Keep the running tool call outside its group

A run of adjacent calls is drawn as one collapsed card, which hid the one
thing worth seeing without opening anything: the command the session is
running right now. It is a row of its own while it runs and folds back into
the run when it ends.

Grouping stays a display decision, so the pieces a running call cuts a run
into are keyed there. The first piece keeps the run's name -- that name is
what survives a page of history landing in front of it -- and later pieces
take their own first call's id behind it, since the call a run was named
after can itself be the one running.

The echo rig's /tools gap now runs between a call's start and its end rather
than between one call and the next, which is where a real session's time goes
and what makes the running state observable at all.

Checked with ktfmtFormat, compileDebugKotlin, testDebugUnitTest (new
ToolRowsTest) and lintDebug, cargo fmt/clippy/test, and on the emulator
against the sandbox: "Called 2 tools" with the live Bash card beneath it.
This commit is contained in:
iris-ai committed 2026-09-15 01:13:03 -04:00
1 parent b9b777acaf
commit 036eb375aa
5 files changed
+172 -45

No files matched your search

+8
View File
@@ -1098,6 +1098,14 @@ dev-updater (Kotlin 2.4.x, CMP 1.11.x, JDK 21).
4. **Session screen** — the core: 4. **Session screen** — the core:
- The transcript rendered from the event stream: markdown, inline images, - The transcript rendered from the event stream: markdown, inline images,
tool cards, question cards. tool cards, question cards.
- **A run of adjacent tool calls is one collapsed card, except for the
call still running** (2026-09-15). What the session is doing right now
is the one thing worth seeing without opening anything, and a heading
counting it hides it; the call folds back into its run the moment it
ends, which is the moment it stops being what is happening. Grouping is
a display decision (`groupToolRuns`) and a cut run's pieces are keyed
there — the first piece keeps the run's name, since that name is what
survives a page of history landing in front of it.
- **Anything that is a note *about* the conversation rather than a turn in - **Anything that is a note *about* the conversation rather than a turn in
it is closed by default** — a tool call, a peer message, a memory note. it is closed by default** — a tool call, a peer message, a memory note.
Open-ness is the screen's, never the card's: a card that remembered for Open-ness is the screen's, never the card's: a card that remembered for
@@ -302,10 +302,10 @@ fun SessionScreen(
// words were in. This is whatever was true as of the last frame. // words were in. This is whatever was true as of the last frame.
val selecting = selection.selectedTexts.isNotEmpty() val selecting = selection.selectedTexts.isNotEmpty()
var expandedTools by remember { mutableStateOf(setOf<String>()) } var expandedTools by remember { mutableStateOf(setOf<String>()) }
// Which runs of adjacent tool calls are open. Keyed by the first call's id, so a group survives // Which runs of adjacent tool calls are open, by the group row's own key, so a group survives
// more calls arriving after it. // more calls arriving after it.
var expandedGroups by remember { mutableStateOf(setOf<String>()) } var expandedGroups by remember { mutableStateOf(setOf<String>()) }
// Runs already drawn as a group, so the transition into one is noticed exactly once. // Calls already drawn inside a group, so being folded into one is noticed exactly once each.
var everGrouped by remember { mutableStateOf(setOf<String>()) } var everGrouped by remember { mutableStateOf(setOf<String>()) }
// Which messages from other agents are open, by the seq that identifies their row. Closed by // Which messages from other agents are open, by the seq that identifies their row. Closed by
// default, which is the rule for anything new in this transcript. // default, which is the rule for anything new in this transcript.
@@ -692,19 +692,23 @@ fun SessionScreen(
} }
} }
// A call opened on its own stays open when a second call in the same run turns it into a group. // A call opened on its own stays open when it is folded into a group -- either because a second
// Until this, watching a Bash call and having the session make another one shut the one being // call in the same run arrived, or because it finished and rejoined the run it was running
// read and folded it behind "Called 2 tools". // outside of. Until this, watching a Bash call and having the session make another one shut the
// one being read and folded it behind "Called 2 tools".
// //
// Considered once per run, at the moment it first becomes a group, and never again: after that // Per call rather than per group, because a group outlives the calls joining it: considered at
// the group's own toggle owns it. // the moment each call first lands inside one, and never again, so the reader who then shuts
// the group has shut it.
LaunchedEffect(rows) { LaunchedEffect(rows) {
val fresh = rows.filterIsInstance<TranscriptRow.Tools>().filter { it.id !in everGrouped } val fresh =
rows.filterIsInstance<TranscriptRow.Tools>().flatMap { group ->
group.calls.filter { it.id !in everGrouped }.map { group.key to it.id }
}
if (fresh.isEmpty()) return@LaunchedEffect if (fresh.isEmpty()) return@LaunchedEffect
expandedGroups = expandedGroups =
expandedGroups + expandedGroups + fresh.filter { it.second in expandedTools }.map { it.first }
fresh.filter { group -> group.calls.any { it.id in expandedTools } }.map { it.id } everGrouped = everGrouped + fresh.map { it.second }
everGrouped = everGrouped + fresh.map { it.id }
} }
// A compaction reports nothing about its own progress -- measured against the CLI, which says // A compaction reports nothing about its own progress -- measured against the CLI, which says
@@ -1624,13 +1628,13 @@ fun SessionScreen(
is TranscriptRow.Tools -> is TranscriptRow.Tools ->
ToolGroup( ToolGroup(
group = row, group = row,
expanded = row.id in expandedGroups, expanded = row.key in expandedGroups,
onToggle = { onToggle = {
toggleAnchored(row) { toggleAnchored(row) {
expandedGroups = expandedGroups =
if (row.id in expandedGroups) if (row.key in expandedGroups)
expandedGroups - row.id expandedGroups - row.key
else expandedGroups + row.id else expandedGroups + row.key
} }
}, },
isToolExpanded = { it in expandedTools }, isToolExpanded = { it in expandedTools },
@@ -61,7 +61,9 @@ sealed class TranscriptRow {
* *
* A tool row therefore keys on [TranscriptItem.ToolRun.runId] rather than on a sequence number, * A tool row therefore keys on [TranscriptItem.ToolRun.runId] rather than on a sequence number,
* and it is the *same* value whether the run is drawn as one card or as a group. Which value * and it is the *same* value whether the run is drawn as one card or as a group. Which value
* that is belongs to the item ([TranscriptItem.key]), not to a `when` here. * that is belongs to the item ([TranscriptItem.key]) everywhere a row is one thing; where
* [groupRuns] cuts a run into several rows it is the one deciding, and it says so by handing
* each piece its key.
*/ */
abstract val key: Any abstract val key: Any
@@ -75,23 +77,15 @@ sealed class TranscriptRow {
*/ */
abstract val startSeq: Long abstract val startSeq: Long
data class Single(val item: TranscriptItem) : TranscriptRow() { data class Single(val item: TranscriptItem, override val key: Any = item.key) :
override val key: Any TranscriptRow() {
get() = item.key
override val startSeq: Long override val startSeq: Long
get() = item.seq get() = item.seq
} }
/** Two or more calls with nothing between them; drawn as one collapsed card. */ /** Two or more calls with nothing between them; drawn as one collapsed card. */
data class Tools(val calls: List<TranscriptItem.ToolRun>) : TranscriptRow() { data class Tools(val calls: List<TranscriptItem.ToolRun>, override val key: String) :
/** The run's own name, which every call in it already carries. */ TranscriptRow() {
val id: String
get() = calls.first().runId
override val key: Any
get() = id
override val startSeq: Long override val startSeq: Long
get() = calls.first().seq get() = calls.first().seq
} }
@@ -102,6 +96,12 @@ sealed class TranscriptRow {
* *
* A single call is left alone: "Called 1 tool" hides a card to say the same thing in more words, * A single call is left alone: "Called 1 tool" hides a card to say the same thing in more words,
* and the run this exists for is the burst of five greps nobody wants to scroll past. * and the run this exists for is the burst of five greps nobody wants to scroll past.
*
* A call that has not finished is left alone too, wherever in its run it sits. What the session is
* doing *now* is the one thing worth seeing without opening anything, and grouping it hides the
* running command behind a heading that counts it. The call rejoins its run when it ends, which is
* the moment it stops being what is happening and becomes history -- and it is a row at the live
* end of the transcript, so nothing above the reader moves when it does.
*/ */
fun groupToolRuns(items: List<TranscriptItem>): List<TranscriptRow> = fun groupToolRuns(items: List<TranscriptItem>): List<TranscriptRow> =
DebugStats.timed("grouped tool runs") { groupRuns(items) } DebugStats.timed("grouped tool runs") { groupRuns(items) }
@@ -109,13 +109,24 @@ fun groupToolRuns(items: List<TranscriptItem>): List<TranscriptRow> =
private fun groupRuns(items: List<TranscriptItem>): List<TranscriptRow> { private fun groupRuns(items: List<TranscriptItem>): List<TranscriptRow> {
val rows = mutableListOf<TranscriptRow>() val rows = mutableListOf<TranscriptRow>()
var run = mutableListOf<TranscriptItem.ToolRun>() var run = mutableListOf<TranscriptItem.ToolRun>()
// The run being walked, and how many rows it has produced so far -- which is what decides the
// keys below.
var runId: String? = null
var emitted = 0
fun flush() { fun flush() {
when (run.size) { val first = run.firstOrNull() ?: return
0 -> {} // The first row a run produces keeps the run's name, which is the name that survives a page
1 -> rows += TranscriptRow.Single(run.first()) // of history landing in front of it ([adoptRun]); losing it is the transcript stepping
else -> rows += TranscriptRow.Tools(run.toList()) // under whoever is reading. The pieces a running call cuts off the back of the run have no
} // such name, so each takes its own first call's id, which is unique because a call id is.
// The run's name goes in front of it because the two can otherwise be the same string: the
// call a run was named after can itself be the one still running.
val key = if (emitted == 0) first.runId else "${first.runId}/${first.id}"
rows +=
if (run.size == 1) TranscriptRow.Single(first, key)
else TranscriptRow.Tools(run.toList(), key)
emitted++
run = mutableListOf() run = mutableListOf()
} }
@@ -124,11 +135,20 @@ private fun groupRuns(items: List<TranscriptItem>): List<TranscriptRow> {
// Adjacency is the same answer most of the time and a worse one at the edges: a call // Adjacency is the same answer most of the time and a worse one at the edges: a call
// arriving next to an existing run, or a page of history arriving in front of one, both // arriving next to an existing run, or a page of history arriving in front of one, both
// change which call is *first*. // change which call is *first*.
if (item is TranscriptItem.ToolRun && (run.isEmpty() || run.first().runId == item.runId)) { val call = item as? TranscriptItem.ToolRun
run += item if (call == null || call.runId != runId) {
} else {
flush() flush()
if (item is TranscriptItem.ToolRun) run += item else rows += TranscriptRow.Single(item) runId = call?.runId
emitted = 0
}
when {
call == null -> rows += TranscriptRow.Single(item)
call.done -> run += call
else -> {
flush()
run += call
flush()
}
} }
} }
flush() flush()
@@ -0,0 +1,93 @@
package com.example.aiapp
import kotlin.test.Test
import kotlin.test.assertEquals
import kotlin.test.assertTrue
/**
* How a run of tool calls is cut into rows: the call still running is drawn on its own, and every
* piece the cut leaves behind still has a key of its own -- two rows sharing one key take the app
* down, and a key that moves takes the reader's place with it.
*/
class ToolRowsTest {
private var seq = 0L
private fun call(id: String, runId: String = id, done: Boolean = true) =
TranscriptItem.ToolRun(
seq = ++seq,
id = id,
runId = runId,
tool = "Bash",
input = "{}",
output = if (done) "ok" else "",
done = done,
)
private fun shape(rows: List<TranscriptRow>) = rows.map { row ->
when (row) {
is TranscriptRow.Tools -> row.calls.map { it.id }
is TranscriptRow.Single -> listOf((row.item as TranscriptItem.ToolRun).id)
}
}
private fun assertKeysDistinct(rows: List<TranscriptRow>) =
assertEquals(rows.size, rows.map { it.key }.toSet().size, "$rows")
@Test
fun the_call_still_running_is_a_row_of_its_own() {
val rows =
groupToolRuns(
listOf(call("a"), call("b", runId = "a"), call("c", runId = "a", done = false))
)
assertEquals(listOf(listOf("a", "b"), listOf("c")), shape(rows))
assertKeysDistinct(rows)
}
@Test
fun a_call_running_in_the_middle_of_its_run_splits_the_group_in_two() {
val rows =
groupToolRuns(
listOf(
call("a"),
call("b", runId = "a", done = false),
call("c", runId = "a"),
call("d", runId = "a"),
)
)
assertEquals(listOf(listOf("a"), listOf("b"), listOf("c", "d")), shape(rows))
assertKeysDistinct(rows)
}
/**
* The one case where the run's name is a call that is not in the run's first row: a page of
* history joined onto a run whose own first call is still going ([joinPages] renames the older
* calls to the newer run's name). Both rows would key on that name.
*/
@Test
fun the_run_keeps_its_name_even_when_the_call_it_is_named_after_is_the_one_running() {
val rows = groupToolRuns(listOf(call("a", runId = "b"), call("b", done = false)))
assertEquals(listOf(listOf("a"), listOf("b")), shape(rows))
assertKeysDistinct(rows)
assertEquals("b", rows.first().key)
}
/** What happens the moment a command finishes: it folds back into the run it was cut out of. */
@Test
fun a_call_that_finishes_rejoins_its_run_without_moving_the_run() {
val running = listOf(call("a"), call("b", runId = "a", done = false))
val finished = listOf(running[0], (running[1] as TranscriptItem.ToolRun).copy(done = true))
val before = groupToolRuns(running)
val after = groupToolRuns(finished)
assertEquals(listOf(listOf("a", "b")), shape(after))
// The run keeps the key it was drawn under, so the list rebuilds a row rather than losing
// its anchor.
assertEquals(before.first().key, after.first().key)
}
@Test
fun a_run_of_finished_calls_is_still_one_group() {
val rows = groupToolRuns(listOf(call("a"), call("b", runId = "a"), call("c", runId = "a")))
assertEquals(listOf(listOf("a", "b", "c")), shape(rows))
assertTrue(rows.single() is TranscriptRow.Tools, "$rows")
}
}
+10 -8
View File
@@ -601,10 +601,15 @@ impl EchoDriver {
.and_then(|w| w.parse().ok()) .and_then(|w| w.parse().ok())
.unwrap_or(3usize) .unwrap_or(3usize)
.clamp(2, 12); .clamp(2, 12);
// How long to wait between calls, default none. A run that arrives // How long each call spends running, default none. A run that
// all at once cannot exercise a run *growing*: the case worth // arrives all at once cannot exercise a run *growing*: the case
// watching is a call somebody has opened and is reading when the // worth watching is a call somebody has opened and is reading when
// next one turns it into a group. // the next one turns it into a group. Spent between the call's start
// and its end rather than between one call and the next, because
// that is where a real session's time goes -- and a call is drawn
// outside its group while it runs, which is a state nothing could
// see while every call here ended a few milliseconds after it
// started.
let gap = Duration::from_secs( let gap = Duration::from_secs(
words words
.next() .next()
@@ -700,9 +705,6 @@ impl EchoDriver {
if let Some((count, gap)) = many_tools { if let Some((count, gap)) = many_tools {
for i in 1..=count { for i in 1..=count {
if i > 1 {
tokio::time::sleep(gap).await;
}
let id = format!("t-{}", super::random_hex()); let id = format!("t-{}", super::random_hex());
send(Event::ToolStart { send(Event::ToolStart {
id: id.clone(), id: id.clone(),
@@ -733,7 +735,7 @@ impl EchoDriver {
}); });
} }
} }
tokio::time::sleep(DELTA_DELAY).await; tokio::time::sleep(DELTA_DELAY + gap).await;
send(Event::ToolEnd { send(Event::ToolEnd {
id, id,
output: format!("call {i} finished"), output: format!("call {i} finished"),