A tool call can say it failed, and what it is for, without a renderer

P1b's pure half (docs/RUST.md). Three pieces, all testable with no
widget in sight:

- `event_model::Event::ToolEnd` gains `is_error`, read from the CLI's own
  `tool_result` field by both the live translator and the import replay
  (`import::tool_result_is_error`, one reader so the two cannot disagree
  about the same conversation). Without it a result is all a card has,
  and a broken call draws exactly as confidently as one that worked --
  the missing state, not a wrong one. `#[serde(default)]`, so an older
  transcript reads back as "not reported to have failed".
- `client_core::transcript_fold::ToolState`: Running, Deciding,
  Succeeded, Failed, NoResult. The pair it exists for is the last two
  against Succeeded-with-empty-output -- a call that printed nothing and
  a call whose result never arrived leave the same empty string, and only
  the session's status separates "still going" from "nobody found out".
- `client_core::tool_summary::parse_tool_input` and
  `client_core::durations`: `ToolInput.kt`'s subject/description/timeout
  split and `Durations.kt`'s span formatting, ported with their tests.

The echo driver's three-call run now has a failing middle call, so the
failed appearance is reachable from `ui-sandbox.sh` at all.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
irisandClaude Fable 5.1 committed 2026-09-06 19:46:21 -04:00
1 parent 69525bd131
commit 9079276ec8
9 files changed
+679 -4

No files matched your search

+100
View File
@@ -0,0 +1,100 @@
//! A span of milliseconds, written the way somebody reads it -- the port
//! of `Durations.kt`'s `formatMillis`/`formatMillisText`, with its tests.
//!
//! Only the tool-timeout half is here. `formatSpan` (the usage
//! countdown's rounding-up rule) belongs with whatever draws the usage
//! bar, and nothing in this crate needs it yet.
/// A span of milliseconds, written the way somebody reads it.
///
/// A tool's timeout arrives as `480000`, which nobody reads as eight
/// minutes. The rule has two halves, because a short span and a long one
/// are read for different things. Under a minute the question is "roughly
/// how long", so only the largest unit is shown and a fraction carries the
/// rest -- `2.5s`. At a minute or more the question is "how long exactly",
/// so every unit with something in it is written out -- `5d 12h 4m`. Empty
/// units are left out rather than written as zero.
///
/// Sub-second precision is dropped past a minute: nothing that takes days
/// is measured in milliseconds.
pub fn format_millis(ms: i64) -> String {
if ms < 0 {
return format!("-{}", format_millis(-ms));
}
if ms < 1000 {
return format!("{ms}ms");
}
if ms < 60_000 {
let tenths = (ms + 50) / 100;
let (whole, rest) = (tenths / 10, tenths % 10);
return if rest == 0 {
format!("{whole}s")
} else {
format!("{whole}.{rest}s")
};
}
let seconds = ms / 1000;
[
("d", seconds / 86_400),
("h", seconds / 3600 % 24),
("m", seconds / 60 % 60),
("s", seconds % 60),
]
.iter()
.filter(|(_, n)| *n > 0)
.map(|(unit, n)| format!("{n}{unit}"))
.collect::<Vec<_>>()
.join(" ")
}
/// `text` as a span when it is a whole number of milliseconds, and
/// unchanged when it is not.
pub fn format_millis_text(text: &str) -> String {
match text.trim().parse::<i64>() {
Ok(ms) => format_millis(ms),
Err(_) => text.to_string(),
}
}
#[cfg(test)]
mod tests {
use super::*;
/// The two ways a span of time is written here, and the rule each of
/// them follows -- ported from `DurationsTest.kt`, whose doc says why:
/// both are read off a screen to make a decision, so what matters is
/// that the shortest form that answers the question is what appears.
#[test]
fn under_a_minute_is_the_largest_unit_alone() {
assert_eq!(format_millis(30), "30ms");
assert_eq!(format_millis(999), "999ms");
assert_eq!(format_millis(1000), "1s");
assert_eq!(format_millis(2500), "2.5s");
// One decimal, rounded rather than cut: 2.46s is nearer two and a
// half than two and four.
assert_eq!(format_millis(2460), "2.5s");
assert_eq!(format_millis(59_900), "59.9s");
}
#[test]
fn a_minute_or_more_is_every_unit_that_has_something_in_it() {
// The figure this rule was written for: a tool timeout, which
// arrives as milliseconds and is unreadable as 480000.
assert_eq!(format_millis(480_000), "8m");
assert_eq!(format_millis(60_000), "1m");
assert_eq!(format_millis(90_000), "1m 30s");
assert_eq!(format_millis(475_440_000), "5d 12h 4m");
// Empty units are left out rather than written as zero: the labels
// say which is which, and "5d 0h 4m" is only longer.
assert_eq!(format_millis(432_240_000), "5d 4m");
}
#[test]
fn only_a_whole_number_of_milliseconds_is_rewritten() {
assert_eq!(format_millis_text(" 480000 "), "8m");
// A timeout a tool expressed some other way is its own words,
// passed through rather than guessed at.
assert_eq!(format_millis_text("2 minutes"), "2 minutes");
assert_eq!(format_millis_text(""), "");
}
}
+2
View File
@@ -5,11 +5,13 @@
pub mod ansi;
pub mod api;
pub mod config;
pub mod durations;
pub mod event_stream;
pub mod highlight;
pub mod markdown_blocks;
pub mod notifications;
pub mod sse;
pub mod tool_summary;
pub mod transcript_cache;
pub mod transcript_fold;
pub mod transcript_source;
+244
View File
@@ -0,0 +1,244 @@
//! A tool call's input, read rather than dumped -- the port of
//! `ToolInput.kt`'s `parseToolInput`, which is what both the collapsed
//! card's one-line summary and the expanded card's key/value list are
//! derived from.
//!
//! Every tool's input arrives as JSON, and showing it raw makes the reader
//! parse `{"command":"…","timeout":120000}` themselves to find the one
//! line they care about. So the fields that carry the meaning are pulled
//! out, and anything left over is still shown, because dropping a field
//! would be claiming the tool has no other input when it might.
//!
//! Pure, and here rather than in the widget crate, for the reason the rest
//! of this crate exists: the derivation is the same on a phone and on a
//! desktop, and it is testable without a renderer.
use crate::durations::format_millis_text;
use crate::highlight::Language;
use serde_json::{Map, Value};
/// A tool call's input, split into the parts a card draws separately.
#[derive(Debug, Clone, PartialEq, Eq, Default)]
pub struct ToolInput {
/// The thing that will actually be run or read, if this tool has one.
pub subject: Option<String>,
/// The language [`ToolInput::subject`] is written in, for
/// highlighting.
pub language: Option<Language>,
/// The tool's own one-line summary, when it wrote one.
pub description: Option<String>,
/// How long the call may take, in the largest units it fits. Shown
/// apart because it is a limit on the call rather than part of what
/// the call does.
pub timeout: Option<String>,
/// Everything else, as `name: value` lines. Never dropped.
pub rest: Vec<String>,
}
impl ToolInput {
/// The one line to show when there is only room for one: what this
/// call is for.
pub fn title(&self) -> Option<&str> {
self.description
.as_deref()
.or(self.subject.as_deref())
// A subject that is only whitespace would draw as an empty
// summary line, which reads as a tool with nothing to say
// rather than as one whose subject was blank.
.filter(|t| !t.trim().is_empty())
}
}
/// Which field of which tool is the subject.
///
/// A table rather than a chain of `if`s: adding a tool is a row, and the
/// shape stops any of them from being the special case that gets its own
/// code path. Unknown tools fall through to "no subject, everything is
/// rest".
const SUBJECTS: &[(&str, &str, Option<Language>)] = &[
("Bash", "command", Some(Language::Shell)),
("Read", "file_path", None),
("Write", "file_path", None),
("Edit", "file_path", None),
("Glob", "pattern", None),
("Grep", "pattern", None),
("WebFetch", "url", None),
];
/// Fields that are the tool's own prose about itself rather than input to
/// it.
const DESCRIPTIONS: &[&str] = &["description", "prompt"];
/// One JSON value as the Kotlin's `JSONObject.optString`/`get` wrote it: a
/// string is its own characters, anything else is its JSON form.
///
/// One function rather than two, because the same coercion decides both
/// what a subject reads as and what a leftover field's value reads as, and
/// two copies would eventually disagree about a number.
fn as_text(value: &Value) -> String {
match value {
Value::String(s) => s.clone(),
other => other.to_string(),
}
}
fn non_blank(value: Option<&Value>) -> Option<String> {
let text = as_text(value?);
(!text.trim().is_empty()).then_some(text)
}
/// Split `input` (a tool call's JSON) into the parts a card draws.
///
/// Input that is not a JSON object -- older transcripts and some tools
/// send a bare string -- is still the input, so it is still shown, as the
/// whole of `rest`.
pub fn parse_tool_input(tool: &str, input: &str) -> ToolInput {
let Ok(Value::Object(json)) = serde_json::from_str::<Value>(input) else {
return ToolInput {
rest: match input.trim().is_empty() {
true => Vec::new(),
false => vec![input.to_string()],
},
..ToolInput::default()
};
};
parse_object(tool, &json)
}
fn parse_object(tool: &str, json: &Map<String, Value>) -> ToolInput {
let (subject_key, language) = SUBJECTS
.iter()
.find(|(name, ..)| *name == tool)
.map(|(_, key, language)| (Some(*key), *language))
.unwrap_or((None, None));
let subject = subject_key.and_then(|key| non_blank(json.get(key)));
let description = DESCRIPTIONS
.iter()
.find_map(|key| non_blank(json.get(*key)));
let timeout = non_blank(json.get("timeout")).map(|t| format_millis_text(&t));
// Sorted, so the leftovers are in the same order every time this call
// is drawn rather than in whatever order the JSON happened to arrive
// in. A field is left out only when it is already drawn somewhere
// else on the card.
let mut keys: Vec<&String> = json
.keys()
.filter(|k| Some(k.as_str()) != subject_key || subject.is_none())
.filter(|k| !DESCRIPTIONS.contains(&k.as_str()) || description.is_none())
.filter(|k| k.as_str() != "timeout" || timeout.is_none())
.collect();
keys.sort();
let rest = keys
.into_iter()
.map(|key| format!("{key}: {}", as_text(&json[key])))
.collect();
ToolInput {
subject,
language,
description,
timeout,
rest,
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn each_tool_in_the_table_has_its_own_subject() {
// One assertion per row of `SUBJECTS`, because the table is the
// whole of the rule and a row lost in an edit would otherwise
// only show up as a card with no summary line.
let cases = [
("Bash", r#"{"command":"ls -la"}"#, "ls -la"),
("Read", r#"{"file_path":"/tmp/x.rs"}"#, "/tmp/x.rs"),
("Write", r#"{"file_path":"/tmp/y.rs"}"#, "/tmp/y.rs"),
("Edit", r#"{"file_path":"/tmp/z.rs"}"#, "/tmp/z.rs"),
("Glob", r#"{"pattern":"**/*.rs"}"#, "**/*.rs"),
("Grep", r#"{"pattern":"fn main"}"#, "fn main"),
("WebFetch", r#"{"url":"https://x/y"}"#, "https://x/y"),
];
for (tool, input, expected) in cases {
let parsed = parse_tool_input(tool, input);
assert_eq!(parsed.subject.as_deref(), Some(expected), "{tool}");
assert_eq!(parsed.title(), Some(expected), "{tool}");
assert!(parsed.rest.is_empty(), "{tool}: {:?}", parsed.rest);
}
assert_eq!(
parse_tool_input("Bash", r#"{"command":"ls"}"#).language,
Some(Language::Shell),
"a Bash command is shell, and is the one row that names a language"
);
}
#[test]
fn a_tools_own_description_is_what_the_one_line_says() {
// The description wins over the subject: it is the tool's own
// prose about what this call is for, which is what a reader
// scanning a collapsed run is looking for.
let parsed = parse_tool_input(
"Bash",
r#"{"command":"cargo test -p iris","description":"Run the iris tests"}"#,
);
assert_eq!(parsed.title(), Some("Run the iris tests"));
assert_eq!(parsed.subject.as_deref(), Some("cargo test -p iris"));
assert!(parsed.rest.is_empty(), "{:?}", parsed.rest);
}
#[test]
fn a_timeout_is_read_as_a_span_and_kept_apart_from_the_rest() {
let parsed = parse_tool_input("Bash", r#"{"command":"sleep 500","timeout":480000}"#);
assert_eq!(parsed.timeout.as_deref(), Some("8m"));
assert!(parsed.rest.is_empty(), "{:?}", parsed.rest);
}
#[test]
fn every_field_not_drawn_elsewhere_is_still_shown() {
// The half the "never dropped" promise is about: a tool this
// build has never heard of has no subject, so *everything* is
// rest -- and a known tool's extra fields are too.
let parsed = parse_tool_input(
"Edit",
r#"{"file_path":"/a.rs","old_string":"x","new_string":"y","replace_all":true}"#,
);
assert_eq!(
parsed.rest,
vec![
"new_string: y".to_string(),
"old_string: x".to_string(),
"replace_all: true".to_string(),
],
"sorted, and a non-string value written as JSON"
);
let unknown = parse_tool_input("SomeNewTool", r#"{"b":2,"a":"one"}"#);
assert_eq!(unknown.subject, None);
assert_eq!(unknown.rest, vec!["a: one".to_string(), "b: 2".to_string()]);
}
#[test]
fn input_that_is_not_an_object_is_still_the_input() {
// Older transcripts and some tools send a bare string; a card
// that dropped it would claim the call had no input at all.
assert_eq!(
parse_tool_input("Bash", "just a string").rest,
vec!["just a string".to_string()]
);
assert_eq!(parse_tool_input("Bash", " ").rest, Vec::<String>::new());
assert_eq!(parse_tool_input("Bash", "").title(), None);
}
#[test]
fn a_blank_subject_is_no_subject_rather_than_an_empty_summary_line() {
let parsed = parse_tool_input("Bash", r#"{"command":" ","other":1}"#);
assert_eq!(parsed.subject, None);
assert_eq!(parsed.title(), None);
// Not dropped just because it was blank -- it is still a field
// the call carried.
assert_eq!(
parsed.rest,
vec!["command: ".to_string(), "other: 1".to_string()]
);
}
}
+241 -2
View File
@@ -78,6 +78,11 @@ pub enum TranscriptItem {
input: String,
output: String,
done: bool,
/// Whether the result that arrived said the call failed
/// ([`Event::ToolEnd`]'s `is_error`). Meaningless while `done` is
/// false, and [`ToolState::of`] is the only thing that reads the
/// pair, so the two cannot be combined wrongly at a call site.
failed: bool,
asks: Vec<QuestionCard>,
images: Vec<String>,
},
@@ -352,6 +357,7 @@ pub fn join_pages(earlier: &[TranscriptItem], later: &[TranscriptItem]) -> Vec<T
let &TranscriptItem::ToolRun {
ref output,
done,
failed,
asks: ref half_asks,
images: ref half_images,
..
@@ -367,6 +373,7 @@ pub fn join_pages(earlier: &[TranscriptItem], later: &[TranscriptItem]) -> Vec<T
input,
output: output.clone(),
done,
failed,
// Kept from both halves: a question or an image can be
// attached to either, depending on which side of the
// boundary its event fell.
@@ -562,6 +569,7 @@ pub fn fold_event(items: &[TranscriptItem], entry: &SeqEvent) -> Vec<TranscriptI
input: input.to_string(),
output: String::new(),
done: false,
failed: false,
asks: Vec::new(),
images: Vec::new(),
});
@@ -572,15 +580,23 @@ pub fn fold_event(items: &[TranscriptItem], entry: &SeqEvent) -> Vec<TranscriptI
*out = output.clone();
}
}),
Event::ToolEnd { id, output } => {
Event::ToolEnd {
id,
output,
is_error,
} => {
if items.iter().any(|i| i.as_tool_run() == Some(id.as_str())) {
update_tool(items, id, |item| {
if let TranscriptItem::ToolRun {
output: out, done, ..
output: out,
done,
failed,
..
} = item
{
*out = output.clone();
*done = true;
*failed = *is_error;
}
})
} else {
@@ -594,6 +610,7 @@ pub fn fold_event(items: &[TranscriptItem], entry: &SeqEvent) -> Vec<TranscriptI
input: String::new(),
output: output.clone(),
done: true,
failed: *is_error,
asks: Vec::new(),
images: Vec::new(),
});
@@ -650,6 +667,7 @@ pub fn fold_event(items: &[TranscriptItem], entry: &SeqEvent) -> Vec<TranscriptI
input,
output,
done,
failed,
images,
} if asks.iter().any(|a| &a.id == id) => {
for ask in asks.iter_mut() {
@@ -665,6 +683,7 @@ pub fn fold_event(items: &[TranscriptItem], entry: &SeqEvent) -> Vec<TranscriptI
input,
output,
done,
failed,
asks,
images,
}
@@ -750,6 +769,74 @@ pub fn fold_event(items: &[TranscriptItem], entry: &SeqEvent) -> Vec<TranscriptI
}
}
/// What became of one tool call -- every state a card has to be able to
/// draw, including the two that are not answers.
///
/// The pair this enum exists for is [`ToolState::Succeeded`] against
/// [`ToolState::NoResult`]. A call that finished having printed nothing
/// and a call whose result never arrived both leave an empty `output`,
/// and drawing them the same way states a verdict nobody reached: "it
/// worked and said nothing" reads as a fact, where the truth is that the
/// turn ended before anything came back.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ToolState {
/// Started, no result yet, and the session is still working -- the
/// ordinary state of a call in flight.
Running,
/// Stopped on the reader: a permission or question this call carries
/// has not been answered, so nothing is happening until somebody
/// answers it. Distinct from [`Self::Running`] because whose move it
/// is differs, which is the Compose card's "your turn".
Deciding,
/// A result arrived and the tool did not report a failure.
Succeeded,
/// A result arrived and the tool reported that the call failed
/// (`is_error`).
Failed,
/// No result ever arrived and the session is not working any more --
/// the turn was interrupted, or the process went away. Not a verdict
/// on the call: it says only that nobody found out.
NoResult,
}
impl ToolState {
/// The state of one call. `session_working` is
/// [`session_working`]'s answer for the session this call is in --
/// the only thing here that is not a property of the call itself, and
/// what separates "still running" from "never came back".
///
/// Written once, over the fields rather than per call site, because
/// the five states are decided by four conditions and every place
/// that re-derived a subset of them got a different subset.
pub fn of(item: &TranscriptItem, session_working: bool) -> Option<Self> {
let TranscriptItem::ToolRun {
done, failed, asks, ..
} = item
else {
return None;
};
debug_assert!(
!failed || *done,
"a call cannot have failed before its result arrived"
);
Some(if asks.iter().any(|ask| ask.answers.is_empty()) {
// Ahead of `done`: a call waiting on permission has not
// finished either, and which of the two the reader is being
// told about is the one they can act on.
Self::Deciding
} else if !*done {
match session_working {
true => Self::Running,
false => Self::NoResult,
}
} else if *failed {
Self::Failed
} else {
Self::Succeeded
})
}
}
/// One row as the transcript draws it: a run of consecutive tool calls, or
/// anything else. Ported from `ToolRows.kt`'s `TranscriptRow` and
/// `groupToolRuns` -- the Compose card rendering in that file is not part
@@ -989,6 +1076,7 @@ mod tests {
Event::ToolEnd {
id: "x".to_string(),
output: "done".to_string(),
is_error: false,
},
)]);
assert_eq!(
@@ -1001,6 +1089,7 @@ mod tests {
input: String::new(),
output: "done".to_string(),
done: true,
failed: false,
asks: Vec::new(),
images: Vec::new(),
}]
@@ -1166,6 +1255,7 @@ mod tests {
Event::ToolEnd {
id: id.to_string(),
output: output.to_string(),
is_error: false,
},
)
}
@@ -1211,6 +1301,7 @@ mod tests {
input: "{}".to_string(),
output: "the result".to_string(),
done: true,
failed: false,
asks: Vec::new(),
images: Vec::new(),
}],
@@ -1269,6 +1360,7 @@ mod tests {
input: "{}".to_string(),
output: String::new(),
done: false,
failed: false,
asks: Vec::new(),
images: Vec::new(),
}];
@@ -1279,3 +1371,150 @@ mod tests {
}
}
}
/// [`ToolState`] is what a card colours itself by, so each of its five
/// states is asserted from the events that actually produce it rather than
/// from a hand-built item -- a mapping that agreed with a fixture and
/// disagreed with the fold would be invisible until it was on screen.
#[cfg(test)]
mod tool_state_tests {
use super::*;
fn event(seq: u64, e: Event) -> SeqEvent {
SeqEvent {
seq,
ts: 0.0,
event: e,
}
}
fn fold_all(events: &[SeqEvent]) -> Vec<TranscriptItem> {
events
.iter()
.fold(Vec::new(), |items, e| fold_event(&items, e))
}
fn start(id: &str) -> SeqEvent {
event(
1,
Event::ToolStart {
id: id.to_string(),
tool: "Bash".to_string(),
input: serde_json::json!({"command": "ls"}),
},
)
}
fn end(id: &str, output: &str, is_error: bool) -> SeqEvent {
event(
2,
Event::ToolEnd {
id: id.to_string(),
output: output.to_string(),
is_error,
},
)
}
fn state_of(events: &[SeqEvent], session_working: bool) -> ToolState {
let items = fold_all(events);
ToolState::of(&items[0], session_working).expect("the fixture's first item is a tool call")
}
#[test]
fn a_result_that_arrived_is_read_from_is_error() {
assert_eq!(
state_of(&[start("a"), end("a", "ok", false)], false),
ToolState::Succeeded
);
assert_eq!(
state_of(&[start("a"), end("a", "No such file", true)], false),
ToolState::Failed
);
}
/// The pair this enum exists for. Both calls have an empty `output`
/// and nothing else distinguishes them, so a card that only looked at
/// the text would draw the interrupted one as a call that ran fine and
/// printed nothing.
#[test]
fn a_call_that_printed_nothing_is_not_a_call_that_never_answered() {
assert_eq!(
state_of(&[start("a"), end("a", "", false)], false),
ToolState::Succeeded,
"a result arrived; it was empty"
);
assert_eq!(
state_of(&[start("a")], false),
ToolState::NoResult,
"no result, and the session is not working any more"
);
}
/// The same call, mid-turn: still running rather than abandoned. The
/// only thing separating the two is the session's own status, which is
/// why `of` takes it.
#[test]
fn no_result_while_the_session_works_is_still_running() {
assert_eq!(state_of(&[start("a")], true), ToolState::Running);
}
#[test]
fn an_unanswered_ask_is_the_readers_move_whatever_else_is_true() {
let asking = event(
3,
Event::Question {
id: "q1".to_string(),
prompt: "Allow?".to_string(),
header: None,
options: vec![QuestionOption {
label: "Allow".to_string(),
description: None,
preview: None,
}],
multi_select: false,
about: Some("a".to_string()),
},
);
let answered = event(
4,
Event::Answered {
id: "q1".to_string(),
answers: vec!["Allow".to_string()],
},
);
// Ahead of both "still running" and "no result": the reader can
// act on this one, and cannot act on either of those.
assert_eq!(
state_of(&[start("a"), asking.clone()], true),
ToolState::Deciding
);
assert_eq!(
state_of(&[start("a"), asking.clone()], false),
ToolState::Deciding
);
assert_eq!(
state_of(
&[start("a"), asking, answered, end("a", "ok", false)],
false
),
ToolState::Succeeded,
"once it is answered the call is an ordinary one again"
);
}
#[test]
fn nothing_but_a_tool_call_has_a_tool_state() {
assert_eq!(
ToolState::of(
&TranscriptItem::UserMsg {
seq: 1,
text: "hi".to_string(),
attachments: Vec::new(),
},
true
),
None
);
}
}