chat: one tool-call summary chip per turn, opening the call list

The chat transcript no longer renders a chip per tool call: `ToolCalls`
(`android/.../ui/chat/ChatToolCalls.kt`) now draws a single summary chip
reading `N tools called` (`1 tool called` for one) as the turn's only
tool-call surface. While the turn streams the chip shows the call count
growing and the existing spinner while any call is in flight, plus the
existing `· long` mark for a running call past `TOOL_LONG_RUNNING_MILLIS`;
both are gated on the turn still streaming (`live = streaming` in
`AgentTurnItem`), so a settled chip never spins or fakes a long mark.

Tapping the chip opens the existing `ToolPanel` as a popup listing every
call of the turn, in the order the agent made them: one independently
collapsible row per call naming it, showing its one-line detail, its
`+N -M` edit stat and its duration, expanding in place to the
pretty-printed INPUT, the synthesised edit diff, then the ERROR block for
a failed call or the OUTPUT (`(no output)` when empty). The list is
bounded by the new `TOOL_PANEL_MAX_HEIGHT_DP` (`Dimens.kt`, purpose-named)
and scrolls inside the panel, so many calls never grow the popup past the
window nor re-flow the transcript. The chip announces the count ("N tool
calls, expand for details") through the app's accessibility treatment.

Pure logic lives in `ChatFormat.kt` (`toolCallsLabel`,
`toolCallsAnnouncement`, `toolCallCount`, `toolsRunning`, `toolCallRow`)
with unit cases added to the existing `ChatFormatTest`; no raw dimension
was introduced, so `DimensTest` is unchanged.

Checked for staleness: the packaged authoring skill references
(`internal/author/skill/references/`) and the packaged templates carry no
Android-chat content, so neither needs updating — this change is
`android/`-only. The shared docs this change makes stale
(`android/README.md`'s "Tool-call chips and the expandable panel" section
and `docs/chat-verification.md`'s tool-chip row) are recorded in
`docs/notes/chat-tool-group.md` for the single `docs-fold` writer rather
than edited here.
This commit is contained in:
2026-10-18 09:58:28 +01:00
parent 5a4cfc2e9c
commit 7fbbd95374
6 changed files with 418 additions and 124 deletions
@@ -80,6 +80,34 @@ internal fun formatToolDuration(millis: Long): String = when {
/** The chip label for a tool call: its name, or a placeholder when unnamed. */
internal fun toolLabel(part: ChatPart): String = part.name.ifBlank { "tool" }
/**
* A turn's one summary chip: `N tools called`, and `1 tool called` for a single
* call — a count of one is never pluralised. The chip is the transcript's only
* tool-call surface, so this label stands in for every call of the turn.
*/
internal fun toolCallsLabel(count: Int): String =
if (count == 1) "1 tool called" else "$count tools called"
/**
* What the summary chip announces to a screen reader, since the individual
* calls only exist inside the popup: the count, plus how to reach the detail.
* Singular and plural follow [toolCallsLabel] so the two never disagree.
*/
internal fun toolCallsAnnouncement(count: Int): String =
if (count == 1) "1 tool call, expand for details" else "$count tool calls, expand for details"
/**
* How many tool calls [parts] holds. Counted by kind rather than by
* `parts.size`, so a turn that also carries text or reasoning never inflates
* the number the summary chip shows.
*/
internal fun toolCallCount(parts: List<ChatPart>): Int =
parts.count { it.type == ChatProtocol.PART_TOOL }
/** True while any tool call in [parts] is still in flight. */
internal fun toolsRunning(parts: List<ChatPart>): Boolean =
parts.any { it.type == ChatProtocol.PART_TOOL && ChatProtocol.toolRunning(it.state) }
/**
* The in-flight tool calls across a conversation's [messages], in transcript
* order, so the conversation's indicator and the session row read the same set
@@ -131,6 +159,36 @@ internal fun toolDetail(part: ChatPart): String? {
return null
}
/**
* Everything one row of the tool panel shows for a single call, derived once so
* the row, the summary chip and their unit test agree: the call's real name (the
* name the earlier frames taught it, only falling back to `tool` when none was
* ever seen), its one-line detail, its duration, the pretty-printed INPUT, the
* OUTPUT, the ERROR text for a failed call and the synthesised edit diff.
* [error] is null for a call that did not fail, and [diff] is null for every
* non-edit call, so the row shows only the sections that have content.
*/
internal data class ToolCallRow(
val name: String,
val detail: String?,
val duration: String?,
val input: String,
val output: String,
val error: String?,
val diff: ToolDiff?,
)
/** Builds the one row a turn's tool panel shows for [part]. */
internal fun toolCallRow(part: ChatPart): ToolCallRow = ToolCallRow(
name = toolLabel(part),
detail = toolDetail(part),
duration = toolDurationMillis(part)?.let(::formatToolDuration),
input = prettyToolInput(part.state?.input),
output = toolOutputText(part.state),
error = if (toolFailed(part.state)) toolErrorText(part.state) else null,
diff = toolEditDiff(part.state?.input),
)
/**
* How long a running [part] has been going at [nowMillis]. The protocol's own
* timing wins whenever the part carries it (a running call from stored history
@@ -4,14 +4,15 @@ import androidx.compose.foundation.clickable
import androidx.compose.foundation.layout.Arrangement
import androidx.compose.foundation.layout.Box
import androidx.compose.foundation.layout.Column
import androidx.compose.foundation.layout.ExperimentalLayoutApi
import androidx.compose.foundation.layout.FlowRow
import androidx.compose.foundation.layout.Row
import androidx.compose.foundation.layout.fillMaxWidth
import androidx.compose.foundation.layout.heightIn
import androidx.compose.foundation.layout.padding
import androidx.compose.foundation.layout.size
import androidx.compose.foundation.layout.widthIn
import androidx.compose.foundation.rememberScrollState
import androidx.compose.foundation.shape.RoundedCornerShape
import androidx.compose.foundation.verticalScroll
import androidx.compose.material.icons.Icons
import androidx.compose.material.icons.filled.Build
import androidx.compose.material.icons.filled.Close
@@ -25,6 +26,7 @@ import androidx.compose.material3.IconButton
import androidx.compose.material3.MaterialTheme
import androidx.compose.material3.Surface
import androidx.compose.material3.Text
import androidx.compose.material3.minimumInteractiveComponentSize
import androidx.compose.runtime.Composable
import androidx.compose.runtime.getValue
import androidx.compose.runtime.mutableStateOf
@@ -33,6 +35,9 @@ import androidx.compose.runtime.setValue
import androidx.compose.ui.Alignment
import androidx.compose.ui.Modifier
import androidx.compose.ui.platform.LocalDensity
import androidx.compose.ui.semantics.Role
import androidx.compose.ui.semantics.contentDescription
import androidx.compose.ui.semantics.semantics
import androidx.compose.ui.text.style.TextOverflow
import androidx.compose.ui.unit.IntOffset
import androidx.compose.ui.unit.IntRect
@@ -48,6 +53,7 @@ import net.awan.pipelines.ui.components.CodeBlock
import net.awan.pipelines.ui.theme.ICON_BUTTON_SMALL_DP
import net.awan.pipelines.ui.theme.ICON_INLINE_DP
import net.awan.pipelines.ui.theme.ICON_STATUS_DP
import net.awan.pipelines.ui.theme.TOOL_PANEL_MAX_HEIGHT_DP
import net.awan.pipelines.ui.theme.TOOL_PANEL_MAX_WIDTH_DP
/** The tool panel floats over the transcript; these bound its width and offsets. */
@@ -57,109 +63,115 @@ private val TOOL_PANEL_MARGIN = 8.dp
private val TOOL_PANEL_ELEVATION = 12.dp
/**
* One turn's tool calls as a single group of compact chips, mirroring the
* reference transcript's `_ToolCalls`/`_ToolChip` (research §1.3): each call
* shows its name (or the edited file), its duration and a spinner while it runs,
* and the chips of the turn wrap together on as few lines as they fit instead of
* each taking a row.
* One turn's tool calls behind a single summary chip: `N tools called`, with the
* existing spinner while any call is still in flight and a `· long` mark while a
* running call has passed [TOOL_LONG_RUNNING_MILLIS]. A settled turn never shows
* either mark, so a finished turn cannot fake work it is not doing.
*
* Tapping a chip opens its INPUT (pretty JSON) / OUTPUT (monospace) panel as an
* overlay anchored to that chip (research §1.3 `_buildPanel`), dismissed by its
* own X or by tapping outside. Because the panel is a popup and not part of the
* transcript layout, opening or closing it never re-flows the conversation nor
* moves its scroll position — the transcript stays exactly where the operator
* left it.
* The chip is the transcript's only tool-call surface: tapping it opens the tool
* panel as a popup listing every call of the turn in the order the agent made
* them, each row collapsing independently. Because the panel is a popup and not
* part of the transcript layout, opening or closing it never re-flows the
* conversation nor moves its scroll position.
*
* An `edit`-family call expands to a diff card (research §1.3 `_CodeRun` /
* `_DiffView`): the `+N -M` stat on the chip and a green/red unified diff derived
* from the arguments, since the brokered protocol carries no diff field. An
* `error` call is visibly failed — an error-tinted chip and error icon — and its
* panel shows the error text.
* [elapsedMillis] is the live age of a call — the conversation supplies it from
* its own clock, so the `long` mark is derived from real elapsed time rather
* than guessed. [live] is whether the turn is still streaming: a stored call can
* carry a stale `running` status long after the turn ended, and a settled chip
* must not spin (or wear the `long` mark) on that.
*/
@OptIn(ExperimentalLayoutApi::class)
@Composable
internal fun ToolCalls(parts: List<ChatPart>) {
// One id at a time: the reference shows a single panel, and a focusable popup
// dismisses the previous one on open. No key is needed: an id that is no
// longer among the calls simply matches nothing, so a streaming rebuild keeps
// the panel the operator just opened without allocating a key list per frame.
var openId by remember { mutableStateOf<String?>(null) }
FlowRow(
modifier = Modifier.fillMaxWidth().padding(vertical = 4.dp),
horizontalArrangement = Arrangement.spacedBy(CHAT_ITEM_GAP),
verticalArrangement = Arrangement.spacedBy(CHAT_ITEM_GAP),
) {
parts.forEach { part ->
ToolChip(
part = part,
expanded = openId == part.id,
onToggle = { openId = if (openId == part.id) null else part.id },
onDismiss = { openId = null },
)
internal fun ToolCalls(parts: List<ChatPart>, live: Boolean, elapsedMillis: (ChatPart) -> Long = { 0L }) {
val calls = parts.filter { it.type == ChatProtocol.PART_TOOL }
if (calls.isEmpty()) return
// One panel at a time: the chip toggles it, and the panel's own X or a tap
// outside dismisses it. No key is needed: the calls belong to this turn.
var open by remember { mutableStateOf(false) }
val running = live && toolsRunning(calls)
val failed = calls.any { toolFailed(it.state) }
// The "long" mark is a running-call signal only: it is meaningful on the
// summary chip while a call is in flight, so a settled chip never wears it.
val long = running && calls.any { toolLongRunning(elapsedMillis(it)) }
Box(Modifier.fillMaxWidth().padding(vertical = CHAT_ITEM_GAP)) {
ToolCallsChip(
count = toolCallCount(calls),
running = running,
failed = failed,
long = long,
expanded = open,
onToggle = { open = !open },
)
if (open) {
Popup(
popupPositionProvider = rememberToolPanelPositionProvider(),
onDismissRequest = { open = false },
properties = PopupProperties(focusable = true),
) {
ToolPanel(calls = calls, onDismiss = { open = false })
}
}
}
}
/** A compact tool-call chip: kind icon, label, diff stat, duration and chevron. */
/**
* The turn's one summary chip: the call count, the running spinner (or a state
* icon once settled), the optional `long` mark and a chevron naming the popup
* it opens. Its content description carries the count and how to reach the
* calls, so the summary is understandable without opening the panel.
*/
@Composable
private fun ToolChip(part: ChatPart, expanded: Boolean, onToggle: () -> Unit, onDismiss: () -> Unit) {
val running = ChatProtocol.toolRunning(part.state)
val failed = toolFailed(part.state)
val diff = toolEditDiff(part.state?.input)
Box {
Surface(
shape = RoundedCornerShape(CHAT_ITEM_GAP + 2.dp),
color = if (failed) {
MaterialTheme.colorScheme.errorContainer
} else {
MaterialTheme.colorScheme.surfaceContainerHigh
},
modifier = Modifier.clickable(onClick = onToggle),
private fun ToolCallsChip(
count: Int,
running: Boolean,
failed: Boolean,
long: Boolean,
expanded: Boolean,
onToggle: () -> Unit,
) {
Surface(
shape = RoundedCornerShape(CHAT_ITEM_GAP + 2.dp),
color = if (failed) {
MaterialTheme.colorScheme.errorContainer
} else {
MaterialTheme.colorScheme.surfaceContainerHigh
},
modifier = Modifier
.semantics { contentDescription = toolCallsAnnouncement(count) }
.clickable(onClick = onToggle),
) {
Row(
verticalAlignment = Alignment.CenterVertically,
horizontalArrangement = Arrangement.spacedBy(CHAT_ITEM_GAP),
modifier = Modifier.padding(horizontal = CHAT_ITEM_GAP + 2.dp, vertical = CHAT_ITEM_GAP),
) {
Row(
verticalAlignment = Alignment.CenterVertically,
horizontalArrangement = Arrangement.spacedBy(CHAT_ITEM_GAP),
modifier = Modifier.padding(horizontal = CHAT_ITEM_GAP + 2.dp, vertical = CHAT_ITEM_GAP),
) {
if (running) {
CircularProgressIndicator(Modifier.size(ICON_STATUS_DP), strokeWidth = 2.dp)
} else {
Icon(
imageVector = when {
failed -> Icons.Default.ErrorOutline
diff != null -> Icons.Default.Edit
else -> Icons.Default.Build
},
contentDescription = if (failed) "Failed tool call" else null,
modifier = Modifier.size(16.dp),
tint = if (failed) MaterialTheme.colorScheme.error else MaterialTheme.colorScheme.onSurfaceVariant,
)
}
Text(
text = diff?.file?.takeIf { it.isNotBlank() } ?: toolLabel(part),
style = MaterialTheme.typography.labelMedium,
maxLines = 1,
overflow = TextOverflow.Ellipsis,
)
diff?.let { DiffStat(it) }
toolDurationMillis(part)?.let {
Text(formatToolDuration(it), style = MaterialTheme.typography.labelMedium, color = MaterialTheme.colorScheme.onSurfaceVariant)
}
if (running) {
CircularProgressIndicator(Modifier.size(ICON_STATUS_DP), strokeWidth = 2.dp)
} else {
Icon(
imageVector = if (expanded) Icons.Default.KeyboardArrowUp else Icons.Default.KeyboardArrowDown,
contentDescription = if (expanded) "Collapse tool call" else "Expand tool call",
imageVector = if (failed) Icons.Default.ErrorOutline else Icons.Default.Build,
contentDescription = null,
modifier = Modifier.size(ICON_INLINE_DP),
tint = if (failed) MaterialTheme.colorScheme.error else MaterialTheme.colorScheme.onSurfaceVariant,
)
}
}
if (expanded) {
Popup(
popupPositionProvider = rememberToolPanelPositionProvider(),
onDismissRequest = onDismiss,
properties = PopupProperties(focusable = true),
) {
ToolPanel(part = part, onDismiss = onDismiss)
Text(
text = toolCallsLabel(count),
style = MaterialTheme.typography.labelMedium,
maxLines = 1,
overflow = TextOverflow.Ellipsis,
)
if (long) {
Text(
text = "· long",
style = MaterialTheme.typography.labelMedium,
color = MaterialTheme.colorScheme.error,
)
}
Icon(
imageVector = if (expanded) Icons.Default.KeyboardArrowUp else Icons.Default.KeyboardArrowDown,
contentDescription = if (expanded) "Hide tool calls" else "Show tool calls",
modifier = Modifier.size(ICON_INLINE_DP),
)
}
}
}
@@ -198,17 +210,15 @@ private fun rememberToolPanelPositionProvider(): PopupPositionProvider {
}
/**
* The expanded panel of one tool call, mirroring the reference panel
* (`_buildPanel`, research §1.3): a header naming the call and its duration with
* a close X, then the pretty-printed INPUT and the monospace OUTPUT (or the
* synthesised edit diff). Both content blocks carry the shared whole-block copy
* action from [CodeBlock]. A failed call always renders its error text in an
* ERROR block — even an edit-family call, whose diff is still shown alongside it.
* The turn's whole tool history as one panel, grown from the single-call panel: a
* header naming the count with a close X, then one collapsible row per call in
* the order the agent made them. The list is bounded ([TOOL_PANEL_MAX_HEIGHT_DP])
* and scrolls inside the panel, so a turn with many calls neither grows the popup
* past the window nor moves the transcript behind it.
*/
@Composable
private fun ToolPanel(part: ChatPart, onDismiss: () -> Unit, modifier: Modifier = Modifier) {
val diff = toolEditDiff(part.state?.input)
val failed = toolFailed(part.state)
private fun ToolPanel(calls: List<ChatPart>, onDismiss: () -> Unit, modifier: Modifier = Modifier) {
var expandedIds by remember { mutableStateOf<Set<String>>(emptySet()) }
Surface(
shape = RoundedCornerShape(CHAT_ITEM_GAP),
color = MaterialTheme.colorScheme.surfaceContainer,
@@ -221,35 +231,110 @@ private fun ToolPanel(part: ChatPart, onDismiss: () -> Unit, modifier: Modifier
) {
Row(verticalAlignment = Alignment.CenterVertically, horizontalArrangement = Arrangement.spacedBy(CHAT_ITEM_GAP)) {
Icon(
imageVector = if (failed) Icons.Default.ErrorOutline else Icons.Default.Build,
imageVector = Icons.Default.Build,
contentDescription = null,
modifier = Modifier.size(16.dp),
tint = if (failed) MaterialTheme.colorScheme.error else MaterialTheme.colorScheme.primary,
modifier = Modifier.size(ICON_INLINE_DP),
tint = MaterialTheme.colorScheme.primary,
)
Text(
text = diff?.file?.takeIf { it.isNotBlank() } ?: toolLabel(part),
text = toolCallsLabel(calls.size),
style = MaterialTheme.typography.labelLarge,
maxLines = 1,
overflow = TextOverflow.Ellipsis,
modifier = Modifier.weight(1f),
)
diff?.let { DiffStat(it) }
toolDurationMillis(part)?.let {
Text(formatToolDuration(it), style = MaterialTheme.typography.labelMedium, color = MaterialTheme.colorScheme.onSurfaceVariant)
}
IconButton(onClick = onDismiss, modifier = Modifier.size(ICON_BUTTON_SMALL_DP)) {
Icon(Icons.Default.Close, contentDescription = "Close tool call", modifier = Modifier.size(16.dp))
Icon(Icons.Default.Close, contentDescription = "Close tool calls", modifier = Modifier.size(ICON_INLINE_DP))
}
}
if (diff != null) {
DiffView(diff)
} else {
CodeBlock(label = "INPUT", text = prettyToolInput(part.state?.input))
}
if (failed) {
CodeBlock(label = "ERROR", text = toolErrorText(part.state))
} else if (diff == null) {
CodeBlock(label = "OUTPUT", text = toolOutputText(part.state))
Column(
verticalArrangement = Arrangement.spacedBy(CHAT_ITEM_GAP),
modifier = Modifier
.fillMaxWidth()
.heightIn(max = TOOL_PANEL_MAX_HEIGHT_DP)
.verticalScroll(rememberScrollState()),
) {
calls.forEach { part ->
val row = toolCallRow(part)
ToolCallRowItem(
row = row,
expanded = part.id in expandedIds,
onToggle = {
expandedIds = if (part.id in expandedIds) expandedIds - part.id else expandedIds + part.id
},
)
}
}
}
}
}
/**
* One call inside the panel: a header naming the call, its `+N -M` stat for an
* edit and its duration, and — expanded in place — the pretty-printed INPUT, the
* synthesised diff for an edit-family call, then the ERROR block for a failed
* call or the OUTPUT (or `(no output)`) for a settled one. Rows expand and
* collapse independently, so comparing two calls never loses the other's detail.
*/
@Composable
private fun ToolCallRowItem(row: ToolCallRow, expanded: Boolean, onToggle: () -> Unit) {
Column(Modifier.fillMaxWidth()) {
Row(
verticalAlignment = Alignment.CenterVertically,
horizontalArrangement = Arrangement.spacedBy(CHAT_ITEM_GAP),
modifier = Modifier
.fillMaxWidth()
.minimumInteractiveComponentSize()
.clickable(role = Role.Button, onClick = onToggle)
.padding(vertical = CHAT_ITEM_GAP),
) {
Icon(
imageVector = when {
row.error != null -> Icons.Default.ErrorOutline
row.diff != null -> Icons.Default.Edit
else -> Icons.Default.Build
},
contentDescription = if (row.error != null) "Failed tool call" else null,
modifier = Modifier.size(ICON_INLINE_DP),
tint = if (row.error != null) MaterialTheme.colorScheme.error else MaterialTheme.colorScheme.onSurfaceVariant,
)
Column(Modifier.weight(1f)) {
Text(
text = row.name,
style = MaterialTheme.typography.labelMedium,
maxLines = 1,
overflow = TextOverflow.Ellipsis,
)
row.detail?.let {
Text(
text = it,
style = MaterialTheme.typography.labelSmall,
color = MaterialTheme.colorScheme.onSurfaceVariant,
maxLines = 1,
overflow = TextOverflow.Ellipsis,
)
}
}
row.diff?.let { DiffStat(it) }
row.duration?.let {
Text(it, style = MaterialTheme.typography.labelMedium, color = MaterialTheme.colorScheme.onSurfaceVariant)
}
Icon(
imageVector = if (expanded) Icons.Default.KeyboardArrowUp else Icons.Default.KeyboardArrowDown,
contentDescription = if (expanded) "Collapse ${row.name}" else "Expand ${row.name}",
modifier = Modifier.size(ICON_INLINE_DP),
)
}
if (expanded) {
Column(
verticalArrangement = Arrangement.spacedBy(CHAT_ITEM_GAP),
modifier = Modifier.fillMaxWidth().padding(start = ICON_INLINE_DP + CHAT_ITEM_GAP, bottom = CHAT_ITEM_GAP),
) {
CodeBlock(label = "INPUT", text = row.input)
row.diff?.let { DiffView(it) }
if (row.error != null) {
CodeBlock(label = "ERROR", text = row.error)
} else {
CodeBlock(label = "OUTPUT", text = row.output)
}
}
}
}
@@ -195,6 +195,7 @@ internal fun ChatTranscript(
reasoning = reasoning,
streaming = row.key == streamingTurnKey,
runningSince = thread.runningSince,
toolStartedAt = thread.toolStartedAt,
onToggleReasoning = onToggleReasoning,
)
}
@@ -434,10 +435,11 @@ private fun UserMessageItem(message: ChatMessage) {
/**
* One agent reply as a single bubble: the turn's one collapsible thinking block
* first, then its tool-call chips in order, then its answer text. A turn's
* first, then its one tool-call summary chip, then its answer text. A turn's
* reasoning and answer are never mixed — the fold in [ChatTurn] keeps them
* separate, so the answer is never drawn inside the thinking block — and the
* whole reply is one item however many model steps produced it.
* whole reply is one item however many model steps produced it. The individual
* calls live inside the chip's popup, not in the transcript.
*/
@Composable
private fun AgentTurnItem(
@@ -445,15 +447,17 @@ private fun AgentTurnItem(
reasoning: ReasoningExpansion,
streaming: Boolean,
runningSince: Long,
toolStartedAt: Map<String, Long>,
onToggleReasoning: (String) -> Unit,
) {
// The turn's own timing for the collapsed thinking line: a streaming turn
// counts live from its own start (the same clock the running indicator uses,
// so the two never disagree), a settled turn shows the span its parts' timing
// envelopes report, and a turn with neither shows no duration rather than a
// guessed one.
// guessed one. The same tick ages the summary chip's running calls, so its
// "long" mark never comes from a stale clock.
val now = if (streaming && runningSince > 0) rememberElapsedNow(runningSince) else 0L
val elapsedMillis: Long? = if (streaming && runningSince > 0) {
val now = rememberElapsedNow(runningSince)
if (runningSince < now) now - runningSince else 0L
} else {
turn.elapsedMillis
@@ -477,7 +481,15 @@ private fun AgentTurnItem(
elapsedMillis = elapsedMillis,
onToggle = { onToggleReasoning(turnKey) },
)
if (turn.tools.isNotEmpty()) ToolCalls(turn.tools)
if (turn.tools.isNotEmpty()) {
ToolCalls(
parts = turn.tools,
live = streaming,
elapsedMillis = { part ->
runningToolElapsedMillis(part, toolStartedAt[part.id] ?: 0L, now)
},
)
}
if (turn.answer.isNotBlank()) MarkdownText(turn.answer)
}
}
@@ -235,6 +235,13 @@ internal val ICON_BUTTON_SMALL_DP = 28.dp
/** The floating tool panel's maximum width; it floats over the transcript. */
internal val TOOL_PANEL_MAX_WIDTH_DP = 420.dp
/**
* How tall the tool panel's list of calls may grow before it scrolls inside the
* panel. Bounded so a turn with many calls never grows the popup past the window
* or pushes its own placement above the screen.
*/
internal val TOOL_PANEL_MAX_HEIGHT_DP = 440.dp
/** A chat row's title column stops here so the close action keeps its place. */
internal val CHAT_TITLE_MAX_WIDTH_DP = 160.dp
@@ -179,4 +179,88 @@ class ChatFormatTest {
assertEquals("scratch", sessionTagLabel("scratch"))
assertNull(sessionTagLabel(""))
}
@Test
fun theToolSummaryReadsAsACallCountAndNeverSaysOneTools() {
// The turn's one chip: the singular is a whole word, not a pluralised count.
assertEquals("1 tool called", toolCallsLabel(1))
assertEquals("6 tools called", toolCallsLabel(6))
// The same count is what a screen reader hears, plus how to reach the calls.
assertEquals("1 tool call, expand for details", toolCallsAnnouncement(1))
assertEquals("6 tool calls, expand for details", toolCallsAnnouncement(6))
}
@Test
fun theSummaryCountCountsOnlyTheTurnsToolCalls() {
val parts = listOf(
ChatPart(type = ChatProtocol.PART_TEXT, id = "t1", text = "answer"),
ChatPart(type = ChatProtocol.PART_TOOL, id = "c1", name = "shell"),
ChatPart(type = ChatProtocol.PART_TOOL, id = "c2", name = "grep"),
)
// Text and reasoning ride in the same turn but never inflate the chip.
assertEquals(2, toolCallCount(parts))
assertEquals(0, toolCallCount(parts.filter { it.type == ChatProtocol.PART_TEXT }))
// One call still in flight keeps the chip live.
assertTrue(toolsRunning(listOf(parts[1].copy(state = ChatToolState(status = ChatProtocol.TOOL_RUNNING)))))
assertFalse(toolsRunning(listOf(parts[1].copy(state = ChatToolState(status = ChatProtocol.TOOL_COMPLETED)))))
}
@Test
fun aToolCallRowDerivesItsNameDetailDurationInputAndOutput() {
val part = ChatPart(
type = ChatProtocol.PART_TOOL,
id = "c1",
name = "shell",
state = ChatToolState(
status = ChatProtocol.TOOL_COMPLETED,
input = JsonObject(mapOf("command" to JsonPrimitive("ls"))),
content = listOf(ChatToolContent(type = "text", text = "a.txt")),
),
time = ChatTime(created = 1_000, ran = 1_000, completed = 2_500),
)
val row = toolCallRow(part)
assertEquals("shell", row.name)
assertEquals("ls", row.detail)
assertEquals("1.5s", row.duration)
assertTrue(row.input.contains("\"command\""))
assertEquals("a.txt", row.output)
// A successful non-edit call has neither an error block nor a diff.
assertNull(row.error)
assertNull(row.diff)
}
@Test
fun aFailedEditRowCarriesItsErrorAndSynthesisedDiff() {
val part = ChatPart(
type = ChatProtocol.PART_TOOL,
id = "c2",
name = "edit",
state = ChatToolState(
status = ChatProtocol.TOOL_ERROR,
input = JsonObject(
mapOf(
"filePath" to JsonPrimitive("/tmp/a.txt"),
"oldString" to JsonPrimitive("one"),
"newString" to JsonPrimitive("two"),
),
),
error = ChatError(type = "boom", message = "exploded"),
),
)
val row = toolCallRow(part)
assertEquals("edit", row.name)
// The ERROR block wins over any output, and the diff is still shown beside it.
assertEquals("exploded", row.error)
assertEquals("/tmp/a.txt", row.diff?.file)
assertEquals(1, row.diff?.added)
assertEquals(1, row.diff?.removed)
// A call with no completion time shows no duration rather than a guess.
assertNull(row.duration)
}
@Test
fun aToolCallRowKeepsTheRealNameAndOnlyFallsBackWhenNoneWasEverSeen() {
assertEquals("shell", toolCallRow(ChatPart(type = ChatProtocol.PART_TOOL, id = "c1", name = "shell")).name)
assertEquals("tool", toolCallRow(ChatPart(type = ChatProtocol.PART_TOOL, id = "c2")).name)
}
}
+48
View File
@@ -0,0 +1,48 @@
# Chat transcript: one tool-call summary chip per agent turn
## What this item changed
`android/` only — the chat transcript's tool-call surface.
- **One chip per turn.** `ui/chat/ChatToolCalls.kt` no longer renders a chip per
call in a `FlowRow`; `ToolCalls` renders a single summary chip reading
`N tools called` (`1 tool called` for one), the transcript's only tool-call
surface. The individual calls are not chips any more.
- **Live summary.** The chip shows the existing spinner while any call is in
flight and a `· long` mark while a running call has passed
`TOOL_LONG_RUNNING_MILLIS`; both are gated on the turn still streaming
(`AgentTurnItem` passes `live = streaming`), so a settled turn never shows the
spinner or a long mark for a stored call whose status is stale.
- **Popup list.** Tapping the chip opens `ToolPanel` as a bounded
(`TOOL_PANEL_MAX_HEIGHT_DP`), scrollable list of one collapsible row per call in
the order the agent made them. A row shows the call's name, its one-line detail,
its `+N -M` stat for an edit and its duration, and expands in place to the
pretty-printed INPUT, the synthesised diff for an `edit`-family call, then the
ERROR block for a failed call or the OUTPUT (or `(no output)`). Rows expand and
collapse independently, and the popup still does not re-flow or scroll the
transcript.
- **Pure logic.** `ui/chat/ChatFormat.kt` adds `toolCallsLabel`,
`toolCallsAnnouncement` (the chip's accessibility line, e.g. "6 tool calls,
expand for details"), `toolCallCount`, `toolsRunning` and `toolCallRow` (the
per-row detail derivation), with cases added to `ChatFormatTest`.
- **Geometry.** `ui/theme/Dimens.kt` adds `TOOL_PANEL_MAX_HEIGHT_DP`; no raw
literal was introduced, so `DimensTest` is unchanged.
## Shared documents this change makes stale
`android/README.md` describes the surface this item changed. The note is written
here rather than edited in place; the `docs-fold` writer should update it.
- `android/README.md` — the section **"Tool-call chips and the expandable panel
(16 Oct 2026)"** and its **Grouped chips** bullet ("A run of consecutive `tool`
parts in one assistant turn is now one `FlowRow` of compact chips") describe the
per-call chips this item removed. It should say the turn renders one summary
chip (`N tools called`) as the transcript's only tool-call surface, and that
tapping it opens one popup listing every call of the turn as independently
collapsible rows, with the chip's spinner/`long` marks gated on the turn still
streaming. The **Overlay panel, scroll preserved** bullet stays true.
- `docs/chat-verification.md:140` — the "Conversation — tool chip expanding" row
names `ChatToolCalls.kt` and the per-chip panel; it should name the summary chip
and the per-turn popup list, and the pure tests now include the label, count,
announcement and per-row derivations in `ChatFormatTest` (the placement maths in
`ToolPanelPositionProviderTest` is unchanged).