feat(chat): read-only subagent sessions surface (#187)

.maestro/flows/subagents.yaml

@@ -0,0 +1,104 @@
+appId: cc.agentlabs.opencode
+name: Subagents section - parent detail lists children with live status (#187)
+---
+# Read-only subagent surface (issue #187): a parent detail screen shows a
+# collapsible Subagents section listing its child sessions with a live status
+# badge. AC1 (child list) + AC3-at-render (root list stays roots-only) at the
+# UI layer; the data path is covered by
+# src/lib/session-list.integration.test.ts (subagents block).
+#
+# The subagents flow's mock is started by scripts/run-e2e-flows.sh itself
+# (port 4101, --seed-subagents) before the flows run, so no workflow-file
+# change is needed for it. --seed-subagents pre-populates one parent + two
+# children, all in /mock/project:
+#   seed-parent     "Parent Session" (updated now-60s)
+#   seed-child-done "Done Subagent", parentID seed-parent, agent explore,
+#                   model opencode/deepseek-v4-flash-free, time.idle set
+#   seed-child-idle "Idle Subagent", parentID seed-parent, no agent/model/idle
+#
+# Why this can actually FAIL if #187 regresses: the section loads children via
+# loadChildren() (GET /session pool, filtered client-side on parentID). If the
+# filter, the mount, or the store wiring breaks, "Subagents (2)" never appears
+# or the child rows never render. If the roots filter regresses, the child
+# titles leak onto the list screen — this flow opens the parent from a list
+# that must show "Parent Session" only.
+#
+# Known limit: the LIVE session.created routing (a child created while viewing
+# its parent appears in the section, never in the root list) is NOT covered
+# here — the mock emits no session.created broadcast and no UI creates a
+# child, so there is no way to trigger it from a flow. That branch is verified
+# by inspection; a live check needs a real server with a task-tool subagent.
+# The flow also does not assert the Running badge: the mock never emits a
+# session.status busy for a child, so no child is live-running in this harness.
+# NEWER flow (non-blocking): see scripts/run-e2e-flows.sh.
+
+- launchApp:
+    clearState: true
+- assertVisible:
+    id: "telemetry-consent-card"
+- tapOn:
+    id: "telemetry-decline-button"
+
+- assertVisible:
+    text: "No Connection"
+- tapOn:
+    id: "add-connection-button"
+- tapOn:
+    id: "connect-ip-input"
+- inputText: "127.0.0.1:4101"
+- hideKeyboard
+- tapOn:
+    id: "connect-submit-button"
+
+# 40s margin: see activation-positive.yaml — the connect handshake's own
+# fetches are individually capped at 30s (src/lib/sdk.ts REQUEST_TIMEOUT_MS).
+- extendedWaitUntil:
+    visible:
+      id: "connection-status-dot"
+    timeout: 40000
+- takeScreenshot: subagents-S1_connected
+
+# The parent must appear on the root list; the children must not leak onto it.
+- extendedWaitUntil:
+    visible:
+      text: "Parent Session"
+    timeout: 15000
+- takeScreenshot: subagents-S2_parent_listed
+
+# Open the parent.
+- tapOn:
+    id: "session-item-seed-parent"
+- extendedWaitUntil:
+    visible:
+      id: "chat-message-input"
+    timeout: 15000
+- assertVisible:
+    text: "Parent Session"
+# Confirms GET /session/seed-parent/message actually resolved (empty state,
+# not stuck loading / errored).
+- assertVisible:
+    text: "Start a conversation"
+
+# The section header renders collapsed with the child count once loadChildren
+# resolves (the count starts at 0, so wait for the loaded value).
+- extendedWaitUntil:
+    visible:
+      text: "Subagents (2)"
+    timeout: 15000
+- takeScreenshot: subagents-S3_section_header
+
+# Expand and check both rows: titles, the agent · model line only where the
+# server provides it, and the Done pill for the finished child.
+- tapOn:
+    text: "Subagents (2)"
+- extendedWaitUntil:
+    visible:
+      text: "Done Subagent"
+    timeout: 15000
+- assertVisible:
+    text: "Idle Subagent"
+- assertVisible:
+    text: "explore · opencode/deepseek-v4-flash-free"
+- assertVisible:
+    text: "Done"
+- takeScreenshot: subagents-S4_children_expanded

app/session/[id].tsx

@@ -33,6 +33,7 @@ import {
   ImageAttachments,
   SessionInfo,
   MessageTextSheet,
+  SubagentSection,
   type SlashCommand,
   type Attachment,
 } from "../../src/components/chat"
@@ -783,6 +784,11 @@ export default function SessionScreen() {
           </View>
         )}
 
+        {/* Subagent children (#187) — above the message list, outside the
+            FlatList. Read-only; a load failure shows its own error + retry
+            and never blocks the transcript. */}
+        {id ? <SubagentSection parentId={id} isDark={isDark} /> : null}
+
         {isLoading ? (
           <View style={s.loading}>
             <ActivityIndicator size="large" color={isDark ? "#ffffff" : "#0a0a0a"} />

scripts/run-e2e-flows.sh

@@ -29,16 +29,36 @@ APK="android/app/build/outputs/apk/release/app-release.apk"
 CORE_FLOWS=(activation-positive activation-negative-401)
 # demo.yaml (offline "Try a demo" mode) needs no mock server — it never
 # connects — so it's safe to run alongside the others here unmodified.
-NEWER_FLOWS=(directory-picker all-sessions variant-picker diff-scroll demo)
+NEWER_FLOWS=(directory-picker all-sessions variant-picker diff-scroll demo subagents)
 mkdir -p "$ROOT/artifacts/screenshots" "$ROOT/artifacts/diag"
 
 echo "== installing APK =="
 adb install "$APK"
 
 echo "== forwarding mock ports into the emulator (adb reverse) =="
-for p in 4096 4097 4098 4099 4100; do adb reverse "tcp:$p" "tcp:$p"; done
+for p in 4096 4097 4098 4099 4100 4101; do adb reverse "tcp:$p" "tcp:$p"; done
 adb reverse --list
 
+# The subagents flow needs its own mock instance (port 4101, --seed-subagents:
+# one parent + two children, for subagents.yaml / issue #187). It starts here,
+# not in activation-e2e.yml, so this flow needs no workflow-file change: the
+# workflow only starts ports 4096-4100. Same node invocation the workflow uses
+# for the other mocks; wait until its health endpoint answers.
+echo "== starting subagents mock on 4101 =="
+nohup node tests/fixtures/mock-opencode-server.ts --port 4101 --seed-subagents > /tmp/mock-4101.log 2>&1 &
+for i in $(seq 1 30); do
+  if curl -sf --connect-timeout 1 -m 3 "http://127.0.0.1:4101/global/health" > /dev/null 2>&1; then
+    echo "subagents mock on 4101 is up"
+    break
+  fi
+  if [ "$i" = "30" ]; then
+    echo "::error::subagents mock server on port 4101 did not come up in 30s"
+    cat /tmp/mock-4101.log || true
+    exit 1
+  fi
+  sleep 1
+done
+
 # Decisive probe: can the EMULATOR actually reach the host mock via 127.0.0.1?
 # If any in-emulator method prints the health JSON, the network path is good
 # and any flow failure is app-side; if none succeed, attribution falls back to
@@ -96,6 +116,7 @@ adb logcat -c || true            # clear, so the captured log is just this run
 dump_diag() {
   echo "== capturing diagnostics =="
   adb logcat -d > "$ROOT/artifacts/diag/logcat.txt" 2>&1 || true
+  cp /tmp/mock-4101.log "$ROOT/artifacts/diag/mock-4101.log" 2>/dev/null || true
   # Maestro writes per-run debug (UI hierarchy + commands) under ~/.maestro/tests
   cp -r "$HOME/.maestro/tests" "$ROOT/artifacts/diag/maestro-tests" 2>/dev/null || true
 }