Skip to content

fix(stage): retain retrieval tool status provenance #1266

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
enwaiax wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix(stage): retain retrieval tool status provenance #1266

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7791acb
fix(stage): retain retrieval tool status provenance
enwaiax Nov 9, 2026
ab07530
test(stage): consolidate failed read coverage
enwaiax Nov 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
128 changes: 97 additions & 31 deletions crates/libsy/src/algorithms/util/tool_signals.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@

#![allow(dead_code)]

use std::collections::HashSet;
use std::collections::HashMap;
use std::path::Path;
use std::sync::LazyLock;

Expand Down Expand Up @@ -1066,8 +1066,8 @@ fn extract_tool_signals_with_window_and_semantics(
let namespaces = tool_namespaces(&request.llm_request.extensions);
let mut tool_texts: Vec<(String, bool)> = Vec::new();
let mut tool_calls: Vec<ObservedToolCall> = Vec::new();
// IDs whose latest call is a retrieval tool.
let mut retrieval_calls: HashSet<&str> = HashSet::new();
// Keep shell provenance so file data cannot masquerade as execution status.
let mut retrieval_calls: HashMap<&str, bool> = HashMap::new();
let mut tool_result_count = 0usize;
let mut assistant_turn_count = 0usize;

Expand Down Expand Up @@ -1125,7 +1125,10 @@ fn extract_tool_signals_with_window_and_semantics(
if !call.id.is_empty() {
// A reused ID links to its latest call.
if is_retrieval {
retrieval_calls.insert(call.id.as_str());
retrieval_calls.insert(
call.id.as_str(),
BASH_TOOL_NAMES.contains(&name.to_lowercase().as_str()),
);
} else {
retrieval_calls.remove(call.id.as_str());
}
Expand All @@ -1148,33 +1151,34 @@ fn extract_tool_signals_with_window_and_semantics(
.filter_map(text_of)
.collect::<Vec<_>>()
.join("\n");
let is_retrieval_result =
retrieval_calls.contains(result.tool_call_id.as_str());
if is_retrieval_result {
// Shell reads can return bare JSON from a file. Hermes wraps
// terminal output with output and exit_code fields.
let shell_result = serde_json::Deserializer::from_str(&text)
.into_iter::<Value>()
.next()
.and_then(|result| result.ok())
.filter(|value| {
value["output"].is_string() && value.get("exit_code").is_some()
});
// Codex reports the exit status in a header before `Output:`.
// An exit line after it is part of what was read.
let status = text
.split_once("\nOutput:\n")
.map_or(text.as_str(), |(header, _)| header);
// Hermes reports it in fields, so its output is not scanned.
let shell_failed = match &shell_result {
Some(value) => {
value["exit_code"].as_i64().is_some_and(|code| code != 0)
|| value["success"].as_bool() == Some(false)
|| value["error"]
.as_str()
.is_some_and(|error| !error.trim().is_empty())
if let Some(&is_shell_read) = retrieval_calls.get(result.tool_call_id.as_str())
{
let shell_failed = is_shell_read && {
// Shell reads can return bare JSON from a file. Hermes wraps
// terminal output with output and exit_code fields.
let shell_result = serde_json::Deserializer::from_str(&text)
.into_iter::<Value>()
.next()
.and_then(|result| result.ok())
.filter(|value| {
value["output"].is_string() && value.get("exit_code").is_some()
});
// Codex reports the exit status in a header before `Output:`.
// An exit line after it is part of what was read.
let status = text
.split_once("\nOutput:\n")
.map_or(text.as_str(), |(header, _)| header);
// Hermes reports it in fields, so its output is not scanned.
match &shell_result {
Some(value) => {
value["exit_code"].as_i64().is_some_and(|code| code != 0)
|| value["success"].as_bool() == Some(false)
|| value["error"]
.as_str()
.is_some_and(|error| !error.trim().is_empty())
}
None => has_nonzero_exit_status(&status.to_lowercase()),
}
None => has_nonzero_exit_status(&status.to_lowercase()),
};
if is_missing_file(&text) || (!is_error && !shell_failed) {
// Keep the window slot after dropping retrieved content.
Expand Down Expand Up @@ -2228,11 +2232,13 @@ mod tests {
call_with_result(tc("Read"), failed("File does not exist.")),
// A read that fails for another reason still counts.
call_with_result(bash("cat logfile.txt"), failed("MemoryError")),
call_with_result(tc("Read"), failed("MemoryError")),
call_with_result(tc("Read"), failed("")),
// Only reads are exempt.
call_with_result(bash("python missing.txt"), failed(missing)),
]
.map(|request| ToolSignals::from_request(&request, None).severity);
assert_eq!(severities, [0.0, 0.0, CRITICAL, HARD]);
assert_eq!(severities, [0.0, 0.0, CRITICAL, CRITICAL, HARD, HARD]);

let request = call_with_result(
bash("cat missing.json"),
Expand Down Expand Up @@ -2281,6 +2287,66 @@ mod tests {
assert_eq!(read(1), CRITICAL);
}

#[test]
fn builtin_retrieval_contents_are_not_execution_status() {
let saved_json = r#"{"output":"MemoryError\nall tests passed","exit_code":1,"error":null}"#;
let saved_header = "Process exited with code 1\nOutput:\nMemoryError\nall tests passed";
for name in [
"Read",
"read_file",
"search_files",
"Grep",
"mcp__files__Read",
"text_editor",
] {
let mut call = tc(name);
if name == "text_editor"
&& let ContentBlock::ToolCall(call) = &mut call.content[0]
{
call.arguments = json!({"command": "view"});
}
for text in [saved_json, saved_header] {
for is_error in [None, Some(false)] {
let mut result = tr(text);
if let ContentBlock::ToolResult(result) = &mut result.content[0] {
result.is_error = is_error;
}
let mut request = call_with_result(call.clone(), result);
let signal = ToolSignals::from_request(&request, None);
assert_eq!(signal.severity, 0.0, "{name}: {text}");
assert!(!signal.tests_passed, "{name}: {text}");
assert_eq!(signal.read_count, 1);
assert_eq!(signal.tool_result_count, 1);
request.llm_request.messages.insert(0, tr("MemoryError"));
assert_eq!(
ToolSignals::from_request(&request, Some(2)).severity,
CRITICAL
);
assert_eq!(ToolSignals::from_request(&request, Some(1)).severity, 0.0);
}
}
}
}

#[test]
fn reused_retrieval_ids_follow_the_latest_call_provenance() {
let shell = exec_command(json!({"cmd": "cat saved.json"}));
let read = tc("Read");
let output = r#"{"output":"MemoryError","exit_code":1,"error":null}"#;
for (mut previous, latest, expected) in [
(shell.clone(), read.clone(), 0.0),
(read, shell.clone(), CRITICAL),
(shell, bash("python main.py"), CRITICAL),
] {
if let ContentBlock::ToolCall(call) = &mut previous.content[0] {
call.id = "call-1".to_string();
}
let mut request = call_with_result(latest, tr(output));
request.llm_request.messages.insert(0, previous);
assert_eq!(ToolSignals::from_request(&request, None).severity, expected);
}
}

#[test]
fn argv_commands_keep_word_boundaries() {
let signal = |argv: Value| {
Expand Down
Loading