().unwrap_or("").chars().take(100).collect();
|
||||
log(&format!("Result: {}", preview));
|
||||
}
|
||||
|
||||
messages.push(Message::tool_result(&call.id, &output));
|
||||
}
|
||||
continue;
|
||||
}
|
||||
|
||||
// Text-only response — step complete
|
||||
let text = msg.content_text().to_string();
|
||||
if text.is_empty() && !has_content {
|
||||
log("empty response, retrying");
|
||||
messages.push(Message::user(
|
||||
"[system] Your previous response was empty. Please respond with text or use a tool."
|
||||
));
|
||||
continue;
|
||||
}
|
||||
|
||||
log(&format!("\n=== RESPONSE ===\n\n{}", text));
|
||||
|
||||
// If there are more prompts, check bail condition and inject the next one
|
||||
if next_prompt_idx < prompts.len() {
|
||||
// Run bail check before continuing to next step
|
||||
if let Some(ref check) = bail_fn {
|
||||
check(next_prompt_idx)?;
|
||||
}
|
||||
// Update provenance for the new phase
|
||||
if let Some(phase) = phases.get(next_prompt_idx) {
|
||||
*provenance.borrow_mut() = format!("agent:{}:{}", agent, phase);
|
||||
}
|
||||
messages.push(Message::assistant(&text));
|
||||
let next = &prompts[next_prompt_idx];
|
||||
next_prompt_idx += 1;
|
||||
log(&format!("\n=== STEP {}/{} ===\n", next_prompt_idx, prompts.len()));
|
||||
messages.push(Message::user(next));
|
||||
continue;
|
||||
}
|
||||
|
||||
return Ok(text);
|
||||
}
|
||||
|
||||
Err(format!("agent exceeded {} tool turns", max_turns))
|
||||
}
|
||||
|
||||
/// Synchronous wrapper — runs the async function on a dedicated thread
|
||||
/// with its own tokio runtime. Safe to call from any context.
|
||||
pub fn call_api_with_tools_sync(
|
||||
agent: &str,
|
||||
prompts: &[String],
|
||||
phases: &[String],
|
||||
temperature: Option<f32>,
|
||||
priority: i32,
|
||||
tools: &[agent_tools::Tool],
|
||||
bail_fn: Option<&(dyn Fn(usize) -> Result<(), String> + Sync)>,
|
||||
log: &(dyn Fn(&str) + Sync),
|
||||
) -> Result<String, String> {
|
||||
std::thread::scope(|s| {
|
||||
s.spawn(|| {
|
||||
let rt = tokio::runtime::Builder::new_current_thread()
|
||||
.enable_all()
|
||||
.build()
|
||||
.map_err(|e| format!("tokio runtime: {}", e))?;
|
||||
rt.block_on(
|
||||
call_api_with_tools(agent, prompts, phases, temperature, priority, tools, bail_fn, log)
|
||||
)
|
||||
}).join().unwrap()
|
||||
})
|
||||
}
|
||||
|
|
@ -210,7 +210,7 @@ pub fn link_audit(store: &mut Store, apply: bool) -> Result<AuditStats, String>
|
|||
// Run batches in parallel via rayon
|
||||
let batch_results: Vec<_> = batch_data.par_iter()
|
||||
.map(|(batch_idx, batch_infos, prompt)| {
|
||||
let response = super::api::call_api_with_tools_sync(
|
||||
let response = crate::agent::oneshot::call_api_with_tools_sync(
|
||||
"audit", &[prompt.clone()], &[], None, 10, &[], None, &|_| {});
|
||||
let completed = done.fetch_add(1, Ordering::Relaxed) + 1;
|
||||
eprint!("\r Batches: {}/{} done", completed, total_batches);
|
||||
|
|
|
|||
|
|
@ -284,7 +284,7 @@ fn generate_digest(
|
|||
.filter(|t| def.tools.iter().any(|w| w == &t.name))
|
||||
.collect()
|
||||
};
|
||||
let digest = super::api::call_api_with_tools_sync(
|
||||
let digest = crate::agent::oneshot::call_api_with_tools_sync(
|
||||
&def.agent, &prompts, &phases, def.temperature, def.priority,
|
||||
&tools, None, &log)?;
|
||||
|
||||
|
|
|
|||
|
|
@ -17,7 +17,6 @@
|
|||
//
|
||||
// The session hook (context injection, agent orchestration) moved to claude/hook.
|
||||
|
||||
pub mod api;
|
||||
pub mod audit;
|
||||
pub mod consolidate;
|
||||
pub mod daemon;
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue