fix(ui): thinking indicator never clears on a failed task

[arjun-scale]

Problem

Opening a task whose turn failed shows "Thinking ..." forever. The error toast fires, but the thread still reads as though the agent were working. It happens in two ways:

• the task fails during the current session, and
• a previously-failed task is opened or reloaded, with no client-local error state at all.

Screen.Recording.2026-08-24.at.2.48.31.PM.mov

Cause

shouldShowThinkingForLastPair decided purely from this client's in-session RPC state and the message list. On a cold load neither mentions the failure: rpcStatus starts 'idle', the subscription's first onMessagesChange writes 'success', and a failed turn leaves a user message with no agent message — so the check fell into its "waiting for first response" branch and returned true indefinitely.

The UI already knew the task had failed; the indicator never consulted it.

Fix

Consult the task's own status, the authoritative signal available in both cases:

// Only a running task can be working on a reply.
if (isTaskEnded(task?.status)) return false;

prompt-input.tsx already made the same RUNNING comparison to decide whether sending is allowed, so the rule is extracted to isTaskEnded in lib/task-utils.ts and shared rather than spelled two ways. Comparing against RUNNING also covers every settled status — including DELETED — without enumerating them.

Testing

components/task-messages/task-messages.test.tsx renders the cold-load shape that was broken: a user message, no reply, and rpcStatus: 'success' as the subscription leaves it. Asserts the indicator shows on RUNNING and hides on every ended status. Removing the fix fails all five hide cases.

isTaskEnded is unit-tested for running, ended, and unknown statuses.

Also verified manually against a live agent locally whose turns fail — the indicator clears both when the task fails in-session and on a cold reload.

97 tests pass. npm run typecheck, eslint, and prettier --check clean.

🤖 Generated with Claude Code

Greptile Summary

The PR centralizes task-ended detection and uses authoritative task status to clear the thinking indicator after terminal outcomes.

• Adds shared isTaskEnded handling to the message view and prompt input.
• Adds coverage for running and terminal task statuses.

Confidence Score: 5/5

The PR appears safe to merge because no blocking failure remains within the eligible follow-up review scope.

No blocking failure remains.

Important Files Changed

Filename	Overview
agentex-ui/lib/task-utils.ts	Adds the shared task-ended predicate used by task workspace components.
agentex-ui/components/task-messages/task-messages.tsx	Uses authoritative task status to suppress the thinking indicator once a task has ended.
agentex-ui/components/primary-content/prompt-input.tsx	Replaces its inline status comparison with the shared task-ended helper.
agentex-ui/components/task-messages/task-messages.test.tsx	Covers thinking-indicator behavior for running and terminal cold-loaded tasks.
agentex-ui/lib/task-utils.test.ts	Adds unit coverage for known, terminal, and unavailable task statuses.

_{Reviews (2): Last reviewed commit: "fix(ui): thinking indicator never clears..." | Re-trigger Greptile}

Context used (3)

• Knowledge Base — Task lifecycle and execution data
• Knowledge Base — Task messages, events, and streaming
• Knowledge Base — Console task workspace

[arjun-scale]

Thanks for the review! Merging