fix(agent): retry LLM completion on malformed 2xx JSON body (#5351)

## Problem

A provider can return HTTP 200 with a **truncated JSON body** — cleanly
closed connection, correct framing, content cut off mid-value. Both LLM
HTTP loops treated this as a terminal error on the first attempt:
`AgentError::Llm("json: EOF while parsing a value")`, surfaced as code
-32000 at the ACP boundary, killing the agent turn before it produced
anything.

Observed live in a tb2.1 bench trial (write-compressor, tb21-twins-1):
deepseek via OpenRouter returned a truncated body, the agent died
mid-prompt with 0 turns completed, and the trial scored 0 on a provider
hiccup.

Meanwhile the same loops already retry timeouts, 429s, 5xxs, 499s, and
mid-body stream stalls — a truncated-but-complete body was the one
transient upstream fault that fell through to terminal.

## Fix

In both `post()` and `openrouter_post()`
(`crates/buzz-agent/src/llm.rs`): when the fully-received success body
fails `serde_json::from_slice`, `continue` the **existing** retry loop
instead of returning terminal — same `MAX_RETRIES` (3) bound, same
`backoff_with_jitter`. On exhaustion, the error goes through
`terminal_llm_error` so it carries cumulative duration + attempt count
like every other retried failure (previously the `json:` error carried
neither).

`post_anthropic` routes through `post()`, so
Anthropic/OpenAI/Databricks/mesh and OpenRouter are all covered.

## Why this cannot re-run a tool call

Hard requirement: tool calls are not idempotent, and this change must
not introduce any possibility of replaying one.

1. **The retry lives inside the HTTP POST helper, below the parse
boundary.** Tool calls are only ever extracted from a *successfully
parsed* response value
(`parse_openai`/`parse_anthropic`/`parse_responses`, all downstream of
these helpers' `Ok` return). A malformed body never parses, therefore no
tool call was ever extracted from it, therefore nothing downstream of it
ever dispatched.
2. **What is re-sent is the completion request itself** — the identical
`body_bytes` captured once at function entry. Sending a completion
request executes no tools; it asks the model for the next message.
3. **Same safety class as existing behavior.** The loop already re-sends
this identical request on 429/5xx/timeout/stream-stall; this adds one
more transient-fault arm to the same loop with the same bytes.

## Tests

Three new tests mirroring the existing 499/dropped-connection fixtures
(raw `TcpListener` stubs):
- `post_retries_malformed_json_body_and_succeeds` — truncated 200 body
on attempt 1, valid JSON on attempt 2; asserts success and **exactly 2**
server-side requests
- `post_exhausts_retries_on_persistent_malformed_json` —
always-truncated body; asserts exactly `MAX_RETRIES` attempts and a
terminal error carrying `json:` + cumulative/attempt context
- `openrouter_post_retries_malformed_json_body_and_succeeds` — same
recovery through OpenRouter's separate loop

Full `cargo test -p buzz-agent` green at e7a5d7bb (430 lib + all
integration targets, 0 failures); `cargo fmt` + `clippy --all-targets`
clean.

Originating conversation: buzz-benchmarking channel, thread 397a992d.

Signed-off-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Co-authored-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
This commit is contained in:
Tyler
2026-08-08 17:29:06 -04:00
committed by GitHub
co-authored by Eva
parent f029deafae
commit 5bf78671f4
+283 -3
View File
@@ -2203,8 +2203,33 @@ where
}
}
}
return serde_json::from_slice(&buf)
.map_err(|e| PostError::Agent(AgentError::Llm(format!("json: {e}"))));
// A 2xx body that fails to parse (e.g. a provider flushing a
// truncated JSON document and closing the stream cleanly) is a
// transient upstream fault, not a request problem: retry it like a
// 5xx. Safe to re-send — a malformed body produced no parsed
// response, so no tool call was ever extracted from it, and the
// retried request is the identical completion POST captured in
// `body_bytes` at function entry.
match serde_json::from_slice(&buf) {
Ok(value) => return Ok(value),
Err(e) => {
if attempt + 1 < MAX_RETRIES {
tracing::warn!(
attempt = attempt + 1,
max_attempts = MAX_RETRIES,
error = %e,
"llm: malformed response body, retrying"
);
backoff_with_jitter(attempt).await;
continue;
}
return Err(PostError::Agent(terminal_llm_error(
call_start.elapsed(),
attempt + 1,
&format!("json: {e}"),
)));
}
}
}
// Unreachable in practice: every iteration either returns or continues.
// A fallthrough here would mean MAX_RETRIES was 0, which is rejected at
@@ -2610,7 +2635,31 @@ async fn openrouter_post(
}
}
}
return serde_json::from_slice(&buf).map_err(|e| AgentError::Llm(format!("json: {e}")));
// Same malformed-body retry as the shared `post()` terminal: a
// truncated 2xx JSON body is transient upstream trouble, and
// re-sending is provably tool-safe — nothing was parsed, so no tool
// call could have been extracted, and the retry re-uses the
// identical `body_bytes` completion request.
match serde_json::from_slice(&buf) {
Ok(value) => return Ok(value),
Err(e) => {
if attempt + 1 < MAX_RETRIES {
tracing::warn!(
attempt = attempt + 1,
max_attempts = MAX_RETRIES,
error = %e,
"llm: openrouter malformed response body, retrying"
);
backoff_with_jitter(attempt).await;
continue;
}
return Err(terminal_llm_error(
call_start.elapsed(),
attempt + 1,
&format!("json: {e}"),
));
}
}
}
Err(terminal_llm_error(
call_start.elapsed(),
@@ -4791,6 +4840,237 @@ mod tests {
);
}
/// Regression (write-compressor, tb21-twins-1): a provider returning
/// HTTP 200 with a *truncated* JSON body (cleanly closed, correct
/// framing, unparseable content) previously surfaced as a terminal
/// `AgentError::Llm("json: EOF while parsing a value ...")` on the very
/// first attempt, killing the agent turn. A malformed body is transient
/// upstream trouble and must be retried like a 5xx. Re-sending is
/// tool-safe: nothing was parsed, so no tool call was extracted from the
/// bad body, and the retry replays the identical completion request.
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn post_retries_malformed_json_body_and_succeeds() {
use std::sync::atomic::{AtomicU32, Ordering};
use std::sync::Arc;
use tokio::io::{AsyncReadExt, AsyncWriteExt};
use tokio::net::TcpListener;
let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
let url = format!("http://{}/v1/x", listener.local_addr().unwrap());
let accepts = Arc::new(AtomicU32::new(0));
let accepts_srv = accepts.clone();
tokio::spawn(async move {
loop {
let (mut sock, _) = match listener.accept().await {
Ok(p) => p,
Err(_) => return,
};
let n = accepts_srv.fetch_add(1, Ordering::SeqCst);
let mut buf = Vec::new();
let mut tmp = [0u8; 4096];
while !buf.windows(4).any(|w| w == b"\r\n\r\n") {
match sock.read(&mut tmp).await {
Ok(0) | Err(_) => return,
Ok(k) => buf.extend_from_slice(&tmp[..k]),
}
}
if n == 0 {
// First attempt: 200 OK with a truncated JSON document.
// Content-Length matches the bytes actually sent, so the
// body read completes cleanly — the fault is purely that
// the JSON is cut off mid-value.
let body = "{\"choices\":[{\"mess";
let resp = format!(
"HTTP/1.1 200 OK\r\nContent-Type: application/json\r\n\
Content-Length: {}\r\nConnection: close\r\n\r\n{}",
body.len(),
body,
);
let _ = sock.write_all(resp.as_bytes()).await;
let _ = sock.shutdown().await;
continue;
}
// Subsequent attempts: complete valid JSON.
let body = "{\"ok\":true}";
let resp = format!(
"HTTP/1.1 200 OK\r\nContent-Type: application/json\r\n\
Content-Length: {}\r\nConnection: close\r\n\r\n{}",
body.len(),
body,
);
let _ = sock.write_all(resp.as_bytes()).await;
let _ = sock.shutdown().await;
}
});
let client = Client::builder()
.timeout(Duration::from_secs(5))
.build()
.unwrap();
let out = post(
&client,
&url,
&serde_json::json!({}),
false,
Duration::from_secs(5),
|b| b,
)
.await
.expect("post should succeed after retrying the malformed body");
assert_eq!(out, serde_json::json!({ "ok": true }));
assert_eq!(
accepts.load(Ordering::SeqCst),
2,
"server must see exactly 2 attempts (malformed body retried once)"
);
}
/// A persistently malformed 200 body exhausts MAX_RETRIES and surfaces
/// the terminal error with the `json:` detail plus cumulative
/// duration/attempt count — never an early first-attempt death.
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn post_exhausts_retries_on_persistent_malformed_json() {
use std::sync::atomic::{AtomicU32, Ordering};
use std::sync::Arc;
use tokio::io::{AsyncReadExt, AsyncWriteExt};
use tokio::net::TcpListener;
let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
let url = format!("http://{}/v1/x", listener.local_addr().unwrap());
let accepts = Arc::new(AtomicU32::new(0));
let accepts_srv = accepts.clone();
tokio::spawn(async move {
loop {
let (mut sock, _) = match listener.accept().await {
Ok(p) => p,
Err(_) => return,
};
accepts_srv.fetch_add(1, Ordering::SeqCst);
let mut buf = Vec::new();
let mut tmp = [0u8; 4096];
while !buf.windows(4).any(|w| w == b"\r\n\r\n") {
match sock.read(&mut tmp).await {
Ok(0) | Err(_) => return,
Ok(k) => buf.extend_from_slice(&tmp[..k]),
}
}
let body = "{\"choices\":[{\"mess";
let resp = format!(
"HTTP/1.1 200 OK\r\nContent-Type: application/json\r\n\
Content-Length: {}\r\nConnection: close\r\n\r\n{}",
body.len(),
body,
);
let _ = sock.write_all(resp.as_bytes()).await;
let _ = sock.shutdown().await;
}
});
let client = Client::builder()
.timeout(Duration::from_secs(5))
.build()
.unwrap();
let err = post(
&client,
&url,
&serde_json::json!({}),
false,
Duration::from_secs(5),
|b| b,
)
.await
.unwrap_err();
match &err {
PostError::Agent(AgentError::Llm(msg)) => {
assert!(
msg.contains("json:"),
"expected the json parse detail, got: {msg}"
);
assert!(
msg.contains("cumulative") && msg.contains("3 attempts"),
"expected cumulative duration + exact attempt count, got: {msg}"
);
}
other => panic!("expected PostError::Agent(AgentError::Llm), got: {other:?}"),
}
assert_eq!(
accepts.load(Ordering::SeqCst),
MAX_RETRIES,
"server must see exactly MAX_RETRIES attempts — malformed bodies must be retried"
);
}
/// Same regression coverage for `openrouter_post`, which carries its own
/// retry loop: a truncated 200 body on the first attempt is retried and
/// the call succeeds on the second.
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
async fn openrouter_post_retries_malformed_json_body_and_succeeds() {
use std::sync::atomic::{AtomicU32, Ordering};
use std::sync::Arc;
use tokio::io::{AsyncReadExt, AsyncWriteExt};
use tokio::net::TcpListener;
let listener = TcpListener::bind("127.0.0.1:0").await.unwrap();
let url = format!("http://{}/v1/x", listener.local_addr().unwrap());
let accepts = Arc::new(AtomicU32::new(0));
let accepts_srv = accepts.clone();
tokio::spawn(async move {
loop {
let (mut sock, _) = match listener.accept().await {
Ok(p) => p,
Err(_) => return,
};
let n = accepts_srv.fetch_add(1, Ordering::SeqCst);
let mut buf = Vec::new();
let mut tmp = [0u8; 4096];
while !buf.windows(4).any(|w| w == b"\r\n\r\n") {
match sock.read(&mut tmp).await {
Ok(0) | Err(_) => return,
Ok(k) => buf.extend_from_slice(&tmp[..k]),
}
}
if n == 0 {
let body = "{\"choices\":[{\"mess";
let resp = format!(
"HTTP/1.1 200 OK\r\nContent-Type: application/json\r\n\
Content-Length: {}\r\nConnection: close\r\n\r\n{}",
body.len(),
body,
);
let _ = sock.write_all(resp.as_bytes()).await;
let _ = sock.shutdown().await;
continue;
}
let body = r#"{"choices":[{"message":{"role":"assistant","content":"ok"},"finish_reason":"stop"}]}"#;
let resp = format!(
"HTTP/1.1 200 OK\r\nContent-Type: application/json\r\n\
Content-Length: {}\r\nConnection: close\r\n\r\n{}",
body.len(),
body,
);
let _ = sock.write_all(resp.as_bytes()).await;
let _ = sock.shutdown().await;
}
});
let client = Client::builder()
.timeout(Duration::from_secs(5))
.build()
.unwrap();
let out = openrouter_post(&client, &url, &json!({}), "key", Duration::from_secs(5))
.await
.expect("openrouter_post should succeed after retrying the malformed body");
assert!(out.is_object(), "expected a JSON object: {out:?}");
assert_eq!(
accepts.load(Ordering::SeqCst),
2,
"server must see exactly 2 attempts (malformed body retried once)"
);
}
/// A body-read timeout on the first attempt triggers a retry under an
/// escalated budget, and the call succeeds on the second attempt.
///